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 $
70 #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>
89 #include <sys/socket.h>
90 #include <sys/socketvar.h>
91 #include <sys/protosw.h>
92 #include <sys/random.h>
93 #include <sys/in_cksum.h>
96 #include <net/route.h>
98 #include <net/netisr.h>
101 #include <netinet/in.h>
102 #include <netinet/in_systm.h>
103 #include <netinet/ip.h>
104 #include <netinet/ip6.h>
105 #include <netinet/in_pcb.h>
106 #include <netinet6/in6_pcb.h>
107 #include <netinet/in_var.h>
108 #include <netinet/ip_var.h>
109 #include <netinet6/ip6_var.h>
110 #include <netinet/ip_icmp.h>
112 #include <netinet/icmp6.h>
114 #include <netinet/tcp.h>
115 #include <netinet/tcp_fsm.h>
116 #include <netinet/tcp_seq.h>
117 #include <netinet/tcp_timer.h>
118 #include <netinet/tcp_timer2.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>
129 #include <netproto/key/key.h>
131 #include <netinet6/ipsec6.h>
136 #include <netproto/ipsec/ipsec.h>
138 #include <netproto/ipsec/ipsec6.h>
144 #include <machine/smp.h>
146 #include <sys/msgport2.h>
147 #include <sys/mplock2.h>
148 #include <net/netmsg2.h>
150 #if !defined(KTR_TCP)
151 #define KTR_TCP KTR_ALL
154 KTR_INFO_MASTER(tcp);
155 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
157 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
158 #define logtcp(name) KTR_LOG(tcp_ ## name)
161 struct inpcbinfo tcbinfo[MAXCPU];
162 struct tcpcbackqhead tcpcbackq[MAXCPU];
164 static struct lwkt_token tcp_port_token =
165 LWKT_TOKEN_INITIALIZER(tcp_port_token);
167 int tcp_mssdflt = TCP_MSS;
168 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
169 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
172 int tcp_v6mssdflt = TCP6_MSS;
173 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
174 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
178 * Minimum MSS we accept and use. This prevents DoS attacks where
179 * we are forced to a ridiculous low MSS like 20 and send hundreds
180 * of packets instead of one. The effect scales with the available
181 * bandwidth and quickly saturates the CPU and network interface
182 * with packet generation and sending. Set to zero to disable MINMSS
183 * checking. This setting prevents us from sending too small packets.
185 int tcp_minmss = TCP_MINMSS;
186 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
187 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
190 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
191 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
192 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
195 int tcp_do_rfc1323 = 1;
196 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
197 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
199 static int tcp_tcbhashsize = 0;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
201 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
203 static int do_tcpdrain = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
205 "Enable tcp_drain routine for extra help when low on mbufs");
207 static int icmp_may_rst = 1;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
209 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
211 static int tcp_isn_reseed_interval = 0;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
213 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
216 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
217 * by default, but with generous values which should allow maximal
218 * bandwidth. In particular, the slop defaults to 50 (5 packets).
220 * The reason for doing this is that the limiter is the only mechanism we
221 * have which seems to do a really good job preventing receiver RX rings
222 * on network interfaces from getting blown out. Even though GigE/10GigE
223 * is supposed to flow control it looks like either it doesn't actually
224 * do it or Open Source drivers do not properly enable it.
226 * People using the limiter to reduce bottlenecks on slower WAN connections
227 * should set the slop to 20 (2 packets).
229 static int tcp_inflight_enable = 1;
230 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
231 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
233 static int tcp_inflight_debug = 0;
234 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
235 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
237 static int tcp_inflight_min = 6144;
238 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
239 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
241 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
243 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
245 static int tcp_inflight_stab = 50;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
247 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
249 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
250 static struct malloc_pipe tcptemp_mpipe;
252 static void tcp_willblock(void);
253 static void tcp_notify (struct inpcb *, int);
255 struct tcp_stats tcpstats_percpu[MAXCPU];
258 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
262 for (cpu = 0; cpu < ncpus; ++cpu) {
263 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
264 sizeof(struct tcp_stats))))
266 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
267 sizeof(struct tcp_stats))))
273 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
274 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
276 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
277 &tcpstat, tcp_stats, "TCP statistics");
281 * Target size of TCP PCB hash tables. Must be a power of two.
283 * Note that this can be overridden by the kernel environment
284 * variable net.inet.tcp.tcbhashsize
287 #define TCBHASHSIZE 512
291 * This is the actual shape of what we allocate using the zone
292 * allocator. Doing it this way allows us to protect both structures
293 * using the same generation count, and also eliminates the overhead
294 * of allocating tcpcbs separately. By hiding the structure here,
295 * we avoid changing most of the rest of the code (although it needs
296 * to be changed, eventually, for greater efficiency).
299 #define ALIGNM1 (ALIGNMENT - 1)
303 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
306 struct tcp_callout inp_tp_rexmt;
307 struct tcp_callout inp_tp_persist;
308 struct tcp_callout inp_tp_keep;
309 struct tcp_callout inp_tp_2msl;
310 struct tcp_callout inp_tp_delack;
311 struct netmsg_tcp_timer inp_tp_timermsg;
322 struct inpcbporthead *porthashbase;
323 struct inpcbinfo *ticb;
325 int hashsize = TCBHASHSIZE;
329 * note: tcptemp is used for keepalives, and it is ok for an
330 * allocation to fail so do not specify MPF_INT.
332 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
333 25, -1, 0, NULL, NULL, NULL);
335 tcp_delacktime = TCPTV_DELACK;
336 tcp_keepinit = TCPTV_KEEP_INIT;
337 tcp_keepidle = TCPTV_KEEP_IDLE;
338 tcp_keepintvl = TCPTV_KEEPINTVL;
339 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
341 tcp_rexmit_min = TCPTV_MIN;
342 tcp_rexmit_slop = TCPTV_CPU_VAR;
344 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
345 if (!powerof2(hashsize)) {
346 kprintf("WARNING: TCB hash size not a power of 2\n");
347 hashsize = 512; /* safe default */
349 tcp_tcbhashsize = hashsize;
350 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
352 for (cpu = 0; cpu < ncpus2; cpu++) {
353 ticb = &tcbinfo[cpu];
354 in_pcbinfo_init(ticb);
356 ticb->hashbase = hashinit(hashsize, M_PCB,
358 ticb->porthashbase = porthashbase;
359 ticb->porthashmask = porthashmask;
360 ticb->porttoken = &tcp_port_token;
362 ticb->porthashbase = hashinit(hashsize, M_PCB,
363 &ticb->porthashmask);
365 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
366 &ticb->wildcardhashmask);
367 ticb->ipi_size = sizeof(struct inp_tp);
368 TAILQ_INIT(&tcpcbackq[cpu]);
371 tcp_reass_maxseg = nmbclusters / 16;
372 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
375 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
377 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
379 if (max_protohdr < TCP_MINPROTOHDR)
380 max_protohdr = TCP_MINPROTOHDR;
381 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
383 #undef TCP_MINPROTOHDR
386 * Initialize TCP statistics counters for each CPU.
389 for (cpu = 0; cpu < ncpus; ++cpu) {
390 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
393 bzero(&tcpstat, sizeof(struct tcp_stats));
397 netisr_register_rollup(tcp_willblock);
404 int cpu = mycpu->gd_cpuid;
406 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
407 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
408 tp->t_flags &= ~TF_ONOUTPUTQ;
409 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
415 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
416 * tcp_template used to store this data in mbufs, but we now recopy it out
417 * of the tcpcb each time to conserve mbufs.
420 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
422 struct inpcb *inp = tp->t_inpcb;
423 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
426 if (inp->inp_vflag & INP_IPV6) {
429 ip6 = (struct ip6_hdr *)ip_ptr;
430 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
431 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
432 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
433 (IPV6_VERSION & IPV6_VERSION_MASK);
434 ip6->ip6_nxt = IPPROTO_TCP;
435 ip6->ip6_plen = sizeof(struct tcphdr);
436 ip6->ip6_src = inp->in6p_laddr;
437 ip6->ip6_dst = inp->in6p_faddr;
442 struct ip *ip = (struct ip *) ip_ptr;
444 ip->ip_vhl = IP_VHL_BORING;
451 ip->ip_p = IPPROTO_TCP;
452 ip->ip_src = inp->inp_laddr;
453 ip->ip_dst = inp->inp_faddr;
454 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
456 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
459 tcp_hdr->th_sport = inp->inp_lport;
460 tcp_hdr->th_dport = inp->inp_fport;
465 tcp_hdr->th_flags = 0;
471 * Create template to be used to send tcp packets on a connection.
472 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
473 * use for this function is in keepalives, which use tcp_respond.
476 tcp_maketemplate(struct tcpcb *tp)
480 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
482 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
487 tcp_freetemplate(struct tcptemp *tmp)
489 mpipe_free(&tcptemp_mpipe, tmp);
493 * Send a single message to the TCP at address specified by
494 * the given TCP/IP header. If m == NULL, then we make a copy
495 * of the tcpiphdr at ti and send directly to the addressed host.
496 * This is used to force keep alive messages out using the TCP
497 * template for a connection. If flags are given then we send
498 * a message back to the TCP which originated the * segment ti,
499 * and discard the mbuf containing it and any other attached mbufs.
501 * In any case the ack and sequence number of the transmitted
502 * segment are as specified by the parameters.
504 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
507 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
508 tcp_seq ack, tcp_seq seq, int flags)
512 struct route *ro = NULL;
514 struct ip *ip = ipgen;
517 struct route_in6 *ro6 = NULL;
518 struct route_in6 sro6;
519 struct ip6_hdr *ip6 = ipgen;
520 boolean_t use_tmpro = TRUE;
522 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
524 const boolean_t isipv6 = FALSE;
528 if (!(flags & TH_RST)) {
529 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
532 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
533 win = (long)TCP_MAXWIN << tp->rcv_scale;
536 * Don't use the route cache of a listen socket,
537 * it is not MPSAFE; use temporary route cache.
539 if (tp->t_state != TCPS_LISTEN) {
541 ro6 = &tp->t_inpcb->in6p_route;
543 ro = &tp->t_inpcb->inp_route;
550 bzero(ro6, sizeof *ro6);
553 bzero(ro, sizeof *ro);
557 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
561 m->m_data += max_linkhdr;
563 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
564 ip6 = mtod(m, struct ip6_hdr *);
565 nth = (struct tcphdr *)(ip6 + 1);
567 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
568 ip = mtod(m, struct ip *);
569 nth = (struct tcphdr *)(ip + 1);
571 bcopy(th, nth, sizeof(struct tcphdr));
576 m->m_data = (caddr_t)ipgen;
577 /* m_len is set later */
579 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
581 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
582 nth = (struct tcphdr *)(ip6 + 1);
584 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
585 nth = (struct tcphdr *)(ip + 1);
589 * this is usually a case when an extension header
590 * exists between the IPv6 header and the
593 nth->th_sport = th->th_sport;
594 nth->th_dport = th->th_dport;
596 xchg(nth->th_dport, nth->th_sport, n_short);
601 ip6->ip6_vfc = IPV6_VERSION;
602 ip6->ip6_nxt = IPPROTO_TCP;
603 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
604 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
606 tlen += sizeof(struct tcpiphdr);
608 ip->ip_ttl = ip_defttl;
611 m->m_pkthdr.len = tlen;
612 m->m_pkthdr.rcvif = NULL;
613 nth->th_seq = htonl(seq);
614 nth->th_ack = htonl(ack);
616 nth->th_off = sizeof(struct tcphdr) >> 2;
617 nth->th_flags = flags;
619 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
621 nth->th_win = htons((u_short)win);
625 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
626 sizeof(struct ip6_hdr),
627 tlen - sizeof(struct ip6_hdr));
628 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
629 (ro6 && ro6->ro_rt) ?
630 ro6->ro_rt->rt_ifp : NULL);
632 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
633 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
634 m->m_pkthdr.csum_flags = CSUM_TCP;
635 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
638 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
639 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
642 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
643 tp ? tp->t_inpcb : NULL);
644 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
649 ipflags |= IP_DEBUGROUTE;
650 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
651 if ((ro == &sro) && (ro->ro_rt != NULL)) {
659 * Create a new TCP control block, making an
660 * empty reassembly queue and hooking it to the argument
661 * protocol control block. The `inp' parameter must have
662 * come from the zone allocator set up in tcp_init().
665 tcp_newtcpcb(struct inpcb *inp)
670 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
672 const boolean_t isipv6 = FALSE;
675 it = (struct inp_tp *)inp;
677 bzero(tp, sizeof(struct tcpcb));
678 LIST_INIT(&tp->t_segq);
679 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
681 /* Set up our timeouts. */
682 tp->tt_rexmt = &it->inp_tp_rexmt;
683 tp->tt_persist = &it->inp_tp_persist;
684 tp->tt_keep = &it->inp_tp_keep;
685 tp->tt_2msl = &it->inp_tp_2msl;
686 tp->tt_delack = &it->inp_tp_delack;
690 * Zero out timer message. We don't create it here,
691 * since the current CPU may not be the owner of this
694 tp->tt_msg = &it->inp_tp_timermsg;
695 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
697 tp->t_keepinit = tcp_keepinit;
698 tp->t_keepidle = tcp_keepidle;
699 tp->t_keepintvl = tcp_keepintvl;
700 tp->t_keepcnt = tcp_keepcnt;
701 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
704 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
705 tp->t_inpcb = inp; /* XXX */
706 tp->t_state = TCPS_CLOSED;
708 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
709 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
710 * reasonable initial retransmit time.
712 tp->t_srtt = TCPTV_SRTTBASE;
714 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
715 tp->t_rttmin = tcp_rexmit_min;
716 tp->t_rxtcur = TCPTV_RTOBASE;
717 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
718 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
719 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
720 tp->t_rcvtime = ticks;
722 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
723 * because the socket may be bound to an IPv6 wildcard address,
724 * which may match an IPv4-mapped IPv6 address.
726 inp->inp_ip_ttl = ip_defttl;
728 tcp_sack_tcpcb_init(tp);
729 return (tp); /* XXX */
733 * Drop a TCP connection, reporting the specified error.
734 * If connection is synchronized, then send a RST to peer.
737 tcp_drop(struct tcpcb *tp, int error)
739 struct socket *so = tp->t_inpcb->inp_socket;
741 if (TCPS_HAVERCVDSYN(tp->t_state)) {
742 tp->t_state = TCPS_CLOSED;
744 tcpstat.tcps_drops++;
746 tcpstat.tcps_conndrops++;
747 if (error == ETIMEDOUT && tp->t_softerror)
748 error = tp->t_softerror;
749 so->so_error = error;
750 return (tcp_close(tp));
755 struct netmsg_listen_detach {
756 struct netmsg_base base;
761 tcp_listen_detach_handler(netmsg_t msg)
763 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
764 struct tcpcb *tp = nmsg->nm_tp;
765 int cpu = mycpuid, nextcpu;
767 if (tp->t_flags & TF_LISTEN)
768 syncache_destroy(tp);
770 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
773 if (nextcpu < ncpus2)
774 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg);
776 lwkt_replymsg(&nmsg->base.lmsg, 0);
782 * Close a TCP control block:
783 * discard all space held by the tcp
784 * discard internet protocol block
785 * wake up any sleepers
788 tcp_close(struct tcpcb *tp)
791 struct inpcb *inp = tp->t_inpcb;
792 struct socket *so = inp->inp_socket;
794 boolean_t dosavessthresh;
796 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
797 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
799 const boolean_t isipv6 = FALSE;
804 * INP_WILDCARD_MP indicates that listen(2) has been called on
805 * this socket. This implies:
806 * - A wildcard inp's hash is replicated for each protocol thread.
807 * - Syncache for this inp grows independently in each protocol
809 * - There is more than one cpu
811 * We have to chain a message to the rest of the protocol threads
812 * to cleanup the wildcard hash and the syncache. The cleanup
813 * in the current protocol thread is defered till the end of this
817 * After cleanup the inp's hash and syncache entries, this inp will
818 * no longer be available to the rest of the protocol threads, so we
819 * are safe to whack the inp in the following code.
821 if (inp->inp_flags & INP_WILDCARD_MP) {
822 struct netmsg_listen_detach nmsg;
824 KKASSERT(so->so_port == cpu_portfn(0));
825 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
826 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
828 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
829 MSGF_PRIORITY, tcp_listen_detach_handler);
831 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0);
833 inp->inp_flags &= ~INP_WILDCARD_MP;
837 KKASSERT(tp->t_state != TCPS_TERMINATING);
838 tp->t_state = TCPS_TERMINATING;
841 * Make sure that all of our timers are stopped before we
842 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
843 * timers are never used. If timer message is never created
844 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
846 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
847 tcp_callout_stop(tp, tp->tt_rexmt);
848 tcp_callout_stop(tp, tp->tt_persist);
849 tcp_callout_stop(tp, tp->tt_keep);
850 tcp_callout_stop(tp, tp->tt_2msl);
851 tcp_callout_stop(tp, tp->tt_delack);
854 if (tp->t_flags & TF_ONOUTPUTQ) {
855 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
856 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
857 tp->t_flags &= ~TF_ONOUTPUTQ;
861 * If we got enough samples through the srtt filter,
862 * save the rtt and rttvar in the routing entry.
863 * 'Enough' is arbitrarily defined as the 16 samples.
864 * 16 samples is enough for the srtt filter to converge
865 * to within 5% of the correct value; fewer samples and
866 * we could save a very bogus rtt.
868 * Don't update the default route's characteristics and don't
869 * update anything that the user "locked".
871 if (tp->t_rttupdated >= 16) {
875 struct sockaddr_in6 *sin6;
877 if ((rt = inp->in6p_route.ro_rt) == NULL)
879 sin6 = (struct sockaddr_in6 *)rt_key(rt);
880 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
883 if ((rt = inp->inp_route.ro_rt) == NULL ||
884 ((struct sockaddr_in *)rt_key(rt))->
885 sin_addr.s_addr == INADDR_ANY)
888 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
889 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
890 if (rt->rt_rmx.rmx_rtt && i)
892 * filter this update to half the old & half
893 * the new values, converting scale.
894 * See route.h and tcp_var.h for a
895 * description of the scaling constants.
898 (rt->rt_rmx.rmx_rtt + i) / 2;
900 rt->rt_rmx.rmx_rtt = i;
901 tcpstat.tcps_cachedrtt++;
903 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
905 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
906 if (rt->rt_rmx.rmx_rttvar && i)
907 rt->rt_rmx.rmx_rttvar =
908 (rt->rt_rmx.rmx_rttvar + i) / 2;
910 rt->rt_rmx.rmx_rttvar = i;
911 tcpstat.tcps_cachedrttvar++;
914 * The old comment here said:
915 * update the pipelimit (ssthresh) if it has been updated
916 * already or if a pipesize was specified & the threshhold
917 * got below half the pipesize. I.e., wait for bad news
918 * before we start updating, then update on both good
921 * But we want to save the ssthresh even if no pipesize is
922 * specified explicitly in the route, because such
923 * connections still have an implicit pipesize specified
924 * by the global tcp_sendspace. In the absence of a reliable
925 * way to calculate the pipesize, it will have to do.
927 i = tp->snd_ssthresh;
928 if (rt->rt_rmx.rmx_sendpipe != 0)
929 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
931 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
932 if (dosavessthresh ||
933 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
934 (rt->rt_rmx.rmx_ssthresh != 0))) {
936 * convert the limit from user data bytes to
937 * packets then to packet data bytes.
939 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
944 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
945 sizeof(struct tcpiphdr));
946 if (rt->rt_rmx.rmx_ssthresh)
947 rt->rt_rmx.rmx_ssthresh =
948 (rt->rt_rmx.rmx_ssthresh + i) / 2;
950 rt->rt_rmx.rmx_ssthresh = i;
951 tcpstat.tcps_cachedssthresh++;
956 /* free the reassembly queue, if any */
957 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
958 LIST_REMOVE(q, tqe_q);
961 atomic_add_int(&tcp_reass_qsize, -1);
963 /* throw away SACK blocks in scoreboard*/
965 tcp_sack_cleanup(&tp->scb);
967 inp->inp_ppcb = NULL;
968 soisdisconnected(so);
969 /* note: pcb detached later on */
971 tcp_destroy_timermsg(tp);
973 if (tp->t_flags & TF_LISTEN)
974 syncache_destroy(tp);
978 * pcbdetach removes any wildcard hash entry on the current CPU.
987 tcpstat.tcps_closed++;
992 tcp_drain_oncpu(struct inpcbhead *head)
994 struct inpcb *marker;
997 struct tseg_qent *te;
1000 * Allows us to block while running the list
1002 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1003 marker->inp_flags |= INP_PLACEMARKER;
1004 LIST_INSERT_HEAD(head, marker, inp_list);
1006 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1007 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1008 (tcpb = intotcpcb(inpb)) != NULL &&
1009 (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1010 LIST_REMOVE(te, tqe_q);
1013 atomic_add_int(&tcp_reass_qsize, -1);
1016 LIST_REMOVE(marker, inp_list);
1017 LIST_INSERT_AFTER(inpb, marker, inp_list);
1020 LIST_REMOVE(marker, inp_list);
1021 kfree(marker, M_TEMP);
1025 struct netmsg_tcp_drain {
1026 struct netmsg_base base;
1027 struct inpcbhead *nm_head;
1031 tcp_drain_handler(netmsg_t msg)
1033 struct netmsg_tcp_drain *nm = (void *)msg;
1035 tcp_drain_oncpu(nm->nm_head);
1036 lwkt_replymsg(&nm->base.lmsg, 0);
1051 * Walk the tcpbs, if existing, and flush the reassembly queue,
1052 * if there is one...
1053 * XXX: The "Net/3" implementation doesn't imply that the TCP
1054 * reassembly queue should be flushed, but in a situation
1055 * where we're really low on mbufs, this is potentially
1059 for (cpu = 0; cpu < ncpus2; cpu++) {
1060 struct netmsg_tcp_drain *nm;
1062 if (cpu == mycpu->gd_cpuid) {
1063 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1065 nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1066 M_LWKTMSG, M_NOWAIT);
1069 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1070 0, tcp_drain_handler);
1071 nm->nm_head = &tcbinfo[cpu].pcblisthead;
1072 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.lmsg);
1076 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1081 * Notify a tcp user of an asynchronous error;
1082 * store error as soft error, but wake up user
1083 * (for now, won't do anything until can select for soft error).
1085 * Do not wake up user since there currently is no mechanism for
1086 * reporting soft errors (yet - a kqueue filter may be added).
1089 tcp_notify(struct inpcb *inp, int error)
1091 struct tcpcb *tp = intotcpcb(inp);
1094 * Ignore some errors if we are hooked up.
1095 * If connection hasn't completed, has retransmitted several times,
1096 * and receives a second error, give up now. This is better
1097 * than waiting a long time to establish a connection that
1098 * can never complete.
1100 if (tp->t_state == TCPS_ESTABLISHED &&
1101 (error == EHOSTUNREACH || error == ENETUNREACH ||
1102 error == EHOSTDOWN)) {
1104 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1106 tcp_drop(tp, error);
1108 tp->t_softerror = error;
1110 wakeup(&so->so_timeo);
1117 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1120 struct inpcb *marker;
1129 * The process of preparing the TCB list is too time-consuming and
1130 * resource-intensive to repeat twice on every request.
1132 if (req->oldptr == NULL) {
1133 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1134 gd = globaldata_find(ccpu);
1135 n += tcbinfo[gd->gd_cpuid].ipi_count;
1137 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1141 if (req->newptr != NULL)
1144 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1145 marker->inp_flags |= INP_PLACEMARKER;
1148 * OK, now we're committed to doing something. Run the inpcb list
1149 * for each cpu in the system and construct the output. Use a
1150 * list placemarker to deal with list changes occuring during
1151 * copyout blockages (but otherwise depend on being on the correct
1152 * cpu to avoid races).
1154 origcpu = mycpu->gd_cpuid;
1155 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1161 cpu_id = (origcpu + ccpu) % ncpus;
1162 if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1164 rgd = globaldata_find(cpu_id);
1165 lwkt_setcpu_self(rgd);
1167 n = tcbinfo[cpu_id].ipi_count;
1169 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1171 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1173 * process a snapshot of pcbs, ignoring placemarkers
1174 * and using our own to allow SYSCTL_OUT to block.
1176 LIST_REMOVE(marker, inp_list);
1177 LIST_INSERT_AFTER(inp, marker, inp_list);
1179 if (inp->inp_flags & INP_PLACEMARKER)
1181 if (prison_xinpcb(req->td, inp))
1184 xt.xt_len = sizeof xt;
1185 bcopy(inp, &xt.xt_inp, sizeof *inp);
1186 inp_ppcb = inp->inp_ppcb;
1187 if (inp_ppcb != NULL)
1188 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1190 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1191 if (inp->inp_socket)
1192 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1193 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1197 LIST_REMOVE(marker, inp_list);
1198 if (error == 0 && i < n) {
1199 bzero(&xt, sizeof xt);
1200 xt.xt_len = sizeof xt;
1202 error = SYSCTL_OUT(req, &xt, sizeof xt);
1211 * Make sure we are on the same cpu we were on originally, since
1212 * higher level callers expect this. Also don't pollute caches with
1213 * migrated userland data by (eventually) returning to userland
1214 * on a different cpu.
1216 lwkt_setcpu_self(globaldata_find(origcpu));
1217 kfree(marker, M_TEMP);
1221 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1222 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1225 tcp_getcred(SYSCTL_HANDLER_ARGS)
1227 struct sockaddr_in addrs[2];
1232 error = priv_check(req->td, PRIV_ROOT);
1235 error = SYSCTL_IN(req, addrs, sizeof addrs);
1239 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1240 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1241 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1242 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1243 if (inp == NULL || inp->inp_socket == NULL) {
1247 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1253 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1254 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1258 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1260 struct sockaddr_in6 addrs[2];
1263 boolean_t mapped = FALSE;
1265 error = priv_check(req->td, PRIV_ROOT);
1268 error = SYSCTL_IN(req, addrs, sizeof addrs);
1271 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1272 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1279 inp = in_pcblookup_hash(&tcbinfo[0],
1280 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1282 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1286 inp = in6_pcblookup_hash(&tcbinfo[0],
1287 &addrs[1].sin6_addr, addrs[1].sin6_port,
1288 &addrs[0].sin6_addr, addrs[0].sin6_port,
1291 if (inp == NULL || inp->inp_socket == NULL) {
1295 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1301 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1303 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1306 struct netmsg_tcp_notify {
1307 struct netmsg_base base;
1308 void (*nm_notify)(struct inpcb *, int);
1309 struct in_addr nm_faddr;
1314 tcp_notifyall_oncpu(netmsg_t msg)
1316 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1319 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1320 nm->nm_arg, nm->nm_notify);
1322 nextcpu = mycpuid + 1;
1323 if (nextcpu < ncpus2)
1324 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg);
1326 lwkt_replymsg(&nm->base.lmsg, 0);
1330 tcp_ctlinput(netmsg_t msg)
1332 int cmd = msg->ctlinput.nm_cmd;
1333 struct sockaddr *sa = msg->ctlinput.nm_arg;
1334 struct ip *ip = msg->ctlinput.nm_extra;
1336 struct in_addr faddr;
1339 void (*notify)(struct inpcb *, int) = tcp_notify;
1343 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1347 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1348 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1351 arg = inetctlerrmap[cmd];
1352 if (cmd == PRC_QUENCH) {
1353 notify = tcp_quench;
1354 } else if (icmp_may_rst &&
1355 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1356 cmd == PRC_UNREACH_PORT ||
1357 cmd == PRC_TIMXCEED_INTRANS) &&
1359 notify = tcp_drop_syn_sent;
1360 } else if (cmd == PRC_MSGSIZE) {
1361 struct icmp *icmp = (struct icmp *)
1362 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1364 arg = ntohs(icmp->icmp_nextmtu);
1365 notify = tcp_mtudisc;
1366 } else if (PRC_IS_REDIRECT(cmd)) {
1368 notify = in_rtchange;
1369 } else if (cmd == PRC_HOSTDEAD) {
1375 th = (struct tcphdr *)((caddr_t)ip +
1376 (IP_VHL_HL(ip->ip_vhl) << 2));
1377 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1378 ip->ip_src.s_addr, th->th_sport);
1379 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1380 ip->ip_src, th->th_sport, 0, NULL);
1381 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1382 icmpseq = htonl(th->th_seq);
1383 tp = intotcpcb(inp);
1384 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1385 SEQ_LT(icmpseq, tp->snd_max))
1386 (*notify)(inp, arg);
1388 struct in_conninfo inc;
1390 inc.inc_fport = th->th_dport;
1391 inc.inc_lport = th->th_sport;
1392 inc.inc_faddr = faddr;
1393 inc.inc_laddr = ip->ip_src;
1397 syncache_unreach(&inc, th);
1401 struct netmsg_tcp_notify *nm;
1403 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1404 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1405 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1406 0, tcp_notifyall_oncpu);
1407 nm->nm_faddr = faddr;
1409 nm->nm_notify = notify;
1411 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg);
1414 lwkt_replymsg(&msg->lmsg, 0);
1420 tcp6_ctlinput(netmsg_t msg)
1422 int cmd = msg->ctlinput.nm_cmd;
1423 struct sockaddr *sa = msg->ctlinput.nm_arg;
1424 void *d = msg->ctlinput.nm_extra;
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)) {
1444 if (cmd == PRC_QUENCH)
1445 notify = tcp_quench;
1446 else if (cmd == PRC_MSGSIZE) {
1447 struct ip6ctlparam *ip6cp = d;
1448 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1450 arg = ntohl(icmp6->icmp6_mtu);
1451 notify = tcp_mtudisc;
1452 } else if (!PRC_IS_REDIRECT(cmd) &&
1453 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1457 /* if the parameter is from icmp6, decode it. */
1459 ip6cp = (struct ip6ctlparam *)d;
1461 ip6 = ip6cp->ip6c_ip6;
1462 off = ip6cp->ip6c_off;
1463 sa6_src = ip6cp->ip6c_src;
1467 off = 0; /* fool gcc */
1472 struct in_conninfo inc;
1474 * XXX: We assume that when IPV6 is non NULL,
1475 * M and OFF are valid.
1478 /* check if we can safely examine src and dst ports */
1479 if (m->m_pkthdr.len < off + sizeof *thp)
1482 bzero(&th, sizeof th);
1483 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1485 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1486 (struct sockaddr *)ip6cp->ip6c_src,
1487 th.th_sport, cmd, arg, notify);
1489 inc.inc_fport = th.th_dport;
1490 inc.inc_lport = th.th_sport;
1491 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1492 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1494 syncache_unreach(&inc, &th);
1496 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1497 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1500 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1506 * Following is where TCP initial sequence number generation occurs.
1508 * There are two places where we must use initial sequence numbers:
1509 * 1. In SYN-ACK packets.
1510 * 2. In SYN packets.
1512 * All ISNs for SYN-ACK packets are generated by the syncache. See
1513 * tcp_syncache.c for details.
1515 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1516 * depends on this property. In addition, these ISNs should be
1517 * unguessable so as to prevent connection hijacking. To satisfy
1518 * the requirements of this situation, the algorithm outlined in
1519 * RFC 1948 is used to generate sequence numbers.
1521 * Implementation details:
1523 * Time is based off the system timer, and is corrected so that it
1524 * increases by one megabyte per second. This allows for proper
1525 * recycling on high speed LANs while still leaving over an hour
1528 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1529 * between seeding of isn_secret. This is normally set to zero,
1530 * as reseeding should not be necessary.
1534 #define ISN_BYTES_PER_SECOND 1048576
1536 u_char isn_secret[32];
1537 int isn_last_reseed;
1541 tcp_new_isn(struct tcpcb *tp)
1543 u_int32_t md5_buffer[4];
1546 /* Seed if this is the first use, reseed if requested. */
1547 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1548 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1550 read_random_unlimited(&isn_secret, sizeof isn_secret);
1551 isn_last_reseed = ticks;
1554 /* Compute the md5 hash and return the ISN. */
1556 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1557 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1559 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1560 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1561 sizeof(struct in6_addr));
1562 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1563 sizeof(struct in6_addr));
1567 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1568 sizeof(struct in_addr));
1569 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1570 sizeof(struct in_addr));
1572 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1573 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1574 new_isn = (tcp_seq) md5_buffer[0];
1575 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1580 * When a source quench is received, close congestion window
1581 * to one segment. We will gradually open it again as we proceed.
1584 tcp_quench(struct inpcb *inp, int error)
1586 struct tcpcb *tp = intotcpcb(inp);
1589 tp->snd_cwnd = tp->t_maxseg;
1595 * When a specific ICMP unreachable message is received and the
1596 * connection state is SYN-SENT, drop the connection. This behavior
1597 * is controlled by the icmp_may_rst sysctl.
1600 tcp_drop_syn_sent(struct inpcb *inp, int error)
1602 struct tcpcb *tp = intotcpcb(inp);
1604 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1605 tcp_drop(tp, error);
1609 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1610 * based on the new value in the route. Also nudge TCP to send something,
1611 * since we know the packet we just sent was dropped.
1612 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1615 tcp_mtudisc(struct inpcb *inp, int mtu)
1617 struct tcpcb *tp = intotcpcb(inp);
1619 struct socket *so = inp->inp_socket;
1622 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1624 const boolean_t isipv6 = FALSE;
1631 * If no MTU is provided in the ICMP message, use the
1632 * next lower likely value, as specified in RFC 1191.
1637 oldmtu = tp->t_maxopd +
1639 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1640 sizeof(struct tcpiphdr));
1641 mtu = ip_next_mtu(oldmtu, 0);
1645 rt = tcp_rtlookup6(&inp->inp_inc);
1647 rt = tcp_rtlookup(&inp->inp_inc);
1649 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1650 mtu = rt->rt_rmx.rmx_mtu;
1654 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1655 sizeof(struct tcpiphdr));
1658 * XXX - The following conditional probably violates the TCP
1659 * spec. The problem is that, since we don't know the
1660 * other end's MSS, we are supposed to use a conservative
1661 * default. But, if we do that, then MTU discovery will
1662 * never actually take place, because the conservative
1663 * default is much less than the MTUs typically seen
1664 * on the Internet today. For the moment, we'll sweep
1665 * this under the carpet.
1667 * The conservative default might not actually be a problem
1668 * if the only case this occurs is when sending an initial
1669 * SYN with options and data to a host we've never talked
1670 * to before. Then, they will reply with an MSS value which
1671 * will get recorded and the new parameters should get
1672 * recomputed. For Further Study.
1674 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1675 maxopd = rt->rt_rmx.rmx_mssopt;
1679 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1680 sizeof(struct tcpiphdr));
1682 if (tp->t_maxopd <= maxopd)
1684 tp->t_maxopd = maxopd;
1687 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1688 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1689 mss -= TCPOLEN_TSTAMP_APPA;
1691 /* round down to multiple of MCLBYTES */
1692 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1694 mss &= ~(MCLBYTES - 1);
1697 mss = (mss / MCLBYTES) * MCLBYTES;
1700 if (so->so_snd.ssb_hiwat < mss)
1701 mss = so->so_snd.ssb_hiwat;
1705 tp->snd_nxt = tp->snd_una;
1707 tcpstat.tcps_mturesent++;
1711 * Look-up the routing entry to the peer of this inpcb. If no route
1712 * is found and it cannot be allocated the return NULL. This routine
1713 * is called by TCP routines that access the rmx structure and by tcp_mss
1714 * to get the interface MTU.
1717 tcp_rtlookup(struct in_conninfo *inc)
1719 struct route *ro = &inc->inc_route;
1721 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1722 /* No route yet, so try to acquire one */
1723 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1725 * unused portions of the structure MUST be zero'd
1726 * out because rtalloc() treats it as opaque data
1728 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1729 ro->ro_dst.sa_family = AF_INET;
1730 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1731 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1741 tcp_rtlookup6(struct in_conninfo *inc)
1743 struct route_in6 *ro6 = &inc->inc6_route;
1745 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1746 /* No route yet, so try to acquire one */
1747 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1749 * unused portions of the structure MUST be zero'd
1750 * out because rtalloc() treats it as opaque data
1752 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1753 ro6->ro_dst.sin6_family = AF_INET6;
1754 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1755 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1756 rtalloc((struct route *)ro6);
1759 return (ro6->ro_rt);
1764 /* compute ESP/AH header size for TCP, including outer IP header. */
1766 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1774 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1776 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1781 if (inp->inp_vflag & INP_IPV6) {
1782 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1784 th = (struct tcphdr *)(ip6 + 1);
1785 m->m_pkthdr.len = m->m_len =
1786 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1787 tcp_fillheaders(tp, ip6, th);
1788 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1792 ip = mtod(m, struct ip *);
1793 th = (struct tcphdr *)(ip + 1);
1794 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1795 tcp_fillheaders(tp, ip, th);
1796 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1805 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1807 * This code attempts to calculate the bandwidth-delay product as a
1808 * means of determining the optimal window size to maximize bandwidth,
1809 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1810 * routers. This code also does a fairly good job keeping RTTs in check
1811 * across slow links like modems. We implement an algorithm which is very
1812 * similar (but not meant to be) TCP/Vegas. The code operates on the
1813 * transmitter side of a TCP connection and so only effects the transmit
1814 * side of the connection.
1816 * BACKGROUND: TCP makes no provision for the management of buffer space
1817 * at the end points or at the intermediate routers and switches. A TCP
1818 * stream, whether using NewReno or not, will eventually buffer as
1819 * many packets as it is able and the only reason this typically works is
1820 * due to the fairly small default buffers made available for a connection
1821 * (typicaly 16K or 32K). As machines use larger windows and/or window
1822 * scaling it is now fairly easy for even a single TCP connection to blow-out
1823 * all available buffer space not only on the local interface, but on
1824 * intermediate routers and switches as well. NewReno makes a misguided
1825 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1826 * then backing off, then steadily increasing the window again until another
1827 * failure occurs, ad-infinitum. This results in terrible oscillation that
1828 * is only made worse as network loads increase and the idea of intentionally
1829 * blowing out network buffers is, frankly, a terrible way to manage network
1832 * It is far better to limit the transmit window prior to the failure
1833 * condition being achieved. There are two general ways to do this: First
1834 * you can 'scan' through different transmit window sizes and locate the
1835 * point where the RTT stops increasing, indicating that you have filled the
1836 * pipe, then scan backwards until you note that RTT stops decreasing, then
1837 * repeat ad-infinitum. This method works in principle but has severe
1838 * implementation issues due to RTT variances, timer granularity, and
1839 * instability in the algorithm which can lead to many false positives and
1840 * create oscillations as well as interact badly with other TCP streams
1841 * implementing the same algorithm.
1843 * The second method is to limit the window to the bandwidth delay product
1844 * of the link. This is the method we implement. RTT variances and our
1845 * own manipulation of the congestion window, bwnd, can potentially
1846 * destabilize the algorithm. For this reason we have to stabilize the
1847 * elements used to calculate the window. We do this by using the minimum
1848 * observed RTT, the long term average of the observed bandwidth, and
1849 * by adding two segments worth of slop. It isn't perfect but it is able
1850 * to react to changing conditions and gives us a very stable basis on
1851 * which to extend the algorithm.
1854 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1862 * If inflight_enable is disabled in the middle of a tcp connection,
1863 * make sure snd_bwnd is effectively disabled.
1865 if (!tcp_inflight_enable) {
1866 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1867 tp->snd_bandwidth = 0;
1872 * Validate the delta time. If a connection is new or has been idle
1873 * a long time we have to reset the bandwidth calculator.
1876 delta_ticks = save_ticks - tp->t_bw_rtttime;
1877 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1878 tp->t_bw_rtttime = ticks;
1879 tp->t_bw_rtseq = ack_seq;
1880 if (tp->snd_bandwidth == 0)
1881 tp->snd_bandwidth = tcp_inflight_min;
1884 if (delta_ticks == 0)
1888 * Sanity check, plus ignore pure window update acks.
1890 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1894 * Figure out the bandwidth. Due to the tick granularity this
1895 * is a very rough number and it MUST be averaged over a fairly
1896 * long period of time. XXX we need to take into account a link
1897 * that is not using all available bandwidth, but for now our
1898 * slop will ramp us up if this case occurs and the bandwidth later
1901 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1902 tp->t_bw_rtttime = save_ticks;
1903 tp->t_bw_rtseq = ack_seq;
1904 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1906 tp->snd_bandwidth = bw;
1909 * Calculate the semi-static bandwidth delay product, plus two maximal
1910 * segments. The additional slop puts us squarely in the sweet
1911 * spot and also handles the bandwidth run-up case. Without the
1912 * slop we could be locking ourselves into a lower bandwidth.
1914 * Situations Handled:
1915 * (1) Prevents over-queueing of packets on LANs, especially on
1916 * high speed LANs, allowing larger TCP buffers to be
1917 * specified, and also does a good job preventing
1918 * over-queueing of packets over choke points like modems
1919 * (at least for the transmit side).
1921 * (2) Is able to handle changing network loads (bandwidth
1922 * drops so bwnd drops, bandwidth increases so bwnd
1925 * (3) Theoretically should stabilize in the face of multiple
1926 * connections implementing the same algorithm (this may need
1929 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1930 * be adjusted with a sysctl but typically only needs to be on
1931 * very slow connections. A value no smaller then 5 should
1932 * be used, but only reduce this default if you have no other
1936 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1937 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1938 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1941 if (tcp_inflight_debug > 0) {
1943 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1945 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1946 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1949 if ((long)bwnd < tcp_inflight_min)
1950 bwnd = tcp_inflight_min;
1951 if (bwnd > tcp_inflight_max)
1952 bwnd = tcp_inflight_max;
1953 if ((long)bwnd < tp->t_maxseg * 2)
1954 bwnd = tp->t_maxseg * 2;
1955 tp->snd_bwnd = bwnd;
1958 #ifdef TCP_SIGNATURE
1960 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1962 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1963 * When called from tcp_input(), we can be sure that th_sum has been
1964 * zeroed out and verified already.
1966 * Return 0 if successful, otherwise return -1.
1968 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1969 * search with the destination IP address, and a 'magic SPI' to be
1970 * determined by the application. This is hardcoded elsewhere to 1179
1971 * right now. Another branch of this code exists which uses the SPD to
1972 * specify per-application flows but it is unstable.
1975 tcpsignature_compute(
1976 struct mbuf *m, /* mbuf chain */
1977 int len, /* length of TCP data */
1978 int optlen, /* length of TCP options */
1979 u_char *buf, /* storage for MD5 digest */
1980 u_int direction) /* direction of flow */
1982 struct ippseudo ippseudo;
1986 struct ipovly *ipovly;
1987 struct secasvar *sav;
1990 struct ip6_hdr *ip6;
1991 struct in6_addr in6;
1997 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
1998 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2000 * Extract the destination from the IP header in the mbuf.
2002 ip = mtod(m, struct ip *);
2004 ip6 = NULL; /* Make the compiler happy. */
2007 * Look up an SADB entry which matches the address found in
2010 switch (IP_VHL_V(ip->ip_vhl)) {
2012 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2013 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2016 case (IPV6_VERSION >> 4):
2017 ip6 = mtod(m, struct ip6_hdr *);
2018 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2019 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2028 kprintf("%s: SADB lookup failed\n", __func__);
2034 * Step 1: Update MD5 hash with IP pseudo-header.
2036 * XXX The ippseudo header MUST be digested in network byte order,
2037 * or else we'll fail the regression test. Assume all fields we've
2038 * been doing arithmetic on have been in host byte order.
2039 * XXX One cannot depend on ipovly->ih_len here. When called from
2040 * tcp_output(), the underlying ip_len member has not yet been set.
2042 switch (IP_VHL_V(ip->ip_vhl)) {
2044 ipovly = (struct ipovly *)ip;
2045 ippseudo.ippseudo_src = ipovly->ih_src;
2046 ippseudo.ippseudo_dst = ipovly->ih_dst;
2047 ippseudo.ippseudo_pad = 0;
2048 ippseudo.ippseudo_p = IPPROTO_TCP;
2049 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2050 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2051 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2052 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2056 * RFC 2385, 2.0 Proposal
2057 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2058 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2059 * extended next header value (to form 32 bits), and 32-bit segment
2061 * Note: Upper-Layer Packet Length comes before Next Header.
2063 case (IPV6_VERSION >> 4):
2065 in6_clearscope(&in6);
2066 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2068 in6_clearscope(&in6);
2069 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2070 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2071 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2073 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2074 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2075 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2077 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2078 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2079 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2088 * Step 2: Update MD5 hash with TCP header, excluding options.
2089 * The TCP checksum must be set to zero.
2091 savecsum = th->th_sum;
2093 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2094 th->th_sum = savecsum;
2096 * Step 3: Update MD5 hash with TCP segment data.
2097 * Use m_apply() to avoid an early m_pullup().
2100 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2102 * Step 4: Update MD5 hash with shared secret.
2104 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2105 MD5Final(buf, &ctx);
2106 key_sa_recordxfer(sav, m);
2112 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2115 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2118 #endif /* TCP_SIGNATURE */