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 int tcp_mssdflt = TCP_MSS;
166 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
167 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
170 int tcp_v6mssdflt = TCP6_MSS;
171 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
172 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
176 * Minimum MSS we accept and use. This prevents DoS attacks where
177 * we are forced to a ridiculous low MSS like 20 and send hundreds
178 * of packets instead of one. The effect scales with the available
179 * bandwidth and quickly saturates the CPU and network interface
180 * with packet generation and sending. Set to zero to disable MINMSS
181 * checking. This setting prevents us from sending too small packets.
183 int tcp_minmss = TCP_MINMSS;
184 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
185 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
188 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
189 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
190 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
193 int tcp_do_rfc1323 = 1;
194 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
195 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
197 static int tcp_tcbhashsize = 0;
198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
199 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
201 static int do_tcpdrain = 1;
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
203 "Enable tcp_drain routine for extra help when low on mbufs");
205 static int icmp_may_rst = 1;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
207 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
209 static int tcp_isn_reseed_interval = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
211 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
214 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
215 * by default, but with generous values which should allow maximal
216 * bandwidth. In particular, the slop defaults to 50 (5 packets).
218 * The reason for doing this is that the limiter is the only mechanism we
219 * have which seems to do a really good job preventing receiver RX rings
220 * on network interfaces from getting blown out. Even though GigE/10GigE
221 * is supposed to flow control it looks like either it doesn't actually
222 * do it or Open Source drivers do not properly enable it.
224 * People using the limiter to reduce bottlenecks on slower WAN connections
225 * should set the slop to 20 (2 packets).
227 static int tcp_inflight_enable = 1;
228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
229 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
231 static int tcp_inflight_debug = 0;
232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
233 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
235 static int tcp_inflight_min = 6144;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
237 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
239 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
240 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
241 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
243 static int tcp_inflight_stab = 50;
244 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
245 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
247 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
248 static struct malloc_pipe tcptemp_mpipe;
250 static void tcp_willblock(void);
251 static void tcp_notify (struct inpcb *, int);
253 struct tcp_stats tcpstats_percpu[MAXCPU];
256 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
260 for (cpu = 0; cpu < ncpus; ++cpu) {
261 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
262 sizeof(struct tcp_stats))))
264 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
265 sizeof(struct tcp_stats))))
271 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
272 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
274 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
275 &tcpstat, tcp_stats, "TCP statistics");
279 * Target size of TCP PCB hash tables. Must be a power of two.
281 * Note that this can be overridden by the kernel environment
282 * variable net.inet.tcp.tcbhashsize
285 #define TCBHASHSIZE 512
289 * This is the actual shape of what we allocate using the zone
290 * allocator. Doing it this way allows us to protect both structures
291 * using the same generation count, and also eliminates the overhead
292 * of allocating tcpcbs separately. By hiding the structure here,
293 * we avoid changing most of the rest of the code (although it needs
294 * to be changed, eventually, for greater efficiency).
297 #define ALIGNM1 (ALIGNMENT - 1)
301 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
304 struct tcp_callout inp_tp_rexmt;
305 struct tcp_callout inp_tp_persist;
306 struct tcp_callout inp_tp_keep;
307 struct tcp_callout inp_tp_2msl;
308 struct tcp_callout inp_tp_delack;
309 struct netmsg_tcp_timer inp_tp_timermsg;
320 struct inpcbinfo *ticb;
321 int hashsize = TCBHASHSIZE;
325 * note: tcptemp is used for keepalives, and it is ok for an
326 * allocation to fail so do not specify MPF_INT.
328 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
331 tcp_delacktime = TCPTV_DELACK;
332 tcp_keepinit = TCPTV_KEEP_INIT;
333 tcp_keepidle = TCPTV_KEEP_IDLE;
334 tcp_keepintvl = TCPTV_KEEPINTVL;
335 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
337 tcp_rexmit_min = TCPTV_MIN;
338 tcp_rexmit_slop = TCPTV_CPU_VAR;
340 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
341 if (!powerof2(hashsize)) {
342 kprintf("WARNING: TCB hash size not a power of 2\n");
343 hashsize = 512; /* safe default */
345 tcp_tcbhashsize = hashsize;
347 for (cpu = 0; cpu < ncpus2; cpu++) {
348 ticb = &tcbinfo[cpu];
349 in_pcbinfo_init(ticb);
351 ticb->hashbase = hashinit(hashsize, M_PCB,
353 ticb->porthashbase = hashinit(hashsize, M_PCB,
354 &ticb->porthashmask);
355 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
356 &ticb->wildcardhashmask);
357 ticb->ipi_size = sizeof(struct inp_tp);
358 TAILQ_INIT(&tcpcbackq[cpu]);
361 tcp_reass_maxseg = nmbclusters / 16;
362 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
365 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
367 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
369 if (max_protohdr < TCP_MINPROTOHDR)
370 max_protohdr = TCP_MINPROTOHDR;
371 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
373 #undef TCP_MINPROTOHDR
376 * Initialize TCP statistics counters for each CPU.
379 for (cpu = 0; cpu < ncpus; ++cpu) {
380 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
383 bzero(&tcpstat, sizeof(struct tcp_stats));
387 netisr_register_rollup(tcp_willblock);
394 int cpu = mycpu->gd_cpuid;
396 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
397 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
398 tp->t_flags &= ~TF_ONOUTPUTQ;
399 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
405 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
406 * tcp_template used to store this data in mbufs, but we now recopy it out
407 * of the tcpcb each time to conserve mbufs.
410 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
412 struct inpcb *inp = tp->t_inpcb;
413 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
416 if (inp->inp_vflag & INP_IPV6) {
419 ip6 = (struct ip6_hdr *)ip_ptr;
420 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
421 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
422 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
423 (IPV6_VERSION & IPV6_VERSION_MASK);
424 ip6->ip6_nxt = IPPROTO_TCP;
425 ip6->ip6_plen = sizeof(struct tcphdr);
426 ip6->ip6_src = inp->in6p_laddr;
427 ip6->ip6_dst = inp->in6p_faddr;
432 struct ip *ip = (struct ip *) ip_ptr;
434 ip->ip_vhl = IP_VHL_BORING;
441 ip->ip_p = IPPROTO_TCP;
442 ip->ip_src = inp->inp_laddr;
443 ip->ip_dst = inp->inp_faddr;
444 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
446 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
449 tcp_hdr->th_sport = inp->inp_lport;
450 tcp_hdr->th_dport = inp->inp_fport;
455 tcp_hdr->th_flags = 0;
461 * Create template to be used to send tcp packets on a connection.
462 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
463 * use for this function is in keepalives, which use tcp_respond.
466 tcp_maketemplate(struct tcpcb *tp)
470 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
472 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
477 tcp_freetemplate(struct tcptemp *tmp)
479 mpipe_free(&tcptemp_mpipe, tmp);
483 * Send a single message to the TCP at address specified by
484 * the given TCP/IP header. If m == NULL, then we make a copy
485 * of the tcpiphdr at ti and send directly to the addressed host.
486 * This is used to force keep alive messages out using the TCP
487 * template for a connection. If flags are given then we send
488 * a message back to the TCP which originated the * segment ti,
489 * and discard the mbuf containing it and any other attached mbufs.
491 * In any case the ack and sequence number of the transmitted
492 * segment are as specified by the parameters.
494 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
497 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
498 tcp_seq ack, tcp_seq seq, int flags)
502 struct route *ro = NULL;
504 struct ip *ip = ipgen;
507 struct route_in6 *ro6 = NULL;
508 struct route_in6 sro6;
509 struct ip6_hdr *ip6 = ipgen;
510 boolean_t use_tmpro = TRUE;
512 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
514 const boolean_t isipv6 = FALSE;
518 if (!(flags & TH_RST)) {
519 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
522 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
523 win = (long)TCP_MAXWIN << tp->rcv_scale;
526 * Don't use the route cache of a listen socket,
527 * it is not MPSAFE; use temporary route cache.
529 if (tp->t_state != TCPS_LISTEN) {
531 ro6 = &tp->t_inpcb->in6p_route;
533 ro = &tp->t_inpcb->inp_route;
540 bzero(ro6, sizeof *ro6);
543 bzero(ro, sizeof *ro);
547 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
551 m->m_data += max_linkhdr;
553 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
554 ip6 = mtod(m, struct ip6_hdr *);
555 nth = (struct tcphdr *)(ip6 + 1);
557 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
558 ip = mtod(m, struct ip *);
559 nth = (struct tcphdr *)(ip + 1);
561 bcopy(th, nth, sizeof(struct tcphdr));
566 m->m_data = (caddr_t)ipgen;
567 /* m_len is set later */
569 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
571 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
572 nth = (struct tcphdr *)(ip6 + 1);
574 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
575 nth = (struct tcphdr *)(ip + 1);
579 * this is usually a case when an extension header
580 * exists between the IPv6 header and the
583 nth->th_sport = th->th_sport;
584 nth->th_dport = th->th_dport;
586 xchg(nth->th_dport, nth->th_sport, n_short);
591 ip6->ip6_vfc = IPV6_VERSION;
592 ip6->ip6_nxt = IPPROTO_TCP;
593 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
594 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
596 tlen += sizeof(struct tcpiphdr);
598 ip->ip_ttl = ip_defttl;
601 m->m_pkthdr.len = tlen;
602 m->m_pkthdr.rcvif = NULL;
603 nth->th_seq = htonl(seq);
604 nth->th_ack = htonl(ack);
606 nth->th_off = sizeof(struct tcphdr) >> 2;
607 nth->th_flags = flags;
609 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
611 nth->th_win = htons((u_short)win);
615 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
616 sizeof(struct ip6_hdr),
617 tlen - sizeof(struct ip6_hdr));
618 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
619 (ro6 && ro6->ro_rt) ?
620 ro6->ro_rt->rt_ifp : NULL);
622 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
623 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
624 m->m_pkthdr.csum_flags = CSUM_TCP;
625 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
628 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
629 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
632 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
633 tp ? tp->t_inpcb : NULL);
634 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
639 ipflags |= IP_DEBUGROUTE;
640 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
641 if ((ro == &sro) && (ro->ro_rt != NULL)) {
649 * Create a new TCP control block, making an
650 * empty reassembly queue and hooking it to the argument
651 * protocol control block. The `inp' parameter must have
652 * come from the zone allocator set up in tcp_init().
655 tcp_newtcpcb(struct inpcb *inp)
660 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
662 const boolean_t isipv6 = FALSE;
665 it = (struct inp_tp *)inp;
667 bzero(tp, sizeof(struct tcpcb));
668 LIST_INIT(&tp->t_segq);
669 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
671 /* Set up our timeouts. */
672 tp->tt_rexmt = &it->inp_tp_rexmt;
673 tp->tt_persist = &it->inp_tp_persist;
674 tp->tt_keep = &it->inp_tp_keep;
675 tp->tt_2msl = &it->inp_tp_2msl;
676 tp->tt_delack = &it->inp_tp_delack;
680 * Zero out timer message. We don't create it here,
681 * since the current CPU may not be the owner of this
684 tp->tt_msg = &it->inp_tp_timermsg;
685 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
688 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
689 tp->t_inpcb = inp; /* XXX */
690 tp->t_state = TCPS_CLOSED;
692 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
693 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
694 * reasonable initial retransmit time.
696 tp->t_srtt = TCPTV_SRTTBASE;
698 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
699 tp->t_rttmin = tcp_rexmit_min;
700 tp->t_rxtcur = TCPTV_RTOBASE;
701 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
702 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
703 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
704 tp->t_rcvtime = ticks;
706 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
707 * because the socket may be bound to an IPv6 wildcard address,
708 * which may match an IPv4-mapped IPv6 address.
710 inp->inp_ip_ttl = ip_defttl;
712 tcp_sack_tcpcb_init(tp);
713 return (tp); /* XXX */
717 * Drop a TCP connection, reporting the specified error.
718 * If connection is synchronized, then send a RST to peer.
721 tcp_drop(struct tcpcb *tp, int error)
723 struct socket *so = tp->t_inpcb->inp_socket;
725 if (TCPS_HAVERCVDSYN(tp->t_state)) {
726 tp->t_state = TCPS_CLOSED;
728 tcpstat.tcps_drops++;
730 tcpstat.tcps_conndrops++;
731 if (error == ETIMEDOUT && tp->t_softerror)
732 error = tp->t_softerror;
733 so->so_error = error;
734 return (tcp_close(tp));
739 struct netmsg_remwildcard {
740 struct netmsg nm_netmsg;
741 struct inpcb *nm_inp;
742 struct inpcbinfo *nm_pcbinfo;
751 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
752 * inp can be detached. We do this by cycling through the cpus, ending up
753 * on the cpu controlling the inp last and then doing the disconnect.
756 in_pcbremwildcardhash_handler(struct netmsg *msg0)
758 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
761 cpu = msg->nm_pcbinfo->cpu;
763 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
764 /* note: detach removes any wildcard hash entry */
767 in6_pcbdetach(msg->nm_inp);
770 in_pcbdetach(msg->nm_inp);
771 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
773 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
774 cpu = (cpu + 1) % ncpus2;
775 msg->nm_pcbinfo = &tcbinfo[cpu];
776 lwkt_forwardmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
783 * Close a TCP control block:
784 * discard all space held by the tcp
785 * discard internet protocol block
786 * wake up any sleepers
789 tcp_close(struct tcpcb *tp)
792 struct inpcb *inp = tp->t_inpcb;
793 struct socket *so = inp->inp_socket;
795 boolean_t dosavessthresh;
800 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
801 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
803 const boolean_t isipv6 = FALSE;
807 * The tp is not instantly destroyed in the wildcard case. Setting
808 * the state to TCPS_TERMINATING will prevent the TCP stack from
809 * messing with it, though it should be noted that this change may
810 * not take effect on other cpus until we have chained the wildcard
813 * XXX we currently depend on the BGL to synchronize the tp->t_state
814 * update and prevent other tcp protocol threads from accepting new
815 * connections on the listen socket we might be trying to close down.
817 KKASSERT(tp->t_state != TCPS_TERMINATING);
818 tp->t_state = TCPS_TERMINATING;
821 * Make sure that all of our timers are stopped before we
822 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
823 * timers are never used. If timer message is never created
824 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
826 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
827 tcp_callout_stop(tp, tp->tt_rexmt);
828 tcp_callout_stop(tp, tp->tt_persist);
829 tcp_callout_stop(tp, tp->tt_keep);
830 tcp_callout_stop(tp, tp->tt_2msl);
831 tcp_callout_stop(tp, tp->tt_delack);
834 if (tp->t_flags & TF_ONOUTPUTQ) {
835 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
836 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
837 tp->t_flags &= ~TF_ONOUTPUTQ;
841 * If we got enough samples through the srtt filter,
842 * save the rtt and rttvar in the routing entry.
843 * 'Enough' is arbitrarily defined as the 16 samples.
844 * 16 samples is enough for the srtt filter to converge
845 * to within 5% of the correct value; fewer samples and
846 * we could save a very bogus rtt.
848 * Don't update the default route's characteristics and don't
849 * update anything that the user "locked".
851 if (tp->t_rttupdated >= 16) {
855 struct sockaddr_in6 *sin6;
857 if ((rt = inp->in6p_route.ro_rt) == NULL)
859 sin6 = (struct sockaddr_in6 *)rt_key(rt);
860 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
863 if ((rt = inp->inp_route.ro_rt) == NULL ||
864 ((struct sockaddr_in *)rt_key(rt))->
865 sin_addr.s_addr == INADDR_ANY)
868 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
869 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
870 if (rt->rt_rmx.rmx_rtt && i)
872 * filter this update to half the old & half
873 * the new values, converting scale.
874 * See route.h and tcp_var.h for a
875 * description of the scaling constants.
878 (rt->rt_rmx.rmx_rtt + i) / 2;
880 rt->rt_rmx.rmx_rtt = i;
881 tcpstat.tcps_cachedrtt++;
883 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
885 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
886 if (rt->rt_rmx.rmx_rttvar && i)
887 rt->rt_rmx.rmx_rttvar =
888 (rt->rt_rmx.rmx_rttvar + i) / 2;
890 rt->rt_rmx.rmx_rttvar = i;
891 tcpstat.tcps_cachedrttvar++;
894 * The old comment here said:
895 * update the pipelimit (ssthresh) if it has been updated
896 * already or if a pipesize was specified & the threshhold
897 * got below half the pipesize. I.e., wait for bad news
898 * before we start updating, then update on both good
901 * But we want to save the ssthresh even if no pipesize is
902 * specified explicitly in the route, because such
903 * connections still have an implicit pipesize specified
904 * by the global tcp_sendspace. In the absence of a reliable
905 * way to calculate the pipesize, it will have to do.
907 i = tp->snd_ssthresh;
908 if (rt->rt_rmx.rmx_sendpipe != 0)
909 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
911 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
912 if (dosavessthresh ||
913 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
914 (rt->rt_rmx.rmx_ssthresh != 0))) {
916 * convert the limit from user data bytes to
917 * packets then to packet data bytes.
919 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
924 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
925 sizeof(struct tcpiphdr));
926 if (rt->rt_rmx.rmx_ssthresh)
927 rt->rt_rmx.rmx_ssthresh =
928 (rt->rt_rmx.rmx_ssthresh + i) / 2;
930 rt->rt_rmx.rmx_ssthresh = i;
931 tcpstat.tcps_cachedssthresh++;
936 /* free the reassembly queue, if any */
937 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
938 LIST_REMOVE(q, tqe_q);
941 atomic_add_int(&tcp_reass_qsize, -1);
943 /* throw away SACK blocks in scoreboard*/
945 tcp_sack_cleanup(&tp->scb);
947 inp->inp_ppcb = NULL;
948 soisdisconnected(so);
949 /* note: pcb detached later on */
951 tcp_destroy_timermsg(tp);
952 if (tp->t_flags & TF_SYNCACHE)
953 syncache_destroy(tp);
956 * Discard the inp. In the SMP case a wildcard inp's hash (created
957 * by a listen socket or an INADDR_ANY udp socket) is replicated
958 * for each protocol thread and must be removed in the context of
959 * that thread. This is accomplished by chaining the message
962 * If the inp is not wildcarded we simply detach, which will remove
963 * the any hashes still present for this inp.
966 if (inp->inp_flags & INP_WILDCARD_MP) {
967 struct netmsg_remwildcard *msg;
969 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
970 msg = kmalloc(sizeof(struct netmsg_remwildcard),
971 M_LWKTMSG, M_INTWAIT);
972 netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
973 0, in_pcbremwildcardhash_handler);
975 msg->nm_isinet6 = isafinet6;
978 msg->nm_pcbinfo = &tcbinfo[cpu];
979 lwkt_sendmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
983 /* note: detach removes any wildcard hash entry */
991 tcpstat.tcps_closed++;
996 tcp_drain_oncpu(struct inpcbhead *head)
998 struct inpcb *marker;
1001 struct tseg_qent *te;
1004 * Allows us to block while running the list
1006 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1007 marker->inp_flags |= INP_PLACEMARKER;
1008 LIST_INSERT_HEAD(head, marker, inp_list);
1010 while ((inpb = LIST_NEXT(marker, inp_list)) != NULL) {
1011 if ((inpb->inp_flags & INP_PLACEMARKER) == 0 &&
1012 (tcpb = intotcpcb(inpb)) != NULL &&
1013 (te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1014 LIST_REMOVE(te, tqe_q);
1017 atomic_add_int(&tcp_reass_qsize, -1);
1020 LIST_REMOVE(marker, inp_list);
1021 LIST_INSERT_AFTER(inpb, marker, inp_list);
1024 LIST_REMOVE(marker, inp_list);
1025 kfree(marker, M_TEMP);
1029 struct netmsg_tcp_drain {
1030 struct netmsg nm_netmsg;
1031 struct inpcbhead *nm_head;
1035 tcp_drain_handler(netmsg_t netmsg)
1037 struct netmsg_tcp_drain *nm = (void *)netmsg;
1039 tcp_drain_oncpu(nm->nm_head);
1040 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1055 * Walk the tcpbs, if existing, and flush the reassembly queue,
1056 * if there is one...
1057 * XXX: The "Net/3" implementation doesn't imply that the TCP
1058 * reassembly queue should be flushed, but in a situation
1059 * where we're really low on mbufs, this is potentially
1063 for (cpu = 0; cpu < ncpus2; cpu++) {
1064 struct netmsg_tcp_drain *msg;
1066 if (cpu == mycpu->gd_cpuid) {
1067 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1069 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1070 M_LWKTMSG, M_NOWAIT);
1073 netmsg_init(&msg->nm_netmsg, NULL, &netisr_afree_rport,
1074 0, tcp_drain_handler);
1075 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1076 lwkt_sendmsg(cpu_portfn(cpu), &msg->nm_netmsg.nm_lmsg);
1080 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1085 * Notify a tcp user of an asynchronous error;
1086 * store error as soft error, but wake up user
1087 * (for now, won't do anything until can select for soft error).
1089 * Do not wake up user since there currently is no mechanism for
1090 * reporting soft errors (yet - a kqueue filter may be added).
1093 tcp_notify(struct inpcb *inp, int error)
1095 struct tcpcb *tp = intotcpcb(inp);
1098 * Ignore some errors if we are hooked up.
1099 * If connection hasn't completed, has retransmitted several times,
1100 * and receives a second error, give up now. This is better
1101 * than waiting a long time to establish a connection that
1102 * can never complete.
1104 if (tp->t_state == TCPS_ESTABLISHED &&
1105 (error == EHOSTUNREACH || error == ENETUNREACH ||
1106 error == EHOSTDOWN)) {
1108 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1110 tcp_drop(tp, error);
1112 tp->t_softerror = error;
1114 wakeup(&so->so_timeo);
1121 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1124 struct inpcb *marker;
1134 * The process of preparing the TCB list is too time-consuming and
1135 * resource-intensive to repeat twice on every request.
1137 if (req->oldptr == NULL) {
1138 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1139 gd = globaldata_find(ccpu);
1140 n += tcbinfo[gd->gd_cpuid].ipi_count;
1142 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1146 if (req->newptr != NULL)
1149 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1150 marker->inp_flags |= INP_PLACEMARKER;
1153 * OK, now we're committed to doing something. Run the inpcb list
1154 * for each cpu in the system and construct the output. Use a
1155 * list placemarker to deal with list changes occuring during
1156 * copyout blockages (but otherwise depend on being on the correct
1157 * cpu to avoid races).
1159 origcpu = mycpu->gd_cpuid;
1160 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1166 cpu_id = (origcpu + ccpu) % ncpus;
1167 if ((smp_active_mask & (1 << cpu_id)) == 0)
1169 rgd = globaldata_find(cpu_id);
1170 lwkt_setcpu_self(rgd);
1172 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1173 n = tcbinfo[cpu_id].ipi_count;
1175 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1177 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1179 * process a snapshot of pcbs, ignoring placemarkers
1180 * and using our own to allow SYSCTL_OUT to block.
1182 LIST_REMOVE(marker, inp_list);
1183 LIST_INSERT_AFTER(inp, marker, inp_list);
1185 if (inp->inp_flags & INP_PLACEMARKER)
1187 if (inp->inp_gencnt > gencnt)
1189 if (prison_xinpcb(req->td, inp))
1192 xt.xt_len = sizeof xt;
1193 bcopy(inp, &xt.xt_inp, sizeof *inp);
1194 inp_ppcb = inp->inp_ppcb;
1195 if (inp_ppcb != NULL)
1196 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1198 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1199 if (inp->inp_socket)
1200 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1201 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1205 LIST_REMOVE(marker, inp_list);
1206 if (error == 0 && i < n) {
1207 bzero(&xt, sizeof xt);
1208 xt.xt_len = sizeof xt;
1210 error = SYSCTL_OUT(req, &xt, sizeof xt);
1219 * Make sure we are on the same cpu we were on originally, since
1220 * higher level callers expect this. Also don't pollute caches with
1221 * migrated userland data by (eventually) returning to userland
1222 * on a different cpu.
1224 lwkt_setcpu_self(globaldata_find(origcpu));
1225 kfree(marker, M_TEMP);
1229 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1230 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1233 tcp_getcred(SYSCTL_HANDLER_ARGS)
1235 struct sockaddr_in addrs[2];
1240 error = priv_check(req->td, PRIV_ROOT);
1243 error = SYSCTL_IN(req, addrs, sizeof addrs);
1247 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1248 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1249 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1250 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1251 if (inp == NULL || inp->inp_socket == NULL) {
1255 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1261 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1262 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1266 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1268 struct sockaddr_in6 addrs[2];
1271 boolean_t mapped = FALSE;
1273 error = priv_check(req->td, PRIV_ROOT);
1276 error = SYSCTL_IN(req, addrs, sizeof addrs);
1279 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1280 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1287 inp = in_pcblookup_hash(&tcbinfo[0],
1288 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1290 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1294 inp = in6_pcblookup_hash(&tcbinfo[0],
1295 &addrs[1].sin6_addr, addrs[1].sin6_port,
1296 &addrs[0].sin6_addr, addrs[0].sin6_port,
1299 if (inp == NULL || inp->inp_socket == NULL) {
1303 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1309 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1311 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1314 struct netmsg_tcp_notify {
1315 struct netmsg nm_nmsg;
1316 void (*nm_notify)(struct inpcb *, int);
1317 struct in_addr nm_faddr;
1322 tcp_notifyall_oncpu(struct netmsg *netmsg)
1324 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1327 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1328 nmsg->nm_arg, nmsg->nm_notify);
1330 nextcpu = mycpuid + 1;
1331 if (nextcpu < ncpus2)
1332 lwkt_forwardmsg(cpu_portfn(nextcpu), &netmsg->nm_lmsg);
1334 lwkt_replymsg(&netmsg->nm_lmsg, 0);
1338 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1340 struct ip *ip = vip;
1342 struct in_addr faddr;
1345 void (*notify)(struct inpcb *, int) = tcp_notify;
1349 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1353 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1354 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1357 arg = inetctlerrmap[cmd];
1358 if (cmd == PRC_QUENCH) {
1359 notify = tcp_quench;
1360 } else if (icmp_may_rst &&
1361 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1362 cmd == PRC_UNREACH_PORT ||
1363 cmd == PRC_TIMXCEED_INTRANS) &&
1365 notify = tcp_drop_syn_sent;
1366 } else if (cmd == PRC_MSGSIZE) {
1367 struct icmp *icmp = (struct icmp *)
1368 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1370 arg = ntohs(icmp->icmp_nextmtu);
1371 notify = tcp_mtudisc;
1372 } else if (PRC_IS_REDIRECT(cmd)) {
1374 notify = in_rtchange;
1375 } else if (cmd == PRC_HOSTDEAD) {
1381 th = (struct tcphdr *)((caddr_t)ip +
1382 (IP_VHL_HL(ip->ip_vhl) << 2));
1383 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1384 ip->ip_src.s_addr, th->th_sport);
1385 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1386 ip->ip_src, th->th_sport, 0, NULL);
1387 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1388 icmpseq = htonl(th->th_seq);
1389 tp = intotcpcb(inp);
1390 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1391 SEQ_LT(icmpseq, tp->snd_max))
1392 (*notify)(inp, arg);
1394 struct in_conninfo inc;
1396 inc.inc_fport = th->th_dport;
1397 inc.inc_lport = th->th_sport;
1398 inc.inc_faddr = faddr;
1399 inc.inc_laddr = ip->ip_src;
1403 syncache_unreach(&inc, th);
1407 struct netmsg_tcp_notify nmsg;
1409 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1410 netmsg_init(&nmsg.nm_nmsg, NULL, &curthread->td_msgport,
1411 0, tcp_notifyall_oncpu);
1412 nmsg.nm_faddr = faddr;
1414 nmsg.nm_notify = notify;
1416 lwkt_domsg(cpu_portfn(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1422 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1425 void (*notify) (struct inpcb *, int) = tcp_notify;
1426 struct ip6_hdr *ip6;
1428 struct ip6ctlparam *ip6cp = NULL;
1429 const struct sockaddr_in6 *sa6_src = NULL;
1431 struct tcp_portonly {
1437 if (sa->sa_family != AF_INET6 ||
1438 sa->sa_len != sizeof(struct sockaddr_in6))
1442 if (cmd == PRC_QUENCH)
1443 notify = tcp_quench;
1444 else if (cmd == PRC_MSGSIZE) {
1445 struct ip6ctlparam *ip6cp = d;
1446 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1448 arg = ntohl(icmp6->icmp6_mtu);
1449 notify = tcp_mtudisc;
1450 } else if (!PRC_IS_REDIRECT(cmd) &&
1451 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1455 /* if the parameter is from icmp6, decode it. */
1457 ip6cp = (struct ip6ctlparam *)d;
1459 ip6 = ip6cp->ip6c_ip6;
1460 off = ip6cp->ip6c_off;
1461 sa6_src = ip6cp->ip6c_src;
1465 off = 0; /* fool gcc */
1470 struct in_conninfo inc;
1472 * XXX: We assume that when IPV6 is non NULL,
1473 * M and OFF are valid.
1476 /* check if we can safely examine src and dst ports */
1477 if (m->m_pkthdr.len < off + sizeof *thp)
1480 bzero(&th, sizeof th);
1481 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1483 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1484 (struct sockaddr *)ip6cp->ip6c_src,
1485 th.th_sport, cmd, arg, notify);
1487 inc.inc_fport = th.th_dport;
1488 inc.inc_lport = th.th_sport;
1489 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1490 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1492 syncache_unreach(&inc, &th);
1494 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1495 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1500 * Following is where TCP initial sequence number generation occurs.
1502 * There are two places where we must use initial sequence numbers:
1503 * 1. In SYN-ACK packets.
1504 * 2. In SYN packets.
1506 * All ISNs for SYN-ACK packets are generated by the syncache. See
1507 * tcp_syncache.c for details.
1509 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1510 * depends on this property. In addition, these ISNs should be
1511 * unguessable so as to prevent connection hijacking. To satisfy
1512 * the requirements of this situation, the algorithm outlined in
1513 * RFC 1948 is used to generate sequence numbers.
1515 * Implementation details:
1517 * Time is based off the system timer, and is corrected so that it
1518 * increases by one megabyte per second. This allows for proper
1519 * recycling on high speed LANs while still leaving over an hour
1522 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1523 * between seeding of isn_secret. This is normally set to zero,
1524 * as reseeding should not be necessary.
1528 #define ISN_BYTES_PER_SECOND 1048576
1530 u_char isn_secret[32];
1531 int isn_last_reseed;
1535 tcp_new_isn(struct tcpcb *tp)
1537 u_int32_t md5_buffer[4];
1540 /* Seed if this is the first use, reseed if requested. */
1541 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1542 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1544 read_random_unlimited(&isn_secret, sizeof isn_secret);
1545 isn_last_reseed = ticks;
1548 /* Compute the md5 hash and return the ISN. */
1550 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1551 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1553 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1554 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1555 sizeof(struct in6_addr));
1556 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1557 sizeof(struct in6_addr));
1561 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1562 sizeof(struct in_addr));
1563 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1564 sizeof(struct in_addr));
1566 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1567 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1568 new_isn = (tcp_seq) md5_buffer[0];
1569 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1574 * When a source quench is received, close congestion window
1575 * to one segment. We will gradually open it again as we proceed.
1578 tcp_quench(struct inpcb *inp, int error)
1580 struct tcpcb *tp = intotcpcb(inp);
1583 tp->snd_cwnd = tp->t_maxseg;
1589 * When a specific ICMP unreachable message is received and the
1590 * connection state is SYN-SENT, drop the connection. This behavior
1591 * is controlled by the icmp_may_rst sysctl.
1594 tcp_drop_syn_sent(struct inpcb *inp, int error)
1596 struct tcpcb *tp = intotcpcb(inp);
1598 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1599 tcp_drop(tp, error);
1603 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1604 * based on the new value in the route. Also nudge TCP to send something,
1605 * since we know the packet we just sent was dropped.
1606 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1609 tcp_mtudisc(struct inpcb *inp, int mtu)
1611 struct tcpcb *tp = intotcpcb(inp);
1613 struct socket *so = inp->inp_socket;
1616 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1618 const boolean_t isipv6 = FALSE;
1625 * If no MTU is provided in the ICMP message, use the
1626 * next lower likely value, as specified in RFC 1191.
1631 oldmtu = tp->t_maxopd +
1633 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1634 sizeof(struct tcpiphdr));
1635 mtu = ip_next_mtu(oldmtu, 0);
1639 rt = tcp_rtlookup6(&inp->inp_inc);
1641 rt = tcp_rtlookup(&inp->inp_inc);
1643 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1644 mtu = rt->rt_rmx.rmx_mtu;
1648 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1649 sizeof(struct tcpiphdr));
1652 * XXX - The following conditional probably violates the TCP
1653 * spec. The problem is that, since we don't know the
1654 * other end's MSS, we are supposed to use a conservative
1655 * default. But, if we do that, then MTU discovery will
1656 * never actually take place, because the conservative
1657 * default is much less than the MTUs typically seen
1658 * on the Internet today. For the moment, we'll sweep
1659 * this under the carpet.
1661 * The conservative default might not actually be a problem
1662 * if the only case this occurs is when sending an initial
1663 * SYN with options and data to a host we've never talked
1664 * to before. Then, they will reply with an MSS value which
1665 * will get recorded and the new parameters should get
1666 * recomputed. For Further Study.
1668 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1669 maxopd = rt->rt_rmx.rmx_mssopt;
1673 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1674 sizeof(struct tcpiphdr));
1676 if (tp->t_maxopd <= maxopd)
1678 tp->t_maxopd = maxopd;
1681 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1682 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1683 mss -= TCPOLEN_TSTAMP_APPA;
1685 /* round down to multiple of MCLBYTES */
1686 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1688 mss &= ~(MCLBYTES - 1);
1691 mss = (mss / MCLBYTES) * MCLBYTES;
1694 if (so->so_snd.ssb_hiwat < mss)
1695 mss = so->so_snd.ssb_hiwat;
1699 tp->snd_nxt = tp->snd_una;
1701 tcpstat.tcps_mturesent++;
1705 * Look-up the routing entry to the peer of this inpcb. If no route
1706 * is found and it cannot be allocated the return NULL. This routine
1707 * is called by TCP routines that access the rmx structure and by tcp_mss
1708 * to get the interface MTU.
1711 tcp_rtlookup(struct in_conninfo *inc)
1713 struct route *ro = &inc->inc_route;
1715 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1716 /* No route yet, so try to acquire one */
1717 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1719 * unused portions of the structure MUST be zero'd
1720 * out because rtalloc() treats it as opaque data
1722 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1723 ro->ro_dst.sa_family = AF_INET;
1724 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1725 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1735 tcp_rtlookup6(struct in_conninfo *inc)
1737 struct route_in6 *ro6 = &inc->inc6_route;
1739 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1740 /* No route yet, so try to acquire one */
1741 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1743 * unused portions of the structure MUST be zero'd
1744 * out because rtalloc() treats it as opaque data
1746 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1747 ro6->ro_dst.sin6_family = AF_INET6;
1748 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1749 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1750 rtalloc((struct route *)ro6);
1753 return (ro6->ro_rt);
1758 /* compute ESP/AH header size for TCP, including outer IP header. */
1760 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1768 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1770 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1775 if (inp->inp_vflag & INP_IPV6) {
1776 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1778 th = (struct tcphdr *)(ip6 + 1);
1779 m->m_pkthdr.len = m->m_len =
1780 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1781 tcp_fillheaders(tp, ip6, th);
1782 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1786 ip = mtod(m, struct ip *);
1787 th = (struct tcphdr *)(ip + 1);
1788 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1789 tcp_fillheaders(tp, ip, th);
1790 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1799 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1801 * This code attempts to calculate the bandwidth-delay product as a
1802 * means of determining the optimal window size to maximize bandwidth,
1803 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1804 * routers. This code also does a fairly good job keeping RTTs in check
1805 * across slow links like modems. We implement an algorithm which is very
1806 * similar (but not meant to be) TCP/Vegas. The code operates on the
1807 * transmitter side of a TCP connection and so only effects the transmit
1808 * side of the connection.
1810 * BACKGROUND: TCP makes no provision for the management of buffer space
1811 * at the end points or at the intermediate routers and switches. A TCP
1812 * stream, whether using NewReno or not, will eventually buffer as
1813 * many packets as it is able and the only reason this typically works is
1814 * due to the fairly small default buffers made available for a connection
1815 * (typicaly 16K or 32K). As machines use larger windows and/or window
1816 * scaling it is now fairly easy for even a single TCP connection to blow-out
1817 * all available buffer space not only on the local interface, but on
1818 * intermediate routers and switches as well. NewReno makes a misguided
1819 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1820 * then backing off, then steadily increasing the window again until another
1821 * failure occurs, ad-infinitum. This results in terrible oscillation that
1822 * is only made worse as network loads increase and the idea of intentionally
1823 * blowing out network buffers is, frankly, a terrible way to manage network
1826 * It is far better to limit the transmit window prior to the failure
1827 * condition being achieved. There are two general ways to do this: First
1828 * you can 'scan' through different transmit window sizes and locate the
1829 * point where the RTT stops increasing, indicating that you have filled the
1830 * pipe, then scan backwards until you note that RTT stops decreasing, then
1831 * repeat ad-infinitum. This method works in principle but has severe
1832 * implementation issues due to RTT variances, timer granularity, and
1833 * instability in the algorithm which can lead to many false positives and
1834 * create oscillations as well as interact badly with other TCP streams
1835 * implementing the same algorithm.
1837 * The second method is to limit the window to the bandwidth delay product
1838 * of the link. This is the method we implement. RTT variances and our
1839 * own manipulation of the congestion window, bwnd, can potentially
1840 * destabilize the algorithm. For this reason we have to stabilize the
1841 * elements used to calculate the window. We do this by using the minimum
1842 * observed RTT, the long term average of the observed bandwidth, and
1843 * by adding two segments worth of slop. It isn't perfect but it is able
1844 * to react to changing conditions and gives us a very stable basis on
1845 * which to extend the algorithm.
1848 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1856 * If inflight_enable is disabled in the middle of a tcp connection,
1857 * make sure snd_bwnd is effectively disabled.
1859 if (!tcp_inflight_enable) {
1860 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1861 tp->snd_bandwidth = 0;
1866 * Validate the delta time. If a connection is new or has been idle
1867 * a long time we have to reset the bandwidth calculator.
1870 delta_ticks = save_ticks - tp->t_bw_rtttime;
1871 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1872 tp->t_bw_rtttime = ticks;
1873 tp->t_bw_rtseq = ack_seq;
1874 if (tp->snd_bandwidth == 0)
1875 tp->snd_bandwidth = tcp_inflight_min;
1878 if (delta_ticks == 0)
1882 * Sanity check, plus ignore pure window update acks.
1884 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1888 * Figure out the bandwidth. Due to the tick granularity this
1889 * is a very rough number and it MUST be averaged over a fairly
1890 * long period of time. XXX we need to take into account a link
1891 * that is not using all available bandwidth, but for now our
1892 * slop will ramp us up if this case occurs and the bandwidth later
1895 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1896 tp->t_bw_rtttime = save_ticks;
1897 tp->t_bw_rtseq = ack_seq;
1898 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1900 tp->snd_bandwidth = bw;
1903 * Calculate the semi-static bandwidth delay product, plus two maximal
1904 * segments. The additional slop puts us squarely in the sweet
1905 * spot and also handles the bandwidth run-up case. Without the
1906 * slop we could be locking ourselves into a lower bandwidth.
1908 * Situations Handled:
1909 * (1) Prevents over-queueing of packets on LANs, especially on
1910 * high speed LANs, allowing larger TCP buffers to be
1911 * specified, and also does a good job preventing
1912 * over-queueing of packets over choke points like modems
1913 * (at least for the transmit side).
1915 * (2) Is able to handle changing network loads (bandwidth
1916 * drops so bwnd drops, bandwidth increases so bwnd
1919 * (3) Theoretically should stabilize in the face of multiple
1920 * connections implementing the same algorithm (this may need
1923 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1924 * be adjusted with a sysctl but typically only needs to be on
1925 * very slow connections. A value no smaller then 5 should
1926 * be used, but only reduce this default if you have no other
1930 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1931 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1932 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1935 if (tcp_inflight_debug > 0) {
1937 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1939 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1940 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1943 if ((long)bwnd < tcp_inflight_min)
1944 bwnd = tcp_inflight_min;
1945 if (bwnd > tcp_inflight_max)
1946 bwnd = tcp_inflight_max;
1947 if ((long)bwnd < tp->t_maxseg * 2)
1948 bwnd = tp->t_maxseg * 2;
1949 tp->snd_bwnd = bwnd;
1952 #ifdef TCP_SIGNATURE
1954 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1956 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1957 * When called from tcp_input(), we can be sure that th_sum has been
1958 * zeroed out and verified already.
1960 * Return 0 if successful, otherwise return -1.
1962 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1963 * search with the destination IP address, and a 'magic SPI' to be
1964 * determined by the application. This is hardcoded elsewhere to 1179
1965 * right now. Another branch of this code exists which uses the SPD to
1966 * specify per-application flows but it is unstable.
1969 tcpsignature_compute(
1970 struct mbuf *m, /* mbuf chain */
1971 int len, /* length of TCP data */
1972 int optlen, /* length of TCP options */
1973 u_char *buf, /* storage for MD5 digest */
1974 u_int direction) /* direction of flow */
1976 struct ippseudo ippseudo;
1980 struct ipovly *ipovly;
1981 struct secasvar *sav;
1984 struct ip6_hdr *ip6;
1985 struct in6_addr in6;
1991 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
1992 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
1994 * Extract the destination from the IP header in the mbuf.
1996 ip = mtod(m, struct ip *);
1998 ip6 = NULL; /* Make the compiler happy. */
2001 * Look up an SADB entry which matches the address found in
2004 switch (IP_VHL_V(ip->ip_vhl)) {
2006 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2007 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2010 case (IPV6_VERSION >> 4):
2011 ip6 = mtod(m, struct ip6_hdr *);
2012 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2013 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2022 kprintf("%s: SADB lookup failed\n", __func__);
2028 * Step 1: Update MD5 hash with IP pseudo-header.
2030 * XXX The ippseudo header MUST be digested in network byte order,
2031 * or else we'll fail the regression test. Assume all fields we've
2032 * been doing arithmetic on have been in host byte order.
2033 * XXX One cannot depend on ipovly->ih_len here. When called from
2034 * tcp_output(), the underlying ip_len member has not yet been set.
2036 switch (IP_VHL_V(ip->ip_vhl)) {
2038 ipovly = (struct ipovly *)ip;
2039 ippseudo.ippseudo_src = ipovly->ih_src;
2040 ippseudo.ippseudo_dst = ipovly->ih_dst;
2041 ippseudo.ippseudo_pad = 0;
2042 ippseudo.ippseudo_p = IPPROTO_TCP;
2043 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2044 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2045 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2046 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2050 * RFC 2385, 2.0 Proposal
2051 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2052 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2053 * extended next header value (to form 32 bits), and 32-bit segment
2055 * Note: Upper-Layer Packet Length comes before Next Header.
2057 case (IPV6_VERSION >> 4):
2059 in6_clearscope(&in6);
2060 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2062 in6_clearscope(&in6);
2063 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2064 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2065 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2067 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2068 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2069 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2071 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2072 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2073 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2082 * Step 2: Update MD5 hash with TCP header, excluding options.
2083 * The TCP checksum must be set to zero.
2085 savecsum = th->th_sum;
2087 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2088 th->th_sum = savecsum;
2090 * Step 3: Update MD5 hash with TCP segment data.
2091 * Use m_apply() to avoid an early m_pullup().
2094 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2096 * Step 4: Update MD5 hash with shared secret.
2098 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2099 MD5Final(buf, &ctx);
2100 key_sa_recordxfer(sav, m);
2106 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2109 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2112 #endif /* TCP_SIGNATURE */