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 int tcp_low_rtobase = 1;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW,
251 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)");
253 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
254 static struct malloc_pipe tcptemp_mpipe;
256 static void tcp_willblock(void);
257 static void tcp_notify (struct inpcb *, int);
259 struct tcp_stats tcpstats_percpu[MAXCPU];
262 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
266 for (cpu = 0; cpu < ncpus; ++cpu) {
267 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
268 sizeof(struct tcp_stats))))
270 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
271 sizeof(struct tcp_stats))))
277 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
278 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
280 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
281 &tcpstat, tcp_stats, "TCP statistics");
285 * Target size of TCP PCB hash tables. Must be a power of two.
287 * Note that this can be overridden by the kernel environment
288 * variable net.inet.tcp.tcbhashsize
291 #define TCBHASHSIZE 512
295 * This is the actual shape of what we allocate using the zone
296 * allocator. Doing it this way allows us to protect both structures
297 * using the same generation count, and also eliminates the overhead
298 * of allocating tcpcbs separately. By hiding the structure here,
299 * we avoid changing most of the rest of the code (although it needs
300 * to be changed, eventually, for greater efficiency).
303 #define ALIGNM1 (ALIGNMENT - 1)
307 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
310 struct tcp_callout inp_tp_rexmt;
311 struct tcp_callout inp_tp_persist;
312 struct tcp_callout inp_tp_keep;
313 struct tcp_callout inp_tp_2msl;
314 struct tcp_callout inp_tp_delack;
315 struct netmsg_tcp_timer inp_tp_timermsg;
326 struct inpcbporthead *porthashbase;
327 struct inpcbinfo *ticb;
329 int hashsize = TCBHASHSIZE;
333 * note: tcptemp is used for keepalives, and it is ok for an
334 * allocation to fail so do not specify MPF_INT.
336 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
337 25, -1, 0, NULL, NULL, NULL);
339 tcp_delacktime = TCPTV_DELACK;
340 tcp_keepinit = TCPTV_KEEP_INIT;
341 tcp_keepidle = TCPTV_KEEP_IDLE;
342 tcp_keepintvl = TCPTV_KEEPINTVL;
343 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
345 tcp_rexmit_min = TCPTV_MIN;
346 tcp_rexmit_slop = TCPTV_CPU_VAR;
348 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
349 if (!powerof2(hashsize)) {
350 kprintf("WARNING: TCB hash size not a power of 2\n");
351 hashsize = 512; /* safe default */
353 tcp_tcbhashsize = hashsize;
354 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
356 for (cpu = 0; cpu < ncpus2; cpu++) {
357 ticb = &tcbinfo[cpu];
358 in_pcbinfo_init(ticb);
360 ticb->hashbase = hashinit(hashsize, M_PCB,
362 ticb->porthashbase = porthashbase;
363 ticb->porthashmask = porthashmask;
364 ticb->porttoken = &tcp_port_token;
366 ticb->porthashbase = hashinit(hashsize, M_PCB,
367 &ticb->porthashmask);
369 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
370 &ticb->wildcardhashmask);
371 ticb->ipi_size = sizeof(struct inp_tp);
372 TAILQ_INIT(&tcpcbackq[cpu]);
375 tcp_reass_maxseg = nmbclusters / 16;
376 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
379 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
381 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
383 if (max_protohdr < TCP_MINPROTOHDR)
384 max_protohdr = TCP_MINPROTOHDR;
385 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
387 #undef TCP_MINPROTOHDR
390 * Initialize TCP statistics counters for each CPU.
393 for (cpu = 0; cpu < ncpus; ++cpu) {
394 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
397 bzero(&tcpstat, sizeof(struct tcp_stats));
401 netisr_register_rollup(tcp_willblock);
408 int cpu = mycpu->gd_cpuid;
410 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
411 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
412 tp->t_flags &= ~TF_ONOUTPUTQ;
413 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
419 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
420 * tcp_template used to store this data in mbufs, but we now recopy it out
421 * of the tcpcb each time to conserve mbufs.
424 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
426 struct inpcb *inp = tp->t_inpcb;
427 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
430 if (inp->inp_vflag & INP_IPV6) {
433 ip6 = (struct ip6_hdr *)ip_ptr;
434 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
435 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
436 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
437 (IPV6_VERSION & IPV6_VERSION_MASK);
438 ip6->ip6_nxt = IPPROTO_TCP;
439 ip6->ip6_plen = sizeof(struct tcphdr);
440 ip6->ip6_src = inp->in6p_laddr;
441 ip6->ip6_dst = inp->in6p_faddr;
446 struct ip *ip = (struct ip *) ip_ptr;
448 ip->ip_vhl = IP_VHL_BORING;
455 ip->ip_p = IPPROTO_TCP;
456 ip->ip_src = inp->inp_laddr;
457 ip->ip_dst = inp->inp_faddr;
458 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
460 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
463 tcp_hdr->th_sport = inp->inp_lport;
464 tcp_hdr->th_dport = inp->inp_fport;
469 tcp_hdr->th_flags = 0;
475 * Create template to be used to send tcp packets on a connection.
476 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
477 * use for this function is in keepalives, which use tcp_respond.
480 tcp_maketemplate(struct tcpcb *tp)
484 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
486 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
491 tcp_freetemplate(struct tcptemp *tmp)
493 mpipe_free(&tcptemp_mpipe, tmp);
497 * Send a single message to the TCP at address specified by
498 * the given TCP/IP header. If m == NULL, then we make a copy
499 * of the tcpiphdr at ti and send directly to the addressed host.
500 * This is used to force keep alive messages out using the TCP
501 * template for a connection. If flags are given then we send
502 * a message back to the TCP which originated the * segment ti,
503 * and discard the mbuf containing it and any other attached mbufs.
505 * In any case the ack and sequence number of the transmitted
506 * segment are as specified by the parameters.
508 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
511 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
512 tcp_seq ack, tcp_seq seq, int flags)
516 struct route *ro = NULL;
518 struct ip *ip = ipgen;
521 struct route_in6 *ro6 = NULL;
522 struct route_in6 sro6;
523 struct ip6_hdr *ip6 = ipgen;
524 boolean_t use_tmpro = TRUE;
526 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
528 const boolean_t isipv6 = FALSE;
532 if (!(flags & TH_RST)) {
533 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
536 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
537 win = (long)TCP_MAXWIN << tp->rcv_scale;
540 * Don't use the route cache of a listen socket,
541 * it is not MPSAFE; use temporary route cache.
543 if (tp->t_state != TCPS_LISTEN) {
545 ro6 = &tp->t_inpcb->in6p_route;
547 ro = &tp->t_inpcb->inp_route;
554 bzero(ro6, sizeof *ro6);
557 bzero(ro, sizeof *ro);
561 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
565 m->m_data += max_linkhdr;
567 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
568 ip6 = mtod(m, struct ip6_hdr *);
569 nth = (struct tcphdr *)(ip6 + 1);
571 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
572 ip = mtod(m, struct ip *);
573 nth = (struct tcphdr *)(ip + 1);
575 bcopy(th, nth, sizeof(struct tcphdr));
580 m->m_data = (caddr_t)ipgen;
581 /* m_len is set later */
583 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
585 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
586 nth = (struct tcphdr *)(ip6 + 1);
588 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
589 nth = (struct tcphdr *)(ip + 1);
593 * this is usually a case when an extension header
594 * exists between the IPv6 header and the
597 nth->th_sport = th->th_sport;
598 nth->th_dport = th->th_dport;
600 xchg(nth->th_dport, nth->th_sport, n_short);
605 ip6->ip6_vfc = IPV6_VERSION;
606 ip6->ip6_nxt = IPPROTO_TCP;
607 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
608 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
610 tlen += sizeof(struct tcpiphdr);
612 ip->ip_ttl = ip_defttl;
615 m->m_pkthdr.len = tlen;
616 m->m_pkthdr.rcvif = NULL;
617 nth->th_seq = htonl(seq);
618 nth->th_ack = htonl(ack);
620 nth->th_off = sizeof(struct tcphdr) >> 2;
621 nth->th_flags = flags;
623 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
625 nth->th_win = htons((u_short)win);
629 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
630 sizeof(struct ip6_hdr),
631 tlen - sizeof(struct ip6_hdr));
632 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
633 (ro6 && ro6->ro_rt) ?
634 ro6->ro_rt->rt_ifp : NULL);
636 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
637 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
638 m->m_pkthdr.csum_flags = CSUM_TCP;
639 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
642 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
643 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
646 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
647 tp ? tp->t_inpcb : NULL);
648 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
653 ipflags |= IP_DEBUGROUTE;
654 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
655 if ((ro == &sro) && (ro->ro_rt != NULL)) {
663 * Create a new TCP control block, making an
664 * empty reassembly queue and hooking it to the argument
665 * protocol control block. The `inp' parameter must have
666 * come from the zone allocator set up in tcp_init().
669 tcp_newtcpcb(struct inpcb *inp)
674 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
676 const boolean_t isipv6 = FALSE;
679 it = (struct inp_tp *)inp;
681 bzero(tp, sizeof(struct tcpcb));
682 LIST_INIT(&tp->t_segq);
683 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
685 /* Set up our timeouts. */
686 tp->tt_rexmt = &it->inp_tp_rexmt;
687 tp->tt_persist = &it->inp_tp_persist;
688 tp->tt_keep = &it->inp_tp_keep;
689 tp->tt_2msl = &it->inp_tp_2msl;
690 tp->tt_delack = &it->inp_tp_delack;
694 * Zero out timer message. We don't create it here,
695 * since the current CPU may not be the owner of this
698 tp->tt_msg = &it->inp_tp_timermsg;
699 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
701 tp->t_keepinit = tcp_keepinit;
702 tp->t_keepidle = tcp_keepidle;
703 tp->t_keepintvl = tcp_keepintvl;
704 tp->t_keepcnt = tcp_keepcnt;
705 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
708 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
709 tp->t_inpcb = inp; /* XXX */
710 tp->t_state = TCPS_CLOSED;
712 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
713 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
714 * reasonable initial retransmit time.
716 tp->t_srtt = TCPTV_SRTTBASE;
718 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
719 tp->t_rttmin = tcp_rexmit_min;
720 tp->t_rxtcur = TCPTV_RTOBASE;
721 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
722 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
723 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
724 tp->t_rcvtime = ticks;
726 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
727 * because the socket may be bound to an IPv6 wildcard address,
728 * which may match an IPv4-mapped IPv6 address.
730 inp->inp_ip_ttl = ip_defttl;
732 tcp_sack_tcpcb_init(tp);
733 return (tp); /* XXX */
737 * Drop a TCP connection, reporting the specified error.
738 * If connection is synchronized, then send a RST to peer.
741 tcp_drop(struct tcpcb *tp, int error)
743 struct socket *so = tp->t_inpcb->inp_socket;
745 if (TCPS_HAVERCVDSYN(tp->t_state)) {
746 tp->t_state = TCPS_CLOSED;
748 tcpstat.tcps_drops++;
750 tcpstat.tcps_conndrops++;
751 if (error == ETIMEDOUT && tp->t_softerror)
752 error = tp->t_softerror;
753 so->so_error = error;
754 return (tcp_close(tp));
759 struct netmsg_listen_detach {
760 struct netmsg_base base;
765 tcp_listen_detach_handler(netmsg_t msg)
767 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
768 struct tcpcb *tp = nmsg->nm_tp;
769 int cpu = mycpuid, nextcpu;
771 if (tp->t_flags & TF_LISTEN)
772 syncache_destroy(tp);
774 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
777 if (nextcpu < ncpus2)
778 lwkt_forwardmsg(cpu_portfn(nextcpu), &nmsg->base.lmsg);
780 lwkt_replymsg(&nmsg->base.lmsg, 0);
786 * Close a TCP control block:
787 * discard all space held by the tcp
788 * discard internet protocol block
789 * wake up any sleepers
792 tcp_close(struct tcpcb *tp)
795 struct inpcb *inp = tp->t_inpcb;
796 struct socket *so = inp->inp_socket;
798 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;
808 * INP_WILDCARD_MP indicates that listen(2) has been called on
809 * this socket. This implies:
810 * - A wildcard inp's hash is replicated for each protocol thread.
811 * - Syncache for this inp grows independently in each protocol
813 * - There is more than one cpu
815 * We have to chain a message to the rest of the protocol threads
816 * to cleanup the wildcard hash and the syncache. The cleanup
817 * in the current protocol thread is defered till the end of this
821 * After cleanup the inp's hash and syncache entries, this inp will
822 * no longer be available to the rest of the protocol threads, so we
823 * are safe to whack the inp in the following code.
825 if (inp->inp_flags & INP_WILDCARD_MP) {
826 struct netmsg_listen_detach nmsg;
828 KKASSERT(so->so_port == cpu_portfn(0));
829 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
830 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
832 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
833 MSGF_PRIORITY, tcp_listen_detach_handler);
835 lwkt_domsg(cpu_portfn(1), &nmsg.base.lmsg, 0);
837 inp->inp_flags &= ~INP_WILDCARD_MP;
841 KKASSERT(tp->t_state != TCPS_TERMINATING);
842 tp->t_state = TCPS_TERMINATING;
845 * Make sure that all of our timers are stopped before we
846 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
847 * timers are never used. If timer message is never created
848 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
850 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
851 tcp_callout_stop(tp, tp->tt_rexmt);
852 tcp_callout_stop(tp, tp->tt_persist);
853 tcp_callout_stop(tp, tp->tt_keep);
854 tcp_callout_stop(tp, tp->tt_2msl);
855 tcp_callout_stop(tp, tp->tt_delack);
858 if (tp->t_flags & TF_ONOUTPUTQ) {
859 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
860 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
861 tp->t_flags &= ~TF_ONOUTPUTQ;
865 * If we got enough samples through the srtt filter,
866 * save the rtt and rttvar in the routing entry.
867 * 'Enough' is arbitrarily defined as the 16 samples.
868 * 16 samples is enough for the srtt filter to converge
869 * to within 5% of the correct value; fewer samples and
870 * we could save a very bogus rtt.
872 * Don't update the default route's characteristics and don't
873 * update anything that the user "locked".
875 if (tp->t_rttupdated >= 16) {
879 struct sockaddr_in6 *sin6;
881 if ((rt = inp->in6p_route.ro_rt) == NULL)
883 sin6 = (struct sockaddr_in6 *)rt_key(rt);
884 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
887 if ((rt = inp->inp_route.ro_rt) == NULL ||
888 ((struct sockaddr_in *)rt_key(rt))->
889 sin_addr.s_addr == INADDR_ANY)
892 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
893 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
894 if (rt->rt_rmx.rmx_rtt && i)
896 * filter this update to half the old & half
897 * the new values, converting scale.
898 * See route.h and tcp_var.h for a
899 * description of the scaling constants.
902 (rt->rt_rmx.rmx_rtt + i) / 2;
904 rt->rt_rmx.rmx_rtt = i;
905 tcpstat.tcps_cachedrtt++;
907 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
909 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
910 if (rt->rt_rmx.rmx_rttvar && i)
911 rt->rt_rmx.rmx_rttvar =
912 (rt->rt_rmx.rmx_rttvar + i) / 2;
914 rt->rt_rmx.rmx_rttvar = i;
915 tcpstat.tcps_cachedrttvar++;
918 * The old comment here said:
919 * update the pipelimit (ssthresh) if it has been updated
920 * already or if a pipesize was specified & the threshhold
921 * got below half the pipesize. I.e., wait for bad news
922 * before we start updating, then update on both good
925 * But we want to save the ssthresh even if no pipesize is
926 * specified explicitly in the route, because such
927 * connections still have an implicit pipesize specified
928 * by the global tcp_sendspace. In the absence of a reliable
929 * way to calculate the pipesize, it will have to do.
931 i = tp->snd_ssthresh;
932 if (rt->rt_rmx.rmx_sendpipe != 0)
933 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
935 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
936 if (dosavessthresh ||
937 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
938 (rt->rt_rmx.rmx_ssthresh != 0))) {
940 * convert the limit from user data bytes to
941 * packets then to packet data bytes.
943 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
948 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
949 sizeof(struct tcpiphdr));
950 if (rt->rt_rmx.rmx_ssthresh)
951 rt->rt_rmx.rmx_ssthresh =
952 (rt->rt_rmx.rmx_ssthresh + i) / 2;
954 rt->rt_rmx.rmx_ssthresh = i;
955 tcpstat.tcps_cachedssthresh++;
960 /* free the reassembly queue, if any */
961 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
962 LIST_REMOVE(q, tqe_q);
965 atomic_add_int(&tcp_reass_qsize, -1);
967 /* throw away SACK blocks in scoreboard*/
969 tcp_sack_cleanup(&tp->scb);
971 inp->inp_ppcb = NULL;
972 soisdisconnected(so);
973 /* note: pcb detached later on */
975 tcp_destroy_timermsg(tp);
977 if (tp->t_flags & TF_LISTEN)
978 syncache_destroy(tp);
982 * pcbdetach removes any wildcard hash entry on the current CPU.
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_base base;
1031 struct inpcbhead *nm_head;
1035 tcp_drain_handler(netmsg_t msg)
1037 struct netmsg_tcp_drain *nm = (void *)msg;
1039 tcp_drain_oncpu(nm->nm_head);
1040 lwkt_replymsg(&nm->base.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 *nm;
1066 if (cpu == mycpu->gd_cpuid) {
1067 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1069 nm = kmalloc(sizeof(struct netmsg_tcp_drain),
1070 M_LWKTMSG, M_NOWAIT);
1073 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1074 0, tcp_drain_handler);
1075 nm->nm_head = &tcbinfo[cpu].pcblisthead;
1076 lwkt_sendmsg(cpu_portfn(cpu), &nm->base.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;
1133 * The process of preparing the TCB list is too time-consuming and
1134 * resource-intensive to repeat twice on every request.
1136 if (req->oldptr == NULL) {
1137 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1138 gd = globaldata_find(ccpu);
1139 n += tcbinfo[gd->gd_cpuid].ipi_count;
1141 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1145 if (req->newptr != NULL)
1148 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1149 marker->inp_flags |= INP_PLACEMARKER;
1152 * OK, now we're committed to doing something. Run the inpcb list
1153 * for each cpu in the system and construct the output. Use a
1154 * list placemarker to deal with list changes occuring during
1155 * copyout blockages (but otherwise depend on being on the correct
1156 * cpu to avoid races).
1158 origcpu = mycpu->gd_cpuid;
1159 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1165 cpu_id = (origcpu + ccpu) % ncpus;
1166 if ((smp_active_mask & CPUMASK(cpu_id)) == 0)
1168 rgd = globaldata_find(cpu_id);
1169 lwkt_setcpu_self(rgd);
1171 n = tcbinfo[cpu_id].ipi_count;
1173 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1175 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1177 * process a snapshot of pcbs, ignoring placemarkers
1178 * and using our own to allow SYSCTL_OUT to block.
1180 LIST_REMOVE(marker, inp_list);
1181 LIST_INSERT_AFTER(inp, marker, inp_list);
1183 if (inp->inp_flags & INP_PLACEMARKER)
1185 if (prison_xinpcb(req->td, inp))
1188 xt.xt_len = sizeof xt;
1189 bcopy(inp, &xt.xt_inp, sizeof *inp);
1190 inp_ppcb = inp->inp_ppcb;
1191 if (inp_ppcb != NULL)
1192 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1194 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1195 if (inp->inp_socket)
1196 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1197 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1201 LIST_REMOVE(marker, inp_list);
1202 if (error == 0 && i < n) {
1203 bzero(&xt, sizeof xt);
1204 xt.xt_len = sizeof xt;
1206 error = SYSCTL_OUT(req, &xt, sizeof xt);
1215 * Make sure we are on the same cpu we were on originally, since
1216 * higher level callers expect this. Also don't pollute caches with
1217 * migrated userland data by (eventually) returning to userland
1218 * on a different cpu.
1220 lwkt_setcpu_self(globaldata_find(origcpu));
1221 kfree(marker, M_TEMP);
1225 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1226 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1229 tcp_getcred(SYSCTL_HANDLER_ARGS)
1231 struct sockaddr_in addrs[2];
1236 error = priv_check(req->td, PRIV_ROOT);
1239 error = SYSCTL_IN(req, addrs, sizeof addrs);
1243 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1244 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1245 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1246 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1247 if (inp == NULL || inp->inp_socket == NULL) {
1251 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1257 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1258 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1262 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1264 struct sockaddr_in6 addrs[2];
1267 boolean_t mapped = FALSE;
1269 error = priv_check(req->td, PRIV_ROOT);
1272 error = SYSCTL_IN(req, addrs, sizeof addrs);
1275 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1276 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1283 inp = in_pcblookup_hash(&tcbinfo[0],
1284 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1286 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1290 inp = in6_pcblookup_hash(&tcbinfo[0],
1291 &addrs[1].sin6_addr, addrs[1].sin6_port,
1292 &addrs[0].sin6_addr, addrs[0].sin6_port,
1295 if (inp == NULL || inp->inp_socket == NULL) {
1299 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1305 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1307 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1310 struct netmsg_tcp_notify {
1311 struct netmsg_base base;
1312 void (*nm_notify)(struct inpcb *, int);
1313 struct in_addr nm_faddr;
1318 tcp_notifyall_oncpu(netmsg_t msg)
1320 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1323 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nm->nm_faddr,
1324 nm->nm_arg, nm->nm_notify);
1326 nextcpu = mycpuid + 1;
1327 if (nextcpu < ncpus2)
1328 lwkt_forwardmsg(cpu_portfn(nextcpu), &nm->base.lmsg);
1330 lwkt_replymsg(&nm->base.lmsg, 0);
1334 tcp_ctlinput(netmsg_t msg)
1336 int cmd = msg->ctlinput.nm_cmd;
1337 struct sockaddr *sa = msg->ctlinput.nm_arg;
1338 struct ip *ip = msg->ctlinput.nm_extra;
1340 struct in_addr faddr;
1343 void (*notify)(struct inpcb *, int) = tcp_notify;
1347 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1351 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1352 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1355 arg = inetctlerrmap[cmd];
1356 if (cmd == PRC_QUENCH) {
1357 notify = tcp_quench;
1358 } else if (icmp_may_rst &&
1359 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1360 cmd == PRC_UNREACH_PORT ||
1361 cmd == PRC_TIMXCEED_INTRANS) &&
1363 notify = tcp_drop_syn_sent;
1364 } else if (cmd == PRC_MSGSIZE) {
1365 struct icmp *icmp = (struct icmp *)
1366 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1368 arg = ntohs(icmp->icmp_nextmtu);
1369 notify = tcp_mtudisc;
1370 } else if (PRC_IS_REDIRECT(cmd)) {
1372 notify = in_rtchange;
1373 } else if (cmd == PRC_HOSTDEAD) {
1379 th = (struct tcphdr *)((caddr_t)ip +
1380 (IP_VHL_HL(ip->ip_vhl) << 2));
1381 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1382 ip->ip_src.s_addr, th->th_sport);
1383 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1384 ip->ip_src, th->th_sport, 0, NULL);
1385 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1386 icmpseq = htonl(th->th_seq);
1387 tp = intotcpcb(inp);
1388 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1389 SEQ_LT(icmpseq, tp->snd_max))
1390 (*notify)(inp, arg);
1392 struct in_conninfo inc;
1394 inc.inc_fport = th->th_dport;
1395 inc.inc_lport = th->th_sport;
1396 inc.inc_faddr = faddr;
1397 inc.inc_laddr = ip->ip_src;
1401 syncache_unreach(&inc, th);
1405 struct netmsg_tcp_notify *nm;
1407 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1408 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1409 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1410 0, tcp_notifyall_oncpu);
1411 nm->nm_faddr = faddr;
1413 nm->nm_notify = notify;
1415 lwkt_sendmsg(cpu_portfn(0), &nm->base.lmsg);
1418 lwkt_replymsg(&msg->lmsg, 0);
1424 tcp6_ctlinput(netmsg_t msg)
1426 int cmd = msg->ctlinput.nm_cmd;
1427 struct sockaddr *sa = msg->ctlinput.nm_arg;
1428 void *d = msg->ctlinput.nm_extra;
1430 void (*notify) (struct inpcb *, int) = tcp_notify;
1431 struct ip6_hdr *ip6;
1433 struct ip6ctlparam *ip6cp = NULL;
1434 const struct sockaddr_in6 *sa6_src = NULL;
1436 struct tcp_portonly {
1442 if (sa->sa_family != AF_INET6 ||
1443 sa->sa_len != sizeof(struct sockaddr_in6)) {
1448 if (cmd == PRC_QUENCH)
1449 notify = tcp_quench;
1450 else if (cmd == PRC_MSGSIZE) {
1451 struct ip6ctlparam *ip6cp = d;
1452 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1454 arg = ntohl(icmp6->icmp6_mtu);
1455 notify = tcp_mtudisc;
1456 } else if (!PRC_IS_REDIRECT(cmd) &&
1457 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1461 /* if the parameter is from icmp6, decode it. */
1463 ip6cp = (struct ip6ctlparam *)d;
1465 ip6 = ip6cp->ip6c_ip6;
1466 off = ip6cp->ip6c_off;
1467 sa6_src = ip6cp->ip6c_src;
1471 off = 0; /* fool gcc */
1476 struct in_conninfo inc;
1478 * XXX: We assume that when IPV6 is non NULL,
1479 * M and OFF are valid.
1482 /* check if we can safely examine src and dst ports */
1483 if (m->m_pkthdr.len < off + sizeof *thp)
1486 bzero(&th, sizeof th);
1487 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1489 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1490 (struct sockaddr *)ip6cp->ip6c_src,
1491 th.th_sport, cmd, arg, notify);
1493 inc.inc_fport = th.th_dport;
1494 inc.inc_lport = th.th_sport;
1495 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1496 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1498 syncache_unreach(&inc, &th);
1500 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1501 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1504 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1510 * Following is where TCP initial sequence number generation occurs.
1512 * There are two places where we must use initial sequence numbers:
1513 * 1. In SYN-ACK packets.
1514 * 2. In SYN packets.
1516 * All ISNs for SYN-ACK packets are generated by the syncache. See
1517 * tcp_syncache.c for details.
1519 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1520 * depends on this property. In addition, these ISNs should be
1521 * unguessable so as to prevent connection hijacking. To satisfy
1522 * the requirements of this situation, the algorithm outlined in
1523 * RFC 1948 is used to generate sequence numbers.
1525 * Implementation details:
1527 * Time is based off the system timer, and is corrected so that it
1528 * increases by one megabyte per second. This allows for proper
1529 * recycling on high speed LANs while still leaving over an hour
1532 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1533 * between seeding of isn_secret. This is normally set to zero,
1534 * as reseeding should not be necessary.
1538 #define ISN_BYTES_PER_SECOND 1048576
1540 u_char isn_secret[32];
1541 int isn_last_reseed;
1545 tcp_new_isn(struct tcpcb *tp)
1547 u_int32_t md5_buffer[4];
1550 /* Seed if this is the first use, reseed if requested. */
1551 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1552 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1554 read_random_unlimited(&isn_secret, sizeof isn_secret);
1555 isn_last_reseed = ticks;
1558 /* Compute the md5 hash and return the ISN. */
1560 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1561 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1563 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1564 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1565 sizeof(struct in6_addr));
1566 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1567 sizeof(struct in6_addr));
1571 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1572 sizeof(struct in_addr));
1573 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1574 sizeof(struct in_addr));
1576 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1577 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1578 new_isn = (tcp_seq) md5_buffer[0];
1579 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1584 * When a source quench is received, close congestion window
1585 * to one segment. We will gradually open it again as we proceed.
1588 tcp_quench(struct inpcb *inp, int error)
1590 struct tcpcb *tp = intotcpcb(inp);
1593 tp->snd_cwnd = tp->t_maxseg;
1599 * When a specific ICMP unreachable message is received and the
1600 * connection state is SYN-SENT, drop the connection. This behavior
1601 * is controlled by the icmp_may_rst sysctl.
1604 tcp_drop_syn_sent(struct inpcb *inp, int error)
1606 struct tcpcb *tp = intotcpcb(inp);
1608 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1609 tcp_drop(tp, error);
1613 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1614 * based on the new value in the route. Also nudge TCP to send something,
1615 * since we know the packet we just sent was dropped.
1616 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1619 tcp_mtudisc(struct inpcb *inp, int mtu)
1621 struct tcpcb *tp = intotcpcb(inp);
1623 struct socket *so = inp->inp_socket;
1626 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1628 const boolean_t isipv6 = FALSE;
1635 * If no MTU is provided in the ICMP message, use the
1636 * next lower likely value, as specified in RFC 1191.
1641 oldmtu = tp->t_maxopd +
1643 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1644 sizeof(struct tcpiphdr));
1645 mtu = ip_next_mtu(oldmtu, 0);
1649 rt = tcp_rtlookup6(&inp->inp_inc);
1651 rt = tcp_rtlookup(&inp->inp_inc);
1653 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1654 mtu = rt->rt_rmx.rmx_mtu;
1658 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1659 sizeof(struct tcpiphdr));
1662 * XXX - The following conditional probably violates the TCP
1663 * spec. The problem is that, since we don't know the
1664 * other end's MSS, we are supposed to use a conservative
1665 * default. But, if we do that, then MTU discovery will
1666 * never actually take place, because the conservative
1667 * default is much less than the MTUs typically seen
1668 * on the Internet today. For the moment, we'll sweep
1669 * this under the carpet.
1671 * The conservative default might not actually be a problem
1672 * if the only case this occurs is when sending an initial
1673 * SYN with options and data to a host we've never talked
1674 * to before. Then, they will reply with an MSS value which
1675 * will get recorded and the new parameters should get
1676 * recomputed. For Further Study.
1678 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1679 maxopd = rt->rt_rmx.rmx_mssopt;
1683 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1684 sizeof(struct tcpiphdr));
1686 if (tp->t_maxopd <= maxopd)
1688 tp->t_maxopd = maxopd;
1691 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1692 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1693 mss -= TCPOLEN_TSTAMP_APPA;
1695 /* round down to multiple of MCLBYTES */
1696 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1698 mss &= ~(MCLBYTES - 1);
1701 mss = (mss / MCLBYTES) * MCLBYTES;
1704 if (so->so_snd.ssb_hiwat < mss)
1705 mss = so->so_snd.ssb_hiwat;
1709 tp->snd_nxt = tp->snd_una;
1711 tcpstat.tcps_mturesent++;
1715 * Look-up the routing entry to the peer of this inpcb. If no route
1716 * is found and it cannot be allocated the return NULL. This routine
1717 * is called by TCP routines that access the rmx structure and by tcp_mss
1718 * to get the interface MTU.
1721 tcp_rtlookup(struct in_conninfo *inc)
1723 struct route *ro = &inc->inc_route;
1725 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1726 /* No route yet, so try to acquire one */
1727 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1729 * unused portions of the structure MUST be zero'd
1730 * out because rtalloc() treats it as opaque data
1732 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1733 ro->ro_dst.sa_family = AF_INET;
1734 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1735 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1745 tcp_rtlookup6(struct in_conninfo *inc)
1747 struct route_in6 *ro6 = &inc->inc6_route;
1749 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1750 /* No route yet, so try to acquire one */
1751 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1753 * unused portions of the structure MUST be zero'd
1754 * out because rtalloc() treats it as opaque data
1756 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1757 ro6->ro_dst.sin6_family = AF_INET6;
1758 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1759 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1760 rtalloc((struct route *)ro6);
1763 return (ro6->ro_rt);
1768 /* compute ESP/AH header size for TCP, including outer IP header. */
1770 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1778 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1780 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1785 if (inp->inp_vflag & INP_IPV6) {
1786 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1788 th = (struct tcphdr *)(ip6 + 1);
1789 m->m_pkthdr.len = m->m_len =
1790 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1791 tcp_fillheaders(tp, ip6, th);
1792 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1796 ip = mtod(m, struct ip *);
1797 th = (struct tcphdr *)(ip + 1);
1798 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1799 tcp_fillheaders(tp, ip, th);
1800 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1809 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1811 * This code attempts to calculate the bandwidth-delay product as a
1812 * means of determining the optimal window size to maximize bandwidth,
1813 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1814 * routers. This code also does a fairly good job keeping RTTs in check
1815 * across slow links like modems. We implement an algorithm which is very
1816 * similar (but not meant to be) TCP/Vegas. The code operates on the
1817 * transmitter side of a TCP connection and so only effects the transmit
1818 * side of the connection.
1820 * BACKGROUND: TCP makes no provision for the management of buffer space
1821 * at the end points or at the intermediate routers and switches. A TCP
1822 * stream, whether using NewReno or not, will eventually buffer as
1823 * many packets as it is able and the only reason this typically works is
1824 * due to the fairly small default buffers made available for a connection
1825 * (typicaly 16K or 32K). As machines use larger windows and/or window
1826 * scaling it is now fairly easy for even a single TCP connection to blow-out
1827 * all available buffer space not only on the local interface, but on
1828 * intermediate routers and switches as well. NewReno makes a misguided
1829 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1830 * then backing off, then steadily increasing the window again until another
1831 * failure occurs, ad-infinitum. This results in terrible oscillation that
1832 * is only made worse as network loads increase and the idea of intentionally
1833 * blowing out network buffers is, frankly, a terrible way to manage network
1836 * It is far better to limit the transmit window prior to the failure
1837 * condition being achieved. There are two general ways to do this: First
1838 * you can 'scan' through different transmit window sizes and locate the
1839 * point where the RTT stops increasing, indicating that you have filled the
1840 * pipe, then scan backwards until you note that RTT stops decreasing, then
1841 * repeat ad-infinitum. This method works in principle but has severe
1842 * implementation issues due to RTT variances, timer granularity, and
1843 * instability in the algorithm which can lead to many false positives and
1844 * create oscillations as well as interact badly with other TCP streams
1845 * implementing the same algorithm.
1847 * The second method is to limit the window to the bandwidth delay product
1848 * of the link. This is the method we implement. RTT variances and our
1849 * own manipulation of the congestion window, bwnd, can potentially
1850 * destabilize the algorithm. For this reason we have to stabilize the
1851 * elements used to calculate the window. We do this by using the minimum
1852 * observed RTT, the long term average of the observed bandwidth, and
1853 * by adding two segments worth of slop. It isn't perfect but it is able
1854 * to react to changing conditions and gives us a very stable basis on
1855 * which to extend the algorithm.
1858 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1866 * If inflight_enable is disabled in the middle of a tcp connection,
1867 * make sure snd_bwnd is effectively disabled.
1869 if (!tcp_inflight_enable) {
1870 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1871 tp->snd_bandwidth = 0;
1876 * Validate the delta time. If a connection is new or has been idle
1877 * a long time we have to reset the bandwidth calculator.
1880 delta_ticks = save_ticks - tp->t_bw_rtttime;
1881 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1882 tp->t_bw_rtttime = ticks;
1883 tp->t_bw_rtseq = ack_seq;
1884 if (tp->snd_bandwidth == 0)
1885 tp->snd_bandwidth = tcp_inflight_min;
1888 if (delta_ticks == 0)
1892 * Sanity check, plus ignore pure window update acks.
1894 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1898 * Figure out the bandwidth. Due to the tick granularity this
1899 * is a very rough number and it MUST be averaged over a fairly
1900 * long period of time. XXX we need to take into account a link
1901 * that is not using all available bandwidth, but for now our
1902 * slop will ramp us up if this case occurs and the bandwidth later
1905 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1906 tp->t_bw_rtttime = save_ticks;
1907 tp->t_bw_rtseq = ack_seq;
1908 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1910 tp->snd_bandwidth = bw;
1913 * Calculate the semi-static bandwidth delay product, plus two maximal
1914 * segments. The additional slop puts us squarely in the sweet
1915 * spot and also handles the bandwidth run-up case. Without the
1916 * slop we could be locking ourselves into a lower bandwidth.
1918 * Situations Handled:
1919 * (1) Prevents over-queueing of packets on LANs, especially on
1920 * high speed LANs, allowing larger TCP buffers to be
1921 * specified, and also does a good job preventing
1922 * over-queueing of packets over choke points like modems
1923 * (at least for the transmit side).
1925 * (2) Is able to handle changing network loads (bandwidth
1926 * drops so bwnd drops, bandwidth increases so bwnd
1929 * (3) Theoretically should stabilize in the face of multiple
1930 * connections implementing the same algorithm (this may need
1933 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1934 * be adjusted with a sysctl but typically only needs to be on
1935 * very slow connections. A value no smaller then 5 should
1936 * be used, but only reduce this default if you have no other
1940 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1941 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
1942 tcp_inflight_stab * (int)tp->t_maxseg / 10;
1945 if (tcp_inflight_debug > 0) {
1947 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
1949 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
1950 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
1953 if ((long)bwnd < tcp_inflight_min)
1954 bwnd = tcp_inflight_min;
1955 if (bwnd > tcp_inflight_max)
1956 bwnd = tcp_inflight_max;
1957 if ((long)bwnd < tp->t_maxseg * 2)
1958 bwnd = tp->t_maxseg * 2;
1959 tp->snd_bwnd = bwnd;
1962 #ifdef TCP_SIGNATURE
1964 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
1966 * We do this over ip, tcphdr, segment data, and the key in the SADB.
1967 * When called from tcp_input(), we can be sure that th_sum has been
1968 * zeroed out and verified already.
1970 * Return 0 if successful, otherwise return -1.
1972 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
1973 * search with the destination IP address, and a 'magic SPI' to be
1974 * determined by the application. This is hardcoded elsewhere to 1179
1975 * right now. Another branch of this code exists which uses the SPD to
1976 * specify per-application flows but it is unstable.
1979 tcpsignature_compute(
1980 struct mbuf *m, /* mbuf chain */
1981 int len, /* length of TCP data */
1982 int optlen, /* length of TCP options */
1983 u_char *buf, /* storage for MD5 digest */
1984 u_int direction) /* direction of flow */
1986 struct ippseudo ippseudo;
1990 struct ipovly *ipovly;
1991 struct secasvar *sav;
1994 struct ip6_hdr *ip6;
1995 struct in6_addr in6;
2001 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2002 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2004 * Extract the destination from the IP header in the mbuf.
2006 ip = mtod(m, struct ip *);
2008 ip6 = NULL; /* Make the compiler happy. */
2011 * Look up an SADB entry which matches the address found in
2014 switch (IP_VHL_V(ip->ip_vhl)) {
2016 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2017 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2020 case (IPV6_VERSION >> 4):
2021 ip6 = mtod(m, struct ip6_hdr *);
2022 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2023 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2032 kprintf("%s: SADB lookup failed\n", __func__);
2038 * Step 1: Update MD5 hash with IP pseudo-header.
2040 * XXX The ippseudo header MUST be digested in network byte order,
2041 * or else we'll fail the regression test. Assume all fields we've
2042 * been doing arithmetic on have been in host byte order.
2043 * XXX One cannot depend on ipovly->ih_len here. When called from
2044 * tcp_output(), the underlying ip_len member has not yet been set.
2046 switch (IP_VHL_V(ip->ip_vhl)) {
2048 ipovly = (struct ipovly *)ip;
2049 ippseudo.ippseudo_src = ipovly->ih_src;
2050 ippseudo.ippseudo_dst = ipovly->ih_dst;
2051 ippseudo.ippseudo_pad = 0;
2052 ippseudo.ippseudo_p = IPPROTO_TCP;
2053 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2054 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2055 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2056 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2060 * RFC 2385, 2.0 Proposal
2061 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2062 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2063 * extended next header value (to form 32 bits), and 32-bit segment
2065 * Note: Upper-Layer Packet Length comes before Next Header.
2067 case (IPV6_VERSION >> 4):
2069 in6_clearscope(&in6);
2070 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2072 in6_clearscope(&in6);
2073 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2074 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2075 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2077 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2078 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2079 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2081 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2082 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2083 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2092 * Step 2: Update MD5 hash with TCP header, excluding options.
2093 * The TCP checksum must be set to zero.
2095 savecsum = th->th_sum;
2097 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2098 th->th_sum = savecsum;
2100 * Step 3: Update MD5 hash with TCP segment data.
2101 * Use m_apply() to avoid an early m_pullup().
2104 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2106 * Step 4: Update MD5 hash with shared secret.
2108 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2109 MD5Final(buf, &ctx);
2110 key_sa_recordxfer(sav, m);
2116 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2119 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2122 #endif /* TCP_SIGNATURE */