2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of The DragonFly Project nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific, prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the University of
49 * California, Berkeley and its contributors.
50 * 4. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
66 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
67 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
68 * $DragonFly: src/sys/netinet/tcp_subr.c,v 1.63 2008/11/11 10:46:58 sephe Exp $
71 #include "opt_compat.h"
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/callout.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 #include <sys/malloc.h>
82 #include <sys/mpipe.h>
85 #include <sys/domain.h>
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 <vm/vm_zone.h>
98 #include <net/route.h>
100 #include <net/netisr.h>
103 #include <netinet/in.h>
104 #include <netinet/in_systm.h>
105 #include <netinet/ip.h>
106 #include <netinet/ip6.h>
107 #include <netinet/in_pcb.h>
108 #include <netinet6/in6_pcb.h>
109 #include <netinet/in_var.h>
110 #include <netinet/ip_var.h>
111 #include <netinet6/ip6_var.h>
112 #include <netinet/ip_icmp.h>
114 #include <netinet/icmp6.h>
116 #include <netinet/tcp.h>
117 #include <netinet/tcp_fsm.h>
118 #include <netinet/tcp_seq.h>
119 #include <netinet/tcp_timer.h>
120 #include <netinet/tcp_timer2.h>
121 #include <netinet/tcp_var.h>
122 #include <netinet6/tcp6_var.h>
123 #include <netinet/tcpip.h>
125 #include <netinet/tcp_debug.h>
127 #include <netinet6/ip6protosw.h>
130 #include <netinet6/ipsec.h>
132 #include <netinet6/ipsec6.h>
137 #include <netproto/ipsec/ipsec.h>
139 #include <netproto/ipsec/ipsec6.h>
145 #include <sys/msgport2.h>
146 #include <machine/smp.h>
148 #include <net/netmsg2.h>
150 #if !defined(KTR_TCP)
151 #define KTR_TCP KTR_ALL
153 KTR_INFO_MASTER(tcp);
154 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
155 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
156 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
157 #define logtcp(name) KTR_LOG(tcp_ ## name)
159 struct inpcbinfo tcbinfo[MAXCPU];
160 struct tcpcbackqhead tcpcbackq[MAXCPU];
162 int tcp_mpsafe_proto = 0;
163 TUNABLE_INT("net.inet.tcp.mpsafe_proto", &tcp_mpsafe_proto);
165 static int tcp_mpsafe_thread = NETMSG_SERVICE_ADAPTIVE;
166 TUNABLE_INT("net.inet.tcp.mpsafe_thread", &tcp_mpsafe_thread);
167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, mpsafe_thread, CTLFLAG_RW,
168 &tcp_mpsafe_thread, 0,
169 "0:BGL, 1:Adaptive BGL, 2:No BGL(experimental)");
171 int tcp_mssdflt = TCP_MSS;
172 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
173 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
176 int tcp_v6mssdflt = TCP6_MSS;
177 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
178 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
182 * Minimum MSS we accept and use. This prevents DoS attacks where
183 * we are forced to a ridiculous low MSS like 20 and send hundreds
184 * of packets instead of one. The effect scales with the available
185 * bandwidth and quickly saturates the CPU and network interface
186 * with packet generation and sending. Set to zero to disable MINMSS
187 * checking. This setting prevents us from sending too small packets.
189 int tcp_minmss = TCP_MINMSS;
190 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
191 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
194 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
195 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
196 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
199 int tcp_do_rfc1323 = 1;
200 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
201 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
203 int tcp_do_rfc1644 = 0;
204 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
205 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
207 static int tcp_tcbhashsize = 0;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
209 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
211 static int do_tcpdrain = 1;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
213 "Enable tcp_drain routine for extra help when low on mbufs");
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
217 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
219 static int icmp_may_rst = 1;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
221 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
223 static int tcp_isn_reseed_interval = 0;
224 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
225 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
228 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
229 * by default, but with generous values which should allow maximal
230 * bandwidth. In particular, the slop defaults to 50 (5 packets).
232 * The reason for doing this is that the limiter is the only mechanism we
233 * have which seems to do a really good job preventing receiver RX rings
234 * on network interfaces from getting blown out. Even though GigE/10GigE
235 * is supposed to flow control it looks like either it doesn't actually
236 * do it or Open Source drivers do not properly enable it.
238 * People using the limiter to reduce bottlenecks on slower WAN connections
239 * should set the slop to 20 (2 packets).
241 static int tcp_inflight_enable = 1;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
243 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
245 static int tcp_inflight_debug = 0;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
247 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
249 static int tcp_inflight_min = 6144;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
251 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
253 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
254 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
255 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
257 static int tcp_inflight_stab = 50;
258 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
259 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 3 packets)");
261 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
262 static struct malloc_pipe tcptemp_mpipe;
264 static void tcp_willblock(int);
265 static void tcp_cleartaocache (void);
266 static void tcp_notify (struct inpcb *, int);
268 struct tcp_stats tcpstats_percpu[MAXCPU];
271 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
275 for (cpu = 0; cpu < ncpus; ++cpu) {
276 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
277 sizeof(struct tcp_stats))))
279 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
280 sizeof(struct tcp_stats))))
286 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
287 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
289 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
290 &tcpstat, tcp_stats, "TCP statistics");
294 * Target size of TCP PCB hash tables. Must be a power of two.
296 * Note that this can be overridden by the kernel environment
297 * variable net.inet.tcp.tcbhashsize
300 #define TCBHASHSIZE 512
304 * This is the actual shape of what we allocate using the zone
305 * allocator. Doing it this way allows us to protect both structures
306 * using the same generation count, and also eliminates the overhead
307 * of allocating tcpcbs separately. By hiding the structure here,
308 * we avoid changing most of the rest of the code (although it needs
309 * to be changed, eventually, for greater efficiency).
312 #define ALIGNM1 (ALIGNMENT - 1)
316 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
319 struct tcp_callout inp_tp_rexmt;
320 struct tcp_callout inp_tp_persist;
321 struct tcp_callout inp_tp_keep;
322 struct tcp_callout inp_tp_2msl;
323 struct tcp_callout inp_tp_delack;
324 struct netmsg_tcp_timer inp_tp_timermsg;
335 struct inpcbporthead *porthashbase;
337 struct vm_zone *ipi_zone;
338 int hashsize = TCBHASHSIZE;
342 * note: tcptemp is used for keepalives, and it is ok for an
343 * allocation to fail so do not specify MPF_INT.
345 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
351 tcp_delacktime = TCPTV_DELACK;
352 tcp_keepinit = TCPTV_KEEP_INIT;
353 tcp_keepidle = TCPTV_KEEP_IDLE;
354 tcp_keepintvl = TCPTV_KEEPINTVL;
355 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
357 tcp_rexmit_min = TCPTV_MIN;
358 tcp_rexmit_slop = TCPTV_CPU_VAR;
360 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
361 if (!powerof2(hashsize)) {
362 kprintf("WARNING: TCB hash size not a power of 2\n");
363 hashsize = 512; /* safe default */
365 tcp_tcbhashsize = hashsize;
366 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
367 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
370 for (cpu = 0; cpu < ncpus2; cpu++) {
371 in_pcbinfo_init(&tcbinfo[cpu]);
372 tcbinfo[cpu].cpu = cpu;
373 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
374 &tcbinfo[cpu].hashmask);
375 tcbinfo[cpu].porthashbase = porthashbase;
376 tcbinfo[cpu].porthashmask = porthashmask;
377 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
378 &tcbinfo[cpu].wildcardhashmask);
379 tcbinfo[cpu].ipi_zone = ipi_zone;
380 TAILQ_INIT(&tcpcbackq[cpu]);
383 tcp_reass_maxseg = nmbclusters / 16;
384 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
387 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
389 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
391 if (max_protohdr < TCP_MINPROTOHDR)
392 max_protohdr = TCP_MINPROTOHDR;
393 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
395 #undef TCP_MINPROTOHDR
398 * Initialize TCP statistics counters for each CPU.
401 for (cpu = 0; cpu < ncpus; ++cpu) {
402 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
405 bzero(&tcpstat, sizeof(struct tcp_stats));
413 tcpmsg_service_loop(void *dummy)
419 * Thread was started with TDF_MPSAFE
423 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
426 mplocked = netmsg_service(msg, tcp_mpsafe_thread,
428 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
431 tcp_willblock(mplocked);
437 tcp_willblock(int mplocked)
440 int cpu = mycpu->gd_cpuid;
443 if (!mplocked && !tcp_mpsafe_proto) {
444 if (TAILQ_EMPTY(&tcpcbackq[cpu]))
452 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
453 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
454 tp->t_flags &= ~TF_ONOUTPUTQ;
455 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
465 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
466 * tcp_template used to store this data in mbufs, but we now recopy it out
467 * of the tcpcb each time to conserve mbufs.
470 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
472 struct inpcb *inp = tp->t_inpcb;
473 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
476 if (inp->inp_vflag & INP_IPV6) {
479 ip6 = (struct ip6_hdr *)ip_ptr;
480 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
481 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
482 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
483 (IPV6_VERSION & IPV6_VERSION_MASK);
484 ip6->ip6_nxt = IPPROTO_TCP;
485 ip6->ip6_plen = sizeof(struct tcphdr);
486 ip6->ip6_src = inp->in6p_laddr;
487 ip6->ip6_dst = inp->in6p_faddr;
492 struct ip *ip = (struct ip *) ip_ptr;
494 ip->ip_vhl = IP_VHL_BORING;
501 ip->ip_p = IPPROTO_TCP;
502 ip->ip_src = inp->inp_laddr;
503 ip->ip_dst = inp->inp_faddr;
504 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
506 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
509 tcp_hdr->th_sport = inp->inp_lport;
510 tcp_hdr->th_dport = inp->inp_fport;
515 tcp_hdr->th_flags = 0;
521 * Create template to be used to send tcp packets on a connection.
522 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
523 * use for this function is in keepalives, which use tcp_respond.
526 tcp_maketemplate(struct tcpcb *tp)
530 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
532 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
537 tcp_freetemplate(struct tcptemp *tmp)
539 mpipe_free(&tcptemp_mpipe, tmp);
543 * Send a single message to the TCP at address specified by
544 * the given TCP/IP header. If m == NULL, then we make a copy
545 * of the tcpiphdr at ti and send directly to the addressed host.
546 * This is used to force keep alive messages out using the TCP
547 * template for a connection. If flags are given then we send
548 * a message back to the TCP which originated the * segment ti,
549 * and discard the mbuf containing it and any other attached mbufs.
551 * In any case the ack and sequence number of the transmitted
552 * segment are as specified by the parameters.
554 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
557 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
558 tcp_seq ack, tcp_seq seq, int flags)
562 struct route *ro = NULL;
564 struct ip *ip = ipgen;
567 struct route_in6 *ro6 = NULL;
568 struct route_in6 sro6;
569 struct ip6_hdr *ip6 = ipgen;
570 boolean_t use_tmpro = TRUE;
572 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
574 const boolean_t isipv6 = FALSE;
578 if (!(flags & TH_RST)) {
579 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
582 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
583 win = (long)TCP_MAXWIN << tp->rcv_scale;
586 * Don't use the route cache of a listen socket,
587 * it is not MPSAFE; use temporary route cache.
589 if (tp->t_state != TCPS_LISTEN) {
591 ro6 = &tp->t_inpcb->in6p_route;
593 ro = &tp->t_inpcb->inp_route;
600 bzero(ro6, sizeof *ro6);
603 bzero(ro, sizeof *ro);
607 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
611 m->m_data += max_linkhdr;
613 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
614 ip6 = mtod(m, struct ip6_hdr *);
615 nth = (struct tcphdr *)(ip6 + 1);
617 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
618 ip = mtod(m, struct ip *);
619 nth = (struct tcphdr *)(ip + 1);
621 bcopy(th, nth, sizeof(struct tcphdr));
626 m->m_data = (caddr_t)ipgen;
627 /* m_len is set later */
629 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
631 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
632 nth = (struct tcphdr *)(ip6 + 1);
634 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
635 nth = (struct tcphdr *)(ip + 1);
639 * this is usually a case when an extension header
640 * exists between the IPv6 header and the
643 nth->th_sport = th->th_sport;
644 nth->th_dport = th->th_dport;
646 xchg(nth->th_dport, nth->th_sport, n_short);
651 ip6->ip6_vfc = IPV6_VERSION;
652 ip6->ip6_nxt = IPPROTO_TCP;
653 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
654 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
656 tlen += sizeof(struct tcpiphdr);
658 ip->ip_ttl = ip_defttl;
661 m->m_pkthdr.len = tlen;
662 m->m_pkthdr.rcvif = NULL;
663 nth->th_seq = htonl(seq);
664 nth->th_ack = htonl(ack);
666 nth->th_off = sizeof(struct tcphdr) >> 2;
667 nth->th_flags = flags;
669 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
671 nth->th_win = htons((u_short)win);
675 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
676 sizeof(struct ip6_hdr),
677 tlen - sizeof(struct ip6_hdr));
678 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
679 (ro6 && ro6->ro_rt) ?
680 ro6->ro_rt->rt_ifp : NULL);
682 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
683 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
684 m->m_pkthdr.csum_flags = CSUM_TCP;
685 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
688 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
689 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
692 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
693 tp ? tp->t_inpcb : NULL);
694 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
699 ipflags |= IP_DEBUGROUTE;
700 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
701 if ((ro == &sro) && (ro->ro_rt != NULL)) {
709 * Create a new TCP control block, making an
710 * empty reassembly queue and hooking it to the argument
711 * protocol control block. The `inp' parameter must have
712 * come from the zone allocator set up in tcp_init().
715 tcp_newtcpcb(struct inpcb *inp)
720 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
722 const boolean_t isipv6 = FALSE;
725 it = (struct inp_tp *)inp;
727 bzero(tp, sizeof(struct tcpcb));
728 LIST_INIT(&tp->t_segq);
729 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
731 /* Set up our timeouts. */
732 tp->tt_rexmt = &it->inp_tp_rexmt;
733 tp->tt_persist = &it->inp_tp_persist;
734 tp->tt_keep = &it->inp_tp_keep;
735 tp->tt_2msl = &it->inp_tp_2msl;
736 tp->tt_delack = &it->inp_tp_delack;
740 * Zero out timer message. We don't create it here,
741 * since the current CPU may not be the owner of this
744 tp->tt_msg = &it->inp_tp_timermsg;
745 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
748 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
750 tp->t_flags |= TF_REQ_CC;
751 tp->t_inpcb = inp; /* XXX */
752 tp->t_state = TCPS_CLOSED;
754 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
755 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
756 * reasonable initial retransmit time.
758 tp->t_srtt = TCPTV_SRTTBASE;
760 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
761 tp->t_rttmin = tcp_rexmit_min;
762 tp->t_rxtcur = TCPTV_RTOBASE;
763 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
764 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
765 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
766 tp->t_rcvtime = ticks;
768 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
769 * because the socket may be bound to an IPv6 wildcard address,
770 * which may match an IPv4-mapped IPv6 address.
772 inp->inp_ip_ttl = ip_defttl;
774 tcp_sack_tcpcb_init(tp);
775 return (tp); /* XXX */
779 * Drop a TCP connection, reporting the specified error.
780 * If connection is synchronized, then send a RST to peer.
783 tcp_drop(struct tcpcb *tp, int error)
785 struct socket *so = tp->t_inpcb->inp_socket;
787 if (TCPS_HAVERCVDSYN(tp->t_state)) {
788 tp->t_state = TCPS_CLOSED;
790 tcpstat.tcps_drops++;
792 tcpstat.tcps_conndrops++;
793 if (error == ETIMEDOUT && tp->t_softerror)
794 error = tp->t_softerror;
795 so->so_error = error;
796 return (tcp_close(tp));
801 struct netmsg_remwildcard {
802 struct netmsg nm_netmsg;
803 struct inpcb *nm_inp;
804 struct inpcbinfo *nm_pcbinfo;
813 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
814 * inp can be detached. We do this by cycling through the cpus, ending up
815 * on the cpu controlling the inp last and then doing the disconnect.
818 in_pcbremwildcardhash_handler(struct netmsg *msg0)
820 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
823 cpu = msg->nm_pcbinfo->cpu;
825 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
826 /* note: detach removes any wildcard hash entry */
829 in6_pcbdetach(msg->nm_inp);
832 in_pcbdetach(msg->nm_inp);
833 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
835 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
836 cpu = (cpu + 1) % ncpus2;
837 msg->nm_pcbinfo = &tcbinfo[cpu];
838 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
845 * Close a TCP control block:
846 * discard all space held by the tcp
847 * discard internet protocol block
848 * wake up any sleepers
851 tcp_close(struct tcpcb *tp)
854 struct inpcb *inp = tp->t_inpcb;
855 struct socket *so = inp->inp_socket;
857 boolean_t dosavessthresh;
862 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
863 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
865 const boolean_t isipv6 = FALSE;
869 * The tp is not instantly destroyed in the wildcard case. Setting
870 * the state to TCPS_TERMINATING will prevent the TCP stack from
871 * messing with it, though it should be noted that this change may
872 * not take effect on other cpus until we have chained the wildcard
875 * XXX we currently depend on the BGL to synchronize the tp->t_state
876 * update and prevent other tcp protocol threads from accepting new
877 * connections on the listen socket we might be trying to close down.
879 KKASSERT(tp->t_state != TCPS_TERMINATING);
880 tp->t_state = TCPS_TERMINATING;
883 * Make sure that all of our timers are stopped before we
884 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
885 * timers are never used. If timer message is never created
886 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
888 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
889 tcp_callout_stop(tp, tp->tt_rexmt);
890 tcp_callout_stop(tp, tp->tt_persist);
891 tcp_callout_stop(tp, tp->tt_keep);
892 tcp_callout_stop(tp, tp->tt_2msl);
893 tcp_callout_stop(tp, tp->tt_delack);
896 if (tp->t_flags & TF_ONOUTPUTQ) {
897 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
898 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
899 tp->t_flags &= ~TF_ONOUTPUTQ;
903 * If we got enough samples through the srtt filter,
904 * save the rtt and rttvar in the routing entry.
905 * 'Enough' is arbitrarily defined as the 16 samples.
906 * 16 samples is enough for the srtt filter to converge
907 * to within 5% of the correct value; fewer samples and
908 * we could save a very bogus rtt.
910 * Don't update the default route's characteristics and don't
911 * update anything that the user "locked".
913 if (tp->t_rttupdated >= 16) {
917 struct sockaddr_in6 *sin6;
919 if ((rt = inp->in6p_route.ro_rt) == NULL)
921 sin6 = (struct sockaddr_in6 *)rt_key(rt);
922 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
925 if ((rt = inp->inp_route.ro_rt) == NULL ||
926 ((struct sockaddr_in *)rt_key(rt))->
927 sin_addr.s_addr == INADDR_ANY)
930 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
931 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
932 if (rt->rt_rmx.rmx_rtt && i)
934 * filter this update to half the old & half
935 * the new values, converting scale.
936 * See route.h and tcp_var.h for a
937 * description of the scaling constants.
940 (rt->rt_rmx.rmx_rtt + i) / 2;
942 rt->rt_rmx.rmx_rtt = i;
943 tcpstat.tcps_cachedrtt++;
945 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
947 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
948 if (rt->rt_rmx.rmx_rttvar && i)
949 rt->rt_rmx.rmx_rttvar =
950 (rt->rt_rmx.rmx_rttvar + i) / 2;
952 rt->rt_rmx.rmx_rttvar = i;
953 tcpstat.tcps_cachedrttvar++;
956 * The old comment here said:
957 * update the pipelimit (ssthresh) if it has been updated
958 * already or if a pipesize was specified & the threshhold
959 * got below half the pipesize. I.e., wait for bad news
960 * before we start updating, then update on both good
963 * But we want to save the ssthresh even if no pipesize is
964 * specified explicitly in the route, because such
965 * connections still have an implicit pipesize specified
966 * by the global tcp_sendspace. In the absence of a reliable
967 * way to calculate the pipesize, it will have to do.
969 i = tp->snd_ssthresh;
970 if (rt->rt_rmx.rmx_sendpipe != 0)
971 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
973 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
974 if (dosavessthresh ||
975 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
976 (rt->rt_rmx.rmx_ssthresh != 0))) {
978 * convert the limit from user data bytes to
979 * packets then to packet data bytes.
981 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
986 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
987 sizeof(struct tcpiphdr));
988 if (rt->rt_rmx.rmx_ssthresh)
989 rt->rt_rmx.rmx_ssthresh =
990 (rt->rt_rmx.rmx_ssthresh + i) / 2;
992 rt->rt_rmx.rmx_ssthresh = i;
993 tcpstat.tcps_cachedssthresh++;
998 /* free the reassembly queue, if any */
999 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
1000 LIST_REMOVE(q, tqe_q);
1005 /* throw away SACK blocks in scoreboard*/
1006 if (TCP_DO_SACK(tp))
1007 tcp_sack_cleanup(&tp->scb);
1009 inp->inp_ppcb = NULL;
1010 soisdisconnected(so);
1012 tcp_destroy_timermsg(tp);
1015 * Discard the inp. In the SMP case a wildcard inp's hash (created
1016 * by a listen socket or an INADDR_ANY udp socket) is replicated
1017 * for each protocol thread and must be removed in the context of
1018 * that thread. This is accomplished by chaining the message
1021 * If the inp is not wildcarded we simply detach, which will remove
1022 * the any hashes still present for this inp.
1025 if (inp->inp_flags & INP_WILDCARD_MP) {
1026 struct netmsg_remwildcard *msg;
1028 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1029 msg = kmalloc(sizeof(struct netmsg_remwildcard),
1030 M_LWKTMSG, M_INTWAIT);
1031 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1032 in_pcbremwildcardhash_handler);
1034 msg->nm_isinet6 = isafinet6;
1037 msg->nm_pcbinfo = &tcbinfo[cpu];
1038 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1042 /* note: detach removes any wildcard hash entry */
1050 tcpstat.tcps_closed++;
1054 static __inline void
1055 tcp_drain_oncpu(struct inpcbhead *head)
1059 struct tseg_qent *te;
1061 LIST_FOREACH(inpb, head, inp_list) {
1062 if (inpb->inp_flags & INP_PLACEMARKER)
1064 if ((tcpb = intotcpcb(inpb))) {
1065 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1066 LIST_REMOVE(te, tqe_q);
1076 struct netmsg_tcp_drain {
1077 struct netmsg nm_netmsg;
1078 struct inpcbhead *nm_head;
1082 tcp_drain_handler(netmsg_t netmsg)
1084 struct netmsg_tcp_drain *nm = (void *)netmsg;
1086 tcp_drain_oncpu(nm->nm_head);
1087 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1102 * Walk the tcpbs, if existing, and flush the reassembly queue,
1103 * if there is one...
1104 * XXX: The "Net/3" implementation doesn't imply that the TCP
1105 * reassembly queue should be flushed, but in a situation
1106 * where we're really low on mbufs, this is potentially
1110 for (cpu = 0; cpu < ncpus2; cpu++) {
1111 struct netmsg_tcp_drain *msg;
1113 if (cpu == mycpu->gd_cpuid) {
1114 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1116 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1117 M_LWKTMSG, M_NOWAIT);
1120 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1122 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1123 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1127 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1132 * Notify a tcp user of an asynchronous error;
1133 * store error as soft error, but wake up user
1134 * (for now, won't do anything until can select for soft error).
1136 * Do not wake up user since there currently is no mechanism for
1137 * reporting soft errors (yet - a kqueue filter may be added).
1140 tcp_notify(struct inpcb *inp, int error)
1142 struct tcpcb *tp = intotcpcb(inp);
1145 * Ignore some errors if we are hooked up.
1146 * If connection hasn't completed, has retransmitted several times,
1147 * and receives a second error, give up now. This is better
1148 * than waiting a long time to establish a connection that
1149 * can never complete.
1151 if (tp->t_state == TCPS_ESTABLISHED &&
1152 (error == EHOSTUNREACH || error == ENETUNREACH ||
1153 error == EHOSTDOWN)) {
1155 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1157 tcp_drop(tp, error);
1159 tp->t_softerror = error;
1161 wakeup(&so->so_timeo);
1168 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1171 struct inpcb *marker;
1181 * The process of preparing the TCB list is too time-consuming and
1182 * resource-intensive to repeat twice on every request.
1184 if (req->oldptr == NULL) {
1185 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1186 gd = globaldata_find(ccpu);
1187 n += tcbinfo[gd->gd_cpuid].ipi_count;
1189 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1193 if (req->newptr != NULL)
1196 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1197 marker->inp_flags |= INP_PLACEMARKER;
1200 * OK, now we're committed to doing something. Run the inpcb list
1201 * for each cpu in the system and construct the output. Use a
1202 * list placemarker to deal with list changes occuring during
1203 * copyout blockages (but otherwise depend on being on the correct
1204 * cpu to avoid races).
1206 origcpu = mycpu->gd_cpuid;
1207 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1213 cpu_id = (origcpu + ccpu) % ncpus;
1214 if ((smp_active_mask & (1 << cpu_id)) == 0)
1216 rgd = globaldata_find(cpu_id);
1217 lwkt_setcpu_self(rgd);
1219 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1220 n = tcbinfo[cpu_id].ipi_count;
1222 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1224 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1226 * process a snapshot of pcbs, ignoring placemarkers
1227 * and using our own to allow SYSCTL_OUT to block.
1229 LIST_REMOVE(marker, inp_list);
1230 LIST_INSERT_AFTER(inp, marker, inp_list);
1232 if (inp->inp_flags & INP_PLACEMARKER)
1234 if (inp->inp_gencnt > gencnt)
1236 if (prison_xinpcb(req->td, inp))
1239 xt.xt_len = sizeof xt;
1240 bcopy(inp, &xt.xt_inp, sizeof *inp);
1241 inp_ppcb = inp->inp_ppcb;
1242 if (inp_ppcb != NULL)
1243 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1245 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1246 if (inp->inp_socket)
1247 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1248 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1252 LIST_REMOVE(marker, inp_list);
1253 if (error == 0 && i < n) {
1254 bzero(&xt, sizeof xt);
1255 xt.xt_len = sizeof xt;
1257 error = SYSCTL_OUT(req, &xt, sizeof xt);
1266 * Make sure we are on the same cpu we were on originally, since
1267 * higher level callers expect this. Also don't pollute caches with
1268 * migrated userland data by (eventually) returning to userland
1269 * on a different cpu.
1271 lwkt_setcpu_self(globaldata_find(origcpu));
1272 kfree(marker, M_TEMP);
1276 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1277 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1280 tcp_getcred(SYSCTL_HANDLER_ARGS)
1282 struct sockaddr_in addrs[2];
1287 error = priv_check(req->td, PRIV_ROOT);
1290 error = SYSCTL_IN(req, addrs, sizeof addrs);
1294 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1295 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1296 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1297 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1298 if (inp == NULL || inp->inp_socket == NULL) {
1302 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1308 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1309 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1313 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1315 struct sockaddr_in6 addrs[2];
1318 boolean_t mapped = FALSE;
1320 error = priv_check(req->td, PRIV_ROOT);
1323 error = SYSCTL_IN(req, addrs, sizeof addrs);
1326 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1327 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1334 inp = in_pcblookup_hash(&tcbinfo[0],
1335 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1337 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1341 inp = in6_pcblookup_hash(&tcbinfo[0],
1342 &addrs[1].sin6_addr, addrs[1].sin6_port,
1343 &addrs[0].sin6_addr, addrs[0].sin6_port,
1346 if (inp == NULL || inp->inp_socket == NULL) {
1350 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1356 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1358 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1361 struct netmsg_tcp_notify {
1362 struct netmsg nm_nmsg;
1363 void (*nm_notify)(struct inpcb *, int);
1364 struct in_addr nm_faddr;
1369 tcp_notifyall_oncpu(struct netmsg *netmsg)
1371 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1374 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1375 nmsg->nm_arg, nmsg->nm_notify);
1377 nextcpu = mycpuid + 1;
1378 if (nextcpu < ncpus2)
1379 lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1381 lwkt_replymsg(&netmsg->nm_lmsg, 0);
1385 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1387 struct ip *ip = vip;
1389 struct in_addr faddr;
1392 void (*notify)(struct inpcb *, int) = tcp_notify;
1396 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1400 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1401 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1404 arg = inetctlerrmap[cmd];
1405 if (cmd == PRC_QUENCH) {
1406 notify = tcp_quench;
1407 } else if (icmp_may_rst &&
1408 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1409 cmd == PRC_UNREACH_PORT ||
1410 cmd == PRC_TIMXCEED_INTRANS) &&
1412 notify = tcp_drop_syn_sent;
1413 } else if (cmd == PRC_MSGSIZE) {
1414 struct icmp *icmp = (struct icmp *)
1415 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1417 arg = ntohs(icmp->icmp_nextmtu);
1418 notify = tcp_mtudisc;
1419 } else if (PRC_IS_REDIRECT(cmd)) {
1421 notify = in_rtchange;
1422 } else if (cmd == PRC_HOSTDEAD) {
1428 th = (struct tcphdr *)((caddr_t)ip +
1429 (IP_VHL_HL(ip->ip_vhl) << 2));
1430 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1431 ip->ip_src.s_addr, th->th_sport);
1432 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1433 ip->ip_src, th->th_sport, 0, NULL);
1434 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1435 icmpseq = htonl(th->th_seq);
1436 tp = intotcpcb(inp);
1437 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1438 SEQ_LT(icmpseq, tp->snd_max))
1439 (*notify)(inp, arg);
1441 struct in_conninfo inc;
1443 inc.inc_fport = th->th_dport;
1444 inc.inc_lport = th->th_sport;
1445 inc.inc_faddr = faddr;
1446 inc.inc_laddr = ip->ip_src;
1450 syncache_unreach(&inc, th);
1454 struct netmsg_tcp_notify nmsg;
1456 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1457 netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0,
1458 tcp_notifyall_oncpu);
1459 nmsg.nm_faddr = faddr;
1461 nmsg.nm_notify = notify;
1463 lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1469 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1472 void (*notify) (struct inpcb *, int) = tcp_notify;
1473 struct ip6_hdr *ip6;
1475 struct ip6ctlparam *ip6cp = NULL;
1476 const struct sockaddr_in6 *sa6_src = NULL;
1478 struct tcp_portonly {
1484 if (sa->sa_family != AF_INET6 ||
1485 sa->sa_len != sizeof(struct sockaddr_in6))
1489 if (cmd == PRC_QUENCH)
1490 notify = tcp_quench;
1491 else if (cmd == PRC_MSGSIZE) {
1492 struct ip6ctlparam *ip6cp = d;
1493 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1495 arg = ntohl(icmp6->icmp6_mtu);
1496 notify = tcp_mtudisc;
1497 } else if (!PRC_IS_REDIRECT(cmd) &&
1498 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1502 /* if the parameter is from icmp6, decode it. */
1504 ip6cp = (struct ip6ctlparam *)d;
1506 ip6 = ip6cp->ip6c_ip6;
1507 off = ip6cp->ip6c_off;
1508 sa6_src = ip6cp->ip6c_src;
1512 off = 0; /* fool gcc */
1517 struct in_conninfo inc;
1519 * XXX: We assume that when IPV6 is non NULL,
1520 * M and OFF are valid.
1523 /* check if we can safely examine src and dst ports */
1524 if (m->m_pkthdr.len < off + sizeof *thp)
1527 bzero(&th, sizeof th);
1528 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1530 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1531 (struct sockaddr *)ip6cp->ip6c_src,
1532 th.th_sport, cmd, arg, notify);
1534 inc.inc_fport = th.th_dport;
1535 inc.inc_lport = th.th_sport;
1536 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1537 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1539 syncache_unreach(&inc, &th);
1541 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1542 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1547 * Following is where TCP initial sequence number generation occurs.
1549 * There are two places where we must use initial sequence numbers:
1550 * 1. In SYN-ACK packets.
1551 * 2. In SYN packets.
1553 * All ISNs for SYN-ACK packets are generated by the syncache. See
1554 * tcp_syncache.c for details.
1556 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1557 * depends on this property. In addition, these ISNs should be
1558 * unguessable so as to prevent connection hijacking. To satisfy
1559 * the requirements of this situation, the algorithm outlined in
1560 * RFC 1948 is used to generate sequence numbers.
1562 * Implementation details:
1564 * Time is based off the system timer, and is corrected so that it
1565 * increases by one megabyte per second. This allows for proper
1566 * recycling on high speed LANs while still leaving over an hour
1569 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1570 * between seeding of isn_secret. This is normally set to zero,
1571 * as reseeding should not be necessary.
1575 #define ISN_BYTES_PER_SECOND 1048576
1577 u_char isn_secret[32];
1578 int isn_last_reseed;
1582 tcp_new_isn(struct tcpcb *tp)
1584 u_int32_t md5_buffer[4];
1587 /* Seed if this is the first use, reseed if requested. */
1588 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1589 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1591 read_random_unlimited(&isn_secret, sizeof isn_secret);
1592 isn_last_reseed = ticks;
1595 /* Compute the md5 hash and return the ISN. */
1597 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1598 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1600 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1601 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1602 sizeof(struct in6_addr));
1603 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1604 sizeof(struct in6_addr));
1608 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1609 sizeof(struct in_addr));
1610 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1611 sizeof(struct in_addr));
1613 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1614 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1615 new_isn = (tcp_seq) md5_buffer[0];
1616 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1621 * When a source quench is received, close congestion window
1622 * to one segment. We will gradually open it again as we proceed.
1625 tcp_quench(struct inpcb *inp, int error)
1627 struct tcpcb *tp = intotcpcb(inp);
1630 tp->snd_cwnd = tp->t_maxseg;
1636 * When a specific ICMP unreachable message is received and the
1637 * connection state is SYN-SENT, drop the connection. This behavior
1638 * is controlled by the icmp_may_rst sysctl.
1641 tcp_drop_syn_sent(struct inpcb *inp, int error)
1643 struct tcpcb *tp = intotcpcb(inp);
1645 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1646 tcp_drop(tp, error);
1650 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1651 * based on the new value in the route. Also nudge TCP to send something,
1652 * since we know the packet we just sent was dropped.
1653 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1656 tcp_mtudisc(struct inpcb *inp, int mtu)
1658 struct tcpcb *tp = intotcpcb(inp);
1660 struct socket *so = inp->inp_socket;
1663 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1665 const boolean_t isipv6 = FALSE;
1672 * If no MTU is provided in the ICMP message, use the
1673 * next lower likely value, as specified in RFC 1191.
1678 oldmtu = tp->t_maxopd +
1680 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1681 sizeof(struct tcpiphdr));
1682 mtu = ip_next_mtu(oldmtu, 0);
1686 rt = tcp_rtlookup6(&inp->inp_inc);
1688 rt = tcp_rtlookup(&inp->inp_inc);
1690 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1692 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1693 mtu = rt->rt_rmx.rmx_mtu;
1697 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1698 sizeof(struct tcpiphdr));
1701 * XXX - The following conditional probably violates the TCP
1702 * spec. The problem is that, since we don't know the
1703 * other end's MSS, we are supposed to use a conservative
1704 * default. But, if we do that, then MTU discovery will
1705 * never actually take place, because the conservative
1706 * default is much less than the MTUs typically seen
1707 * on the Internet today. For the moment, we'll sweep
1708 * this under the carpet.
1710 * The conservative default might not actually be a problem
1711 * if the only case this occurs is when sending an initial
1712 * SYN with options and data to a host we've never talked
1713 * to before. Then, they will reply with an MSS value which
1714 * will get recorded and the new parameters should get
1715 * recomputed. For Further Study.
1717 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1718 maxopd = taop->tao_mssopt;
1722 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1723 sizeof(struct tcpiphdr));
1725 if (tp->t_maxopd <= maxopd)
1727 tp->t_maxopd = maxopd;
1730 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1731 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1732 mss -= TCPOLEN_TSTAMP_APPA;
1734 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1735 (TF_REQ_CC | TF_RCVD_CC))
1736 mss -= TCPOLEN_CC_APPA;
1738 /* round down to multiple of MCLBYTES */
1739 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1741 mss &= ~(MCLBYTES - 1);
1744 mss = (mss / MCLBYTES) * MCLBYTES;
1747 if (so->so_snd.ssb_hiwat < mss)
1748 mss = so->so_snd.ssb_hiwat;
1752 tp->snd_nxt = tp->snd_una;
1754 tcpstat.tcps_mturesent++;
1758 * Look-up the routing entry to the peer of this inpcb. If no route
1759 * is found and it cannot be allocated the return NULL. This routine
1760 * is called by TCP routines that access the rmx structure and by tcp_mss
1761 * to get the interface MTU.
1764 tcp_rtlookup(struct in_conninfo *inc)
1766 struct route *ro = &inc->inc_route;
1768 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1769 /* No route yet, so try to acquire one */
1770 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1772 * unused portions of the structure MUST be zero'd
1773 * out because rtalloc() treats it as opaque data
1775 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1776 ro->ro_dst.sa_family = AF_INET;
1777 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1778 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1788 tcp_rtlookup6(struct in_conninfo *inc)
1790 struct route_in6 *ro6 = &inc->inc6_route;
1792 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1793 /* No route yet, so try to acquire one */
1794 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1796 * unused portions of the structure MUST be zero'd
1797 * out because rtalloc() treats it as opaque data
1799 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1800 ro6->ro_dst.sin6_family = AF_INET6;
1801 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1802 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1803 rtalloc((struct route *)ro6);
1806 return (ro6->ro_rt);
1811 /* compute ESP/AH header size for TCP, including outer IP header. */
1813 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1821 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1823 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1828 if (inp->inp_vflag & INP_IPV6) {
1829 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1831 th = (struct tcphdr *)(ip6 + 1);
1832 m->m_pkthdr.len = m->m_len =
1833 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1834 tcp_fillheaders(tp, ip6, th);
1835 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1839 ip = mtod(m, struct ip *);
1840 th = (struct tcphdr *)(ip + 1);
1841 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1842 tcp_fillheaders(tp, ip, th);
1843 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1852 * Return a pointer to the cached information about the remote host.
1853 * The cached information is stored in the protocol specific part of
1854 * the route metrics.
1857 tcp_gettaocache(struct in_conninfo *inc)
1862 if (inc->inc_isipv6)
1863 rt = tcp_rtlookup6(inc);
1866 rt = tcp_rtlookup(inc);
1868 /* Make sure this is a host route and is up. */
1870 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1873 return (rmx_taop(rt->rt_rmx));
1877 * Clear all the TAO cache entries, called from tcp_init.
1880 * This routine is just an empty one, because we assume that the routing
1881 * routing tables are initialized at the same time when TCP, so there is
1882 * nothing in the cache left over.
1885 tcp_cleartaocache(void)
1890 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1892 * This code attempts to calculate the bandwidth-delay product as a
1893 * means of determining the optimal window size to maximize bandwidth,
1894 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1895 * routers. This code also does a fairly good job keeping RTTs in check
1896 * across slow links like modems. We implement an algorithm which is very
1897 * similar (but not meant to be) TCP/Vegas. The code operates on the
1898 * transmitter side of a TCP connection and so only effects the transmit
1899 * side of the connection.
1901 * BACKGROUND: TCP makes no provision for the management of buffer space
1902 * at the end points or at the intermediate routers and switches. A TCP
1903 * stream, whether using NewReno or not, will eventually buffer as
1904 * many packets as it is able and the only reason this typically works is
1905 * due to the fairly small default buffers made available for a connection
1906 * (typicaly 16K or 32K). As machines use larger windows and/or window
1907 * scaling it is now fairly easy for even a single TCP connection to blow-out
1908 * all available buffer space not only on the local interface, but on
1909 * intermediate routers and switches as well. NewReno makes a misguided
1910 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1911 * then backing off, then steadily increasing the window again until another
1912 * failure occurs, ad-infinitum. This results in terrible oscillation that
1913 * is only made worse as network loads increase and the idea of intentionally
1914 * blowing out network buffers is, frankly, a terrible way to manage network
1917 * It is far better to limit the transmit window prior to the failure
1918 * condition being achieved. There are two general ways to do this: First
1919 * you can 'scan' through different transmit window sizes and locate the
1920 * point where the RTT stops increasing, indicating that you have filled the
1921 * pipe, then scan backwards until you note that RTT stops decreasing, then
1922 * repeat ad-infinitum. This method works in principle but has severe
1923 * implementation issues due to RTT variances, timer granularity, and
1924 * instability in the algorithm which can lead to many false positives and
1925 * create oscillations as well as interact badly with other TCP streams
1926 * implementing the same algorithm.
1928 * The second method is to limit the window to the bandwidth delay product
1929 * of the link. This is the method we implement. RTT variances and our
1930 * own manipulation of the congestion window, bwnd, can potentially
1931 * destabilize the algorithm. For this reason we have to stabilize the
1932 * elements used to calculate the window. We do this by using the minimum
1933 * observed RTT, the long term average of the observed bandwidth, and
1934 * by adding two segments worth of slop. It isn't perfect but it is able
1935 * to react to changing conditions and gives us a very stable basis on
1936 * which to extend the algorithm.
1939 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1947 * If inflight_enable is disabled in the middle of a tcp connection,
1948 * make sure snd_bwnd is effectively disabled.
1950 if (!tcp_inflight_enable) {
1951 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1952 tp->snd_bandwidth = 0;
1957 * Validate the delta time. If a connection is new or has been idle
1958 * a long time we have to reset the bandwidth calculator.
1961 delta_ticks = save_ticks - tp->t_bw_rtttime;
1962 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1963 tp->t_bw_rtttime = ticks;
1964 tp->t_bw_rtseq = ack_seq;
1965 if (tp->snd_bandwidth == 0)
1966 tp->snd_bandwidth = tcp_inflight_min;
1969 if (delta_ticks == 0)
1973 * Sanity check, plus ignore pure window update acks.
1975 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1979 * Figure out the bandwidth. Due to the tick granularity this
1980 * is a very rough number and it MUST be averaged over a fairly
1981 * long period of time. XXX we need to take into account a link
1982 * that is not using all available bandwidth, but for now our
1983 * slop will ramp us up if this case occurs and the bandwidth later
1986 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1987 tp->t_bw_rtttime = save_ticks;
1988 tp->t_bw_rtseq = ack_seq;
1989 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1991 tp->snd_bandwidth = bw;
1994 * Calculate the semi-static bandwidth delay product, plus two maximal
1995 * segments. The additional slop puts us squarely in the sweet
1996 * spot and also handles the bandwidth run-up case. Without the
1997 * slop we could be locking ourselves into a lower bandwidth.
1999 * Situations Handled:
2000 * (1) Prevents over-queueing of packets on LANs, especially on
2001 * high speed LANs, allowing larger TCP buffers to be
2002 * specified, and also does a good job preventing
2003 * over-queueing of packets over choke points like modems
2004 * (at least for the transmit side).
2006 * (2) Is able to handle changing network loads (bandwidth
2007 * drops so bwnd drops, bandwidth increases so bwnd
2010 * (3) Theoretically should stabilize in the face of multiple
2011 * connections implementing the same algorithm (this may need
2014 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2015 * be adjusted with a sysctl but typically only needs to be on
2016 * very slow connections. A value no smaller then 5 should
2017 * be used, but only reduce this default if you have no other
2021 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
2022 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
2023 tcp_inflight_stab * (int)tp->t_maxseg / 10;
2026 if (tcp_inflight_debug > 0) {
2028 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2030 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2031 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
2034 if ((long)bwnd < tcp_inflight_min)
2035 bwnd = tcp_inflight_min;
2036 if (bwnd > tcp_inflight_max)
2037 bwnd = tcp_inflight_max;
2038 if ((long)bwnd < tp->t_maxseg * 2)
2039 bwnd = tp->t_maxseg * 2;
2040 tp->snd_bwnd = bwnd;