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 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
183 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
184 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
187 int tcp_do_rfc1323 = 1;
188 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
189 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
191 int tcp_do_rfc1644 = 0;
192 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1644, rfc1644, CTLFLAG_RW,
193 &tcp_do_rfc1644, 0, "Enable rfc1644 (TTCP) extensions");
195 static int tcp_tcbhashsize = 0;
196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
197 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
199 static int do_tcpdrain = 1;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
201 "Enable tcp_drain routine for extra help when low on mbufs");
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_RD,
205 &tcbinfo[0].ipi_count, 0, "Number of active PCBs");
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. Note that the default lower bound of
217 * 1024 exists only for debugging. A good production default would be
218 * something like 6100.
220 static int tcp_inflight_enable = 0;
221 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
222 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
224 static int tcp_inflight_debug = 0;
225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
226 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
228 static int tcp_inflight_min = 6144;
229 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
230 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
232 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
233 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
234 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
236 static int tcp_inflight_stab = 20;
237 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
238 &tcp_inflight_stab, 0, "Slop in maximal packets / 10 (20 = 2 packets)");
240 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
241 static struct malloc_pipe tcptemp_mpipe;
243 static void tcp_willblock(int);
244 static void tcp_cleartaocache (void);
245 static void tcp_notify (struct inpcb *, int);
247 struct tcp_stats tcpstats_percpu[MAXCPU];
250 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
254 for (cpu = 0; cpu < ncpus; ++cpu) {
255 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
256 sizeof(struct tcp_stats))))
258 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
259 sizeof(struct tcp_stats))))
265 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
266 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
268 SYSCTL_STRUCT(_net_inet_tcp, TCPCTL_STATS, stats, CTLFLAG_RW,
269 &tcpstat, tcp_stats, "TCP statistics");
273 * Target size of TCP PCB hash tables. Must be a power of two.
275 * Note that this can be overridden by the kernel environment
276 * variable net.inet.tcp.tcbhashsize
279 #define TCBHASHSIZE 512
283 * This is the actual shape of what we allocate using the zone
284 * allocator. Doing it this way allows us to protect both structures
285 * using the same generation count, and also eliminates the overhead
286 * of allocating tcpcbs separately. By hiding the structure here,
287 * we avoid changing most of the rest of the code (although it needs
288 * to be changed, eventually, for greater efficiency).
291 #define ALIGNM1 (ALIGNMENT - 1)
295 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
298 struct tcp_callout inp_tp_rexmt;
299 struct tcp_callout inp_tp_persist;
300 struct tcp_callout inp_tp_keep;
301 struct tcp_callout inp_tp_2msl;
302 struct tcp_callout inp_tp_delack;
303 struct netmsg_tcp_timer inp_tp_timermsg;
314 struct inpcbporthead *porthashbase;
316 struct vm_zone *ipi_zone;
317 int hashsize = TCBHASHSIZE;
321 * note: tcptemp is used for keepalives, and it is ok for an
322 * allocation to fail so do not specify MPF_INT.
324 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
330 tcp_delacktime = TCPTV_DELACK;
331 tcp_keepinit = TCPTV_KEEP_INIT;
332 tcp_keepidle = TCPTV_KEEP_IDLE;
333 tcp_keepintvl = TCPTV_KEEPINTVL;
334 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
336 tcp_rexmit_min = TCPTV_MIN;
337 tcp_rexmit_slop = TCPTV_CPU_VAR;
339 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
340 if (!powerof2(hashsize)) {
341 kprintf("WARNING: TCB hash size not a power of 2\n");
342 hashsize = 512; /* safe default */
344 tcp_tcbhashsize = hashsize;
345 porthashbase = hashinit(hashsize, M_PCB, &porthashmask);
346 ipi_zone = zinit("tcpcb", sizeof(struct inp_tp), maxsockets,
349 for (cpu = 0; cpu < ncpus2; cpu++) {
350 in_pcbinfo_init(&tcbinfo[cpu]);
351 tcbinfo[cpu].cpu = cpu;
352 tcbinfo[cpu].hashbase = hashinit(hashsize, M_PCB,
353 &tcbinfo[cpu].hashmask);
354 tcbinfo[cpu].porthashbase = porthashbase;
355 tcbinfo[cpu].porthashmask = porthashmask;
356 tcbinfo[cpu].wildcardhashbase = hashinit(hashsize, M_PCB,
357 &tcbinfo[cpu].wildcardhashmask);
358 tcbinfo[cpu].ipi_zone = ipi_zone;
359 TAILQ_INIT(&tcpcbackq[cpu]);
362 tcp_reass_maxseg = nmbclusters / 16;
363 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
366 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
368 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
370 if (max_protohdr < TCP_MINPROTOHDR)
371 max_protohdr = TCP_MINPROTOHDR;
372 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
374 #undef TCP_MINPROTOHDR
377 * Initialize TCP statistics counters for each CPU.
380 for (cpu = 0; cpu < ncpus; ++cpu) {
381 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
384 bzero(&tcpstat, sizeof(struct tcp_stats));
392 tcpmsg_service_loop(void *dummy)
398 * Thread was started with TDF_MPSAFE
402 while ((msg = lwkt_waitport(&curthread->td_msgport, 0))) {
405 mplocked = netmsg_service(msg, tcp_mpsafe_thread,
407 } while ((msg = lwkt_getport(&curthread->td_msgport)) != NULL);
410 tcp_willblock(mplocked);
416 tcp_willblock(int mplocked)
419 int cpu = mycpu->gd_cpuid;
422 if (!mplocked && !tcp_mpsafe_proto) {
423 if (TAILQ_EMPTY(&tcpcbackq[cpu]))
431 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu])) != NULL) {
432 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
433 tp->t_flags &= ~TF_ONOUTPUTQ;
434 TAILQ_REMOVE(&tcpcbackq[cpu], tp, t_outputq);
444 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
445 * tcp_template used to store this data in mbufs, but we now recopy it out
446 * of the tcpcb each time to conserve mbufs.
449 tcp_fillheaders(struct tcpcb *tp, void *ip_ptr, void *tcp_ptr)
451 struct inpcb *inp = tp->t_inpcb;
452 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
455 if (inp->inp_vflag & INP_IPV6) {
458 ip6 = (struct ip6_hdr *)ip_ptr;
459 ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) |
460 (inp->in6p_flowinfo & IPV6_FLOWINFO_MASK);
461 ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) |
462 (IPV6_VERSION & IPV6_VERSION_MASK);
463 ip6->ip6_nxt = IPPROTO_TCP;
464 ip6->ip6_plen = sizeof(struct tcphdr);
465 ip6->ip6_src = inp->in6p_laddr;
466 ip6->ip6_dst = inp->in6p_faddr;
471 struct ip *ip = (struct ip *) ip_ptr;
473 ip->ip_vhl = IP_VHL_BORING;
480 ip->ip_p = IPPROTO_TCP;
481 ip->ip_src = inp->inp_laddr;
482 ip->ip_dst = inp->inp_faddr;
483 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
485 htons(sizeof(struct tcphdr) + IPPROTO_TCP));
488 tcp_hdr->th_sport = inp->inp_lport;
489 tcp_hdr->th_dport = inp->inp_fport;
494 tcp_hdr->th_flags = 0;
500 * Create template to be used to send tcp packets on a connection.
501 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
502 * use for this function is in keepalives, which use tcp_respond.
505 tcp_maketemplate(struct tcpcb *tp)
509 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
511 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t);
516 tcp_freetemplate(struct tcptemp *tmp)
518 mpipe_free(&tcptemp_mpipe, tmp);
522 * Send a single message to the TCP at address specified by
523 * the given TCP/IP header. If m == NULL, then we make a copy
524 * of the tcpiphdr at ti and send directly to the addressed host.
525 * This is used to force keep alive messages out using the TCP
526 * template for a connection. If flags are given then we send
527 * a message back to the TCP which originated the * segment ti,
528 * and discard the mbuf containing it and any other attached mbufs.
530 * In any case the ack and sequence number of the transmitted
531 * segment are as specified by the parameters.
533 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
536 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
537 tcp_seq ack, tcp_seq seq, int flags)
541 struct route *ro = NULL;
543 struct ip *ip = ipgen;
546 struct route_in6 *ro6 = NULL;
547 struct route_in6 sro6;
548 struct ip6_hdr *ip6 = ipgen;
549 boolean_t use_tmpro = TRUE;
551 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
553 const boolean_t isipv6 = FALSE;
557 if (!(flags & TH_RST)) {
558 win = ssb_space(&tp->t_inpcb->inp_socket->so_rcv);
559 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
560 win = (long)TCP_MAXWIN << tp->rcv_scale;
563 * Don't use the route cache of a listen socket,
564 * it is not MPSAFE; use temporary route cache.
566 if (tp->t_state != TCPS_LISTEN) {
568 ro6 = &tp->t_inpcb->in6p_route;
570 ro = &tp->t_inpcb->inp_route;
577 bzero(ro6, sizeof *ro6);
580 bzero(ro, sizeof *ro);
584 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
588 m->m_data += max_linkhdr;
590 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
591 ip6 = mtod(m, struct ip6_hdr *);
592 nth = (struct tcphdr *)(ip6 + 1);
594 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
595 ip = mtod(m, struct ip *);
596 nth = (struct tcphdr *)(ip + 1);
598 bcopy(th, nth, sizeof(struct tcphdr));
603 m->m_data = (caddr_t)ipgen;
604 /* m_len is set later */
606 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
608 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
609 nth = (struct tcphdr *)(ip6 + 1);
611 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
612 nth = (struct tcphdr *)(ip + 1);
616 * this is usually a case when an extension header
617 * exists between the IPv6 header and the
620 nth->th_sport = th->th_sport;
621 nth->th_dport = th->th_dport;
623 xchg(nth->th_dport, nth->th_sport, n_short);
628 ip6->ip6_vfc = IPV6_VERSION;
629 ip6->ip6_nxt = IPPROTO_TCP;
630 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
631 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
633 tlen += sizeof(struct tcpiphdr);
635 ip->ip_ttl = ip_defttl;
638 m->m_pkthdr.len = tlen;
639 m->m_pkthdr.rcvif = NULL;
640 nth->th_seq = htonl(seq);
641 nth->th_ack = htonl(ack);
643 nth->th_off = sizeof(struct tcphdr) >> 2;
644 nth->th_flags = flags;
646 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
648 nth->th_win = htons((u_short)win);
652 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
653 sizeof(struct ip6_hdr),
654 tlen - sizeof(struct ip6_hdr));
655 ip6->ip6_hlim = in6_selecthlim(tp ? tp->t_inpcb : NULL,
656 (ro6 && ro6->ro_rt) ?
657 ro6->ro_rt->rt_ifp : NULL);
659 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
660 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
661 m->m_pkthdr.csum_flags = CSUM_TCP;
662 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
665 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
666 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
669 ip6_output(m, NULL, ro6, ipflags, NULL, NULL,
670 tp ? tp->t_inpcb : NULL);
671 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
676 ipflags |= IP_DEBUGROUTE;
677 ip_output(m, NULL, ro, ipflags, NULL, tp ? tp->t_inpcb : NULL);
678 if ((ro == &sro) && (ro->ro_rt != NULL)) {
686 * Create a new TCP control block, making an
687 * empty reassembly queue and hooking it to the argument
688 * protocol control block. The `inp' parameter must have
689 * come from the zone allocator set up in tcp_init().
692 tcp_newtcpcb(struct inpcb *inp)
697 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
699 const boolean_t isipv6 = FALSE;
702 it = (struct inp_tp *)inp;
704 bzero(tp, sizeof(struct tcpcb));
705 LIST_INIT(&tp->t_segq);
706 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
708 /* Set up our timeouts. */
709 tp->tt_rexmt = &it->inp_tp_rexmt;
710 tp->tt_persist = &it->inp_tp_persist;
711 tp->tt_keep = &it->inp_tp_keep;
712 tp->tt_2msl = &it->inp_tp_2msl;
713 tp->tt_delack = &it->inp_tp_delack;
717 * Zero out timer message. We don't create it here,
718 * since the current CPU may not be the owner of this
721 tp->tt_msg = &it->inp_tp_timermsg;
722 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
725 tp->t_flags = (TF_REQ_SCALE | TF_REQ_TSTMP);
727 tp->t_flags |= TF_REQ_CC;
728 tp->t_inpcb = inp; /* XXX */
729 tp->t_state = TCPS_CLOSED;
731 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
732 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
733 * reasonable initial retransmit time.
735 tp->t_srtt = TCPTV_SRTTBASE;
737 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
738 tp->t_rttmin = tcp_rexmit_min;
739 tp->t_rxtcur = TCPTV_RTOBASE;
740 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
741 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
742 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
743 tp->t_rcvtime = ticks;
745 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
746 * because the socket may be bound to an IPv6 wildcard address,
747 * which may match an IPv4-mapped IPv6 address.
749 inp->inp_ip_ttl = ip_defttl;
751 tcp_sack_tcpcb_init(tp);
752 return (tp); /* XXX */
756 * Drop a TCP connection, reporting the specified error.
757 * If connection is synchronized, then send a RST to peer.
760 tcp_drop(struct tcpcb *tp, int error)
762 struct socket *so = tp->t_inpcb->inp_socket;
764 if (TCPS_HAVERCVDSYN(tp->t_state)) {
765 tp->t_state = TCPS_CLOSED;
767 tcpstat.tcps_drops++;
769 tcpstat.tcps_conndrops++;
770 if (error == ETIMEDOUT && tp->t_softerror)
771 error = tp->t_softerror;
772 so->so_error = error;
773 return (tcp_close(tp));
778 struct netmsg_remwildcard {
779 struct netmsg nm_netmsg;
780 struct inpcb *nm_inp;
781 struct inpcbinfo *nm_pcbinfo;
790 * Wildcard inpcb's on SMP boxes must be removed from all cpus before the
791 * inp can be detached. We do this by cycling through the cpus, ending up
792 * on the cpu controlling the inp last and then doing the disconnect.
795 in_pcbremwildcardhash_handler(struct netmsg *msg0)
797 struct netmsg_remwildcard *msg = (struct netmsg_remwildcard *)msg0;
800 cpu = msg->nm_pcbinfo->cpu;
802 if (cpu == msg->nm_inp->inp_pcbinfo->cpu) {
803 /* note: detach removes any wildcard hash entry */
806 in6_pcbdetach(msg->nm_inp);
809 in_pcbdetach(msg->nm_inp);
810 lwkt_replymsg(&msg->nm_netmsg.nm_lmsg, 0);
812 in_pcbremwildcardhash_oncpu(msg->nm_inp, msg->nm_pcbinfo);
813 cpu = (cpu + 1) % ncpus2;
814 msg->nm_pcbinfo = &tcbinfo[cpu];
815 lwkt_forwardmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
822 * Close a TCP control block:
823 * discard all space held by the tcp
824 * discard internet protocol block
825 * wake up any sleepers
828 tcp_close(struct tcpcb *tp)
831 struct inpcb *inp = tp->t_inpcb;
832 struct socket *so = inp->inp_socket;
834 boolean_t dosavessthresh;
839 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) != 0);
840 boolean_t isafinet6 = (INP_CHECK_SOCKAF(so, AF_INET6) != 0);
842 const boolean_t isipv6 = FALSE;
846 * The tp is not instantly destroyed in the wildcard case. Setting
847 * the state to TCPS_TERMINATING will prevent the TCP stack from
848 * messing with it, though it should be noted that this change may
849 * not take effect on other cpus until we have chained the wildcard
852 * XXX we currently depend on the BGL to synchronize the tp->t_state
853 * update and prevent other tcp protocol threads from accepting new
854 * connections on the listen socket we might be trying to close down.
856 KKASSERT(tp->t_state != TCPS_TERMINATING);
857 tp->t_state = TCPS_TERMINATING;
860 * Make sure that all of our timers are stopped before we
861 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
862 * timers are never used. If timer message is never created
863 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
865 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
866 tcp_callout_stop(tp, tp->tt_rexmt);
867 tcp_callout_stop(tp, tp->tt_persist);
868 tcp_callout_stop(tp, tp->tt_keep);
869 tcp_callout_stop(tp, tp->tt_2msl);
870 tcp_callout_stop(tp, tp->tt_delack);
873 if (tp->t_flags & TF_ONOUTPUTQ) {
874 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
875 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu], tp, t_outputq);
876 tp->t_flags &= ~TF_ONOUTPUTQ;
880 * If we got enough samples through the srtt filter,
881 * save the rtt and rttvar in the routing entry.
882 * 'Enough' is arbitrarily defined as the 16 samples.
883 * 16 samples is enough for the srtt filter to converge
884 * to within 5% of the correct value; fewer samples and
885 * we could save a very bogus rtt.
887 * Don't update the default route's characteristics and don't
888 * update anything that the user "locked".
890 if (tp->t_rttupdated >= 16) {
894 struct sockaddr_in6 *sin6;
896 if ((rt = inp->in6p_route.ro_rt) == NULL)
898 sin6 = (struct sockaddr_in6 *)rt_key(rt);
899 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
902 if ((rt = inp->inp_route.ro_rt) == NULL ||
903 ((struct sockaddr_in *)rt_key(rt))->
904 sin_addr.s_addr == INADDR_ANY)
907 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
908 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
909 if (rt->rt_rmx.rmx_rtt && i)
911 * filter this update to half the old & half
912 * the new values, converting scale.
913 * See route.h and tcp_var.h for a
914 * description of the scaling constants.
917 (rt->rt_rmx.rmx_rtt + i) / 2;
919 rt->rt_rmx.rmx_rtt = i;
920 tcpstat.tcps_cachedrtt++;
922 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
924 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
925 if (rt->rt_rmx.rmx_rttvar && i)
926 rt->rt_rmx.rmx_rttvar =
927 (rt->rt_rmx.rmx_rttvar + i) / 2;
929 rt->rt_rmx.rmx_rttvar = i;
930 tcpstat.tcps_cachedrttvar++;
933 * The old comment here said:
934 * update the pipelimit (ssthresh) if it has been updated
935 * already or if a pipesize was specified & the threshhold
936 * got below half the pipesize. I.e., wait for bad news
937 * before we start updating, then update on both good
940 * But we want to save the ssthresh even if no pipesize is
941 * specified explicitly in the route, because such
942 * connections still have an implicit pipesize specified
943 * by the global tcp_sendspace. In the absence of a reliable
944 * way to calculate the pipesize, it will have to do.
946 i = tp->snd_ssthresh;
947 if (rt->rt_rmx.rmx_sendpipe != 0)
948 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
950 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
951 if (dosavessthresh ||
952 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
953 (rt->rt_rmx.rmx_ssthresh != 0))) {
955 * convert the limit from user data bytes to
956 * packets then to packet data bytes.
958 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
963 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
964 sizeof(struct tcpiphdr));
965 if (rt->rt_rmx.rmx_ssthresh)
966 rt->rt_rmx.rmx_ssthresh =
967 (rt->rt_rmx.rmx_ssthresh + i) / 2;
969 rt->rt_rmx.rmx_ssthresh = i;
970 tcpstat.tcps_cachedssthresh++;
975 /* free the reassembly queue, if any */
976 while((q = LIST_FIRST(&tp->t_segq)) != NULL) {
977 LIST_REMOVE(q, tqe_q);
982 /* throw away SACK blocks in scoreboard*/
984 tcp_sack_cleanup(&tp->scb);
986 inp->inp_ppcb = NULL;
987 soisdisconnected(so);
989 tcp_destroy_timermsg(tp);
992 * Discard the inp. In the SMP case a wildcard inp's hash (created
993 * by a listen socket or an INADDR_ANY udp socket) is replicated
994 * for each protocol thread and must be removed in the context of
995 * that thread. This is accomplished by chaining the message
998 * If the inp is not wildcarded we simply detach, which will remove
999 * the any hashes still present for this inp.
1002 if (inp->inp_flags & INP_WILDCARD_MP) {
1003 struct netmsg_remwildcard *msg;
1005 cpu = (inp->inp_pcbinfo->cpu + 1) % ncpus2;
1006 msg = kmalloc(sizeof(struct netmsg_remwildcard),
1007 M_LWKTMSG, M_INTWAIT);
1008 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1009 in_pcbremwildcardhash_handler);
1011 msg->nm_isinet6 = isafinet6;
1014 msg->nm_pcbinfo = &tcbinfo[cpu];
1015 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1019 /* note: detach removes any wildcard hash entry */
1027 tcpstat.tcps_closed++;
1031 static __inline void
1032 tcp_drain_oncpu(struct inpcbhead *head)
1036 struct tseg_qent *te;
1038 LIST_FOREACH(inpb, head, inp_list) {
1039 if (inpb->inp_flags & INP_PLACEMARKER)
1041 if ((tcpb = intotcpcb(inpb))) {
1042 while ((te = LIST_FIRST(&tcpb->t_segq)) != NULL) {
1043 LIST_REMOVE(te, tqe_q);
1053 struct netmsg_tcp_drain {
1054 struct netmsg nm_netmsg;
1055 struct inpcbhead *nm_head;
1059 tcp_drain_handler(netmsg_t netmsg)
1061 struct netmsg_tcp_drain *nm = (void *)netmsg;
1063 tcp_drain_oncpu(nm->nm_head);
1064 lwkt_replymsg(&nm->nm_netmsg.nm_lmsg, 0);
1079 * Walk the tcpbs, if existing, and flush the reassembly queue,
1080 * if there is one...
1081 * XXX: The "Net/3" implementation doesn't imply that the TCP
1082 * reassembly queue should be flushed, but in a situation
1083 * where we're really low on mbufs, this is potentially
1087 for (cpu = 0; cpu < ncpus2; cpu++) {
1088 struct netmsg_tcp_drain *msg;
1090 if (cpu == mycpu->gd_cpuid) {
1091 tcp_drain_oncpu(&tcbinfo[cpu].pcblisthead);
1093 msg = kmalloc(sizeof(struct netmsg_tcp_drain),
1094 M_LWKTMSG, M_NOWAIT);
1097 netmsg_init(&msg->nm_netmsg, &netisr_afree_rport, 0,
1099 msg->nm_head = &tcbinfo[cpu].pcblisthead;
1100 lwkt_sendmsg(tcp_cport(cpu), &msg->nm_netmsg.nm_lmsg);
1104 tcp_drain_oncpu(&tcbinfo[0].pcblisthead);
1109 * Notify a tcp user of an asynchronous error;
1110 * store error as soft error, but wake up user
1111 * (for now, won't do anything until can select for soft error).
1113 * Do not wake up user since there currently is no mechanism for
1114 * reporting soft errors (yet - a kqueue filter may be added).
1117 tcp_notify(struct inpcb *inp, int error)
1119 struct tcpcb *tp = intotcpcb(inp);
1122 * Ignore some errors if we are hooked up.
1123 * If connection hasn't completed, has retransmitted several times,
1124 * and receives a second error, give up now. This is better
1125 * than waiting a long time to establish a connection that
1126 * can never complete.
1128 if (tp->t_state == TCPS_ESTABLISHED &&
1129 (error == EHOSTUNREACH || error == ENETUNREACH ||
1130 error == EHOSTDOWN)) {
1132 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1134 tcp_drop(tp, error);
1136 tp->t_softerror = error;
1138 wakeup(&so->so_timeo);
1145 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1148 struct inpcb *marker;
1158 * The process of preparing the TCB list is too time-consuming and
1159 * resource-intensive to repeat twice on every request.
1161 if (req->oldptr == NULL) {
1162 for (ccpu = 0; ccpu < ncpus; ++ccpu) {
1163 gd = globaldata_find(ccpu);
1164 n += tcbinfo[gd->gd_cpuid].ipi_count;
1166 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1170 if (req->newptr != NULL)
1173 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1174 marker->inp_flags |= INP_PLACEMARKER;
1177 * OK, now we're committed to doing something. Run the inpcb list
1178 * for each cpu in the system and construct the output. Use a
1179 * list placemarker to deal with list changes occuring during
1180 * copyout blockages (but otherwise depend on being on the correct
1181 * cpu to avoid races).
1183 origcpu = mycpu->gd_cpuid;
1184 for (ccpu = 1; ccpu <= ncpus && error == 0; ++ccpu) {
1190 cpu_id = (origcpu + ccpu) % ncpus;
1191 if ((smp_active_mask & (1 << cpu_id)) == 0)
1193 rgd = globaldata_find(cpu_id);
1194 lwkt_setcpu_self(rgd);
1196 gencnt = tcbinfo[cpu_id].ipi_gencnt;
1197 n = tcbinfo[cpu_id].ipi_count;
1199 LIST_INSERT_HEAD(&tcbinfo[cpu_id].pcblisthead, marker, inp_list);
1201 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1203 * process a snapshot of pcbs, ignoring placemarkers
1204 * and using our own to allow SYSCTL_OUT to block.
1206 LIST_REMOVE(marker, inp_list);
1207 LIST_INSERT_AFTER(inp, marker, inp_list);
1209 if (inp->inp_flags & INP_PLACEMARKER)
1211 if (inp->inp_gencnt > gencnt)
1213 if (prison_xinpcb(req->td, inp))
1216 xt.xt_len = sizeof xt;
1217 bcopy(inp, &xt.xt_inp, sizeof *inp);
1218 inp_ppcb = inp->inp_ppcb;
1219 if (inp_ppcb != NULL)
1220 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1222 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1223 if (inp->inp_socket)
1224 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1225 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1229 LIST_REMOVE(marker, inp_list);
1230 if (error == 0 && i < n) {
1231 bzero(&xt, sizeof xt);
1232 xt.xt_len = sizeof xt;
1234 error = SYSCTL_OUT(req, &xt, sizeof xt);
1243 * Make sure we are on the same cpu we were on originally, since
1244 * higher level callers expect this. Also don't pollute caches with
1245 * migrated userland data by (eventually) returning to userland
1246 * on a different cpu.
1248 lwkt_setcpu_self(globaldata_find(origcpu));
1249 kfree(marker, M_TEMP);
1253 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1254 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1257 tcp_getcred(SYSCTL_HANDLER_ARGS)
1259 struct sockaddr_in addrs[2];
1264 error = priv_check(req->td, PRIV_ROOT);
1267 error = SYSCTL_IN(req, addrs, sizeof addrs);
1271 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1272 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1273 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1274 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1275 if (inp == NULL || inp->inp_socket == NULL) {
1279 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1285 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1286 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1290 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1292 struct sockaddr_in6 addrs[2];
1295 boolean_t mapped = FALSE;
1297 error = priv_check(req->td, PRIV_ROOT);
1300 error = SYSCTL_IN(req, addrs, sizeof addrs);
1303 if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) {
1304 if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr))
1311 inp = in_pcblookup_hash(&tcbinfo[0],
1312 *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12],
1314 *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12],
1318 inp = in6_pcblookup_hash(&tcbinfo[0],
1319 &addrs[1].sin6_addr, addrs[1].sin6_port,
1320 &addrs[0].sin6_addr, addrs[0].sin6_port,
1323 if (inp == NULL || inp->inp_socket == NULL) {
1327 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1333 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1335 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1338 struct netmsg_tcp_notify {
1339 struct netmsg nm_nmsg;
1340 void (*nm_notify)(struct inpcb *, int);
1341 struct in_addr nm_faddr;
1346 tcp_notifyall_oncpu(struct netmsg *netmsg)
1348 struct netmsg_tcp_notify *nmsg = (struct netmsg_tcp_notify *)netmsg;
1351 in_pcbnotifyall(&tcbinfo[mycpuid].pcblisthead, nmsg->nm_faddr,
1352 nmsg->nm_arg, nmsg->nm_notify);
1354 nextcpu = mycpuid + 1;
1355 if (nextcpu < ncpus2)
1356 lwkt_forwardmsg(tcp_cport(nextcpu), &netmsg->nm_lmsg);
1358 lwkt_replymsg(&netmsg->nm_lmsg, 0);
1362 tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip)
1364 struct ip *ip = vip;
1366 struct in_addr faddr;
1369 void (*notify)(struct inpcb *, int) = tcp_notify;
1373 if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1377 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1378 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1381 arg = inetctlerrmap[cmd];
1382 if (cmd == PRC_QUENCH) {
1383 notify = tcp_quench;
1384 } else if (icmp_may_rst &&
1385 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1386 cmd == PRC_UNREACH_PORT ||
1387 cmd == PRC_TIMXCEED_INTRANS) &&
1389 notify = tcp_drop_syn_sent;
1390 } else if (cmd == PRC_MSGSIZE) {
1391 struct icmp *icmp = (struct icmp *)
1392 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1394 arg = ntohs(icmp->icmp_nextmtu);
1395 notify = tcp_mtudisc;
1396 } else if (PRC_IS_REDIRECT(cmd)) {
1398 notify = in_rtchange;
1399 } else if (cmd == PRC_HOSTDEAD) {
1405 th = (struct tcphdr *)((caddr_t)ip +
1406 (IP_VHL_HL(ip->ip_vhl) << 2));
1407 cpu = tcp_addrcpu(faddr.s_addr, th->th_dport,
1408 ip->ip_src.s_addr, th->th_sport);
1409 inp = in_pcblookup_hash(&tcbinfo[cpu], faddr, th->th_dport,
1410 ip->ip_src, th->th_sport, 0, NULL);
1411 if ((inp != NULL) && (inp->inp_socket != NULL)) {
1412 icmpseq = htonl(th->th_seq);
1413 tp = intotcpcb(inp);
1414 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1415 SEQ_LT(icmpseq, tp->snd_max))
1416 (*notify)(inp, arg);
1418 struct in_conninfo inc;
1420 inc.inc_fport = th->th_dport;
1421 inc.inc_lport = th->th_sport;
1422 inc.inc_faddr = faddr;
1423 inc.inc_laddr = ip->ip_src;
1427 syncache_unreach(&inc, th);
1431 struct netmsg_tcp_notify nmsg;
1433 KKASSERT(&curthread->td_msgport == cpu_portfn(0));
1434 netmsg_init(&nmsg.nm_nmsg, &curthread->td_msgport, 0,
1435 tcp_notifyall_oncpu);
1436 nmsg.nm_faddr = faddr;
1438 nmsg.nm_notify = notify;
1440 lwkt_domsg(tcp_cport(0), &nmsg.nm_nmsg.nm_lmsg, 0);
1446 tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d)
1449 void (*notify) (struct inpcb *, int) = tcp_notify;
1450 struct ip6_hdr *ip6;
1452 struct ip6ctlparam *ip6cp = NULL;
1453 const struct sockaddr_in6 *sa6_src = NULL;
1455 struct tcp_portonly {
1461 if (sa->sa_family != AF_INET6 ||
1462 sa->sa_len != sizeof(struct sockaddr_in6))
1466 if (cmd == PRC_QUENCH)
1467 notify = tcp_quench;
1468 else if (cmd == PRC_MSGSIZE) {
1469 struct ip6ctlparam *ip6cp = d;
1470 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1472 arg = ntohl(icmp6->icmp6_mtu);
1473 notify = tcp_mtudisc;
1474 } else if (!PRC_IS_REDIRECT(cmd) &&
1475 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1479 /* if the parameter is from icmp6, decode it. */
1481 ip6cp = (struct ip6ctlparam *)d;
1483 ip6 = ip6cp->ip6c_ip6;
1484 off = ip6cp->ip6c_off;
1485 sa6_src = ip6cp->ip6c_src;
1489 off = 0; /* fool gcc */
1494 struct in_conninfo inc;
1496 * XXX: We assume that when IPV6 is non NULL,
1497 * M and OFF are valid.
1500 /* check if we can safely examine src and dst ports */
1501 if (m->m_pkthdr.len < off + sizeof *thp)
1504 bzero(&th, sizeof th);
1505 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1507 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, th.th_dport,
1508 (struct sockaddr *)ip6cp->ip6c_src,
1509 th.th_sport, cmd, arg, notify);
1511 inc.inc_fport = th.th_dport;
1512 inc.inc_lport = th.th_sport;
1513 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1514 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1516 syncache_unreach(&inc, &th);
1518 in6_pcbnotify(&tcbinfo[0].pcblisthead, sa, 0,
1519 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1524 * Following is where TCP initial sequence number generation occurs.
1526 * There are two places where we must use initial sequence numbers:
1527 * 1. In SYN-ACK packets.
1528 * 2. In SYN packets.
1530 * All ISNs for SYN-ACK packets are generated by the syncache. See
1531 * tcp_syncache.c for details.
1533 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1534 * depends on this property. In addition, these ISNs should be
1535 * unguessable so as to prevent connection hijacking. To satisfy
1536 * the requirements of this situation, the algorithm outlined in
1537 * RFC 1948 is used to generate sequence numbers.
1539 * Implementation details:
1541 * Time is based off the system timer, and is corrected so that it
1542 * increases by one megabyte per second. This allows for proper
1543 * recycling on high speed LANs while still leaving over an hour
1546 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1547 * between seeding of isn_secret. This is normally set to zero,
1548 * as reseeding should not be necessary.
1552 #define ISN_BYTES_PER_SECOND 1048576
1554 u_char isn_secret[32];
1555 int isn_last_reseed;
1559 tcp_new_isn(struct tcpcb *tp)
1561 u_int32_t md5_buffer[4];
1564 /* Seed if this is the first use, reseed if requested. */
1565 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1566 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1568 read_random_unlimited(&isn_secret, sizeof isn_secret);
1569 isn_last_reseed = ticks;
1572 /* Compute the md5 hash and return the ISN. */
1574 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1575 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1577 if (tp->t_inpcb->inp_vflag & INP_IPV6) {
1578 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1579 sizeof(struct in6_addr));
1580 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1581 sizeof(struct in6_addr));
1585 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1586 sizeof(struct in_addr));
1587 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1588 sizeof(struct in_addr));
1590 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1591 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1592 new_isn = (tcp_seq) md5_buffer[0];
1593 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1598 * When a source quench is received, close congestion window
1599 * to one segment. We will gradually open it again as we proceed.
1602 tcp_quench(struct inpcb *inp, int error)
1604 struct tcpcb *tp = intotcpcb(inp);
1607 tp->snd_cwnd = tp->t_maxseg;
1613 * When a specific ICMP unreachable message is received and the
1614 * connection state is SYN-SENT, drop the connection. This behavior
1615 * is controlled by the icmp_may_rst sysctl.
1618 tcp_drop_syn_sent(struct inpcb *inp, int error)
1620 struct tcpcb *tp = intotcpcb(inp);
1622 if ((tp != NULL) && (tp->t_state == TCPS_SYN_SENT))
1623 tcp_drop(tp, error);
1627 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1628 * based on the new value in the route. Also nudge TCP to send something,
1629 * since we know the packet we just sent was dropped.
1630 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1633 tcp_mtudisc(struct inpcb *inp, int mtu)
1635 struct tcpcb *tp = intotcpcb(inp);
1637 struct socket *so = inp->inp_socket;
1640 boolean_t isipv6 = ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0);
1642 const boolean_t isipv6 = FALSE;
1649 * If no MTU is provided in the ICMP message, use the
1650 * next lower likely value, as specified in RFC 1191.
1655 oldmtu = tp->t_maxopd +
1657 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1658 sizeof(struct tcpiphdr));
1659 mtu = ip_next_mtu(oldmtu, 0);
1663 rt = tcp_rtlookup6(&inp->inp_inc);
1665 rt = tcp_rtlookup(&inp->inp_inc);
1667 struct rmxp_tao *taop = rmx_taop(rt->rt_rmx);
1669 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1670 mtu = rt->rt_rmx.rmx_mtu;
1674 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1675 sizeof(struct tcpiphdr));
1678 * XXX - The following conditional probably violates the TCP
1679 * spec. The problem is that, since we don't know the
1680 * other end's MSS, we are supposed to use a conservative
1681 * default. But, if we do that, then MTU discovery will
1682 * never actually take place, because the conservative
1683 * default is much less than the MTUs typically seen
1684 * on the Internet today. For the moment, we'll sweep
1685 * this under the carpet.
1687 * The conservative default might not actually be a problem
1688 * if the only case this occurs is when sending an initial
1689 * SYN with options and data to a host we've never talked
1690 * to before. Then, they will reply with an MSS value which
1691 * will get recorded and the new parameters should get
1692 * recomputed. For Further Study.
1694 if (taop->tao_mssopt != 0 && taop->tao_mssopt < maxopd)
1695 maxopd = taop->tao_mssopt;
1699 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1700 sizeof(struct tcpiphdr));
1702 if (tp->t_maxopd <= maxopd)
1704 tp->t_maxopd = maxopd;
1707 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1708 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1709 mss -= TCPOLEN_TSTAMP_APPA;
1711 if ((tp->t_flags & (TF_REQ_CC | TF_RCVD_CC | TF_NOOPT)) ==
1712 (TF_REQ_CC | TF_RCVD_CC))
1713 mss -= TCPOLEN_CC_APPA;
1715 /* round down to multiple of MCLBYTES */
1716 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1718 mss &= ~(MCLBYTES - 1);
1721 mss = (mss / MCLBYTES) * MCLBYTES;
1724 if (so->so_snd.ssb_hiwat < mss)
1725 mss = so->so_snd.ssb_hiwat;
1729 tp->snd_nxt = tp->snd_una;
1731 tcpstat.tcps_mturesent++;
1735 * Look-up the routing entry to the peer of this inpcb. If no route
1736 * is found and it cannot be allocated the return NULL. This routine
1737 * is called by TCP routines that access the rmx structure and by tcp_mss
1738 * to get the interface MTU.
1741 tcp_rtlookup(struct in_conninfo *inc)
1743 struct route *ro = &inc->inc_route;
1745 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1746 /* No route yet, so try to acquire one */
1747 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1749 * unused portions of the structure MUST be zero'd
1750 * out because rtalloc() treats it as opaque data
1752 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1753 ro->ro_dst.sa_family = AF_INET;
1754 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1755 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1765 tcp_rtlookup6(struct in_conninfo *inc)
1767 struct route_in6 *ro6 = &inc->inc6_route;
1769 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1770 /* No route yet, so try to acquire one */
1771 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1773 * unused portions of the structure MUST be zero'd
1774 * out because rtalloc() treats it as opaque data
1776 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1777 ro6->ro_dst.sin6_family = AF_INET6;
1778 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1779 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1780 rtalloc((struct route *)ro6);
1783 return (ro6->ro_rt);
1788 /* compute ESP/AH header size for TCP, including outer IP header. */
1790 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1798 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1800 MGETHDR(m, MB_DONTWAIT, MT_DATA);
1805 if (inp->inp_vflag & INP_IPV6) {
1806 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1808 th = (struct tcphdr *)(ip6 + 1);
1809 m->m_pkthdr.len = m->m_len =
1810 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1811 tcp_fillheaders(tp, ip6, th);
1812 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1816 ip = mtod(m, struct ip *);
1817 th = (struct tcphdr *)(ip + 1);
1818 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1819 tcp_fillheaders(tp, ip, th);
1820 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1829 * Return a pointer to the cached information about the remote host.
1830 * The cached information is stored in the protocol specific part of
1831 * the route metrics.
1834 tcp_gettaocache(struct in_conninfo *inc)
1839 if (inc->inc_isipv6)
1840 rt = tcp_rtlookup6(inc);
1843 rt = tcp_rtlookup(inc);
1845 /* Make sure this is a host route and is up. */
1847 (rt->rt_flags & (RTF_UP | RTF_HOST)) != (RTF_UP | RTF_HOST))
1850 return (rmx_taop(rt->rt_rmx));
1854 * Clear all the TAO cache entries, called from tcp_init.
1857 * This routine is just an empty one, because we assume that the routing
1858 * routing tables are initialized at the same time when TCP, so there is
1859 * nothing in the cache left over.
1862 tcp_cleartaocache(void)
1867 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1869 * This code attempts to calculate the bandwidth-delay product as a
1870 * means of determining the optimal window size to maximize bandwidth,
1871 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1872 * routers. This code also does a fairly good job keeping RTTs in check
1873 * across slow links like modems. We implement an algorithm which is very
1874 * similar (but not meant to be) TCP/Vegas. The code operates on the
1875 * transmitter side of a TCP connection and so only effects the transmit
1876 * side of the connection.
1878 * BACKGROUND: TCP makes no provision for the management of buffer space
1879 * at the end points or at the intermediate routers and switches. A TCP
1880 * stream, whether using NewReno or not, will eventually buffer as
1881 * many packets as it is able and the only reason this typically works is
1882 * due to the fairly small default buffers made available for a connection
1883 * (typicaly 16K or 32K). As machines use larger windows and/or window
1884 * scaling it is now fairly easy for even a single TCP connection to blow-out
1885 * all available buffer space not only on the local interface, but on
1886 * intermediate routers and switches as well. NewReno makes a misguided
1887 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1888 * then backing off, then steadily increasing the window again until another
1889 * failure occurs, ad-infinitum. This results in terrible oscillation that
1890 * is only made worse as network loads increase and the idea of intentionally
1891 * blowing out network buffers is, frankly, a terrible way to manage network
1894 * It is far better to limit the transmit window prior to the failure
1895 * condition being achieved. There are two general ways to do this: First
1896 * you can 'scan' through different transmit window sizes and locate the
1897 * point where the RTT stops increasing, indicating that you have filled the
1898 * pipe, then scan backwards until you note that RTT stops decreasing, then
1899 * repeat ad-infinitum. This method works in principle but has severe
1900 * implementation issues due to RTT variances, timer granularity, and
1901 * instability in the algorithm which can lead to many false positives and
1902 * create oscillations as well as interact badly with other TCP streams
1903 * implementing the same algorithm.
1905 * The second method is to limit the window to the bandwidth delay product
1906 * of the link. This is the method we implement. RTT variances and our
1907 * own manipulation of the congestion window, bwnd, can potentially
1908 * destabilize the algorithm. For this reason we have to stabilize the
1909 * elements used to calculate the window. We do this by using the minimum
1910 * observed RTT, the long term average of the observed bandwidth, and
1911 * by adding two segments worth of slop. It isn't perfect but it is able
1912 * to react to changing conditions and gives us a very stable basis on
1913 * which to extend the algorithm.
1916 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1924 * If inflight_enable is disabled in the middle of a tcp connection,
1925 * make sure snd_bwnd is effectively disabled.
1927 if (!tcp_inflight_enable) {
1928 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
1929 tp->snd_bandwidth = 0;
1934 * Validate the delta time. If a connection is new or has been idle
1935 * a long time we have to reset the bandwidth calculator.
1938 delta_ticks = save_ticks - tp->t_bw_rtttime;
1939 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
1940 tp->t_bw_rtttime = ticks;
1941 tp->t_bw_rtseq = ack_seq;
1942 if (tp->snd_bandwidth == 0)
1943 tp->snd_bandwidth = tcp_inflight_min;
1946 if (delta_ticks == 0)
1950 * Sanity check, plus ignore pure window update acks.
1952 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
1956 * Figure out the bandwidth. Due to the tick granularity this
1957 * is a very rough number and it MUST be averaged over a fairly
1958 * long period of time. XXX we need to take into account a link
1959 * that is not using all available bandwidth, but for now our
1960 * slop will ramp us up if this case occurs and the bandwidth later
1963 bw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
1964 tp->t_bw_rtttime = save_ticks;
1965 tp->t_bw_rtseq = ack_seq;
1966 bw = ((int64_t)tp->snd_bandwidth * 15 + bw) >> 4;
1968 tp->snd_bandwidth = bw;
1971 * Calculate the semi-static bandwidth delay product, plus two maximal
1972 * segments. The additional slop puts us squarely in the sweet
1973 * spot and also handles the bandwidth run-up case. Without the
1974 * slop we could be locking ourselves into a lower bandwidth.
1976 * Situations Handled:
1977 * (1) Prevents over-queueing of packets on LANs, especially on
1978 * high speed LANs, allowing larger TCP buffers to be
1979 * specified, and also does a good job preventing
1980 * over-queueing of packets over choke points like modems
1981 * (at least for the transmit side).
1983 * (2) Is able to handle changing network loads (bandwidth
1984 * drops so bwnd drops, bandwidth increases so bwnd
1987 * (3) Theoretically should stabilize in the face of multiple
1988 * connections implementing the same algorithm (this may need
1991 * (4) Stability value (defaults to 20 = 2 maximal packets) can
1992 * be adjusted with a sysctl but typically only needs to be on
1993 * very slow connections. A value no smaller then 5 should
1994 * be used, but only reduce this default if you have no other
1998 #define USERTT ((tp->t_srtt + tp->t_rttbest) / 2)
1999 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
2000 tcp_inflight_stab * (int)tp->t_maxseg / 10;
2003 if (tcp_inflight_debug > 0) {
2005 if ((u_int)(ticks - ltime) >= hz / tcp_inflight_debug) {
2007 kprintf("%p bw %ld rttbest %d srtt %d bwnd %ld\n",
2008 tp, bw, tp->t_rttbest, tp->t_srtt, bwnd);
2011 if ((long)bwnd < tcp_inflight_min)
2012 bwnd = tcp_inflight_min;
2013 if (bwnd > tcp_inflight_max)
2014 bwnd = tcp_inflight_max;
2015 if ((long)bwnd < tp->t_maxseg * 2)
2016 bwnd = tp->t_maxseg * 2;
2017 tp->snd_bwnd = bwnd;