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. Neither the name of the University nor the names of its contributors
47 * may be used to endorse or promote products derived from this software
48 * without specific prior written permission.
50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
62 * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95
63 * $FreeBSD: src/sys/netinet/tcp_subr.c,v 1.73.2.31 2003/01/24 05:11:34 sam Exp $
67 #include "opt_inet6.h"
68 #include "opt_ipsec.h"
69 #include "opt_tcpdebug.h"
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/callout.h>
74 #include <sys/kernel.h>
75 #include <sys/sysctl.h>
76 #include <sys/malloc.h>
77 #include <sys/mpipe.h>
80 #include <sys/domain.h>
84 #include <sys/socket.h>
85 #include <sys/socketops.h>
86 #include <sys/socketvar.h>
87 #include <sys/protosw.h>
88 #include <sys/random.h>
89 #include <sys/in_cksum.h>
92 #include <net/route.h>
94 #include <net/netisr2.h>
97 #include <netinet/in.h>
98 #include <netinet/in_systm.h>
99 #include <netinet/ip.h>
100 #include <netinet/ip6.h>
101 #include <netinet/in_pcb.h>
102 #include <netinet6/in6_pcb.h>
103 #include <netinet/in_var.h>
104 #include <netinet/ip_var.h>
105 #include <netinet6/ip6_var.h>
106 #include <netinet/ip_icmp.h>
108 #include <netinet/icmp6.h>
110 #include <netinet/tcp.h>
111 #include <netinet/tcp_fsm.h>
112 #include <netinet/tcp_seq.h>
113 #include <netinet/tcp_timer.h>
114 #include <netinet/tcp_timer2.h>
115 #include <netinet/tcp_var.h>
116 #include <netinet6/tcp6_var.h>
117 #include <netinet/tcpip.h>
119 #include <netinet/tcp_debug.h>
121 #include <netinet6/ip6protosw.h>
124 #include <netinet6/ipsec.h>
125 #include <netproto/key/key.h>
127 #include <netinet6/ipsec6.h>
132 #include <netproto/ipsec/ipsec.h>
134 #include <netproto/ipsec/ipsec6.h>
140 #include <machine/smp.h>
142 #include <sys/msgport2.h>
143 #include <sys/mplock2.h>
144 #include <net/netmsg2.h>
146 #if !defined(KTR_TCP)
147 #define KTR_TCP KTR_ALL
150 KTR_INFO_MASTER(tcp);
151 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
152 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
153 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
154 #define logtcp(name) KTR_LOG(tcp_ ## name)
157 #define TCP_IW_MAXSEGS_DFLT 4
158 #define TCP_IW_CAPSEGS_DFLT 4
160 struct tcp_reass_pcpu {
162 struct netmsg_base drain_nmsg;
165 struct inpcbinfo tcbinfo[MAXCPU];
166 struct tcpcbackq tcpcbackq[MAXCPU];
167 struct tcp_reass_pcpu tcp_reassq[MAXCPU];
169 int tcp_mssdflt = TCP_MSS;
170 SYSCTL_INT(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_RW,
171 &tcp_mssdflt, 0, "Default TCP Maximum Segment Size");
174 int tcp_v6mssdflt = TCP6_MSS;
175 SYSCTL_INT(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_RW,
176 &tcp_v6mssdflt, 0, "Default TCP Maximum Segment Size for IPv6");
180 * Minimum MSS we accept and use. This prevents DoS attacks where
181 * we are forced to a ridiculous low MSS like 20 and send hundreds
182 * of packets instead of one. The effect scales with the available
183 * bandwidth and quickly saturates the CPU and network interface
184 * with packet generation and sending. Set to zero to disable MINMSS
185 * checking. This setting prevents us from sending too small packets.
187 int tcp_minmss = TCP_MINMSS;
188 SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_RW,
189 &tcp_minmss , 0, "Minmum TCP Maximum Segment Size");
192 static int tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
193 SYSCTL_INT(_net_inet_tcp, TCPCTL_RTTDFLT, rttdflt, CTLFLAG_RW,
194 &tcp_rttdflt, 0, "Default maximum TCP Round Trip Time");
197 int tcp_do_rfc1323 = 1;
198 SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_RW,
199 &tcp_do_rfc1323, 0, "Enable rfc1323 (high performance TCP) extensions");
201 static int tcp_tcbhashsize = 0;
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RD,
203 &tcp_tcbhashsize, 0, "Size of TCP control block hashtable");
205 static int do_tcpdrain = 1;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0,
207 "Enable tcp_drain routine for extra help when low on mbufs");
209 static int icmp_may_rst = 1;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_RW, &icmp_may_rst, 0,
211 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
213 static int tcp_isn_reseed_interval = 0;
214 SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_RW,
215 &tcp_isn_reseed_interval, 0, "Seconds between reseeding of ISN secret");
218 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
219 * by default, but with generous values which should allow maximal
220 * bandwidth. In particular, the slop defaults to 50 (5 packets).
222 * The reason for doing this is that the limiter is the only mechanism we
223 * have which seems to do a really good job preventing receiver RX rings
224 * on network interfaces from getting blown out. Even though GigE/10GigE
225 * is supposed to flow control it looks like either it doesn't actually
226 * do it or Open Source drivers do not properly enable it.
228 * People using the limiter to reduce bottlenecks on slower WAN connections
229 * should set the slop to 20 (2 packets).
231 static int tcp_inflight_enable = 1;
232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_enable, CTLFLAG_RW,
233 &tcp_inflight_enable, 0, "Enable automatic TCP inflight data limiting");
235 static int tcp_inflight_debug = 0;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_debug, CTLFLAG_RW,
237 &tcp_inflight_debug, 0, "Debug TCP inflight calculations");
240 * NOTE: tcp_inflight_start is essentially the starting receive window
241 * for a connection. If set too low then fetches over tcp
242 * connections will take noticably longer to ramp-up over
243 * high-latency connections. 6144 is too low for a default,
244 * use something more reasonable.
246 static int tcp_inflight_start = 33792;
247 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_start, CTLFLAG_RW,
248 &tcp_inflight_start, 0, "Start value for TCP inflight window");
250 static int tcp_inflight_min = 6144;
251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_min, CTLFLAG_RW,
252 &tcp_inflight_min, 0, "Lower bound for TCP inflight window");
254 static int tcp_inflight_max = TCP_MAXWIN << TCP_MAX_WINSHIFT;
255 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_max, CTLFLAG_RW,
256 &tcp_inflight_max, 0, "Upper bound for TCP inflight window");
258 static int tcp_inflight_stab = 50;
259 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_stab, CTLFLAG_RW,
260 &tcp_inflight_stab, 0, "Fudge bw 1/10% (50=5%)");
262 static int tcp_inflight_adjrtt = 2;
263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, inflight_adjrtt, CTLFLAG_RW,
264 &tcp_inflight_adjrtt, 0, "Slop for rtt 1/(hz*32)");
266 static int tcp_do_rfc3390 = 1;
267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
269 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
271 static u_long tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
272 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwmaxsegs, CTLFLAG_RW,
273 &tcp_iw_maxsegs, 0, "TCP IW segments max");
275 static u_long tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
276 SYSCTL_ULONG(_net_inet_tcp, OID_AUTO, iwcapsegs, CTLFLAG_RW,
277 &tcp_iw_capsegs, 0, "TCP IW segments");
279 int tcp_low_rtobase = 1;
280 SYSCTL_INT(_net_inet_tcp, OID_AUTO, low_rtobase, CTLFLAG_RW,
281 &tcp_low_rtobase, 0, "Lowering the Initial RTO (RFC 6298)");
283 static int tcp_do_ncr = 1;
284 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr, CTLFLAG_RW,
285 &tcp_do_ncr, 0, "Non-Congestion Robustness (RFC 4653)");
287 int tcp_ncr_linklocal = 0;
288 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_linklocal, CTLFLAG_RW,
289 &tcp_ncr_linklocal, 0,
290 "Enable Non-Congestion Robustness (RFC 4653) on link local network");
292 int tcp_ncr_rxtthresh_max = 16;
293 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ncr_rxtthresh_max, CTLFLAG_RW,
294 &tcp_ncr_rxtthresh_max, 0,
295 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
297 static MALLOC_DEFINE(M_TCPTEMP, "tcptemp", "TCP Templates for Keepalives");
298 static struct malloc_pipe tcptemp_mpipe;
300 static void tcp_willblock(void);
301 static void tcp_notify (struct inpcb *, int);
303 struct tcp_stats tcpstats_percpu[MAXCPU] __cachealign;
304 struct tcp_state_count tcpstate_count[MAXCPU] __cachealign;
306 static void tcp_drain_dispatch(netmsg_t nmsg);
309 sysctl_tcpstats(SYSCTL_HANDLER_ARGS)
313 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
314 if ((error = SYSCTL_OUT(req, &tcpstats_percpu[cpu],
315 sizeof(struct tcp_stats))))
317 if ((error = SYSCTL_IN(req, &tcpstats_percpu[cpu],
318 sizeof(struct tcp_stats))))
324 SYSCTL_PROC(_net_inet_tcp, TCPCTL_STATS, stats, (CTLTYPE_OPAQUE | CTLFLAG_RW),
325 0, 0, sysctl_tcpstats, "S,tcp_stats", "TCP statistics");
328 * Target size of TCP PCB hash tables. Must be a power of two.
330 * Note that this can be overridden by the kernel environment
331 * variable net.inet.tcp.tcbhashsize
334 #define TCBHASHSIZE 512
338 * This is the actual shape of what we allocate using the zone
339 * allocator. Doing it this way allows us to protect both structures
340 * using the same generation count, and also eliminates the overhead
341 * of allocating tcpcbs separately. By hiding the structure here,
342 * we avoid changing most of the rest of the code (although it needs
343 * to be changed, eventually, for greater efficiency).
346 #define ALIGNM1 (ALIGNMENT - 1)
350 char align[(sizeof(struct inpcb) + ALIGNM1) & ~ALIGNM1];
353 struct tcp_callout inp_tp_rexmt;
354 struct tcp_callout inp_tp_persist;
355 struct tcp_callout inp_tp_keep;
356 struct tcp_callout inp_tp_2msl;
357 struct tcp_callout inp_tp_delack;
358 struct netmsg_tcp_timer inp_tp_timermsg;
359 struct netmsg_base inp_tp_sndmore;
370 struct inpcbportinfo *portinfo;
371 struct inpcbinfo *ticb;
372 int hashsize = TCBHASHSIZE;
376 * note: tcptemp is used for keepalives, and it is ok for an
377 * allocation to fail so do not specify MPF_INT.
379 mpipe_init(&tcptemp_mpipe, M_TCPTEMP, sizeof(struct tcptemp),
380 25, -1, 0, NULL, NULL, NULL);
382 tcp_delacktime = TCPTV_DELACK;
383 tcp_keepinit = TCPTV_KEEP_INIT;
384 tcp_keepidle = TCPTV_KEEP_IDLE;
385 tcp_keepintvl = TCPTV_KEEPINTVL;
386 tcp_maxpersistidle = TCPTV_KEEP_IDLE;
388 tcp_rexmit_min = TCPTV_MIN;
389 if (tcp_rexmit_min < 1) /* if kern.hz is too low */
391 tcp_rexmit_slop = TCPTV_CPU_VAR;
393 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize);
394 if (!powerof2(hashsize)) {
395 kprintf("WARNING: TCB hash size not a power of 2\n");
396 hashsize = 512; /* safe default */
398 tcp_tcbhashsize = hashsize;
400 portinfo = kmalloc_cachealign(sizeof(*portinfo) * netisr_ncpus, M_PCB,
403 for (cpu = 0; cpu < netisr_ncpus; cpu++) {
404 ticb = &tcbinfo[cpu];
405 in_pcbinfo_init(ticb, cpu, FALSE);
406 ticb->hashbase = hashinit(hashsize, M_PCB,
408 in_pcbportinfo_init(&portinfo[cpu], hashsize, cpu);
409 in_pcbportinfo_set(ticb, portinfo, netisr_ncpus);
410 ticb->wildcardhashbase = hashinit(hashsize, M_PCB,
411 &ticb->wildcardhashmask);
412 ticb->localgrphashbase = hashinit(hashsize, M_PCB,
413 &ticb->localgrphashmask);
414 ticb->ipi_size = sizeof(struct inp_tp);
415 TAILQ_INIT(&tcpcbackq[cpu].head);
418 tcp_reass_maxseg = nmbclusters / 16;
419 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg);
422 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
424 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
426 if (max_protohdr < TCP_MINPROTOHDR)
427 max_protohdr = TCP_MINPROTOHDR;
428 if (max_linkhdr + TCP_MINPROTOHDR > MHLEN)
430 #undef TCP_MINPROTOHDR
433 * Initialize TCP statistics counters for each CPU.
435 for (cpu = 0; cpu < netisr_ncpus; ++cpu)
436 bzero(&tcpstats_percpu[cpu], sizeof(struct tcp_stats));
439 * Initialize netmsgs for TCP drain
441 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
442 netmsg_init(&tcp_reassq[cpu].drain_nmsg, NULL,
443 &netisr_adone_rport, MSGF_PRIORITY, tcp_drain_dispatch);
447 netisr_register_rollup(tcp_willblock, NETISR_ROLLUP_PRIO_TCP);
456 while ((tp = TAILQ_FIRST(&tcpcbackq[cpu].head)) != NULL) {
457 KKASSERT(tp->t_flags & TF_ONOUTPUTQ);
458 tp->t_flags &= ~TF_ONOUTPUTQ;
459 TAILQ_REMOVE(&tcpcbackq[cpu].head, 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, boolean_t tso)
472 struct inpcb *inp = tp->t_inpcb;
473 struct tcphdr *tcp_hdr = (struct tcphdr *)tcp_ptr;
476 if (INP_ISIPV6(inp)) {
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;
495 ip->ip_vhl = IP_VHL_BORING;
502 ip->ip_p = IPPROTO_TCP;
503 ip->ip_src = inp->inp_laddr;
504 ip->ip_dst = inp->inp_faddr;
507 plen = htons(IPPROTO_TCP);
509 plen = htons(sizeof(struct tcphdr) + IPPROTO_TCP);
510 tcp_hdr->th_sum = in_pseudo(ip->ip_src.s_addr,
511 ip->ip_dst.s_addr, plen);
514 tcp_hdr->th_sport = inp->inp_lport;
515 tcp_hdr->th_dport = inp->inp_fport;
520 tcp_hdr->th_flags = 0;
526 * Create template to be used to send tcp packets on a connection.
527 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
528 * use for this function is in keepalives, which use tcp_respond.
531 tcp_maketemplate(struct tcpcb *tp)
535 if ((tmp = mpipe_alloc_nowait(&tcptemp_mpipe)) == NULL)
537 tcp_fillheaders(tp, &tmp->tt_ipgen, &tmp->tt_t, FALSE);
542 tcp_freetemplate(struct tcptemp *tmp)
544 mpipe_free(&tcptemp_mpipe, tmp);
548 * Send a single message to the TCP at address specified by
549 * the given TCP/IP header. If m == NULL, then we make a copy
550 * of the tcpiphdr at ti and send directly to the addressed host.
551 * This is used to force keep alive messages out using the TCP
552 * template for a connection. If flags are given then we send
553 * a message back to the TCP which originated the * segment ti,
554 * and discard the mbuf containing it and any other attached mbufs.
556 * In any case the ack and sequence number of the transmitted
557 * segment are as specified by the parameters.
559 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
562 tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m,
563 tcp_seq ack, tcp_seq seq, int flags)
567 struct route *ro = NULL;
569 struct ip *ip = ipgen;
572 struct route_in6 *ro6 = NULL;
573 struct route_in6 sro6;
574 struct ip6_hdr *ip6 = ipgen;
575 struct inpcb *inp = NULL;
576 boolean_t use_tmpro = TRUE;
578 boolean_t isipv6 = (IP_VHL_V(ip->ip_vhl) == 6);
580 const boolean_t isipv6 = FALSE;
585 if (!(flags & TH_RST)) {
586 win = ssb_space(&inp->inp_socket->so_rcv);
589 if (win > (long)TCP_MAXWIN << tp->rcv_scale)
590 win = (long)TCP_MAXWIN << tp->rcv_scale;
593 * Don't use the route cache of a listen socket,
594 * it is not MPSAFE; use temporary route cache.
596 if (tp->t_state != TCPS_LISTEN) {
598 ro6 = &inp->in6p_route;
600 ro = &inp->inp_route;
607 bzero(ro6, sizeof *ro6);
610 bzero(ro, sizeof *ro);
614 m = m_gethdr(M_NOWAIT, MT_HEADER);
618 m->m_data += max_linkhdr;
620 bcopy(ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr));
621 ip6 = mtod(m, struct ip6_hdr *);
622 nth = (struct tcphdr *)(ip6 + 1);
624 bcopy(ip, mtod(m, caddr_t), sizeof(struct ip));
625 ip = mtod(m, struct ip *);
626 nth = (struct tcphdr *)(ip + 1);
628 bcopy(th, nth, sizeof(struct tcphdr));
633 m->m_data = (caddr_t)ipgen;
634 /* m_len is set later */
636 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
638 xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr);
639 nth = (struct tcphdr *)(ip6 + 1);
641 xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, n_long);
642 nth = (struct tcphdr *)(ip + 1);
646 * this is usually a case when an extension header
647 * exists between the IPv6 header and the
650 nth->th_sport = th->th_sport;
651 nth->th_dport = th->th_dport;
653 xchg(nth->th_dport, nth->th_sport, n_short);
658 ip6->ip6_vfc = IPV6_VERSION;
659 ip6->ip6_nxt = IPPROTO_TCP;
660 ip6->ip6_plen = htons((u_short)(sizeof(struct tcphdr) + tlen));
661 tlen += sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
663 tlen += sizeof(struct tcpiphdr);
665 ip->ip_ttl = ip_defttl;
668 m->m_pkthdr.len = tlen;
669 m->m_pkthdr.rcvif = NULL;
670 nth->th_seq = htonl(seq);
671 nth->th_ack = htonl(ack);
673 nth->th_off = sizeof(struct tcphdr) >> 2;
674 nth->th_flags = flags;
676 nth->th_win = htons((u_short) (win >> tp->rcv_scale));
678 nth->th_win = htons((u_short)win);
682 nth->th_sum = in6_cksum(m, IPPROTO_TCP,
683 sizeof(struct ip6_hdr),
684 tlen - sizeof(struct ip6_hdr));
685 ip6->ip6_hlim = in6_selecthlim(inp,
686 (ro6 && ro6->ro_rt) ? ro6->ro_rt->rt_ifp : NULL);
688 nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
689 htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p)));
690 m->m_pkthdr.csum_flags = CSUM_TCP;
691 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
692 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr);
695 if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG))
696 tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0);
699 ip6_output(m, NULL, ro6, ipflags, NULL, NULL, inp);
700 if ((ro6 == &sro6) && (ro6->ro_rt != NULL)) {
705 if (inp != NULL && (inp->inp_flags & INP_HASH))
706 m_sethash(m, inp->inp_hashval);
707 ipflags |= IP_DEBUGROUTE;
708 ip_output(m, NULL, ro, ipflags, NULL, inp);
709 if ((ro == &sro) && (ro->ro_rt != NULL)) {
717 * Create a new TCP control block, making an
718 * empty reassembly queue and hooking it to the argument
719 * protocol control block. The `inp' parameter must have
720 * come from the zone allocator set up in tcp_init().
723 tcp_newtcpcb(struct inpcb *inp)
728 boolean_t isipv6 = INP_ISIPV6(inp);
730 const boolean_t isipv6 = FALSE;
733 it = (struct inp_tp *)inp;
735 bzero(tp, sizeof(struct tcpcb));
736 TAILQ_INIT(&tp->t_segq);
737 tp->t_maxseg = tp->t_maxopd = isipv6 ? tcp_v6mssdflt : tcp_mssdflt;
738 tp->t_rxtthresh = tcprexmtthresh;
740 /* Set up our timeouts. */
741 tp->tt_rexmt = &it->inp_tp_rexmt;
742 tp->tt_persist = &it->inp_tp_persist;
743 tp->tt_keep = &it->inp_tp_keep;
744 tp->tt_2msl = &it->inp_tp_2msl;
745 tp->tt_delack = &it->inp_tp_delack;
749 * Zero out timer message. We don't create it here,
750 * since the current CPU may not be the owner of this
753 tp->tt_msg = &it->inp_tp_timermsg;
754 bzero(tp->tt_msg, sizeof(*tp->tt_msg));
756 tp->t_keepinit = tcp_keepinit;
757 tp->t_keepidle = tcp_keepidle;
758 tp->t_keepintvl = tcp_keepintvl;
759 tp->t_keepcnt = tcp_keepcnt;
760 tp->t_maxidle = tp->t_keepintvl * tp->t_keepcnt;
763 tp->t_flags |= TF_NCR;
765 tp->t_flags |= (TF_REQ_SCALE | TF_REQ_TSTMP);
767 tp->t_inpcb = inp; /* XXX */
770 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
771 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
772 * reasonable initial retransmit time.
774 tp->t_srtt = TCPTV_SRTTBASE;
776 ((TCPTV_RTOBASE - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4;
777 tp->t_rttmin = tcp_rexmit_min;
778 tp->t_rxtcur = TCPTV_RTOBASE;
779 tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
780 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
781 tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT;
782 tp->snd_last = ticks;
783 tp->t_rcvtime = ticks;
785 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
786 * because the socket may be bound to an IPv6 wildcard address,
787 * which may match an IPv4-mapped IPv6 address.
789 inp->inp_ip_ttl = ip_defttl;
791 tcp_sack_tcpcb_init(tp);
793 tp->tt_sndmore = &it->inp_tp_sndmore;
798 * Drop a TCP connection, reporting the specified error.
799 * If connection is synchronized, then send a RST to peer.
802 tcp_drop(struct tcpcb *tp, int error)
804 struct socket *so = tp->t_inpcb->inp_socket;
806 if (TCPS_HAVERCVDSYN(tp->t_state)) {
807 TCP_STATE_CHANGE(tp, TCPS_CLOSED);
809 tcpstat.tcps_drops++;
811 tcpstat.tcps_conndrops++;
812 if (error == ETIMEDOUT && tp->t_softerror)
813 error = tp->t_softerror;
814 so->so_error = error;
815 return (tcp_close(tp));
818 struct netmsg_listen_detach {
819 struct netmsg_base base;
821 struct tcpcb *nm_tp_inh;
825 tcp_listen_detach_handler(netmsg_t msg)
827 struct netmsg_listen_detach *nmsg = (struct netmsg_listen_detach *)msg;
828 struct tcpcb *tp = nmsg->nm_tp;
829 int cpu = mycpuid, nextcpu;
831 if (tp->t_flags & TF_LISTEN) {
832 syncache_destroy(tp, nmsg->nm_tp_inh);
833 tcp_pcbport_merge_oncpu(tp);
836 in_pcbremwildcardhash_oncpu(tp->t_inpcb, &tcbinfo[cpu]);
839 if (nextcpu < netisr_ncpus)
840 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nmsg->base.lmsg);
842 lwkt_replymsg(&nmsg->base.lmsg, 0);
846 * Close a TCP control block:
847 * discard all space held by the tcp
848 * discard internet protocol block
849 * wake up any sleepers
852 tcp_close(struct tcpcb *tp)
855 struct inpcb *inp = tp->t_inpcb;
856 struct inpcb *inp_inh = NULL;
857 struct tcpcb *tp_inh = NULL;
858 struct socket *so = inp->inp_socket;
860 boolean_t dosavessthresh;
862 boolean_t isipv6 = INP_ISIPV6(inp);
864 const boolean_t isipv6 = FALSE;
867 if (tp->t_flags & TF_LISTEN) {
869 * Pending socket/syncache inheritance
871 * If this is a listen(2) socket, find another listen(2)
872 * socket in the same local group, which could inherit
873 * the syncache and sockets pending on the completion
874 * and incompletion queues.
877 * Currently the inheritance could only happen on the
878 * listen(2) sockets w/ SO_REUSEPORT set.
881 inp_inh = in_pcblocalgroup_last(&tcbinfo[0], inp);
883 tp_inh = intotcpcb(inp_inh);
887 * INP_WILDCARD indicates that listen(2) has been called on
888 * this socket. This implies:
889 * - A wildcard inp's hash is replicated for each protocol thread.
890 * - Syncache for this inp grows independently in each protocol
892 * - There is more than one cpu
894 * We have to chain a message to the rest of the protocol threads
895 * to cleanup the wildcard hash and the syncache. The cleanup
896 * in the current protocol thread is defered till the end of this
897 * function (syncache_destroy and in_pcbdetach).
900 * After cleanup the inp's hash and syncache entries, this inp will
901 * no longer be available to the rest of the protocol threads, so we
902 * are safe to whack the inp in the following code.
904 if ((inp->inp_flags & INP_WILDCARD) && netisr_ncpus > 1) {
905 struct netmsg_listen_detach nmsg;
907 KKASSERT(so->so_port == netisr_cpuport(0));
909 KKASSERT(inp->inp_pcbinfo == &tcbinfo[0]);
911 netmsg_init(&nmsg.base, NULL, &curthread->td_msgport,
912 MSGF_PRIORITY, tcp_listen_detach_handler);
914 nmsg.nm_tp_inh = tp_inh;
915 lwkt_domsg(netisr_cpuport(1), &nmsg.base.lmsg, 0);
921 * Make sure that all of our timers are stopped before we
922 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
923 * timers are never used. If timer message is never created
924 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
926 if (tp->tt_msg != NULL && tp->tt_msg->tt_tcb != NULL) {
927 tcp_callout_stop(tp, tp->tt_rexmt);
928 tcp_callout_stop(tp, tp->tt_persist);
929 tcp_callout_stop(tp, tp->tt_keep);
930 tcp_callout_stop(tp, tp->tt_2msl);
931 tcp_callout_stop(tp, tp->tt_delack);
934 if (tp->t_flags & TF_ONOUTPUTQ) {
935 KKASSERT(tp->tt_cpu == mycpu->gd_cpuid);
936 TAILQ_REMOVE(&tcpcbackq[tp->tt_cpu].head, tp, t_outputq);
937 tp->t_flags &= ~TF_ONOUTPUTQ;
941 * If we got enough samples through the srtt filter,
942 * save the rtt and rttvar in the routing entry.
943 * 'Enough' is arbitrarily defined as the 16 samples.
944 * 16 samples is enough for the srtt filter to converge
945 * to within 5% of the correct value; fewer samples and
946 * we could save a very bogus rtt.
948 * Don't update the default route's characteristics and don't
949 * update anything that the user "locked".
951 if (tp->t_rttupdated >= 16) {
955 struct sockaddr_in6 *sin6;
957 if ((rt = inp->in6p_route.ro_rt) == NULL)
959 sin6 = (struct sockaddr_in6 *)rt_key(rt);
960 if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr))
963 if ((rt = inp->inp_route.ro_rt) == NULL ||
964 ((struct sockaddr_in *)rt_key(rt))->
965 sin_addr.s_addr == INADDR_ANY)
968 if (!(rt->rt_rmx.rmx_locks & RTV_RTT)) {
969 i = tp->t_srtt * (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
970 if (rt->rt_rmx.rmx_rtt && i)
972 * filter this update to half the old & half
973 * the new values, converting scale.
974 * See route.h and tcp_var.h for a
975 * description of the scaling constants.
978 (rt->rt_rmx.rmx_rtt + i) / 2;
980 rt->rt_rmx.rmx_rtt = i;
981 tcpstat.tcps_cachedrtt++;
983 if (!(rt->rt_rmx.rmx_locks & RTV_RTTVAR)) {
985 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
986 if (rt->rt_rmx.rmx_rttvar && i)
987 rt->rt_rmx.rmx_rttvar =
988 (rt->rt_rmx.rmx_rttvar + i) / 2;
990 rt->rt_rmx.rmx_rttvar = i;
991 tcpstat.tcps_cachedrttvar++;
994 * The old comment here said:
995 * update the pipelimit (ssthresh) if it has been updated
996 * already or if a pipesize was specified & the threshhold
997 * got below half the pipesize. I.e., wait for bad news
998 * before we start updating, then update on both good
1001 * But we want to save the ssthresh even if no pipesize is
1002 * specified explicitly in the route, because such
1003 * connections still have an implicit pipesize specified
1004 * by the global tcp_sendspace. In the absence of a reliable
1005 * way to calculate the pipesize, it will have to do.
1007 i = tp->snd_ssthresh;
1008 if (rt->rt_rmx.rmx_sendpipe != 0)
1009 dosavessthresh = (i < rt->rt_rmx.rmx_sendpipe/2);
1011 dosavessthresh = (i < so->so_snd.ssb_hiwat/2);
1012 if (dosavessthresh ||
1013 (!(rt->rt_rmx.rmx_locks & RTV_SSTHRESH) && (i != 0) &&
1014 (rt->rt_rmx.rmx_ssthresh != 0))) {
1016 * convert the limit from user data bytes to
1017 * packets then to packet data bytes.
1019 i = (i + tp->t_maxseg / 2) / tp->t_maxseg;
1024 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1025 sizeof(struct tcpiphdr));
1026 if (rt->rt_rmx.rmx_ssthresh)
1027 rt->rt_rmx.rmx_ssthresh =
1028 (rt->rt_rmx.rmx_ssthresh + i) / 2;
1030 rt->rt_rmx.rmx_ssthresh = i;
1031 tcpstat.tcps_cachedssthresh++;
1036 /* free the reassembly queue, if any */
1037 while((q = TAILQ_FIRST(&tp->t_segq)) != NULL) {
1038 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1041 atomic_add_int(&tcp_reass_qsize, -1);
1043 /* throw away SACK blocks in scoreboard*/
1044 if (TCP_DO_SACK(tp))
1045 tcp_sack_destroy(&tp->scb);
1047 inp->inp_ppcb = NULL;
1048 soisdisconnected(so);
1049 /* note: pcb detached later on */
1051 tcp_destroy_timermsg(tp);
1052 tcp_output_cancel(tp);
1054 if (tp->t_flags & TF_LISTEN) {
1055 syncache_destroy(tp, tp_inh);
1056 tcp_pcbport_merge_oncpu(tp);
1057 tcp_pcbport_destroy(tp);
1058 if (inp_inh != NULL && inp_inh->inp_socket != NULL) {
1060 * Pending sockets inheritance only needs
1061 * to be done once in the current thread,
1064 soinherit(so, inp_inh->inp_socket);
1067 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache is not destroyed"));
1069 so_async_rcvd_drop(so);
1070 /* Drop the reference for the asynchronized pru_rcvd */
1075 * - Remove self from listen tcpcb per-cpu port cache _before_
1077 * - pcbdetach removes any wildcard hash entry on the current CPU.
1079 tcp_pcbport_remove(inp);
1087 tcpstat.tcps_closed++;
1092 * Walk the tcpbs, if existing, and flush the reassembly queue,
1093 * if there is one...
1096 tcp_drain_oncpu(struct inpcbinfo *pcbinfo)
1098 struct inpcbhead *head = &pcbinfo->pcblisthead;
1102 * Since we run in netisr, it is MP safe, even if
1103 * we block during the inpcb list iteration, i.e.
1104 * we don't need to use inpcb marker here.
1106 ASSERT_NETISR_NCPUS(pcbinfo->cpu);
1108 LIST_FOREACH(inpb, head, inp_list) {
1110 struct tseg_qent *te;
1112 if (inpb->inp_flags & INP_PLACEMARKER)
1115 tcpb = intotcpcb(inpb);
1116 KASSERT(tcpb != NULL, ("tcp_drain_oncpu: tcpb is NULL"));
1118 if ((te = TAILQ_FIRST(&tcpb->t_segq)) != NULL) {
1119 TAILQ_REMOVE(&tcpb->t_segq, te, tqe_q);
1120 if (te->tqe_th->th_flags & TH_FIN)
1121 tcpb->t_flags &= ~TF_QUEDFIN;
1124 atomic_add_int(&tcp_reass_qsize, -1);
1131 tcp_drain_dispatch(netmsg_t nmsg)
1134 lwkt_replymsg(&nmsg->lmsg, 0); /* reply ASAP */
1137 tcp_drain_oncpu(&tcbinfo[mycpuid]);
1138 tcp_reassq[mycpuid].draining = 0;
1142 tcp_drain_ipi(void *arg __unused)
1145 struct lwkt_msg *msg = &tcp_reassq[cpu].drain_nmsg.lmsg;
1148 if (msg->ms_flags & MSGF_DONE)
1149 lwkt_sendmsg_oncpu(netisr_cpuport(cpu), msg);
1162 if (tcp_reass_qsize == 0)
1165 CPUMASK_ASSBMASK(mask, netisr_ncpus);
1166 CPUMASK_ANDMASK(mask, smp_active_mask);
1169 if (IN_NETISR_NCPUS(cpu)) {
1170 tcp_drain_oncpu(&tcbinfo[cpu]);
1171 CPUMASK_NANDBIT(mask, cpu);
1174 if (tcp_reass_qsize < netisr_ncpus) {
1175 /* Does not worth the trouble. */
1179 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
1180 if (!CPUMASK_TESTBIT(mask, cpu))
1183 if (tcp_reassq[cpu].draining) {
1184 /* Draining; skip this cpu. */
1185 CPUMASK_NANDBIT(mask, cpu);
1188 tcp_reassq[cpu].draining = 1;
1191 if (CPUMASK_TESTNZERO(mask))
1192 lwkt_send_ipiq_mask(mask, tcp_drain_ipi, NULL);
1196 * Notify a tcp user of an asynchronous error;
1197 * store error as soft error, but wake up user
1198 * (for now, won't do anything until can select for soft error).
1200 * Do not wake up user since there currently is no mechanism for
1201 * reporting soft errors (yet - a kqueue filter may be added).
1204 tcp_notify(struct inpcb *inp, int error)
1206 struct tcpcb *tp = intotcpcb(inp);
1209 * Ignore some errors if we are hooked up.
1210 * If connection hasn't completed, has retransmitted several times,
1211 * and receives a second error, give up now. This is better
1212 * than waiting a long time to establish a connection that
1213 * can never complete.
1215 if (tp->t_state == TCPS_ESTABLISHED &&
1216 (error == EHOSTUNREACH || error == ENETUNREACH ||
1217 error == EHOSTDOWN)) {
1219 } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 &&
1221 tcp_drop(tp, error);
1223 tp->t_softerror = error;
1225 wakeup(&so->so_timeo);
1232 tcp_pcblist(SYSCTL_HANDLER_ARGS)
1235 struct inpcb *marker;
1243 * The process of preparing the TCB list is too time-consuming and
1244 * resource-intensive to repeat twice on every request.
1246 if (req->oldptr == NULL) {
1247 for (ccpu = 0; ccpu < netisr_ncpus; ++ccpu)
1248 n += tcbinfo[ccpu].ipi_count;
1249 req->oldidx = (n + n/8 + 10) * sizeof(struct xtcpcb);
1253 if (req->newptr != NULL)
1256 marker = kmalloc(sizeof(struct inpcb), M_TEMP, M_WAITOK|M_ZERO);
1257 marker->inp_flags |= INP_PLACEMARKER;
1260 * OK, now we're committed to doing something. Run the inpcb list
1261 * for each cpu in the system and construct the output. Use a
1262 * list placemarker to deal with list changes occuring during
1263 * copyout blockages (but otherwise depend on being on the correct
1264 * cpu to avoid races).
1266 origcpu = mycpu->gd_cpuid;
1267 for (ccpu = 0; ccpu < netisr_ncpus && error == 0; ++ccpu) {
1271 lwkt_migratecpu(ccpu);
1273 n = tcbinfo[ccpu].ipi_count;
1275 LIST_INSERT_HEAD(&tcbinfo[ccpu].pcblisthead, marker, inp_list);
1277 while ((inp = LIST_NEXT(marker, inp_list)) != NULL && i < n) {
1279 * process a snapshot of pcbs, ignoring placemarkers
1280 * and using our own to allow SYSCTL_OUT to block.
1282 LIST_REMOVE(marker, inp_list);
1283 LIST_INSERT_AFTER(inp, marker, inp_list);
1285 if (inp->inp_flags & INP_PLACEMARKER)
1287 if (prison_xinpcb(req->td, inp))
1290 xt.xt_len = sizeof xt;
1291 bcopy(inp, &xt.xt_inp, sizeof *inp);
1292 inp_ppcb = inp->inp_ppcb;
1293 if (inp_ppcb != NULL)
1294 bcopy(inp_ppcb, &xt.xt_tp, sizeof xt.xt_tp);
1296 bzero(&xt.xt_tp, sizeof xt.xt_tp);
1297 if (inp->inp_socket)
1298 sotoxsocket(inp->inp_socket, &xt.xt_socket);
1299 if ((error = SYSCTL_OUT(req, &xt, sizeof xt)) != 0)
1303 LIST_REMOVE(marker, inp_list);
1304 if (error == 0 && i < n) {
1305 bzero(&xt, sizeof xt);
1306 xt.xt_len = sizeof xt;
1308 error = SYSCTL_OUT(req, &xt, sizeof xt);
1317 * Make sure we are on the same cpu we were on originally, since
1318 * higher level callers expect this. Also don't pollute caches with
1319 * migrated userland data by (eventually) returning to userland
1320 * on a different cpu.
1322 lwkt_migratecpu(origcpu);
1323 kfree(marker, M_TEMP);
1327 SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLFLAG_RD, 0, 0,
1328 tcp_pcblist, "S,xtcpcb", "List of active TCP connections");
1331 tcp_getcred(SYSCTL_HANDLER_ARGS)
1333 struct sockaddr_in addrs[2];
1334 struct ucred cred0, *cred = NULL;
1336 int cpu, origcpu, error;
1338 error = priv_check(req->td, PRIV_ROOT);
1341 error = SYSCTL_IN(req, addrs, sizeof addrs);
1346 cpu = tcp_addrcpu(addrs[1].sin_addr.s_addr, addrs[1].sin_port,
1347 addrs[0].sin_addr.s_addr, addrs[0].sin_port);
1349 lwkt_migratecpu(cpu);
1351 inp = in_pcblookup_hash(&tcbinfo[cpu], addrs[1].sin_addr,
1352 addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, 0, NULL);
1353 if (inp == NULL || inp->inp_socket == NULL) {
1355 } else if (inp->inp_socket->so_cred != NULL) {
1356 cred0 = *(inp->inp_socket->so_cred);
1360 lwkt_migratecpu(origcpu);
1365 return SYSCTL_OUT(req, cred, sizeof(struct ucred));
1368 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1369 0, 0, tcp_getcred, "S,ucred", "Get the ucred of a TCP connection");
1373 tcp6_getcred(SYSCTL_HANDLER_ARGS)
1375 struct sockaddr_in6 addrs[2];
1379 error = priv_check(req->td, PRIV_ROOT);
1382 error = SYSCTL_IN(req, addrs, sizeof addrs);
1386 inp = in6_pcblookup_hash(&tcbinfo[0],
1387 &addrs[1].sin6_addr, addrs[1].sin6_port,
1388 &addrs[0].sin6_addr, addrs[0].sin6_port, 0, NULL);
1389 if (inp == NULL || inp->inp_socket == NULL) {
1393 error = SYSCTL_OUT(req, inp->inp_socket->so_cred, sizeof(struct ucred));
1399 SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, (CTLTYPE_OPAQUE | CTLFLAG_RW),
1401 tcp6_getcred, "S,ucred", "Get the ucred of a TCP6 connection");
1404 struct netmsg_tcp_notify {
1405 struct netmsg_base base;
1406 inp_notify_t nm_notify;
1407 struct in_addr nm_faddr;
1412 tcp_notifyall_oncpu(netmsg_t msg)
1414 struct netmsg_tcp_notify *nm = (struct netmsg_tcp_notify *)msg;
1417 ASSERT_NETISR_NCPUS(mycpuid);
1419 in_pcbnotifyall(&tcbinfo[mycpuid], nm->nm_faddr,
1420 nm->nm_arg, nm->nm_notify);
1422 nextcpu = mycpuid + 1;
1423 if (nextcpu < netisr_ncpus)
1424 lwkt_forwardmsg(netisr_cpuport(nextcpu), &nm->base.lmsg);
1426 lwkt_replymsg(&nm->base.lmsg, 0);
1430 tcp_get_inpnotify(int cmd, const struct sockaddr *sa,
1431 int *arg, struct ip **ip0, int *cpuid)
1433 struct ip *ip = *ip0;
1434 struct in_addr faddr;
1435 inp_notify_t notify = tcp_notify;
1437 faddr = ((const struct sockaddr_in *)sa)->sin_addr;
1438 if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY)
1441 *arg = inetctlerrmap[cmd];
1442 if (cmd == PRC_QUENCH) {
1443 notify = tcp_quench;
1444 } else if (icmp_may_rst &&
1445 (cmd == PRC_UNREACH_ADMIN_PROHIB ||
1446 cmd == PRC_UNREACH_PORT ||
1447 cmd == PRC_TIMXCEED_INTRANS) &&
1449 notify = tcp_drop_syn_sent;
1450 } else if (cmd == PRC_MSGSIZE) {
1451 const struct icmp *icmp = (const struct icmp *)
1452 ((caddr_t)ip - offsetof(struct icmp, icmp_ip));
1454 *arg = ntohs(icmp->icmp_nextmtu);
1455 notify = tcp_mtudisc;
1456 } else if (PRC_IS_REDIRECT(cmd)) {
1458 notify = in_rtchange;
1459 } else if (cmd == PRC_HOSTDEAD) {
1461 } else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) {
1465 if (cpuid != NULL) {
1467 /* Go through all effective netisr CPUs. */
1468 *cpuid = netisr_ncpus;
1470 const struct tcphdr *th;
1472 th = (const struct tcphdr *)
1473 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2));
1474 *cpuid = tcp_addrcpu(faddr.s_addr, th->th_dport,
1475 ip->ip_src.s_addr, th->th_sport);
1484 tcp_ctlinput(netmsg_t msg)
1486 int cmd = msg->ctlinput.nm_cmd;
1487 struct sockaddr *sa = msg->ctlinput.nm_arg;
1488 struct ip *ip = msg->ctlinput.nm_extra;
1489 struct in_addr faddr;
1490 inp_notify_t notify;
1493 ASSERT_NETISR_NCPUS(mycpuid);
1495 notify = tcp_get_inpnotify(cmd, sa, &arg, &ip, &cpuid);
1499 faddr = ((struct sockaddr_in *)sa)->sin_addr;
1501 const struct tcphdr *th;
1504 if (cpuid != mycpuid)
1507 th = (const struct tcphdr *)
1508 ((caddr_t)ip + (IP_VHL_HL(ip->ip_vhl) << 2));
1509 inp = in_pcblookup_hash(&tcbinfo[mycpuid], faddr, th->th_dport,
1510 ip->ip_src, th->th_sport, 0, NULL);
1511 if (inp != NULL && inp->inp_socket != NULL) {
1512 tcp_seq icmpseq = htonl(th->th_seq);
1513 struct tcpcb *tp = intotcpcb(inp);
1515 if (SEQ_GEQ(icmpseq, tp->snd_una) &&
1516 SEQ_LT(icmpseq, tp->snd_max))
1519 struct in_conninfo inc;
1521 inc.inc_fport = th->th_dport;
1522 inc.inc_lport = th->th_sport;
1523 inc.inc_faddr = faddr;
1524 inc.inc_laddr = ip->ip_src;
1528 syncache_unreach(&inc, th);
1530 } else if (msg->ctlinput.nm_direct) {
1531 if (cpuid != netisr_ncpus && cpuid != mycpuid)
1534 in_pcbnotifyall(&tcbinfo[mycpuid], faddr, arg, notify);
1536 struct netmsg_tcp_notify *nm;
1539 nm = kmalloc(sizeof(*nm), M_LWKTMSG, M_INTWAIT);
1540 netmsg_init(&nm->base, NULL, &netisr_afree_rport,
1541 0, tcp_notifyall_oncpu);
1542 nm->nm_faddr = faddr;
1544 nm->nm_notify = notify;
1546 lwkt_sendmsg(netisr_cpuport(0), &nm->base.lmsg);
1549 lwkt_replymsg(&msg->lmsg, 0);
1555 tcp6_ctlinput(netmsg_t msg)
1557 int cmd = msg->ctlinput.nm_cmd;
1558 struct sockaddr *sa = msg->ctlinput.nm_arg;
1559 void *d = msg->ctlinput.nm_extra;
1561 inp_notify_t notify = tcp_notify;
1562 struct ip6_hdr *ip6;
1564 struct ip6ctlparam *ip6cp = NULL;
1565 const struct sockaddr_in6 *sa6_src = NULL;
1567 struct tcp_portonly {
1573 if (sa->sa_family != AF_INET6 ||
1574 sa->sa_len != sizeof(struct sockaddr_in6)) {
1579 if (cmd == PRC_QUENCH)
1580 notify = tcp_quench;
1581 else if (cmd == PRC_MSGSIZE) {
1582 struct ip6ctlparam *ip6cp = d;
1583 struct icmp6_hdr *icmp6 = ip6cp->ip6c_icmp6;
1585 arg = ntohl(icmp6->icmp6_mtu);
1586 notify = tcp_mtudisc;
1587 } else if (!PRC_IS_REDIRECT(cmd) &&
1588 ((unsigned)cmd > PRC_NCMDS || inet6ctlerrmap[cmd] == 0)) {
1592 /* if the parameter is from icmp6, decode it. */
1594 ip6cp = (struct ip6ctlparam *)d;
1596 ip6 = ip6cp->ip6c_ip6;
1597 off = ip6cp->ip6c_off;
1598 sa6_src = ip6cp->ip6c_src;
1602 off = 0; /* fool gcc */
1607 struct in_conninfo inc;
1609 * XXX: We assume that when IPV6 is non NULL,
1610 * M and OFF are valid.
1613 /* check if we can safely examine src and dst ports */
1614 if (m->m_pkthdr.len < off + sizeof *thp)
1617 bzero(&th, sizeof th);
1618 m_copydata(m, off, sizeof *thp, (caddr_t)&th);
1620 in6_pcbnotify(&tcbinfo[0], sa, th.th_dport,
1621 (struct sockaddr *)ip6cp->ip6c_src,
1622 th.th_sport, cmd, arg, notify);
1624 inc.inc_fport = th.th_dport;
1625 inc.inc_lport = th.th_sport;
1626 inc.inc6_faddr = ((struct sockaddr_in6 *)sa)->sin6_addr;
1627 inc.inc6_laddr = ip6cp->ip6c_src->sin6_addr;
1629 syncache_unreach(&inc, &th);
1631 in6_pcbnotify(&tcbinfo[0], sa, 0,
1632 (const struct sockaddr *)sa6_src, 0, cmd, arg, notify);
1635 lwkt_replymsg(&msg->ctlinput.base.lmsg, 0);
1641 * Following is where TCP initial sequence number generation occurs.
1643 * There are two places where we must use initial sequence numbers:
1644 * 1. In SYN-ACK packets.
1645 * 2. In SYN packets.
1647 * All ISNs for SYN-ACK packets are generated by the syncache. See
1648 * tcp_syncache.c for details.
1650 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1651 * depends on this property. In addition, these ISNs should be
1652 * unguessable so as to prevent connection hijacking. To satisfy
1653 * the requirements of this situation, the algorithm outlined in
1654 * RFC 1948 is used to generate sequence numbers.
1656 * Implementation details:
1658 * Time is based off the system timer, and is corrected so that it
1659 * increases by one megabyte per second. This allows for proper
1660 * recycling on high speed LANs while still leaving over an hour
1663 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1664 * between seeding of isn_secret. This is normally set to zero,
1665 * as reseeding should not be necessary.
1669 #define ISN_BYTES_PER_SECOND 1048576
1671 u_char isn_secret[32];
1672 int isn_last_reseed;
1676 tcp_new_isn(struct tcpcb *tp)
1678 u_int32_t md5_buffer[4];
1681 /* Seed if this is the first use, reseed if requested. */
1682 if ((isn_last_reseed == 0) || ((tcp_isn_reseed_interval > 0) &&
1683 (((u_int)isn_last_reseed + (u_int)tcp_isn_reseed_interval*hz)
1685 read_random_unlimited(&isn_secret, sizeof isn_secret);
1686 isn_last_reseed = ticks;
1689 /* Compute the md5 hash and return the ISN. */
1691 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_fport, sizeof(u_short));
1692 MD5Update(&isn_ctx, (u_char *)&tp->t_inpcb->inp_lport, sizeof(u_short));
1694 if (INP_ISIPV6(tp->t_inpcb)) {
1695 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_faddr,
1696 sizeof(struct in6_addr));
1697 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->in6p_laddr,
1698 sizeof(struct in6_addr));
1702 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_faddr,
1703 sizeof(struct in_addr));
1704 MD5Update(&isn_ctx, (u_char *) &tp->t_inpcb->inp_laddr,
1705 sizeof(struct in_addr));
1707 MD5Update(&isn_ctx, (u_char *) &isn_secret, sizeof(isn_secret));
1708 MD5Final((u_char *) &md5_buffer, &isn_ctx);
1709 new_isn = (tcp_seq) md5_buffer[0];
1710 new_isn += ticks * (ISN_BYTES_PER_SECOND / hz);
1715 * When a source quench is received, close congestion window
1716 * to one segment. We will gradually open it again as we proceed.
1719 tcp_quench(struct inpcb *inp, int error)
1721 struct tcpcb *tp = intotcpcb(inp);
1723 KASSERT(tp != NULL, ("tcp_quench: tp is NULL"));
1724 tp->snd_cwnd = tp->t_maxseg;
1729 * When a specific ICMP unreachable message is received and the
1730 * connection state is SYN-SENT, drop the connection. This behavior
1731 * is controlled by the icmp_may_rst sysctl.
1734 tcp_drop_syn_sent(struct inpcb *inp, int error)
1736 struct tcpcb *tp = intotcpcb(inp);
1738 KASSERT(tp != NULL, ("tcp_drop_syn_sent: tp is NULL"));
1739 if (tp->t_state == TCPS_SYN_SENT)
1740 tcp_drop(tp, error);
1744 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1745 * based on the new value in the route. Also nudge TCP to send something,
1746 * since we know the packet we just sent was dropped.
1747 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1750 tcp_mtudisc(struct inpcb *inp, int mtu)
1752 struct tcpcb *tp = intotcpcb(inp);
1754 struct socket *so = inp->inp_socket;
1757 boolean_t isipv6 = INP_ISIPV6(inp);
1759 const boolean_t isipv6 = FALSE;
1762 KASSERT(tp != NULL, ("tcp_mtudisc: tp is NULL"));
1765 * If no MTU is provided in the ICMP message, use the
1766 * next lower likely value, as specified in RFC 1191.
1771 oldmtu = tp->t_maxopd +
1773 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1774 sizeof(struct tcpiphdr));
1775 mtu = ip_next_mtu(oldmtu, 0);
1779 rt = tcp_rtlookup6(&inp->inp_inc);
1781 rt = tcp_rtlookup(&inp->inp_inc);
1783 if (rt->rt_rmx.rmx_mtu != 0 && rt->rt_rmx.rmx_mtu < mtu)
1784 mtu = rt->rt_rmx.rmx_mtu;
1788 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1789 sizeof(struct tcpiphdr));
1792 * XXX - The following conditional probably violates the TCP
1793 * spec. The problem is that, since we don't know the
1794 * other end's MSS, we are supposed to use a conservative
1795 * default. But, if we do that, then MTU discovery will
1796 * never actually take place, because the conservative
1797 * default is much less than the MTUs typically seen
1798 * on the Internet today. For the moment, we'll sweep
1799 * this under the carpet.
1801 * The conservative default might not actually be a problem
1802 * if the only case this occurs is when sending an initial
1803 * SYN with options and data to a host we've never talked
1804 * to before. Then, they will reply with an MSS value which
1805 * will get recorded and the new parameters should get
1806 * recomputed. For Further Study.
1808 if (rt->rt_rmx.rmx_mssopt && rt->rt_rmx.rmx_mssopt < maxopd)
1809 maxopd = rt->rt_rmx.rmx_mssopt;
1813 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
1814 sizeof(struct tcpiphdr));
1816 if (tp->t_maxopd <= maxopd)
1818 tp->t_maxopd = maxopd;
1821 if ((tp->t_flags & (TF_REQ_TSTMP | TF_RCVD_TSTMP | TF_NOOPT)) ==
1822 (TF_REQ_TSTMP | TF_RCVD_TSTMP))
1823 mss -= TCPOLEN_TSTAMP_APPA;
1825 /* round down to multiple of MCLBYTES */
1826 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1828 mss &= ~(MCLBYTES - 1);
1831 mss = (mss / MCLBYTES) * MCLBYTES;
1834 if (so->so_snd.ssb_hiwat < mss)
1835 mss = so->so_snd.ssb_hiwat;
1839 tp->snd_nxt = tp->snd_una;
1841 tcpstat.tcps_mturesent++;
1845 * Look-up the routing entry to the peer of this inpcb. If no route
1846 * is found and it cannot be allocated the return NULL. This routine
1847 * is called by TCP routines that access the rmx structure and by tcp_mss
1848 * to get the interface MTU.
1851 tcp_rtlookup(struct in_conninfo *inc)
1853 struct route *ro = &inc->inc_route;
1855 if (ro->ro_rt == NULL || !(ro->ro_rt->rt_flags & RTF_UP)) {
1856 /* No route yet, so try to acquire one */
1857 if (inc->inc_faddr.s_addr != INADDR_ANY) {
1859 * unused portions of the structure MUST be zero'd
1860 * out because rtalloc() treats it as opaque data
1862 bzero(&ro->ro_dst, sizeof(struct sockaddr_in));
1863 ro->ro_dst.sa_family = AF_INET;
1864 ro->ro_dst.sa_len = sizeof(struct sockaddr_in);
1865 ((struct sockaddr_in *) &ro->ro_dst)->sin_addr =
1875 tcp_rtlookup6(struct in_conninfo *inc)
1877 struct route_in6 *ro6 = &inc->inc6_route;
1879 if (ro6->ro_rt == NULL || !(ro6->ro_rt->rt_flags & RTF_UP)) {
1880 /* No route yet, so try to acquire one */
1881 if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) {
1883 * unused portions of the structure MUST be zero'd
1884 * out because rtalloc() treats it as opaque data
1886 bzero(&ro6->ro_dst, sizeof(struct sockaddr_in6));
1887 ro6->ro_dst.sin6_family = AF_INET6;
1888 ro6->ro_dst.sin6_len = sizeof(struct sockaddr_in6);
1889 ro6->ro_dst.sin6_addr = inc->inc6_faddr;
1890 rtalloc((struct route *)ro6);
1893 return (ro6->ro_rt);
1898 /* compute ESP/AH header size for TCP, including outer IP header. */
1900 ipsec_hdrsiz_tcp(struct tcpcb *tp)
1908 if ((tp == NULL) || ((inp = tp->t_inpcb) == NULL))
1910 MGETHDR(m, M_NOWAIT, MT_DATA);
1915 if (INP_ISIPV6(inp)) {
1916 struct ip6_hdr *ip6 = mtod(m, struct ip6_hdr *);
1918 th = (struct tcphdr *)(ip6 + 1);
1919 m->m_pkthdr.len = m->m_len =
1920 sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1921 tcp_fillheaders(tp, ip6, th, FALSE);
1922 hdrsiz = ipsec6_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1926 ip = mtod(m, struct ip *);
1927 th = (struct tcphdr *)(ip + 1);
1928 m->m_pkthdr.len = m->m_len = sizeof(struct tcpiphdr);
1929 tcp_fillheaders(tp, ip, th, FALSE);
1930 hdrsiz = ipsec4_hdrsiz(m, IPSEC_DIR_OUTBOUND, inp);
1939 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1941 * This code attempts to calculate the bandwidth-delay product as a
1942 * means of determining the optimal window size to maximize bandwidth,
1943 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1944 * routers. This code also does a fairly good job keeping RTTs in check
1945 * across slow links like modems. We implement an algorithm which is very
1946 * similar (but not meant to be) TCP/Vegas. The code operates on the
1947 * transmitter side of a TCP connection and so only effects the transmit
1948 * side of the connection.
1950 * BACKGROUND: TCP makes no provision for the management of buffer space
1951 * at the end points or at the intermediate routers and switches. A TCP
1952 * stream, whether using NewReno or not, will eventually buffer as
1953 * many packets as it is able and the only reason this typically works is
1954 * due to the fairly small default buffers made available for a connection
1955 * (typicaly 16K or 32K). As machines use larger windows and/or window
1956 * scaling it is now fairly easy for even a single TCP connection to blow-out
1957 * all available buffer space not only on the local interface, but on
1958 * intermediate routers and switches as well. NewReno makes a misguided
1959 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1960 * then backing off, then steadily increasing the window again until another
1961 * failure occurs, ad-infinitum. This results in terrible oscillation that
1962 * is only made worse as network loads increase and the idea of intentionally
1963 * blowing out network buffers is, frankly, a terrible way to manage network
1966 * It is far better to limit the transmit window prior to the failure
1967 * condition being achieved. There are two general ways to do this: First
1968 * you can 'scan' through different transmit window sizes and locate the
1969 * point where the RTT stops increasing, indicating that you have filled the
1970 * pipe, then scan backwards until you note that RTT stops decreasing, then
1971 * repeat ad-infinitum. This method works in principle but has severe
1972 * implementation issues due to RTT variances, timer granularity, and
1973 * instability in the algorithm which can lead to many false positives and
1974 * create oscillations as well as interact badly with other TCP streams
1975 * implementing the same algorithm.
1977 * The second method is to limit the window to the bandwidth delay product
1978 * of the link. This is the method we implement. RTT variances and our
1979 * own manipulation of the congestion window, bwnd, can potentially
1980 * destabilize the algorithm. For this reason we have to stabilize the
1981 * elements used to calculate the window. We do this by using the minimum
1982 * observed RTT, the long term average of the observed bandwidth, and
1983 * by adding two segments worth of slop. It isn't perfect but it is able
1984 * to react to changing conditions and gives us a very stable basis on
1985 * which to extend the algorithm.
1988 tcp_xmit_bandwidth_limit(struct tcpcb *tp, tcp_seq ack_seq)
1997 * If inflight_enable is disabled in the middle of a tcp connection,
1998 * make sure snd_bwnd is effectively disabled.
2000 if (!tcp_inflight_enable) {
2001 tp->snd_bwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT;
2002 tp->snd_bandwidth = 0;
2007 * Validate the delta time. If a connection is new or has been idle
2008 * a long time we have to reset the bandwidth calculator.
2012 delta_ticks = save_ticks - tp->t_bw_rtttime;
2013 if (tp->t_bw_rtttime == 0 || delta_ticks < 0 || delta_ticks > hz * 10) {
2014 tp->t_bw_rtttime = save_ticks;
2015 tp->t_bw_rtseq = ack_seq;
2016 if (tp->snd_bandwidth == 0)
2017 tp->snd_bandwidth = tcp_inflight_start;
2022 * A delta of at least 1 tick is required. Waiting 2 ticks will
2023 * result in better (bw) accuracy. More than that and the ramp-up
2026 if (delta_ticks == 0 || delta_ticks == 1)
2030 * Sanity check, plus ignore pure window update acks.
2032 if ((int)(ack_seq - tp->t_bw_rtseq) <= 0)
2036 * Figure out the bandwidth. Due to the tick granularity this
2037 * is a very rough number and it MUST be averaged over a fairly
2038 * long period of time. XXX we need to take into account a link
2039 * that is not using all available bandwidth, but for now our
2040 * slop will ramp us up if this case occurs and the bandwidth later
2043 ibw = (int64_t)(ack_seq - tp->t_bw_rtseq) * hz / delta_ticks;
2044 tp->t_bw_rtttime = save_ticks;
2045 tp->t_bw_rtseq = ack_seq;
2046 bw = ((int64_t)tp->snd_bandwidth * 15 + ibw) >> 4;
2048 tp->snd_bandwidth = bw;
2051 * Calculate the semi-static bandwidth delay product, plus two maximal
2052 * segments. The additional slop puts us squarely in the sweet
2053 * spot and also handles the bandwidth run-up case. Without the
2054 * slop we could be locking ourselves into a lower bandwidth.
2056 * At very high speeds the bw calculation can become overly sensitive
2057 * and error prone when delta_ticks is low (e.g. usually 1). To deal
2058 * with the problem the stab must be scaled to the bw. A stab of 50
2059 * (the default) increases the bw for the purposes of the bwnd
2060 * calculation by 5%.
2062 * Situations Handled:
2063 * (1) Prevents over-queueing of packets on LANs, especially on
2064 * high speed LANs, allowing larger TCP buffers to be
2065 * specified, and also does a good job preventing
2066 * over-queueing of packets over choke points like modems
2067 * (at least for the transmit side).
2069 * (2) Is able to handle changing network loads (bandwidth
2070 * drops so bwnd drops, bandwidth increases so bwnd
2073 * (3) Theoretically should stabilize in the face of multiple
2074 * connections implementing the same algorithm (this may need
2077 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2078 * be adjusted with a sysctl but typically only needs to be on
2079 * very slow connections. A value no smaller then 5 should
2080 * be used, but only reduce this default if you have no other
2084 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2085 bw += bw * tcp_inflight_stab / 1000;
2086 bwnd = (int64_t)bw * USERTT / (hz << TCP_RTT_SHIFT) +
2087 (int)tp->t_maxseg * 2;
2090 if (tcp_inflight_debug > 0) {
2092 if ((u_int)(save_ticks - ltime) >= hz / tcp_inflight_debug) {
2094 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2095 "bwnd %ld delta %d snd_win %ld\n",
2096 tp, ibw, bw, tp->t_rttvar, tp->t_srtt,
2097 bwnd, delta_ticks, tp->snd_wnd);
2100 if ((long)bwnd < tcp_inflight_min)
2101 bwnd = tcp_inflight_min;
2102 if (bwnd > tcp_inflight_max)
2103 bwnd = tcp_inflight_max;
2104 if ((long)bwnd < tp->t_maxseg * 2)
2105 bwnd = tp->t_maxseg * 2;
2106 tp->snd_bwnd = bwnd;
2110 tcp_rmx_iwsegs(struct tcpcb *tp, u_long *maxsegs, u_long *capsegs)
2113 struct inpcb *inp = tp->t_inpcb;
2115 boolean_t isipv6 = INP_ISIPV6(inp);
2117 const boolean_t isipv6 = FALSE;
2121 if (tcp_iw_maxsegs < TCP_IW_MAXSEGS_DFLT)
2122 tcp_iw_maxsegs = TCP_IW_MAXSEGS_DFLT;
2123 if (tcp_iw_capsegs < TCP_IW_CAPSEGS_DFLT)
2124 tcp_iw_capsegs = TCP_IW_CAPSEGS_DFLT;
2127 rt = tcp_rtlookup6(&inp->inp_inc);
2129 rt = tcp_rtlookup(&inp->inp_inc);
2131 rt->rt_rmx.rmx_iwmaxsegs < TCP_IW_MAXSEGS_DFLT ||
2132 rt->rt_rmx.rmx_iwcapsegs < TCP_IW_CAPSEGS_DFLT) {
2133 *maxsegs = tcp_iw_maxsegs;
2134 *capsegs = tcp_iw_capsegs;
2137 *maxsegs = rt->rt_rmx.rmx_iwmaxsegs;
2138 *capsegs = rt->rt_rmx.rmx_iwcapsegs;
2142 tcp_initial_window(struct tcpcb *tp)
2144 if (tcp_do_rfc3390) {
2147 * "If the SYN or SYN/ACK is lost, the initial window
2148 * used by a sender after a correctly transmitted SYN
2149 * MUST be one segment consisting of MSS bytes."
2151 * However, we do something a little bit more aggressive
2152 * then RFC3390 here:
2153 * - Only if time spent in the SYN or SYN|ACK retransmition
2154 * >= 3 seconds, the IW is reduced. We do this mainly
2155 * because when RFC3390 is published, the initial RTO is
2156 * still 3 seconds (the threshold we test here), while
2157 * after RFC6298, the initial RTO is 1 second. This
2158 * behaviour probably still falls within the spirit of
2160 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2161 * Mainly to avoid sender and receiver deadlock until
2162 * delayed ACK timer expires. And even RFC2581 does not
2163 * try to reduce IW upon SYN or SYN|ACK retransmition
2167 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2169 if (tp->t_rxtsyn >= TCPTV_RTOBASE3) {
2170 return (2 * tp->t_maxseg);
2172 u_long maxsegs, capsegs;
2174 tcp_rmx_iwsegs(tp, &maxsegs, &capsegs);
2175 return min(maxsegs * tp->t_maxseg,
2176 max(2 * tp->t_maxseg, capsegs * 1460));
2180 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2182 * Mainly to avoid sender and receiver deadlock
2183 * until delayed ACK timer expires.
2185 return (2 * tp->t_maxseg);
2189 #ifdef TCP_SIGNATURE
2191 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2193 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2194 * When called from tcp_input(), we can be sure that th_sum has been
2195 * zeroed out and verified already.
2197 * Return 0 if successful, otherwise return -1.
2199 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2200 * search with the destination IP address, and a 'magic SPI' to be
2201 * determined by the application. This is hardcoded elsewhere to 1179
2202 * right now. Another branch of this code exists which uses the SPD to
2203 * specify per-application flows but it is unstable.
2206 tcpsignature_compute(
2207 struct mbuf *m, /* mbuf chain */
2208 int len, /* length of TCP data */
2209 int optlen, /* length of TCP options */
2210 u_char *buf, /* storage for MD5 digest */
2211 u_int direction) /* direction of flow */
2213 struct ippseudo ippseudo;
2217 struct ipovly *ipovly;
2218 struct secasvar *sav;
2221 struct ip6_hdr *ip6;
2222 struct in6_addr in6;
2228 KASSERT(m != NULL, ("passed NULL mbuf. Game over."));
2229 KASSERT(buf != NULL, ("passed NULL storage pointer for MD5 signature"));
2231 * Extract the destination from the IP header in the mbuf.
2233 ip = mtod(m, struct ip *);
2235 ip6 = NULL; /* Make the compiler happy. */
2238 * Look up an SADB entry which matches the address found in
2241 switch (IP_VHL_V(ip->ip_vhl)) {
2243 sav = key_allocsa(AF_INET, (caddr_t)&ip->ip_src, (caddr_t)&ip->ip_dst,
2244 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2247 case (IPV6_VERSION >> 4):
2248 ip6 = mtod(m, struct ip6_hdr *);
2249 sav = key_allocsa(AF_INET6, (caddr_t)&ip6->ip6_src, (caddr_t)&ip6->ip6_dst,
2250 IPPROTO_TCP, htonl(TCP_SIG_SPI));
2259 kprintf("%s: SADB lookup failed\n", __func__);
2265 * Step 1: Update MD5 hash with IP pseudo-header.
2267 * XXX The ippseudo header MUST be digested in network byte order,
2268 * or else we'll fail the regression test. Assume all fields we've
2269 * been doing arithmetic on have been in host byte order.
2270 * XXX One cannot depend on ipovly->ih_len here. When called from
2271 * tcp_output(), the underlying ip_len member has not yet been set.
2273 switch (IP_VHL_V(ip->ip_vhl)) {
2275 ipovly = (struct ipovly *)ip;
2276 ippseudo.ippseudo_src = ipovly->ih_src;
2277 ippseudo.ippseudo_dst = ipovly->ih_dst;
2278 ippseudo.ippseudo_pad = 0;
2279 ippseudo.ippseudo_p = IPPROTO_TCP;
2280 ippseudo.ippseudo_len = htons(len + sizeof(struct tcphdr) + optlen);
2281 MD5Update(&ctx, (char *)&ippseudo, sizeof(struct ippseudo));
2282 th = (struct tcphdr *)((u_char *)ip + sizeof(struct ip));
2283 doff = sizeof(struct ip) + sizeof(struct tcphdr) + optlen;
2287 * RFC 2385, 2.0 Proposal
2288 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2289 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2290 * extended next header value (to form 32 bits), and 32-bit segment
2292 * Note: Upper-Layer Packet Length comes before Next Header.
2294 case (IPV6_VERSION >> 4):
2296 in6_clearscope(&in6);
2297 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2299 in6_clearscope(&in6);
2300 MD5Update(&ctx, (char *)&in6, sizeof(struct in6_addr));
2301 plen = htonl(len + sizeof(struct tcphdr) + optlen);
2302 MD5Update(&ctx, (char *)&plen, sizeof(uint32_t));
2304 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2305 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2306 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2308 MD5Update(&ctx, (char *)&nhdr, sizeof(uint8_t));
2309 th = (struct tcphdr *)((u_char *)ip6 + sizeof(struct ip6_hdr));
2310 doff = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + optlen;
2319 * Step 2: Update MD5 hash with TCP header, excluding options.
2320 * The TCP checksum must be set to zero.
2322 savecsum = th->th_sum;
2324 MD5Update(&ctx, (char *)th, sizeof(struct tcphdr));
2325 th->th_sum = savecsum;
2327 * Step 3: Update MD5 hash with TCP segment data.
2328 * Use m_apply() to avoid an early m_pullup().
2331 m_apply(m, doff, len, tcpsignature_apply, &ctx);
2333 * Step 4: Update MD5 hash with shared secret.
2335 MD5Update(&ctx, _KEYBUF(sav->key_auth), _KEYLEN(sav->key_auth));
2336 MD5Final(buf, &ctx);
2337 key_sa_recordxfer(sav, m);
2343 tcpsignature_apply(void *fstate, void *data, unsigned int len)
2346 MD5Update((MD5_CTX *)fstate, (unsigned char *)data, len);
2349 #endif /* TCP_SIGNATURE */
2352 tcp_drop_sysctl_dispatch(netmsg_t nmsg)
2354 struct lwkt_msg *lmsg = &nmsg->lmsg;
2355 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2356 struct sockaddr_storage *addrs = lmsg->u.ms_resultp;
2358 struct sockaddr_in *fin, *lin;
2360 struct sockaddr_in6 *fin6, *lin6;
2361 struct in6_addr f6, l6;
2365 switch (addrs[0].ss_family) {
2368 fin6 = (struct sockaddr_in6 *)&addrs[0];
2369 lin6 = (struct sockaddr_in6 *)&addrs[1];
2370 error = in6_embedscope(&f6, fin6, NULL, NULL);
2373 error = in6_embedscope(&l6, lin6, NULL, NULL);
2376 inp = in6_pcblookup_hash(&tcbinfo[mycpuid], &f6,
2377 fin6->sin6_port, &l6, lin6->sin6_port, FALSE, NULL);
2382 fin = (struct sockaddr_in *)&addrs[0];
2383 lin = (struct sockaddr_in *)&addrs[1];
2384 inp = in_pcblookup_hash(&tcbinfo[mycpuid], fin->sin_addr,
2385 fin->sin_port, lin->sin_addr, lin->sin_port, FALSE, NULL);
2390 * Must not reach here, since the address family was
2391 * checked in sysctl handler.
2393 panic("unknown address family %d", addrs[0].ss_family);
2396 struct tcpcb *tp = intotcpcb(inp);
2398 KASSERT((inp->inp_flags & INP_WILDCARD) == 0,
2399 ("in wildcard hash"));
2400 KASSERT(tp != NULL, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2401 KASSERT((tp->t_flags & TF_LISTEN) == 0, ("listen socket"));
2402 tcp_drop(tp, ECONNABORTED);
2410 lwkt_replymsg(lmsg, error);
2414 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS)
2416 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2417 struct sockaddr_storage addrs[2];
2418 struct sockaddr_in *fin, *lin;
2420 struct sockaddr_in6 *fin6, *lin6;
2422 struct netmsg_base nmsg;
2423 struct lwkt_msg *lmsg = &nmsg.lmsg;
2424 struct lwkt_port *port = NULL;
2433 if (req->oldptr != NULL || req->oldlen != 0)
2435 if (req->newptr == NULL)
2437 if (req->newlen < sizeof(addrs))
2439 error = SYSCTL_IN(req, &addrs, sizeof(addrs));
2443 switch (addrs[0].ss_family) {
2446 fin6 = (struct sockaddr_in6 *)&addrs[0];
2447 lin6 = (struct sockaddr_in6 *)&addrs[1];
2448 if (fin6->sin6_len != sizeof(struct sockaddr_in6) ||
2449 lin6->sin6_len != sizeof(struct sockaddr_in6))
2451 if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr) ||
2452 IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr))
2453 return (EADDRNOTAVAIL);
2455 error = sa6_embedscope(fin6, V_ip6_use_defzone);
2458 error = sa6_embedscope(lin6, V_ip6_use_defzone);
2462 port = tcp6_addrport();
2467 fin = (struct sockaddr_in *)&addrs[0];
2468 lin = (struct sockaddr_in *)&addrs[1];
2469 if (fin->sin_len != sizeof(struct sockaddr_in) ||
2470 lin->sin_len != sizeof(struct sockaddr_in))
2472 port = tcp_addrport(fin->sin_addr.s_addr, fin->sin_port,
2473 lin->sin_addr.s_addr, lin->sin_port);
2480 netmsg_init(&nmsg, NULL, &curthread->td_msgport, 0,
2481 tcp_drop_sysctl_dispatch);
2482 lmsg->u.ms_resultp = addrs;
2483 return lwkt_domsg(port, lmsg, 0);
2486 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, drop,
2487 CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP, NULL,
2488 0, sysctl_tcp_drop, "", "Drop TCP connection");
2491 sysctl_tcps_count(SYSCTL_HANDLER_ARGS)
2493 u_long state_count[TCP_NSTATES];
2496 memset(state_count, 0, sizeof(state_count));
2497 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
2500 for (i = 0; i < TCP_NSTATES; ++i)
2501 state_count[i] += tcpstate_count[cpu].tcps_count[i];
2504 return sysctl_handle_opaque(oidp, state_count, sizeof(state_count), req);
2506 SYSCTL_PROC(_net_inet_tcp, OID_AUTO, state_count,
2507 CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0,
2508 sysctl_tcps_count, "LU", "TCP connection counts by state");
2511 tcp_pcbport_create(struct tcpcb *tp)
2515 KASSERT((tp->t_flags & TF_LISTEN) && tp->t_state == TCPS_LISTEN,
2516 ("not a listen tcpcb"));
2518 KASSERT(tp->t_pcbport == NULL, ("tcpcb port cache was created"));
2519 tp->t_pcbport = kmalloc_cachealign(
2520 sizeof(struct tcp_pcbport) * netisr_ncpus, M_PCB, M_WAITOK);
2522 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
2523 struct inpcbport *phd;
2525 phd = &tp->t_pcbport[cpu].t_phd;
2526 LIST_INIT(&phd->phd_pcblist);
2527 /* Though, not used ... */
2528 phd->phd_port = tp->t_inpcb->inp_lport;
2533 tcp_pcbport_merge_oncpu(struct tcpcb *tp)
2535 struct inpcbport *phd;
2539 KASSERT(cpu < netisr_ncpus, ("invalid cpu%d", cpu));
2540 phd = &tp->t_pcbport[cpu].t_phd;
2542 while ((inp = LIST_FIRST(&phd->phd_pcblist)) != NULL) {
2543 KASSERT(inp->inp_phd == phd && inp->inp_porthash == NULL,
2544 ("not on tcpcb port cache"));
2545 LIST_REMOVE(inp, inp_portlist);
2546 in_pcbinsporthash_lport(inp);
2547 KASSERT(inp->inp_phd == tp->t_inpcb->inp_phd &&
2548 inp->inp_porthash == tp->t_inpcb->inp_porthash,
2549 ("tcpcb port cache merge failed"));
2554 tcp_pcbport_destroy(struct tcpcb *tp)
2559 for (cpu = 0; cpu < netisr_ncpus; ++cpu) {
2560 KASSERT(LIST_EMPTY(&tp->t_pcbport[cpu].t_phd.phd_pcblist),
2561 ("tcpcb port cache is not empty"));
2564 kfree(tp->t_pcbport, M_PCB);
2565 tp->t_pcbport = NULL;