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 $
66 #include "opt_compat.h"
68 #include "opt_inet6.h"
69 #include "opt_ipsec.h"
70 #include "opt_tcpdebug.h"
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/callout.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/malloc.h>
78 #include <sys/mpipe.h>
81 #include <sys/domain.h>
85 #include <sys/socket.h>
86 #include <sys/socketops.h>
87 #include <sys/socketvar.h>
88 #include <sys/protosw.h>
89 #include <sys/random.h>
90 #include <sys/in_cksum.h>
93 #include <net/route.h>
95 #include <net/netisr2.h>
98 #include <netinet/in.h>
99 #include <netinet/in_systm.h>
100 #include <netinet/ip.h>
101 #include <netinet/ip6.h>
102 #include <netinet/in_pcb.h>
103 #include <netinet6/in6_pcb.h>
104 #include <netinet/in_var.h>
105 #include <netinet/ip_var.h>
106 #include <netinet6/ip6_var.h>
107 #include <netinet/ip_icmp.h>
109 #include <netinet/icmp6.h>
111 #include <netinet/tcp.h>
112 #include <netinet/tcp_fsm.h>
113 #include <netinet/tcp_seq.h>
114 #include <netinet/tcp_timer.h>
115 #include <netinet/tcp_timer2.h>
116 #include <netinet/tcp_var.h>
117 #include <netinet6/tcp6_var.h>
118 #include <netinet/tcpip.h>
120 #include <netinet/tcp_debug.h>
122 #include <netinet6/ip6protosw.h>
125 #include <netinet6/ipsec.h>
126 #include <netproto/key/key.h>
128 #include <netinet6/ipsec6.h>
133 #include <netproto/ipsec/ipsec.h>
135 #include <netproto/ipsec/ipsec6.h>
141 #include <machine/smp.h>
143 #include <sys/msgport2.h>
144 #include <sys/mplock2.h>
145 #include <net/netmsg2.h>
147 #if !defined(KTR_TCP)
148 #define KTR_TCP KTR_ALL
151 KTR_INFO_MASTER(tcp);
152 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
153 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
154 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
155 #define logtcp(name) KTR_LOG(tcp_ ## name)
158 #define TCP_IW_MAXSEGS_DFLT 4
159 #define TCP_IW_CAPSEGS_DFLT 4
161 struct inpcbinfo tcbinfo
[MAXCPU
];
162 struct tcpcbackqhead tcpcbackq
[MAXCPU
];
164 int tcp_mssdflt
= TCP_MSS
;
165 SYSCTL_INT(_net_inet_tcp
, TCPCTL_MSSDFLT
, mssdflt
, CTLFLAG_RW
,
166 &tcp_mssdflt
, 0, "Default TCP Maximum Segment Size");
169 int tcp_v6mssdflt
= TCP6_MSS
;
170 SYSCTL_INT(_net_inet_tcp
, TCPCTL_V6MSSDFLT
, v6mssdflt
, CTLFLAG_RW
,
171 &tcp_v6mssdflt
, 0, "Default TCP Maximum Segment Size for IPv6");
175 * Minimum MSS we accept and use. This prevents DoS attacks where
176 * we are forced to a ridiculous low MSS like 20 and send hundreds
177 * of packets instead of one. The effect scales with the available
178 * bandwidth and quickly saturates the CPU and network interface
179 * with packet generation and sending. Set to zero to disable MINMSS
180 * checking. This setting prevents us from sending too small packets.
182 int tcp_minmss
= TCP_MINMSS
;
183 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, minmss
, CTLFLAG_RW
,
184 &tcp_minmss
, 0, "Minmum TCP Maximum Segment Size");
187 static int tcp_rttdflt
= TCPTV_SRTTDFLT
/ PR_SLOWHZ
;
188 SYSCTL_INT(_net_inet_tcp
, TCPCTL_RTTDFLT
, rttdflt
, CTLFLAG_RW
,
189 &tcp_rttdflt
, 0, "Default maximum TCP Round Trip Time");
192 int tcp_do_rfc1323
= 1;
193 SYSCTL_INT(_net_inet_tcp
, TCPCTL_DO_RFC1323
, rfc1323
, CTLFLAG_RW
,
194 &tcp_do_rfc1323
, 0, "Enable rfc1323 (high performance TCP) extensions");
196 static int tcp_tcbhashsize
= 0;
197 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, tcbhashsize
, CTLFLAG_RD
,
198 &tcp_tcbhashsize
, 0, "Size of TCP control block hashtable");
200 static int do_tcpdrain
= 1;
201 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, do_tcpdrain
, CTLFLAG_RW
, &do_tcpdrain
, 0,
202 "Enable tcp_drain routine for extra help when low on mbufs");
204 static int icmp_may_rst
= 1;
205 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, icmp_may_rst
, CTLFLAG_RW
, &icmp_may_rst
, 0,
206 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
208 static int tcp_isn_reseed_interval
= 0;
209 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, isn_reseed_interval
, CTLFLAG_RW
,
210 &tcp_isn_reseed_interval
, 0, "Seconds between reseeding of ISN secret");
213 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
214 * by default, but with generous values which should allow maximal
215 * bandwidth. In particular, the slop defaults to 50 (5 packets).
217 * The reason for doing this is that the limiter is the only mechanism we
218 * have which seems to do a really good job preventing receiver RX rings
219 * on network interfaces from getting blown out. Even though GigE/10GigE
220 * is supposed to flow control it looks like either it doesn't actually
221 * do it or Open Source drivers do not properly enable it.
223 * People using the limiter to reduce bottlenecks on slower WAN connections
224 * should set the slop to 20 (2 packets).
226 static int tcp_inflight_enable
= 1;
227 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_enable
, CTLFLAG_RW
,
228 &tcp_inflight_enable
, 0, "Enable automatic TCP inflight data limiting");
230 static int tcp_inflight_debug
= 0;
231 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_debug
, CTLFLAG_RW
,
232 &tcp_inflight_debug
, 0, "Debug TCP inflight calculations");
235 * NOTE: tcp_inflight_start is essentially the starting receive window
236 * for a connection. If set too low then fetches over tcp
237 * connections will take noticably longer to ramp-up over
238 * high-latency connections. 6144 is too low for a default,
239 * use something more reasonable.
241 static int tcp_inflight_start
= 33792;
242 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_start
, CTLFLAG_RW
,
243 &tcp_inflight_start
, 0, "Start value for TCP inflight window");
245 static int tcp_inflight_min
= 6144;
246 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_min
, CTLFLAG_RW
,
247 &tcp_inflight_min
, 0, "Lower bound for TCP inflight window");
249 static int tcp_inflight_max
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
250 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_max
, CTLFLAG_RW
,
251 &tcp_inflight_max
, 0, "Upper bound for TCP inflight window");
253 static int tcp_inflight_stab
= 50;
254 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_stab
, CTLFLAG_RW
,
255 &tcp_inflight_stab
, 0, "Fudge bw 1/10% (50=5%)");
257 static int tcp_inflight_adjrtt
= 2;
258 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_adjrtt
, CTLFLAG_RW
,
259 &tcp_inflight_adjrtt
, 0, "Slop for rtt 1/(hz*32)");
261 static int tcp_do_rfc3390
= 1;
262 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, rfc3390
, CTLFLAG_RW
,
264 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
266 static u_long tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
267 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwmaxsegs
, CTLFLAG_RW
,
268 &tcp_iw_maxsegs
, 0, "TCP IW segments max");
270 static u_long tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
271 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwcapsegs
, CTLFLAG_RW
,
272 &tcp_iw_capsegs
, 0, "TCP IW segments");
274 int tcp_low_rtobase
= 1;
275 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, low_rtobase
, CTLFLAG_RW
,
276 &tcp_low_rtobase
, 0, "Lowering the Initial RTO (RFC 6298)");
278 static int tcp_do_ncr
= 1;
279 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr
, CTLFLAG_RW
,
280 &tcp_do_ncr
, 0, "Non-Congestion Robustness (RFC 4653)");
282 int tcp_ncr_rxtthresh_max
= 16;
283 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_rxtthresh_max
, CTLFLAG_RW
,
284 &tcp_ncr_rxtthresh_max
, 0,
285 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
287 static MALLOC_DEFINE(M_TCPTEMP
, "tcptemp", "TCP Templates for Keepalives");
288 static struct malloc_pipe tcptemp_mpipe
;
290 static void tcp_willblock(void);
291 static void tcp_notify (struct inpcb
*, int);
293 struct tcp_stats tcpstats_percpu
[MAXCPU
] __cachealign
;
294 struct tcp_state_count tcpstate_count
[MAXCPU
] __cachealign
;
296 static struct netmsg_base tcp_drain_netmsg
[MAXCPU
];
297 static void tcp_drain_dispatch(netmsg_t nmsg
);
300 sysctl_tcpstats(SYSCTL_HANDLER_ARGS
)
304 for (cpu
= 0; cpu
< ncpus2
; ++cpu
) {
305 if ((error
= SYSCTL_OUT(req
, &tcpstats_percpu
[cpu
],
306 sizeof(struct tcp_stats
))))
308 if ((error
= SYSCTL_IN(req
, &tcpstats_percpu
[cpu
],
309 sizeof(struct tcp_stats
))))
315 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_STATS
, stats
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
316 0, 0, sysctl_tcpstats
, "S,tcp_stats", "TCP statistics");
319 * Target size of TCP PCB hash tables. Must be a power of two.
321 * Note that this can be overridden by the kernel environment
322 * variable net.inet.tcp.tcbhashsize
325 #define TCBHASHSIZE 512
329 * This is the actual shape of what we allocate using the zone
330 * allocator. Doing it this way allows us to protect both structures
331 * using the same generation count, and also eliminates the overhead
332 * of allocating tcpcbs separately. By hiding the structure here,
333 * we avoid changing most of the rest of the code (although it needs
334 * to be changed, eventually, for greater efficiency).
337 #define ALIGNM1 (ALIGNMENT - 1)
341 char align
[(sizeof(struct inpcb
) + ALIGNM1
) & ~ALIGNM1
];
344 struct tcp_callout inp_tp_rexmt
;
345 struct tcp_callout inp_tp_persist
;
346 struct tcp_callout inp_tp_keep
;
347 struct tcp_callout inp_tp_2msl
;
348 struct tcp_callout inp_tp_delack
;
349 struct netmsg_tcp_timer inp_tp_timermsg
;
350 struct netmsg_base inp_tp_sndmore
;
361 struct inpcbportinfo
*portinfo
;
362 struct inpcbinfo
*ticb
;
363 int hashsize
= TCBHASHSIZE
;
367 * note: tcptemp is used for keepalives, and it is ok for an
368 * allocation to fail so do not specify MPF_INT.
370 mpipe_init(&tcptemp_mpipe
, M_TCPTEMP
, sizeof(struct tcptemp
),
371 25, -1, 0, NULL
, NULL
, NULL
);
373 tcp_delacktime
= TCPTV_DELACK
;
374 tcp_keepinit
= TCPTV_KEEP_INIT
;
375 tcp_keepidle
= TCPTV_KEEP_IDLE
;
376 tcp_keepintvl
= TCPTV_KEEPINTVL
;
377 tcp_maxpersistidle
= TCPTV_KEEP_IDLE
;
379 tcp_rexmit_min
= TCPTV_MIN
;
380 tcp_rexmit_slop
= TCPTV_CPU_VAR
;
382 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize
);
383 if (!powerof2(hashsize
)) {
384 kprintf("WARNING: TCB hash size not a power of 2\n");
385 hashsize
= 512; /* safe default */
387 tcp_tcbhashsize
= hashsize
;
389 portinfo
= kmalloc_cachealign(sizeof(*portinfo
) * ncpus2
, M_PCB
,
392 for (cpu
= 0; cpu
< ncpus2
; cpu
++) {
393 ticb
= &tcbinfo
[cpu
];
394 in_pcbinfo_init(ticb
, cpu
, FALSE
);
395 ticb
->hashbase
= hashinit(hashsize
, M_PCB
,
397 in_pcbportinfo_init(&portinfo
[cpu
], hashsize
, cpu
);
398 ticb
->portinfo
= portinfo
;
399 ticb
->portinfo_mask
= ncpus2_mask
;
400 ticb
->wildcardhashbase
= hashinit(hashsize
, M_PCB
,
401 &ticb
->wildcardhashmask
);
402 ticb
->localgrphashbase
= hashinit(hashsize
, M_PCB
,
403 &ticb
->localgrphashmask
);
404 ticb
->ipi_size
= sizeof(struct inp_tp
);
405 TAILQ_INIT(&tcpcbackq
[cpu
]);
408 tcp_reass_maxseg
= nmbclusters
/ 16;
409 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg
);
412 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
414 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
416 if (max_protohdr
< TCP_MINPROTOHDR
)
417 max_protohdr
= TCP_MINPROTOHDR
;
418 if (max_linkhdr
+ TCP_MINPROTOHDR
> MHLEN
)
420 #undef TCP_MINPROTOHDR
423 * Initialize TCP statistics counters for each CPU.
425 for (cpu
= 0; cpu
< ncpus2
; ++cpu
)
426 bzero(&tcpstats_percpu
[cpu
], sizeof(struct tcp_stats
));
429 * Initialize netmsgs for TCP drain
431 for (cpu
= 0; cpu
< ncpus2
; ++cpu
) {
432 netmsg_init(&tcp_drain_netmsg
[cpu
], NULL
, &netisr_adone_rport
,
433 MSGF_PRIORITY
, tcp_drain_dispatch
);
437 netisr_register_rollup(tcp_willblock
, NETISR_ROLLUP_PRIO_TCP
);
444 int cpu
= mycpu
->gd_cpuid
;
446 while ((tp
= TAILQ_FIRST(&tcpcbackq
[cpu
])) != NULL
) {
447 KKASSERT(tp
->t_flags
& TF_ONOUTPUTQ
);
448 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
449 TAILQ_REMOVE(&tcpcbackq
[cpu
], tp
, t_outputq
);
455 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
456 * tcp_template used to store this data in mbufs, but we now recopy it out
457 * of the tcpcb each time to conserve mbufs.
460 tcp_fillheaders(struct tcpcb
*tp
, void *ip_ptr
, void *tcp_ptr
, boolean_t tso
)
462 struct inpcb
*inp
= tp
->t_inpcb
;
463 struct tcphdr
*tcp_hdr
= (struct tcphdr
*)tcp_ptr
;
466 if (INP_ISIPV6(inp
)) {
469 ip6
= (struct ip6_hdr
*)ip_ptr
;
470 ip6
->ip6_flow
= (ip6
->ip6_flow
& ~IPV6_FLOWINFO_MASK
) |
471 (inp
->in6p_flowinfo
& IPV6_FLOWINFO_MASK
);
472 ip6
->ip6_vfc
= (ip6
->ip6_vfc
& ~IPV6_VERSION_MASK
) |
473 (IPV6_VERSION
& IPV6_VERSION_MASK
);
474 ip6
->ip6_nxt
= IPPROTO_TCP
;
475 ip6
->ip6_plen
= sizeof(struct tcphdr
);
476 ip6
->ip6_src
= inp
->in6p_laddr
;
477 ip6
->ip6_dst
= inp
->in6p_faddr
;
482 struct ip
*ip
= (struct ip
*) ip_ptr
;
485 ip
->ip_vhl
= IP_VHL_BORING
;
492 ip
->ip_p
= IPPROTO_TCP
;
493 ip
->ip_src
= inp
->inp_laddr
;
494 ip
->ip_dst
= inp
->inp_faddr
;
497 plen
= htons(IPPROTO_TCP
);
499 plen
= htons(sizeof(struct tcphdr
) + IPPROTO_TCP
);
500 tcp_hdr
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
,
501 ip
->ip_dst
.s_addr
, plen
);
504 tcp_hdr
->th_sport
= inp
->inp_lport
;
505 tcp_hdr
->th_dport
= inp
->inp_fport
;
510 tcp_hdr
->th_flags
= 0;
516 * Create template to be used to send tcp packets on a connection.
517 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
518 * use for this function is in keepalives, which use tcp_respond.
521 tcp_maketemplate(struct tcpcb
*tp
)
525 if ((tmp
= mpipe_alloc_nowait(&tcptemp_mpipe
)) == NULL
)
527 tcp_fillheaders(tp
, &tmp
->tt_ipgen
, &tmp
->tt_t
, FALSE
);
532 tcp_freetemplate(struct tcptemp
*tmp
)
534 mpipe_free(&tcptemp_mpipe
, tmp
);
538 * Send a single message to the TCP at address specified by
539 * the given TCP/IP header. If m == NULL, then we make a copy
540 * of the tcpiphdr at ti and send directly to the addressed host.
541 * This is used to force keep alive messages out using the TCP
542 * template for a connection. If flags are given then we send
543 * a message back to the TCP which originated the * segment ti,
544 * and discard the mbuf containing it and any other attached mbufs.
546 * In any case the ack and sequence number of the transmitted
547 * segment are as specified by the parameters.
549 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
552 tcp_respond(struct tcpcb
*tp
, void *ipgen
, struct tcphdr
*th
, struct mbuf
*m
,
553 tcp_seq ack
, tcp_seq seq
, int flags
)
557 struct route
*ro
= NULL
;
559 struct ip
*ip
= ipgen
;
562 struct route_in6
*ro6
= NULL
;
563 struct route_in6 sro6
;
564 struct ip6_hdr
*ip6
= ipgen
;
565 boolean_t use_tmpro
= TRUE
;
567 boolean_t isipv6
= (IP_VHL_V(ip
->ip_vhl
) == 6);
569 const boolean_t isipv6
= FALSE
;
573 if (!(flags
& TH_RST
)) {
574 win
= ssb_space(&tp
->t_inpcb
->inp_socket
->so_rcv
);
577 if (win
> (long)TCP_MAXWIN
<< tp
->rcv_scale
)
578 win
= (long)TCP_MAXWIN
<< tp
->rcv_scale
;
581 * Don't use the route cache of a listen socket,
582 * it is not MPSAFE; use temporary route cache.
584 if (tp
->t_state
!= TCPS_LISTEN
) {
586 ro6
= &tp
->t_inpcb
->in6p_route
;
588 ro
= &tp
->t_inpcb
->inp_route
;
595 bzero(ro6
, sizeof *ro6
);
598 bzero(ro
, sizeof *ro
);
602 m
= m_gethdr(M_NOWAIT
, MT_HEADER
);
606 m
->m_data
+= max_linkhdr
;
608 bcopy(ip6
, mtod(m
, caddr_t
), sizeof(struct ip6_hdr
));
609 ip6
= mtod(m
, struct ip6_hdr
*);
610 nth
= (struct tcphdr
*)(ip6
+ 1);
612 bcopy(ip
, mtod(m
, caddr_t
), sizeof(struct ip
));
613 ip
= mtod(m
, struct ip
*);
614 nth
= (struct tcphdr
*)(ip
+ 1);
616 bcopy(th
, nth
, sizeof(struct tcphdr
));
621 m
->m_data
= (caddr_t
)ipgen
;
622 /* m_len is set later */
624 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
626 xchg(ip6
->ip6_dst
, ip6
->ip6_src
, struct in6_addr
);
627 nth
= (struct tcphdr
*)(ip6
+ 1);
629 xchg(ip
->ip_dst
.s_addr
, ip
->ip_src
.s_addr
, n_long
);
630 nth
= (struct tcphdr
*)(ip
+ 1);
634 * this is usually a case when an extension header
635 * exists between the IPv6 header and the
638 nth
->th_sport
= th
->th_sport
;
639 nth
->th_dport
= th
->th_dport
;
641 xchg(nth
->th_dport
, nth
->th_sport
, n_short
);
646 ip6
->ip6_vfc
= IPV6_VERSION
;
647 ip6
->ip6_nxt
= IPPROTO_TCP
;
648 ip6
->ip6_plen
= htons((u_short
)(sizeof(struct tcphdr
) + tlen
));
649 tlen
+= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
);
651 tlen
+= sizeof(struct tcpiphdr
);
653 ip
->ip_ttl
= ip_defttl
;
656 m
->m_pkthdr
.len
= tlen
;
657 m
->m_pkthdr
.rcvif
= NULL
;
658 nth
->th_seq
= htonl(seq
);
659 nth
->th_ack
= htonl(ack
);
661 nth
->th_off
= sizeof(struct tcphdr
) >> 2;
662 nth
->th_flags
= flags
;
664 nth
->th_win
= htons((u_short
) (win
>> tp
->rcv_scale
));
666 nth
->th_win
= htons((u_short
)win
);
670 nth
->th_sum
= in6_cksum(m
, IPPROTO_TCP
,
671 sizeof(struct ip6_hdr
),
672 tlen
- sizeof(struct ip6_hdr
));
673 ip6
->ip6_hlim
= in6_selecthlim(tp
? tp
->t_inpcb
: NULL
,
674 (ro6
&& ro6
->ro_rt
) ?
675 ro6
->ro_rt
->rt_ifp
: NULL
);
677 nth
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
, ip
->ip_dst
.s_addr
,
678 htons((u_short
)(tlen
- sizeof(struct ip
) + ip
->ip_p
)));
679 m
->m_pkthdr
.csum_flags
= CSUM_TCP
;
680 m
->m_pkthdr
.csum_data
= offsetof(struct tcphdr
, th_sum
);
681 m
->m_pkthdr
.csum_thlen
= sizeof(struct tcphdr
);
684 if (tp
== NULL
|| (tp
->t_inpcb
->inp_socket
->so_options
& SO_DEBUG
))
685 tcp_trace(TA_OUTPUT
, 0, tp
, mtod(m
, void *), th
, 0);
688 ip6_output(m
, NULL
, ro6
, ipflags
, NULL
, NULL
,
689 tp
? tp
->t_inpcb
: NULL
);
690 if ((ro6
== &sro6
) && (ro6
->ro_rt
!= NULL
)) {
695 ipflags
|= IP_DEBUGROUTE
;
696 ip_output(m
, NULL
, ro
, ipflags
, NULL
, tp
? tp
->t_inpcb
: NULL
);
697 if ((ro
== &sro
) && (ro
->ro_rt
!= NULL
)) {
705 * Create a new TCP control block, making an
706 * empty reassembly queue and hooking it to the argument
707 * protocol control block. The `inp' parameter must have
708 * come from the zone allocator set up in tcp_init().
711 tcp_newtcpcb(struct inpcb
*inp
)
716 boolean_t isipv6
= INP_ISIPV6(inp
);
718 const boolean_t isipv6
= FALSE
;
721 it
= (struct inp_tp
*)inp
;
723 bzero(tp
, sizeof(struct tcpcb
));
724 TAILQ_INIT(&tp
->t_segq
);
725 tp
->t_maxseg
= tp
->t_maxopd
= isipv6
? tcp_v6mssdflt
: tcp_mssdflt
;
726 tp
->t_rxtthresh
= tcprexmtthresh
;
728 /* Set up our timeouts. */
729 tp
->tt_rexmt
= &it
->inp_tp_rexmt
;
730 tp
->tt_persist
= &it
->inp_tp_persist
;
731 tp
->tt_keep
= &it
->inp_tp_keep
;
732 tp
->tt_2msl
= &it
->inp_tp_2msl
;
733 tp
->tt_delack
= &it
->inp_tp_delack
;
737 * Zero out timer message. We don't create it here,
738 * since the current CPU may not be the owner of this
741 tp
->tt_msg
= &it
->inp_tp_timermsg
;
742 bzero(tp
->tt_msg
, sizeof(*tp
->tt_msg
));
744 tp
->t_keepinit
= tcp_keepinit
;
745 tp
->t_keepidle
= tcp_keepidle
;
746 tp
->t_keepintvl
= tcp_keepintvl
;
747 tp
->t_keepcnt
= tcp_keepcnt
;
748 tp
->t_maxidle
= tp
->t_keepintvl
* tp
->t_keepcnt
;
751 tp
->t_flags
|= TF_NCR
;
753 tp
->t_flags
|= (TF_REQ_SCALE
| TF_REQ_TSTMP
);
755 tp
->t_inpcb
= inp
; /* XXX */
758 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
759 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
760 * reasonable initial retransmit time.
762 tp
->t_srtt
= TCPTV_SRTTBASE
;
764 ((TCPTV_RTOBASE
- TCPTV_SRTTBASE
) << TCP_RTTVAR_SHIFT
) / 4;
765 tp
->t_rttmin
= tcp_rexmit_min
;
766 tp
->t_rxtcur
= TCPTV_RTOBASE
;
767 tp
->snd_cwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
768 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
769 tp
->snd_ssthresh
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
770 tp
->snd_last
= ticks
;
771 tp
->t_rcvtime
= ticks
;
773 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
774 * because the socket may be bound to an IPv6 wildcard address,
775 * which may match an IPv4-mapped IPv6 address.
777 inp
->inp_ip_ttl
= ip_defttl
;
779 tcp_sack_tcpcb_init(tp
);
781 tp
->tt_sndmore
= &it
->inp_tp_sndmore
;
786 * Drop a TCP connection, reporting the specified error.
787 * If connection is synchronized, then send a RST to peer.
790 tcp_drop(struct tcpcb
*tp
, int error
)
792 struct socket
*so
= tp
->t_inpcb
->inp_socket
;
794 if (TCPS_HAVERCVDSYN(tp
->t_state
)) {
795 TCP_STATE_CHANGE(tp
, TCPS_CLOSED
);
797 tcpstat
.tcps_drops
++;
799 tcpstat
.tcps_conndrops
++;
800 if (error
== ETIMEDOUT
&& tp
->t_softerror
)
801 error
= tp
->t_softerror
;
802 so
->so_error
= error
;
803 return (tcp_close(tp
));
806 struct netmsg_listen_detach
{
807 struct netmsg_base base
;
809 struct tcpcb
*nm_tp_inh
;
813 tcp_listen_detach_handler(netmsg_t msg
)
815 struct netmsg_listen_detach
*nmsg
= (struct netmsg_listen_detach
*)msg
;
816 struct tcpcb
*tp
= nmsg
->nm_tp
;
817 int cpu
= mycpuid
, nextcpu
;
819 if (tp
->t_flags
& TF_LISTEN
) {
820 syncache_destroy(tp
, nmsg
->nm_tp_inh
);
821 tcp_pcbport_merge_oncpu(tp
);
824 in_pcbremwildcardhash_oncpu(tp
->t_inpcb
, &tcbinfo
[cpu
]);
827 if (nextcpu
< ncpus2
)
828 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nmsg
->base
.lmsg
);
830 lwkt_replymsg(&nmsg
->base
.lmsg
, 0);
834 * Close a TCP control block:
835 * discard all space held by the tcp
836 * discard internet protocol block
837 * wake up any sleepers
840 tcp_close(struct tcpcb
*tp
)
843 struct inpcb
*inp
= tp
->t_inpcb
;
844 struct inpcb
*inp_inh
= NULL
;
845 struct tcpcb
*tp_inh
= NULL
;
846 struct socket
*so
= inp
->inp_socket
;
848 boolean_t dosavessthresh
;
850 boolean_t isipv6
= INP_ISIPV6(inp
);
852 const boolean_t isipv6
= FALSE
;
855 if (tp
->t_flags
& TF_LISTEN
) {
857 * Pending socket/syncache inheritance
859 * If this is a listen(2) socket, find another listen(2)
860 * socket in the same local group, which could inherit
861 * the syncache and sockets pending on the completion
862 * and incompletion queues.
865 * Currently the inheritance could only happen on the
866 * listen(2) sockets w/ SO_REUSEPORT set.
869 inp_inh
= in_pcblocalgroup_last(&tcbinfo
[0], inp
);
871 tp_inh
= intotcpcb(inp_inh
);
875 * INP_WILDCARD indicates that listen(2) has been called on
876 * this socket. This implies:
877 * - A wildcard inp's hash is replicated for each protocol thread.
878 * - Syncache for this inp grows independently in each protocol
880 * - There is more than one cpu
882 * We have to chain a message to the rest of the protocol threads
883 * to cleanup the wildcard hash and the syncache. The cleanup
884 * in the current protocol thread is defered till the end of this
885 * function (syncache_destroy and in_pcbdetach).
888 * After cleanup the inp's hash and syncache entries, this inp will
889 * no longer be available to the rest of the protocol threads, so we
890 * are safe to whack the inp in the following code.
892 if ((inp
->inp_flags
& INP_WILDCARD
) && ncpus2
> 1) {
893 struct netmsg_listen_detach nmsg
;
895 KKASSERT(so
->so_port
== netisr_cpuport(0));
897 KKASSERT(inp
->inp_pcbinfo
== &tcbinfo
[0]);
899 netmsg_init(&nmsg
.base
, NULL
, &curthread
->td_msgport
,
900 MSGF_PRIORITY
, tcp_listen_detach_handler
);
902 nmsg
.nm_tp_inh
= tp_inh
;
903 lwkt_domsg(netisr_cpuport(1), &nmsg
.base
.lmsg
, 0);
909 * Make sure that all of our timers are stopped before we
910 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
911 * timers are never used. If timer message is never created
912 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
914 if (tp
->tt_msg
!= NULL
&& tp
->tt_msg
->tt_tcb
!= NULL
) {
915 tcp_callout_stop(tp
, tp
->tt_rexmt
);
916 tcp_callout_stop(tp
, tp
->tt_persist
);
917 tcp_callout_stop(tp
, tp
->tt_keep
);
918 tcp_callout_stop(tp
, tp
->tt_2msl
);
919 tcp_callout_stop(tp
, tp
->tt_delack
);
922 if (tp
->t_flags
& TF_ONOUTPUTQ
) {
923 KKASSERT(tp
->tt_cpu
== mycpu
->gd_cpuid
);
924 TAILQ_REMOVE(&tcpcbackq
[tp
->tt_cpu
], tp
, t_outputq
);
925 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
929 * If we got enough samples through the srtt filter,
930 * save the rtt and rttvar in the routing entry.
931 * 'Enough' is arbitrarily defined as the 16 samples.
932 * 16 samples is enough for the srtt filter to converge
933 * to within 5% of the correct value; fewer samples and
934 * we could save a very bogus rtt.
936 * Don't update the default route's characteristics and don't
937 * update anything that the user "locked".
939 if (tp
->t_rttupdated
>= 16) {
943 struct sockaddr_in6
*sin6
;
945 if ((rt
= inp
->in6p_route
.ro_rt
) == NULL
)
947 sin6
= (struct sockaddr_in6
*)rt_key(rt
);
948 if (IN6_IS_ADDR_UNSPECIFIED(&sin6
->sin6_addr
))
951 if ((rt
= inp
->inp_route
.ro_rt
) == NULL
||
952 ((struct sockaddr_in
*)rt_key(rt
))->
953 sin_addr
.s_addr
== INADDR_ANY
)
956 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTT
)) {
957 i
= tp
->t_srtt
* (RTM_RTTUNIT
/ (hz
* TCP_RTT_SCALE
));
958 if (rt
->rt_rmx
.rmx_rtt
&& i
)
960 * filter this update to half the old & half
961 * the new values, converting scale.
962 * See route.h and tcp_var.h for a
963 * description of the scaling constants.
966 (rt
->rt_rmx
.rmx_rtt
+ i
) / 2;
968 rt
->rt_rmx
.rmx_rtt
= i
;
969 tcpstat
.tcps_cachedrtt
++;
971 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTTVAR
)) {
973 (RTM_RTTUNIT
/ (hz
* TCP_RTTVAR_SCALE
));
974 if (rt
->rt_rmx
.rmx_rttvar
&& i
)
975 rt
->rt_rmx
.rmx_rttvar
=
976 (rt
->rt_rmx
.rmx_rttvar
+ i
) / 2;
978 rt
->rt_rmx
.rmx_rttvar
= i
;
979 tcpstat
.tcps_cachedrttvar
++;
982 * The old comment here said:
983 * update the pipelimit (ssthresh) if it has been updated
984 * already or if a pipesize was specified & the threshhold
985 * got below half the pipesize. I.e., wait for bad news
986 * before we start updating, then update on both good
989 * But we want to save the ssthresh even if no pipesize is
990 * specified explicitly in the route, because such
991 * connections still have an implicit pipesize specified
992 * by the global tcp_sendspace. In the absence of a reliable
993 * way to calculate the pipesize, it will have to do.
995 i
= tp
->snd_ssthresh
;
996 if (rt
->rt_rmx
.rmx_sendpipe
!= 0)
997 dosavessthresh
= (i
< rt
->rt_rmx
.rmx_sendpipe
/2);
999 dosavessthresh
= (i
< so
->so_snd
.ssb_hiwat
/2);
1000 if (dosavessthresh
||
1001 (!(rt
->rt_rmx
.rmx_locks
& RTV_SSTHRESH
) && (i
!= 0) &&
1002 (rt
->rt_rmx
.rmx_ssthresh
!= 0))) {
1004 * convert the limit from user data bytes to
1005 * packets then to packet data bytes.
1007 i
= (i
+ tp
->t_maxseg
/ 2) / tp
->t_maxseg
;
1012 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1013 sizeof(struct tcpiphdr
));
1014 if (rt
->rt_rmx
.rmx_ssthresh
)
1015 rt
->rt_rmx
.rmx_ssthresh
=
1016 (rt
->rt_rmx
.rmx_ssthresh
+ i
) / 2;
1018 rt
->rt_rmx
.rmx_ssthresh
= i
;
1019 tcpstat
.tcps_cachedssthresh
++;
1024 /* free the reassembly queue, if any */
1025 while((q
= TAILQ_FIRST(&tp
->t_segq
)) != NULL
) {
1026 TAILQ_REMOVE(&tp
->t_segq
, q
, tqe_q
);
1029 atomic_add_int(&tcp_reass_qsize
, -1);
1031 /* throw away SACK blocks in scoreboard*/
1032 if (TCP_DO_SACK(tp
))
1033 tcp_sack_destroy(&tp
->scb
);
1035 inp
->inp_ppcb
= NULL
;
1036 soisdisconnected(so
);
1037 /* note: pcb detached later on */
1039 tcp_destroy_timermsg(tp
);
1040 tcp_output_cancel(tp
);
1042 if (tp
->t_flags
& TF_LISTEN
) {
1043 syncache_destroy(tp
, tp_inh
);
1044 tcp_pcbport_merge_oncpu(tp
);
1045 tcp_pcbport_destroy(tp
);
1046 if (inp_inh
!= NULL
&& inp_inh
->inp_socket
!= NULL
) {
1048 * Pending sockets inheritance only needs
1049 * to be done once in the current thread,
1052 soinherit(so
, inp_inh
->inp_socket
);
1055 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache is not destroyed"));
1057 so_async_rcvd_drop(so
);
1058 /* Drop the reference for the asynchronized pru_rcvd */
1063 * - Remove self from listen tcpcb per-cpu port cache _before_
1065 * - pcbdetach removes any wildcard hash entry on the current CPU.
1067 tcp_pcbport_remove(inp
);
1075 tcpstat
.tcps_closed
++;
1079 static __inline
void
1080 tcp_drain_oncpu(struct inpcbinfo
*pcbinfo
)
1082 struct inpcbhead
*head
= &pcbinfo
->pcblisthead
;
1086 * Since we run in netisr, it is MP safe, even if
1087 * we block during the inpcb list iteration, i.e.
1088 * we don't need to use inpcb marker here.
1090 ASSERT_IN_NETISR(pcbinfo
->cpu
);
1092 LIST_FOREACH(inpb
, head
, inp_list
) {
1094 struct tseg_qent
*te
;
1096 if (inpb
->inp_flags
& INP_PLACEMARKER
)
1099 tcpb
= intotcpcb(inpb
);
1100 KASSERT(tcpb
!= NULL
, ("tcp_drain_oncpu: tcpb is NULL"));
1102 if ((te
= TAILQ_FIRST(&tcpb
->t_segq
)) != NULL
) {
1103 TAILQ_REMOVE(&tcpb
->t_segq
, te
, tqe_q
);
1104 if (te
->tqe_th
->th_flags
& TH_FIN
)
1105 tcpb
->t_flags
&= ~TF_QUEDFIN
;
1108 atomic_add_int(&tcp_reass_qsize
, -1);
1115 tcp_drain_dispatch(netmsg_t nmsg
)
1118 lwkt_replymsg(&nmsg
->lmsg
, 0); /* reply ASAP */
1121 tcp_drain_oncpu(&tcbinfo
[mycpuid
]);
1125 tcp_drain_ipi(void *arg __unused
)
1128 struct lwkt_msg
*msg
= &tcp_drain_netmsg
[cpu
].lmsg
;
1131 if (msg
->ms_flags
& MSGF_DONE
)
1132 lwkt_sendmsg_oncpu(netisr_cpuport(cpu
), msg
);
1145 * Walk the tcpbs, if existing, and flush the reassembly queue,
1146 * if there is one...
1147 * XXX: The "Net/3" implementation doesn't imply that the TCP
1148 * reassembly queue should be flushed, but in a situation
1149 * where we're really low on mbufs, this is potentially
1151 * YYY: We may consider run tcp_drain_oncpu directly here,
1152 * however, that will require M_WAITOK memory allocation
1153 * for the inpcb marker.
1155 CPUMASK_ASSBMASK(mask
, ncpus2
);
1156 CPUMASK_ANDMASK(mask
, smp_active_mask
);
1157 if (CPUMASK_TESTNZERO(mask
))
1158 lwkt_send_ipiq_mask(mask
, tcp_drain_ipi
, NULL
);
1162 * Notify a tcp user of an asynchronous error;
1163 * store error as soft error, but wake up user
1164 * (for now, won't do anything until can select for soft error).
1166 * Do not wake up user since there currently is no mechanism for
1167 * reporting soft errors (yet - a kqueue filter may be added).
1170 tcp_notify(struct inpcb
*inp
, int error
)
1172 struct tcpcb
*tp
= intotcpcb(inp
);
1175 * Ignore some errors if we are hooked up.
1176 * If connection hasn't completed, has retransmitted several times,
1177 * and receives a second error, give up now. This is better
1178 * than waiting a long time to establish a connection that
1179 * can never complete.
1181 if (tp
->t_state
== TCPS_ESTABLISHED
&&
1182 (error
== EHOSTUNREACH
|| error
== ENETUNREACH
||
1183 error
== EHOSTDOWN
)) {
1185 } else if (tp
->t_state
< TCPS_ESTABLISHED
&& tp
->t_rxtshift
> 3 &&
1187 tcp_drop(tp
, error
);
1189 tp
->t_softerror
= error
;
1191 wakeup(&so
->so_timeo
);
1198 tcp_pcblist(SYSCTL_HANDLER_ARGS
)
1201 struct inpcb
*marker
;
1209 * The process of preparing the TCB list is too time-consuming and
1210 * resource-intensive to repeat twice on every request.
1212 if (req
->oldptr
== NULL
) {
1213 for (ccpu
= 0; ccpu
< ncpus2
; ++ccpu
)
1214 n
+= tcbinfo
[ccpu
].ipi_count
;
1215 req
->oldidx
= (n
+ n
/8 + 10) * sizeof(struct xtcpcb
);
1219 if (req
->newptr
!= NULL
)
1222 marker
= kmalloc(sizeof(struct inpcb
), M_TEMP
, M_WAITOK
|M_ZERO
);
1223 marker
->inp_flags
|= INP_PLACEMARKER
;
1226 * OK, now we're committed to doing something. Run the inpcb list
1227 * for each cpu in the system and construct the output. Use a
1228 * list placemarker to deal with list changes occuring during
1229 * copyout blockages (but otherwise depend on being on the correct
1230 * cpu to avoid races).
1232 origcpu
= mycpu
->gd_cpuid
;
1233 for (ccpu
= 0; ccpu
< ncpus2
&& error
== 0; ++ccpu
) {
1237 lwkt_migratecpu(ccpu
);
1239 n
= tcbinfo
[ccpu
].ipi_count
;
1241 LIST_INSERT_HEAD(&tcbinfo
[ccpu
].pcblisthead
, marker
, inp_list
);
1243 while ((inp
= LIST_NEXT(marker
, inp_list
)) != NULL
&& i
< n
) {
1245 * process a snapshot of pcbs, ignoring placemarkers
1246 * and using our own to allow SYSCTL_OUT to block.
1248 LIST_REMOVE(marker
, inp_list
);
1249 LIST_INSERT_AFTER(inp
, marker
, inp_list
);
1251 if (inp
->inp_flags
& INP_PLACEMARKER
)
1253 if (prison_xinpcb(req
->td
, inp
))
1256 xt
.xt_len
= sizeof xt
;
1257 bcopy(inp
, &xt
.xt_inp
, sizeof *inp
);
1258 inp_ppcb
= inp
->inp_ppcb
;
1259 if (inp_ppcb
!= NULL
)
1260 bcopy(inp_ppcb
, &xt
.xt_tp
, sizeof xt
.xt_tp
);
1262 bzero(&xt
.xt_tp
, sizeof xt
.xt_tp
);
1263 if (inp
->inp_socket
)
1264 sotoxsocket(inp
->inp_socket
, &xt
.xt_socket
);
1265 if ((error
= SYSCTL_OUT(req
, &xt
, sizeof xt
)) != 0)
1269 LIST_REMOVE(marker
, inp_list
);
1270 if (error
== 0 && i
< n
) {
1271 bzero(&xt
, sizeof xt
);
1272 xt
.xt_len
= sizeof xt
;
1274 error
= SYSCTL_OUT(req
, &xt
, sizeof xt
);
1283 * Make sure we are on the same cpu we were on originally, since
1284 * higher level callers expect this. Also don't pollute caches with
1285 * migrated userland data by (eventually) returning to userland
1286 * on a different cpu.
1288 lwkt_migratecpu(origcpu
);
1289 kfree(marker
, M_TEMP
);
1293 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_PCBLIST
, pcblist
, CTLFLAG_RD
, 0, 0,
1294 tcp_pcblist
, "S,xtcpcb", "List of active TCP connections");
1297 tcp_getcred(SYSCTL_HANDLER_ARGS
)
1299 struct sockaddr_in addrs
[2];
1300 struct ucred cred0
, *cred
= NULL
;
1302 int cpu
, origcpu
, error
;
1304 error
= priv_check(req
->td
, PRIV_ROOT
);
1307 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1312 cpu
= tcp_addrcpu(addrs
[1].sin_addr
.s_addr
, addrs
[1].sin_port
,
1313 addrs
[0].sin_addr
.s_addr
, addrs
[0].sin_port
);
1315 lwkt_migratecpu(cpu
);
1317 inp
= in_pcblookup_hash(&tcbinfo
[cpu
], addrs
[1].sin_addr
,
1318 addrs
[1].sin_port
, addrs
[0].sin_addr
, addrs
[0].sin_port
, 0, NULL
);
1319 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1321 } else if (inp
->inp_socket
->so_cred
!= NULL
) {
1322 cred0
= *(inp
->inp_socket
->so_cred
);
1326 lwkt_migratecpu(origcpu
);
1331 return SYSCTL_OUT(req
, cred
, sizeof(struct ucred
));
1334 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1335 0, 0, tcp_getcred
, "S,ucred", "Get the ucred of a TCP connection");
1339 tcp6_getcred(SYSCTL_HANDLER_ARGS
)
1341 struct sockaddr_in6 addrs
[2];
1345 error
= priv_check(req
->td
, PRIV_ROOT
);
1348 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1352 inp
= in6_pcblookup_hash(&tcbinfo
[0],
1353 &addrs
[1].sin6_addr
, addrs
[1].sin6_port
,
1354 &addrs
[0].sin6_addr
, addrs
[0].sin6_port
, 0, NULL
);
1355 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1359 error
= SYSCTL_OUT(req
, inp
->inp_socket
->so_cred
, sizeof(struct ucred
));
1365 SYSCTL_PROC(_net_inet6_tcp6
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1367 tcp6_getcred
, "S,ucred", "Get the ucred of a TCP6 connection");
1370 struct netmsg_tcp_notify
{
1371 struct netmsg_base base
;
1372 inp_notify_t nm_notify
;
1373 struct in_addr nm_faddr
;
1378 tcp_notifyall_oncpu(netmsg_t msg
)
1380 struct netmsg_tcp_notify
*nm
= (struct netmsg_tcp_notify
*)msg
;
1383 in_pcbnotifyall(&tcbinfo
[mycpuid
], nm
->nm_faddr
,
1384 nm
->nm_arg
, nm
->nm_notify
);
1386 nextcpu
= mycpuid
+ 1;
1387 if (nextcpu
< ncpus2
)
1388 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nm
->base
.lmsg
);
1390 lwkt_replymsg(&nm
->base
.lmsg
, 0);
1394 tcp_get_inpnotify(int cmd
, const struct sockaddr
*sa
,
1395 int *arg
, struct ip
**ip0
, int *cpuid
)
1397 struct ip
*ip
= *ip0
;
1398 struct in_addr faddr
;
1399 inp_notify_t notify
= tcp_notify
;
1401 faddr
= ((const struct sockaddr_in
*)sa
)->sin_addr
;
1402 if (sa
->sa_family
!= AF_INET
|| faddr
.s_addr
== INADDR_ANY
)
1405 *arg
= inetctlerrmap
[cmd
];
1406 if (cmd
== PRC_QUENCH
) {
1407 notify
= tcp_quench
;
1408 } else if (icmp_may_rst
&&
1409 (cmd
== PRC_UNREACH_ADMIN_PROHIB
||
1410 cmd
== PRC_UNREACH_PORT
||
1411 cmd
== PRC_TIMXCEED_INTRANS
) &&
1413 notify
= tcp_drop_syn_sent
;
1414 } else if (cmd
== PRC_MSGSIZE
) {
1415 const struct icmp
*icmp
= (const struct icmp
*)
1416 ((caddr_t
)ip
- offsetof(struct icmp
, icmp_ip
));
1418 *arg
= ntohs(icmp
->icmp_nextmtu
);
1419 notify
= tcp_mtudisc
;
1420 } else if (PRC_IS_REDIRECT(cmd
)) {
1422 notify
= in_rtchange
;
1423 } else if (cmd
== PRC_HOSTDEAD
) {
1425 } else if ((unsigned)cmd
>= PRC_NCMDS
|| inetctlerrmap
[cmd
] == 0) {
1429 if (cpuid
!= NULL
) {
1431 /* Go through all CPUs */
1434 const struct tcphdr
*th
;
1436 th
= (const struct tcphdr
*)
1437 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1438 *cpuid
= tcp_addrcpu(faddr
.s_addr
, th
->th_dport
,
1439 ip
->ip_src
.s_addr
, th
->th_sport
);
1448 tcp_ctlinput(netmsg_t msg
)
1450 int cmd
= msg
->ctlinput
.nm_cmd
;
1451 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1452 struct ip
*ip
= msg
->ctlinput
.nm_extra
;
1453 struct in_addr faddr
;
1454 inp_notify_t notify
;
1457 notify
= tcp_get_inpnotify(cmd
, sa
, &arg
, &ip
, &cpuid
);
1461 faddr
= ((struct sockaddr_in
*)sa
)->sin_addr
;
1463 const struct tcphdr
*th
;
1466 if (cpuid
!= mycpuid
)
1469 th
= (const struct tcphdr
*)
1470 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1471 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], faddr
, th
->th_dport
,
1472 ip
->ip_src
, th
->th_sport
, 0, NULL
);
1473 if (inp
!= NULL
&& inp
->inp_socket
!= NULL
) {
1474 tcp_seq icmpseq
= htonl(th
->th_seq
);
1475 struct tcpcb
*tp
= intotcpcb(inp
);
1477 if (SEQ_GEQ(icmpseq
, tp
->snd_una
) &&
1478 SEQ_LT(icmpseq
, tp
->snd_max
))
1481 struct in_conninfo inc
;
1483 inc
.inc_fport
= th
->th_dport
;
1484 inc
.inc_lport
= th
->th_sport
;
1485 inc
.inc_faddr
= faddr
;
1486 inc
.inc_laddr
= ip
->ip_src
;
1490 syncache_unreach(&inc
, th
);
1492 } else if (msg
->ctlinput
.nm_direct
) {
1493 if (cpuid
!= ncpus
&& cpuid
!= mycpuid
)
1495 if (mycpuid
>= ncpus2
)
1498 in_pcbnotifyall(&tcbinfo
[mycpuid
], faddr
, arg
, notify
);
1500 struct netmsg_tcp_notify
*nm
;
1502 ASSERT_IN_NETISR(0);
1503 nm
= kmalloc(sizeof(*nm
), M_LWKTMSG
, M_INTWAIT
);
1504 netmsg_init(&nm
->base
, NULL
, &netisr_afree_rport
,
1505 0, tcp_notifyall_oncpu
);
1506 nm
->nm_faddr
= faddr
;
1508 nm
->nm_notify
= notify
;
1510 lwkt_sendmsg(netisr_cpuport(0), &nm
->base
.lmsg
);
1513 lwkt_replymsg(&msg
->lmsg
, 0);
1519 tcp6_ctlinput(netmsg_t msg
)
1521 int cmd
= msg
->ctlinput
.nm_cmd
;
1522 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1523 void *d
= msg
->ctlinput
.nm_extra
;
1525 inp_notify_t notify
= tcp_notify
;
1526 struct ip6_hdr
*ip6
;
1528 struct ip6ctlparam
*ip6cp
= NULL
;
1529 const struct sockaddr_in6
*sa6_src
= NULL
;
1531 struct tcp_portonly
{
1537 if (sa
->sa_family
!= AF_INET6
||
1538 sa
->sa_len
!= sizeof(struct sockaddr_in6
)) {
1543 if (cmd
== PRC_QUENCH
)
1544 notify
= tcp_quench
;
1545 else if (cmd
== PRC_MSGSIZE
) {
1546 struct ip6ctlparam
*ip6cp
= d
;
1547 struct icmp6_hdr
*icmp6
= ip6cp
->ip6c_icmp6
;
1549 arg
= ntohl(icmp6
->icmp6_mtu
);
1550 notify
= tcp_mtudisc
;
1551 } else if (!PRC_IS_REDIRECT(cmd
) &&
1552 ((unsigned)cmd
> PRC_NCMDS
|| inet6ctlerrmap
[cmd
] == 0)) {
1556 /* if the parameter is from icmp6, decode it. */
1558 ip6cp
= (struct ip6ctlparam
*)d
;
1560 ip6
= ip6cp
->ip6c_ip6
;
1561 off
= ip6cp
->ip6c_off
;
1562 sa6_src
= ip6cp
->ip6c_src
;
1566 off
= 0; /* fool gcc */
1571 struct in_conninfo inc
;
1573 * XXX: We assume that when IPV6 is non NULL,
1574 * M and OFF are valid.
1577 /* check if we can safely examine src and dst ports */
1578 if (m
->m_pkthdr
.len
< off
+ sizeof *thp
)
1581 bzero(&th
, sizeof th
);
1582 m_copydata(m
, off
, sizeof *thp
, (caddr_t
)&th
);
1584 in6_pcbnotify(&tcbinfo
[0], sa
, th
.th_dport
,
1585 (struct sockaddr
*)ip6cp
->ip6c_src
,
1586 th
.th_sport
, cmd
, arg
, notify
);
1588 inc
.inc_fport
= th
.th_dport
;
1589 inc
.inc_lport
= th
.th_sport
;
1590 inc
.inc6_faddr
= ((struct sockaddr_in6
*)sa
)->sin6_addr
;
1591 inc
.inc6_laddr
= ip6cp
->ip6c_src
->sin6_addr
;
1593 syncache_unreach(&inc
, &th
);
1595 in6_pcbnotify(&tcbinfo
[0], sa
, 0,
1596 (const struct sockaddr
*)sa6_src
, 0, cmd
, arg
, notify
);
1599 lwkt_replymsg(&msg
->ctlinput
.base
.lmsg
, 0);
1605 * Following is where TCP initial sequence number generation occurs.
1607 * There are two places where we must use initial sequence numbers:
1608 * 1. In SYN-ACK packets.
1609 * 2. In SYN packets.
1611 * All ISNs for SYN-ACK packets are generated by the syncache. See
1612 * tcp_syncache.c for details.
1614 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1615 * depends on this property. In addition, these ISNs should be
1616 * unguessable so as to prevent connection hijacking. To satisfy
1617 * the requirements of this situation, the algorithm outlined in
1618 * RFC 1948 is used to generate sequence numbers.
1620 * Implementation details:
1622 * Time is based off the system timer, and is corrected so that it
1623 * increases by one megabyte per second. This allows for proper
1624 * recycling on high speed LANs while still leaving over an hour
1627 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1628 * between seeding of isn_secret. This is normally set to zero,
1629 * as reseeding should not be necessary.
1633 #define ISN_BYTES_PER_SECOND 1048576
1635 u_char isn_secret
[32];
1636 int isn_last_reseed
;
1640 tcp_new_isn(struct tcpcb
*tp
)
1642 u_int32_t md5_buffer
[4];
1645 /* Seed if this is the first use, reseed if requested. */
1646 if ((isn_last_reseed
== 0) || ((tcp_isn_reseed_interval
> 0) &&
1647 (((u_int
)isn_last_reseed
+ (u_int
)tcp_isn_reseed_interval
*hz
)
1649 read_random_unlimited(&isn_secret
, sizeof isn_secret
);
1650 isn_last_reseed
= ticks
;
1653 /* Compute the md5 hash and return the ISN. */
1655 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_fport
, sizeof(u_short
));
1656 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_lport
, sizeof(u_short
));
1658 if (INP_ISIPV6(tp
->t_inpcb
)) {
1659 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_faddr
,
1660 sizeof(struct in6_addr
));
1661 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_laddr
,
1662 sizeof(struct in6_addr
));
1666 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_faddr
,
1667 sizeof(struct in_addr
));
1668 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_laddr
,
1669 sizeof(struct in_addr
));
1671 MD5Update(&isn_ctx
, (u_char
*) &isn_secret
, sizeof(isn_secret
));
1672 MD5Final((u_char
*) &md5_buffer
, &isn_ctx
);
1673 new_isn
= (tcp_seq
) md5_buffer
[0];
1674 new_isn
+= ticks
* (ISN_BYTES_PER_SECOND
/ hz
);
1679 * When a source quench is received, close congestion window
1680 * to one segment. We will gradually open it again as we proceed.
1683 tcp_quench(struct inpcb
*inp
, int error
)
1685 struct tcpcb
*tp
= intotcpcb(inp
);
1687 KASSERT(tp
!= NULL
, ("tcp_quench: tp is NULL"));
1688 tp
->snd_cwnd
= tp
->t_maxseg
;
1693 * When a specific ICMP unreachable message is received and the
1694 * connection state is SYN-SENT, drop the connection. This behavior
1695 * is controlled by the icmp_may_rst sysctl.
1698 tcp_drop_syn_sent(struct inpcb
*inp
, int error
)
1700 struct tcpcb
*tp
= intotcpcb(inp
);
1702 KASSERT(tp
!= NULL
, ("tcp_drop_syn_sent: tp is NULL"));
1703 if (tp
->t_state
== TCPS_SYN_SENT
)
1704 tcp_drop(tp
, error
);
1708 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1709 * based on the new value in the route. Also nudge TCP to send something,
1710 * since we know the packet we just sent was dropped.
1711 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1714 tcp_mtudisc(struct inpcb
*inp
, int mtu
)
1716 struct tcpcb
*tp
= intotcpcb(inp
);
1718 struct socket
*so
= inp
->inp_socket
;
1721 boolean_t isipv6
= INP_ISIPV6(inp
);
1723 const boolean_t isipv6
= FALSE
;
1726 KASSERT(tp
!= NULL
, ("tcp_mtudisc: tp is NULL"));
1729 * If no MTU is provided in the ICMP message, use the
1730 * next lower likely value, as specified in RFC 1191.
1735 oldmtu
= tp
->t_maxopd
+
1737 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1738 sizeof(struct tcpiphdr
));
1739 mtu
= ip_next_mtu(oldmtu
, 0);
1743 rt
= tcp_rtlookup6(&inp
->inp_inc
);
1745 rt
= tcp_rtlookup(&inp
->inp_inc
);
1747 if (rt
->rt_rmx
.rmx_mtu
!= 0 && rt
->rt_rmx
.rmx_mtu
< mtu
)
1748 mtu
= rt
->rt_rmx
.rmx_mtu
;
1752 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1753 sizeof(struct tcpiphdr
));
1756 * XXX - The following conditional probably violates the TCP
1757 * spec. The problem is that, since we don't know the
1758 * other end's MSS, we are supposed to use a conservative
1759 * default. But, if we do that, then MTU discovery will
1760 * never actually take place, because the conservative
1761 * default is much less than the MTUs typically seen
1762 * on the Internet today. For the moment, we'll sweep
1763 * this under the carpet.
1765 * The conservative default might not actually be a problem
1766 * if the only case this occurs is when sending an initial
1767 * SYN with options and data to a host we've never talked
1768 * to before. Then, they will reply with an MSS value which
1769 * will get recorded and the new parameters should get
1770 * recomputed. For Further Study.
1772 if (rt
->rt_rmx
.rmx_mssopt
&& rt
->rt_rmx
.rmx_mssopt
< maxopd
)
1773 maxopd
= rt
->rt_rmx
.rmx_mssopt
;
1777 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1778 sizeof(struct tcpiphdr
));
1780 if (tp
->t_maxopd
<= maxopd
)
1782 tp
->t_maxopd
= maxopd
;
1785 if ((tp
->t_flags
& (TF_REQ_TSTMP
| TF_RCVD_TSTMP
| TF_NOOPT
)) ==
1786 (TF_REQ_TSTMP
| TF_RCVD_TSTMP
))
1787 mss
-= TCPOLEN_TSTAMP_APPA
;
1789 /* round down to multiple of MCLBYTES */
1790 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1792 mss
&= ~(MCLBYTES
- 1);
1795 mss
= (mss
/ MCLBYTES
) * MCLBYTES
;
1798 if (so
->so_snd
.ssb_hiwat
< mss
)
1799 mss
= so
->so_snd
.ssb_hiwat
;
1803 tp
->snd_nxt
= tp
->snd_una
;
1805 tcpstat
.tcps_mturesent
++;
1809 * Look-up the routing entry to the peer of this inpcb. If no route
1810 * is found and it cannot be allocated the return NULL. This routine
1811 * is called by TCP routines that access the rmx structure and by tcp_mss
1812 * to get the interface MTU.
1815 tcp_rtlookup(struct in_conninfo
*inc
)
1817 struct route
*ro
= &inc
->inc_route
;
1819 if (ro
->ro_rt
== NULL
|| !(ro
->ro_rt
->rt_flags
& RTF_UP
)) {
1820 /* No route yet, so try to acquire one */
1821 if (inc
->inc_faddr
.s_addr
!= INADDR_ANY
) {
1823 * unused portions of the structure MUST be zero'd
1824 * out because rtalloc() treats it as opaque data
1826 bzero(&ro
->ro_dst
, sizeof(struct sockaddr_in
));
1827 ro
->ro_dst
.sa_family
= AF_INET
;
1828 ro
->ro_dst
.sa_len
= sizeof(struct sockaddr_in
);
1829 ((struct sockaddr_in
*) &ro
->ro_dst
)->sin_addr
=
1839 tcp_rtlookup6(struct in_conninfo
*inc
)
1841 struct route_in6
*ro6
= &inc
->inc6_route
;
1843 if (ro6
->ro_rt
== NULL
|| !(ro6
->ro_rt
->rt_flags
& RTF_UP
)) {
1844 /* No route yet, so try to acquire one */
1845 if (!IN6_IS_ADDR_UNSPECIFIED(&inc
->inc6_faddr
)) {
1847 * unused portions of the structure MUST be zero'd
1848 * out because rtalloc() treats it as opaque data
1850 bzero(&ro6
->ro_dst
, sizeof(struct sockaddr_in6
));
1851 ro6
->ro_dst
.sin6_family
= AF_INET6
;
1852 ro6
->ro_dst
.sin6_len
= sizeof(struct sockaddr_in6
);
1853 ro6
->ro_dst
.sin6_addr
= inc
->inc6_faddr
;
1854 rtalloc((struct route
*)ro6
);
1857 return (ro6
->ro_rt
);
1862 /* compute ESP/AH header size for TCP, including outer IP header. */
1864 ipsec_hdrsiz_tcp(struct tcpcb
*tp
)
1872 if ((tp
== NULL
) || ((inp
= tp
->t_inpcb
) == NULL
))
1874 MGETHDR(m
, M_NOWAIT
, MT_DATA
);
1879 if (INP_ISIPV6(inp
)) {
1880 struct ip6_hdr
*ip6
= mtod(m
, struct ip6_hdr
*);
1882 th
= (struct tcphdr
*)(ip6
+ 1);
1883 m
->m_pkthdr
.len
= m
->m_len
=
1884 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
);
1885 tcp_fillheaders(tp
, ip6
, th
, FALSE
);
1886 hdrsiz
= ipsec6_hdrsiz(m
, IPSEC_DIR_OUTBOUND
, inp
);
1890 ip
= mtod(m
, struct ip
*);
1891 th
= (struct tcphdr
*)(ip
+ 1);
1892 m
->m_pkthdr
.len
= m
->m_len
= sizeof(struct tcpiphdr
);
1893 tcp_fillheaders(tp
, ip
, th
, FALSE
);
1894 hdrsiz
= ipsec4_hdrsiz(m
, IPSEC_DIR_OUTBOUND
, inp
);
1903 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1905 * This code attempts to calculate the bandwidth-delay product as a
1906 * means of determining the optimal window size to maximize bandwidth,
1907 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1908 * routers. This code also does a fairly good job keeping RTTs in check
1909 * across slow links like modems. We implement an algorithm which is very
1910 * similar (but not meant to be) TCP/Vegas. The code operates on the
1911 * transmitter side of a TCP connection and so only effects the transmit
1912 * side of the connection.
1914 * BACKGROUND: TCP makes no provision for the management of buffer space
1915 * at the end points or at the intermediate routers and switches. A TCP
1916 * stream, whether using NewReno or not, will eventually buffer as
1917 * many packets as it is able and the only reason this typically works is
1918 * due to the fairly small default buffers made available for a connection
1919 * (typicaly 16K or 32K). As machines use larger windows and/or window
1920 * scaling it is now fairly easy for even a single TCP connection to blow-out
1921 * all available buffer space not only on the local interface, but on
1922 * intermediate routers and switches as well. NewReno makes a misguided
1923 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1924 * then backing off, then steadily increasing the window again until another
1925 * failure occurs, ad-infinitum. This results in terrible oscillation that
1926 * is only made worse as network loads increase and the idea of intentionally
1927 * blowing out network buffers is, frankly, a terrible way to manage network
1930 * It is far better to limit the transmit window prior to the failure
1931 * condition being achieved. There are two general ways to do this: First
1932 * you can 'scan' through different transmit window sizes and locate the
1933 * point where the RTT stops increasing, indicating that you have filled the
1934 * pipe, then scan backwards until you note that RTT stops decreasing, then
1935 * repeat ad-infinitum. This method works in principle but has severe
1936 * implementation issues due to RTT variances, timer granularity, and
1937 * instability in the algorithm which can lead to many false positives and
1938 * create oscillations as well as interact badly with other TCP streams
1939 * implementing the same algorithm.
1941 * The second method is to limit the window to the bandwidth delay product
1942 * of the link. This is the method we implement. RTT variances and our
1943 * own manipulation of the congestion window, bwnd, can potentially
1944 * destabilize the algorithm. For this reason we have to stabilize the
1945 * elements used to calculate the window. We do this by using the minimum
1946 * observed RTT, the long term average of the observed bandwidth, and
1947 * by adding two segments worth of slop. It isn't perfect but it is able
1948 * to react to changing conditions and gives us a very stable basis on
1949 * which to extend the algorithm.
1952 tcp_xmit_bandwidth_limit(struct tcpcb
*tp
, tcp_seq ack_seq
)
1961 * If inflight_enable is disabled in the middle of a tcp connection,
1962 * make sure snd_bwnd is effectively disabled.
1964 if (!tcp_inflight_enable
) {
1965 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
1966 tp
->snd_bandwidth
= 0;
1971 * Validate the delta time. If a connection is new or has been idle
1972 * a long time we have to reset the bandwidth calculator.
1976 delta_ticks
= save_ticks
- tp
->t_bw_rtttime
;
1977 if (tp
->t_bw_rtttime
== 0 || delta_ticks
< 0 || delta_ticks
> hz
* 10) {
1978 tp
->t_bw_rtttime
= save_ticks
;
1979 tp
->t_bw_rtseq
= ack_seq
;
1980 if (tp
->snd_bandwidth
== 0)
1981 tp
->snd_bandwidth
= tcp_inflight_start
;
1986 * A delta of at least 1 tick is required. Waiting 2 ticks will
1987 * result in better (bw) accuracy. More than that and the ramp-up
1990 if (delta_ticks
== 0 || delta_ticks
== 1)
1994 * Sanity check, plus ignore pure window update acks.
1996 if ((int)(ack_seq
- tp
->t_bw_rtseq
) <= 0)
2000 * Figure out the bandwidth. Due to the tick granularity this
2001 * is a very rough number and it MUST be averaged over a fairly
2002 * long period of time. XXX we need to take into account a link
2003 * that is not using all available bandwidth, but for now our
2004 * slop will ramp us up if this case occurs and the bandwidth later
2007 ibw
= (int64_t)(ack_seq
- tp
->t_bw_rtseq
) * hz
/ delta_ticks
;
2008 tp
->t_bw_rtttime
= save_ticks
;
2009 tp
->t_bw_rtseq
= ack_seq
;
2010 bw
= ((int64_t)tp
->snd_bandwidth
* 15 + ibw
) >> 4;
2012 tp
->snd_bandwidth
= bw
;
2015 * Calculate the semi-static bandwidth delay product, plus two maximal
2016 * segments. The additional slop puts us squarely in the sweet
2017 * spot and also handles the bandwidth run-up case. Without the
2018 * slop we could be locking ourselves into a lower bandwidth.
2020 * At very high speeds the bw calculation can become overly sensitive
2021 * and error prone when delta_ticks is low (e.g. usually 1). To deal
2022 * with the problem the stab must be scaled to the bw. A stab of 50
2023 * (the default) increases the bw for the purposes of the bwnd
2024 * calculation by 5%.
2026 * Situations Handled:
2027 * (1) Prevents over-queueing of packets on LANs, especially on
2028 * high speed LANs, allowing larger TCP buffers to be
2029 * specified, and also does a good job preventing
2030 * over-queueing of packets over choke points like modems
2031 * (at least for the transmit side).
2033 * (2) Is able to handle changing network loads (bandwidth
2034 * drops so bwnd drops, bandwidth increases so bwnd
2037 * (3) Theoretically should stabilize in the face of multiple
2038 * connections implementing the same algorithm (this may need
2041 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2042 * be adjusted with a sysctl but typically only needs to be on
2043 * very slow connections. A value no smaller then 5 should
2044 * be used, but only reduce this default if you have no other
2048 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2049 bw
+= bw
* tcp_inflight_stab
/ 1000;
2050 bwnd
= (int64_t)bw
* USERTT
/ (hz
<< TCP_RTT_SHIFT
) +
2051 (int)tp
->t_maxseg
* 2;
2054 if (tcp_inflight_debug
> 0) {
2056 if ((u_int
)(save_ticks
- ltime
) >= hz
/ tcp_inflight_debug
) {
2058 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2059 "bwnd %ld delta %d snd_win %ld\n",
2060 tp
, ibw
, bw
, tp
->t_rttvar
, tp
->t_srtt
,
2061 bwnd
, delta_ticks
, tp
->snd_wnd
);
2064 if ((long)bwnd
< tcp_inflight_min
)
2065 bwnd
= tcp_inflight_min
;
2066 if (bwnd
> tcp_inflight_max
)
2067 bwnd
= tcp_inflight_max
;
2068 if ((long)bwnd
< tp
->t_maxseg
* 2)
2069 bwnd
= tp
->t_maxseg
* 2;
2070 tp
->snd_bwnd
= bwnd
;
2074 tcp_rmx_iwsegs(struct tcpcb
*tp
, u_long
*maxsegs
, u_long
*capsegs
)
2077 struct inpcb
*inp
= tp
->t_inpcb
;
2079 boolean_t isipv6
= INP_ISIPV6(inp
);
2081 const boolean_t isipv6
= FALSE
;
2085 if (tcp_iw_maxsegs
< TCP_IW_MAXSEGS_DFLT
)
2086 tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
2087 if (tcp_iw_capsegs
< TCP_IW_CAPSEGS_DFLT
)
2088 tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
2091 rt
= tcp_rtlookup6(&inp
->inp_inc
);
2093 rt
= tcp_rtlookup(&inp
->inp_inc
);
2095 rt
->rt_rmx
.rmx_iwmaxsegs
< TCP_IW_MAXSEGS_DFLT
||
2096 rt
->rt_rmx
.rmx_iwcapsegs
< TCP_IW_CAPSEGS_DFLT
) {
2097 *maxsegs
= tcp_iw_maxsegs
;
2098 *capsegs
= tcp_iw_capsegs
;
2101 *maxsegs
= rt
->rt_rmx
.rmx_iwmaxsegs
;
2102 *capsegs
= rt
->rt_rmx
.rmx_iwcapsegs
;
2106 tcp_initial_window(struct tcpcb
*tp
)
2108 if (tcp_do_rfc3390
) {
2111 * "If the SYN or SYN/ACK is lost, the initial window
2112 * used by a sender after a correctly transmitted SYN
2113 * MUST be one segment consisting of MSS bytes."
2115 * However, we do something a little bit more aggressive
2116 * then RFC3390 here:
2117 * - Only if time spent in the SYN or SYN|ACK retransmition
2118 * >= 3 seconds, the IW is reduced. We do this mainly
2119 * because when RFC3390 is published, the initial RTO is
2120 * still 3 seconds (the threshold we test here), while
2121 * after RFC6298, the initial RTO is 1 second. This
2122 * behaviour probably still falls within the spirit of
2124 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2125 * Mainly to avoid sender and receiver deadlock until
2126 * delayed ACK timer expires. And even RFC2581 does not
2127 * try to reduce IW upon SYN or SYN|ACK retransmition
2131 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2133 if (tp
->t_rxtsyn
>= TCPTV_RTOBASE3
) {
2134 return (2 * tp
->t_maxseg
);
2136 u_long maxsegs
, capsegs
;
2138 tcp_rmx_iwsegs(tp
, &maxsegs
, &capsegs
);
2139 return min(maxsegs
* tp
->t_maxseg
,
2140 max(2 * tp
->t_maxseg
, capsegs
* 1460));
2144 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2146 * Mainly to avoid sender and receiver deadlock
2147 * until delayed ACK timer expires.
2149 return (2 * tp
->t_maxseg
);
2153 #ifdef TCP_SIGNATURE
2155 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2157 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2158 * When called from tcp_input(), we can be sure that th_sum has been
2159 * zeroed out and verified already.
2161 * Return 0 if successful, otherwise return -1.
2163 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2164 * search with the destination IP address, and a 'magic SPI' to be
2165 * determined by the application. This is hardcoded elsewhere to 1179
2166 * right now. Another branch of this code exists which uses the SPD to
2167 * specify per-application flows but it is unstable.
2170 tcpsignature_compute(
2171 struct mbuf
*m
, /* mbuf chain */
2172 int len
, /* length of TCP data */
2173 int optlen
, /* length of TCP options */
2174 u_char
*buf
, /* storage for MD5 digest */
2175 u_int direction
) /* direction of flow */
2177 struct ippseudo ippseudo
;
2181 struct ipovly
*ipovly
;
2182 struct secasvar
*sav
;
2185 struct ip6_hdr
*ip6
;
2186 struct in6_addr in6
;
2192 KASSERT(m
!= NULL
, ("passed NULL mbuf. Game over."));
2193 KASSERT(buf
!= NULL
, ("passed NULL storage pointer for MD5 signature"));
2195 * Extract the destination from the IP header in the mbuf.
2197 ip
= mtod(m
, struct ip
*);
2199 ip6
= NULL
; /* Make the compiler happy. */
2202 * Look up an SADB entry which matches the address found in
2205 switch (IP_VHL_V(ip
->ip_vhl
)) {
2207 sav
= key_allocsa(AF_INET
, (caddr_t
)&ip
->ip_src
, (caddr_t
)&ip
->ip_dst
,
2208 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2211 case (IPV6_VERSION
>> 4):
2212 ip6
= mtod(m
, struct ip6_hdr
*);
2213 sav
= key_allocsa(AF_INET6
, (caddr_t
)&ip6
->ip6_src
, (caddr_t
)&ip6
->ip6_dst
,
2214 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2223 kprintf("%s: SADB lookup failed\n", __func__
);
2229 * Step 1: Update MD5 hash with IP pseudo-header.
2231 * XXX The ippseudo header MUST be digested in network byte order,
2232 * or else we'll fail the regression test. Assume all fields we've
2233 * been doing arithmetic on have been in host byte order.
2234 * XXX One cannot depend on ipovly->ih_len here. When called from
2235 * tcp_output(), the underlying ip_len member has not yet been set.
2237 switch (IP_VHL_V(ip
->ip_vhl
)) {
2239 ipovly
= (struct ipovly
*)ip
;
2240 ippseudo
.ippseudo_src
= ipovly
->ih_src
;
2241 ippseudo
.ippseudo_dst
= ipovly
->ih_dst
;
2242 ippseudo
.ippseudo_pad
= 0;
2243 ippseudo
.ippseudo_p
= IPPROTO_TCP
;
2244 ippseudo
.ippseudo_len
= htons(len
+ sizeof(struct tcphdr
) + optlen
);
2245 MD5Update(&ctx
, (char *)&ippseudo
, sizeof(struct ippseudo
));
2246 th
= (struct tcphdr
*)((u_char
*)ip
+ sizeof(struct ip
));
2247 doff
= sizeof(struct ip
) + sizeof(struct tcphdr
) + optlen
;
2251 * RFC 2385, 2.0 Proposal
2252 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2253 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2254 * extended next header value (to form 32 bits), and 32-bit segment
2256 * Note: Upper-Layer Packet Length comes before Next Header.
2258 case (IPV6_VERSION
>> 4):
2260 in6_clearscope(&in6
);
2261 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2263 in6_clearscope(&in6
);
2264 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2265 plen
= htonl(len
+ sizeof(struct tcphdr
) + optlen
);
2266 MD5Update(&ctx
, (char *)&plen
, sizeof(uint32_t));
2268 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2269 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2270 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2272 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2273 th
= (struct tcphdr
*)((u_char
*)ip6
+ sizeof(struct ip6_hdr
));
2274 doff
= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) + optlen
;
2283 * Step 2: Update MD5 hash with TCP header, excluding options.
2284 * The TCP checksum must be set to zero.
2286 savecsum
= th
->th_sum
;
2288 MD5Update(&ctx
, (char *)th
, sizeof(struct tcphdr
));
2289 th
->th_sum
= savecsum
;
2291 * Step 3: Update MD5 hash with TCP segment data.
2292 * Use m_apply() to avoid an early m_pullup().
2295 m_apply(m
, doff
, len
, tcpsignature_apply
, &ctx
);
2297 * Step 4: Update MD5 hash with shared secret.
2299 MD5Update(&ctx
, _KEYBUF(sav
->key_auth
), _KEYLEN(sav
->key_auth
));
2300 MD5Final(buf
, &ctx
);
2301 key_sa_recordxfer(sav
, m
);
2307 tcpsignature_apply(void *fstate
, void *data
, unsigned int len
)
2310 MD5Update((MD5_CTX
*)fstate
, (unsigned char *)data
, len
);
2313 #endif /* TCP_SIGNATURE */
2316 tcp_drop_sysctl_dispatch(netmsg_t nmsg
)
2318 struct lwkt_msg
*lmsg
= &nmsg
->lmsg
;
2319 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2320 struct sockaddr_storage
*addrs
= lmsg
->u
.ms_resultp
;
2322 struct sockaddr_in
*fin
, *lin
;
2324 struct sockaddr_in6
*fin6
, *lin6
;
2325 struct in6_addr f6
, l6
;
2329 switch (addrs
[0].ss_family
) {
2332 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2333 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2334 error
= in6_embedscope(&f6
, fin6
, NULL
, NULL
);
2337 error
= in6_embedscope(&l6
, lin6
, NULL
, NULL
);
2340 inp
= in6_pcblookup_hash(&tcbinfo
[mycpuid
], &f6
,
2341 fin6
->sin6_port
, &l6
, lin6
->sin6_port
, FALSE
, NULL
);
2346 fin
= (struct sockaddr_in
*)&addrs
[0];
2347 lin
= (struct sockaddr_in
*)&addrs
[1];
2348 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], fin
->sin_addr
,
2349 fin
->sin_port
, lin
->sin_addr
, lin
->sin_port
, FALSE
, NULL
);
2354 * Must not reach here, since the address family was
2355 * checked in sysctl handler.
2357 panic("unknown address family %d", addrs
[0].ss_family
);
2360 struct tcpcb
*tp
= intotcpcb(inp
);
2362 KASSERT((inp
->inp_flags
& INP_WILDCARD
) == 0,
2363 ("in wildcard hash"));
2364 KASSERT(tp
!= NULL
, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2365 KASSERT((tp
->t_flags
& TF_LISTEN
) == 0, ("listen socket"));
2366 tcp_drop(tp
, ECONNABORTED
);
2374 lwkt_replymsg(lmsg
, error
);
2378 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS
)
2380 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2381 struct sockaddr_storage addrs
[2];
2382 struct sockaddr_in
*fin
, *lin
;
2384 struct sockaddr_in6
*fin6
, *lin6
;
2386 struct netmsg_base nmsg
;
2387 struct lwkt_msg
*lmsg
= &nmsg
.lmsg
;
2388 struct lwkt_port
*port
= NULL
;
2397 if (req
->oldptr
!= NULL
|| req
->oldlen
!= 0)
2399 if (req
->newptr
== NULL
)
2401 if (req
->newlen
< sizeof(addrs
))
2403 error
= SYSCTL_IN(req
, &addrs
, sizeof(addrs
));
2407 switch (addrs
[0].ss_family
) {
2410 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2411 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2412 if (fin6
->sin6_len
!= sizeof(struct sockaddr_in6
) ||
2413 lin6
->sin6_len
!= sizeof(struct sockaddr_in6
))
2415 if (IN6_IS_ADDR_V4MAPPED(&fin6
->sin6_addr
) ||
2416 IN6_IS_ADDR_V4MAPPED(&lin6
->sin6_addr
))
2417 return (EADDRNOTAVAIL
);
2419 error
= sa6_embedscope(fin6
, V_ip6_use_defzone
);
2422 error
= sa6_embedscope(lin6
, V_ip6_use_defzone
);
2426 port
= tcp6_addrport();
2431 fin
= (struct sockaddr_in
*)&addrs
[0];
2432 lin
= (struct sockaddr_in
*)&addrs
[1];
2433 if (fin
->sin_len
!= sizeof(struct sockaddr_in
) ||
2434 lin
->sin_len
!= sizeof(struct sockaddr_in
))
2436 port
= tcp_addrport(fin
->sin_addr
.s_addr
, fin
->sin_port
,
2437 lin
->sin_addr
.s_addr
, lin
->sin_port
);
2444 netmsg_init(&nmsg
, NULL
, &curthread
->td_msgport
, 0,
2445 tcp_drop_sysctl_dispatch
);
2446 lmsg
->u
.ms_resultp
= addrs
;
2447 return lwkt_domsg(port
, lmsg
, 0);
2450 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, drop
,
2451 CTLTYPE_STRUCT
| CTLFLAG_WR
| CTLFLAG_SKIP
, NULL
,
2452 0, sysctl_tcp_drop
, "", "Drop TCP connection");
2455 sysctl_tcps_count(SYSCTL_HANDLER_ARGS
)
2457 u_long state_count
[TCP_NSTATES
];
2460 memset(state_count
, 0, sizeof(state_count
));
2461 for (cpu
= 0; cpu
< ncpus2
; ++cpu
) {
2464 for (i
= 0; i
< TCP_NSTATES
; ++i
)
2465 state_count
[i
] += tcpstate_count
[cpu
].tcps_count
[i
];
2468 return sysctl_handle_opaque(oidp
, state_count
, sizeof(state_count
), req
);
2470 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, state_count
,
2471 CTLTYPE_OPAQUE
| CTLFLAG_RD
, NULL
, 0,
2472 sysctl_tcps_count
, "LU", "TCP connection counts by state");
2475 tcp_pcbport_create(struct tcpcb
*tp
)
2479 KASSERT((tp
->t_flags
& TF_LISTEN
) && tp
->t_state
== TCPS_LISTEN
,
2480 ("not a listen tcpcb"));
2482 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache was created"));
2483 tp
->t_pcbport
= kmalloc_cachealign(sizeof(struct tcp_pcbport
) * ncpus2
,
2486 for (cpu
= 0; cpu
< ncpus2
; ++cpu
) {
2487 struct inpcbport
*phd
;
2489 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2490 LIST_INIT(&phd
->phd_pcblist
);
2491 /* Though, not used ... */
2492 phd
->phd_port
= tp
->t_inpcb
->inp_lport
;
2497 tcp_pcbport_merge_oncpu(struct tcpcb
*tp
)
2499 struct inpcbport
*phd
;
2503 KASSERT(cpu
< ncpus2
, ("invalid cpu%d", cpu
));
2504 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2506 while ((inp
= LIST_FIRST(&phd
->phd_pcblist
)) != NULL
) {
2507 KASSERT(inp
->inp_phd
== phd
&& inp
->inp_porthash
== NULL
,
2508 ("not on tcpcb port cache"));
2509 LIST_REMOVE(inp
, inp_portlist
);
2510 in_pcbinsporthash_lport(inp
);
2511 KASSERT(inp
->inp_phd
== tp
->t_inpcb
->inp_phd
&&
2512 inp
->inp_porthash
== tp
->t_inpcb
->inp_porthash
,
2513 ("tcpcb port cache merge failed"));
2518 tcp_pcbport_destroy(struct tcpcb
*tp
)
2523 for (cpu
= 0; cpu
< ncpus2
; ++cpu
) {
2524 KASSERT(LIST_EMPTY(&tp
->t_pcbport
[cpu
].t_phd
.phd_pcblist
),
2525 ("tcpcb port cache is not empty"));
2528 kfree(tp
->t_pcbport
, M_PCB
);
2529 tp
->t_pcbport
= NULL
;