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 inpcbinfo tcbinfo
[MAXCPU
];
161 struct tcpcbackq tcpcbackq
[MAXCPU
];
163 int tcp_mssdflt
= TCP_MSS
;
164 SYSCTL_INT(_net_inet_tcp
, TCPCTL_MSSDFLT
, mssdflt
, CTLFLAG_RW
,
165 &tcp_mssdflt
, 0, "Default TCP Maximum Segment Size");
168 int tcp_v6mssdflt
= TCP6_MSS
;
169 SYSCTL_INT(_net_inet_tcp
, TCPCTL_V6MSSDFLT
, v6mssdflt
, CTLFLAG_RW
,
170 &tcp_v6mssdflt
, 0, "Default TCP Maximum Segment Size for IPv6");
174 * Minimum MSS we accept and use. This prevents DoS attacks where
175 * we are forced to a ridiculous low MSS like 20 and send hundreds
176 * of packets instead of one. The effect scales with the available
177 * bandwidth and quickly saturates the CPU and network interface
178 * with packet generation and sending. Set to zero to disable MINMSS
179 * checking. This setting prevents us from sending too small packets.
181 int tcp_minmss
= TCP_MINMSS
;
182 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, minmss
, CTLFLAG_RW
,
183 &tcp_minmss
, 0, "Minmum TCP Maximum Segment Size");
186 static int tcp_rttdflt
= TCPTV_SRTTDFLT
/ PR_SLOWHZ
;
187 SYSCTL_INT(_net_inet_tcp
, TCPCTL_RTTDFLT
, rttdflt
, CTLFLAG_RW
,
188 &tcp_rttdflt
, 0, "Default maximum TCP Round Trip Time");
191 int tcp_do_rfc1323
= 1;
192 SYSCTL_INT(_net_inet_tcp
, TCPCTL_DO_RFC1323
, rfc1323
, CTLFLAG_RW
,
193 &tcp_do_rfc1323
, 0, "Enable rfc1323 (high performance TCP) extensions");
195 static int tcp_tcbhashsize
= 0;
196 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, tcbhashsize
, CTLFLAG_RD
,
197 &tcp_tcbhashsize
, 0, "Size of TCP control block hashtable");
199 static int do_tcpdrain
= 1;
200 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, do_tcpdrain
, CTLFLAG_RW
, &do_tcpdrain
, 0,
201 "Enable tcp_drain routine for extra help when low on mbufs");
203 static int icmp_may_rst
= 1;
204 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, icmp_may_rst
, CTLFLAG_RW
, &icmp_may_rst
, 0,
205 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
207 static int tcp_isn_reseed_interval
= 0;
208 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, isn_reseed_interval
, CTLFLAG_RW
,
209 &tcp_isn_reseed_interval
, 0, "Seconds between reseeding of ISN secret");
212 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
213 * by default, but with generous values which should allow maximal
214 * bandwidth. In particular, the slop defaults to 50 (5 packets).
216 * The reason for doing this is that the limiter is the only mechanism we
217 * have which seems to do a really good job preventing receiver RX rings
218 * on network interfaces from getting blown out. Even though GigE/10GigE
219 * is supposed to flow control it looks like either it doesn't actually
220 * do it or Open Source drivers do not properly enable it.
222 * People using the limiter to reduce bottlenecks on slower WAN connections
223 * should set the slop to 20 (2 packets).
225 static int tcp_inflight_enable
= 1;
226 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_enable
, CTLFLAG_RW
,
227 &tcp_inflight_enable
, 0, "Enable automatic TCP inflight data limiting");
229 static int tcp_inflight_debug
= 0;
230 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_debug
, CTLFLAG_RW
,
231 &tcp_inflight_debug
, 0, "Debug TCP inflight calculations");
234 * NOTE: tcp_inflight_start is essentially the starting receive window
235 * for a connection. If set too low then fetches over tcp
236 * connections will take noticably longer to ramp-up over
237 * high-latency connections. 6144 is too low for a default,
238 * use something more reasonable.
240 static int tcp_inflight_start
= 33792;
241 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_start
, CTLFLAG_RW
,
242 &tcp_inflight_start
, 0, "Start value for TCP inflight window");
244 static int tcp_inflight_min
= 6144;
245 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_min
, CTLFLAG_RW
,
246 &tcp_inflight_min
, 0, "Lower bound for TCP inflight window");
248 static int tcp_inflight_max
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
249 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_max
, CTLFLAG_RW
,
250 &tcp_inflight_max
, 0, "Upper bound for TCP inflight window");
252 static int tcp_inflight_stab
= 50;
253 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_stab
, CTLFLAG_RW
,
254 &tcp_inflight_stab
, 0, "Fudge bw 1/10% (50=5%)");
256 static int tcp_inflight_adjrtt
= 2;
257 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_adjrtt
, CTLFLAG_RW
,
258 &tcp_inflight_adjrtt
, 0, "Slop for rtt 1/(hz*32)");
260 static int tcp_do_rfc3390
= 1;
261 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, rfc3390
, CTLFLAG_RW
,
263 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
265 static u_long tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
266 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwmaxsegs
, CTLFLAG_RW
,
267 &tcp_iw_maxsegs
, 0, "TCP IW segments max");
269 static u_long tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
270 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwcapsegs
, CTLFLAG_RW
,
271 &tcp_iw_capsegs
, 0, "TCP IW segments");
273 int tcp_low_rtobase
= 1;
274 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, low_rtobase
, CTLFLAG_RW
,
275 &tcp_low_rtobase
, 0, "Lowering the Initial RTO (RFC 6298)");
277 static int tcp_do_ncr
= 1;
278 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr
, CTLFLAG_RW
,
279 &tcp_do_ncr
, 0, "Non-Congestion Robustness (RFC 4653)");
281 int tcp_ncr_linklocal
= 0;
282 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_linklocal
, CTLFLAG_RW
,
283 &tcp_ncr_linklocal
, 0,
284 "Enable Non-Congestion Robustness (RFC 4653) on link local network");
286 int tcp_ncr_rxtthresh_max
= 16;
287 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_rxtthresh_max
, CTLFLAG_RW
,
288 &tcp_ncr_rxtthresh_max
, 0,
289 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
291 static MALLOC_DEFINE(M_TCPTEMP
, "tcptemp", "TCP Templates for Keepalives");
292 static struct malloc_pipe tcptemp_mpipe
;
294 static void tcp_willblock(void);
295 static void tcp_notify (struct inpcb
*, int);
297 struct tcp_stats tcpstats_percpu
[MAXCPU
] __cachealign
;
298 struct tcp_state_count tcpstate_count
[MAXCPU
] __cachealign
;
300 static struct netmsg_base tcp_drain_netmsg
[MAXCPU
];
301 static void tcp_drain_dispatch(netmsg_t nmsg
);
304 sysctl_tcpstats(SYSCTL_HANDLER_ARGS
)
308 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
309 if ((error
= SYSCTL_OUT(req
, &tcpstats_percpu
[cpu
],
310 sizeof(struct tcp_stats
))))
312 if ((error
= SYSCTL_IN(req
, &tcpstats_percpu
[cpu
],
313 sizeof(struct tcp_stats
))))
319 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_STATS
, stats
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
320 0, 0, sysctl_tcpstats
, "S,tcp_stats", "TCP statistics");
323 * Target size of TCP PCB hash tables. Must be a power of two.
325 * Note that this can be overridden by the kernel environment
326 * variable net.inet.tcp.tcbhashsize
329 #define TCBHASHSIZE 512
333 * This is the actual shape of what we allocate using the zone
334 * allocator. Doing it this way allows us to protect both structures
335 * using the same generation count, and also eliminates the overhead
336 * of allocating tcpcbs separately. By hiding the structure here,
337 * we avoid changing most of the rest of the code (although it needs
338 * to be changed, eventually, for greater efficiency).
341 #define ALIGNM1 (ALIGNMENT - 1)
345 char align
[(sizeof(struct inpcb
) + ALIGNM1
) & ~ALIGNM1
];
348 struct tcp_callout inp_tp_rexmt
;
349 struct tcp_callout inp_tp_persist
;
350 struct tcp_callout inp_tp_keep
;
351 struct tcp_callout inp_tp_2msl
;
352 struct tcp_callout inp_tp_delack
;
353 struct netmsg_tcp_timer inp_tp_timermsg
;
354 struct netmsg_base inp_tp_sndmore
;
365 struct inpcbportinfo
*portinfo
;
366 struct inpcbinfo
*ticb
;
367 int hashsize
= TCBHASHSIZE
;
371 * note: tcptemp is used for keepalives, and it is ok for an
372 * allocation to fail so do not specify MPF_INT.
374 mpipe_init(&tcptemp_mpipe
, M_TCPTEMP
, sizeof(struct tcptemp
),
375 25, -1, 0, NULL
, NULL
, NULL
);
377 tcp_delacktime
= TCPTV_DELACK
;
378 tcp_keepinit
= TCPTV_KEEP_INIT
;
379 tcp_keepidle
= TCPTV_KEEP_IDLE
;
380 tcp_keepintvl
= TCPTV_KEEPINTVL
;
381 tcp_maxpersistidle
= TCPTV_KEEP_IDLE
;
383 tcp_rexmit_min
= TCPTV_MIN
;
384 if (tcp_rexmit_min
< 1) /* if kern.hz is too low */
386 tcp_rexmit_slop
= TCPTV_CPU_VAR
;
388 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize
);
389 if (!powerof2(hashsize
)) {
390 kprintf("WARNING: TCB hash size not a power of 2\n");
391 hashsize
= 512; /* safe default */
393 tcp_tcbhashsize
= hashsize
;
395 portinfo
= kmalloc_cachealign(sizeof(*portinfo
) * netisr_ncpus
, M_PCB
,
398 for (cpu
= 0; cpu
< netisr_ncpus
; cpu
++) {
399 ticb
= &tcbinfo
[cpu
];
400 in_pcbinfo_init(ticb
, cpu
, FALSE
);
401 ticb
->hashbase
= hashinit(hashsize
, M_PCB
,
403 in_pcbportinfo_init(&portinfo
[cpu
], hashsize
, cpu
);
404 in_pcbportinfo_set(ticb
, portinfo
, netisr_ncpus
);
405 ticb
->wildcardhashbase
= hashinit(hashsize
, M_PCB
,
406 &ticb
->wildcardhashmask
);
407 ticb
->localgrphashbase
= hashinit(hashsize
, M_PCB
,
408 &ticb
->localgrphashmask
);
409 ticb
->ipi_size
= sizeof(struct inp_tp
);
410 TAILQ_INIT(&tcpcbackq
[cpu
].head
);
413 tcp_reass_maxseg
= nmbclusters
/ 16;
414 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg
);
417 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
419 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
421 if (max_protohdr
< TCP_MINPROTOHDR
)
422 max_protohdr
= TCP_MINPROTOHDR
;
423 if (max_linkhdr
+ TCP_MINPROTOHDR
> MHLEN
)
425 #undef TCP_MINPROTOHDR
428 * Initialize TCP statistics counters for each CPU.
430 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
)
431 bzero(&tcpstats_percpu
[cpu
], sizeof(struct tcp_stats
));
434 * Initialize netmsgs for TCP drain
436 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
437 netmsg_init(&tcp_drain_netmsg
[cpu
], NULL
, &netisr_adone_rport
,
438 MSGF_PRIORITY
, tcp_drain_dispatch
);
442 netisr_register_rollup(tcp_willblock
, NETISR_ROLLUP_PRIO_TCP
);
451 while ((tp
= TAILQ_FIRST(&tcpcbackq
[cpu
].head
)) != NULL
) {
452 KKASSERT(tp
->t_flags
& TF_ONOUTPUTQ
);
453 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
454 TAILQ_REMOVE(&tcpcbackq
[cpu
].head
, tp
, t_outputq
);
460 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
461 * tcp_template used to store this data in mbufs, but we now recopy it out
462 * of the tcpcb each time to conserve mbufs.
465 tcp_fillheaders(struct tcpcb
*tp
, void *ip_ptr
, void *tcp_ptr
, boolean_t tso
)
467 struct inpcb
*inp
= tp
->t_inpcb
;
468 struct tcphdr
*tcp_hdr
= (struct tcphdr
*)tcp_ptr
;
471 if (INP_ISIPV6(inp
)) {
474 ip6
= (struct ip6_hdr
*)ip_ptr
;
475 ip6
->ip6_flow
= (ip6
->ip6_flow
& ~IPV6_FLOWINFO_MASK
) |
476 (inp
->in6p_flowinfo
& IPV6_FLOWINFO_MASK
);
477 ip6
->ip6_vfc
= (ip6
->ip6_vfc
& ~IPV6_VERSION_MASK
) |
478 (IPV6_VERSION
& IPV6_VERSION_MASK
);
479 ip6
->ip6_nxt
= IPPROTO_TCP
;
480 ip6
->ip6_plen
= sizeof(struct tcphdr
);
481 ip6
->ip6_src
= inp
->in6p_laddr
;
482 ip6
->ip6_dst
= inp
->in6p_faddr
;
487 struct ip
*ip
= (struct ip
*) ip_ptr
;
490 ip
->ip_vhl
= IP_VHL_BORING
;
497 ip
->ip_p
= IPPROTO_TCP
;
498 ip
->ip_src
= inp
->inp_laddr
;
499 ip
->ip_dst
= inp
->inp_faddr
;
502 plen
= htons(IPPROTO_TCP
);
504 plen
= htons(sizeof(struct tcphdr
) + IPPROTO_TCP
);
505 tcp_hdr
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
,
506 ip
->ip_dst
.s_addr
, plen
);
509 tcp_hdr
->th_sport
= inp
->inp_lport
;
510 tcp_hdr
->th_dport
= inp
->inp_fport
;
515 tcp_hdr
->th_flags
= 0;
521 * Create template to be used to send tcp packets on a connection.
522 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
523 * use for this function is in keepalives, which use tcp_respond.
526 tcp_maketemplate(struct tcpcb
*tp
)
530 if ((tmp
= mpipe_alloc_nowait(&tcptemp_mpipe
)) == NULL
)
532 tcp_fillheaders(tp
, &tmp
->tt_ipgen
, &tmp
->tt_t
, FALSE
);
537 tcp_freetemplate(struct tcptemp
*tmp
)
539 mpipe_free(&tcptemp_mpipe
, tmp
);
543 * Send a single message to the TCP at address specified by
544 * the given TCP/IP header. If m == NULL, then we make a copy
545 * of the tcpiphdr at ti and send directly to the addressed host.
546 * This is used to force keep alive messages out using the TCP
547 * template for a connection. If flags are given then we send
548 * a message back to the TCP which originated the * segment ti,
549 * and discard the mbuf containing it and any other attached mbufs.
551 * In any case the ack and sequence number of the transmitted
552 * segment are as specified by the parameters.
554 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
557 tcp_respond(struct tcpcb
*tp
, void *ipgen
, struct tcphdr
*th
, struct mbuf
*m
,
558 tcp_seq ack
, tcp_seq seq
, int flags
)
562 struct route
*ro
= NULL
;
564 struct ip
*ip
= ipgen
;
567 struct route_in6
*ro6
= NULL
;
568 struct route_in6 sro6
;
569 struct ip6_hdr
*ip6
= ipgen
;
570 struct inpcb
*inp
= NULL
;
571 boolean_t use_tmpro
= TRUE
;
573 boolean_t isipv6
= (IP_VHL_V(ip
->ip_vhl
) == 6);
575 const boolean_t isipv6
= FALSE
;
580 if (!(flags
& TH_RST
)) {
581 win
= ssb_space(&inp
->inp_socket
->so_rcv
);
584 if (win
> (long)TCP_MAXWIN
<< tp
->rcv_scale
)
585 win
= (long)TCP_MAXWIN
<< tp
->rcv_scale
;
588 * Don't use the route cache of a listen socket,
589 * it is not MPSAFE; use temporary route cache.
591 if (tp
->t_state
!= TCPS_LISTEN
) {
593 ro6
= &inp
->in6p_route
;
595 ro
= &inp
->inp_route
;
602 bzero(ro6
, sizeof *ro6
);
605 bzero(ro
, sizeof *ro
);
609 m
= m_gethdr(M_NOWAIT
, MT_HEADER
);
613 m
->m_data
+= max_linkhdr
;
615 bcopy(ip6
, mtod(m
, caddr_t
), sizeof(struct ip6_hdr
));
616 ip6
= mtod(m
, struct ip6_hdr
*);
617 nth
= (struct tcphdr
*)(ip6
+ 1);
619 bcopy(ip
, mtod(m
, caddr_t
), sizeof(struct ip
));
620 ip
= mtod(m
, struct ip
*);
621 nth
= (struct tcphdr
*)(ip
+ 1);
623 bcopy(th
, nth
, sizeof(struct tcphdr
));
628 m
->m_data
= (caddr_t
)ipgen
;
629 /* m_len is set later */
631 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
633 xchg(ip6
->ip6_dst
, ip6
->ip6_src
, struct in6_addr
);
634 nth
= (struct tcphdr
*)(ip6
+ 1);
636 xchg(ip
->ip_dst
.s_addr
, ip
->ip_src
.s_addr
, n_long
);
637 nth
= (struct tcphdr
*)(ip
+ 1);
641 * this is usually a case when an extension header
642 * exists between the IPv6 header and the
645 nth
->th_sport
= th
->th_sport
;
646 nth
->th_dport
= th
->th_dport
;
648 xchg(nth
->th_dport
, nth
->th_sport
, n_short
);
653 ip6
->ip6_vfc
= IPV6_VERSION
;
654 ip6
->ip6_nxt
= IPPROTO_TCP
;
655 ip6
->ip6_plen
= htons((u_short
)(sizeof(struct tcphdr
) + tlen
));
656 tlen
+= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
);
658 tlen
+= sizeof(struct tcpiphdr
);
660 ip
->ip_ttl
= ip_defttl
;
663 m
->m_pkthdr
.len
= tlen
;
664 m
->m_pkthdr
.rcvif
= NULL
;
665 nth
->th_seq
= htonl(seq
);
666 nth
->th_ack
= htonl(ack
);
668 nth
->th_off
= sizeof(struct tcphdr
) >> 2;
669 nth
->th_flags
= flags
;
671 nth
->th_win
= htons((u_short
) (win
>> tp
->rcv_scale
));
673 nth
->th_win
= htons((u_short
)win
);
677 nth
->th_sum
= in6_cksum(m
, IPPROTO_TCP
,
678 sizeof(struct ip6_hdr
),
679 tlen
- sizeof(struct ip6_hdr
));
680 ip6
->ip6_hlim
= in6_selecthlim(inp
,
681 (ro6
&& ro6
->ro_rt
) ? ro6
->ro_rt
->rt_ifp
: NULL
);
683 nth
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
, ip
->ip_dst
.s_addr
,
684 htons((u_short
)(tlen
- sizeof(struct ip
) + ip
->ip_p
)));
685 m
->m_pkthdr
.csum_flags
= CSUM_TCP
;
686 m
->m_pkthdr
.csum_data
= offsetof(struct tcphdr
, th_sum
);
687 m
->m_pkthdr
.csum_thlen
= sizeof(struct tcphdr
);
690 if (tp
== NULL
|| (inp
->inp_socket
->so_options
& SO_DEBUG
))
691 tcp_trace(TA_OUTPUT
, 0, tp
, mtod(m
, void *), th
, 0);
694 ip6_output(m
, NULL
, ro6
, ipflags
, NULL
, NULL
, inp
);
695 if ((ro6
== &sro6
) && (ro6
->ro_rt
!= NULL
)) {
700 if (inp
!= NULL
&& (inp
->inp_flags
& INP_HASH
))
701 m_sethash(m
, inp
->inp_hashval
);
702 ipflags
|= IP_DEBUGROUTE
;
703 ip_output(m
, NULL
, ro
, ipflags
, NULL
, inp
);
704 if ((ro
== &sro
) && (ro
->ro_rt
!= NULL
)) {
712 * Create a new TCP control block, making an
713 * empty reassembly queue and hooking it to the argument
714 * protocol control block. The `inp' parameter must have
715 * come from the zone allocator set up in tcp_init().
718 tcp_newtcpcb(struct inpcb
*inp
)
723 boolean_t isipv6
= INP_ISIPV6(inp
);
725 const boolean_t isipv6
= FALSE
;
728 it
= (struct inp_tp
*)inp
;
730 bzero(tp
, sizeof(struct tcpcb
));
731 TAILQ_INIT(&tp
->t_segq
);
732 tp
->t_maxseg
= tp
->t_maxopd
= isipv6
? tcp_v6mssdflt
: tcp_mssdflt
;
733 tp
->t_rxtthresh
= tcprexmtthresh
;
735 /* Set up our timeouts. */
736 tp
->tt_rexmt
= &it
->inp_tp_rexmt
;
737 tp
->tt_persist
= &it
->inp_tp_persist
;
738 tp
->tt_keep
= &it
->inp_tp_keep
;
739 tp
->tt_2msl
= &it
->inp_tp_2msl
;
740 tp
->tt_delack
= &it
->inp_tp_delack
;
744 * Zero out timer message. We don't create it here,
745 * since the current CPU may not be the owner of this
748 tp
->tt_msg
= &it
->inp_tp_timermsg
;
749 bzero(tp
->tt_msg
, sizeof(*tp
->tt_msg
));
751 tp
->t_keepinit
= tcp_keepinit
;
752 tp
->t_keepidle
= tcp_keepidle
;
753 tp
->t_keepintvl
= tcp_keepintvl
;
754 tp
->t_keepcnt
= tcp_keepcnt
;
755 tp
->t_maxidle
= tp
->t_keepintvl
* tp
->t_keepcnt
;
758 tp
->t_flags
|= TF_NCR
;
760 tp
->t_flags
|= (TF_REQ_SCALE
| TF_REQ_TSTMP
);
762 tp
->t_inpcb
= inp
; /* XXX */
765 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
766 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
767 * reasonable initial retransmit time.
769 tp
->t_srtt
= TCPTV_SRTTBASE
;
771 ((TCPTV_RTOBASE
- TCPTV_SRTTBASE
) << TCP_RTTVAR_SHIFT
) / 4;
772 tp
->t_rttmin
= tcp_rexmit_min
;
773 tp
->t_rxtcur
= TCPTV_RTOBASE
;
774 tp
->snd_cwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
775 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
776 tp
->snd_ssthresh
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
777 tp
->snd_last
= ticks
;
778 tp
->t_rcvtime
= ticks
;
780 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
781 * because the socket may be bound to an IPv6 wildcard address,
782 * which may match an IPv4-mapped IPv6 address.
784 inp
->inp_ip_ttl
= ip_defttl
;
786 tcp_sack_tcpcb_init(tp
);
788 tp
->tt_sndmore
= &it
->inp_tp_sndmore
;
793 * Drop a TCP connection, reporting the specified error.
794 * If connection is synchronized, then send a RST to peer.
797 tcp_drop(struct tcpcb
*tp
, int error
)
799 struct socket
*so
= tp
->t_inpcb
->inp_socket
;
801 if (TCPS_HAVERCVDSYN(tp
->t_state
)) {
802 TCP_STATE_CHANGE(tp
, TCPS_CLOSED
);
804 tcpstat
.tcps_drops
++;
806 tcpstat
.tcps_conndrops
++;
807 if (error
== ETIMEDOUT
&& tp
->t_softerror
)
808 error
= tp
->t_softerror
;
809 so
->so_error
= error
;
810 return (tcp_close(tp
));
813 struct netmsg_listen_detach
{
814 struct netmsg_base base
;
816 struct tcpcb
*nm_tp_inh
;
820 tcp_listen_detach_handler(netmsg_t msg
)
822 struct netmsg_listen_detach
*nmsg
= (struct netmsg_listen_detach
*)msg
;
823 struct tcpcb
*tp
= nmsg
->nm_tp
;
824 int cpu
= mycpuid
, nextcpu
;
826 if (tp
->t_flags
& TF_LISTEN
) {
827 syncache_destroy(tp
, nmsg
->nm_tp_inh
);
828 tcp_pcbport_merge_oncpu(tp
);
831 in_pcbremwildcardhash_oncpu(tp
->t_inpcb
, &tcbinfo
[cpu
]);
834 if (nextcpu
< netisr_ncpus
)
835 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nmsg
->base
.lmsg
);
837 lwkt_replymsg(&nmsg
->base
.lmsg
, 0);
841 * Close a TCP control block:
842 * discard all space held by the tcp
843 * discard internet protocol block
844 * wake up any sleepers
847 tcp_close(struct tcpcb
*tp
)
850 struct inpcb
*inp
= tp
->t_inpcb
;
851 struct inpcb
*inp_inh
= NULL
;
852 struct tcpcb
*tp_inh
= NULL
;
853 struct socket
*so
= inp
->inp_socket
;
855 boolean_t dosavessthresh
;
857 boolean_t isipv6
= INP_ISIPV6(inp
);
859 const boolean_t isipv6
= FALSE
;
862 if (tp
->t_flags
& TF_LISTEN
) {
864 * Pending socket/syncache inheritance
866 * If this is a listen(2) socket, find another listen(2)
867 * socket in the same local group, which could inherit
868 * the syncache and sockets pending on the completion
869 * and incompletion queues.
872 * Currently the inheritance could only happen on the
873 * listen(2) sockets w/ SO_REUSEPORT set.
876 inp_inh
= in_pcblocalgroup_last(&tcbinfo
[0], inp
);
878 tp_inh
= intotcpcb(inp_inh
);
882 * INP_WILDCARD indicates that listen(2) has been called on
883 * this socket. This implies:
884 * - A wildcard inp's hash is replicated for each protocol thread.
885 * - Syncache for this inp grows independently in each protocol
887 * - There is more than one cpu
889 * We have to chain a message to the rest of the protocol threads
890 * to cleanup the wildcard hash and the syncache. The cleanup
891 * in the current protocol thread is defered till the end of this
892 * function (syncache_destroy and in_pcbdetach).
895 * After cleanup the inp's hash and syncache entries, this inp will
896 * no longer be available to the rest of the protocol threads, so we
897 * are safe to whack the inp in the following code.
899 if ((inp
->inp_flags
& INP_WILDCARD
) && netisr_ncpus
> 1) {
900 struct netmsg_listen_detach nmsg
;
902 KKASSERT(so
->so_port
== netisr_cpuport(0));
904 KKASSERT(inp
->inp_pcbinfo
== &tcbinfo
[0]);
906 netmsg_init(&nmsg
.base
, NULL
, &curthread
->td_msgport
,
907 MSGF_PRIORITY
, tcp_listen_detach_handler
);
909 nmsg
.nm_tp_inh
= tp_inh
;
910 lwkt_domsg(netisr_cpuport(1), &nmsg
.base
.lmsg
, 0);
916 * Make sure that all of our timers are stopped before we
917 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
918 * timers are never used. If timer message is never created
919 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
921 if (tp
->tt_msg
!= NULL
&& tp
->tt_msg
->tt_tcb
!= NULL
) {
922 tcp_callout_stop(tp
, tp
->tt_rexmt
);
923 tcp_callout_stop(tp
, tp
->tt_persist
);
924 tcp_callout_stop(tp
, tp
->tt_keep
);
925 tcp_callout_stop(tp
, tp
->tt_2msl
);
926 tcp_callout_stop(tp
, tp
->tt_delack
);
929 if (tp
->t_flags
& TF_ONOUTPUTQ
) {
930 KKASSERT(tp
->tt_cpu
== mycpu
->gd_cpuid
);
931 TAILQ_REMOVE(&tcpcbackq
[tp
->tt_cpu
].head
, tp
, t_outputq
);
932 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
936 * If we got enough samples through the srtt filter,
937 * save the rtt and rttvar in the routing entry.
938 * 'Enough' is arbitrarily defined as the 16 samples.
939 * 16 samples is enough for the srtt filter to converge
940 * to within 5% of the correct value; fewer samples and
941 * we could save a very bogus rtt.
943 * Don't update the default route's characteristics and don't
944 * update anything that the user "locked".
946 if (tp
->t_rttupdated
>= 16) {
950 struct sockaddr_in6
*sin6
;
952 if ((rt
= inp
->in6p_route
.ro_rt
) == NULL
)
954 sin6
= (struct sockaddr_in6
*)rt_key(rt
);
955 if (IN6_IS_ADDR_UNSPECIFIED(&sin6
->sin6_addr
))
958 if ((rt
= inp
->inp_route
.ro_rt
) == NULL
||
959 ((struct sockaddr_in
*)rt_key(rt
))->
960 sin_addr
.s_addr
== INADDR_ANY
)
963 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTT
)) {
964 i
= tp
->t_srtt
* (RTM_RTTUNIT
/ (hz
* TCP_RTT_SCALE
));
965 if (rt
->rt_rmx
.rmx_rtt
&& i
)
967 * filter this update to half the old & half
968 * the new values, converting scale.
969 * See route.h and tcp_var.h for a
970 * description of the scaling constants.
973 (rt
->rt_rmx
.rmx_rtt
+ i
) / 2;
975 rt
->rt_rmx
.rmx_rtt
= i
;
976 tcpstat
.tcps_cachedrtt
++;
978 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTTVAR
)) {
980 (RTM_RTTUNIT
/ (hz
* TCP_RTTVAR_SCALE
));
981 if (rt
->rt_rmx
.rmx_rttvar
&& i
)
982 rt
->rt_rmx
.rmx_rttvar
=
983 (rt
->rt_rmx
.rmx_rttvar
+ i
) / 2;
985 rt
->rt_rmx
.rmx_rttvar
= i
;
986 tcpstat
.tcps_cachedrttvar
++;
989 * The old comment here said:
990 * update the pipelimit (ssthresh) if it has been updated
991 * already or if a pipesize was specified & the threshhold
992 * got below half the pipesize. I.e., wait for bad news
993 * before we start updating, then update on both good
996 * But we want to save the ssthresh even if no pipesize is
997 * specified explicitly in the route, because such
998 * connections still have an implicit pipesize specified
999 * by the global tcp_sendspace. In the absence of a reliable
1000 * way to calculate the pipesize, it will have to do.
1002 i
= tp
->snd_ssthresh
;
1003 if (rt
->rt_rmx
.rmx_sendpipe
!= 0)
1004 dosavessthresh
= (i
< rt
->rt_rmx
.rmx_sendpipe
/2);
1006 dosavessthresh
= (i
< so
->so_snd
.ssb_hiwat
/2);
1007 if (dosavessthresh
||
1008 (!(rt
->rt_rmx
.rmx_locks
& RTV_SSTHRESH
) && (i
!= 0) &&
1009 (rt
->rt_rmx
.rmx_ssthresh
!= 0))) {
1011 * convert the limit from user data bytes to
1012 * packets then to packet data bytes.
1014 i
= (i
+ tp
->t_maxseg
/ 2) / tp
->t_maxseg
;
1019 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1020 sizeof(struct tcpiphdr
));
1021 if (rt
->rt_rmx
.rmx_ssthresh
)
1022 rt
->rt_rmx
.rmx_ssthresh
=
1023 (rt
->rt_rmx
.rmx_ssthresh
+ i
) / 2;
1025 rt
->rt_rmx
.rmx_ssthresh
= i
;
1026 tcpstat
.tcps_cachedssthresh
++;
1031 /* free the reassembly queue, if any */
1032 while((q
= TAILQ_FIRST(&tp
->t_segq
)) != NULL
) {
1033 TAILQ_REMOVE(&tp
->t_segq
, q
, tqe_q
);
1036 atomic_add_int(&tcp_reass_qsize
, -1);
1038 /* throw away SACK blocks in scoreboard*/
1039 if (TCP_DO_SACK(tp
))
1040 tcp_sack_destroy(&tp
->scb
);
1042 inp
->inp_ppcb
= NULL
;
1043 soisdisconnected(so
);
1044 /* note: pcb detached later on */
1046 tcp_destroy_timermsg(tp
);
1047 tcp_output_cancel(tp
);
1049 if (tp
->t_flags
& TF_LISTEN
) {
1050 syncache_destroy(tp
, tp_inh
);
1051 tcp_pcbport_merge_oncpu(tp
);
1052 tcp_pcbport_destroy(tp
);
1053 if (inp_inh
!= NULL
&& inp_inh
->inp_socket
!= NULL
) {
1055 * Pending sockets inheritance only needs
1056 * to be done once in the current thread,
1059 soinherit(so
, inp_inh
->inp_socket
);
1062 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache is not destroyed"));
1064 so_async_rcvd_drop(so
);
1065 /* Drop the reference for the asynchronized pru_rcvd */
1070 * - Remove self from listen tcpcb per-cpu port cache _before_
1072 * - pcbdetach removes any wildcard hash entry on the current CPU.
1074 tcp_pcbport_remove(inp
);
1082 tcpstat
.tcps_closed
++;
1086 static __inline
void
1087 tcp_drain_oncpu(struct inpcbinfo
*pcbinfo
)
1089 struct inpcbhead
*head
= &pcbinfo
->pcblisthead
;
1093 * Since we run in netisr, it is MP safe, even if
1094 * we block during the inpcb list iteration, i.e.
1095 * we don't need to use inpcb marker here.
1097 ASSERT_IN_NETISR(pcbinfo
->cpu
);
1099 LIST_FOREACH(inpb
, head
, inp_list
) {
1101 struct tseg_qent
*te
;
1103 if (inpb
->inp_flags
& INP_PLACEMARKER
)
1106 tcpb
= intotcpcb(inpb
);
1107 KASSERT(tcpb
!= NULL
, ("tcp_drain_oncpu: tcpb is NULL"));
1109 if ((te
= TAILQ_FIRST(&tcpb
->t_segq
)) != NULL
) {
1110 TAILQ_REMOVE(&tcpb
->t_segq
, te
, tqe_q
);
1111 if (te
->tqe_th
->th_flags
& TH_FIN
)
1112 tcpb
->t_flags
&= ~TF_QUEDFIN
;
1115 atomic_add_int(&tcp_reass_qsize
, -1);
1122 tcp_drain_dispatch(netmsg_t nmsg
)
1125 lwkt_replymsg(&nmsg
->lmsg
, 0); /* reply ASAP */
1128 tcp_drain_oncpu(&tcbinfo
[mycpuid
]);
1132 tcp_drain_ipi(void *arg __unused
)
1135 struct lwkt_msg
*msg
= &tcp_drain_netmsg
[cpu
].lmsg
;
1138 if (msg
->ms_flags
& MSGF_DONE
)
1139 lwkt_sendmsg_oncpu(netisr_cpuport(cpu
), msg
);
1152 * Walk the tcpbs, if existing, and flush the reassembly queue,
1153 * if there is one...
1154 * XXX: The "Net/3" implementation doesn't imply that the TCP
1155 * reassembly queue should be flushed, but in a situation
1156 * where we're really low on mbufs, this is potentially
1158 * YYY: We may consider run tcp_drain_oncpu directly here,
1159 * however, that will require M_WAITOK memory allocation
1160 * for the inpcb marker.
1162 CPUMASK_ASSBMASK(mask
, netisr_ncpus
);
1163 CPUMASK_ANDMASK(mask
, smp_active_mask
);
1164 if (CPUMASK_TESTNZERO(mask
))
1165 lwkt_send_ipiq_mask(mask
, tcp_drain_ipi
, NULL
);
1169 * Notify a tcp user of an asynchronous error;
1170 * store error as soft error, but wake up user
1171 * (for now, won't do anything until can select for soft error).
1173 * Do not wake up user since there currently is no mechanism for
1174 * reporting soft errors (yet - a kqueue filter may be added).
1177 tcp_notify(struct inpcb
*inp
, int error
)
1179 struct tcpcb
*tp
= intotcpcb(inp
);
1182 * Ignore some errors if we are hooked up.
1183 * If connection hasn't completed, has retransmitted several times,
1184 * and receives a second error, give up now. This is better
1185 * than waiting a long time to establish a connection that
1186 * can never complete.
1188 if (tp
->t_state
== TCPS_ESTABLISHED
&&
1189 (error
== EHOSTUNREACH
|| error
== ENETUNREACH
||
1190 error
== EHOSTDOWN
)) {
1192 } else if (tp
->t_state
< TCPS_ESTABLISHED
&& tp
->t_rxtshift
> 3 &&
1194 tcp_drop(tp
, error
);
1196 tp
->t_softerror
= error
;
1198 wakeup(&so
->so_timeo
);
1205 tcp_pcblist(SYSCTL_HANDLER_ARGS
)
1208 struct inpcb
*marker
;
1216 * The process of preparing the TCB list is too time-consuming and
1217 * resource-intensive to repeat twice on every request.
1219 if (req
->oldptr
== NULL
) {
1220 for (ccpu
= 0; ccpu
< netisr_ncpus
; ++ccpu
)
1221 n
+= tcbinfo
[ccpu
].ipi_count
;
1222 req
->oldidx
= (n
+ n
/8 + 10) * sizeof(struct xtcpcb
);
1226 if (req
->newptr
!= NULL
)
1229 marker
= kmalloc(sizeof(struct inpcb
), M_TEMP
, M_WAITOK
|M_ZERO
);
1230 marker
->inp_flags
|= INP_PLACEMARKER
;
1233 * OK, now we're committed to doing something. Run the inpcb list
1234 * for each cpu in the system and construct the output. Use a
1235 * list placemarker to deal with list changes occuring during
1236 * copyout blockages (but otherwise depend on being on the correct
1237 * cpu to avoid races).
1239 origcpu
= mycpu
->gd_cpuid
;
1240 for (ccpu
= 0; ccpu
< netisr_ncpus
&& error
== 0; ++ccpu
) {
1244 lwkt_migratecpu(ccpu
);
1246 n
= tcbinfo
[ccpu
].ipi_count
;
1248 LIST_INSERT_HEAD(&tcbinfo
[ccpu
].pcblisthead
, marker
, inp_list
);
1250 while ((inp
= LIST_NEXT(marker
, inp_list
)) != NULL
&& i
< n
) {
1252 * process a snapshot of pcbs, ignoring placemarkers
1253 * and using our own to allow SYSCTL_OUT to block.
1255 LIST_REMOVE(marker
, inp_list
);
1256 LIST_INSERT_AFTER(inp
, marker
, inp_list
);
1258 if (inp
->inp_flags
& INP_PLACEMARKER
)
1260 if (prison_xinpcb(req
->td
, inp
))
1263 xt
.xt_len
= sizeof xt
;
1264 bcopy(inp
, &xt
.xt_inp
, sizeof *inp
);
1265 inp_ppcb
= inp
->inp_ppcb
;
1266 if (inp_ppcb
!= NULL
)
1267 bcopy(inp_ppcb
, &xt
.xt_tp
, sizeof xt
.xt_tp
);
1269 bzero(&xt
.xt_tp
, sizeof xt
.xt_tp
);
1270 if (inp
->inp_socket
)
1271 sotoxsocket(inp
->inp_socket
, &xt
.xt_socket
);
1272 if ((error
= SYSCTL_OUT(req
, &xt
, sizeof xt
)) != 0)
1276 LIST_REMOVE(marker
, inp_list
);
1277 if (error
== 0 && i
< n
) {
1278 bzero(&xt
, sizeof xt
);
1279 xt
.xt_len
= sizeof xt
;
1281 error
= SYSCTL_OUT(req
, &xt
, sizeof xt
);
1290 * Make sure we are on the same cpu we were on originally, since
1291 * higher level callers expect this. Also don't pollute caches with
1292 * migrated userland data by (eventually) returning to userland
1293 * on a different cpu.
1295 lwkt_migratecpu(origcpu
);
1296 kfree(marker
, M_TEMP
);
1300 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_PCBLIST
, pcblist
, CTLFLAG_RD
, 0, 0,
1301 tcp_pcblist
, "S,xtcpcb", "List of active TCP connections");
1304 tcp_getcred(SYSCTL_HANDLER_ARGS
)
1306 struct sockaddr_in addrs
[2];
1307 struct ucred cred0
, *cred
= NULL
;
1309 int cpu
, origcpu
, error
;
1311 error
= priv_check(req
->td
, PRIV_ROOT
);
1314 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1319 cpu
= tcp_addrcpu(addrs
[1].sin_addr
.s_addr
, addrs
[1].sin_port
,
1320 addrs
[0].sin_addr
.s_addr
, addrs
[0].sin_port
);
1322 lwkt_migratecpu(cpu
);
1324 inp
= in_pcblookup_hash(&tcbinfo
[cpu
], addrs
[1].sin_addr
,
1325 addrs
[1].sin_port
, addrs
[0].sin_addr
, addrs
[0].sin_port
, 0, NULL
);
1326 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1328 } else if (inp
->inp_socket
->so_cred
!= NULL
) {
1329 cred0
= *(inp
->inp_socket
->so_cred
);
1333 lwkt_migratecpu(origcpu
);
1338 return SYSCTL_OUT(req
, cred
, sizeof(struct ucred
));
1341 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1342 0, 0, tcp_getcred
, "S,ucred", "Get the ucred of a TCP connection");
1346 tcp6_getcred(SYSCTL_HANDLER_ARGS
)
1348 struct sockaddr_in6 addrs
[2];
1352 error
= priv_check(req
->td
, PRIV_ROOT
);
1355 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1359 inp
= in6_pcblookup_hash(&tcbinfo
[0],
1360 &addrs
[1].sin6_addr
, addrs
[1].sin6_port
,
1361 &addrs
[0].sin6_addr
, addrs
[0].sin6_port
, 0, NULL
);
1362 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1366 error
= SYSCTL_OUT(req
, inp
->inp_socket
->so_cred
, sizeof(struct ucred
));
1372 SYSCTL_PROC(_net_inet6_tcp6
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1374 tcp6_getcred
, "S,ucred", "Get the ucred of a TCP6 connection");
1377 struct netmsg_tcp_notify
{
1378 struct netmsg_base base
;
1379 inp_notify_t nm_notify
;
1380 struct in_addr nm_faddr
;
1385 tcp_notifyall_oncpu(netmsg_t msg
)
1387 struct netmsg_tcp_notify
*nm
= (struct netmsg_tcp_notify
*)msg
;
1390 in_pcbnotifyall(&tcbinfo
[mycpuid
], nm
->nm_faddr
,
1391 nm
->nm_arg
, nm
->nm_notify
);
1393 nextcpu
= mycpuid
+ 1;
1394 if (nextcpu
< netisr_ncpus
)
1395 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nm
->base
.lmsg
);
1397 lwkt_replymsg(&nm
->base
.lmsg
, 0);
1401 tcp_get_inpnotify(int cmd
, const struct sockaddr
*sa
,
1402 int *arg
, struct ip
**ip0
, int *cpuid
)
1404 struct ip
*ip
= *ip0
;
1405 struct in_addr faddr
;
1406 inp_notify_t notify
= tcp_notify
;
1408 faddr
= ((const struct sockaddr_in
*)sa
)->sin_addr
;
1409 if (sa
->sa_family
!= AF_INET
|| faddr
.s_addr
== INADDR_ANY
)
1412 *arg
= inetctlerrmap
[cmd
];
1413 if (cmd
== PRC_QUENCH
) {
1414 notify
= tcp_quench
;
1415 } else if (icmp_may_rst
&&
1416 (cmd
== PRC_UNREACH_ADMIN_PROHIB
||
1417 cmd
== PRC_UNREACH_PORT
||
1418 cmd
== PRC_TIMXCEED_INTRANS
) &&
1420 notify
= tcp_drop_syn_sent
;
1421 } else if (cmd
== PRC_MSGSIZE
) {
1422 const struct icmp
*icmp
= (const struct icmp
*)
1423 ((caddr_t
)ip
- offsetof(struct icmp
, icmp_ip
));
1425 *arg
= ntohs(icmp
->icmp_nextmtu
);
1426 notify
= tcp_mtudisc
;
1427 } else if (PRC_IS_REDIRECT(cmd
)) {
1429 notify
= in_rtchange
;
1430 } else if (cmd
== PRC_HOSTDEAD
) {
1432 } else if ((unsigned)cmd
>= PRC_NCMDS
|| inetctlerrmap
[cmd
] == 0) {
1436 if (cpuid
!= NULL
) {
1438 /* Go through all CPUs */
1441 const struct tcphdr
*th
;
1443 th
= (const struct tcphdr
*)
1444 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1445 *cpuid
= tcp_addrcpu(faddr
.s_addr
, th
->th_dport
,
1446 ip
->ip_src
.s_addr
, th
->th_sport
);
1455 tcp_ctlinput(netmsg_t msg
)
1457 int cmd
= msg
->ctlinput
.nm_cmd
;
1458 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1459 struct ip
*ip
= msg
->ctlinput
.nm_extra
;
1460 struct in_addr faddr
;
1461 inp_notify_t notify
;
1464 notify
= tcp_get_inpnotify(cmd
, sa
, &arg
, &ip
, &cpuid
);
1468 faddr
= ((struct sockaddr_in
*)sa
)->sin_addr
;
1470 const struct tcphdr
*th
;
1473 if (cpuid
!= mycpuid
)
1476 th
= (const struct tcphdr
*)
1477 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1478 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], faddr
, th
->th_dport
,
1479 ip
->ip_src
, th
->th_sport
, 0, NULL
);
1480 if (inp
!= NULL
&& inp
->inp_socket
!= NULL
) {
1481 tcp_seq icmpseq
= htonl(th
->th_seq
);
1482 struct tcpcb
*tp
= intotcpcb(inp
);
1484 if (SEQ_GEQ(icmpseq
, tp
->snd_una
) &&
1485 SEQ_LT(icmpseq
, tp
->snd_max
))
1488 struct in_conninfo inc
;
1490 inc
.inc_fport
= th
->th_dport
;
1491 inc
.inc_lport
= th
->th_sport
;
1492 inc
.inc_faddr
= faddr
;
1493 inc
.inc_laddr
= ip
->ip_src
;
1497 syncache_unreach(&inc
, th
);
1499 } else if (msg
->ctlinput
.nm_direct
) {
1500 if (cpuid
!= ncpus
&& cpuid
!= mycpuid
)
1502 if (mycpuid
>= netisr_ncpus
)
1505 in_pcbnotifyall(&tcbinfo
[mycpuid
], faddr
, arg
, notify
);
1507 struct netmsg_tcp_notify
*nm
;
1509 ASSERT_IN_NETISR(0);
1510 nm
= kmalloc(sizeof(*nm
), M_LWKTMSG
, M_INTWAIT
);
1511 netmsg_init(&nm
->base
, NULL
, &netisr_afree_rport
,
1512 0, tcp_notifyall_oncpu
);
1513 nm
->nm_faddr
= faddr
;
1515 nm
->nm_notify
= notify
;
1517 lwkt_sendmsg(netisr_cpuport(0), &nm
->base
.lmsg
);
1520 lwkt_replymsg(&msg
->lmsg
, 0);
1526 tcp6_ctlinput(netmsg_t msg
)
1528 int cmd
= msg
->ctlinput
.nm_cmd
;
1529 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1530 void *d
= msg
->ctlinput
.nm_extra
;
1532 inp_notify_t notify
= tcp_notify
;
1533 struct ip6_hdr
*ip6
;
1535 struct ip6ctlparam
*ip6cp
= NULL
;
1536 const struct sockaddr_in6
*sa6_src
= NULL
;
1538 struct tcp_portonly
{
1544 if (sa
->sa_family
!= AF_INET6
||
1545 sa
->sa_len
!= sizeof(struct sockaddr_in6
)) {
1550 if (cmd
== PRC_QUENCH
)
1551 notify
= tcp_quench
;
1552 else if (cmd
== PRC_MSGSIZE
) {
1553 struct ip6ctlparam
*ip6cp
= d
;
1554 struct icmp6_hdr
*icmp6
= ip6cp
->ip6c_icmp6
;
1556 arg
= ntohl(icmp6
->icmp6_mtu
);
1557 notify
= tcp_mtudisc
;
1558 } else if (!PRC_IS_REDIRECT(cmd
) &&
1559 ((unsigned)cmd
> PRC_NCMDS
|| inet6ctlerrmap
[cmd
] == 0)) {
1563 /* if the parameter is from icmp6, decode it. */
1565 ip6cp
= (struct ip6ctlparam
*)d
;
1567 ip6
= ip6cp
->ip6c_ip6
;
1568 off
= ip6cp
->ip6c_off
;
1569 sa6_src
= ip6cp
->ip6c_src
;
1573 off
= 0; /* fool gcc */
1578 struct in_conninfo inc
;
1580 * XXX: We assume that when IPV6 is non NULL,
1581 * M and OFF are valid.
1584 /* check if we can safely examine src and dst ports */
1585 if (m
->m_pkthdr
.len
< off
+ sizeof *thp
)
1588 bzero(&th
, sizeof th
);
1589 m_copydata(m
, off
, sizeof *thp
, (caddr_t
)&th
);
1591 in6_pcbnotify(&tcbinfo
[0], sa
, th
.th_dport
,
1592 (struct sockaddr
*)ip6cp
->ip6c_src
,
1593 th
.th_sport
, cmd
, arg
, notify
);
1595 inc
.inc_fport
= th
.th_dport
;
1596 inc
.inc_lport
= th
.th_sport
;
1597 inc
.inc6_faddr
= ((struct sockaddr_in6
*)sa
)->sin6_addr
;
1598 inc
.inc6_laddr
= ip6cp
->ip6c_src
->sin6_addr
;
1600 syncache_unreach(&inc
, &th
);
1602 in6_pcbnotify(&tcbinfo
[0], sa
, 0,
1603 (const struct sockaddr
*)sa6_src
, 0, cmd
, arg
, notify
);
1606 lwkt_replymsg(&msg
->ctlinput
.base
.lmsg
, 0);
1612 * Following is where TCP initial sequence number generation occurs.
1614 * There are two places where we must use initial sequence numbers:
1615 * 1. In SYN-ACK packets.
1616 * 2. In SYN packets.
1618 * All ISNs for SYN-ACK packets are generated by the syncache. See
1619 * tcp_syncache.c for details.
1621 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1622 * depends on this property. In addition, these ISNs should be
1623 * unguessable so as to prevent connection hijacking. To satisfy
1624 * the requirements of this situation, the algorithm outlined in
1625 * RFC 1948 is used to generate sequence numbers.
1627 * Implementation details:
1629 * Time is based off the system timer, and is corrected so that it
1630 * increases by one megabyte per second. This allows for proper
1631 * recycling on high speed LANs while still leaving over an hour
1634 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1635 * between seeding of isn_secret. This is normally set to zero,
1636 * as reseeding should not be necessary.
1640 #define ISN_BYTES_PER_SECOND 1048576
1642 u_char isn_secret
[32];
1643 int isn_last_reseed
;
1647 tcp_new_isn(struct tcpcb
*tp
)
1649 u_int32_t md5_buffer
[4];
1652 /* Seed if this is the first use, reseed if requested. */
1653 if ((isn_last_reseed
== 0) || ((tcp_isn_reseed_interval
> 0) &&
1654 (((u_int
)isn_last_reseed
+ (u_int
)tcp_isn_reseed_interval
*hz
)
1656 read_random_unlimited(&isn_secret
, sizeof isn_secret
);
1657 isn_last_reseed
= ticks
;
1660 /* Compute the md5 hash and return the ISN. */
1662 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_fport
, sizeof(u_short
));
1663 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_lport
, sizeof(u_short
));
1665 if (INP_ISIPV6(tp
->t_inpcb
)) {
1666 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_faddr
,
1667 sizeof(struct in6_addr
));
1668 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_laddr
,
1669 sizeof(struct in6_addr
));
1673 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_faddr
,
1674 sizeof(struct in_addr
));
1675 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_laddr
,
1676 sizeof(struct in_addr
));
1678 MD5Update(&isn_ctx
, (u_char
*) &isn_secret
, sizeof(isn_secret
));
1679 MD5Final((u_char
*) &md5_buffer
, &isn_ctx
);
1680 new_isn
= (tcp_seq
) md5_buffer
[0];
1681 new_isn
+= ticks
* (ISN_BYTES_PER_SECOND
/ hz
);
1686 * When a source quench is received, close congestion window
1687 * to one segment. We will gradually open it again as we proceed.
1690 tcp_quench(struct inpcb
*inp
, int error
)
1692 struct tcpcb
*tp
= intotcpcb(inp
);
1694 KASSERT(tp
!= NULL
, ("tcp_quench: tp is NULL"));
1695 tp
->snd_cwnd
= tp
->t_maxseg
;
1700 * When a specific ICMP unreachable message is received and the
1701 * connection state is SYN-SENT, drop the connection. This behavior
1702 * is controlled by the icmp_may_rst sysctl.
1705 tcp_drop_syn_sent(struct inpcb
*inp
, int error
)
1707 struct tcpcb
*tp
= intotcpcb(inp
);
1709 KASSERT(tp
!= NULL
, ("tcp_drop_syn_sent: tp is NULL"));
1710 if (tp
->t_state
== TCPS_SYN_SENT
)
1711 tcp_drop(tp
, error
);
1715 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1716 * based on the new value in the route. Also nudge TCP to send something,
1717 * since we know the packet we just sent was dropped.
1718 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1721 tcp_mtudisc(struct inpcb
*inp
, int mtu
)
1723 struct tcpcb
*tp
= intotcpcb(inp
);
1725 struct socket
*so
= inp
->inp_socket
;
1728 boolean_t isipv6
= INP_ISIPV6(inp
);
1730 const boolean_t isipv6
= FALSE
;
1733 KASSERT(tp
!= NULL
, ("tcp_mtudisc: tp is NULL"));
1736 * If no MTU is provided in the ICMP message, use the
1737 * next lower likely value, as specified in RFC 1191.
1742 oldmtu
= tp
->t_maxopd
+
1744 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1745 sizeof(struct tcpiphdr
));
1746 mtu
= ip_next_mtu(oldmtu
, 0);
1750 rt
= tcp_rtlookup6(&inp
->inp_inc
);
1752 rt
= tcp_rtlookup(&inp
->inp_inc
);
1754 if (rt
->rt_rmx
.rmx_mtu
!= 0 && rt
->rt_rmx
.rmx_mtu
< mtu
)
1755 mtu
= rt
->rt_rmx
.rmx_mtu
;
1759 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1760 sizeof(struct tcpiphdr
));
1763 * XXX - The following conditional probably violates the TCP
1764 * spec. The problem is that, since we don't know the
1765 * other end's MSS, we are supposed to use a conservative
1766 * default. But, if we do that, then MTU discovery will
1767 * never actually take place, because the conservative
1768 * default is much less than the MTUs typically seen
1769 * on the Internet today. For the moment, we'll sweep
1770 * this under the carpet.
1772 * The conservative default might not actually be a problem
1773 * if the only case this occurs is when sending an initial
1774 * SYN with options and data to a host we've never talked
1775 * to before. Then, they will reply with an MSS value which
1776 * will get recorded and the new parameters should get
1777 * recomputed. For Further Study.
1779 if (rt
->rt_rmx
.rmx_mssopt
&& rt
->rt_rmx
.rmx_mssopt
< maxopd
)
1780 maxopd
= rt
->rt_rmx
.rmx_mssopt
;
1784 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1785 sizeof(struct tcpiphdr
));
1787 if (tp
->t_maxopd
<= maxopd
)
1789 tp
->t_maxopd
= maxopd
;
1792 if ((tp
->t_flags
& (TF_REQ_TSTMP
| TF_RCVD_TSTMP
| TF_NOOPT
)) ==
1793 (TF_REQ_TSTMP
| TF_RCVD_TSTMP
))
1794 mss
-= TCPOLEN_TSTAMP_APPA
;
1796 /* round down to multiple of MCLBYTES */
1797 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1799 mss
&= ~(MCLBYTES
- 1);
1802 mss
= (mss
/ MCLBYTES
) * MCLBYTES
;
1805 if (so
->so_snd
.ssb_hiwat
< mss
)
1806 mss
= so
->so_snd
.ssb_hiwat
;
1810 tp
->snd_nxt
= tp
->snd_una
;
1812 tcpstat
.tcps_mturesent
++;
1816 * Look-up the routing entry to the peer of this inpcb. If no route
1817 * is found and it cannot be allocated the return NULL. This routine
1818 * is called by TCP routines that access the rmx structure and by tcp_mss
1819 * to get the interface MTU.
1822 tcp_rtlookup(struct in_conninfo
*inc
)
1824 struct route
*ro
= &inc
->inc_route
;
1826 if (ro
->ro_rt
== NULL
|| !(ro
->ro_rt
->rt_flags
& RTF_UP
)) {
1827 /* No route yet, so try to acquire one */
1828 if (inc
->inc_faddr
.s_addr
!= INADDR_ANY
) {
1830 * unused portions of the structure MUST be zero'd
1831 * out because rtalloc() treats it as opaque data
1833 bzero(&ro
->ro_dst
, sizeof(struct sockaddr_in
));
1834 ro
->ro_dst
.sa_family
= AF_INET
;
1835 ro
->ro_dst
.sa_len
= sizeof(struct sockaddr_in
);
1836 ((struct sockaddr_in
*) &ro
->ro_dst
)->sin_addr
=
1846 tcp_rtlookup6(struct in_conninfo
*inc
)
1848 struct route_in6
*ro6
= &inc
->inc6_route
;
1850 if (ro6
->ro_rt
== NULL
|| !(ro6
->ro_rt
->rt_flags
& RTF_UP
)) {
1851 /* No route yet, so try to acquire one */
1852 if (!IN6_IS_ADDR_UNSPECIFIED(&inc
->inc6_faddr
)) {
1854 * unused portions of the structure MUST be zero'd
1855 * out because rtalloc() treats it as opaque data
1857 bzero(&ro6
->ro_dst
, sizeof(struct sockaddr_in6
));
1858 ro6
->ro_dst
.sin6_family
= AF_INET6
;
1859 ro6
->ro_dst
.sin6_len
= sizeof(struct sockaddr_in6
);
1860 ro6
->ro_dst
.sin6_addr
= inc
->inc6_faddr
;
1861 rtalloc((struct route
*)ro6
);
1864 return (ro6
->ro_rt
);
1869 /* compute ESP/AH header size for TCP, including outer IP header. */
1871 ipsec_hdrsiz_tcp(struct tcpcb
*tp
)
1879 if ((tp
== NULL
) || ((inp
= tp
->t_inpcb
) == NULL
))
1881 MGETHDR(m
, M_NOWAIT
, MT_DATA
);
1886 if (INP_ISIPV6(inp
)) {
1887 struct ip6_hdr
*ip6
= mtod(m
, struct ip6_hdr
*);
1889 th
= (struct tcphdr
*)(ip6
+ 1);
1890 m
->m_pkthdr
.len
= m
->m_len
=
1891 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
);
1892 tcp_fillheaders(tp
, ip6
, th
, FALSE
);
1893 hdrsiz
= ipsec6_hdrsiz(m
, IPSEC_DIR_OUTBOUND
, inp
);
1897 ip
= mtod(m
, struct ip
*);
1898 th
= (struct tcphdr
*)(ip
+ 1);
1899 m
->m_pkthdr
.len
= m
->m_len
= sizeof(struct tcpiphdr
);
1900 tcp_fillheaders(tp
, ip
, th
, FALSE
);
1901 hdrsiz
= ipsec4_hdrsiz(m
, IPSEC_DIR_OUTBOUND
, inp
);
1910 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1912 * This code attempts to calculate the bandwidth-delay product as a
1913 * means of determining the optimal window size to maximize bandwidth,
1914 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1915 * routers. This code also does a fairly good job keeping RTTs in check
1916 * across slow links like modems. We implement an algorithm which is very
1917 * similar (but not meant to be) TCP/Vegas. The code operates on the
1918 * transmitter side of a TCP connection and so only effects the transmit
1919 * side of the connection.
1921 * BACKGROUND: TCP makes no provision for the management of buffer space
1922 * at the end points or at the intermediate routers and switches. A TCP
1923 * stream, whether using NewReno or not, will eventually buffer as
1924 * many packets as it is able and the only reason this typically works is
1925 * due to the fairly small default buffers made available for a connection
1926 * (typicaly 16K or 32K). As machines use larger windows and/or window
1927 * scaling it is now fairly easy for even a single TCP connection to blow-out
1928 * all available buffer space not only on the local interface, but on
1929 * intermediate routers and switches as well. NewReno makes a misguided
1930 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1931 * then backing off, then steadily increasing the window again until another
1932 * failure occurs, ad-infinitum. This results in terrible oscillation that
1933 * is only made worse as network loads increase and the idea of intentionally
1934 * blowing out network buffers is, frankly, a terrible way to manage network
1937 * It is far better to limit the transmit window prior to the failure
1938 * condition being achieved. There are two general ways to do this: First
1939 * you can 'scan' through different transmit window sizes and locate the
1940 * point where the RTT stops increasing, indicating that you have filled the
1941 * pipe, then scan backwards until you note that RTT stops decreasing, then
1942 * repeat ad-infinitum. This method works in principle but has severe
1943 * implementation issues due to RTT variances, timer granularity, and
1944 * instability in the algorithm which can lead to many false positives and
1945 * create oscillations as well as interact badly with other TCP streams
1946 * implementing the same algorithm.
1948 * The second method is to limit the window to the bandwidth delay product
1949 * of the link. This is the method we implement. RTT variances and our
1950 * own manipulation of the congestion window, bwnd, can potentially
1951 * destabilize the algorithm. For this reason we have to stabilize the
1952 * elements used to calculate the window. We do this by using the minimum
1953 * observed RTT, the long term average of the observed bandwidth, and
1954 * by adding two segments worth of slop. It isn't perfect but it is able
1955 * to react to changing conditions and gives us a very stable basis on
1956 * which to extend the algorithm.
1959 tcp_xmit_bandwidth_limit(struct tcpcb
*tp
, tcp_seq ack_seq
)
1968 * If inflight_enable is disabled in the middle of a tcp connection,
1969 * make sure snd_bwnd is effectively disabled.
1971 if (!tcp_inflight_enable
) {
1972 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
1973 tp
->snd_bandwidth
= 0;
1978 * Validate the delta time. If a connection is new or has been idle
1979 * a long time we have to reset the bandwidth calculator.
1983 delta_ticks
= save_ticks
- tp
->t_bw_rtttime
;
1984 if (tp
->t_bw_rtttime
== 0 || delta_ticks
< 0 || delta_ticks
> hz
* 10) {
1985 tp
->t_bw_rtttime
= save_ticks
;
1986 tp
->t_bw_rtseq
= ack_seq
;
1987 if (tp
->snd_bandwidth
== 0)
1988 tp
->snd_bandwidth
= tcp_inflight_start
;
1993 * A delta of at least 1 tick is required. Waiting 2 ticks will
1994 * result in better (bw) accuracy. More than that and the ramp-up
1997 if (delta_ticks
== 0 || delta_ticks
== 1)
2001 * Sanity check, plus ignore pure window update acks.
2003 if ((int)(ack_seq
- tp
->t_bw_rtseq
) <= 0)
2007 * Figure out the bandwidth. Due to the tick granularity this
2008 * is a very rough number and it MUST be averaged over a fairly
2009 * long period of time. XXX we need to take into account a link
2010 * that is not using all available bandwidth, but for now our
2011 * slop will ramp us up if this case occurs and the bandwidth later
2014 ibw
= (int64_t)(ack_seq
- tp
->t_bw_rtseq
) * hz
/ delta_ticks
;
2015 tp
->t_bw_rtttime
= save_ticks
;
2016 tp
->t_bw_rtseq
= ack_seq
;
2017 bw
= ((int64_t)tp
->snd_bandwidth
* 15 + ibw
) >> 4;
2019 tp
->snd_bandwidth
= bw
;
2022 * Calculate the semi-static bandwidth delay product, plus two maximal
2023 * segments. The additional slop puts us squarely in the sweet
2024 * spot and also handles the bandwidth run-up case. Without the
2025 * slop we could be locking ourselves into a lower bandwidth.
2027 * At very high speeds the bw calculation can become overly sensitive
2028 * and error prone when delta_ticks is low (e.g. usually 1). To deal
2029 * with the problem the stab must be scaled to the bw. A stab of 50
2030 * (the default) increases the bw for the purposes of the bwnd
2031 * calculation by 5%.
2033 * Situations Handled:
2034 * (1) Prevents over-queueing of packets on LANs, especially on
2035 * high speed LANs, allowing larger TCP buffers to be
2036 * specified, and also does a good job preventing
2037 * over-queueing of packets over choke points like modems
2038 * (at least for the transmit side).
2040 * (2) Is able to handle changing network loads (bandwidth
2041 * drops so bwnd drops, bandwidth increases so bwnd
2044 * (3) Theoretically should stabilize in the face of multiple
2045 * connections implementing the same algorithm (this may need
2048 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2049 * be adjusted with a sysctl but typically only needs to be on
2050 * very slow connections. A value no smaller then 5 should
2051 * be used, but only reduce this default if you have no other
2055 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2056 bw
+= bw
* tcp_inflight_stab
/ 1000;
2057 bwnd
= (int64_t)bw
* USERTT
/ (hz
<< TCP_RTT_SHIFT
) +
2058 (int)tp
->t_maxseg
* 2;
2061 if (tcp_inflight_debug
> 0) {
2063 if ((u_int
)(save_ticks
- ltime
) >= hz
/ tcp_inflight_debug
) {
2065 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2066 "bwnd %ld delta %d snd_win %ld\n",
2067 tp
, ibw
, bw
, tp
->t_rttvar
, tp
->t_srtt
,
2068 bwnd
, delta_ticks
, tp
->snd_wnd
);
2071 if ((long)bwnd
< tcp_inflight_min
)
2072 bwnd
= tcp_inflight_min
;
2073 if (bwnd
> tcp_inflight_max
)
2074 bwnd
= tcp_inflight_max
;
2075 if ((long)bwnd
< tp
->t_maxseg
* 2)
2076 bwnd
= tp
->t_maxseg
* 2;
2077 tp
->snd_bwnd
= bwnd
;
2081 tcp_rmx_iwsegs(struct tcpcb
*tp
, u_long
*maxsegs
, u_long
*capsegs
)
2084 struct inpcb
*inp
= tp
->t_inpcb
;
2086 boolean_t isipv6
= INP_ISIPV6(inp
);
2088 const boolean_t isipv6
= FALSE
;
2092 if (tcp_iw_maxsegs
< TCP_IW_MAXSEGS_DFLT
)
2093 tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
2094 if (tcp_iw_capsegs
< TCP_IW_CAPSEGS_DFLT
)
2095 tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
2098 rt
= tcp_rtlookup6(&inp
->inp_inc
);
2100 rt
= tcp_rtlookup(&inp
->inp_inc
);
2102 rt
->rt_rmx
.rmx_iwmaxsegs
< TCP_IW_MAXSEGS_DFLT
||
2103 rt
->rt_rmx
.rmx_iwcapsegs
< TCP_IW_CAPSEGS_DFLT
) {
2104 *maxsegs
= tcp_iw_maxsegs
;
2105 *capsegs
= tcp_iw_capsegs
;
2108 *maxsegs
= rt
->rt_rmx
.rmx_iwmaxsegs
;
2109 *capsegs
= rt
->rt_rmx
.rmx_iwcapsegs
;
2113 tcp_initial_window(struct tcpcb
*tp
)
2115 if (tcp_do_rfc3390
) {
2118 * "If the SYN or SYN/ACK is lost, the initial window
2119 * used by a sender after a correctly transmitted SYN
2120 * MUST be one segment consisting of MSS bytes."
2122 * However, we do something a little bit more aggressive
2123 * then RFC3390 here:
2124 * - Only if time spent in the SYN or SYN|ACK retransmition
2125 * >= 3 seconds, the IW is reduced. We do this mainly
2126 * because when RFC3390 is published, the initial RTO is
2127 * still 3 seconds (the threshold we test here), while
2128 * after RFC6298, the initial RTO is 1 second. This
2129 * behaviour probably still falls within the spirit of
2131 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2132 * Mainly to avoid sender and receiver deadlock until
2133 * delayed ACK timer expires. And even RFC2581 does not
2134 * try to reduce IW upon SYN or SYN|ACK retransmition
2138 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2140 if (tp
->t_rxtsyn
>= TCPTV_RTOBASE3
) {
2141 return (2 * tp
->t_maxseg
);
2143 u_long maxsegs
, capsegs
;
2145 tcp_rmx_iwsegs(tp
, &maxsegs
, &capsegs
);
2146 return min(maxsegs
* tp
->t_maxseg
,
2147 max(2 * tp
->t_maxseg
, capsegs
* 1460));
2151 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2153 * Mainly to avoid sender and receiver deadlock
2154 * until delayed ACK timer expires.
2156 return (2 * tp
->t_maxseg
);
2160 #ifdef TCP_SIGNATURE
2162 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2164 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2165 * When called from tcp_input(), we can be sure that th_sum has been
2166 * zeroed out and verified already.
2168 * Return 0 if successful, otherwise return -1.
2170 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2171 * search with the destination IP address, and a 'magic SPI' to be
2172 * determined by the application. This is hardcoded elsewhere to 1179
2173 * right now. Another branch of this code exists which uses the SPD to
2174 * specify per-application flows but it is unstable.
2177 tcpsignature_compute(
2178 struct mbuf
*m
, /* mbuf chain */
2179 int len
, /* length of TCP data */
2180 int optlen
, /* length of TCP options */
2181 u_char
*buf
, /* storage for MD5 digest */
2182 u_int direction
) /* direction of flow */
2184 struct ippseudo ippseudo
;
2188 struct ipovly
*ipovly
;
2189 struct secasvar
*sav
;
2192 struct ip6_hdr
*ip6
;
2193 struct in6_addr in6
;
2199 KASSERT(m
!= NULL
, ("passed NULL mbuf. Game over."));
2200 KASSERT(buf
!= NULL
, ("passed NULL storage pointer for MD5 signature"));
2202 * Extract the destination from the IP header in the mbuf.
2204 ip
= mtod(m
, struct ip
*);
2206 ip6
= NULL
; /* Make the compiler happy. */
2209 * Look up an SADB entry which matches the address found in
2212 switch (IP_VHL_V(ip
->ip_vhl
)) {
2214 sav
= key_allocsa(AF_INET
, (caddr_t
)&ip
->ip_src
, (caddr_t
)&ip
->ip_dst
,
2215 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2218 case (IPV6_VERSION
>> 4):
2219 ip6
= mtod(m
, struct ip6_hdr
*);
2220 sav
= key_allocsa(AF_INET6
, (caddr_t
)&ip6
->ip6_src
, (caddr_t
)&ip6
->ip6_dst
,
2221 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2230 kprintf("%s: SADB lookup failed\n", __func__
);
2236 * Step 1: Update MD5 hash with IP pseudo-header.
2238 * XXX The ippseudo header MUST be digested in network byte order,
2239 * or else we'll fail the regression test. Assume all fields we've
2240 * been doing arithmetic on have been in host byte order.
2241 * XXX One cannot depend on ipovly->ih_len here. When called from
2242 * tcp_output(), the underlying ip_len member has not yet been set.
2244 switch (IP_VHL_V(ip
->ip_vhl
)) {
2246 ipovly
= (struct ipovly
*)ip
;
2247 ippseudo
.ippseudo_src
= ipovly
->ih_src
;
2248 ippseudo
.ippseudo_dst
= ipovly
->ih_dst
;
2249 ippseudo
.ippseudo_pad
= 0;
2250 ippseudo
.ippseudo_p
= IPPROTO_TCP
;
2251 ippseudo
.ippseudo_len
= htons(len
+ sizeof(struct tcphdr
) + optlen
);
2252 MD5Update(&ctx
, (char *)&ippseudo
, sizeof(struct ippseudo
));
2253 th
= (struct tcphdr
*)((u_char
*)ip
+ sizeof(struct ip
));
2254 doff
= sizeof(struct ip
) + sizeof(struct tcphdr
) + optlen
;
2258 * RFC 2385, 2.0 Proposal
2259 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2260 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2261 * extended next header value (to form 32 bits), and 32-bit segment
2263 * Note: Upper-Layer Packet Length comes before Next Header.
2265 case (IPV6_VERSION
>> 4):
2267 in6_clearscope(&in6
);
2268 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2270 in6_clearscope(&in6
);
2271 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2272 plen
= htonl(len
+ sizeof(struct tcphdr
) + optlen
);
2273 MD5Update(&ctx
, (char *)&plen
, sizeof(uint32_t));
2275 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2276 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2277 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2279 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2280 th
= (struct tcphdr
*)((u_char
*)ip6
+ sizeof(struct ip6_hdr
));
2281 doff
= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) + optlen
;
2290 * Step 2: Update MD5 hash with TCP header, excluding options.
2291 * The TCP checksum must be set to zero.
2293 savecsum
= th
->th_sum
;
2295 MD5Update(&ctx
, (char *)th
, sizeof(struct tcphdr
));
2296 th
->th_sum
= savecsum
;
2298 * Step 3: Update MD5 hash with TCP segment data.
2299 * Use m_apply() to avoid an early m_pullup().
2302 m_apply(m
, doff
, len
, tcpsignature_apply
, &ctx
);
2304 * Step 4: Update MD5 hash with shared secret.
2306 MD5Update(&ctx
, _KEYBUF(sav
->key_auth
), _KEYLEN(sav
->key_auth
));
2307 MD5Final(buf
, &ctx
);
2308 key_sa_recordxfer(sav
, m
);
2314 tcpsignature_apply(void *fstate
, void *data
, unsigned int len
)
2317 MD5Update((MD5_CTX
*)fstate
, (unsigned char *)data
, len
);
2320 #endif /* TCP_SIGNATURE */
2323 tcp_drop_sysctl_dispatch(netmsg_t nmsg
)
2325 struct lwkt_msg
*lmsg
= &nmsg
->lmsg
;
2326 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2327 struct sockaddr_storage
*addrs
= lmsg
->u
.ms_resultp
;
2329 struct sockaddr_in
*fin
, *lin
;
2331 struct sockaddr_in6
*fin6
, *lin6
;
2332 struct in6_addr f6
, l6
;
2336 switch (addrs
[0].ss_family
) {
2339 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2340 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2341 error
= in6_embedscope(&f6
, fin6
, NULL
, NULL
);
2344 error
= in6_embedscope(&l6
, lin6
, NULL
, NULL
);
2347 inp
= in6_pcblookup_hash(&tcbinfo
[mycpuid
], &f6
,
2348 fin6
->sin6_port
, &l6
, lin6
->sin6_port
, FALSE
, NULL
);
2353 fin
= (struct sockaddr_in
*)&addrs
[0];
2354 lin
= (struct sockaddr_in
*)&addrs
[1];
2355 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], fin
->sin_addr
,
2356 fin
->sin_port
, lin
->sin_addr
, lin
->sin_port
, FALSE
, NULL
);
2361 * Must not reach here, since the address family was
2362 * checked in sysctl handler.
2364 panic("unknown address family %d", addrs
[0].ss_family
);
2367 struct tcpcb
*tp
= intotcpcb(inp
);
2369 KASSERT((inp
->inp_flags
& INP_WILDCARD
) == 0,
2370 ("in wildcard hash"));
2371 KASSERT(tp
!= NULL
, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2372 KASSERT((tp
->t_flags
& TF_LISTEN
) == 0, ("listen socket"));
2373 tcp_drop(tp
, ECONNABORTED
);
2381 lwkt_replymsg(lmsg
, error
);
2385 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS
)
2387 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2388 struct sockaddr_storage addrs
[2];
2389 struct sockaddr_in
*fin
, *lin
;
2391 struct sockaddr_in6
*fin6
, *lin6
;
2393 struct netmsg_base nmsg
;
2394 struct lwkt_msg
*lmsg
= &nmsg
.lmsg
;
2395 struct lwkt_port
*port
= NULL
;
2404 if (req
->oldptr
!= NULL
|| req
->oldlen
!= 0)
2406 if (req
->newptr
== NULL
)
2408 if (req
->newlen
< sizeof(addrs
))
2410 error
= SYSCTL_IN(req
, &addrs
, sizeof(addrs
));
2414 switch (addrs
[0].ss_family
) {
2417 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2418 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2419 if (fin6
->sin6_len
!= sizeof(struct sockaddr_in6
) ||
2420 lin6
->sin6_len
!= sizeof(struct sockaddr_in6
))
2422 if (IN6_IS_ADDR_V4MAPPED(&fin6
->sin6_addr
) ||
2423 IN6_IS_ADDR_V4MAPPED(&lin6
->sin6_addr
))
2424 return (EADDRNOTAVAIL
);
2426 error
= sa6_embedscope(fin6
, V_ip6_use_defzone
);
2429 error
= sa6_embedscope(lin6
, V_ip6_use_defzone
);
2433 port
= tcp6_addrport();
2438 fin
= (struct sockaddr_in
*)&addrs
[0];
2439 lin
= (struct sockaddr_in
*)&addrs
[1];
2440 if (fin
->sin_len
!= sizeof(struct sockaddr_in
) ||
2441 lin
->sin_len
!= sizeof(struct sockaddr_in
))
2443 port
= tcp_addrport(fin
->sin_addr
.s_addr
, fin
->sin_port
,
2444 lin
->sin_addr
.s_addr
, lin
->sin_port
);
2451 netmsg_init(&nmsg
, NULL
, &curthread
->td_msgport
, 0,
2452 tcp_drop_sysctl_dispatch
);
2453 lmsg
->u
.ms_resultp
= addrs
;
2454 return lwkt_domsg(port
, lmsg
, 0);
2457 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, drop
,
2458 CTLTYPE_STRUCT
| CTLFLAG_WR
| CTLFLAG_SKIP
, NULL
,
2459 0, sysctl_tcp_drop
, "", "Drop TCP connection");
2462 sysctl_tcps_count(SYSCTL_HANDLER_ARGS
)
2464 u_long state_count
[TCP_NSTATES
];
2467 memset(state_count
, 0, sizeof(state_count
));
2468 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2471 for (i
= 0; i
< TCP_NSTATES
; ++i
)
2472 state_count
[i
] += tcpstate_count
[cpu
].tcps_count
[i
];
2475 return sysctl_handle_opaque(oidp
, state_count
, sizeof(state_count
), req
);
2477 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, state_count
,
2478 CTLTYPE_OPAQUE
| CTLFLAG_RD
, NULL
, 0,
2479 sysctl_tcps_count
, "LU", "TCP connection counts by state");
2482 tcp_pcbport_create(struct tcpcb
*tp
)
2486 KASSERT((tp
->t_flags
& TF_LISTEN
) && tp
->t_state
== TCPS_LISTEN
,
2487 ("not a listen tcpcb"));
2489 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache was created"));
2490 tp
->t_pcbport
= kmalloc_cachealign(
2491 sizeof(struct tcp_pcbport
) * netisr_ncpus
, M_PCB
, M_WAITOK
);
2493 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2494 struct inpcbport
*phd
;
2496 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2497 LIST_INIT(&phd
->phd_pcblist
);
2498 /* Though, not used ... */
2499 phd
->phd_port
= tp
->t_inpcb
->inp_lport
;
2504 tcp_pcbport_merge_oncpu(struct tcpcb
*tp
)
2506 struct inpcbport
*phd
;
2510 KASSERT(cpu
< netisr_ncpus
, ("invalid cpu%d", cpu
));
2511 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2513 while ((inp
= LIST_FIRST(&phd
->phd_pcblist
)) != NULL
) {
2514 KASSERT(inp
->inp_phd
== phd
&& inp
->inp_porthash
== NULL
,
2515 ("not on tcpcb port cache"));
2516 LIST_REMOVE(inp
, inp_portlist
);
2517 in_pcbinsporthash_lport(inp
);
2518 KASSERT(inp
->inp_phd
== tp
->t_inpcb
->inp_phd
&&
2519 inp
->inp_porthash
== tp
->t_inpcb
->inp_porthash
,
2520 ("tcpcb port cache merge failed"));
2525 tcp_pcbport_destroy(struct tcpcb
*tp
)
2530 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2531 KASSERT(LIST_EMPTY(&tp
->t_pcbport
[cpu
].t_phd
.phd_pcblist
),
2532 ("tcpcb port cache is not empty"));
2535 kfree(tp
->t_pcbport
, M_PCB
);
2536 tp
->t_pcbport
= NULL
;