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_tcpdebug.h"
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/callout.h>
73 #include <sys/kernel.h>
74 #include <sys/sysctl.h>
75 #include <sys/malloc.h>
76 #include <sys/mpipe.h>
79 #include <sys/domain.h>
83 #include <sys/socket.h>
84 #include <sys/socketops.h>
85 #include <sys/socketvar.h>
86 #include <sys/protosw.h>
87 #include <sys/random.h>
88 #include <sys/in_cksum.h>
91 #include <net/route.h>
93 #include <net/netisr2.h>
96 #include <netinet/in.h>
97 #include <netinet/in_systm.h>
98 #include <netinet/ip.h>
99 #include <netinet/ip6.h>
100 #include <netinet/in_pcb.h>
101 #include <netinet6/in6_pcb.h>
102 #include <netinet/in_var.h>
103 #include <netinet/ip_var.h>
104 #include <netinet6/ip6_var.h>
105 #include <netinet/ip_icmp.h>
107 #include <netinet/icmp6.h>
109 #include <netinet/tcp.h>
110 #include <netinet/tcp_fsm.h>
111 #include <netinet/tcp_seq.h>
112 #include <netinet/tcp_timer.h>
113 #include <netinet/tcp_timer2.h>
114 #include <netinet/tcp_var.h>
115 #include <netinet6/tcp6_var.h>
116 #include <netinet/tcpip.h>
118 #include <netinet/tcp_debug.h>
120 #include <netinet6/ip6protosw.h>
123 #include <machine/smp.h>
125 #include <sys/msgport2.h>
126 #include <sys/mplock2.h>
127 #include <net/netmsg2.h>
129 #if !defined(KTR_TCP)
130 #define KTR_TCP KTR_ALL
133 KTR_INFO_MASTER(tcp);
134 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
135 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
136 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
137 #define logtcp(name) KTR_LOG(tcp_ ## name)
140 #define TCP_IW_MAXSEGS_DFLT 4
141 #define TCP_IW_CAPSEGS_DFLT 4
143 struct tcp_reass_pcpu
{
145 struct netmsg_base drain_nmsg
;
148 struct inpcbinfo tcbinfo
[MAXCPU
];
149 struct tcpcbackq tcpcbackq
[MAXCPU
];
150 struct tcp_reass_pcpu tcp_reassq
[MAXCPU
];
152 int tcp_mssdflt
= TCP_MSS
;
153 SYSCTL_INT(_net_inet_tcp
, TCPCTL_MSSDFLT
, mssdflt
, CTLFLAG_RW
,
154 &tcp_mssdflt
, 0, "Default TCP Maximum Segment Size");
157 int tcp_v6mssdflt
= TCP6_MSS
;
158 SYSCTL_INT(_net_inet_tcp
, TCPCTL_V6MSSDFLT
, v6mssdflt
, CTLFLAG_RW
,
159 &tcp_v6mssdflt
, 0, "Default TCP Maximum Segment Size for IPv6");
163 * Minimum MSS we accept and use. This prevents DoS attacks where
164 * we are forced to a ridiculous low MSS like 20 and send hundreds
165 * of packets instead of one. The effect scales with the available
166 * bandwidth and quickly saturates the CPU and network interface
167 * with packet generation and sending. Set to zero to disable MINMSS
168 * checking. This setting prevents us from sending too small packets.
170 int tcp_minmss
= TCP_MINMSS
;
171 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, minmss
, CTLFLAG_RW
,
172 &tcp_minmss
, 0, "Minmum TCP Maximum Segment Size");
175 static int tcp_rttdflt
= TCPTV_SRTTDFLT
/ PR_SLOWHZ
;
176 SYSCTL_INT(_net_inet_tcp
, TCPCTL_RTTDFLT
, rttdflt
, CTLFLAG_RW
,
177 &tcp_rttdflt
, 0, "Default maximum TCP Round Trip Time");
180 int tcp_do_rfc1323
= 1;
181 SYSCTL_INT(_net_inet_tcp
, TCPCTL_DO_RFC1323
, rfc1323
, CTLFLAG_RW
,
182 &tcp_do_rfc1323
, 0, "Enable rfc1323 (high performance TCP) extensions");
184 static int tcp_tcbhashsize
= 0;
185 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, tcbhashsize
, CTLFLAG_RD
,
186 &tcp_tcbhashsize
, 0, "Size of TCP control block hashtable");
188 static int do_tcpdrain
= 1;
189 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, do_tcpdrain
, CTLFLAG_RW
, &do_tcpdrain
, 0,
190 "Enable tcp_drain routine for extra help when low on mbufs");
192 static int icmp_may_rst
= 1;
193 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, icmp_may_rst
, CTLFLAG_RW
, &icmp_may_rst
, 0,
194 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
196 static int tcp_isn_reseed_interval
= 0;
197 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, isn_reseed_interval
, CTLFLAG_RW
,
198 &tcp_isn_reseed_interval
, 0, "Seconds between reseeding of ISN secret");
201 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
202 * by default, but with generous values which should allow maximal
203 * bandwidth. In particular, the slop defaults to 50 (5 packets).
205 * The reason for doing this is that the limiter is the only mechanism we
206 * have which seems to do a really good job preventing receiver RX rings
207 * on network interfaces from getting blown out. Even though GigE/10GigE
208 * is supposed to flow control it looks like either it doesn't actually
209 * do it or Open Source drivers do not properly enable it.
211 * People using the limiter to reduce bottlenecks on slower WAN connections
212 * should set the slop to 20 (2 packets).
214 static int tcp_inflight_enable
= 1;
215 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_enable
, CTLFLAG_RW
,
216 &tcp_inflight_enable
, 0, "Enable automatic TCP inflight data limiting");
218 static int tcp_inflight_debug
= 0;
219 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_debug
, CTLFLAG_RW
,
220 &tcp_inflight_debug
, 0, "Debug TCP inflight calculations");
223 * NOTE: tcp_inflight_start is essentially the starting receive window
224 * for a connection. If set too low then fetches over tcp
225 * connections will take noticably longer to ramp-up over
226 * high-latency connections. 6144 is too low for a default,
227 * use something more reasonable.
229 static int tcp_inflight_start
= 33792;
230 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_start
, CTLFLAG_RW
,
231 &tcp_inflight_start
, 0, "Start value for TCP inflight window");
233 static int tcp_inflight_min
= 6144;
234 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_min
, CTLFLAG_RW
,
235 &tcp_inflight_min
, 0, "Lower bound for TCP inflight window");
237 static int tcp_inflight_max
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
238 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_max
, CTLFLAG_RW
,
239 &tcp_inflight_max
, 0, "Upper bound for TCP inflight window");
241 static int tcp_inflight_stab
= 50;
242 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_stab
, CTLFLAG_RW
,
243 &tcp_inflight_stab
, 0, "Fudge bw 1/10% (50=5%)");
245 static int tcp_inflight_adjrtt
= 2;
246 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_adjrtt
, CTLFLAG_RW
,
247 &tcp_inflight_adjrtt
, 0, "Slop for rtt 1/(hz*32)");
249 static int tcp_do_rfc3390
= 1;
250 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, rfc3390
, CTLFLAG_RW
,
252 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
254 static u_long tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
255 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwmaxsegs
, CTLFLAG_RW
,
256 &tcp_iw_maxsegs
, 0, "TCP IW segments max");
258 static u_long tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
259 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwcapsegs
, CTLFLAG_RW
,
260 &tcp_iw_capsegs
, 0, "TCP IW segments");
262 int tcp_low_rtobase
= 1;
263 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, low_rtobase
, CTLFLAG_RW
,
264 &tcp_low_rtobase
, 0, "Lowering the Initial RTO (RFC 6298)");
266 static int tcp_do_ncr
= 1;
267 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr
, CTLFLAG_RW
,
268 &tcp_do_ncr
, 0, "Non-Congestion Robustness (RFC 4653)");
270 int tcp_ncr_linklocal
= 0;
271 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_linklocal
, CTLFLAG_RW
,
272 &tcp_ncr_linklocal
, 0,
273 "Enable Non-Congestion Robustness (RFC 4653) on link local network");
275 int tcp_ncr_rxtthresh_max
= 16;
276 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_rxtthresh_max
, CTLFLAG_RW
,
277 &tcp_ncr_rxtthresh_max
, 0,
278 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
280 static MALLOC_DEFINE(M_TCPTEMP
, "tcptemp", "TCP Templates for Keepalives");
281 static struct malloc_pipe tcptemp_mpipe
;
283 static void tcp_willblock(void);
284 static void tcp_notify (struct inpcb
*, int);
286 struct tcp_stats tcpstats_percpu
[MAXCPU
] __cachealign
;
287 struct tcp_state_count tcpstate_count
[MAXCPU
] __cachealign
;
289 static void tcp_drain_dispatch(netmsg_t nmsg
);
292 sysctl_tcpstats(SYSCTL_HANDLER_ARGS
)
296 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
297 if ((error
= SYSCTL_OUT(req
, &tcpstats_percpu
[cpu
],
298 sizeof(struct tcp_stats
))))
300 if ((error
= SYSCTL_IN(req
, &tcpstats_percpu
[cpu
],
301 sizeof(struct tcp_stats
))))
307 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_STATS
, stats
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
308 0, 0, sysctl_tcpstats
, "S,tcp_stats", "TCP statistics");
311 * Target size of TCP PCB hash tables. Must be a power of two.
313 * Note that this can be overridden by the kernel environment
314 * variable net.inet.tcp.tcbhashsize
317 #define TCBHASHSIZE 512
319 CTASSERT((TCBHASHSIZE
& (TCBHASHSIZE
- 1)) == 0);
322 * This is the actual shape of what we allocate using the zone
323 * allocator. Doing it this way allows us to protect both structures
324 * using the same generation count, and also eliminates the overhead
325 * of allocating tcpcbs separately. By hiding the structure here,
326 * we avoid changing most of the rest of the code (although it needs
327 * to be changed, eventually, for greater efficiency).
330 #define ALIGNM1 (ALIGNMENT - 1)
334 char align
[(sizeof(struct inpcb
) + ALIGNM1
) & ~ALIGNM1
];
337 struct tcp_callout inp_tp_rexmt
;
338 struct tcp_callout inp_tp_persist
;
339 struct tcp_callout inp_tp_keep
;
340 struct tcp_callout inp_tp_2msl
;
341 struct tcp_callout inp_tp_delack
;
342 struct netmsg_tcp_timer inp_tp_timermsg
;
343 struct netmsg_base inp_tp_sndmore
;
354 struct inpcbportinfo
*portinfo
;
355 struct inpcbinfo
*ticb
;
356 int hashsize
= TCBHASHSIZE
, portinfo_hsize
;
360 * note: tcptemp is used for keepalives, and it is ok for an
361 * allocation to fail so do not specify MPF_INT.
363 mpipe_init(&tcptemp_mpipe
, M_TCPTEMP
, sizeof(struct tcptemp
),
364 25, -1, 0, NULL
, NULL
, NULL
);
366 tcp_delacktime
= TCPTV_DELACK
;
367 tcp_keepinit
= TCPTV_KEEP_INIT
;
368 tcp_keepidle
= TCPTV_KEEP_IDLE
;
369 tcp_keepintvl
= TCPTV_KEEPINTVL
;
370 tcp_maxpersistidle
= TCPTV_KEEP_IDLE
;
372 tcp_rexmit_min
= TCPTV_MIN
;
373 if (tcp_rexmit_min
< 1) /* if kern.hz is too low */
375 tcp_rexmit_slop
= TCPTV_CPU_VAR
;
377 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize
);
378 if (!powerof2(hashsize
)) {
379 kprintf("WARNING: TCB hash size not a power of 2\n");
380 hashsize
= TCBHASHSIZE
; /* safe default */
382 tcp_tcbhashsize
= hashsize
;
384 portinfo_hsize
= 65536 / netisr_ncpus
;
385 if (portinfo_hsize
> hashsize
)
386 portinfo_hsize
= hashsize
;
388 portinfo
= kmalloc_cachealign(sizeof(*portinfo
) * netisr_ncpus
, M_PCB
,
391 for (cpu
= 0; cpu
< netisr_ncpus
; cpu
++) {
392 ticb
= &tcbinfo
[cpu
];
393 in_pcbinfo_init(ticb
, cpu
, FALSE
);
394 ticb
->hashbase
= hashinit(hashsize
, M_PCB
,
396 in_pcbportinfo_init(&portinfo
[cpu
], portinfo_hsize
, cpu
);
397 in_pcbportinfo_set(ticb
, portinfo
, netisr_ncpus
);
398 ticb
->wildcardhashbase
= hashinit(hashsize
, M_PCB
,
399 &ticb
->wildcardhashmask
);
400 ticb
->localgrphashbase
= hashinit(hashsize
, M_PCB
,
401 &ticb
->localgrphashmask
);
402 ticb
->ipi_size
= sizeof(struct inp_tp
);
403 TAILQ_INIT(&tcpcbackq
[cpu
].head
);
406 tcp_reass_maxseg
= nmbclusters
/ 16;
407 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg
);
410 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
412 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
414 if (max_protohdr
< TCP_MINPROTOHDR
)
415 max_protohdr
= TCP_MINPROTOHDR
;
416 if (max_linkhdr
+ TCP_MINPROTOHDR
> MHLEN
)
418 #undef TCP_MINPROTOHDR
421 * Initialize TCP statistics counters for each CPU.
423 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
)
424 bzero(&tcpstats_percpu
[cpu
], sizeof(struct tcp_stats
));
427 * Initialize netmsgs for TCP drain
429 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
430 netmsg_init(&tcp_reassq
[cpu
].drain_nmsg
, NULL
,
431 &netisr_adone_rport
, MSGF_PRIORITY
, tcp_drain_dispatch
);
435 netisr_register_rollup(tcp_willblock
, NETISR_ROLLUP_PRIO_TCP
);
444 while ((tp
= TAILQ_FIRST(&tcpcbackq
[cpu
].head
)) != NULL
) {
445 KKASSERT(tp
->t_flags
& TF_ONOUTPUTQ
);
446 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
447 TAILQ_REMOVE(&tcpcbackq
[cpu
].head
, tp
, t_outputq
);
453 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
454 * tcp_template used to store this data in mbufs, but we now recopy it out
455 * of the tcpcb each time to conserve mbufs.
458 tcp_fillheaders(struct tcpcb
*tp
, void *ip_ptr
, void *tcp_ptr
, boolean_t tso
)
460 struct inpcb
*inp
= tp
->t_inpcb
;
461 struct tcphdr
*tcp_hdr
= (struct tcphdr
*)tcp_ptr
;
464 if (INP_ISIPV6(inp
)) {
467 ip6
= (struct ip6_hdr
*)ip_ptr
;
468 ip6
->ip6_flow
= (ip6
->ip6_flow
& ~IPV6_FLOWINFO_MASK
) |
469 (inp
->in6p_flowinfo
& IPV6_FLOWINFO_MASK
);
470 ip6
->ip6_vfc
= (ip6
->ip6_vfc
& ~IPV6_VERSION_MASK
) |
471 (IPV6_VERSION
& IPV6_VERSION_MASK
);
472 ip6
->ip6_nxt
= IPPROTO_TCP
;
473 ip6
->ip6_plen
= sizeof(struct tcphdr
);
474 ip6
->ip6_src
= inp
->in6p_laddr
;
475 ip6
->ip6_dst
= inp
->in6p_faddr
;
480 struct ip
*ip
= (struct ip
*) ip_ptr
;
483 ip
->ip_vhl
= IP_VHL_BORING
;
490 ip
->ip_p
= IPPROTO_TCP
;
491 ip
->ip_src
= inp
->inp_laddr
;
492 ip
->ip_dst
= inp
->inp_faddr
;
495 plen
= htons(IPPROTO_TCP
);
497 plen
= htons(sizeof(struct tcphdr
) + IPPROTO_TCP
);
498 tcp_hdr
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
,
499 ip
->ip_dst
.s_addr
, plen
);
502 tcp_hdr
->th_sport
= inp
->inp_lport
;
503 tcp_hdr
->th_dport
= inp
->inp_fport
;
508 tcp_hdr
->th_flags
= 0;
514 * Create template to be used to send tcp packets on a connection.
515 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
516 * use for this function is in keepalives, which use tcp_respond.
519 tcp_maketemplate(struct tcpcb
*tp
)
523 if ((tmp
= mpipe_alloc_nowait(&tcptemp_mpipe
)) == NULL
)
525 tcp_fillheaders(tp
, &tmp
->tt_ipgen
, &tmp
->tt_t
, FALSE
);
530 tcp_freetemplate(struct tcptemp
*tmp
)
532 mpipe_free(&tcptemp_mpipe
, tmp
);
536 * Send a single message to the TCP at address specified by
537 * the given TCP/IP header. If m == NULL, then we make a copy
538 * of the tcpiphdr at ti and send directly to the addressed host.
539 * This is used to force keep alive messages out using the TCP
540 * template for a connection. If flags are given then we send
541 * a message back to the TCP which originated the * segment ti,
542 * and discard the mbuf containing it and any other attached mbufs.
544 * In any case the ack and sequence number of the transmitted
545 * segment are as specified by the parameters.
547 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
550 tcp_respond(struct tcpcb
*tp
, void *ipgen
, struct tcphdr
*th
, struct mbuf
*m
,
551 tcp_seq ack
, tcp_seq seq
, int flags
)
555 struct route
*ro
= NULL
;
557 struct ip
*ip
= ipgen
;
560 struct route_in6
*ro6
= NULL
;
561 struct route_in6 sro6
;
562 struct ip6_hdr
*ip6
= ipgen
;
563 struct inpcb
*inp
= NULL
;
564 boolean_t use_tmpro
= TRUE
;
566 boolean_t isipv6
= (IP_VHL_V(ip
->ip_vhl
) == 6);
568 const boolean_t isipv6
= FALSE
;
573 if (!(flags
& TH_RST
)) {
574 win
= ssb_space(&inp
->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
= &inp
->in6p_route
;
588 ro
= &inp
->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(inp
,
674 (ro6
&& ro6
->ro_rt
) ? ro6
->ro_rt
->rt_ifp
: NULL
);
676 nth
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
, ip
->ip_dst
.s_addr
,
677 htons((u_short
)(tlen
- sizeof(struct ip
) + ip
->ip_p
)));
678 m
->m_pkthdr
.csum_flags
= CSUM_TCP
;
679 m
->m_pkthdr
.csum_data
= offsetof(struct tcphdr
, th_sum
);
680 m
->m_pkthdr
.csum_thlen
= sizeof(struct tcphdr
);
683 if (tp
== NULL
|| (inp
->inp_socket
->so_options
& SO_DEBUG
))
684 tcp_trace(TA_OUTPUT
, 0, tp
, mtod(m
, void *), th
, 0);
687 ip6_output(m
, NULL
, ro6
, ipflags
, NULL
, NULL
, inp
);
688 if ((ro6
== &sro6
) && (ro6
->ro_rt
!= NULL
)) {
693 if (inp
!= NULL
&& (inp
->inp_flags
& INP_HASH
))
694 m_sethash(m
, inp
->inp_hashval
);
695 ipflags
|= IP_DEBUGROUTE
;
696 ip_output(m
, NULL
, ro
, ipflags
, NULL
, inp
);
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
< netisr_ncpus
)
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
) && netisr_ncpus
> 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_terminate(tp
, tp
->tt_rexmt
);
916 tcp_callout_terminate(tp
, tp
->tt_persist
);
917 tcp_callout_terminate(tp
, tp
->tt_keep
);
918 tcp_callout_terminate(tp
, tp
->tt_2msl
);
919 tcp_callout_terminate(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
].head
, 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
++;
1080 * Walk the tcpbs, if existing, and flush the reassembly queue,
1081 * if there is one...
1084 tcp_drain_oncpu(struct inpcbinfo
*pcbinfo
)
1086 struct inpcbhead
*head
= &pcbinfo
->pcblisthead
;
1090 * Since we run in netisr, it is MP safe, even if
1091 * we block during the inpcb list iteration, i.e.
1092 * we don't need to use inpcb marker here.
1094 ASSERT_NETISR_NCPUS(pcbinfo
->cpu
);
1096 LIST_FOREACH(inpb
, head
, inp_list
) {
1098 struct tseg_qent
*te
;
1100 if (inpb
->inp_flags
& INP_PLACEMARKER
)
1103 tcpb
= intotcpcb(inpb
);
1104 KASSERT(tcpb
!= NULL
, ("tcp_drain_oncpu: tcpb is NULL"));
1106 if ((te
= TAILQ_FIRST(&tcpb
->t_segq
)) != NULL
) {
1107 TAILQ_REMOVE(&tcpb
->t_segq
, te
, tqe_q
);
1108 if (te
->tqe_th
->th_flags
& TH_FIN
)
1109 tcpb
->t_flags
&= ~TF_QUEDFIN
;
1112 atomic_add_int(&tcp_reass_qsize
, -1);
1119 tcp_drain_dispatch(netmsg_t nmsg
)
1122 lwkt_replymsg(&nmsg
->lmsg
, 0); /* reply ASAP */
1125 tcp_drain_oncpu(&tcbinfo
[mycpuid
]);
1126 tcp_reassq
[mycpuid
].draining
= 0;
1130 tcp_drain_ipi(void *arg __unused
)
1133 struct lwkt_msg
*msg
= &tcp_reassq
[cpu
].drain_nmsg
.lmsg
;
1136 if (msg
->ms_flags
& MSGF_DONE
)
1137 lwkt_sendmsg_oncpu(netisr_cpuport(cpu
), msg
);
1150 if (tcp_reass_qsize
== 0)
1153 CPUMASK_ASSBMASK(mask
, netisr_ncpus
);
1154 CPUMASK_ANDMASK(mask
, smp_active_mask
);
1157 if (IN_NETISR_NCPUS(cpu
)) {
1158 tcp_drain_oncpu(&tcbinfo
[cpu
]);
1159 CPUMASK_NANDBIT(mask
, cpu
);
1162 if (tcp_reass_qsize
< netisr_ncpus
) {
1163 /* Does not worth the trouble. */
1167 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
1168 if (!CPUMASK_TESTBIT(mask
, cpu
))
1171 if (tcp_reassq
[cpu
].draining
) {
1172 /* Draining; skip this cpu. */
1173 CPUMASK_NANDBIT(mask
, cpu
);
1176 tcp_reassq
[cpu
].draining
= 1;
1179 if (CPUMASK_TESTNZERO(mask
))
1180 lwkt_send_ipiq_mask(mask
, tcp_drain_ipi
, NULL
);
1184 * Notify a tcp user of an asynchronous error;
1185 * store error as soft error, but wake up user
1186 * (for now, won't do anything until can select for soft error).
1188 * Do not wake up user since there currently is no mechanism for
1189 * reporting soft errors (yet - a kqueue filter may be added).
1192 tcp_notify(struct inpcb
*inp
, int error
)
1194 struct tcpcb
*tp
= intotcpcb(inp
);
1197 * Ignore some errors if we are hooked up.
1198 * If connection hasn't completed, has retransmitted several times,
1199 * and receives a second error, give up now. This is better
1200 * than waiting a long time to establish a connection that
1201 * can never complete.
1203 if (tp
->t_state
== TCPS_ESTABLISHED
&&
1204 (error
== EHOSTUNREACH
|| error
== ENETUNREACH
||
1205 error
== EHOSTDOWN
)) {
1207 } else if (tp
->t_state
< TCPS_ESTABLISHED
&& tp
->t_rxtshift
> 3 &&
1209 tcp_drop(tp
, error
);
1211 tp
->t_softerror
= error
;
1213 wakeup(&so
->so_timeo
);
1220 tcp_pcblist(SYSCTL_HANDLER_ARGS
)
1223 struct inpcb
*marker
;
1231 * The process of preparing the TCB list is too time-consuming and
1232 * resource-intensive to repeat twice on every request.
1234 if (req
->oldptr
== NULL
) {
1235 for (ccpu
= 0; ccpu
< netisr_ncpus
; ++ccpu
)
1236 n
+= tcbinfo
[ccpu
].ipi_count
;
1237 req
->oldidx
= (n
+ n
/8 + 10) * sizeof(struct xtcpcb
);
1241 if (req
->newptr
!= NULL
)
1244 marker
= kmalloc(sizeof(struct inpcb
), M_TEMP
, M_WAITOK
|M_ZERO
);
1245 marker
->inp_flags
|= INP_PLACEMARKER
;
1248 * OK, now we're committed to doing something. Run the inpcb list
1249 * for each cpu in the system and construct the output. Use a
1250 * list placemarker to deal with list changes occuring during
1251 * copyout blockages (but otherwise depend on being on the correct
1252 * cpu to avoid races).
1254 origcpu
= mycpu
->gd_cpuid
;
1255 for (ccpu
= 0; ccpu
< netisr_ncpus
&& error
== 0; ++ccpu
) {
1259 lwkt_migratecpu(ccpu
);
1261 n
= tcbinfo
[ccpu
].ipi_count
;
1263 LIST_INSERT_HEAD(&tcbinfo
[ccpu
].pcblisthead
, marker
, inp_list
);
1265 while ((inp
= LIST_NEXT(marker
, inp_list
)) != NULL
&& i
< n
) {
1267 * process a snapshot of pcbs, ignoring placemarkers
1268 * and using our own to allow SYSCTL_OUT to block.
1270 LIST_REMOVE(marker
, inp_list
);
1271 LIST_INSERT_AFTER(inp
, marker
, inp_list
);
1273 if (inp
->inp_flags
& INP_PLACEMARKER
)
1275 if (prison_xinpcb(req
->td
, inp
))
1278 xt
.xt_len
= sizeof xt
;
1279 bcopy(inp
, &xt
.xt_inp
, sizeof *inp
);
1280 inp_ppcb
= inp
->inp_ppcb
;
1281 if (inp_ppcb
!= NULL
)
1282 bcopy(inp_ppcb
, &xt
.xt_tp
, sizeof xt
.xt_tp
);
1284 bzero(&xt
.xt_tp
, sizeof xt
.xt_tp
);
1285 if (inp
->inp_socket
)
1286 sotoxsocket(inp
->inp_socket
, &xt
.xt_socket
);
1287 if ((error
= SYSCTL_OUT(req
, &xt
, sizeof xt
)) != 0)
1291 LIST_REMOVE(marker
, inp_list
);
1292 if (error
== 0 && i
< n
) {
1293 bzero(&xt
, sizeof xt
);
1294 xt
.xt_len
= sizeof xt
;
1296 error
= SYSCTL_OUT(req
, &xt
, sizeof xt
);
1305 * Make sure we are on the same cpu we were on originally, since
1306 * higher level callers expect this. Also don't pollute caches with
1307 * migrated userland data by (eventually) returning to userland
1308 * on a different cpu.
1310 lwkt_migratecpu(origcpu
);
1311 kfree(marker
, M_TEMP
);
1315 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_PCBLIST
, pcblist
, CTLFLAG_RD
, 0, 0,
1316 tcp_pcblist
, "S,xtcpcb", "List of active TCP connections");
1319 tcp_getcred(SYSCTL_HANDLER_ARGS
)
1321 struct sockaddr_in addrs
[2];
1322 struct ucred cred0
, *cred
= NULL
;
1324 int cpu
, origcpu
, error
;
1326 error
= priv_check(req
->td
, PRIV_ROOT
);
1329 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1334 cpu
= tcp_addrcpu(addrs
[1].sin_addr
.s_addr
, addrs
[1].sin_port
,
1335 addrs
[0].sin_addr
.s_addr
, addrs
[0].sin_port
);
1337 lwkt_migratecpu(cpu
);
1339 inp
= in_pcblookup_hash(&tcbinfo
[cpu
], addrs
[1].sin_addr
,
1340 addrs
[1].sin_port
, addrs
[0].sin_addr
, addrs
[0].sin_port
, 0, NULL
);
1341 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1343 } else if (inp
->inp_socket
->so_cred
!= NULL
) {
1344 cred0
= *(inp
->inp_socket
->so_cred
);
1348 lwkt_migratecpu(origcpu
);
1353 return SYSCTL_OUT(req
, cred
, sizeof(struct ucred
));
1356 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1357 0, 0, tcp_getcred
, "S,ucred", "Get the ucred of a TCP connection");
1361 tcp6_getcred(SYSCTL_HANDLER_ARGS
)
1363 struct sockaddr_in6 addrs
[2];
1367 error
= priv_check(req
->td
, PRIV_ROOT
);
1370 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1374 inp
= in6_pcblookup_hash(&tcbinfo
[0],
1375 &addrs
[1].sin6_addr
, addrs
[1].sin6_port
,
1376 &addrs
[0].sin6_addr
, addrs
[0].sin6_port
, 0, NULL
);
1377 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1381 error
= SYSCTL_OUT(req
, inp
->inp_socket
->so_cred
, sizeof(struct ucred
));
1387 SYSCTL_PROC(_net_inet6_tcp6
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1389 tcp6_getcred
, "S,ucred", "Get the ucred of a TCP6 connection");
1392 struct netmsg_tcp_notify
{
1393 struct netmsg_base base
;
1394 inp_notify_t nm_notify
;
1395 struct in_addr nm_faddr
;
1400 tcp_notifyall_oncpu(netmsg_t msg
)
1402 struct netmsg_tcp_notify
*nm
= (struct netmsg_tcp_notify
*)msg
;
1405 ASSERT_NETISR_NCPUS(mycpuid
);
1407 in_pcbnotifyall(&tcbinfo
[mycpuid
], nm
->nm_faddr
,
1408 nm
->nm_arg
, nm
->nm_notify
);
1410 nextcpu
= mycpuid
+ 1;
1411 if (nextcpu
< netisr_ncpus
)
1412 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nm
->base
.lmsg
);
1414 lwkt_replymsg(&nm
->base
.lmsg
, 0);
1418 tcp_get_inpnotify(int cmd
, const struct sockaddr
*sa
,
1419 int *arg
, struct ip
**ip0
, int *cpuid
)
1421 struct ip
*ip
= *ip0
;
1422 struct in_addr faddr
;
1423 inp_notify_t notify
= tcp_notify
;
1425 faddr
= ((const struct sockaddr_in
*)sa
)->sin_addr
;
1426 if (sa
->sa_family
!= AF_INET
|| faddr
.s_addr
== INADDR_ANY
)
1429 *arg
= inetctlerrmap
[cmd
];
1430 if (cmd
== PRC_QUENCH
) {
1431 notify
= tcp_quench
;
1432 } else if (icmp_may_rst
&&
1433 (cmd
== PRC_UNREACH_ADMIN_PROHIB
||
1434 cmd
== PRC_UNREACH_PORT
||
1435 cmd
== PRC_TIMXCEED_INTRANS
) &&
1437 notify
= tcp_drop_syn_sent
;
1438 } else if (cmd
== PRC_MSGSIZE
) {
1439 const struct icmp
*icmp
= (const struct icmp
*)
1440 ((caddr_t
)ip
- offsetof(struct icmp
, icmp_ip
));
1442 *arg
= ntohs(icmp
->icmp_nextmtu
);
1443 notify
= tcp_mtudisc
;
1444 } else if (PRC_IS_REDIRECT(cmd
)) {
1446 notify
= in_rtchange
;
1447 } else if (cmd
== PRC_HOSTDEAD
) {
1449 } else if ((unsigned)cmd
>= PRC_NCMDS
|| inetctlerrmap
[cmd
] == 0) {
1453 if (cpuid
!= NULL
) {
1455 /* Go through all effective netisr CPUs. */
1456 *cpuid
= netisr_ncpus
;
1458 const struct tcphdr
*th
;
1460 th
= (const struct tcphdr
*)
1461 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1462 *cpuid
= tcp_addrcpu(faddr
.s_addr
, th
->th_dport
,
1463 ip
->ip_src
.s_addr
, th
->th_sport
);
1472 tcp_ctlinput(netmsg_t msg
)
1474 int cmd
= msg
->ctlinput
.nm_cmd
;
1475 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1476 struct ip
*ip
= msg
->ctlinput
.nm_extra
;
1477 struct in_addr faddr
;
1478 inp_notify_t notify
;
1481 ASSERT_NETISR_NCPUS(mycpuid
);
1483 notify
= tcp_get_inpnotify(cmd
, sa
, &arg
, &ip
, &cpuid
);
1487 faddr
= ((struct sockaddr_in
*)sa
)->sin_addr
;
1489 const struct tcphdr
*th
;
1492 if (cpuid
!= mycpuid
)
1495 th
= (const struct tcphdr
*)
1496 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1497 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], faddr
, th
->th_dport
,
1498 ip
->ip_src
, th
->th_sport
, 0, NULL
);
1499 if (inp
!= NULL
&& inp
->inp_socket
!= NULL
) {
1500 tcp_seq icmpseq
= htonl(th
->th_seq
);
1501 struct tcpcb
*tp
= intotcpcb(inp
);
1503 if (SEQ_GEQ(icmpseq
, tp
->snd_una
) &&
1504 SEQ_LT(icmpseq
, tp
->snd_max
))
1507 struct in_conninfo inc
;
1509 inc
.inc_fport
= th
->th_dport
;
1510 inc
.inc_lport
= th
->th_sport
;
1511 inc
.inc_faddr
= faddr
;
1512 inc
.inc_laddr
= ip
->ip_src
;
1516 syncache_unreach(&inc
, th
);
1518 } else if (msg
->ctlinput
.nm_direct
) {
1519 if (cpuid
!= netisr_ncpus
&& cpuid
!= mycpuid
)
1522 in_pcbnotifyall(&tcbinfo
[mycpuid
], faddr
, arg
, notify
);
1524 struct netmsg_tcp_notify
*nm
;
1527 nm
= kmalloc(sizeof(*nm
), M_LWKTMSG
, M_INTWAIT
);
1528 netmsg_init(&nm
->base
, NULL
, &netisr_afree_rport
,
1529 0, tcp_notifyall_oncpu
);
1530 nm
->nm_faddr
= faddr
;
1532 nm
->nm_notify
= notify
;
1534 lwkt_sendmsg(netisr_cpuport(0), &nm
->base
.lmsg
);
1537 lwkt_replymsg(&msg
->lmsg
, 0);
1543 tcp6_ctlinput(netmsg_t msg
)
1545 int cmd
= msg
->ctlinput
.nm_cmd
;
1546 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1547 void *d
= msg
->ctlinput
.nm_extra
;
1549 inp_notify_t notify
= tcp_notify
;
1550 struct ip6_hdr
*ip6
;
1552 struct ip6ctlparam
*ip6cp
= NULL
;
1553 const struct sockaddr_in6
*sa6_src
= NULL
;
1555 struct tcp_portonly
{
1561 if (sa
->sa_family
!= AF_INET6
||
1562 sa
->sa_len
!= sizeof(struct sockaddr_in6
)) {
1567 if (cmd
== PRC_QUENCH
)
1568 notify
= tcp_quench
;
1569 else if (cmd
== PRC_MSGSIZE
) {
1570 struct ip6ctlparam
*ip6cp
= d
;
1571 struct icmp6_hdr
*icmp6
= ip6cp
->ip6c_icmp6
;
1573 arg
= ntohl(icmp6
->icmp6_mtu
);
1574 notify
= tcp_mtudisc
;
1575 } else if (!PRC_IS_REDIRECT(cmd
) &&
1576 ((unsigned)cmd
> PRC_NCMDS
|| inet6ctlerrmap
[cmd
] == 0)) {
1580 /* if the parameter is from icmp6, decode it. */
1582 ip6cp
= (struct ip6ctlparam
*)d
;
1584 ip6
= ip6cp
->ip6c_ip6
;
1585 off
= ip6cp
->ip6c_off
;
1586 sa6_src
= ip6cp
->ip6c_src
;
1590 off
= 0; /* fool gcc */
1595 struct in_conninfo inc
;
1597 * XXX: We assume that when IPV6 is non NULL,
1598 * M and OFF are valid.
1601 /* check if we can safely examine src and dst ports */
1602 if (m
->m_pkthdr
.len
< off
+ sizeof *thp
)
1605 bzero(&th
, sizeof th
);
1606 m_copydata(m
, off
, sizeof *thp
, (caddr_t
)&th
);
1608 in6_pcbnotify(&tcbinfo
[0], sa
, th
.th_dport
,
1609 (struct sockaddr
*)ip6cp
->ip6c_src
,
1610 th
.th_sport
, cmd
, arg
, notify
);
1612 inc
.inc_fport
= th
.th_dport
;
1613 inc
.inc_lport
= th
.th_sport
;
1614 inc
.inc6_faddr
= ((struct sockaddr_in6
*)sa
)->sin6_addr
;
1615 inc
.inc6_laddr
= ip6cp
->ip6c_src
->sin6_addr
;
1617 syncache_unreach(&inc
, &th
);
1619 in6_pcbnotify(&tcbinfo
[0], sa
, 0,
1620 (const struct sockaddr
*)sa6_src
, 0, cmd
, arg
, notify
);
1623 lwkt_replymsg(&msg
->ctlinput
.base
.lmsg
, 0);
1629 * Following is where TCP initial sequence number generation occurs.
1631 * There are two places where we must use initial sequence numbers:
1632 * 1. In SYN-ACK packets.
1633 * 2. In SYN packets.
1635 * All ISNs for SYN-ACK packets are generated by the syncache. See
1636 * tcp_syncache.c for details.
1638 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1639 * depends on this property. In addition, these ISNs should be
1640 * unguessable so as to prevent connection hijacking. To satisfy
1641 * the requirements of this situation, the algorithm outlined in
1642 * RFC 1948 is used to generate sequence numbers.
1644 * Implementation details:
1646 * Time is based off the system timer, and is corrected so that it
1647 * increases by one megabyte per second. This allows for proper
1648 * recycling on high speed LANs while still leaving over an hour
1651 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1652 * between seeding of isn_secret. This is normally set to zero,
1653 * as reseeding should not be necessary.
1657 #define ISN_BYTES_PER_SECOND 1048576
1659 u_char isn_secret
[32];
1660 int isn_last_reseed
;
1664 tcp_new_isn(struct tcpcb
*tp
)
1666 u_int32_t md5_buffer
[4];
1669 /* Seed if this is the first use, reseed if requested. */
1670 if ((isn_last_reseed
== 0) || ((tcp_isn_reseed_interval
> 0) &&
1671 (((u_int
)isn_last_reseed
+ (u_int
)tcp_isn_reseed_interval
*hz
)
1673 read_random_unlimited(&isn_secret
, sizeof isn_secret
);
1674 isn_last_reseed
= ticks
;
1677 /* Compute the md5 hash and return the ISN. */
1679 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_fport
, sizeof(u_short
));
1680 MD5Update(&isn_ctx
, (u_char
*)&tp
->t_inpcb
->inp_lport
, sizeof(u_short
));
1682 if (INP_ISIPV6(tp
->t_inpcb
)) {
1683 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_faddr
,
1684 sizeof(struct in6_addr
));
1685 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->in6p_laddr
,
1686 sizeof(struct in6_addr
));
1690 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_faddr
,
1691 sizeof(struct in_addr
));
1692 MD5Update(&isn_ctx
, (u_char
*) &tp
->t_inpcb
->inp_laddr
,
1693 sizeof(struct in_addr
));
1695 MD5Update(&isn_ctx
, (u_char
*) &isn_secret
, sizeof(isn_secret
));
1696 MD5Final((u_char
*) &md5_buffer
, &isn_ctx
);
1697 new_isn
= (tcp_seq
) md5_buffer
[0];
1698 new_isn
+= ticks
* (ISN_BYTES_PER_SECOND
/ hz
);
1703 * When a source quench is received, close congestion window
1704 * to one segment. We will gradually open it again as we proceed.
1707 tcp_quench(struct inpcb
*inp
, int error
)
1709 struct tcpcb
*tp
= intotcpcb(inp
);
1711 KASSERT(tp
!= NULL
, ("tcp_quench: tp is NULL"));
1712 tp
->snd_cwnd
= tp
->t_maxseg
;
1717 * When a specific ICMP unreachable message is received and the
1718 * connection state is SYN-SENT, drop the connection. This behavior
1719 * is controlled by the icmp_may_rst sysctl.
1722 tcp_drop_syn_sent(struct inpcb
*inp
, int error
)
1724 struct tcpcb
*tp
= intotcpcb(inp
);
1726 KASSERT(tp
!= NULL
, ("tcp_drop_syn_sent: tp is NULL"));
1727 if (tp
->t_state
== TCPS_SYN_SENT
)
1728 tcp_drop(tp
, error
);
1732 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1733 * based on the new value in the route. Also nudge TCP to send something,
1734 * since we know the packet we just sent was dropped.
1735 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1738 tcp_mtudisc(struct inpcb
*inp
, int mtu
)
1740 struct tcpcb
*tp
= intotcpcb(inp
);
1742 struct socket
*so
= inp
->inp_socket
;
1745 boolean_t isipv6
= INP_ISIPV6(inp
);
1747 const boolean_t isipv6
= FALSE
;
1750 KASSERT(tp
!= NULL
, ("tcp_mtudisc: tp is NULL"));
1753 * If no MTU is provided in the ICMP message, use the
1754 * next lower likely value, as specified in RFC 1191.
1759 oldmtu
= tp
->t_maxopd
+
1761 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1762 sizeof(struct tcpiphdr
));
1763 mtu
= ip_next_mtu(oldmtu
, 0);
1767 rt
= tcp_rtlookup6(&inp
->inp_inc
);
1769 rt
= tcp_rtlookup(&inp
->inp_inc
);
1771 if (rt
->rt_rmx
.rmx_mtu
!= 0 && rt
->rt_rmx
.rmx_mtu
< mtu
)
1772 mtu
= rt
->rt_rmx
.rmx_mtu
;
1776 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1777 sizeof(struct tcpiphdr
));
1780 * XXX - The following conditional probably violates the TCP
1781 * spec. The problem is that, since we don't know the
1782 * other end's MSS, we are supposed to use a conservative
1783 * default. But, if we do that, then MTU discovery will
1784 * never actually take place, because the conservative
1785 * default is much less than the MTUs typically seen
1786 * on the Internet today. For the moment, we'll sweep
1787 * this under the carpet.
1789 * The conservative default might not actually be a problem
1790 * if the only case this occurs is when sending an initial
1791 * SYN with options and data to a host we've never talked
1792 * to before. Then, they will reply with an MSS value which
1793 * will get recorded and the new parameters should get
1794 * recomputed. For Further Study.
1796 if (rt
->rt_rmx
.rmx_mssopt
&& rt
->rt_rmx
.rmx_mssopt
< maxopd
)
1797 maxopd
= rt
->rt_rmx
.rmx_mssopt
;
1801 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1802 sizeof(struct tcpiphdr
));
1804 if (tp
->t_maxopd
<= maxopd
)
1806 tp
->t_maxopd
= maxopd
;
1809 if ((tp
->t_flags
& (TF_REQ_TSTMP
| TF_RCVD_TSTMP
| TF_NOOPT
)) ==
1810 (TF_REQ_TSTMP
| TF_RCVD_TSTMP
))
1811 mss
-= TCPOLEN_TSTAMP_APPA
;
1813 /* round down to multiple of MCLBYTES */
1814 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1816 mss
&= ~(MCLBYTES
- 1);
1819 mss
= (mss
/ MCLBYTES
) * MCLBYTES
;
1822 if (so
->so_snd
.ssb_hiwat
< mss
)
1823 mss
= so
->so_snd
.ssb_hiwat
;
1827 tp
->snd_nxt
= tp
->snd_una
;
1829 tcpstat
.tcps_mturesent
++;
1833 * Look-up the routing entry to the peer of this inpcb. If no route
1834 * is found and it cannot be allocated the return NULL. This routine
1835 * is called by TCP routines that access the rmx structure and by tcp_mss
1836 * to get the interface MTU.
1839 tcp_rtlookup(struct in_conninfo
*inc
)
1841 struct route
*ro
= &inc
->inc_route
;
1843 if (ro
->ro_rt
== NULL
|| !(ro
->ro_rt
->rt_flags
& RTF_UP
)) {
1844 /* No route yet, so try to acquire one */
1845 if (inc
->inc_faddr
.s_addr
!= INADDR_ANY
) {
1847 * unused portions of the structure MUST be zero'd
1848 * out because rtalloc() treats it as opaque data
1850 bzero(&ro
->ro_dst
, sizeof(struct sockaddr_in
));
1851 ro
->ro_dst
.sa_family
= AF_INET
;
1852 ro
->ro_dst
.sa_len
= sizeof(struct sockaddr_in
);
1853 ((struct sockaddr_in
*) &ro
->ro_dst
)->sin_addr
=
1863 tcp_rtlookup6(struct in_conninfo
*inc
)
1865 struct route_in6
*ro6
= &inc
->inc6_route
;
1867 if (ro6
->ro_rt
== NULL
|| !(ro6
->ro_rt
->rt_flags
& RTF_UP
)) {
1868 /* No route yet, so try to acquire one */
1869 if (!IN6_IS_ADDR_UNSPECIFIED(&inc
->inc6_faddr
)) {
1871 * unused portions of the structure MUST be zero'd
1872 * out because rtalloc() treats it as opaque data
1874 bzero(&ro6
->ro_dst
, sizeof(struct sockaddr_in6
));
1875 ro6
->ro_dst
.sin6_family
= AF_INET6
;
1876 ro6
->ro_dst
.sin6_len
= sizeof(struct sockaddr_in6
);
1877 ro6
->ro_dst
.sin6_addr
= inc
->inc6_faddr
;
1878 rtalloc((struct route
*)ro6
);
1881 return (ro6
->ro_rt
);
1886 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1888 * This code attempts to calculate the bandwidth-delay product as a
1889 * means of determining the optimal window size to maximize bandwidth,
1890 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1891 * routers. This code also does a fairly good job keeping RTTs in check
1892 * across slow links like modems. We implement an algorithm which is very
1893 * similar (but not meant to be) TCP/Vegas. The code operates on the
1894 * transmitter side of a TCP connection and so only effects the transmit
1895 * side of the connection.
1897 * BACKGROUND: TCP makes no provision for the management of buffer space
1898 * at the end points or at the intermediate routers and switches. A TCP
1899 * stream, whether using NewReno or not, will eventually buffer as
1900 * many packets as it is able and the only reason this typically works is
1901 * due to the fairly small default buffers made available for a connection
1902 * (typicaly 16K or 32K). As machines use larger windows and/or window
1903 * scaling it is now fairly easy for even a single TCP connection to blow-out
1904 * all available buffer space not only on the local interface, but on
1905 * intermediate routers and switches as well. NewReno makes a misguided
1906 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1907 * then backing off, then steadily increasing the window again until another
1908 * failure occurs, ad-infinitum. This results in terrible oscillation that
1909 * is only made worse as network loads increase and the idea of intentionally
1910 * blowing out network buffers is, frankly, a terrible way to manage network
1913 * It is far better to limit the transmit window prior to the failure
1914 * condition being achieved. There are two general ways to do this: First
1915 * you can 'scan' through different transmit window sizes and locate the
1916 * point where the RTT stops increasing, indicating that you have filled the
1917 * pipe, then scan backwards until you note that RTT stops decreasing, then
1918 * repeat ad-infinitum. This method works in principle but has severe
1919 * implementation issues due to RTT variances, timer granularity, and
1920 * instability in the algorithm which can lead to many false positives and
1921 * create oscillations as well as interact badly with other TCP streams
1922 * implementing the same algorithm.
1924 * The second method is to limit the window to the bandwidth delay product
1925 * of the link. This is the method we implement. RTT variances and our
1926 * own manipulation of the congestion window, bwnd, can potentially
1927 * destabilize the algorithm. For this reason we have to stabilize the
1928 * elements used to calculate the window. We do this by using the minimum
1929 * observed RTT, the long term average of the observed bandwidth, and
1930 * by adding two segments worth of slop. It isn't perfect but it is able
1931 * to react to changing conditions and gives us a very stable basis on
1932 * which to extend the algorithm.
1935 tcp_xmit_bandwidth_limit(struct tcpcb
*tp
, tcp_seq ack_seq
)
1944 * If inflight_enable is disabled in the middle of a tcp connection,
1945 * make sure snd_bwnd is effectively disabled.
1947 if (!tcp_inflight_enable
) {
1948 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
1949 tp
->snd_bandwidth
= 0;
1954 * Validate the delta time. If a connection is new or has been idle
1955 * a long time we have to reset the bandwidth calculator.
1959 delta_ticks
= save_ticks
- tp
->t_bw_rtttime
;
1960 if (tp
->t_bw_rtttime
== 0 || delta_ticks
< 0 || delta_ticks
> hz
* 10) {
1961 tp
->t_bw_rtttime
= save_ticks
;
1962 tp
->t_bw_rtseq
= ack_seq
;
1963 if (tp
->snd_bandwidth
== 0)
1964 tp
->snd_bandwidth
= tcp_inflight_start
;
1969 * A delta of at least 1 tick is required. Waiting 2 ticks will
1970 * result in better (bw) accuracy. More than that and the ramp-up
1973 if (delta_ticks
== 0 || delta_ticks
== 1)
1977 * Sanity check, plus ignore pure window update acks.
1979 if ((int)(ack_seq
- tp
->t_bw_rtseq
) <= 0)
1983 * Figure out the bandwidth. Due to the tick granularity this
1984 * is a very rough number and it MUST be averaged over a fairly
1985 * long period of time. XXX we need to take into account a link
1986 * that is not using all available bandwidth, but for now our
1987 * slop will ramp us up if this case occurs and the bandwidth later
1990 ibw
= (int64_t)(ack_seq
- tp
->t_bw_rtseq
) * hz
/ delta_ticks
;
1991 tp
->t_bw_rtttime
= save_ticks
;
1992 tp
->t_bw_rtseq
= ack_seq
;
1993 bw
= ((int64_t)tp
->snd_bandwidth
* 15 + ibw
) >> 4;
1995 tp
->snd_bandwidth
= bw
;
1998 * Calculate the semi-static bandwidth delay product, plus two maximal
1999 * segments. The additional slop puts us squarely in the sweet
2000 * spot and also handles the bandwidth run-up case. Without the
2001 * slop we could be locking ourselves into a lower bandwidth.
2003 * At very high speeds the bw calculation can become overly sensitive
2004 * and error prone when delta_ticks is low (e.g. usually 1). To deal
2005 * with the problem the stab must be scaled to the bw. A stab of 50
2006 * (the default) increases the bw for the purposes of the bwnd
2007 * calculation by 5%.
2009 * Situations Handled:
2010 * (1) Prevents over-queueing of packets on LANs, especially on
2011 * high speed LANs, allowing larger TCP buffers to be
2012 * specified, and also does a good job preventing
2013 * over-queueing of packets over choke points like modems
2014 * (at least for the transmit side).
2016 * (2) Is able to handle changing network loads (bandwidth
2017 * drops so bwnd drops, bandwidth increases so bwnd
2020 * (3) Theoretically should stabilize in the face of multiple
2021 * connections implementing the same algorithm (this may need
2024 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2025 * be adjusted with a sysctl but typically only needs to be on
2026 * very slow connections. A value no smaller then 5 should
2027 * be used, but only reduce this default if you have no other
2031 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2032 bw
+= bw
* tcp_inflight_stab
/ 1000;
2033 bwnd
= (int64_t)bw
* USERTT
/ (hz
<< TCP_RTT_SHIFT
) +
2034 (int)tp
->t_maxseg
* 2;
2037 if (tcp_inflight_debug
> 0) {
2039 if ((u_int
)(save_ticks
- ltime
) >= hz
/ tcp_inflight_debug
) {
2041 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2042 "bwnd %ld delta %d snd_win %ld\n",
2043 tp
, ibw
, bw
, tp
->t_rttvar
, tp
->t_srtt
,
2044 bwnd
, delta_ticks
, tp
->snd_wnd
);
2047 if ((long)bwnd
< tcp_inflight_min
)
2048 bwnd
= tcp_inflight_min
;
2049 if (bwnd
> tcp_inflight_max
)
2050 bwnd
= tcp_inflight_max
;
2051 if ((long)bwnd
< tp
->t_maxseg
* 2)
2052 bwnd
= tp
->t_maxseg
* 2;
2053 tp
->snd_bwnd
= bwnd
;
2057 tcp_rmx_iwsegs(struct tcpcb
*tp
, u_long
*maxsegs
, u_long
*capsegs
)
2060 struct inpcb
*inp
= tp
->t_inpcb
;
2062 boolean_t isipv6
= INP_ISIPV6(inp
);
2064 const boolean_t isipv6
= FALSE
;
2068 if (tcp_iw_maxsegs
< TCP_IW_MAXSEGS_DFLT
)
2069 tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
2070 if (tcp_iw_capsegs
< TCP_IW_CAPSEGS_DFLT
)
2071 tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
2074 rt
= tcp_rtlookup6(&inp
->inp_inc
);
2076 rt
= tcp_rtlookup(&inp
->inp_inc
);
2078 rt
->rt_rmx
.rmx_iwmaxsegs
< TCP_IW_MAXSEGS_DFLT
||
2079 rt
->rt_rmx
.rmx_iwcapsegs
< TCP_IW_CAPSEGS_DFLT
) {
2080 *maxsegs
= tcp_iw_maxsegs
;
2081 *capsegs
= tcp_iw_capsegs
;
2084 *maxsegs
= rt
->rt_rmx
.rmx_iwmaxsegs
;
2085 *capsegs
= rt
->rt_rmx
.rmx_iwcapsegs
;
2089 tcp_initial_window(struct tcpcb
*tp
)
2091 if (tcp_do_rfc3390
) {
2094 * "If the SYN or SYN/ACK is lost, the initial window
2095 * used by a sender after a correctly transmitted SYN
2096 * MUST be one segment consisting of MSS bytes."
2098 * However, we do something a little bit more aggressive
2099 * then RFC3390 here:
2100 * - Only if time spent in the SYN or SYN|ACK retransmition
2101 * >= 3 seconds, the IW is reduced. We do this mainly
2102 * because when RFC3390 is published, the initial RTO is
2103 * still 3 seconds (the threshold we test here), while
2104 * after RFC6298, the initial RTO is 1 second. This
2105 * behaviour probably still falls within the spirit of
2107 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2108 * Mainly to avoid sender and receiver deadlock until
2109 * delayed ACK timer expires. And even RFC2581 does not
2110 * try to reduce IW upon SYN or SYN|ACK retransmition
2114 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2116 if (tp
->t_rxtsyn
>= TCPTV_RTOBASE3
) {
2117 return (2 * tp
->t_maxseg
);
2119 u_long maxsegs
, capsegs
;
2121 tcp_rmx_iwsegs(tp
, &maxsegs
, &capsegs
);
2122 return min(maxsegs
* tp
->t_maxseg
,
2123 max(2 * tp
->t_maxseg
, capsegs
* 1460));
2127 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2129 * Mainly to avoid sender and receiver deadlock
2130 * until delayed ACK timer expires.
2132 return (2 * tp
->t_maxseg
);
2136 #ifdef TCP_SIGNATURE
2138 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2140 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2141 * When called from tcp_input(), we can be sure that th_sum has been
2142 * zeroed out and verified already.
2144 * Return 0 if successful, otherwise return -1.
2146 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2147 * search with the destination IP address, and a 'magic SPI' to be
2148 * determined by the application. This is hardcoded elsewhere to 1179
2149 * right now. Another branch of this code exists which uses the SPD to
2150 * specify per-application flows but it is unstable.
2153 tcpsignature_compute(
2154 struct mbuf
*m
, /* mbuf chain */
2155 int len
, /* length of TCP data */
2156 int optlen
, /* length of TCP options */
2157 u_char
*buf
, /* storage for MD5 digest */
2158 u_int direction
) /* direction of flow */
2160 struct ippseudo ippseudo
;
2164 struct ipovly
*ipovly
;
2165 struct secasvar
*sav
;
2168 struct ip6_hdr
*ip6
;
2169 struct in6_addr in6
;
2175 KASSERT(m
!= NULL
, ("passed NULL mbuf. Game over."));
2176 KASSERT(buf
!= NULL
, ("passed NULL storage pointer for MD5 signature"));
2178 * Extract the destination from the IP header in the mbuf.
2180 ip
= mtod(m
, struct ip
*);
2182 ip6
= NULL
; /* Make the compiler happy. */
2185 * Look up an SADB entry which matches the address found in
2188 switch (IP_VHL_V(ip
->ip_vhl
)) {
2190 sav
= key_allocsa(AF_INET
, (caddr_t
)&ip
->ip_src
, (caddr_t
)&ip
->ip_dst
,
2191 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2194 case (IPV6_VERSION
>> 4):
2195 ip6
= mtod(m
, struct ip6_hdr
*);
2196 sav
= key_allocsa(AF_INET6
, (caddr_t
)&ip6
->ip6_src
, (caddr_t
)&ip6
->ip6_dst
,
2197 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2206 kprintf("%s: SADB lookup failed\n", __func__
);
2212 * Step 1: Update MD5 hash with IP pseudo-header.
2214 * XXX The ippseudo header MUST be digested in network byte order,
2215 * or else we'll fail the regression test. Assume all fields we've
2216 * been doing arithmetic on have been in host byte order.
2217 * XXX One cannot depend on ipovly->ih_len here. When called from
2218 * tcp_output(), the underlying ip_len member has not yet been set.
2220 switch (IP_VHL_V(ip
->ip_vhl
)) {
2222 ipovly
= (struct ipovly
*)ip
;
2223 ippseudo
.ippseudo_src
= ipovly
->ih_src
;
2224 ippseudo
.ippseudo_dst
= ipovly
->ih_dst
;
2225 ippseudo
.ippseudo_pad
= 0;
2226 ippseudo
.ippseudo_p
= IPPROTO_TCP
;
2227 ippseudo
.ippseudo_len
= htons(len
+ sizeof(struct tcphdr
) + optlen
);
2228 MD5Update(&ctx
, (char *)&ippseudo
, sizeof(struct ippseudo
));
2229 th
= (struct tcphdr
*)((u_char
*)ip
+ sizeof(struct ip
));
2230 doff
= sizeof(struct ip
) + sizeof(struct tcphdr
) + optlen
;
2234 * RFC 2385, 2.0 Proposal
2235 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2236 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2237 * extended next header value (to form 32 bits), and 32-bit segment
2239 * Note: Upper-Layer Packet Length comes before Next Header.
2241 case (IPV6_VERSION
>> 4):
2243 in6_clearscope(&in6
);
2244 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2246 in6_clearscope(&in6
);
2247 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2248 plen
= htonl(len
+ sizeof(struct tcphdr
) + optlen
);
2249 MD5Update(&ctx
, (char *)&plen
, sizeof(uint32_t));
2251 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2252 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2253 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2255 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2256 th
= (struct tcphdr
*)((u_char
*)ip6
+ sizeof(struct ip6_hdr
));
2257 doff
= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) + optlen
;
2266 * Step 2: Update MD5 hash with TCP header, excluding options.
2267 * The TCP checksum must be set to zero.
2269 savecsum
= th
->th_sum
;
2271 MD5Update(&ctx
, (char *)th
, sizeof(struct tcphdr
));
2272 th
->th_sum
= savecsum
;
2274 * Step 3: Update MD5 hash with TCP segment data.
2275 * Use m_apply() to avoid an early m_pullup().
2278 m_apply(m
, doff
, len
, tcpsignature_apply
, &ctx
);
2280 * Step 4: Update MD5 hash with shared secret.
2282 MD5Update(&ctx
, _KEYBUF(sav
->key_auth
), _KEYLEN(sav
->key_auth
));
2283 MD5Final(buf
, &ctx
);
2284 key_sa_recordxfer(sav
, m
);
2290 tcpsignature_apply(void *fstate
, void *data
, unsigned int len
)
2293 MD5Update((MD5_CTX
*)fstate
, (unsigned char *)data
, len
);
2296 #endif /* TCP_SIGNATURE */
2299 tcp_drop_sysctl_dispatch(netmsg_t nmsg
)
2301 struct lwkt_msg
*lmsg
= &nmsg
->lmsg
;
2302 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2303 struct sockaddr_storage
*addrs
= lmsg
->u
.ms_resultp
;
2305 struct sockaddr_in
*fin
, *lin
;
2307 struct sockaddr_in6
*fin6
, *lin6
;
2308 struct in6_addr f6
, l6
;
2312 switch (addrs
[0].ss_family
) {
2315 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2316 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2317 error
= in6_embedscope(&f6
, fin6
, NULL
, NULL
);
2320 error
= in6_embedscope(&l6
, lin6
, NULL
, NULL
);
2323 inp
= in6_pcblookup_hash(&tcbinfo
[mycpuid
], &f6
,
2324 fin6
->sin6_port
, &l6
, lin6
->sin6_port
, FALSE
, NULL
);
2329 fin
= (struct sockaddr_in
*)&addrs
[0];
2330 lin
= (struct sockaddr_in
*)&addrs
[1];
2331 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], fin
->sin_addr
,
2332 fin
->sin_port
, lin
->sin_addr
, lin
->sin_port
, FALSE
, NULL
);
2337 * Must not reach here, since the address family was
2338 * checked in sysctl handler.
2340 panic("unknown address family %d", addrs
[0].ss_family
);
2343 struct tcpcb
*tp
= intotcpcb(inp
);
2345 KASSERT((inp
->inp_flags
& INP_WILDCARD
) == 0,
2346 ("in wildcard hash"));
2347 KASSERT(tp
!= NULL
, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2348 KASSERT((tp
->t_flags
& TF_LISTEN
) == 0, ("listen socket"));
2349 tcp_drop(tp
, ECONNABORTED
);
2357 lwkt_replymsg(lmsg
, error
);
2361 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS
)
2363 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2364 struct sockaddr_storage addrs
[2];
2365 struct sockaddr_in
*fin
, *lin
;
2367 struct sockaddr_in6
*fin6
, *lin6
;
2369 struct netmsg_base nmsg
;
2370 struct lwkt_msg
*lmsg
= &nmsg
.lmsg
;
2371 struct lwkt_port
*port
= NULL
;
2380 if (req
->oldptr
!= NULL
|| req
->oldlen
!= 0)
2382 if (req
->newptr
== NULL
)
2384 if (req
->newlen
< sizeof(addrs
))
2386 error
= SYSCTL_IN(req
, &addrs
, sizeof(addrs
));
2390 switch (addrs
[0].ss_family
) {
2393 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2394 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2395 if (fin6
->sin6_len
!= sizeof(struct sockaddr_in6
) ||
2396 lin6
->sin6_len
!= sizeof(struct sockaddr_in6
))
2398 if (IN6_IS_ADDR_V4MAPPED(&fin6
->sin6_addr
) ||
2399 IN6_IS_ADDR_V4MAPPED(&lin6
->sin6_addr
))
2400 return (EADDRNOTAVAIL
);
2402 error
= sa6_embedscope(fin6
, V_ip6_use_defzone
);
2405 error
= sa6_embedscope(lin6
, V_ip6_use_defzone
);
2409 port
= tcp6_addrport();
2414 fin
= (struct sockaddr_in
*)&addrs
[0];
2415 lin
= (struct sockaddr_in
*)&addrs
[1];
2416 if (fin
->sin_len
!= sizeof(struct sockaddr_in
) ||
2417 lin
->sin_len
!= sizeof(struct sockaddr_in
))
2419 port
= tcp_addrport(fin
->sin_addr
.s_addr
, fin
->sin_port
,
2420 lin
->sin_addr
.s_addr
, lin
->sin_port
);
2427 netmsg_init(&nmsg
, NULL
, &curthread
->td_msgport
, 0,
2428 tcp_drop_sysctl_dispatch
);
2429 lmsg
->u
.ms_resultp
= addrs
;
2430 return lwkt_domsg(port
, lmsg
, 0);
2433 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, drop
,
2434 CTLTYPE_STRUCT
| CTLFLAG_WR
| CTLFLAG_SKIP
, NULL
,
2435 0, sysctl_tcp_drop
, "", "Drop TCP connection");
2438 sysctl_tcps_count(SYSCTL_HANDLER_ARGS
)
2440 u_long state_count
[TCP_NSTATES
];
2443 memset(state_count
, 0, sizeof(state_count
));
2444 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2447 for (i
= 0; i
< TCP_NSTATES
; ++i
)
2448 state_count
[i
] += tcpstate_count
[cpu
].tcps_count
[i
];
2451 return sysctl_handle_opaque(oidp
, state_count
, sizeof(state_count
), req
);
2453 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, state_count
,
2454 CTLTYPE_OPAQUE
| CTLFLAG_RD
, NULL
, 0,
2455 sysctl_tcps_count
, "LU", "TCP connection counts by state");
2458 tcp_pcbport_create(struct tcpcb
*tp
)
2462 KASSERT((tp
->t_flags
& TF_LISTEN
) && tp
->t_state
== TCPS_LISTEN
,
2463 ("not a listen tcpcb"));
2465 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache was created"));
2466 tp
->t_pcbport
= kmalloc_cachealign(
2467 sizeof(struct tcp_pcbport
) * netisr_ncpus
, M_PCB
, M_WAITOK
);
2469 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2470 struct inpcbport
*phd
;
2472 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2473 LIST_INIT(&phd
->phd_pcblist
);
2474 /* Though, not used ... */
2475 phd
->phd_port
= tp
->t_inpcb
->inp_lport
;
2480 tcp_pcbport_merge_oncpu(struct tcpcb
*tp
)
2482 struct inpcbport
*phd
;
2486 KASSERT(cpu
< netisr_ncpus
, ("invalid cpu%d", cpu
));
2487 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2489 while ((inp
= LIST_FIRST(&phd
->phd_pcblist
)) != NULL
) {
2490 KASSERT(inp
->inp_phd
== phd
&& inp
->inp_porthash
== NULL
,
2491 ("not on tcpcb port cache"));
2492 LIST_REMOVE(inp
, inp_portlist
);
2493 in_pcbinsporthash_lport(inp
);
2494 KASSERT(inp
->inp_phd
== tp
->t_inpcb
->inp_phd
&&
2495 inp
->inp_porthash
== tp
->t_inpcb
->inp_porthash
,
2496 ("tcpcb port cache merge failed"));
2501 tcp_pcbport_destroy(struct tcpcb
*tp
)
2506 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2507 KASSERT(LIST_EMPTY(&tp
->t_pcbport
[cpu
].t_phd
.phd_pcblist
),
2508 ("tcpcb port cache is not empty"));
2511 kfree(tp
->t_pcbport
, M_PCB
);
2512 tp
->t_pcbport
= NULL
;