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 <net/netmsg2.h>
128 #if !defined(KTR_TCP)
129 #define KTR_TCP KTR_ALL
132 KTR_INFO_MASTER(tcp);
133 KTR_INFO(KTR_TCP, tcp, rxmsg, 0, "tcp getmsg", 0);
134 KTR_INFO(KTR_TCP, tcp, wait, 1, "tcp waitmsg", 0);
135 KTR_INFO(KTR_TCP, tcp, delayed, 2, "tcp execute delayed ops", 0);
136 #define logtcp(name) KTR_LOG(tcp_ ## name)
139 #define TCP_IW_MAXSEGS_DFLT 4
140 #define TCP_IW_CAPSEGS_DFLT 4
142 struct tcp_reass_pcpu
{
144 struct netmsg_base drain_nmsg
;
147 struct inpcbinfo tcbinfo
[MAXCPU
];
148 struct tcpcbackq tcpcbackq
[MAXCPU
];
149 struct tcp_reass_pcpu tcp_reassq
[MAXCPU
];
151 int tcp_mssdflt
= TCP_MSS
;
152 SYSCTL_INT(_net_inet_tcp
, TCPCTL_MSSDFLT
, mssdflt
, CTLFLAG_RW
,
153 &tcp_mssdflt
, 0, "Default TCP Maximum Segment Size");
156 int tcp_v6mssdflt
= TCP6_MSS
;
157 SYSCTL_INT(_net_inet_tcp
, TCPCTL_V6MSSDFLT
, v6mssdflt
, CTLFLAG_RW
,
158 &tcp_v6mssdflt
, 0, "Default TCP Maximum Segment Size for IPv6");
162 * Minimum MSS we accept and use. This prevents DoS attacks where
163 * we are forced to a ridiculous low MSS like 20 and send hundreds
164 * of packets instead of one. The effect scales with the available
165 * bandwidth and quickly saturates the CPU and network interface
166 * with packet generation and sending. Set to zero to disable MINMSS
167 * checking. This setting prevents us from sending too small packets.
169 int tcp_minmss
= TCP_MINMSS
;
170 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, minmss
, CTLFLAG_RW
,
171 &tcp_minmss
, 0, "Minmum TCP Maximum Segment Size");
174 static int tcp_rttdflt
= TCPTV_SRTTDFLT
/ PR_SLOWHZ
;
175 SYSCTL_INT(_net_inet_tcp
, TCPCTL_RTTDFLT
, rttdflt
, CTLFLAG_RW
,
176 &tcp_rttdflt
, 0, "Default maximum TCP Round Trip Time");
179 int tcp_do_rfc1323
= 1;
180 SYSCTL_INT(_net_inet_tcp
, TCPCTL_DO_RFC1323
, rfc1323
, CTLFLAG_RW
,
181 &tcp_do_rfc1323
, 0, "Enable rfc1323 (high performance TCP) extensions");
183 static int tcp_tcbhashsize
= 0;
184 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, tcbhashsize
, CTLFLAG_RD
,
185 &tcp_tcbhashsize
, 0, "Size of TCP control block hashtable");
187 static int do_tcpdrain
= 1;
188 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, do_tcpdrain
, CTLFLAG_RW
, &do_tcpdrain
, 0,
189 "Enable tcp_drain routine for extra help when low on mbufs");
191 static int icmp_may_rst
= 1;
192 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, icmp_may_rst
, CTLFLAG_RW
, &icmp_may_rst
, 0,
193 "Certain ICMP unreachable messages may abort connections in SYN_SENT");
196 * Recommend 20 (6 times in two minutes)
198 * Lower values may cause the sequence space to cycle too quickly and lose
199 * its signed monotonically-increasing nature within the 2-minute TIMEDWAIT
202 static int tcp_isn_reseed_interval
= 20;
203 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, isn_reseed_interval
, CTLFLAG_RW
,
204 &tcp_isn_reseed_interval
, 0, "Seconds between reseeding of ISN secret");
207 * TCP bandwidth limiting sysctls. The inflight limiter is now turned on
208 * by default, but with generous values which should allow maximal
209 * bandwidth. In particular, the slop defaults to 50 (5 packets).
211 * The reason for doing this is that the limiter is the only mechanism we
212 * have which seems to do a really good job preventing receiver RX rings
213 * on network interfaces from getting blown out. Even though GigE/10GigE
214 * is supposed to flow control it looks like either it doesn't actually
215 * do it or Open Source drivers do not properly enable it.
217 * People using the limiter to reduce bottlenecks on slower WAN connections
218 * should set the slop to 20 (2 packets).
220 static int tcp_inflight_enable
= 1;
221 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_enable
, CTLFLAG_RW
,
222 &tcp_inflight_enable
, 0, "Enable automatic TCP inflight data limiting");
224 static int tcp_inflight_debug
= 0;
225 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_debug
, CTLFLAG_RW
,
226 &tcp_inflight_debug
, 0, "Debug TCP inflight calculations");
229 * NOTE: tcp_inflight_start is essentially the starting receive window
230 * for a connection. If set too low then fetches over tcp
231 * connections will take noticably longer to ramp-up over
232 * high-latency connections. 6144 is too low for a default,
233 * use something more reasonable.
235 static int tcp_inflight_start
= 33792;
236 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_start
, CTLFLAG_RW
,
237 &tcp_inflight_start
, 0, "Start value for TCP inflight window");
239 static int tcp_inflight_min
= 6144;
240 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_min
, CTLFLAG_RW
,
241 &tcp_inflight_min
, 0, "Lower bound for TCP inflight window");
243 static int tcp_inflight_max
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
244 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_max
, CTLFLAG_RW
,
245 &tcp_inflight_max
, 0, "Upper bound for TCP inflight window");
247 static int tcp_inflight_stab
= 50;
248 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_stab
, CTLFLAG_RW
,
249 &tcp_inflight_stab
, 0, "Fudge bw 1/10% (50=5%)");
251 static int tcp_inflight_adjrtt
= 2;
252 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, inflight_adjrtt
, CTLFLAG_RW
,
253 &tcp_inflight_adjrtt
, 0, "Slop for rtt 1/(hz*32)");
255 static int tcp_do_rfc3390
= 1;
256 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, rfc3390
, CTLFLAG_RW
,
258 "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
260 static u_long tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
261 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwmaxsegs
, CTLFLAG_RW
,
262 &tcp_iw_maxsegs
, 0, "TCP IW segments max");
264 static u_long tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
265 SYSCTL_ULONG(_net_inet_tcp
, OID_AUTO
, iwcapsegs
, CTLFLAG_RW
,
266 &tcp_iw_capsegs
, 0, "TCP IW segments");
268 int tcp_low_rtobase
= 1;
269 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, low_rtobase
, CTLFLAG_RW
,
270 &tcp_low_rtobase
, 0, "Lowering the Initial RTO (RFC 6298)");
272 static int tcp_do_ncr
= 1;
273 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr
, CTLFLAG_RW
,
274 &tcp_do_ncr
, 0, "Non-Congestion Robustness (RFC 4653)");
276 int tcp_ncr_linklocal
= 0;
277 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_linklocal
, CTLFLAG_RW
,
278 &tcp_ncr_linklocal
, 0,
279 "Enable Non-Congestion Robustness (RFC 4653) on link local network");
281 int tcp_ncr_rxtthresh_max
= 16;
282 SYSCTL_INT(_net_inet_tcp
, OID_AUTO
, ncr_rxtthresh_max
, CTLFLAG_RW
,
283 &tcp_ncr_rxtthresh_max
, 0,
284 "Non-Congestion Robustness (RFC 4653), DupThresh upper limit");
286 static MALLOC_DEFINE(M_TCPTEMP
, "tcptemp", "TCP Templates for Keepalives");
287 static struct malloc_pipe tcptemp_mpipe
;
289 static void tcp_willblock(void);
290 static void tcp_notify (struct inpcb
*, int);
292 struct tcp_stats tcpstats_percpu
[MAXCPU
] __cachealign
;
293 struct tcp_state_count tcpstate_count
[MAXCPU
] __cachealign
;
295 static void tcp_drain_dispatch(netmsg_t nmsg
);
298 sysctl_tcpstats(SYSCTL_HANDLER_ARGS
)
302 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
303 if ((error
= SYSCTL_OUT(req
, &tcpstats_percpu
[cpu
],
304 sizeof(struct tcp_stats
))))
306 if ((error
= SYSCTL_IN(req
, &tcpstats_percpu
[cpu
],
307 sizeof(struct tcp_stats
))))
313 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_STATS
, stats
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
314 0, 0, sysctl_tcpstats
, "S,tcp_stats", "TCP statistics");
317 * Target size of TCP PCB hash tables. Must be a power of two.
319 * Note that this can be overridden by the kernel environment
320 * variable net.inet.tcp.tcbhashsize
323 #define TCBHASHSIZE 512
325 CTASSERT(powerof2(TCBHASHSIZE
));
328 * This is the actual shape of what we allocate using the zone
329 * allocator. Doing it this way allows us to protect both structures
330 * using the same generation count, and also eliminates the overhead
331 * of allocating tcpcbs separately. By hiding the structure here,
332 * we avoid changing most of the rest of the code (although it needs
333 * to be changed, eventually, for greater efficiency).
336 #define ALIGNM1 (ALIGNMENT - 1)
340 char align
[(sizeof(struct inpcb
) + ALIGNM1
) & ~ALIGNM1
];
343 struct tcp_callout inp_tp_rexmt
;
344 struct tcp_callout inp_tp_persist
;
345 struct tcp_callout inp_tp_keep
;
346 struct tcp_callout inp_tp_2msl
;
347 struct tcp_callout inp_tp_delack
;
348 struct netmsg_tcp_timer inp_tp_timermsg
;
349 struct netmsg_base inp_tp_sndmore
;
360 struct inpcbportinfo
*portinfo
;
361 struct inpcbinfo
*ticb
;
362 int hashsize
= TCBHASHSIZE
, portinfo_hsize
;
366 * note: tcptemp is used for keepalives, and it is ok for an
367 * allocation to fail so do not specify MPF_INT.
369 mpipe_init(&tcptemp_mpipe
, M_TCPTEMP
, sizeof(struct tcptemp
),
370 25, -1, 0, NULL
, NULL
, NULL
);
372 tcp_delacktime
= TCPTV_DELACK
;
373 tcp_keepinit
= TCPTV_KEEP_INIT
;
374 tcp_keepidle
= TCPTV_KEEP_IDLE
;
375 tcp_keepintvl
= TCPTV_KEEPINTVL
;
376 tcp_maxpersistidle
= TCPTV_KEEP_IDLE
;
378 tcp_rexmit_min
= TCPTV_MIN
;
379 if (tcp_rexmit_min
< 1) /* if kern.hz is too low */
381 tcp_rexmit_slop
= TCPTV_CPU_VAR
;
383 TUNABLE_INT_FETCH("net.inet.tcp.tcbhashsize", &hashsize
);
384 if (!powerof2(hashsize
)) {
385 kprintf("WARNING: TCB hash size not a power of 2\n");
386 hashsize
= TCBHASHSIZE
; /* safe default */
388 tcp_tcbhashsize
= hashsize
;
390 portinfo_hsize
= 65536 / netisr_ncpus
;
391 if (portinfo_hsize
> hashsize
)
392 portinfo_hsize
= hashsize
;
394 portinfo
= kmalloc(sizeof(*portinfo
) * netisr_ncpus
, M_PCB
,
395 M_WAITOK
| M_CACHEALIGN
);
397 for (cpu
= 0; cpu
< netisr_ncpus
; cpu
++) {
398 ticb
= &tcbinfo
[cpu
];
399 in_pcbinfo_init(ticb
, cpu
, FALSE
);
400 ticb
->hashbase
= hashinit(hashsize
, M_PCB
,
402 in_pcbportinfo_init(&portinfo
[cpu
], portinfo_hsize
, cpu
);
403 in_pcbportinfo_set(ticb
, portinfo
, netisr_ncpus
);
404 ticb
->wildcardhashbase
= hashinit(hashsize
, M_PCB
,
405 &ticb
->wildcardhashmask
);
406 ticb
->localgrphashbase
= hashinit(hashsize
, M_PCB
,
407 &ticb
->localgrphashmask
);
408 ticb
->ipi_size
= sizeof(struct inp_tp
);
409 TAILQ_INIT(&tcpcbackq
[cpu
].head
);
412 tcp_reass_maxseg
= nmbclusters
/ 16;
413 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments", &tcp_reass_maxseg
);
416 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr))
418 #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr))
420 if (max_protohdr
< TCP_MINPROTOHDR
)
421 max_protohdr
= TCP_MINPROTOHDR
;
422 if (max_linkhdr
+ TCP_MINPROTOHDR
> MHLEN
)
424 #undef TCP_MINPROTOHDR
427 * Initialize TCP statistics counters for each CPU.
429 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
)
430 bzero(&tcpstats_percpu
[cpu
], sizeof(struct tcp_stats
));
433 * Initialize netmsgs for TCP drain
435 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
436 netmsg_init(&tcp_reassq
[cpu
].drain_nmsg
, NULL
,
437 &netisr_adone_rport
, MSGF_PRIORITY
, tcp_drain_dispatch
);
441 netisr_register_rollup(tcp_willblock
, NETISR_ROLLUP_PRIO_TCP
);
450 while ((tp
= TAILQ_FIRST(&tcpcbackq
[cpu
].head
)) != NULL
) {
451 KKASSERT(tp
->t_flags
& TF_ONOUTPUTQ
);
452 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
453 TAILQ_REMOVE(&tcpcbackq
[cpu
].head
, tp
, t_outputq
);
459 * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb.
460 * tcp_template used to store this data in mbufs, but we now recopy it out
461 * of the tcpcb each time to conserve mbufs.
464 tcp_fillheaders(struct tcpcb
*tp
, void *ip_ptr
, void *tcp_ptr
, boolean_t tso
)
466 struct inpcb
*inp
= tp
->t_inpcb
;
467 struct tcphdr
*tcp_hdr
= (struct tcphdr
*)tcp_ptr
;
470 if (INP_ISIPV6(inp
)) {
473 ip6
= (struct ip6_hdr
*)ip_ptr
;
474 ip6
->ip6_flow
= (ip6
->ip6_flow
& ~IPV6_FLOWINFO_MASK
) |
475 (inp
->in6p_flowinfo
& IPV6_FLOWINFO_MASK
);
476 ip6
->ip6_vfc
= (ip6
->ip6_vfc
& ~IPV6_VERSION_MASK
) |
477 (IPV6_VERSION
& IPV6_VERSION_MASK
);
478 ip6
->ip6_nxt
= IPPROTO_TCP
;
479 ip6
->ip6_plen
= sizeof(struct tcphdr
);
480 ip6
->ip6_src
= inp
->in6p_laddr
;
481 ip6
->ip6_dst
= inp
->in6p_faddr
;
486 struct ip
*ip
= (struct ip
*) ip_ptr
;
489 ip
->ip_vhl
= IP_VHL_BORING
;
496 ip
->ip_p
= IPPROTO_TCP
;
497 ip
->ip_src
= inp
->inp_laddr
;
498 ip
->ip_dst
= inp
->inp_faddr
;
501 plen
= htons(IPPROTO_TCP
);
503 plen
= htons(sizeof(struct tcphdr
) + IPPROTO_TCP
);
504 tcp_hdr
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
,
505 ip
->ip_dst
.s_addr
, plen
);
508 tcp_hdr
->th_sport
= inp
->inp_lport
;
509 tcp_hdr
->th_dport
= inp
->inp_fport
;
514 tcp_hdr
->th_flags
= 0;
520 * Create template to be used to send tcp packets on a connection.
521 * Allocates an mbuf and fills in a skeletal tcp/ip header. The only
522 * use for this function is in keepalives, which use tcp_respond.
525 tcp_maketemplate(struct tcpcb
*tp
)
529 if ((tmp
= mpipe_alloc_nowait(&tcptemp_mpipe
)) == NULL
)
531 tcp_fillheaders(tp
, &tmp
->tt_ipgen
, &tmp
->tt_t
, FALSE
);
536 tcp_freetemplate(struct tcptemp
*tmp
)
538 mpipe_free(&tcptemp_mpipe
, tmp
);
542 * Send a single message to the TCP at address specified by
543 * the given TCP/IP header. If m == NULL, then we make a copy
544 * of the tcpiphdr at ti and send directly to the addressed host.
545 * This is used to force keep alive messages out using the TCP
546 * template for a connection. If flags are given then we send
547 * a message back to the TCP which originated the * segment ti,
548 * and discard the mbuf containing it and any other attached mbufs.
550 * In any case the ack and sequence number of the transmitted
551 * segment are as specified by the parameters.
553 * NOTE: If m != NULL, then ti must point to *inside* the mbuf.
556 tcp_respond(struct tcpcb
*tp
, void *ipgen
, struct tcphdr
*th
, struct mbuf
*m
,
557 tcp_seq ack
, tcp_seq seq
, int flags
)
561 struct route
*ro
= NULL
;
563 struct ip
*ip
= ipgen
;
566 struct route_in6
*ro6
= NULL
;
567 struct route_in6 sro6
;
568 struct ip6_hdr
*ip6
= ipgen
;
569 struct inpcb
*inp
= NULL
;
570 boolean_t use_tmpro
= TRUE
;
572 boolean_t isipv6
= (IP_VHL_V(ip
->ip_vhl
) == 6);
574 const boolean_t isipv6
= FALSE
;
579 if (!(flags
& TH_RST
)) {
580 win
= ssb_space(&inp
->inp_socket
->so_rcv
);
583 if (win
> (long)TCP_MAXWIN
<< tp
->rcv_scale
)
584 win
= (long)TCP_MAXWIN
<< tp
->rcv_scale
;
587 * Don't use the route cache of a listen socket,
588 * it is not MPSAFE; use temporary route cache.
590 if (tp
->t_state
!= TCPS_LISTEN
) {
592 ro6
= &inp
->in6p_route
;
594 ro
= &inp
->inp_route
;
601 bzero(ro6
, sizeof *ro6
);
604 bzero(ro
, sizeof *ro
);
608 m
= m_gethdr(M_NOWAIT
, MT_HEADER
);
612 m
->m_data
+= max_linkhdr
;
614 bcopy(ip6
, mtod(m
, caddr_t
), sizeof(struct ip6_hdr
));
615 ip6
= mtod(m
, struct ip6_hdr
*);
616 nth
= (struct tcphdr
*)(ip6
+ 1);
618 bcopy(ip
, mtod(m
, caddr_t
), sizeof(struct ip
));
619 ip
= mtod(m
, struct ip
*);
620 nth
= (struct tcphdr
*)(ip
+ 1);
622 bcopy(th
, nth
, sizeof(struct tcphdr
));
627 m
->m_data
= (caddr_t
)ipgen
;
628 /* m_len is set later */
630 #define xchg(a, b, type) { type t; t = a; a = b; b = t; }
632 xchg(ip6
->ip6_dst
, ip6
->ip6_src
, struct in6_addr
);
633 nth
= (struct tcphdr
*)(ip6
+ 1);
635 xchg(ip
->ip_dst
.s_addr
, ip
->ip_src
.s_addr
, n_long
);
636 nth
= (struct tcphdr
*)(ip
+ 1);
640 * this is usually a case when an extension header
641 * exists between the IPv6 header and the
644 nth
->th_sport
= th
->th_sport
;
645 nth
->th_dport
= th
->th_dport
;
647 xchg(nth
->th_dport
, nth
->th_sport
, n_short
);
652 ip6
->ip6_vfc
= IPV6_VERSION
;
653 ip6
->ip6_nxt
= IPPROTO_TCP
;
654 ip6
->ip6_plen
= htons((u_short
)(sizeof(struct tcphdr
) + tlen
));
655 tlen
+= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
);
657 tlen
+= sizeof(struct tcpiphdr
);
658 ip
->ip_len
= htons(tlen
);
659 ip
->ip_ttl
= ip_defttl
;
662 m
->m_pkthdr
.len
= tlen
;
663 m
->m_pkthdr
.rcvif
= NULL
;
664 nth
->th_seq
= htonl(seq
);
665 nth
->th_ack
= htonl(ack
);
667 nth
->th_off
= sizeof(struct tcphdr
) >> 2;
668 nth
->th_flags
= flags
;
670 nth
->th_win
= htons((u_short
) (win
>> tp
->rcv_scale
));
672 nth
->th_win
= htons((u_short
)win
);
676 nth
->th_sum
= in6_cksum(m
, IPPROTO_TCP
,
677 sizeof(struct ip6_hdr
),
678 tlen
- sizeof(struct ip6_hdr
));
679 ip6
->ip6_hlim
= in6_selecthlim(inp
,
680 (ro6
&& ro6
->ro_rt
) ? ro6
->ro_rt
->rt_ifp
: NULL
);
682 nth
->th_sum
= in_pseudo(ip
->ip_src
.s_addr
, ip
->ip_dst
.s_addr
,
683 htons((u_short
)(tlen
- sizeof(struct ip
) + ip
->ip_p
)));
684 m
->m_pkthdr
.csum_flags
= CSUM_TCP
;
685 m
->m_pkthdr
.csum_data
= offsetof(struct tcphdr
, th_sum
);
686 m
->m_pkthdr
.csum_thlen
= sizeof(struct tcphdr
);
689 if (tp
== NULL
|| (inp
->inp_socket
->so_options
& SO_DEBUG
))
690 tcp_trace(TA_OUTPUT
, 0, tp
, mtod(m
, void *), th
, 0);
693 ip6_output(m
, NULL
, ro6
, ipflags
, NULL
, NULL
, inp
);
694 if ((ro6
== &sro6
) && (ro6
->ro_rt
!= NULL
)) {
699 if (inp
!= NULL
&& (inp
->inp_flags
& INP_HASH
))
700 m_sethash(m
, inp
->inp_hashval
);
701 ipflags
|= IP_DEBUGROUTE
;
702 ip_output(m
, NULL
, ro
, ipflags
, NULL
, inp
);
703 if ((ro
== &sro
) && (ro
->ro_rt
!= NULL
)) {
711 * Create a new TCP control block, making an
712 * empty reassembly queue and hooking it to the argument
713 * protocol control block. The `inp' parameter must have
714 * come from the zone allocator set up in tcp_init().
717 tcp_newtcpcb(struct inpcb
*inp
)
722 boolean_t isipv6
= INP_ISIPV6(inp
);
724 const boolean_t isipv6
= FALSE
;
727 it
= (struct inp_tp
*)inp
;
729 bzero(tp
, sizeof(struct tcpcb
));
730 TAILQ_INIT(&tp
->t_segq
);
731 tp
->t_maxseg
= tp
->t_maxopd
= isipv6
? tcp_v6mssdflt
: tcp_mssdflt
;
732 tp
->t_rxtthresh
= tcprexmtthresh
;
734 /* Set up our timeouts. */
735 tp
->tt_rexmt
= &it
->inp_tp_rexmt
;
736 tp
->tt_persist
= &it
->inp_tp_persist
;
737 tp
->tt_keep
= &it
->inp_tp_keep
;
738 tp
->tt_2msl
= &it
->inp_tp_2msl
;
739 tp
->tt_delack
= &it
->inp_tp_delack
;
743 * Zero out timer message. We don't create it here,
744 * since the current CPU may not be the owner of this
747 tp
->tt_msg
= &it
->inp_tp_timermsg
;
748 bzero(tp
->tt_msg
, sizeof(*tp
->tt_msg
));
750 tp
->t_keepinit
= tcp_keepinit
;
751 tp
->t_keepidle
= tcp_keepidle
;
752 tp
->t_keepintvl
= tcp_keepintvl
;
753 tp
->t_keepcnt
= tcp_keepcnt
;
754 tp
->t_maxidle
= tp
->t_keepintvl
* tp
->t_keepcnt
;
757 tp
->t_flags
|= TF_NCR
;
759 tp
->t_flags
|= (TF_REQ_SCALE
| TF_REQ_TSTMP
);
761 tp
->t_inpcb
= inp
; /* XXX */
764 * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no
765 * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives
766 * reasonable initial retransmit time.
768 tp
->t_srtt
= TCPTV_SRTTBASE
;
770 ((TCPTV_RTOBASE
- TCPTV_SRTTBASE
) << TCP_RTTVAR_SHIFT
) / 4;
771 tp
->t_rttmin
= tcp_rexmit_min
;
772 tp
->t_rxtcur
= TCPTV_RTOBASE
;
773 tp
->snd_cwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
774 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
775 tp
->snd_ssthresh
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
776 tp
->snd_last
= ticks
;
777 tp
->t_rcvtime
= ticks
;
779 * IPv4 TTL initialization is necessary for an IPv6 socket as well,
780 * because the socket may be bound to an IPv6 wildcard address,
781 * which may match an IPv4-mapped IPv6 address.
783 inp
->inp_ip_ttl
= ip_defttl
;
785 tcp_sack_tcpcb_init(tp
);
787 tp
->tt_sndmore
= &it
->inp_tp_sndmore
;
792 * Drop a TCP connection, reporting the specified error.
793 * If connection is synchronized, then send a RST to peer.
796 tcp_drop(struct tcpcb
*tp
, int error
)
798 struct socket
*so
= tp
->t_inpcb
->inp_socket
;
800 if (TCPS_HAVERCVDSYN(tp
->t_state
)) {
801 TCP_STATE_CHANGE(tp
, TCPS_CLOSED
);
803 tcpstat
.tcps_drops
++;
805 tcpstat
.tcps_conndrops
++;
806 if (error
== ETIMEDOUT
&& tp
->t_softerror
)
807 error
= tp
->t_softerror
;
808 so
->so_error
= error
;
809 return (tcp_close(tp
));
812 struct netmsg_listen_detach
{
813 struct netmsg_base base
;
815 struct tcpcb
*nm_tp_inh
;
819 tcp_listen_detach_handler(netmsg_t msg
)
821 struct netmsg_listen_detach
*nmsg
= (struct netmsg_listen_detach
*)msg
;
822 struct tcpcb
*tp
= nmsg
->nm_tp
;
823 int cpu
= mycpuid
, nextcpu
;
825 if (tp
->t_flags
& TF_LISTEN
) {
826 syncache_destroy(tp
, nmsg
->nm_tp_inh
);
827 tcp_pcbport_merge_oncpu(tp
);
830 in_pcbremwildcardhash_oncpu(tp
->t_inpcb
, &tcbinfo
[cpu
]);
833 if (nextcpu
< netisr_ncpus
)
834 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nmsg
->base
.lmsg
);
836 lwkt_replymsg(&nmsg
->base
.lmsg
, 0);
840 * Close a TCP control block:
841 * discard all space held by the tcp
842 * discard internet protocol block
843 * wake up any sleepers
846 tcp_close(struct tcpcb
*tp
)
849 struct inpcb
*inp
= tp
->t_inpcb
;
850 struct inpcb
*inp_inh
= NULL
;
851 struct tcpcb
*tp_inh
= NULL
;
852 struct socket
*so
= inp
->inp_socket
;
854 boolean_t dosavessthresh
;
856 boolean_t isipv6
= INP_ISIPV6(inp
);
858 const boolean_t isipv6
= FALSE
;
861 if (tp
->t_flags
& TF_LISTEN
) {
863 * Pending socket/syncache inheritance
865 * If this is a listen(2) socket, find another listen(2)
866 * socket in the same local group, which could inherit
867 * the syncache and sockets pending on the completion
868 * and incompletion queues.
871 * Currently the inheritance could only happen on the
872 * listen(2) sockets w/ SO_REUSEPORT set.
875 inp_inh
= in_pcblocalgroup_last(&tcbinfo
[0], inp
);
877 tp_inh
= intotcpcb(inp_inh
);
881 * INP_WILDCARD indicates that listen(2) has been called on
882 * this socket. This implies:
883 * - A wildcard inp's hash is replicated for each protocol thread.
884 * - Syncache for this inp grows independently in each protocol
886 * - There is more than one cpu
888 * We have to chain a message to the rest of the protocol threads
889 * to cleanup the wildcard hash and the syncache. The cleanup
890 * in the current protocol thread is defered till the end of this
891 * function (syncache_destroy and in_pcbdetach).
894 * After cleanup the inp's hash and syncache entries, this inp will
895 * no longer be available to the rest of the protocol threads, so we
896 * are safe to whack the inp in the following code.
898 if ((inp
->inp_flags
& INP_WILDCARD
) && netisr_ncpus
> 1) {
899 struct netmsg_listen_detach nmsg
;
901 KKASSERT(so
->so_port
== netisr_cpuport(0));
903 KKASSERT(inp
->inp_pcbinfo
== &tcbinfo
[0]);
905 netmsg_init(&nmsg
.base
, NULL
, &curthread
->td_msgport
,
906 MSGF_PRIORITY
, tcp_listen_detach_handler
);
908 nmsg
.nm_tp_inh
= tp_inh
;
909 lwkt_domsg(netisr_cpuport(1), &nmsg
.base
.lmsg
, 0);
915 * Make sure that all of our timers are stopped before we
916 * delete the PCB. For listen TCP socket (tp->tt_msg == NULL),
917 * timers are never used. If timer message is never created
918 * (tp->tt_msg->tt_tcb == NULL), timers are never used too.
920 if (tp
->tt_msg
!= NULL
&& tp
->tt_msg
->tt_tcb
!= NULL
) {
921 tcp_callout_terminate(tp
, tp
->tt_rexmt
);
922 tcp_callout_terminate(tp
, tp
->tt_persist
);
923 tcp_callout_terminate(tp
, tp
->tt_keep
);
924 tcp_callout_terminate(tp
, tp
->tt_2msl
);
925 tcp_callout_terminate(tp
, tp
->tt_delack
);
928 if (tp
->t_flags
& TF_ONOUTPUTQ
) {
929 KKASSERT(tp
->tt_cpu
== mycpu
->gd_cpuid
);
930 TAILQ_REMOVE(&tcpcbackq
[tp
->tt_cpu
].head
, tp
, t_outputq
);
931 tp
->t_flags
&= ~TF_ONOUTPUTQ
;
935 * If we got enough samples through the srtt filter,
936 * save the rtt and rttvar in the routing entry.
937 * 'Enough' is arbitrarily defined as the 16 samples.
938 * 16 samples is enough for the srtt filter to converge
939 * to within 5% of the correct value; fewer samples and
940 * we could save a very bogus rtt.
942 * Don't update the default route's characteristics and don't
943 * update anything that the user "locked".
945 if (tp
->t_rttupdated
>= 16) {
949 struct sockaddr_in6
*sin6
;
951 if ((rt
= inp
->in6p_route
.ro_rt
) == NULL
)
953 sin6
= (struct sockaddr_in6
*)rt_key(rt
);
954 if (IN6_IS_ADDR_UNSPECIFIED(&sin6
->sin6_addr
))
957 if ((rt
= inp
->inp_route
.ro_rt
) == NULL
||
958 ((struct sockaddr_in
*)rt_key(rt
))->
959 sin_addr
.s_addr
== INADDR_ANY
)
962 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTT
)) {
963 i
= tp
->t_srtt
* (RTM_RTTUNIT
/ (hz
* TCP_RTT_SCALE
));
964 if (rt
->rt_rmx
.rmx_rtt
&& i
)
966 * filter this update to half the old & half
967 * the new values, converting scale.
968 * See route.h and tcp_var.h for a
969 * description of the scaling constants.
972 (rt
->rt_rmx
.rmx_rtt
+ i
) / 2;
974 rt
->rt_rmx
.rmx_rtt
= i
;
975 tcpstat
.tcps_cachedrtt
++;
977 if (!(rt
->rt_rmx
.rmx_locks
& RTV_RTTVAR
)) {
979 (RTM_RTTUNIT
/ (hz
* TCP_RTTVAR_SCALE
));
980 if (rt
->rt_rmx
.rmx_rttvar
&& i
)
981 rt
->rt_rmx
.rmx_rttvar
=
982 (rt
->rt_rmx
.rmx_rttvar
+ i
) / 2;
984 rt
->rt_rmx
.rmx_rttvar
= i
;
985 tcpstat
.tcps_cachedrttvar
++;
988 * The old comment here said:
989 * update the pipelimit (ssthresh) if it has been updated
990 * already or if a pipesize was specified & the threshhold
991 * got below half the pipesize. I.e., wait for bad news
992 * before we start updating, then update on both good
995 * But we want to save the ssthresh even if no pipesize is
996 * specified explicitly in the route, because such
997 * connections still have an implicit pipesize specified
998 * by the global tcp_sendspace. In the absence of a reliable
999 * way to calculate the pipesize, it will have to do.
1001 i
= tp
->snd_ssthresh
;
1002 if (rt
->rt_rmx
.rmx_sendpipe
!= 0)
1003 dosavessthresh
= (i
< rt
->rt_rmx
.rmx_sendpipe
/2);
1005 dosavessthresh
= (i
< so
->so_snd
.ssb_hiwat
/2);
1006 if (dosavessthresh
||
1007 (!(rt
->rt_rmx
.rmx_locks
& RTV_SSTHRESH
) && (i
!= 0) &&
1008 (rt
->rt_rmx
.rmx_ssthresh
!= 0))) {
1010 * convert the limit from user data bytes to
1011 * packets then to packet data bytes.
1013 i
= (i
+ tp
->t_maxseg
/ 2) / tp
->t_maxseg
;
1018 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1019 sizeof(struct tcpiphdr
));
1020 if (rt
->rt_rmx
.rmx_ssthresh
)
1021 rt
->rt_rmx
.rmx_ssthresh
=
1022 (rt
->rt_rmx
.rmx_ssthresh
+ i
) / 2;
1024 rt
->rt_rmx
.rmx_ssthresh
= i
;
1025 tcpstat
.tcps_cachedssthresh
++;
1030 /* free the reassembly queue, if any */
1031 while((q
= TAILQ_FIRST(&tp
->t_segq
)) != NULL
) {
1032 TAILQ_REMOVE(&tp
->t_segq
, q
, tqe_q
);
1035 atomic_add_int(&tcp_reass_qsize
, -1);
1037 /* throw away SACK blocks in scoreboard*/
1038 if (TCP_DO_SACK(tp
))
1039 tcp_sack_destroy(&tp
->scb
);
1041 inp
->inp_ppcb
= NULL
;
1042 soisdisconnected(so
);
1043 /* note: pcb detached later on */
1045 tcp_destroy_timermsg(tp
);
1046 tcp_output_cancel(tp
);
1048 if (tp
->t_flags
& TF_LISTEN
) {
1049 syncache_destroy(tp
, tp_inh
);
1050 tcp_pcbport_merge_oncpu(tp
);
1051 tcp_pcbport_destroy(tp
);
1052 if (inp_inh
!= NULL
&& inp_inh
->inp_socket
!= NULL
) {
1054 * Pending sockets inheritance only needs
1055 * to be done once in the current thread,
1058 soinherit(so
, inp_inh
->inp_socket
);
1061 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache is not destroyed"));
1063 so_async_rcvd_drop(so
);
1064 /* Drop the reference for the asynchronized pru_rcvd */
1069 * - Remove self from listen tcpcb per-cpu port cache _before_
1071 * - pcbdetach removes any wildcard hash entry on the current CPU.
1073 tcp_pcbport_remove(inp
);
1081 tcpstat
.tcps_closed
++;
1086 * Walk the tcpbs, if existing, and flush the reassembly queue,
1087 * if there is one...
1090 tcp_drain_oncpu(struct inpcbinfo
*pcbinfo
)
1092 struct inpcbhead
*head
= &pcbinfo
->pcblisthead
;
1096 * Since we run in netisr, it is MP safe, even if
1097 * we block during the inpcb list iteration, i.e.
1098 * we don't need to use inpcb marker here.
1100 ASSERT_NETISR_NCPUS(pcbinfo
->cpu
);
1102 LIST_FOREACH(inpb
, head
, inp_list
) {
1104 struct tseg_qent
*te
;
1106 if (inpb
->inp_flags
& INP_PLACEMARKER
)
1109 tcpb
= intotcpcb(inpb
);
1110 KASSERT(tcpb
!= NULL
, ("tcp_drain_oncpu: tcpb is NULL"));
1112 if ((te
= TAILQ_FIRST(&tcpb
->t_segq
)) != NULL
) {
1113 TAILQ_REMOVE(&tcpb
->t_segq
, te
, tqe_q
);
1114 if (te
->tqe_th
->th_flags
& TH_FIN
)
1115 tcpb
->t_flags
&= ~TF_QUEDFIN
;
1118 atomic_add_int(&tcp_reass_qsize
, -1);
1125 tcp_drain_dispatch(netmsg_t nmsg
)
1128 lwkt_replymsg(&nmsg
->lmsg
, 0); /* reply ASAP */
1131 tcp_drain_oncpu(&tcbinfo
[mycpuid
]);
1132 tcp_reassq
[mycpuid
].draining
= 0;
1136 tcp_drain_ipi(void *arg __unused
)
1139 struct lwkt_msg
*msg
= &tcp_reassq
[cpu
].drain_nmsg
.lmsg
;
1142 if (msg
->ms_flags
& MSGF_DONE
)
1143 lwkt_sendmsg_oncpu(netisr_cpuport(cpu
), msg
);
1156 if (tcp_reass_qsize
== 0)
1159 CPUMASK_ASSBMASK(mask
, netisr_ncpus
);
1160 CPUMASK_ANDMASK(mask
, smp_active_mask
);
1163 if (IN_NETISR_NCPUS(cpu
)) {
1164 tcp_drain_oncpu(&tcbinfo
[cpu
]);
1165 CPUMASK_NANDBIT(mask
, cpu
);
1168 if (tcp_reass_qsize
< netisr_ncpus
) {
1169 /* Does not worth the trouble. */
1173 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
1174 if (!CPUMASK_TESTBIT(mask
, cpu
))
1177 if (tcp_reassq
[cpu
].draining
) {
1178 /* Draining; skip this cpu. */
1179 CPUMASK_NANDBIT(mask
, cpu
);
1182 tcp_reassq
[cpu
].draining
= 1;
1185 if (CPUMASK_TESTNZERO(mask
))
1186 lwkt_send_ipiq_mask(mask
, tcp_drain_ipi
, NULL
);
1190 * Notify a tcp user of an asynchronous error;
1191 * store error as soft error, but wake up user
1192 * (for now, won't do anything until can select for soft error).
1194 * Do not wake up user since there currently is no mechanism for
1195 * reporting soft errors (yet - a kqueue filter may be added).
1198 tcp_notify(struct inpcb
*inp
, int error
)
1200 struct tcpcb
*tp
= intotcpcb(inp
);
1203 * Ignore some errors if we are hooked up.
1204 * If connection hasn't completed, has retransmitted several times,
1205 * and receives a second error, give up now. This is better
1206 * than waiting a long time to establish a connection that
1207 * can never complete.
1209 if (tp
->t_state
== TCPS_ESTABLISHED
&&
1210 (error
== EHOSTUNREACH
|| error
== ENETUNREACH
||
1211 error
== EHOSTDOWN
)) {
1213 } else if (tp
->t_state
< TCPS_ESTABLISHED
&& tp
->t_rxtshift
> 3 &&
1215 tcp_drop(tp
, error
);
1217 tp
->t_softerror
= error
;
1219 wakeup(&so
->so_timeo
);
1226 tcp_pcblist(SYSCTL_HANDLER_ARGS
)
1229 struct inpcb
*marker
;
1237 * The process of preparing the TCB list is too time-consuming and
1238 * resource-intensive to repeat twice on every request.
1240 if (req
->oldptr
== NULL
) {
1241 for (ccpu
= 0; ccpu
< netisr_ncpus
; ++ccpu
)
1242 n
+= tcbinfo
[ccpu
].ipi_count
;
1243 req
->oldidx
= (n
+ n
/8 + 10) * sizeof(struct xtcpcb
);
1247 if (req
->newptr
!= NULL
)
1250 marker
= kmalloc(sizeof(struct inpcb
), M_TEMP
, M_WAITOK
|M_ZERO
);
1251 marker
->inp_flags
|= INP_PLACEMARKER
;
1254 * OK, now we're committed to doing something. Run the inpcb list
1255 * for each cpu in the system and construct the output. Use a
1256 * list placemarker to deal with list changes occuring during
1257 * copyout blockages (but otherwise depend on being on the correct
1258 * cpu to avoid races).
1260 origcpu
= mycpu
->gd_cpuid
;
1261 for (ccpu
= 0; ccpu
< netisr_ncpus
&& error
== 0; ++ccpu
) {
1265 lwkt_migratecpu(ccpu
);
1267 n
= tcbinfo
[ccpu
].ipi_count
;
1269 LIST_INSERT_HEAD(&tcbinfo
[ccpu
].pcblisthead
, marker
, inp_list
);
1271 while ((inp
= LIST_NEXT(marker
, inp_list
)) != NULL
&& i
< n
) {
1273 * process a snapshot of pcbs, ignoring placemarkers
1274 * and using our own to allow SYSCTL_OUT to block.
1276 LIST_REMOVE(marker
, inp_list
);
1277 LIST_INSERT_AFTER(inp
, marker
, inp_list
);
1279 if (inp
->inp_flags
& INP_PLACEMARKER
)
1281 if (prison_xinpcb(req
->td
, inp
))
1284 xt
.xt_len
= sizeof xt
;
1285 bcopy(inp
, &xt
.xt_inp
, sizeof *inp
);
1286 inp_ppcb
= inp
->inp_ppcb
;
1287 if (inp_ppcb
!= NULL
)
1288 bcopy(inp_ppcb
, &xt
.xt_tp
, sizeof xt
.xt_tp
);
1290 bzero(&xt
.xt_tp
, sizeof xt
.xt_tp
);
1291 if (inp
->inp_socket
)
1292 sotoxsocket(inp
->inp_socket
, &xt
.xt_socket
);
1293 if ((error
= SYSCTL_OUT(req
, &xt
, sizeof xt
)) != 0)
1297 LIST_REMOVE(marker
, inp_list
);
1298 if (error
== 0 && i
< n
) {
1299 bzero(&xt
, sizeof xt
);
1300 xt
.xt_len
= sizeof xt
;
1302 error
= SYSCTL_OUT(req
, &xt
, sizeof xt
);
1311 * Make sure we are on the same cpu we were on originally, since
1312 * higher level callers expect this. Also don't pollute caches with
1313 * migrated userland data by (eventually) returning to userland
1314 * on a different cpu.
1316 lwkt_migratecpu(origcpu
);
1317 kfree(marker
, M_TEMP
);
1321 SYSCTL_PROC(_net_inet_tcp
, TCPCTL_PCBLIST
, pcblist
, CTLFLAG_RD
, 0, 0,
1322 tcp_pcblist
, "S,xtcpcb", "List of active TCP connections");
1325 tcp_getcred(SYSCTL_HANDLER_ARGS
)
1327 struct sockaddr_in addrs
[2];
1328 struct ucred cred0
, *cred
= NULL
;
1330 int cpu
, origcpu
, error
;
1332 error
= caps_priv_check_td(req
->td
, SYSCAP_RESTRICTEDROOT
);
1335 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1340 cpu
= tcp_addrcpu(addrs
[1].sin_addr
.s_addr
, addrs
[1].sin_port
,
1341 addrs
[0].sin_addr
.s_addr
, addrs
[0].sin_port
);
1343 lwkt_migratecpu(cpu
);
1345 inp
= in_pcblookup_hash(&tcbinfo
[cpu
], addrs
[1].sin_addr
,
1346 addrs
[1].sin_port
, addrs
[0].sin_addr
, addrs
[0].sin_port
, 0, NULL
);
1347 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1349 } else if (inp
->inp_socket
->so_cred
!= NULL
) {
1350 cred0
= *(inp
->inp_socket
->so_cred
);
1354 lwkt_migratecpu(origcpu
);
1359 return SYSCTL_OUT(req
, cred
, sizeof(struct ucred
));
1362 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1363 0, 0, tcp_getcred
, "S,ucred", "Get the ucred of a TCP connection");
1367 tcp6_getcred(SYSCTL_HANDLER_ARGS
)
1369 struct sockaddr_in6 addrs
[2];
1373 error
= caps_priv_check_td(req
->td
, SYSCAP_RESTRICTEDROOT
);
1376 error
= SYSCTL_IN(req
, addrs
, sizeof addrs
);
1380 inp
= in6_pcblookup_hash(&tcbinfo
[0],
1381 &addrs
[1].sin6_addr
, addrs
[1].sin6_port
,
1382 &addrs
[0].sin6_addr
, addrs
[0].sin6_port
, 0, NULL
);
1383 if (inp
== NULL
|| inp
->inp_socket
== NULL
) {
1387 error
= SYSCTL_OUT(req
, inp
->inp_socket
->so_cred
, sizeof(struct ucred
));
1393 SYSCTL_PROC(_net_inet6_tcp6
, OID_AUTO
, getcred
, (CTLTYPE_OPAQUE
| CTLFLAG_RW
),
1395 tcp6_getcred
, "S,ucred", "Get the ucred of a TCP6 connection");
1398 struct netmsg_tcp_notify
{
1399 struct netmsg_base base
;
1400 inp_notify_t nm_notify
;
1401 struct in_addr nm_faddr
;
1406 tcp_notifyall_oncpu(netmsg_t msg
)
1408 struct netmsg_tcp_notify
*nm
= (struct netmsg_tcp_notify
*)msg
;
1411 ASSERT_NETISR_NCPUS(mycpuid
);
1413 in_pcbnotifyall(&tcbinfo
[mycpuid
], nm
->nm_faddr
,
1414 nm
->nm_arg
, nm
->nm_notify
);
1416 nextcpu
= mycpuid
+ 1;
1417 if (nextcpu
< netisr_ncpus
)
1418 lwkt_forwardmsg(netisr_cpuport(nextcpu
), &nm
->base
.lmsg
);
1420 lwkt_replymsg(&nm
->base
.lmsg
, 0);
1424 tcp_get_inpnotify(int cmd
, const struct sockaddr
*sa
,
1425 int *arg
, struct ip
**ip0
, int *cpuid
)
1427 struct ip
*ip
= *ip0
;
1428 struct in_addr faddr
;
1429 inp_notify_t notify
= tcp_notify
;
1431 faddr
= ((const struct sockaddr_in
*)sa
)->sin_addr
;
1432 if (sa
->sa_family
!= AF_INET
|| faddr
.s_addr
== INADDR_ANY
)
1435 *arg
= inetctlerrmap
[cmd
];
1436 if (cmd
== PRC_QUENCH
) {
1437 notify
= tcp_quench
;
1438 } else if (icmp_may_rst
&&
1439 (cmd
== PRC_UNREACH_ADMIN_PROHIB
||
1440 cmd
== PRC_UNREACH_PORT
||
1441 cmd
== PRC_TIMXCEED_INTRANS
) &&
1443 notify
= tcp_drop_syn_sent
;
1444 } else if (cmd
== PRC_MSGSIZE
) {
1445 const struct icmp
*icmp
= (const struct icmp
*)
1446 ((caddr_t
)ip
- offsetof(struct icmp
, icmp_ip
));
1448 *arg
= ntohs(icmp
->icmp_nextmtu
);
1449 notify
= tcp_mtudisc
;
1450 } else if (PRC_IS_REDIRECT(cmd
)) {
1452 notify
= in_rtchange
;
1453 } else if (cmd
== PRC_HOSTDEAD
) {
1455 } else if ((unsigned)cmd
>= PRC_NCMDS
|| inetctlerrmap
[cmd
] == 0) {
1459 if (cpuid
!= NULL
) {
1461 /* Go through all effective netisr CPUs. */
1462 *cpuid
= netisr_ncpus
;
1464 const struct tcphdr
*th
;
1466 th
= (const struct tcphdr
*)
1467 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1468 *cpuid
= tcp_addrcpu(faddr
.s_addr
, th
->th_dport
,
1469 ip
->ip_src
.s_addr
, th
->th_sport
);
1478 tcp_ctlinput(netmsg_t msg
)
1480 int cmd
= msg
->ctlinput
.nm_cmd
;
1481 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1482 struct ip
*ip
= msg
->ctlinput
.nm_extra
;
1483 struct in_addr faddr
;
1484 inp_notify_t notify
;
1487 ASSERT_NETISR_NCPUS(mycpuid
);
1489 notify
= tcp_get_inpnotify(cmd
, sa
, &arg
, &ip
, &cpuid
);
1493 faddr
= ((struct sockaddr_in
*)sa
)->sin_addr
;
1495 const struct tcphdr
*th
;
1498 if (cpuid
!= mycpuid
)
1501 th
= (const struct tcphdr
*)
1502 ((caddr_t
)ip
+ (IP_VHL_HL(ip
->ip_vhl
) << 2));
1503 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], faddr
, th
->th_dport
,
1504 ip
->ip_src
, th
->th_sport
, 0, NULL
);
1505 if (inp
!= NULL
&& inp
->inp_socket
!= NULL
) {
1506 tcp_seq icmpseq
= htonl(th
->th_seq
);
1507 struct tcpcb
*tp
= intotcpcb(inp
);
1509 if (SEQ_GEQ(icmpseq
, tp
->snd_una
) &&
1510 SEQ_LT(icmpseq
, tp
->snd_max
))
1513 struct in_conninfo inc
;
1515 inc
.inc_fport
= th
->th_dport
;
1516 inc
.inc_lport
= th
->th_sport
;
1517 inc
.inc_faddr
= faddr
;
1518 inc
.inc_laddr
= ip
->ip_src
;
1522 syncache_unreach(&inc
, th
);
1524 } else if (msg
->ctlinput
.nm_direct
) {
1525 if (cpuid
!= netisr_ncpus
&& cpuid
!= mycpuid
)
1528 in_pcbnotifyall(&tcbinfo
[mycpuid
], faddr
, arg
, notify
);
1530 struct netmsg_tcp_notify
*nm
;
1533 nm
= kmalloc(sizeof(*nm
), M_LWKTMSG
, M_INTWAIT
);
1534 netmsg_init(&nm
->base
, NULL
, &netisr_afree_rport
,
1535 0, tcp_notifyall_oncpu
);
1536 nm
->nm_faddr
= faddr
;
1538 nm
->nm_notify
= notify
;
1540 lwkt_sendmsg(netisr_cpuport(0), &nm
->base
.lmsg
);
1543 lwkt_replymsg(&msg
->lmsg
, 0);
1549 tcp6_ctlinput(netmsg_t msg
)
1551 int cmd
= msg
->ctlinput
.nm_cmd
;
1552 struct sockaddr
*sa
= msg
->ctlinput
.nm_arg
;
1553 void *d
= msg
->ctlinput
.nm_extra
;
1555 inp_notify_t notify
= tcp_notify
;
1556 struct ip6_hdr
*ip6
;
1558 struct ip6ctlparam
*ip6cp
= NULL
;
1559 const struct sockaddr_in6
*sa6_src
= NULL
;
1561 struct tcp_portonly
{
1567 if (sa
->sa_family
!= AF_INET6
||
1568 sa
->sa_len
!= sizeof(struct sockaddr_in6
)) {
1573 if (cmd
== PRC_QUENCH
)
1574 notify
= tcp_quench
;
1575 else if (cmd
== PRC_MSGSIZE
) {
1577 * The MTU can be passed via an icmp6 packet or directly
1580 struct ip6ctlparam
*ip6cp
= d
;
1582 if (ip6cp
->ip6c_icmp6
) {
1583 struct icmp6_hdr
*icmp6
= ip6cp
->ip6c_icmp6
;
1584 arg
= ntohl(icmp6
->icmp6_mtu
);
1585 } else if (ip6cp
->ip6c_cmdarg
) {
1586 arg
= *(uint32_t *)ip6cp
->ip6c_cmdarg
;
1590 notify
= tcp_mtudisc
;
1591 } else if (!PRC_IS_REDIRECT(cmd
) &&
1592 ((unsigned)cmd
> PRC_NCMDS
|| inet6ctlerrmap
[cmd
] == 0)) {
1597 * If the parameter is from icmp6, decode it. Note that in the
1598 * mtu shortcut case, the rest of the ip6ctlparam content is
1602 ip6cp
= (struct ip6ctlparam
*)d
;
1604 ip6
= ip6cp
->ip6c_ip6
;
1605 off
= ip6cp
->ip6c_off
;
1606 sa6_src
= ip6cp
->ip6c_src
;
1610 off
= 0; /* fool gcc */
1615 struct in_conninfo inc
;
1617 * XXX: We assume that when IPV6 is non NULL,
1618 * M and OFF are valid.
1621 /* check if we can safely examine src and dst ports */
1622 if (m
->m_pkthdr
.len
< off
+ sizeof *thp
)
1625 bzero(&th
, sizeof th
);
1626 m_copydata(m
, off
, sizeof *thp
, &th
);
1628 in6_pcbnotify(&tcbinfo
[0], sa
, th
.th_dport
,
1629 (struct sockaddr
*)ip6cp
->ip6c_src
,
1630 th
.th_sport
, cmd
, arg
, notify
);
1632 inc
.inc_fport
= th
.th_dport
;
1633 inc
.inc_lport
= th
.th_sport
;
1634 inc
.inc6_faddr
= ((struct sockaddr_in6
*)sa
)->sin6_addr
;
1635 inc
.inc6_laddr
= ip6cp
->ip6c_src
->sin6_addr
;
1637 syncache_unreach(&inc
, &th
);
1639 in6_pcbnotify(&tcbinfo
[0], sa
, 0,
1640 (const struct sockaddr
*)sa6_src
, 0, cmd
, arg
, notify
);
1643 lwkt_replymsg(&msg
->ctlinput
.base
.lmsg
, 0);
1649 * Following is where TCP initial sequence number generation occurs.
1651 * There are two places where we must use initial sequence numbers:
1652 * 1. In SYN-ACK packets.
1653 * 2. In SYN packets.
1655 * All ISNs for SYN-ACK packets are generated by the syncache. See
1656 * tcp_syncache.c for details.
1658 * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling
1659 * depends on this property. In addition, these ISNs should be
1660 * unguessable so as to prevent connection hijacking. To satisfy
1661 * the requirements of this situation, the algorithm outlined in
1662 * RFC 1948 is used to generate sequence numbers.
1664 * Implementation details:
1666 * net.inet.tcp.isn_reseed_interval controls the number of seconds
1667 * between the seeding of isn_secret. On every reseed we jump the
1677 struct tcp_isn tcp_isn_ary
[MAXCPU
];
1680 tcp_new_isn(struct tcpcb
*tp
)
1682 struct tcp_isn
*isn
;
1684 tcp_seq digest
[16 / sizeof(tcp_seq
)];
1687 isn
= &tcp_isn_ary
[mycpuid
];
1690 * Reseed every 20 seconds. 6 reseeds per 2-minute interval in
1691 * order to retain our monotonic offset.
1693 * The initial seed randomizes last_offset with all 32 bits.
1695 * Note that the md5 digest is masked with 0x0FFFFFFF, so we must
1696 * add 1/16 of our full range (1/8 of our signed range) to ensure
1697 * monotonic operation.
1699 if (isn
->last_reseed
== 0 ||
1700 (u_int
)(ticks
- isn
->last_reseed
) > tcp_isn_reseed_interval
* hz
) {
1701 if (isn
->last_reseed
== 0) {
1702 read_random(&isn
->last_offset
,
1703 sizeof(isn
->last_offset
), 1);
1705 read_random(&isn
->secret
, sizeof(isn
->secret
), 1);
1706 isn
->last_reseed
= ticks
;
1707 isn
->last_offset
+= 0x10000000;
1711 * Compute the md5 hash, giving us a deterministic result for the
1712 * port/address pair for any given secret.
1715 MD5Update(&isn
->ctx
, isn
->secret
, sizeof(isn
->secret
));
1716 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->inp_fport
, 2);
1717 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->inp_lport
, 2);
1719 if (INP_ISIPV6(tp
->t_inpcb
)) {
1720 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->in6p_faddr
,
1721 sizeof(struct in6_addr
));
1722 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->in6p_laddr
,
1723 sizeof(struct in6_addr
));
1727 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->inp_faddr
,
1728 sizeof(struct in_addr
));
1729 MD5Update(&isn
->ctx
, (u_char
*)&tp
->t_inpcb
->inp_laddr
,
1730 sizeof(struct in_addr
));
1732 MD5Final((char *)digest
, &isn
->ctx
);
1735 * Add a random component 0-1048575 plus advance by 1048576.
1737 * The sequence space is simply too small, in modern times we also
1738 * must depend on the receive-side being a bit smarter when recycling
1739 * ports in TIME_WAIT.
1741 read_random(&n
, sizeof(n
), 1);
1742 isn
->last_offset
+= (n
& 0x000FFFFF) + 0x00100000;
1743 new_isn
= (digest
[0] & 0x0FFFFFFF) + isn
->last_offset
;
1749 * When a source quench is received, close congestion window
1750 * to one segment. We will gradually open it again as we proceed.
1753 tcp_quench(struct inpcb
*inp
, int error
)
1755 struct tcpcb
*tp
= intotcpcb(inp
);
1757 KASSERT(tp
!= NULL
, ("tcp_quench: tp is NULL"));
1758 tp
->snd_cwnd
= tp
->t_maxseg
;
1763 * When a specific ICMP unreachable message is received and the
1764 * connection state is SYN-SENT, drop the connection. This behavior
1765 * is controlled by the icmp_may_rst sysctl.
1768 tcp_drop_syn_sent(struct inpcb
*inp
, int error
)
1770 struct tcpcb
*tp
= intotcpcb(inp
);
1772 KASSERT(tp
!= NULL
, ("tcp_drop_syn_sent: tp is NULL"));
1773 if (tp
->t_state
== TCPS_SYN_SENT
)
1774 tcp_drop(tp
, error
);
1778 * When a `need fragmentation' ICMP is received, update our idea of the MSS
1779 * based on the new value in the route. Also nudge TCP to send something,
1780 * since we know the packet we just sent was dropped.
1781 * This duplicates some code in the tcp_mss() function in tcp_input.c.
1784 tcp_mtudisc(struct inpcb
*inp
, int mtu
)
1786 struct tcpcb
*tp
= intotcpcb(inp
);
1788 struct socket
*so
= inp
->inp_socket
;
1791 boolean_t isipv6
= INP_ISIPV6(inp
);
1793 const boolean_t isipv6
= FALSE
;
1796 KASSERT(tp
!= NULL
, ("tcp_mtudisc: tp is NULL"));
1799 * If no MTU is provided in the ICMP message, use the
1800 * next lower likely value, as specified in RFC 1191.
1805 oldmtu
= tp
->t_maxopd
+
1807 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1808 sizeof(struct tcpiphdr
));
1809 mtu
= ip_next_mtu(oldmtu
, 0);
1813 rt
= tcp_rtlookup6(&inp
->inp_inc
);
1815 rt
= tcp_rtlookup(&inp
->inp_inc
);
1817 if (rt
->rt_rmx
.rmx_mtu
!= 0 && rt
->rt_rmx
.rmx_mtu
< mtu
)
1818 mtu
= rt
->rt_rmx
.rmx_mtu
;
1822 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1823 sizeof(struct tcpiphdr
));
1826 * XXX - The following conditional probably violates the TCP
1827 * spec. The problem is that, since we don't know the
1828 * other end's MSS, we are supposed to use a conservative
1829 * default. But, if we do that, then MTU discovery will
1830 * never actually take place, because the conservative
1831 * default is much less than the MTUs typically seen
1832 * on the Internet today. For the moment, we'll sweep
1833 * this under the carpet.
1835 * The conservative default might not actually be a problem
1836 * if the only case this occurs is when sending an initial
1837 * SYN with options and data to a host we've never talked
1838 * to before. Then, they will reply with an MSS value which
1839 * will get recorded and the new parameters should get
1840 * recomputed. For Further Study.
1842 if (rt
->rt_rmx
.rmx_mssopt
&& rt
->rt_rmx
.rmx_mssopt
< maxopd
)
1843 maxopd
= rt
->rt_rmx
.rmx_mssopt
;
1847 sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) :
1848 sizeof(struct tcpiphdr
));
1850 if (tp
->t_maxopd
<= maxopd
)
1852 tp
->t_maxopd
= maxopd
;
1855 if ((tp
->t_flags
& (TF_REQ_TSTMP
| TF_RCVD_TSTMP
| TF_NOOPT
)) ==
1856 (TF_REQ_TSTMP
| TF_RCVD_TSTMP
))
1857 mss
-= TCPOLEN_TSTAMP_APPA
;
1859 /* round down to multiple of MCLBYTES */
1860 #if (MCLBYTES & (MCLBYTES - 1)) == 0 /* test if MCLBYTES power of 2 */
1862 mss
&= ~(MCLBYTES
- 1);
1865 mss
= rounddown(mss
, MCLBYTES
);
1868 if (so
->so_snd
.ssb_hiwat
< mss
)
1869 mss
= so
->so_snd
.ssb_hiwat
;
1873 tp
->snd_nxt
= tp
->snd_una
;
1875 tcpstat
.tcps_mturesent
++;
1879 * Look-up the routing entry to the peer of this inpcb. If no route
1880 * is found and it cannot be allocated the return NULL. This routine
1881 * is called by TCP routines that access the rmx structure and by tcp_mss
1882 * to get the interface MTU.
1885 tcp_rtlookup(struct in_conninfo
*inc
)
1887 struct route
*ro
= &inc
->inc_route
;
1889 if (ro
->ro_rt
== NULL
|| !(ro
->ro_rt
->rt_flags
& RTF_UP
)) {
1890 /* No route yet, so try to acquire one */
1891 if (inc
->inc_faddr
.s_addr
!= INADDR_ANY
) {
1893 * unused portions of the structure MUST be zero'd
1894 * out because rtalloc() treats it as opaque data
1896 bzero(&ro
->ro_dst
, sizeof(struct sockaddr_in
));
1897 ro
->ro_dst
.sa_family
= AF_INET
;
1898 ro
->ro_dst
.sa_len
= sizeof(struct sockaddr_in
);
1899 ((struct sockaddr_in
*) &ro
->ro_dst
)->sin_addr
=
1909 tcp_rtlookup6(struct in_conninfo
*inc
)
1911 struct route_in6
*ro6
= &inc
->inc6_route
;
1913 if (ro6
->ro_rt
== NULL
|| !(ro6
->ro_rt
->rt_flags
& RTF_UP
)) {
1914 /* No route yet, so try to acquire one */
1915 if (!IN6_IS_ADDR_UNSPECIFIED(&inc
->inc6_faddr
)) {
1917 * unused portions of the structure MUST be zero'd
1918 * out because rtalloc() treats it as opaque data
1920 bzero(&ro6
->ro_dst
, sizeof(struct sockaddr_in6
));
1921 ro6
->ro_dst
.sin6_family
= AF_INET6
;
1922 ro6
->ro_dst
.sin6_len
= sizeof(struct sockaddr_in6
);
1923 ro6
->ro_dst
.sin6_addr
= inc
->inc6_faddr
;
1924 rtalloc((struct route
*)ro6
);
1927 return (ro6
->ro_rt
);
1932 * TCP BANDWIDTH DELAY PRODUCT WINDOW LIMITING
1934 * This code attempts to calculate the bandwidth-delay product as a
1935 * means of determining the optimal window size to maximize bandwidth,
1936 * minimize RTT, and avoid the over-allocation of buffers on interfaces and
1937 * routers. This code also does a fairly good job keeping RTTs in check
1938 * across slow links like modems. We implement an algorithm which is very
1939 * similar (but not meant to be) TCP/Vegas. The code operates on the
1940 * transmitter side of a TCP connection and so only effects the transmit
1941 * side of the connection.
1943 * BACKGROUND: TCP makes no provision for the management of buffer space
1944 * at the end points or at the intermediate routers and switches. A TCP
1945 * stream, whether using NewReno or not, will eventually buffer as
1946 * many packets as it is able and the only reason this typically works is
1947 * due to the fairly small default buffers made available for a connection
1948 * (typicaly 16K or 32K). As machines use larger windows and/or window
1949 * scaling it is now fairly easy for even a single TCP connection to blow-out
1950 * all available buffer space not only on the local interface, but on
1951 * intermediate routers and switches as well. NewReno makes a misguided
1952 * attempt to 'solve' this problem by waiting for an actual failure to occur,
1953 * then backing off, then steadily increasing the window again until another
1954 * failure occurs, ad-infinitum. This results in terrible oscillation that
1955 * is only made worse as network loads increase and the idea of intentionally
1956 * blowing out network buffers is, frankly, a terrible way to manage network
1959 * It is far better to limit the transmit window prior to the failure
1960 * condition being achieved. There are two general ways to do this: First
1961 * you can 'scan' through different transmit window sizes and locate the
1962 * point where the RTT stops increasing, indicating that you have filled the
1963 * pipe, then scan backwards until you note that RTT stops decreasing, then
1964 * repeat ad-infinitum. This method works in principle but has severe
1965 * implementation issues due to RTT variances, timer granularity, and
1966 * instability in the algorithm which can lead to many false positives and
1967 * create oscillations as well as interact badly with other TCP streams
1968 * implementing the same algorithm.
1970 * The second method is to limit the window to the bandwidth delay product
1971 * of the link. This is the method we implement. RTT variances and our
1972 * own manipulation of the congestion window, bwnd, can potentially
1973 * destabilize the algorithm. For this reason we have to stabilize the
1974 * elements used to calculate the window. We do this by using the minimum
1975 * observed RTT, the long term average of the observed bandwidth, and
1976 * by adding two segments worth of slop. It isn't perfect but it is able
1977 * to react to changing conditions and gives us a very stable basis on
1978 * which to extend the algorithm.
1981 tcp_xmit_bandwidth_limit(struct tcpcb
*tp
, tcp_seq ack_seq
)
1990 * If inflight_enable is disabled in the middle of a tcp connection,
1991 * make sure snd_bwnd is effectively disabled.
1993 if (!tcp_inflight_enable
) {
1994 tp
->snd_bwnd
= TCP_MAXWIN
<< TCP_MAX_WINSHIFT
;
1995 tp
->snd_bandwidth
= 0;
2000 * Validate the delta time. If a connection is new or has been idle
2001 * a long time we have to reset the bandwidth calculator.
2005 delta_ticks
= save_ticks
- tp
->t_bw_rtttime
;
2006 if (tp
->t_bw_rtttime
== 0 || delta_ticks
< 0 || delta_ticks
> hz
* 10) {
2007 tp
->t_bw_rtttime
= save_ticks
;
2008 tp
->t_bw_rtseq
= ack_seq
;
2009 if (tp
->snd_bandwidth
== 0)
2010 tp
->snd_bandwidth
= tcp_inflight_start
;
2015 * A delta of at least 1 tick is required. Waiting 2 ticks will
2016 * result in better (bw) accuracy. More than that and the ramp-up
2019 if (delta_ticks
== 0 || delta_ticks
== 1)
2023 * Sanity check, plus ignore pure window update acks.
2025 if ((int)(ack_seq
- tp
->t_bw_rtseq
) <= 0)
2029 * Figure out the bandwidth. Due to the tick granularity this
2030 * is a very rough number and it MUST be averaged over a fairly
2031 * long period of time. XXX we need to take into account a link
2032 * that is not using all available bandwidth, but for now our
2033 * slop will ramp us up if this case occurs and the bandwidth later
2036 ibw
= (int64_t)(ack_seq
- tp
->t_bw_rtseq
) * hz
/ delta_ticks
;
2037 tp
->t_bw_rtttime
= save_ticks
;
2038 tp
->t_bw_rtseq
= ack_seq
;
2039 bw
= ((int64_t)tp
->snd_bandwidth
* 15 + ibw
) >> 4;
2041 tp
->snd_bandwidth
= bw
;
2044 * Calculate the semi-static bandwidth delay product, plus two maximal
2045 * segments. The additional slop puts us squarely in the sweet
2046 * spot and also handles the bandwidth run-up case. Without the
2047 * slop we could be locking ourselves into a lower bandwidth.
2049 * At very high speeds the bw calculation can become overly sensitive
2050 * and error prone when delta_ticks is low (e.g. usually 1). To deal
2051 * with the problem the stab must be scaled to the bw. A stab of 50
2052 * (the default) increases the bw for the purposes of the bwnd
2053 * calculation by 5%.
2055 * Situations Handled:
2056 * (1) Prevents over-queueing of packets on LANs, especially on
2057 * high speed LANs, allowing larger TCP buffers to be
2058 * specified, and also does a good job preventing
2059 * over-queueing of packets over choke points like modems
2060 * (at least for the transmit side).
2062 * (2) Is able to handle changing network loads (bandwidth
2063 * drops so bwnd drops, bandwidth increases so bwnd
2066 * (3) Theoretically should stabilize in the face of multiple
2067 * connections implementing the same algorithm (this may need
2070 * (4) Stability value (defaults to 20 = 2 maximal packets) can
2071 * be adjusted with a sysctl but typically only needs to be on
2072 * very slow connections. A value no smaller then 5 should
2073 * be used, but only reduce this default if you have no other
2077 #define USERTT ((tp->t_srtt + tp->t_rttvar) + tcp_inflight_adjrtt)
2078 bw
+= bw
* tcp_inflight_stab
/ 1000;
2079 bwnd
= (int64_t)bw
* USERTT
/ (hz
<< TCP_RTT_SHIFT
) +
2080 (int)tp
->t_maxseg
* 2;
2083 if (tcp_inflight_debug
> 0) {
2085 if ((u_int
)(save_ticks
- ltime
) >= hz
/ tcp_inflight_debug
) {
2087 kprintf("%p ibw %ld bw %ld rttvar %d srtt %d "
2088 "bwnd %ld delta %d snd_win %ld\n",
2089 tp
, ibw
, bw
, tp
->t_rttvar
, tp
->t_srtt
,
2090 bwnd
, delta_ticks
, tp
->snd_wnd
);
2093 if ((long)bwnd
< tcp_inflight_min
)
2094 bwnd
= tcp_inflight_min
;
2095 if (bwnd
> tcp_inflight_max
)
2096 bwnd
= tcp_inflight_max
;
2097 if ((long)bwnd
< tp
->t_maxseg
* 2)
2098 bwnd
= tp
->t_maxseg
* 2;
2099 tp
->snd_bwnd
= bwnd
;
2103 tcp_rmx_iwsegs(struct tcpcb
*tp
, u_long
*maxsegs
, u_long
*capsegs
)
2106 struct inpcb
*inp
= tp
->t_inpcb
;
2108 boolean_t isipv6
= INP_ISIPV6(inp
);
2110 const boolean_t isipv6
= FALSE
;
2114 if (tcp_iw_maxsegs
< TCP_IW_MAXSEGS_DFLT
)
2115 tcp_iw_maxsegs
= TCP_IW_MAXSEGS_DFLT
;
2116 if (tcp_iw_capsegs
< TCP_IW_CAPSEGS_DFLT
)
2117 tcp_iw_capsegs
= TCP_IW_CAPSEGS_DFLT
;
2120 rt
= tcp_rtlookup6(&inp
->inp_inc
);
2122 rt
= tcp_rtlookup(&inp
->inp_inc
);
2124 rt
->rt_rmx
.rmx_iwmaxsegs
< TCP_IW_MAXSEGS_DFLT
||
2125 rt
->rt_rmx
.rmx_iwcapsegs
< TCP_IW_CAPSEGS_DFLT
) {
2126 *maxsegs
= tcp_iw_maxsegs
;
2127 *capsegs
= tcp_iw_capsegs
;
2130 *maxsegs
= rt
->rt_rmx
.rmx_iwmaxsegs
;
2131 *capsegs
= rt
->rt_rmx
.rmx_iwcapsegs
;
2135 tcp_initial_window(struct tcpcb
*tp
)
2137 if (tcp_do_rfc3390
) {
2140 * "If the SYN or SYN/ACK is lost, the initial window
2141 * used by a sender after a correctly transmitted SYN
2142 * MUST be one segment consisting of MSS bytes."
2144 * However, we do something a little bit more aggressive
2145 * then RFC3390 here:
2146 * - Only if time spent in the SYN or SYN|ACK retransmition
2147 * >= 3 seconds, the IW is reduced. We do this mainly
2148 * because when RFC3390 is published, the initial RTO is
2149 * still 3 seconds (the threshold we test here), while
2150 * after RFC6298, the initial RTO is 1 second. This
2151 * behaviour probably still falls within the spirit of
2153 * - When IW is reduced, 2*MSS is used instead of 1*MSS.
2154 * Mainly to avoid sender and receiver deadlock until
2155 * delayed ACK timer expires. And even RFC2581 does not
2156 * try to reduce IW upon SYN or SYN|ACK retransmition
2160 * http://tools.ietf.org/html/draft-ietf-tcpm-initcwnd-03
2162 if (tp
->t_rxtsyn
>= TCPTV_RTOBASE3
) {
2163 return (2 * tp
->t_maxseg
);
2165 u_long maxsegs
, capsegs
;
2167 tcp_rmx_iwsegs(tp
, &maxsegs
, &capsegs
);
2168 return min(maxsegs
* tp
->t_maxseg
,
2169 max(2 * tp
->t_maxseg
, capsegs
* 1460));
2173 * Even RFC2581 (back to 1999) allows 2*SMSS IW.
2175 * Mainly to avoid sender and receiver deadlock
2176 * until delayed ACK timer expires.
2178 return (2 * tp
->t_maxseg
);
2182 #ifdef TCP_SIGNATURE
2184 * Compute TCP-MD5 hash of a TCP segment. (RFC2385)
2186 * We do this over ip, tcphdr, segment data, and the key in the SADB.
2187 * When called from tcp_input(), we can be sure that th_sum has been
2188 * zeroed out and verified already.
2190 * Return 0 if successful, otherwise return -1.
2192 * XXX The key is retrieved from the system's PF_KEY SADB, by keying a
2193 * search with the destination IP address, and a 'magic SPI' to be
2194 * determined by the application. This is hardcoded elsewhere to 1179
2195 * right now. Another branch of this code exists which uses the SPD to
2196 * specify per-application flows but it is unstable.
2199 tcpsignature_compute(
2200 struct mbuf
*m
, /* mbuf chain */
2201 int len
, /* length of TCP data */
2202 int optlen
, /* length of TCP options */
2203 u_char
*buf
, /* storage for MD5 digest */
2204 u_int direction
) /* direction of flow */
2206 struct ippseudo ippseudo
;
2210 struct ipovly
*ipovly
;
2211 struct secasvar
*sav
;
2214 struct ip6_hdr
*ip6
;
2215 struct in6_addr in6
;
2221 KASSERT(m
!= NULL
, ("passed NULL mbuf. Game over."));
2222 KASSERT(buf
!= NULL
, ("passed NULL storage pointer for MD5 signature"));
2224 * Extract the destination from the IP header in the mbuf.
2226 ip
= mtod(m
, struct ip
*);
2228 ip6
= NULL
; /* Make the compiler happy. */
2231 * Look up an SADB entry which matches the address found in
2234 switch (IP_VHL_V(ip
->ip_vhl
)) {
2236 sav
= key_allocsa(AF_INET
, (caddr_t
)&ip
->ip_src
, (caddr_t
)&ip
->ip_dst
,
2237 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2240 case (IPV6_VERSION
>> 4):
2241 ip6
= mtod(m
, struct ip6_hdr
*);
2242 sav
= key_allocsa(AF_INET6
, (caddr_t
)&ip6
->ip6_src
, (caddr_t
)&ip6
->ip6_dst
,
2243 IPPROTO_TCP
, htonl(TCP_SIG_SPI
));
2252 kprintf("%s: SADB lookup failed\n", __func__
);
2258 * Step 1: Update MD5 hash with IP pseudo-header.
2260 * XXX The ippseudo header MUST be digested in network byte order,
2261 * or else we'll fail the regression test. Assume all fields we've
2262 * been doing arithmetic on have been in host byte order.
2263 * XXX One cannot depend on ipovly->ih_len here. When called from
2264 * tcp_output(), the underlying ip_len member has not yet been set.
2266 switch (IP_VHL_V(ip
->ip_vhl
)) {
2268 ipovly
= (struct ipovly
*)ip
;
2269 ippseudo
.ippseudo_src
= ipovly
->ih_src
;
2270 ippseudo
.ippseudo_dst
= ipovly
->ih_dst
;
2271 ippseudo
.ippseudo_pad
= 0;
2272 ippseudo
.ippseudo_p
= IPPROTO_TCP
;
2273 ippseudo
.ippseudo_len
= htons(len
+ sizeof(struct tcphdr
) + optlen
);
2274 MD5Update(&ctx
, (char *)&ippseudo
, sizeof(struct ippseudo
));
2275 th
= (struct tcphdr
*)((u_char
*)ip
+ sizeof(struct ip
));
2276 doff
= sizeof(struct ip
) + sizeof(struct tcphdr
) + optlen
;
2280 * RFC 2385, 2.0 Proposal
2281 * For IPv6, the pseudo-header is as described in RFC 2460, namely the
2282 * 128-bit source IPv6 address, 128-bit destination IPv6 address, zero-
2283 * extended next header value (to form 32 bits), and 32-bit segment
2285 * Note: Upper-Layer Packet Length comes before Next Header.
2287 case (IPV6_VERSION
>> 4):
2289 in6_clearscope(&in6
);
2290 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2292 in6_clearscope(&in6
);
2293 MD5Update(&ctx
, (char *)&in6
, sizeof(struct in6_addr
));
2294 plen
= htonl(len
+ sizeof(struct tcphdr
) + optlen
);
2295 MD5Update(&ctx
, (char *)&plen
, sizeof(uint32_t));
2297 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2298 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2299 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2301 MD5Update(&ctx
, (char *)&nhdr
, sizeof(uint8_t));
2302 th
= (struct tcphdr
*)((u_char
*)ip6
+ sizeof(struct ip6_hdr
));
2303 doff
= sizeof(struct ip6_hdr
) + sizeof(struct tcphdr
) + optlen
;
2312 * Step 2: Update MD5 hash with TCP header, excluding options.
2313 * The TCP checksum must be set to zero.
2315 savecsum
= th
->th_sum
;
2317 MD5Update(&ctx
, (char *)th
, sizeof(struct tcphdr
));
2318 th
->th_sum
= savecsum
;
2320 * Step 3: Update MD5 hash with TCP segment data.
2321 * Use m_apply() to avoid an early m_pullup().
2324 m_apply(m
, doff
, len
, tcpsignature_apply
, &ctx
);
2326 * Step 4: Update MD5 hash with shared secret.
2328 MD5Update(&ctx
, _KEYBUF(sav
->key_auth
), _KEYLEN(sav
->key_auth
));
2329 MD5Final(buf
, &ctx
);
2330 key_sa_recordxfer(sav
, m
);
2336 tcpsignature_apply(void *fstate
, void *data
, unsigned int len
)
2339 MD5Update((MD5_CTX
*)fstate
, (unsigned char *)data
, len
);
2342 #endif /* TCP_SIGNATURE */
2345 tcp_drop_sysctl_dispatch(netmsg_t nmsg
)
2347 struct lwkt_msg
*lmsg
= &nmsg
->lmsg
;
2348 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2349 struct sockaddr_storage
*addrs
= lmsg
->u
.ms_resultp
;
2351 struct sockaddr_in
*fin
, *lin
;
2353 struct sockaddr_in6
*fin6
, *lin6
;
2354 struct in6_addr f6
, l6
;
2358 switch (addrs
[0].ss_family
) {
2361 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2362 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2363 error
= in6_embedscope(&f6
, fin6
, NULL
, NULL
);
2366 error
= in6_embedscope(&l6
, lin6
, NULL
, NULL
);
2369 inp
= in6_pcblookup_hash(&tcbinfo
[mycpuid
], &f6
,
2370 fin6
->sin6_port
, &l6
, lin6
->sin6_port
, FALSE
, NULL
);
2375 fin
= (struct sockaddr_in
*)&addrs
[0];
2376 lin
= (struct sockaddr_in
*)&addrs
[1];
2377 inp
= in_pcblookup_hash(&tcbinfo
[mycpuid
], fin
->sin_addr
,
2378 fin
->sin_port
, lin
->sin_addr
, lin
->sin_port
, FALSE
, NULL
);
2383 * Must not reach here, since the address family was
2384 * checked in sysctl handler.
2386 panic("unknown address family %d", addrs
[0].ss_family
);
2389 struct tcpcb
*tp
= intotcpcb(inp
);
2391 KASSERT((inp
->inp_flags
& INP_WILDCARD
) == 0,
2392 ("in wildcard hash"));
2393 KASSERT(tp
!= NULL
, ("tcp_drop_sysctl_dispatch: tp is NULL"));
2394 KASSERT((tp
->t_flags
& TF_LISTEN
) == 0, ("listen socket"));
2395 tcp_drop(tp
, ECONNABORTED
);
2403 lwkt_replymsg(lmsg
, error
);
2407 sysctl_tcp_drop(SYSCTL_HANDLER_ARGS
)
2409 /* addrs[0] is a foreign socket, addrs[1] is a local one. */
2410 struct sockaddr_storage addrs
[2];
2411 struct sockaddr_in
*fin
, *lin
;
2413 struct sockaddr_in6
*fin6
, *lin6
;
2415 struct netmsg_base nmsg
;
2416 struct lwkt_msg
*lmsg
= &nmsg
.lmsg
;
2417 struct lwkt_port
*port
= NULL
;
2426 if (req
->oldptr
!= NULL
|| req
->oldlen
!= 0)
2428 if (req
->newptr
== NULL
)
2430 if (req
->newlen
< sizeof(addrs
))
2432 error
= SYSCTL_IN(req
, &addrs
, sizeof(addrs
));
2436 switch (addrs
[0].ss_family
) {
2439 fin6
= (struct sockaddr_in6
*)&addrs
[0];
2440 lin6
= (struct sockaddr_in6
*)&addrs
[1];
2441 if (fin6
->sin6_len
!= sizeof(struct sockaddr_in6
) ||
2442 lin6
->sin6_len
!= sizeof(struct sockaddr_in6
))
2444 if (IN6_IS_ADDR_V4MAPPED(&fin6
->sin6_addr
) ||
2445 IN6_IS_ADDR_V4MAPPED(&lin6
->sin6_addr
))
2446 return (EADDRNOTAVAIL
);
2448 error
= sa6_embedscope(fin6
, V_ip6_use_defzone
);
2451 error
= sa6_embedscope(lin6
, V_ip6_use_defzone
);
2455 port
= tcp6_addrport();
2460 fin
= (struct sockaddr_in
*)&addrs
[0];
2461 lin
= (struct sockaddr_in
*)&addrs
[1];
2462 if (fin
->sin_len
!= sizeof(struct sockaddr_in
) ||
2463 lin
->sin_len
!= sizeof(struct sockaddr_in
))
2465 port
= tcp_addrport(fin
->sin_addr
.s_addr
, fin
->sin_port
,
2466 lin
->sin_addr
.s_addr
, lin
->sin_port
);
2473 netmsg_init(&nmsg
, NULL
, &curthread
->td_msgport
, 0,
2474 tcp_drop_sysctl_dispatch
);
2475 lmsg
->u
.ms_resultp
= addrs
;
2476 return lwkt_domsg(port
, lmsg
, 0);
2479 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, drop
,
2480 CTLTYPE_STRUCT
| CTLFLAG_WR
| CTLFLAG_SKIP
, NULL
,
2481 0, sysctl_tcp_drop
, "", "Drop TCP connection");
2484 sysctl_tcps_count(SYSCTL_HANDLER_ARGS
)
2486 u_long state_count
[TCP_NSTATES
];
2489 memset(state_count
, 0, sizeof(state_count
));
2490 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2493 for (i
= 0; i
< TCP_NSTATES
; ++i
)
2494 state_count
[i
] += tcpstate_count
[cpu
].tcps_count
[i
];
2497 return sysctl_handle_opaque(oidp
, state_count
, sizeof(state_count
), req
);
2499 SYSCTL_PROC(_net_inet_tcp
, OID_AUTO
, state_count
,
2500 CTLTYPE_OPAQUE
| CTLFLAG_RD
, NULL
, 0,
2501 sysctl_tcps_count
, "LU", "TCP connection counts by state");
2504 tcp_pcbport_create(struct tcpcb
*tp
)
2508 KASSERT((tp
->t_flags
& TF_LISTEN
) && tp
->t_state
== TCPS_LISTEN
,
2509 ("not a listen tcpcb"));
2511 KASSERT(tp
->t_pcbport
== NULL
, ("tcpcb port cache was created"));
2513 kmalloc(sizeof(struct tcp_pcbport
) * netisr_ncpus
,
2515 M_WAITOK
| M_CACHEALIGN
);
2517 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2518 struct inpcbport
*phd
;
2520 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2521 LIST_INIT(&phd
->phd_pcblist
);
2522 /* Though, not used ... */
2523 phd
->phd_port
= tp
->t_inpcb
->inp_lport
;
2528 tcp_pcbport_merge_oncpu(struct tcpcb
*tp
)
2530 struct inpcbport
*phd
;
2534 KASSERT(cpu
< netisr_ncpus
, ("invalid cpu%d", cpu
));
2535 phd
= &tp
->t_pcbport
[cpu
].t_phd
;
2537 while ((inp
= LIST_FIRST(&phd
->phd_pcblist
)) != NULL
) {
2538 KASSERT(inp
->inp_phd
== phd
&& inp
->inp_porthash
== NULL
,
2539 ("not on tcpcb port cache"));
2540 LIST_REMOVE(inp
, inp_portlist
);
2541 in_pcbinsporthash_lport(inp
);
2542 KASSERT(inp
->inp_phd
== tp
->t_inpcb
->inp_phd
&&
2543 inp
->inp_porthash
== tp
->t_inpcb
->inp_porthash
,
2544 ("tcpcb port cache merge failed"));
2549 tcp_pcbport_destroy(struct tcpcb
*tp
)
2554 for (cpu
= 0; cpu
< netisr_ncpus
; ++cpu
) {
2555 KASSERT(LIST_EMPTY(&tp
->t_pcbport
[cpu
].t_phd
.phd_pcblist
),
2556 ("tcpcb port cache is not empty"));
2559 kfree(tp
->t_pcbport
, M_PCB
);
2560 tp
->t_pcbport
= NULL
;