kernel - MPSAFE the protocol drain routines
[dragonfly.git] / sys / netinet / tcp_input.c
1 /*
2  * Copyright (c) 2002, 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2002, 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
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.
19  *
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
31  * SUCH DAMAGE.
32  */
33
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
36  *      The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
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. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *      This product includes software developed by the University of
49  *      California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_input.c,v 1.107.2.38 2003/05/21 04:46:41 cjc Exp $
68  * $DragonFly: src/sys/netinet/tcp_input.c,v 1.68 2008/08/22 09:14:17 sephe Exp $
69  */
70
71 #include "opt_ipfw.h"           /* for ipfw_fwd         */
72 #include "opt_inet.h"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_tcpdebug.h"
76 #include "opt_tcp_input.h"
77
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mbuf.h>
84 #include <sys/proc.h>           /* for proc0 declaration */
85 #include <sys/protosw.h>
86 #include <sys/socket.h>
87 #include <sys/socketvar.h>
88 #include <sys/syslog.h>
89 #include <sys/in_cksum.h>
90
91 #include <machine/cpu.h>        /* before tcp_seq.h, for tcp_random18() */
92 #include <machine/stdarg.h>
93
94 #include <net/if.h>
95 #include <net/route.h>
96
97 #include <netinet/in.h>
98 #include <netinet/in_systm.h>
99 #include <netinet/ip.h>
100 #include <netinet/ip_icmp.h>    /* for ICMP_BANDLIM */
101 #include <netinet/in_var.h>
102 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
103 #include <netinet/in_pcb.h>
104 #include <netinet/ip_var.h>
105 #include <netinet/ip6.h>
106 #include <netinet/icmp6.h>
107 #include <netinet6/nd6.h>
108 #include <netinet6/ip6_var.h>
109 #include <netinet6/in6_pcb.h>
110 #include <netinet/tcp.h>
111 #include <netinet/tcp_fsm.h>
112 #include <netinet/tcp_seq.h>
113 #include <netinet/tcp_timer.h>
114 #include <netinet/tcp_timer2.h>
115 #include <netinet/tcp_var.h>
116 #include <netinet6/tcp6_var.h>
117 #include <netinet/tcpip.h>
118
119 #ifdef TCPDEBUG
120 #include <netinet/tcp_debug.h>
121
122 u_char tcp_saveipgen[40];    /* the size must be of max ip header, now IPv6 */
123 struct tcphdr tcp_savetcp;
124 #endif
125
126 #ifdef FAST_IPSEC
127 #include <netproto/ipsec/ipsec.h>
128 #include <netproto/ipsec/ipsec6.h>
129 #endif
130
131 #ifdef IPSEC
132 #include <netinet6/ipsec.h>
133 #include <netinet6/ipsec6.h>
134 #include <netproto/key/key.h>
135 #endif
136
137 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
138
139 static int log_in_vain = 0;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW,
141     &log_in_vain, 0, "Log all incoming TCP connections");
142
143 static int blackhole = 0;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW,
145     &blackhole, 0, "Do not send RST when dropping refused connections");
146
147 int tcp_delack_enabled = 1;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW,
149     &tcp_delack_enabled, 0,
150     "Delay ACK to try and piggyback it onto a data packet");
151
152 #ifdef TCP_DROP_SYNFIN
153 static int drop_synfin = 0;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW,
155     &drop_synfin, 0, "Drop TCP packets with SYN+FIN set");
156 #endif
157
158 static int tcp_do_limitedtransmit = 1;
159 SYSCTL_INT(_net_inet_tcp, OID_AUTO, limitedtransmit, CTLFLAG_RW,
160     &tcp_do_limitedtransmit, 0, "Enable RFC 3042 (Limited Transmit)");
161
162 static int tcp_do_early_retransmit = 1;
163 SYSCTL_INT(_net_inet_tcp, OID_AUTO, earlyretransmit, CTLFLAG_RW,
164     &tcp_do_early_retransmit, 0, "Early retransmit");
165
166 int tcp_aggregate_acks = 1;
167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, aggregate_acks, CTLFLAG_RW,
168     &tcp_aggregate_acks, 0, "Aggregate built-up acks into one ack");
169
170 int tcp_do_rfc3390 = 1;
171 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
172     &tcp_do_rfc3390, 0,
173     "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
174
175 static int tcp_do_eifel_detect = 1;
176 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
177     &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
178
179 static int tcp_do_abc = 1;
180 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
181     &tcp_do_abc, 0,
182     "TCP Appropriate Byte Counting (RFC 3465)");
183
184 /*
185  * Define as tunable for easy testing with SACK on and off.
186  * Warning:  do not change setting in the middle of an existing active TCP flow,
187  *   else strange things might happen to that flow.
188  */
189 int tcp_do_sack = 1;
190 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
191     &tcp_do_sack, 0, "Enable SACK Algorithms");
192
193 int tcp_do_smartsack = 1;
194 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
195     &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
196
197 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
198     "TCP Segment Reassembly Queue");
199
200 int tcp_reass_maxseg = 0;
201 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
202     &tcp_reass_maxseg, 0,
203     "Global maximum number of TCP Segments in Reassembly Queue");
204
205 int tcp_reass_qsize = 0;
206 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
207     &tcp_reass_qsize, 0,
208     "Global number of TCP Segments currently in Reassembly Queue");
209
210 static int tcp_reass_overflows = 0;
211 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
212     &tcp_reass_overflows, 0,
213     "Global number of TCP Segment Reassembly Queue Overflows");
214
215 int tcp_do_autorcvbuf = 1;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
217     &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
218
219 int tcp_autorcvbuf_inc = 16*1024;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
221     &tcp_autorcvbuf_inc, 0,
222     "Incrementor step size of automatic receive buffer");
223
224 int tcp_autorcvbuf_max = 2*1024*1024;
225 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
226     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
227
228
229 static void      tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t);
230 static void      tcp_pulloutofband(struct socket *,
231                      struct tcphdr *, struct mbuf *, int);
232 static int       tcp_reass(struct tcpcb *, struct tcphdr *, int *,
233                      struct mbuf *);
234 static void      tcp_xmit_timer(struct tcpcb *, int);
235 static void      tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
236 static void      tcp_sack_rexmt(struct tcpcb *, struct tcphdr *);
237
238 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
239 #ifdef INET6
240 #define ND6_HINT(tp) \
241 do { \
242         if ((tp) && (tp)->t_inpcb && \
243             ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
244             (tp)->t_inpcb->in6p_route.ro_rt) \
245                 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
246 } while (0)
247 #else
248 #define ND6_HINT(tp)
249 #endif
250
251 /*
252  * Indicate whether this ack should be delayed.  We can delay the ack if
253  *      - delayed acks are enabled and
254  *      - there is no delayed ack timer in progress and
255  *      - our last ack wasn't a 0-sized window.  We never want to delay
256  *        the ack that opens up a 0-sized window.
257  */
258 #define DELAY_ACK(tp) \
259         (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
260         !(tp->t_flags & TF_RXWIN0SENT))
261
262 #define acceptable_window_update(tp, th, tiwin)                         \
263     (SEQ_LT(tp->snd_wl1, th->th_seq) ||                                 \
264      (tp->snd_wl1 == th->th_seq &&                                      \
265       (SEQ_LT(tp->snd_wl2, th->th_ack) ||                               \
266        (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
267
268 static int
269 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
270 {
271         struct tseg_qent *q;
272         struct tseg_qent *p = NULL;
273         struct tseg_qent *te;
274         struct socket *so = tp->t_inpcb->inp_socket;
275         int flags;
276
277         /*
278          * Call with th == NULL after become established to
279          * force pre-ESTABLISHED data up to user socket.
280          */
281         if (th == NULL)
282                 goto present;
283
284         /*
285          * Limit the number of segments in the reassembly queue to prevent
286          * holding on to too many segments (and thus running out of mbufs).
287          * Make sure to let the missing segment through which caused this
288          * queue.  Always keep one global queue entry spare to be able to
289          * process the missing segment.
290          */
291         if (th->th_seq != tp->rcv_nxt &&
292             tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
293                 tcp_reass_overflows++;
294                 tcpstat.tcps_rcvmemdrop++;
295                 m_freem(m);
296                 /* no SACK block to report */
297                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
298                 return (0);
299         }
300
301         /* Allocate a new queue entry. */
302         MALLOC(te, struct tseg_qent *, sizeof(struct tseg_qent), M_TSEGQ,
303                M_INTWAIT | M_NULLOK);
304         if (te == NULL) {
305                 tcpstat.tcps_rcvmemdrop++;
306                 m_freem(m);
307                 /* no SACK block to report */
308                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
309                 return (0);
310         }
311         atomic_add_int(&tcp_reass_qsize, 1);
312
313         /*
314          * Find a segment which begins after this one does.
315          */
316         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
317                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
318                         break;
319                 p = q;
320         }
321
322         /*
323          * If there is a preceding segment, it may provide some of
324          * our data already.  If so, drop the data from the incoming
325          * segment.  If it provides all of our data, drop us.
326          */
327         if (p != NULL) {
328                 tcp_seq_diff_t i;
329
330                 /* conversion to int (in i) handles seq wraparound */
331                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
332                 if (i > 0) {            /* overlaps preceding segment */
333                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
334                         /* enclosing block starts w/ preceding segment */
335                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
336                         if (i >= *tlenp) {
337                                 /* preceding encloses incoming segment */
338                                 tp->encloseblk.rblk_end = p->tqe_th->th_seq +
339                                     p->tqe_len;
340                                 tcpstat.tcps_rcvduppack++;
341                                 tcpstat.tcps_rcvdupbyte += *tlenp;
342                                 m_freem(m);
343                                 kfree(te, M_TSEGQ);
344                                 atomic_add_int(&tcp_reass_qsize, -1);
345                                 /*
346                                  * Try to present any queued data
347                                  * at the left window edge to the user.
348                                  * This is needed after the 3-WHS
349                                  * completes.
350                                  */
351                                 goto present;   /* ??? */
352                         }
353                         m_adj(m, i);
354                         *tlenp -= i;
355                         th->th_seq += i;
356                         /* incoming segment end is enclosing block end */
357                         tp->encloseblk.rblk_end = th->th_seq + *tlenp +
358                             ((th->th_flags & TH_FIN) != 0);
359                         /* trim end of reported D-SACK block */
360                         tp->reportblk.rblk_end = th->th_seq;
361                 }
362         }
363         tcpstat.tcps_rcvoopack++;
364         tcpstat.tcps_rcvoobyte += *tlenp;
365
366         /*
367          * While we overlap succeeding segments trim them or,
368          * if they are completely covered, dequeue them.
369          */
370         while (q) {
371                 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
372                 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
373                 struct tseg_qent *nq;
374
375                 if (i <= 0)
376                         break;
377                 if (!(tp->t_flags & TF_DUPSEG)) {    /* first time through */
378                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
379                         tp->encloseblk = tp->reportblk;
380                         /* report trailing duplicate D-SACK segment */
381                         tp->reportblk.rblk_start = q->tqe_th->th_seq;
382                 }
383                 if ((tp->t_flags & TF_ENCLOSESEG) &&
384                     SEQ_GT(qend, tp->encloseblk.rblk_end)) {
385                         /* extend enclosing block if one exists */
386                         tp->encloseblk.rblk_end = qend;
387                 }
388                 if (i < q->tqe_len) {
389                         q->tqe_th->th_seq += i;
390                         q->tqe_len -= i;
391                         m_adj(q->tqe_m, i);
392                         break;
393                 }
394
395                 nq = LIST_NEXT(q, tqe_q);
396                 LIST_REMOVE(q, tqe_q);
397                 m_freem(q->tqe_m);
398                 kfree(q, M_TSEGQ);
399                 atomic_add_int(&tcp_reass_qsize, -1);
400                 q = nq;
401         }
402
403         /* Insert the new segment queue entry into place. */
404         te->tqe_m = m;
405         te->tqe_th = th;
406         te->tqe_len = *tlenp;
407
408         /* check if can coalesce with following segment */
409         if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
410                 tcp_seq tend = te->tqe_th->th_seq + te->tqe_len;
411
412                 te->tqe_len += q->tqe_len;
413                 if (q->tqe_th->th_flags & TH_FIN)
414                         te->tqe_th->th_flags |= TH_FIN;
415                 m_cat(te->tqe_m, q->tqe_m);
416                 tp->encloseblk.rblk_end = tend;
417                 /*
418                  * When not reporting a duplicate segment, use
419                  * the larger enclosing block as the SACK block.
420                  */
421                 if (!(tp->t_flags & TF_DUPSEG))
422                         tp->reportblk.rblk_end = tend;
423                 LIST_REMOVE(q, tqe_q);
424                 kfree(q, M_TSEGQ);
425                 atomic_add_int(&tcp_reass_qsize, -1);
426         }
427
428         if (p == NULL) {
429                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
430         } else {
431                 /* check if can coalesce with preceding segment */
432                 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
433                         p->tqe_len += te->tqe_len;
434                         m_cat(p->tqe_m, te->tqe_m);
435                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
436                         /*
437                          * When not reporting a duplicate segment, use
438                          * the larger enclosing block as the SACK block.
439                          */
440                         if (!(tp->t_flags & TF_DUPSEG))
441                                 tp->reportblk.rblk_start = p->tqe_th->th_seq;
442                         kfree(te, M_TSEGQ);
443                         atomic_add_int(&tcp_reass_qsize, -1);
444                 } else {
445                         LIST_INSERT_AFTER(p, te, tqe_q);
446                 }
447         }
448
449 present:
450         /*
451          * Present data to user, advancing rcv_nxt through
452          * completed sequence space.
453          */
454         if (!TCPS_HAVEESTABLISHED(tp->t_state))
455                 return (0);
456         q = LIST_FIRST(&tp->t_segq);
457         if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
458                 return (0);
459         tp->rcv_nxt += q->tqe_len;
460         if (!(tp->t_flags & TF_DUPSEG)) {
461                 /* no SACK block to report since ACK advanced */
462                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
463         }
464         /* no enclosing block to report since ACK advanced */
465         tp->t_flags &= ~TF_ENCLOSESEG;
466         flags = q->tqe_th->th_flags & TH_FIN;
467         LIST_REMOVE(q, tqe_q);
468         KASSERT(LIST_EMPTY(&tp->t_segq) ||
469                 LIST_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
470                 ("segment not coalesced"));
471         if (so->so_state & SS_CANTRCVMORE)
472                 m_freem(q->tqe_m);
473         else
474                 ssb_appendstream(&so->so_rcv, q->tqe_m);
475         kfree(q, M_TSEGQ);
476         atomic_add_int(&tcp_reass_qsize, -1);
477         ND6_HINT(tp);
478         sorwakeup(so);
479         return (flags);
480 }
481
482 /*
483  * TCP input routine, follows pages 65-76 of the
484  * protocol specification dated September, 1981 very closely.
485  */
486 #ifdef INET6
487 int
488 tcp6_input(struct mbuf **mp, int *offp, int proto)
489 {
490         struct mbuf *m = *mp;
491         struct in6_ifaddr *ia6;
492
493         IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
494
495         /*
496          * draft-itojun-ipv6-tcp-to-anycast
497          * better place to put this in?
498          */
499         ia6 = ip6_getdstifaddr(m);
500         if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
501                 struct ip6_hdr *ip6;
502
503                 ip6 = mtod(m, struct ip6_hdr *);
504                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
505                             offsetof(struct ip6_hdr, ip6_dst));
506                 return (IPPROTO_DONE);
507         }
508
509         tcp_input(m, *offp, proto);
510         return (IPPROTO_DONE);
511 }
512 #endif
513
514 void
515 tcp_input(struct mbuf *m, ...)
516 {
517         __va_list ap;
518         int off0, proto;
519         struct tcphdr *th;
520         struct ip *ip = NULL;
521         struct ipovly *ipov;
522         struct inpcb *inp = NULL;
523         u_char *optp = NULL;
524         int optlen = 0;
525         int tlen, off;
526         int len = 0;
527         int drop_hdrlen;
528         struct tcpcb *tp = NULL;
529         int thflags;
530         struct socket *so = 0;
531         int todrop, acked;
532         boolean_t ourfinisacked, needoutput = FALSE;
533         u_long tiwin;
534         int recvwin;
535         struct tcpopt to;               /* options in this segment */
536         struct sockaddr_in *next_hop = NULL;
537         int rstreason; /* For badport_bandlim accounting purposes */
538         int cpu;
539         struct ip6_hdr *ip6 = NULL;
540 #ifdef INET6
541         boolean_t isipv6;
542 #else
543         const boolean_t isipv6 = FALSE;
544 #endif
545 #ifdef TCPDEBUG
546         short ostate = 0;
547 #endif
548
549         __va_start(ap, m);
550         off0 = __va_arg(ap, int);
551         proto = __va_arg(ap, int);
552         __va_end(ap);
553
554         tcpstat.tcps_rcvtotal++;
555
556         if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
557                 struct m_tag *mtag;
558
559                 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
560                 KKASSERT(mtag != NULL);
561                 next_hop = m_tag_data(mtag);
562         }
563
564 #ifdef INET6
565         isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
566 #endif
567
568         if (isipv6) {
569                 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
570                 ip6 = mtod(m, struct ip6_hdr *);
571                 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
572                 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
573                         tcpstat.tcps_rcvbadsum++;
574                         goto drop;
575                 }
576                 th = (struct tcphdr *)((caddr_t)ip6 + off0);
577
578                 /*
579                  * Be proactive about unspecified IPv6 address in source.
580                  * As we use all-zero to indicate unbounded/unconnected pcb,
581                  * unspecified IPv6 address can be used to confuse us.
582                  *
583                  * Note that packets with unspecified IPv6 destination is
584                  * already dropped in ip6_input.
585                  */
586                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
587                         /* XXX stat */
588                         goto drop;
589                 }
590         } else {
591                 /*
592                  * Get IP and TCP header together in first mbuf.
593                  * Note: IP leaves IP header in first mbuf.
594                  */
595                 if (off0 > sizeof(struct ip)) {
596                         ip_stripoptions(m);
597                         off0 = sizeof(struct ip);
598                 }
599                 /* already checked and pulled up in ip_demux() */
600                 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
601                     ("TCP header not in one mbuf: m->m_len %d", m->m_len));
602                 ip = mtod(m, struct ip *);
603                 ipov = (struct ipovly *)ip;
604                 th = (struct tcphdr *)((caddr_t)ip + off0);
605                 tlen = ip->ip_len;
606
607                 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
608                         if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
609                                 th->th_sum = m->m_pkthdr.csum_data;
610                         else
611                                 th->th_sum = in_pseudo(ip->ip_src.s_addr,
612                                                 ip->ip_dst.s_addr,
613                                                 htonl(m->m_pkthdr.csum_data +
614                                                         ip->ip_len +
615                                                         IPPROTO_TCP));
616                         th->th_sum ^= 0xffff;
617                 } else {
618                         /*
619                          * Checksum extended TCP header and data.
620                          */
621                         len = sizeof(struct ip) + tlen;
622                         bzero(ipov->ih_x1, sizeof ipov->ih_x1);
623                         ipov->ih_len = (u_short)tlen;
624                         ipov->ih_len = htons(ipov->ih_len);
625                         th->th_sum = in_cksum(m, len);
626                 }
627                 if (th->th_sum) {
628                         tcpstat.tcps_rcvbadsum++;
629                         goto drop;
630                 }
631 #ifdef INET6
632                 /* Re-initialization for later version check */
633                 ip->ip_v = IPVERSION;
634 #endif
635         }
636
637         /*
638          * Check that TCP offset makes sense,
639          * pull out TCP options and adjust length.              XXX
640          */
641         off = th->th_off << 2;
642         /* already checked and pulled up in ip_demux() */
643         KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
644             ("bad TCP data offset %d (tlen %d)", off, tlen));
645         tlen -= off;    /* tlen is used instead of ti->ti_len */
646         if (off > sizeof(struct tcphdr)) {
647                 if (isipv6) {
648                         IP6_EXTHDR_CHECK(m, off0, off, );
649                         ip6 = mtod(m, struct ip6_hdr *);
650                         th = (struct tcphdr *)((caddr_t)ip6 + off0);
651                 } else {
652                         /* already pulled up in ip_demux() */
653                         KASSERT(m->m_len >= sizeof(struct ip) + off,
654                             ("TCP header and options not in one mbuf: "
655                              "m_len %d, off %d", m->m_len, off));
656                 }
657                 optlen = off - sizeof(struct tcphdr);
658                 optp = (u_char *)(th + 1);
659         }
660         thflags = th->th_flags;
661
662 #ifdef TCP_DROP_SYNFIN
663         /*
664          * If the drop_synfin option is enabled, drop all packets with
665          * both the SYN and FIN bits set. This prevents e.g. nmap from
666          * identifying the TCP/IP stack.
667          *
668          * This is a violation of the TCP specification.
669          */
670         if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
671                 goto drop;
672 #endif
673
674         /*
675          * Convert TCP protocol specific fields to host format.
676          */
677         th->th_seq = ntohl(th->th_seq);
678         th->th_ack = ntohl(th->th_ack);
679         th->th_win = ntohs(th->th_win);
680         th->th_urp = ntohs(th->th_urp);
681
682         /*
683          * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
684          * until after ip6_savecontrol() is called and before other functions
685          * which don't want those proto headers.
686          * Because ip6_savecontrol() is going to parse the mbuf to
687          * search for data to be passed up to user-land, it wants mbuf
688          * parameters to be unchanged.
689          * XXX: the call of ip6_savecontrol() has been obsoleted based on
690          * latest version of the advanced API (20020110).
691          */
692         drop_hdrlen = off0 + off;
693
694         /*
695          * Locate pcb for segment.
696          */
697 findpcb:
698         /* IPFIREWALL_FORWARD section */
699         if (next_hop != NULL && !isipv6) {  /* IPv6 support is not there yet */
700                 /*
701                  * Transparently forwarded. Pretend to be the destination.
702                  * already got one like this?
703                  */
704                 cpu = mycpu->gd_cpuid;
705                 inp = in_pcblookup_hash(&tcbinfo[cpu],
706                                         ip->ip_src, th->th_sport,
707                                         ip->ip_dst, th->th_dport,
708                                         0, m->m_pkthdr.rcvif);
709                 if (!inp) {
710                         /*
711                          * It's new.  Try to find the ambushing socket.
712                          */
713
714                         /*
715                          * The rest of the ipfw code stores the port in
716                          * host order.  XXX
717                          * (The IP address is still in network order.)
718                          */
719                         in_port_t dport = next_hop->sin_port ?
720                                                 htons(next_hop->sin_port) :
721                                                 th->th_dport;
722
723                         cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
724                                           next_hop->sin_addr.s_addr, dport);
725                         inp = in_pcblookup_hash(&tcbinfo[cpu],
726                                                 ip->ip_src, th->th_sport,
727                                                 next_hop->sin_addr, dport,
728                                                 1, m->m_pkthdr.rcvif);
729                 }
730         } else {
731                 if (isipv6) {
732                         inp = in6_pcblookup_hash(&tcbinfo[0],
733                                                  &ip6->ip6_src, th->th_sport,
734                                                  &ip6->ip6_dst, th->th_dport,
735                                                  1, m->m_pkthdr.rcvif);
736                 } else {
737                         cpu = mycpu->gd_cpuid;
738                         inp = in_pcblookup_hash(&tcbinfo[cpu],
739                                                 ip->ip_src, th->th_sport,
740                                                 ip->ip_dst, th->th_dport,
741                                                 1, m->m_pkthdr.rcvif);
742                 }
743         }
744
745         /*
746          * If the state is CLOSED (i.e., TCB does not exist) then
747          * all data in the incoming segment is discarded.
748          * If the TCB exists but is in CLOSED state, it is embryonic,
749          * but should either do a listen or a connect soon.
750          */
751         if (inp == NULL) {
752                 if (log_in_vain) {
753 #ifdef INET6
754                         char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
755 #else
756                         char dbuf[sizeof "aaa.bbb.ccc.ddd"];
757                         char sbuf[sizeof "aaa.bbb.ccc.ddd"];
758 #endif
759                         if (isipv6) {
760                                 strcpy(dbuf, "[");
761                                 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
762                                 strcat(dbuf, "]");
763                                 strcpy(sbuf, "[");
764                                 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
765                                 strcat(sbuf, "]");
766                         } else {
767                                 strcpy(dbuf, inet_ntoa(ip->ip_dst));
768                                 strcpy(sbuf, inet_ntoa(ip->ip_src));
769                         }
770                         switch (log_in_vain) {
771                         case 1:
772                                 if (!(thflags & TH_SYN))
773                                         break;
774                         case 2:
775                                 log(LOG_INFO,
776                                     "Connection attempt to TCP %s:%d "
777                                     "from %s:%d flags:0x%02x\n",
778                                     dbuf, ntohs(th->th_dport), sbuf,
779                                     ntohs(th->th_sport), thflags);
780                                 break;
781                         default:
782                                 break;
783                         }
784                 }
785                 if (blackhole) {
786                         switch (blackhole) {
787                         case 1:
788                                 if (thflags & TH_SYN)
789                                         goto drop;
790                                 break;
791                         case 2:
792                                 goto drop;
793                         default:
794                                 goto drop;
795                         }
796                 }
797                 rstreason = BANDLIM_RST_CLOSEDPORT;
798                 goto dropwithreset;
799         }
800
801 #ifdef IPSEC
802         if (isipv6) {
803                 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
804                         ipsec6stat.in_polvio++;
805                         goto drop;
806                 }
807         } else {
808                 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
809                         ipsecstat.in_polvio++;
810                         goto drop;
811                 }
812         }
813 #endif
814 #ifdef FAST_IPSEC
815         if (isipv6) {
816                 if (ipsec6_in_reject(m, inp))
817                         goto drop;
818         } else {
819                 if (ipsec4_in_reject(m, inp))
820                         goto drop;
821         }
822 #endif
823         /* Check the minimum TTL for socket. */
824 #ifdef INET6
825         if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
826                 goto drop;
827 #endif
828
829         tp = intotcpcb(inp);
830         if (tp == NULL) {
831                 rstreason = BANDLIM_RST_CLOSEDPORT;
832                 goto dropwithreset;
833         }
834         if (tp->t_state <= TCPS_CLOSED)
835                 goto drop;
836
837         /* Unscale the window into a 32-bit value. */
838         if (!(thflags & TH_SYN))
839                 tiwin = th->th_win << tp->snd_scale;
840         else
841                 tiwin = th->th_win;
842
843         so = inp->inp_socket;
844
845 #ifdef TCPDEBUG
846         if (so->so_options & SO_DEBUG) {
847                 ostate = tp->t_state;
848                 if (isipv6)
849                         bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
850                 else
851                         bcopy(ip, tcp_saveipgen, sizeof(*ip));
852                 tcp_savetcp = *th;
853         }
854 #endif
855
856         bzero(&to, sizeof to);
857
858         if (so->so_options & SO_ACCEPTCONN) {
859                 struct in_conninfo inc;
860
861 #ifdef INET6
862                 inc.inc_isipv6 = (isipv6 == TRUE);
863 #endif
864                 if (isipv6) {
865                         inc.inc6_faddr = ip6->ip6_src;
866                         inc.inc6_laddr = ip6->ip6_dst;
867                         inc.inc6_route.ro_rt = NULL;            /* XXX */
868                 } else {
869                         inc.inc_faddr = ip->ip_src;
870                         inc.inc_laddr = ip->ip_dst;
871                         inc.inc_route.ro_rt = NULL;             /* XXX */
872                 }
873                 inc.inc_fport = th->th_sport;
874                 inc.inc_lport = th->th_dport;
875
876                 /*
877                  * If the state is LISTEN then ignore segment if it contains
878                  * a RST.  If the segment contains an ACK then it is bad and
879                  * send a RST.  If it does not contain a SYN then it is not
880                  * interesting; drop it.
881                  *
882                  * If the state is SYN_RECEIVED (syncache) and seg contains
883                  * an ACK, but not for our SYN/ACK, send a RST.  If the seg
884                  * contains a RST, check the sequence number to see if it
885                  * is a valid reset segment.
886                  */
887                 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
888                         if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
889                                 if (!syncache_expand(&inc, th, &so, m)) {
890                                         /*
891                                          * No syncache entry, or ACK was not
892                                          * for our SYN/ACK.  Send a RST.
893                                          */
894                                         tcpstat.tcps_badsyn++;
895                                         rstreason = BANDLIM_RST_OPENPORT;
896                                         goto dropwithreset;
897                                 }
898
899                                 /*
900                                  * Could not complete 3-way handshake,
901                                  * connection is being closed down, and
902                                  * syncache will free mbuf.
903                                  */
904                                 if (so == NULL)
905                                         return;
906
907                                 /*
908                                  * We must be in the correct protocol thread
909                                  * for this connection.
910                                  */
911                                 KKASSERT(so->so_port == &curthread->td_msgport);
912
913                                 /*
914                                  * Socket is created in state SYN_RECEIVED.
915                                  * Continue processing segment.
916                                  */
917                                 inp = so->so_pcb;
918                                 tp = intotcpcb(inp);
919                                 /*
920                                  * This is what would have happened in
921                                  * tcp_output() when the SYN,ACK was sent.
922                                  */
923                                 tp->snd_up = tp->snd_una;
924                                 tp->snd_max = tp->snd_nxt = tp->iss + 1;
925                                 tp->last_ack_sent = tp->rcv_nxt;
926 /*
927  * XXX possible bug - it doesn't appear that tp->snd_wnd is unscaled
928  * until the _second_ ACK is received:
929  *    rcv SYN (set wscale opts)  --> send SYN/ACK, set snd_wnd = window.
930  *    rcv ACK, calculate tiwin --> process SYN_RECEIVED, determine wscale,
931  *        move to ESTAB, set snd_wnd to tiwin.
932  */
933                                 tp->snd_wnd = tiwin;    /* unscaled */
934                                 goto after_listen;
935                         }
936                         if (thflags & TH_RST) {
937                                 syncache_chkrst(&inc, th);
938                                 goto drop;
939                         }
940                         if (thflags & TH_ACK) {
941                                 syncache_badack(&inc);
942                                 tcpstat.tcps_badsyn++;
943                                 rstreason = BANDLIM_RST_OPENPORT;
944                                 goto dropwithreset;
945                         }
946                         goto drop;
947                 }
948
949                 /*
950                  * Segment's flags are (SYN) or (SYN | FIN).
951                  */
952 #ifdef INET6
953                 /*
954                  * If deprecated address is forbidden,
955                  * we do not accept SYN to deprecated interface
956                  * address to prevent any new inbound connection from
957                  * getting established.
958                  * When we do not accept SYN, we send a TCP RST,
959                  * with deprecated source address (instead of dropping
960                  * it).  We compromise it as it is much better for peer
961                  * to send a RST, and RST will be the final packet
962                  * for the exchange.
963                  *
964                  * If we do not forbid deprecated addresses, we accept
965                  * the SYN packet.  RFC2462 does not suggest dropping
966                  * SYN in this case.
967                  * If we decipher RFC2462 5.5.4, it says like this:
968                  * 1. use of deprecated addr with existing
969                  *    communication is okay - "SHOULD continue to be
970                  *    used"
971                  * 2. use of it with new communication:
972                  *   (2a) "SHOULD NOT be used if alternate address
973                  *        with sufficient scope is available"
974                  *   (2b) nothing mentioned otherwise.
975                  * Here we fall into (2b) case as we have no choice in
976                  * our source address selection - we must obey the peer.
977                  *
978                  * The wording in RFC2462 is confusing, and there are
979                  * multiple description text for deprecated address
980                  * handling - worse, they are not exactly the same.
981                  * I believe 5.5.4 is the best one, so we follow 5.5.4.
982                  */
983                 if (isipv6 && !ip6_use_deprecated) {
984                         struct in6_ifaddr *ia6;
985
986                         if ((ia6 = ip6_getdstifaddr(m)) &&
987                             (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
988                                 tp = NULL;
989                                 rstreason = BANDLIM_RST_OPENPORT;
990                                 goto dropwithreset;
991                         }
992                 }
993 #endif
994                 /*
995                  * If it is from this socket, drop it, it must be forged.
996                  * Don't bother responding if the destination was a broadcast.
997                  */
998                 if (th->th_dport == th->th_sport) {
999                         if (isipv6) {
1000                                 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1001                                                        &ip6->ip6_src))
1002                                         goto drop;
1003                         } else {
1004                                 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1005                                         goto drop;
1006                         }
1007                 }
1008                 /*
1009                  * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1010                  *
1011                  * Note that it is quite possible to receive unicast
1012                  * link-layer packets with a broadcast IP address. Use
1013                  * in_broadcast() to find them.
1014                  */
1015                 if (m->m_flags & (M_BCAST | M_MCAST))
1016                         goto drop;
1017                 if (isipv6) {
1018                         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1019                             IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1020                                 goto drop;
1021                 } else {
1022                         if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1023                             IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1024                             ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1025                             in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1026                                 goto drop;
1027                 }
1028                 /*
1029                  * SYN appears to be valid; create compressed TCP state
1030                  * for syncache, or perform t/tcp connection.
1031                  */
1032                 if (so->so_qlen <= so->so_qlimit) {
1033                         tcp_dooptions(&to, optp, optlen, TRUE);
1034                         if (!syncache_add(&inc, &to, th, &so, m))
1035                                 goto drop;
1036
1037                         /*
1038                          * Entry added to syncache, mbuf used to
1039                          * send SYN,ACK packet.
1040                          */
1041                         if (so == NULL)
1042                                 return;
1043
1044                         /*
1045                          * We must be in the correct protocol thread for
1046                          * this connection.
1047                          */
1048                         KKASSERT(so->so_port == &curthread->td_msgport);
1049
1050                         inp = so->so_pcb;
1051                         tp = intotcpcb(inp);
1052                         tp->snd_wnd = tiwin;
1053                         tp->t_starttime = ticks;
1054                         tp->t_state = TCPS_ESTABLISHED;
1055
1056                         /*
1057                          * If there is a FIN, or if there is data and the
1058                          * connection is local, then delay SYN,ACK(SYN) in
1059                          * the hope of piggy-backing it on a response
1060                          * segment.  Otherwise must send ACK now in case
1061                          * the other side is slow starting.
1062                          */
1063                         if (DELAY_ACK(tp) &&
1064                             ((thflags & TH_FIN) ||
1065                              (tlen != 0 &&
1066                               ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
1067                                (!isipv6 && in_localaddr(inp->inp_faddr)))))) {
1068                                 tcp_callout_reset(tp, tp->tt_delack,
1069                                     tcp_delacktime, tcp_timer_delack);
1070                                 tp->t_flags |= TF_NEEDSYN;
1071                         } else {
1072                                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
1073                         }
1074
1075                         tcpstat.tcps_connects++;
1076                         soisconnected(so);
1077                         goto trimthenstep6;
1078                 }
1079                 goto drop;
1080         }
1081
1082 after_listen:
1083         /*
1084          * Should not happen - syncache should pick up these connections.
1085          *
1086          * Once we are past handling listen sockets we must be in the
1087          * correct protocol processing thread.
1088          */
1089         KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1090         KKASSERT(so->so_port == &curthread->td_msgport);
1091
1092         /*
1093          * This is the second part of the MSS DoS prevention code (after
1094          * minmss on the sending side) and it deals with too many too small
1095          * tcp packets in a too short timeframe (1 second).
1096          *
1097          * XXX Removed.  This code was crap.  It does not scale to network
1098          *     speed, and default values break NFS.  Gone.
1099          */
1100         /* REMOVED */
1101
1102         /*
1103          * Segment received on connection.
1104          *
1105          * Reset idle time and keep-alive timer.  Don't waste time if less
1106          * then a second has elapsed.  Only update t_rcvtime for non-SYN
1107          * packets.
1108          *
1109          * Handle the case where one side thinks the connection is established
1110          * but the other side has, say, rebooted without cleaning out the
1111          * connection.   The SYNs could be construed as an attack and wind
1112          * up ignored, but in case it isn't an attack we can validate the
1113          * connection by forcing a keepalive.
1114          */
1115         if (TCPS_HAVEESTABLISHED(tp->t_state) && (ticks - tp->t_rcvtime) > hz) {
1116                 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
1117                         tp->t_flags |= TF_KEEPALIVE;
1118                         tcp_callout_reset(tp, tp->tt_keep, hz / 2,
1119                                           tcp_timer_keep);
1120                 } else {
1121                         tp->t_rcvtime = ticks;
1122                         tp->t_flags &= ~TF_KEEPALIVE;
1123                         tcp_callout_reset(tp, tp->tt_keep, tcp_keepidle,
1124                                           tcp_timer_keep);
1125                 }
1126         }
1127
1128         /*
1129          * Process options.
1130          * XXX this is tradtitional behavior, may need to be cleaned up.
1131          */
1132         tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0);
1133         if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1134                 if (to.to_flags & TOF_SCALE) {
1135                         tp->t_flags |= TF_RCVD_SCALE;
1136                         tp->requested_s_scale = to.to_requested_s_scale;
1137                 }
1138                 if (to.to_flags & TOF_TS) {
1139                         tp->t_flags |= TF_RCVD_TSTMP;
1140                         tp->ts_recent = to.to_tsval;
1141                         tp->ts_recent_age = ticks;
1142                 }
1143                 if (to.to_flags & TOF_MSS)
1144                         tcp_mss(tp, to.to_mss);
1145                 /*
1146                  * Only set the TF_SACK_PERMITTED per-connection flag
1147                  * if we got a SACK_PERMITTED option from the other side
1148                  * and the global tcp_do_sack variable is true.
1149                  */
1150                 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1151                         tp->t_flags |= TF_SACK_PERMITTED;
1152         }
1153
1154         /*
1155          * Header prediction: check for the two common cases
1156          * of a uni-directional data xfer.  If the packet has
1157          * no control flags, is in-sequence, the window didn't
1158          * change and we're not retransmitting, it's a
1159          * candidate.  If the length is zero and the ack moved
1160          * forward, we're the sender side of the xfer.  Just
1161          * free the data acked & wake any higher level process
1162          * that was blocked waiting for space.  If the length
1163          * is non-zero and the ack didn't move, we're the
1164          * receiver side.  If we're getting packets in-order
1165          * (the reassembly queue is empty), add the data to
1166          * the socket buffer and note that we need a delayed ack.
1167          * Make sure that the hidden state-flags are also off.
1168          * Since we check for TCPS_ESTABLISHED above, it can only
1169          * be TH_NEEDSYN.
1170          */
1171         if (tp->t_state == TCPS_ESTABLISHED &&
1172             (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1173             !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1174             (!(to.to_flags & TOF_TS) ||
1175              TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1176             th->th_seq == tp->rcv_nxt &&
1177             tp->snd_nxt == tp->snd_max) {
1178
1179                 /*
1180                  * If last ACK falls within this segment's sequence numbers,
1181                  * record the timestamp.
1182                  * NOTE that the test is modified according to the latest
1183                  * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1184                  */
1185                 if ((to.to_flags & TOF_TS) &&
1186                     SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1187                         tp->ts_recent_age = ticks;
1188                         tp->ts_recent = to.to_tsval;
1189                 }
1190
1191                 if (tlen == 0) {
1192                         if (SEQ_GT(th->th_ack, tp->snd_una) &&
1193                             SEQ_LEQ(th->th_ack, tp->snd_max) &&
1194                             tp->snd_cwnd >= tp->snd_wnd &&
1195                             !IN_FASTRECOVERY(tp)) {
1196                                 /*
1197                                  * This is a pure ack for outstanding data.
1198                                  */
1199                                 ++tcpstat.tcps_predack;
1200                                 /*
1201                                  * "bad retransmit" recovery
1202                                  *
1203                                  * If Eifel detection applies, then
1204                                  * it is deterministic, so use it
1205                                  * unconditionally over the old heuristic.
1206                                  * Otherwise, fall back to the old heuristic.
1207                                  */
1208                                 if (tcp_do_eifel_detect &&
1209                                     (to.to_flags & TOF_TS) && to.to_tsecr &&
1210                                     (tp->t_flags & TF_FIRSTACCACK)) {
1211                                         /* Eifel detection applicable. */
1212                                         if (to.to_tsecr < tp->t_rexmtTS) {
1213                                                 tcp_revert_congestion_state(tp);
1214                                                 ++tcpstat.tcps_eifeldetected;
1215                                         }
1216                                 } else if (tp->t_rxtshift == 1 &&
1217                                            ticks < tp->t_badrxtwin) {
1218                                         tcp_revert_congestion_state(tp);
1219                                         ++tcpstat.tcps_rttdetected;
1220                                 }
1221                                 tp->t_flags &= ~(TF_FIRSTACCACK |
1222                                                  TF_FASTREXMT | TF_EARLYREXMT);
1223                                 /*
1224                                  * Recalculate the retransmit timer / rtt.
1225                                  *
1226                                  * Some machines (certain windows boxes)
1227                                  * send broken timestamp replies during the
1228                                  * SYN+ACK phase, ignore timestamps of 0.
1229                                  */
1230                                 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1231                                         tcp_xmit_timer(tp,
1232                                                        ticks - to.to_tsecr + 1);
1233                                 } else if (tp->t_rtttime &&
1234                                            SEQ_GT(th->th_ack, tp->t_rtseq)) {
1235                                         tcp_xmit_timer(tp,
1236                                                        ticks - tp->t_rtttime);
1237                                 }
1238                                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1239                                 acked = th->th_ack - tp->snd_una;
1240                                 tcpstat.tcps_rcvackpack++;
1241                                 tcpstat.tcps_rcvackbyte += acked;
1242                                 sbdrop(&so->so_snd.sb, acked);
1243                                 tp->snd_recover = th->th_ack - 1;
1244                                 tp->snd_una = th->th_ack;
1245                                 tp->t_dupacks = 0;
1246                                 /*
1247                                  * Update window information.
1248                                  */
1249                                 if (tiwin != tp->snd_wnd &&
1250                                     acceptable_window_update(tp, th, tiwin)) {
1251                                         /* keep track of pure window updates */
1252                                         if (tp->snd_wl2 == th->th_ack &&
1253                                             tiwin > tp->snd_wnd)
1254                                                 tcpstat.tcps_rcvwinupd++;
1255                                         tp->snd_wnd = tiwin;
1256                                         tp->snd_wl1 = th->th_seq;
1257                                         tp->snd_wl2 = th->th_ack;
1258                                         if (tp->snd_wnd > tp->max_sndwnd)
1259                                                 tp->max_sndwnd = tp->snd_wnd;
1260                                 }
1261                                 m_freem(m);
1262                                 ND6_HINT(tp); /* some progress has been done */
1263                                 /*
1264                                  * If all outstanding data are acked, stop
1265                                  * retransmit timer, otherwise restart timer
1266                                  * using current (possibly backed-off) value.
1267                                  * If process is waiting for space,
1268                                  * wakeup/selwakeup/signal.  If data
1269                                  * are ready to send, let tcp_output
1270                                  * decide between more output or persist.
1271                                  */
1272                                 if (tp->snd_una == tp->snd_max) {
1273                                         tcp_callout_stop(tp, tp->tt_rexmt);
1274                                 } else if (!tcp_callout_active(tp,
1275                                             tp->tt_persist)) {
1276                                         tcp_callout_reset(tp, tp->tt_rexmt,
1277                                             tp->t_rxtcur, tcp_timer_rexmt);
1278                                 }
1279                                 sowwakeup(so);
1280                                 if (so->so_snd.ssb_cc > 0)
1281                                         tcp_output(tp);
1282                                 return;
1283                         }
1284                 } else if (tiwin == tp->snd_wnd &&
1285                     th->th_ack == tp->snd_una &&
1286                     LIST_EMPTY(&tp->t_segq) &&
1287                     tlen <= ssb_space(&so->so_rcv)) {
1288                         u_long newsize = 0;     /* automatic sockbuf scaling */
1289                         /*
1290                          * This is a pure, in-sequence data packet
1291                          * with nothing on the reassembly queue and
1292                          * we have enough buffer space to take it.
1293                          */
1294                         ++tcpstat.tcps_preddat;
1295                         tp->rcv_nxt += tlen;
1296                         tcpstat.tcps_rcvpack++;
1297                         tcpstat.tcps_rcvbyte += tlen;
1298                         ND6_HINT(tp);   /* some progress has been done */
1299                 /*
1300                  * Automatic sizing of receive socket buffer.  Often the send
1301                  * buffer size is not optimally adjusted to the actual network
1302                  * conditions at hand (delay bandwidth product).  Setting the
1303                  * buffer size too small limits throughput on links with high
1304                  * bandwidth and high delay (eg. trans-continental/oceanic links).
1305                  *
1306                  * On the receive side the socket buffer memory is only rarely
1307                  * used to any significant extent.  This allows us to be much
1308                  * more aggressive in scaling the receive socket buffer.  For
1309                  * the case that the buffer space is actually used to a large
1310                  * extent and we run out of kernel memory we can simply drop
1311                  * the new segments; TCP on the sender will just retransmit it
1312                  * later.  Setting the buffer size too big may only consume too
1313                  * much kernel memory if the application doesn't read() from
1314                  * the socket or packet loss or reordering makes use of the
1315                  * reassembly queue.
1316                  *
1317                  * The criteria to step up the receive buffer one notch are:
1318                  *  1. the number of bytes received during the time it takes
1319                  *     one timestamp to be reflected back to us (the RTT);
1320                  *  2. received bytes per RTT is within seven eighth of the
1321                  *     current socket buffer size;
1322                  *  3. receive buffer size has not hit maximal automatic size;
1323                  *
1324                  * This algorithm does one step per RTT at most and only if
1325                  * we receive a bulk stream w/o packet losses or reorderings.
1326                  * Shrinking the buffer during idle times is not necessary as
1327                  * it doesn't consume any memory when idle.
1328                  *
1329                  * TODO: Only step up if the application is actually serving
1330                  * the buffer to better manage the socket buffer resources.
1331                  */
1332                         if (tcp_do_autorcvbuf &&
1333                             to.to_tsecr &&
1334                             (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1335                                 if (to.to_tsecr > tp->rfbuf_ts &&
1336                                     to.to_tsecr - tp->rfbuf_ts < hz) {
1337                                         if (tp->rfbuf_cnt >
1338                                             (so->so_rcv.ssb_hiwat / 8 * 7) &&
1339                                             so->so_rcv.ssb_hiwat <
1340                                             tcp_autorcvbuf_max) {
1341                                                 newsize =
1342                                                     ulmin(so->so_rcv.ssb_hiwat +
1343                                                           tcp_autorcvbuf_inc,
1344                                                           tcp_autorcvbuf_max);
1345                                         }
1346                                         /* Start over with next RTT. */
1347                                         tp->rfbuf_ts = 0;
1348                                         tp->rfbuf_cnt = 0;
1349                                 } else
1350                                         tp->rfbuf_cnt += tlen;  /* add up */
1351                         }
1352                         /*
1353                          * Add data to socket buffer.
1354                          */
1355                         if (so->so_state & SS_CANTRCVMORE) {
1356                                 m_freem(m);
1357                         } else {
1358                                 /*
1359                                  * Set new socket buffer size, give up when
1360                                  * limit is reached.
1361                                  *
1362                                  * Adjusting the size can mess up ACK
1363                                  * sequencing when pure window updates are
1364                                  * being avoided (which is the default),
1365                                  * so force an ack.
1366                                  */
1367                                 if (newsize) {
1368                                         tp->t_flags |= TF_RXRESIZED;
1369                                         if (!ssb_reserve(&so->so_rcv, newsize,
1370                                                          so, NULL)) {
1371                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1372                                         }
1373                                         if (newsize >=
1374                                             (TCP_MAXWIN << tp->rcv_scale)) {
1375                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1376                                         }
1377                                 }
1378                                 m_adj(m, drop_hdrlen); /* delayed header drop */
1379                                 ssb_appendstream(&so->so_rcv, m);
1380                         }
1381                         sorwakeup(so);
1382                         /*
1383                          * This code is responsible for most of the ACKs
1384                          * the TCP stack sends back after receiving a data
1385                          * packet.  Note that the DELAY_ACK check fails if
1386                          * the delack timer is already running, which results
1387                          * in an ack being sent every other packet (which is
1388                          * what we want).
1389                          *
1390                          * We then further aggregate acks by not actually
1391                          * sending one until the protocol thread has completed
1392                          * processing the current backlog of packets.  This
1393                          * does not delay the ack any further, but allows us
1394                          * to take advantage of the packet aggregation that
1395                          * high speed NICs do (usually blocks of 8-10 packets)
1396                          * to send a single ack rather then four or five acks,
1397                          * greatly reducing the ack rate, the return channel
1398                          * bandwidth, and the protocol overhead on both ends.
1399                          *
1400                          * Since this also has the effect of slowing down
1401                          * the exponential slow-start ramp-up, systems with 
1402                          * very large bandwidth-delay products might want
1403                          * to turn the feature off.
1404                          */
1405                         if (DELAY_ACK(tp)) {
1406                                 tcp_callout_reset(tp, tp->tt_delack,
1407                                     tcp_delacktime, tcp_timer_delack);
1408                         } else if (tcp_aggregate_acks) {
1409                                 tp->t_flags |= TF_ACKNOW;
1410                                 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1411                                         tp->t_flags |= TF_ONOUTPUTQ;
1412                                         tp->tt_cpu = mycpu->gd_cpuid;
1413                                         TAILQ_INSERT_TAIL(
1414                                             &tcpcbackq[tp->tt_cpu],
1415                                             tp, t_outputq);
1416                                 }
1417                         } else {
1418                                 tp->t_flags |= TF_ACKNOW;
1419                                 tcp_output(tp);
1420                         }
1421                         return;
1422                 }
1423         }
1424
1425         /*
1426          * Calculate amount of space in receive window,
1427          * and then do TCP input processing.
1428          * Receive window is amount of space in rcv queue,
1429          * but not less than advertised window.
1430          */
1431         recvwin = ssb_space(&so->so_rcv);
1432         if (recvwin < 0)
1433                 recvwin = 0;
1434         tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1435
1436         /* Reset receive buffer auto scaling when not in bulk receive mode. */
1437         tp->rfbuf_ts = 0;
1438         tp->rfbuf_cnt = 0;
1439
1440         switch (tp->t_state) {
1441         /*
1442          * If the state is SYN_RECEIVED:
1443          *      if seg contains an ACK, but not for our SYN/ACK, send a RST.
1444          */
1445         case TCPS_SYN_RECEIVED:
1446                 if ((thflags & TH_ACK) &&
1447                     (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1448                      SEQ_GT(th->th_ack, tp->snd_max))) {
1449                                 rstreason = BANDLIM_RST_OPENPORT;
1450                                 goto dropwithreset;
1451                 }
1452                 break;
1453
1454         /*
1455          * If the state is SYN_SENT:
1456          *      if seg contains an ACK, but not for our SYN, drop the input.
1457          *      if seg contains a RST, then drop the connection.
1458          *      if seg does not contain SYN, then drop it.
1459          * Otherwise this is an acceptable SYN segment
1460          *      initialize tp->rcv_nxt and tp->irs
1461          *      if seg contains ack then advance tp->snd_una
1462          *      if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1463          *      arrange for segment to be acked (eventually)
1464          *      continue processing rest of data/controls, beginning with URG
1465          */
1466         case TCPS_SYN_SENT:
1467                 if ((thflags & TH_ACK) &&
1468                     (SEQ_LEQ(th->th_ack, tp->iss) ||
1469                      SEQ_GT(th->th_ack, tp->snd_max))) {
1470                         rstreason = BANDLIM_UNLIMITED;
1471                         goto dropwithreset;
1472                 }
1473                 if (thflags & TH_RST) {
1474                         if (thflags & TH_ACK)
1475                                 tp = tcp_drop(tp, ECONNREFUSED);
1476                         goto drop;
1477                 }
1478                 if (!(thflags & TH_SYN))
1479                         goto drop;
1480                 tp->snd_wnd = th->th_win;       /* initial send window */
1481
1482                 tp->irs = th->th_seq;
1483                 tcp_rcvseqinit(tp);
1484                 if (thflags & TH_ACK) {
1485                         /* Our SYN was acked. */
1486                         tcpstat.tcps_connects++;
1487                         soisconnected(so);
1488                         /* Do window scaling on this connection? */
1489                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1490                             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
1491                                 tp->snd_scale = tp->requested_s_scale;
1492                                 tp->rcv_scale = tp->request_r_scale;
1493                         }
1494                         tp->rcv_adv += tp->rcv_wnd;
1495                         tp->snd_una++;          /* SYN is acked */
1496                         tcp_callout_stop(tp, tp->tt_rexmt);
1497                         /*
1498                          * If there's data, delay ACK; if there's also a FIN
1499                          * ACKNOW will be turned on later.
1500                          */
1501                         if (DELAY_ACK(tp) && tlen != 0) {
1502                                 tcp_callout_reset(tp, tp->tt_delack,
1503                                     tcp_delacktime, tcp_timer_delack);
1504                         } else {
1505                                 tp->t_flags |= TF_ACKNOW;
1506                         }
1507                         /*
1508                          * Received <SYN,ACK> in SYN_SENT[*] state.
1509                          * Transitions:
1510                          *      SYN_SENT  --> ESTABLISHED
1511                          *      SYN_SENT* --> FIN_WAIT_1
1512                          */
1513                         tp->t_starttime = ticks;
1514                         if (tp->t_flags & TF_NEEDFIN) {
1515                                 tp->t_state = TCPS_FIN_WAIT_1;
1516                                 tp->t_flags &= ~TF_NEEDFIN;
1517                                 thflags &= ~TH_SYN;
1518                         } else {
1519                                 tp->t_state = TCPS_ESTABLISHED;
1520                                 tcp_callout_reset(tp, tp->tt_keep, tcp_keepidle,
1521                                     tcp_timer_keep);
1522                         }
1523                 } else {
1524                         /*
1525                          * Received initial SYN in SYN-SENT[*] state =>
1526                          * simultaneous open.
1527                          * Do 3-way handshake:
1528                          *        SYN-SENT -> SYN-RECEIVED
1529                          *        SYN-SENT* -> SYN-RECEIVED*
1530                          */
1531                         tp->t_flags |= TF_ACKNOW;
1532                         tcp_callout_stop(tp, tp->tt_rexmt);
1533                         tp->t_state = TCPS_SYN_RECEIVED;
1534                 }
1535
1536 trimthenstep6:
1537                 /*
1538                  * Advance th->th_seq to correspond to first data byte.
1539                  * If data, trim to stay within window,
1540                  * dropping FIN if necessary.
1541                  */
1542                 th->th_seq++;
1543                 if (tlen > tp->rcv_wnd) {
1544                         todrop = tlen - tp->rcv_wnd;
1545                         m_adj(m, -todrop);
1546                         tlen = tp->rcv_wnd;
1547                         thflags &= ~TH_FIN;
1548                         tcpstat.tcps_rcvpackafterwin++;
1549                         tcpstat.tcps_rcvbyteafterwin += todrop;
1550                 }
1551                 tp->snd_wl1 = th->th_seq - 1;
1552                 tp->rcv_up = th->th_seq;
1553                 /*
1554                  * Client side of transaction: already sent SYN and data.
1555                  * If the remote host used T/TCP to validate the SYN,
1556                  * our data will be ACK'd; if so, enter normal data segment
1557                  * processing in the middle of step 5, ack processing.
1558                  * Otherwise, goto step 6.
1559                  */
1560                 if (thflags & TH_ACK)
1561                         goto process_ACK;
1562
1563                 goto step6;
1564
1565         /*
1566          * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1567          *      do normal processing (we no longer bother with T/TCP).
1568          */
1569         case TCPS_LAST_ACK:
1570         case TCPS_CLOSING:
1571         case TCPS_TIME_WAIT:
1572                 break;  /* continue normal processing */
1573         }
1574
1575         /*
1576          * States other than LISTEN or SYN_SENT.
1577          * First check the RST flag and sequence number since reset segments
1578          * are exempt from the timestamp and connection count tests.  This
1579          * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1580          * below which allowed reset segments in half the sequence space
1581          * to fall though and be processed (which gives forged reset
1582          * segments with a random sequence number a 50 percent chance of
1583          * killing a connection).
1584          * Then check timestamp, if present.
1585          * Then check the connection count, if present.
1586          * Then check that at least some bytes of segment are within
1587          * receive window.  If segment begins before rcv_nxt,
1588          * drop leading data (and SYN); if nothing left, just ack.
1589          *
1590          *
1591          * If the RST bit is set, check the sequence number to see
1592          * if this is a valid reset segment.
1593          * RFC 793 page 37:
1594          *   In all states except SYN-SENT, all reset (RST) segments
1595          *   are validated by checking their SEQ-fields.  A reset is
1596          *   valid if its sequence number is in the window.
1597          * Note: this does not take into account delayed ACKs, so
1598          *   we should test against last_ack_sent instead of rcv_nxt.
1599          *   The sequence number in the reset segment is normally an
1600          *   echo of our outgoing acknowledgement numbers, but some hosts
1601          *   send a reset with the sequence number at the rightmost edge
1602          *   of our receive window, and we have to handle this case.
1603          * If we have multiple segments in flight, the intial reset
1604          * segment sequence numbers will be to the left of last_ack_sent,
1605          * but they will eventually catch up.
1606          * In any case, it never made sense to trim reset segments to
1607          * fit the receive window since RFC 1122 says:
1608          *   4.2.2.12  RST Segment: RFC-793 Section 3.4
1609          *
1610          *    A TCP SHOULD allow a received RST segment to include data.
1611          *
1612          *    DISCUSSION
1613          *         It has been suggested that a RST segment could contain
1614          *         ASCII text that encoded and explained the cause of the
1615          *         RST.  No standard has yet been established for such
1616          *         data.
1617          *
1618          * If the reset segment passes the sequence number test examine
1619          * the state:
1620          *    SYN_RECEIVED STATE:
1621          *      If passive open, return to LISTEN state.
1622          *      If active open, inform user that connection was refused.
1623          *    ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1624          *      Inform user that connection was reset, and close tcb.
1625          *    CLOSING, LAST_ACK STATES:
1626          *      Close the tcb.
1627          *    TIME_WAIT STATE:
1628          *      Drop the segment - see Stevens, vol. 2, p. 964 and
1629          *      RFC 1337.
1630          */
1631         if (thflags & TH_RST) {
1632                 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1633                     SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1634                         switch (tp->t_state) {
1635
1636                         case TCPS_SYN_RECEIVED:
1637                                 so->so_error = ECONNREFUSED;
1638                                 goto close;
1639
1640                         case TCPS_ESTABLISHED:
1641                         case TCPS_FIN_WAIT_1:
1642                         case TCPS_FIN_WAIT_2:
1643                         case TCPS_CLOSE_WAIT:
1644                                 so->so_error = ECONNRESET;
1645                         close:
1646                                 tp->t_state = TCPS_CLOSED;
1647                                 tcpstat.tcps_drops++;
1648                                 tp = tcp_close(tp);
1649                                 break;
1650
1651                         case TCPS_CLOSING:
1652                         case TCPS_LAST_ACK:
1653                                 tp = tcp_close(tp);
1654                                 break;
1655
1656                         case TCPS_TIME_WAIT:
1657                                 break;
1658                         }
1659                 }
1660                 goto drop;
1661         }
1662
1663         /*
1664          * RFC 1323 PAWS: If we have a timestamp reply on this segment
1665          * and it's less than ts_recent, drop it.
1666          */
1667         if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1668             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1669
1670                 /* Check to see if ts_recent is over 24 days old.  */
1671                 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1672                         /*
1673                          * Invalidate ts_recent.  If this segment updates
1674                          * ts_recent, the age will be reset later and ts_recent
1675                          * will get a valid value.  If it does not, setting
1676                          * ts_recent to zero will at least satisfy the
1677                          * requirement that zero be placed in the timestamp
1678                          * echo reply when ts_recent isn't valid.  The
1679                          * age isn't reset until we get a valid ts_recent
1680                          * because we don't want out-of-order segments to be
1681                          * dropped when ts_recent is old.
1682                          */
1683                         tp->ts_recent = 0;
1684                 } else {
1685                         tcpstat.tcps_rcvduppack++;
1686                         tcpstat.tcps_rcvdupbyte += tlen;
1687                         tcpstat.tcps_pawsdrop++;
1688                         if (tlen)
1689                                 goto dropafterack;
1690                         goto drop;
1691                 }
1692         }
1693
1694         /*
1695          * In the SYN-RECEIVED state, validate that the packet belongs to
1696          * this connection before trimming the data to fit the receive
1697          * window.  Check the sequence number versus IRS since we know
1698          * the sequence numbers haven't wrapped.  This is a partial fix
1699          * for the "LAND" DoS attack.
1700          */
1701         if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1702                 rstreason = BANDLIM_RST_OPENPORT;
1703                 goto dropwithreset;
1704         }
1705
1706         todrop = tp->rcv_nxt - th->th_seq;
1707         if (todrop > 0) {
1708                 if (TCP_DO_SACK(tp)) {
1709                         /* Report duplicate segment at head of packet. */
1710                         tp->reportblk.rblk_start = th->th_seq;
1711                         tp->reportblk.rblk_end = th->th_seq + tlen;
1712                         if (thflags & TH_FIN)
1713                                 ++tp->reportblk.rblk_end;
1714                         if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1715                                 tp->reportblk.rblk_end = tp->rcv_nxt;
1716                         tp->t_flags |= (TF_DUPSEG | TF_SACKLEFT | TF_ACKNOW);
1717                 }
1718                 if (thflags & TH_SYN) {
1719                         thflags &= ~TH_SYN;
1720                         th->th_seq++;
1721                         if (th->th_urp > 1)
1722                                 th->th_urp--;
1723                         else
1724                                 thflags &= ~TH_URG;
1725                         todrop--;
1726                 }
1727                 /*
1728                  * Following if statement from Stevens, vol. 2, p. 960.
1729                  */
1730                 if (todrop > tlen ||
1731                     (todrop == tlen && !(thflags & TH_FIN))) {
1732                         /*
1733                          * Any valid FIN must be to the left of the window.
1734                          * At this point the FIN must be a duplicate or out
1735                          * of sequence; drop it.
1736                          */
1737                         thflags &= ~TH_FIN;
1738
1739                         /*
1740                          * Send an ACK to resynchronize and drop any data.
1741                          * But keep on processing for RST or ACK.
1742                          */
1743                         tp->t_flags |= TF_ACKNOW;
1744                         todrop = tlen;
1745                         tcpstat.tcps_rcvduppack++;
1746                         tcpstat.tcps_rcvdupbyte += todrop;
1747                 } else {
1748                         tcpstat.tcps_rcvpartduppack++;
1749                         tcpstat.tcps_rcvpartdupbyte += todrop;
1750                 }
1751                 drop_hdrlen += todrop;  /* drop from the top afterwards */
1752                 th->th_seq += todrop;
1753                 tlen -= todrop;
1754                 if (th->th_urp > todrop)
1755                         th->th_urp -= todrop;
1756                 else {
1757                         thflags &= ~TH_URG;
1758                         th->th_urp = 0;
1759                 }
1760         }
1761
1762         /*
1763          * If new data are received on a connection after the
1764          * user processes are gone, then RST the other end.
1765          */
1766         if ((so->so_state & SS_NOFDREF) &&
1767             tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1768                 tp = tcp_close(tp);
1769                 tcpstat.tcps_rcvafterclose++;
1770                 rstreason = BANDLIM_UNLIMITED;
1771                 goto dropwithreset;
1772         }
1773
1774         /*
1775          * If segment ends after window, drop trailing data
1776          * (and PUSH and FIN); if nothing left, just ACK.
1777          */
1778         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1779         if (todrop > 0) {
1780                 tcpstat.tcps_rcvpackafterwin++;
1781                 if (todrop >= tlen) {
1782                         tcpstat.tcps_rcvbyteafterwin += tlen;
1783                         /*
1784                          * If a new connection request is received
1785                          * while in TIME_WAIT, drop the old connection
1786                          * and start over if the sequence numbers
1787                          * are above the previous ones.
1788                          */
1789                         if (thflags & TH_SYN &&
1790                             tp->t_state == TCPS_TIME_WAIT &&
1791                             SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1792                                 tp = tcp_close(tp);
1793                                 goto findpcb;
1794                         }
1795                         /*
1796                          * If window is closed can only take segments at
1797                          * window edge, and have to drop data and PUSH from
1798                          * incoming segments.  Continue processing, but
1799                          * remember to ack.  Otherwise, drop segment
1800                          * and ack.
1801                          */
1802                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1803                                 tp->t_flags |= TF_ACKNOW;
1804                                 tcpstat.tcps_rcvwinprobe++;
1805                         } else
1806                                 goto dropafterack;
1807                 } else
1808                         tcpstat.tcps_rcvbyteafterwin += todrop;
1809                 m_adj(m, -todrop);
1810                 tlen -= todrop;
1811                 thflags &= ~(TH_PUSH | TH_FIN);
1812         }
1813
1814         /*
1815          * If last ACK falls within this segment's sequence numbers,
1816          * record its timestamp.
1817          * NOTE:
1818          * 1) That the test incorporates suggestions from the latest
1819          *    proposal of the tcplw@cray.com list (Braden 1993/04/26).
1820          * 2) That updating only on newer timestamps interferes with
1821          *    our earlier PAWS tests, so this check should be solely
1822          *    predicated on the sequence space of this segment.
1823          * 3) That we modify the segment boundary check to be
1824          *        Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1825          *    instead of RFC1323's
1826          *        Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1827          *    This modified check allows us to overcome RFC1323's
1828          *    limitations as described in Stevens TCP/IP Illustrated
1829          *    Vol. 2 p.869. In such cases, we can still calculate the
1830          *    RTT correctly when RCV.NXT == Last.ACK.Sent.
1831          */
1832         if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1833             SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1834                                         + ((thflags & TH_SYN) != 0)
1835                                         + ((thflags & TH_FIN) != 0)))) {
1836                 tp->ts_recent_age = ticks;
1837                 tp->ts_recent = to.to_tsval;
1838         }
1839
1840         /*
1841          * If a SYN is in the window, then this is an
1842          * error and we send an RST and drop the connection.
1843          */
1844         if (thflags & TH_SYN) {
1845                 tp = tcp_drop(tp, ECONNRESET);
1846                 rstreason = BANDLIM_UNLIMITED;
1847                 goto dropwithreset;
1848         }
1849
1850         /*
1851          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN
1852          * flag is on (half-synchronized state), then queue data for
1853          * later processing; else drop segment and return.
1854          */
1855         if (!(thflags & TH_ACK)) {
1856                 if (tp->t_state == TCPS_SYN_RECEIVED ||
1857                     (tp->t_flags & TF_NEEDSYN))
1858                         goto step6;
1859                 else
1860                         goto drop;
1861         }
1862
1863         /*
1864          * Ack processing.
1865          */
1866         switch (tp->t_state) {
1867         /*
1868          * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1869          * ESTABLISHED state and continue processing.
1870          * The ACK was checked above.
1871          */
1872         case TCPS_SYN_RECEIVED:
1873
1874                 tcpstat.tcps_connects++;
1875                 soisconnected(so);
1876                 /* Do window scaling? */
1877                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1878                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
1879                         tp->snd_scale = tp->requested_s_scale;
1880                         tp->rcv_scale = tp->request_r_scale;
1881                 }
1882                 /*
1883                  * Make transitions:
1884                  *      SYN-RECEIVED  -> ESTABLISHED
1885                  *      SYN-RECEIVED* -> FIN-WAIT-1
1886                  */
1887                 tp->t_starttime = ticks;
1888                 if (tp->t_flags & TF_NEEDFIN) {
1889                         tp->t_state = TCPS_FIN_WAIT_1;
1890                         tp->t_flags &= ~TF_NEEDFIN;
1891                 } else {
1892                         tp->t_state = TCPS_ESTABLISHED;
1893                         tcp_callout_reset(tp, tp->tt_keep, tcp_keepidle,
1894                             tcp_timer_keep);
1895                 }
1896                 /*
1897                  * If segment contains data or ACK, will call tcp_reass()
1898                  * later; if not, do so now to pass queued data to user.
1899                  */
1900                 if (tlen == 0 && !(thflags & TH_FIN))
1901                         tcp_reass(tp, NULL, NULL, NULL);
1902                 /* fall into ... */
1903
1904         /*
1905          * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
1906          * ACKs.  If the ack is in the range
1907          *      tp->snd_una < th->th_ack <= tp->snd_max
1908          * then advance tp->snd_una to th->th_ack and drop
1909          * data from the retransmission queue.  If this ACK reflects
1910          * more up to date window information we update our window information.
1911          */
1912         case TCPS_ESTABLISHED:
1913         case TCPS_FIN_WAIT_1:
1914         case TCPS_FIN_WAIT_2:
1915         case TCPS_CLOSE_WAIT:
1916         case TCPS_CLOSING:
1917         case TCPS_LAST_ACK:
1918         case TCPS_TIME_WAIT:
1919
1920                 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
1921                         if (TCP_DO_SACK(tp))
1922                                 tcp_sack_update_scoreboard(tp, &to);
1923                         if (tlen != 0 || tiwin != tp->snd_wnd) {
1924                                 tp->t_dupacks = 0;
1925                                 break;
1926                         }
1927                         tcpstat.tcps_rcvdupack++;
1928                         if (!tcp_callout_active(tp, tp->tt_rexmt) ||
1929                             th->th_ack != tp->snd_una) {
1930                                 tp->t_dupacks = 0;
1931                                 break;
1932                         }
1933                         /*
1934                          * We have outstanding data (other than
1935                          * a window probe), this is a completely
1936                          * duplicate ack (ie, window info didn't
1937                          * change), the ack is the biggest we've
1938                          * seen and we've seen exactly our rexmt
1939                          * threshhold of them, so assume a packet
1940                          * has been dropped and retransmit it.
1941                          * Kludge snd_nxt & the congestion
1942                          * window so we send only this one
1943                          * packet.
1944                          */
1945                         if (IN_FASTRECOVERY(tp)) {
1946                                 if (TCP_DO_SACK(tp)) {
1947                                         /* No artifical cwnd inflation. */
1948                                         tcp_sack_rexmt(tp, th);
1949                                 } else {
1950                                         /*
1951                                          * Dup acks mean that packets
1952                                          * have left the network
1953                                          * (they're now cached at the
1954                                          * receiver) so bump cwnd by
1955                                          * the amount in the receiver
1956                                          * to keep a constant cwnd
1957                                          * packets in the network.
1958                                          */
1959                                         tp->snd_cwnd += tp->t_maxseg;
1960                                         tcp_output(tp);
1961                                 }
1962                         } else if (SEQ_LT(th->th_ack, tp->snd_recover)) {
1963                                 tp->t_dupacks = 0;
1964                                 break;
1965                         } else if (++tp->t_dupacks == tcprexmtthresh) {
1966                                 tcp_seq old_snd_nxt;
1967                                 u_int win;
1968
1969 fastretransmit:
1970                                 if (tcp_do_eifel_detect &&
1971                                     (tp->t_flags & TF_RCVD_TSTMP)) {
1972                                         tcp_save_congestion_state(tp);
1973                                         tp->t_flags |= TF_FASTREXMT;
1974                                 }
1975                                 /*
1976                                  * We know we're losing at the current
1977                                  * window size, so do congestion avoidance:
1978                                  * set ssthresh to half the current window
1979                                  * and pull our congestion window back to the
1980                                  * new ssthresh.
1981                                  */
1982                                 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 /
1983                                     tp->t_maxseg;
1984                                 if (win < 2)
1985                                         win = 2;
1986                                 tp->snd_ssthresh = win * tp->t_maxseg;
1987                                 ENTER_FASTRECOVERY(tp);
1988                                 tp->snd_recover = tp->snd_max;
1989                                 tcp_callout_stop(tp, tp->tt_rexmt);
1990                                 tp->t_rtttime = 0;
1991                                 old_snd_nxt = tp->snd_nxt;
1992                                 tp->snd_nxt = th->th_ack;
1993                                 tp->snd_cwnd = tp->t_maxseg;
1994                                 tcp_output(tp);
1995                                 ++tcpstat.tcps_sndfastrexmit;
1996                                 tp->snd_cwnd = tp->snd_ssthresh;
1997                                 tp->rexmt_high = tp->snd_nxt;
1998                                 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
1999                                         tp->snd_nxt = old_snd_nxt;
2000                                 KASSERT(tp->snd_limited <= 2,
2001                                     ("tp->snd_limited too big"));
2002                                 if (TCP_DO_SACK(tp))
2003                                         tcp_sack_rexmt(tp, th);
2004                                 else
2005                                         tp->snd_cwnd += tp->t_maxseg *
2006                                             (tp->t_dupacks - tp->snd_limited);
2007                         } else if (tcp_do_limitedtransmit) {
2008                                 u_long oldcwnd = tp->snd_cwnd;
2009                                 tcp_seq oldsndmax = tp->snd_max;
2010                                 tcp_seq oldsndnxt = tp->snd_nxt;
2011                                 /* outstanding data */
2012                                 uint32_t ownd = tp->snd_max - tp->snd_una;
2013                                 u_int sent;
2014
2015 #define iceildiv(n, d)          (((n)+(d)-1) / (d))
2016
2017                                 KASSERT(tp->t_dupacks == 1 ||
2018                                         tp->t_dupacks == 2,
2019                                     ("dupacks not 1 or 2"));
2020                                 if (tp->t_dupacks == 1)
2021                                         tp->snd_limited = 0;
2022                                 tp->snd_nxt = tp->snd_max;
2023                                 tp->snd_cwnd = ownd +
2024                                     (tp->t_dupacks - tp->snd_limited) *
2025                                     tp->t_maxseg;
2026                                 tcp_output(tp);
2027
2028                                 /*
2029                                  * Other acks may have been processed,
2030                                  * snd_nxt cannot be reset to a value less
2031                                  * then snd_una.
2032                                  */
2033                                 if (SEQ_LT(oldsndnxt, oldsndmax)) {
2034                                     if (SEQ_GT(oldsndnxt, tp->snd_una))
2035                                         tp->snd_nxt = oldsndnxt;
2036                                     else
2037                                         tp->snd_nxt = tp->snd_una;
2038                                 }
2039                                 tp->snd_cwnd = oldcwnd;
2040                                 sent = tp->snd_max - oldsndmax;
2041                                 if (sent > tp->t_maxseg) {
2042                                         KASSERT((tp->t_dupacks == 2 &&
2043                                                  tp->snd_limited == 0) ||
2044                                                 (sent == tp->t_maxseg + 1 &&
2045                                                  tp->t_flags & TF_SENTFIN),
2046                                             ("sent too much"));
2047                                         KASSERT(sent <= tp->t_maxseg * 2,
2048                                             ("sent too many segments"));
2049                                         tp->snd_limited = 2;
2050                                         tcpstat.tcps_sndlimited += 2;
2051                                 } else if (sent > 0) {
2052                                         ++tp->snd_limited;
2053                                         ++tcpstat.tcps_sndlimited;
2054                                 } else if (tcp_do_early_retransmit &&
2055                                     (tcp_do_eifel_detect &&
2056                                      (tp->t_flags & TF_RCVD_TSTMP)) &&
2057                                     ownd < 4 * tp->t_maxseg &&
2058                                     tp->t_dupacks + 1 >=
2059                                       iceildiv(ownd, tp->t_maxseg) &&
2060                                     (!TCP_DO_SACK(tp) ||
2061                                      ownd <= tp->t_maxseg ||
2062                                      tcp_sack_has_sacked(&tp->scb,
2063                                                         ownd - tp->t_maxseg))) {
2064                                         ++tcpstat.tcps_sndearlyrexmit;
2065                                         tp->t_flags |= TF_EARLYREXMT;
2066                                         goto fastretransmit;
2067                                 }
2068                         }
2069                         goto drop;
2070                 }
2071
2072                 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2073                 tp->t_dupacks = 0;
2074                 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2075                         /*
2076                          * Detected optimistic ACK attack.
2077                          * Force slow-start to de-synchronize attack.
2078                          */
2079                         tp->snd_cwnd = tp->t_maxseg;
2080                         tp->snd_wacked = 0;
2081
2082                         tcpstat.tcps_rcvacktoomuch++;
2083                         goto dropafterack;
2084                 }
2085                 /*
2086                  * If we reach this point, ACK is not a duplicate,
2087                  *     i.e., it ACKs something we sent.
2088                  */
2089                 if (tp->t_flags & TF_NEEDSYN) {
2090                         /*
2091                          * T/TCP: Connection was half-synchronized, and our
2092                          * SYN has been ACK'd (so connection is now fully
2093                          * synchronized).  Go to non-starred state,
2094                          * increment snd_una for ACK of SYN, and check if
2095                          * we can do window scaling.
2096                          */
2097                         tp->t_flags &= ~TF_NEEDSYN;
2098                         tp->snd_una++;
2099                         /* Do window scaling? */
2100                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2101                             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
2102                                 tp->snd_scale = tp->requested_s_scale;
2103                                 tp->rcv_scale = tp->request_r_scale;
2104                         }
2105                 }
2106
2107 process_ACK:
2108                 acked = th->th_ack - tp->snd_una;
2109                 tcpstat.tcps_rcvackpack++;
2110                 tcpstat.tcps_rcvackbyte += acked;
2111
2112                 if (tcp_do_eifel_detect && acked > 0 &&
2113                     (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2114                     (tp->t_flags & TF_FIRSTACCACK)) {
2115                         /* Eifel detection applicable. */
2116                         if (to.to_tsecr < tp->t_rexmtTS) {
2117                                 ++tcpstat.tcps_eifeldetected;
2118                                 tcp_revert_congestion_state(tp);
2119                                 if (tp->t_rxtshift == 1 &&
2120                                     ticks >= tp->t_badrxtwin)
2121                                         ++tcpstat.tcps_rttcantdetect;
2122                         }
2123                 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2124                         /*
2125                          * If we just performed our first retransmit,
2126                          * and the ACK arrives within our recovery window,
2127                          * then it was a mistake to do the retransmit
2128                          * in the first place.  Recover our original cwnd
2129                          * and ssthresh, and proceed to transmit where we
2130                          * left off.
2131                          */
2132                         tcp_revert_congestion_state(tp);
2133                         ++tcpstat.tcps_rttdetected;
2134                 }
2135
2136                 /*
2137                  * If we have a timestamp reply, update smoothed
2138                  * round trip time.  If no timestamp is present but
2139                  * transmit timer is running and timed sequence
2140                  * number was acked, update smoothed round trip time.
2141                  * Since we now have an rtt measurement, cancel the
2142                  * timer backoff (cf., Phil Karn's retransmit alg.).
2143                  * Recompute the initial retransmit timer.
2144                  *
2145                  * Some machines (certain windows boxes) send broken
2146                  * timestamp replies during the SYN+ACK phase, ignore
2147                  * timestamps of 0.
2148                  */
2149                 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2150                         tcp_xmit_timer(tp, ticks - to.to_tsecr + 1);
2151                 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2152                         tcp_xmit_timer(tp, ticks - tp->t_rtttime);
2153                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2154
2155                 /*
2156                  * If no data (only SYN) was ACK'd,
2157                  *    skip rest of ACK processing.
2158                  */
2159                 if (acked == 0)
2160                         goto step6;
2161
2162                 /* Stop looking for an acceptable ACK since one was received. */
2163                 tp->t_flags &= ~(TF_FIRSTACCACK | TF_FASTREXMT | TF_EARLYREXMT);
2164
2165                 if (acked > so->so_snd.ssb_cc) {
2166                         tp->snd_wnd -= so->so_snd.ssb_cc;
2167                         sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2168                         ourfinisacked = TRUE;
2169                 } else {
2170                         sbdrop(&so->so_snd.sb, acked);
2171                         tp->snd_wnd -= acked;
2172                         ourfinisacked = FALSE;
2173                 }
2174                 sowwakeup(so);
2175
2176                 /*
2177                  * Update window information.
2178                  * Don't look at window if no ACK:
2179                  * TAC's send garbage on first SYN.
2180                  */
2181                 if (SEQ_LT(tp->snd_wl1, th->th_seq) ||
2182                     (tp->snd_wl1 == th->th_seq &&
2183                      (SEQ_LT(tp->snd_wl2, th->th_ack) ||
2184                       (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)))) {
2185                         /* keep track of pure window updates */
2186                         if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2187                             tiwin > tp->snd_wnd)
2188                                 tcpstat.tcps_rcvwinupd++;
2189                         tp->snd_wnd = tiwin;
2190                         tp->snd_wl1 = th->th_seq;
2191                         tp->snd_wl2 = th->th_ack;
2192                         if (tp->snd_wnd > tp->max_sndwnd)
2193                                 tp->max_sndwnd = tp->snd_wnd;
2194                         needoutput = TRUE;
2195                 }
2196
2197                 tp->snd_una = th->th_ack;
2198                 if (TCP_DO_SACK(tp))
2199                         tcp_sack_update_scoreboard(tp, &to);
2200                 if (IN_FASTRECOVERY(tp)) {
2201                         if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2202                                 EXIT_FASTRECOVERY(tp);
2203                                 needoutput = TRUE;
2204                                 /*
2205                                  * If the congestion window was inflated
2206                                  * to account for the other side's
2207                                  * cached packets, retract it.
2208                                  */
2209                                 if (!TCP_DO_SACK(tp))
2210                                         tp->snd_cwnd = tp->snd_ssthresh;
2211
2212                                 /*
2213                                  * Window inflation should have left us
2214                                  * with approximately snd_ssthresh outstanding
2215                                  * data.  But, in case we would be inclined
2216                                  * to send a burst, better do it using
2217                                  * slow start.
2218                                  */
2219                                 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2220                                            tp->snd_max + 2 * tp->t_maxseg))
2221                                         tp->snd_cwnd =
2222                                             (tp->snd_max - tp->snd_una) +
2223                                             2 * tp->t_maxseg;
2224
2225                                 tp->snd_wacked = 0;
2226                         } else {
2227                                 if (TCP_DO_SACK(tp)) {
2228                                         tp->snd_max_rexmt = tp->snd_max;
2229                                         tcp_sack_rexmt(tp, th);
2230                                 } else {
2231                                         tcp_newreno_partial_ack(tp, th, acked);
2232                                 }
2233                                 needoutput = FALSE;
2234                         }
2235                 } else {
2236                         /*
2237                          * Open the congestion window.  When in slow-start,
2238                          * open exponentially: maxseg per packet.  Otherwise,
2239                          * open linearly: maxseg per window.
2240                          */
2241                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
2242                                 u_int abc_sslimit =
2243                                     (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2244                                      tp->t_maxseg : 2 * tp->t_maxseg);
2245
2246                                 /* slow-start */
2247                                 tp->snd_cwnd += tcp_do_abc ?
2248                                     min(acked, abc_sslimit) : tp->t_maxseg;
2249                         } else {
2250                                 /* linear increase */
2251                                 tp->snd_wacked += tcp_do_abc ? acked :
2252                                     tp->t_maxseg;
2253                                 if (tp->snd_wacked >= tp->snd_cwnd) {
2254                                         tp->snd_wacked -= tp->snd_cwnd;
2255                                         tp->snd_cwnd += tp->t_maxseg;
2256                                 }
2257                         }
2258                         tp->snd_cwnd = min(tp->snd_cwnd,
2259                                            TCP_MAXWIN << tp->snd_scale);
2260                         tp->snd_recover = th->th_ack - 1;
2261                 }
2262                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2263                         tp->snd_nxt = tp->snd_una;
2264
2265                 /*
2266                  * If all outstanding data is acked, stop retransmit
2267                  * timer and remember to restart (more output or persist).
2268                  * If there is more data to be acked, restart retransmit
2269                  * timer, using current (possibly backed-off) value.
2270                  */
2271                 if (th->th_ack == tp->snd_max) {
2272                         tcp_callout_stop(tp, tp->tt_rexmt);
2273                         needoutput = TRUE;
2274                 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2275                         tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2276                             tcp_timer_rexmt);
2277                 }
2278
2279                 switch (tp->t_state) {
2280                 /*
2281                  * In FIN_WAIT_1 STATE in addition to the processing
2282                  * for the ESTABLISHED state if our FIN is now acknowledged
2283                  * then enter FIN_WAIT_2.
2284                  */
2285                 case TCPS_FIN_WAIT_1:
2286                         if (ourfinisacked) {
2287                                 /*
2288                                  * If we can't receive any more
2289                                  * data, then closing user can proceed.
2290                                  * Starting the timer is contrary to the
2291                                  * specification, but if we don't get a FIN
2292                                  * we'll hang forever.
2293                                  */
2294                                 if (so->so_state & SS_CANTRCVMORE) {
2295                                         soisdisconnected(so);
2296                                         tcp_callout_reset(tp, tp->tt_2msl,
2297                                             tcp_maxidle, tcp_timer_2msl);
2298                                 }
2299                                 tp->t_state = TCPS_FIN_WAIT_2;
2300                         }
2301                         break;
2302
2303                 /*
2304                  * In CLOSING STATE in addition to the processing for
2305                  * the ESTABLISHED state if the ACK acknowledges our FIN
2306                  * then enter the TIME-WAIT state, otherwise ignore
2307                  * the segment.
2308                  */
2309                 case TCPS_CLOSING:
2310                         if (ourfinisacked) {
2311                                 tp->t_state = TCPS_TIME_WAIT;
2312                                 tcp_canceltimers(tp);
2313                                 tcp_callout_reset(tp, tp->tt_2msl,
2314                                             2 * tcp_msl, tcp_timer_2msl);
2315                                 soisdisconnected(so);
2316                         }
2317                         break;
2318
2319                 /*
2320                  * In LAST_ACK, we may still be waiting for data to drain
2321                  * and/or to be acked, as well as for the ack of our FIN.
2322                  * If our FIN is now acknowledged, delete the TCB,
2323                  * enter the closed state and return.
2324                  */
2325                 case TCPS_LAST_ACK:
2326                         if (ourfinisacked) {
2327                                 tp = tcp_close(tp);
2328                                 goto drop;
2329                         }
2330                         break;
2331
2332                 /*
2333                  * In TIME_WAIT state the only thing that should arrive
2334                  * is a retransmission of the remote FIN.  Acknowledge
2335                  * it and restart the finack timer.
2336                  */
2337                 case TCPS_TIME_WAIT:
2338                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2339                             tcp_timer_2msl);
2340                         goto dropafterack;
2341                 }
2342         }
2343
2344 step6:
2345         /*
2346          * Update window information.
2347          * Don't look at window if no ACK: TAC's send garbage on first SYN.
2348          */
2349         if ((thflags & TH_ACK) &&
2350             acceptable_window_update(tp, th, tiwin)) {
2351                 /* keep track of pure window updates */
2352                 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2353                     tiwin > tp->snd_wnd)
2354                         tcpstat.tcps_rcvwinupd++;
2355                 tp->snd_wnd = tiwin;
2356                 tp->snd_wl1 = th->th_seq;
2357                 tp->snd_wl2 = th->th_ack;
2358                 if (tp->snd_wnd > tp->max_sndwnd)
2359                         tp->max_sndwnd = tp->snd_wnd;
2360                 needoutput = TRUE;
2361         }
2362
2363         /*
2364          * Process segments with URG.
2365          */
2366         if ((thflags & TH_URG) && th->th_urp &&
2367             !TCPS_HAVERCVDFIN(tp->t_state)) {
2368                 /*
2369                  * This is a kludge, but if we receive and accept
2370                  * random urgent pointers, we'll crash in
2371                  * soreceive.  It's hard to imagine someone
2372                  * actually wanting to send this much urgent data.
2373                  */
2374                 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2375                         th->th_urp = 0;                 /* XXX */
2376                         thflags &= ~TH_URG;             /* XXX */
2377                         goto dodata;                    /* XXX */
2378                 }
2379                 /*
2380                  * If this segment advances the known urgent pointer,
2381                  * then mark the data stream.  This should not happen
2382                  * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2383                  * a FIN has been received from the remote side.
2384                  * In these states we ignore the URG.
2385                  *
2386                  * According to RFC961 (Assigned Protocols),
2387                  * the urgent pointer points to the last octet
2388                  * of urgent data.  We continue, however,
2389                  * to consider it to indicate the first octet
2390                  * of data past the urgent section as the original
2391                  * spec states (in one of two places).
2392                  */
2393                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2394                         tp->rcv_up = th->th_seq + th->th_urp;
2395                         so->so_oobmark = so->so_rcv.ssb_cc +
2396                             (tp->rcv_up - tp->rcv_nxt) - 1;
2397                         if (so->so_oobmark == 0)
2398                                 so->so_state |= SS_RCVATMARK;
2399                         sohasoutofband(so);
2400                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2401                 }
2402                 /*
2403                  * Remove out of band data so doesn't get presented to user.
2404                  * This can happen independent of advancing the URG pointer,
2405                  * but if two URG's are pending at once, some out-of-band
2406                  * data may creep in... ick.
2407                  */
2408                 if (th->th_urp <= (u_long)tlen &&
2409                     !(so->so_options & SO_OOBINLINE)) {
2410                         /* hdr drop is delayed */
2411                         tcp_pulloutofband(so, th, m, drop_hdrlen);
2412                 }
2413         } else {
2414                 /*
2415                  * If no out of band data is expected,
2416                  * pull receive urgent pointer along
2417                  * with the receive window.
2418                  */
2419                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2420                         tp->rcv_up = tp->rcv_nxt;
2421         }
2422
2423 dodata:                                                 /* XXX */
2424         /*
2425          * Process the segment text, merging it into the TCP sequencing queue,
2426          * and arranging for acknowledgment of receipt if necessary.
2427          * This process logically involves adjusting tp->rcv_wnd as data
2428          * is presented to the user (this happens in tcp_usrreq.c,
2429          * case PRU_RCVD).  If a FIN has already been received on this
2430          * connection then we just ignore the text.
2431          */
2432         if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2433                 m_adj(m, drop_hdrlen);  /* delayed header drop */
2434                 /*
2435                  * Insert segment which includes th into TCP reassembly queue
2436                  * with control block tp.  Set thflags to whether reassembly now
2437                  * includes a segment with FIN.  This handles the common case
2438                  * inline (segment is the next to be received on an established
2439                  * connection, and the queue is empty), avoiding linkage into
2440                  * and removal from the queue and repetition of various
2441                  * conversions.
2442                  * Set DELACK for segments received in order, but ack
2443                  * immediately when segments are out of order (so
2444                  * fast retransmit can work).
2445                  */
2446                 if (th->th_seq == tp->rcv_nxt &&
2447                     LIST_EMPTY(&tp->t_segq) &&
2448                     TCPS_HAVEESTABLISHED(tp->t_state)) {
2449                         if (DELAY_ACK(tp)) {
2450                                 tcp_callout_reset(tp, tp->tt_delack,
2451                                     tcp_delacktime, tcp_timer_delack);
2452                         } else {
2453                                 tp->t_flags |= TF_ACKNOW;
2454                         }
2455                         tp->rcv_nxt += tlen;
2456                         thflags = th->th_flags & TH_FIN;
2457                         tcpstat.tcps_rcvpack++;
2458                         tcpstat.tcps_rcvbyte += tlen;
2459                         ND6_HINT(tp);
2460                         if (so->so_state & SS_CANTRCVMORE)
2461                                 m_freem(m);
2462                         else
2463                                 ssb_appendstream(&so->so_rcv, m);
2464                         sorwakeup(so);
2465                 } else {
2466                         if (!(tp->t_flags & TF_DUPSEG)) {
2467                                 /* Initialize SACK report block. */
2468                                 tp->reportblk.rblk_start = th->th_seq;
2469                                 tp->reportblk.rblk_end = th->th_seq + tlen +
2470                                     ((thflags & TH_FIN) != 0);
2471                         }
2472                         thflags = tcp_reass(tp, th, &tlen, m);
2473                         tp->t_flags |= TF_ACKNOW;
2474                 }
2475
2476                 /*
2477                  * Note the amount of data that peer has sent into
2478                  * our window, in order to estimate the sender's
2479                  * buffer size.
2480                  */
2481                 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2482         } else {
2483                 m_freem(m);
2484                 thflags &= ~TH_FIN;
2485         }
2486
2487         /*
2488          * If FIN is received ACK the FIN and let the user know
2489          * that the connection is closing.
2490          */
2491         if (thflags & TH_FIN) {
2492                 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2493                         socantrcvmore(so);
2494                         /*
2495                          * If connection is half-synchronized
2496                          * (ie NEEDSYN flag on) then delay ACK,
2497                          * so it may be piggybacked when SYN is sent.
2498                          * Otherwise, since we received a FIN then no
2499                          * more input can be expected, send ACK now.
2500                          */
2501                         if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2502                                 tcp_callout_reset(tp, tp->tt_delack,
2503                                     tcp_delacktime, tcp_timer_delack);
2504                         } else {
2505                                 tp->t_flags |= TF_ACKNOW;
2506                         }
2507                         tp->rcv_nxt++;
2508                 }
2509
2510                 switch (tp->t_state) {
2511                 /*
2512                  * In SYN_RECEIVED and ESTABLISHED STATES
2513                  * enter the CLOSE_WAIT state.
2514                  */
2515                 case TCPS_SYN_RECEIVED:
2516                         tp->t_starttime = ticks;
2517                         /*FALLTHROUGH*/
2518                 case TCPS_ESTABLISHED:
2519                         tp->t_state = TCPS_CLOSE_WAIT;
2520                         break;
2521
2522                 /*
2523                  * If still in FIN_WAIT_1 STATE FIN has not been acked so
2524                  * enter the CLOSING state.
2525                  */
2526                 case TCPS_FIN_WAIT_1:
2527                         tp->t_state = TCPS_CLOSING;
2528                         break;
2529
2530                 /*
2531                  * In FIN_WAIT_2 state enter the TIME_WAIT state,
2532                  * starting the time-wait timer, turning off the other
2533                  * standard timers.
2534                  */
2535                 case TCPS_FIN_WAIT_2:
2536                         tp->t_state = TCPS_TIME_WAIT;
2537                         tcp_canceltimers(tp);
2538                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2539                                     tcp_timer_2msl);
2540                         soisdisconnected(so);
2541                         break;
2542
2543                 /*
2544                  * In TIME_WAIT state restart the 2 MSL time_wait timer.
2545                  */
2546                 case TCPS_TIME_WAIT:
2547                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2548                             tcp_timer_2msl);
2549                         break;
2550                 }
2551         }
2552
2553 #ifdef TCPDEBUG
2554         if (so->so_options & SO_DEBUG)
2555                 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2556 #endif
2557
2558         /*
2559          * Return any desired output.
2560          */
2561         if (needoutput || (tp->t_flags & TF_ACKNOW))
2562                 tcp_output(tp);
2563         return;
2564
2565 dropafterack:
2566         /*
2567          * Generate an ACK dropping incoming segment if it occupies
2568          * sequence space, where the ACK reflects our state.
2569          *
2570          * We can now skip the test for the RST flag since all
2571          * paths to this code happen after packets containing
2572          * RST have been dropped.
2573          *
2574          * In the SYN-RECEIVED state, don't send an ACK unless the
2575          * segment we received passes the SYN-RECEIVED ACK test.
2576          * If it fails send a RST.  This breaks the loop in the
2577          * "LAND" DoS attack, and also prevents an ACK storm
2578          * between two listening ports that have been sent forged
2579          * SYN segments, each with the source address of the other.
2580          */
2581         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2582             (SEQ_GT(tp->snd_una, th->th_ack) ||
2583              SEQ_GT(th->th_ack, tp->snd_max)) ) {
2584                 rstreason = BANDLIM_RST_OPENPORT;
2585                 goto dropwithreset;
2586         }
2587 #ifdef TCPDEBUG
2588         if (so->so_options & SO_DEBUG)
2589                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2590 #endif
2591         m_freem(m);
2592         tp->t_flags |= TF_ACKNOW;
2593         tcp_output(tp);
2594         return;
2595
2596 dropwithreset:
2597         /*
2598          * Generate a RST, dropping incoming segment.
2599          * Make ACK acceptable to originator of segment.
2600          * Don't bother to respond if destination was broadcast/multicast.
2601          */
2602         if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2603                 goto drop;
2604         if (isipv6) {
2605                 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2606                     IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2607                         goto drop;
2608         } else {
2609                 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2610                     IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2611                     ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2612                     in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2613                         goto drop;
2614         }
2615         /* IPv6 anycast check is done at tcp6_input() */
2616
2617         /*
2618          * Perform bandwidth limiting.
2619          */
2620 #ifdef ICMP_BANDLIM
2621         if (badport_bandlim(rstreason) < 0)
2622                 goto drop;
2623 #endif
2624
2625 #ifdef TCPDEBUG
2626         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2627                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2628 #endif
2629         if (thflags & TH_ACK)
2630                 /* mtod() below is safe as long as hdr dropping is delayed */
2631                 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2632                             TH_RST);
2633         else {
2634                 if (thflags & TH_SYN)
2635                         tlen++;
2636                 /* mtod() below is safe as long as hdr dropping is delayed */
2637                 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2638                             (tcp_seq)0, TH_RST | TH_ACK);
2639         }
2640         return;
2641
2642 drop:
2643         /*
2644          * Drop space held by incoming segment and return.
2645          */
2646 #ifdef TCPDEBUG
2647         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2648                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2649 #endif
2650         m_freem(m);
2651         return;
2652 }
2653
2654 /*
2655  * Parse TCP options and place in tcpopt.
2656  */
2657 static void
2658 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn)
2659 {
2660         int opt, optlen, i;
2661
2662         to->to_flags = 0;
2663         for (; cnt > 0; cnt -= optlen, cp += optlen) {
2664                 opt = cp[0];
2665                 if (opt == TCPOPT_EOL)
2666                         break;
2667                 if (opt == TCPOPT_NOP)
2668                         optlen = 1;
2669                 else {
2670                         if (cnt < 2)
2671                                 break;
2672                         optlen = cp[1];
2673                         if (optlen < 2 || optlen > cnt)
2674                                 break;
2675                 }
2676                 switch (opt) {
2677                 case TCPOPT_MAXSEG:
2678                         if (optlen != TCPOLEN_MAXSEG)
2679                                 continue;
2680                         if (!is_syn)
2681                                 continue;
2682                         to->to_flags |= TOF_MSS;
2683                         bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2684                         to->to_mss = ntohs(to->to_mss);
2685                         break;
2686                 case TCPOPT_WINDOW:
2687                         if (optlen != TCPOLEN_WINDOW)
2688                                 continue;
2689                         if (!is_syn)
2690                                 continue;
2691                         to->to_flags |= TOF_SCALE;
2692                         to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2693                         break;
2694                 case TCPOPT_TIMESTAMP:
2695                         if (optlen != TCPOLEN_TIMESTAMP)
2696                                 continue;
2697                         to->to_flags |= TOF_TS;
2698                         bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2699                         to->to_tsval = ntohl(to->to_tsval);
2700                         bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2701                         to->to_tsecr = ntohl(to->to_tsecr);
2702                         /*
2703                          * If echoed timestamp is later than the current time,
2704                          * fall back to non RFC1323 RTT calculation.
2705                          */
2706                         if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2707                                 to->to_tsecr = 0;
2708                         break;
2709                 case TCPOPT_SACK_PERMITTED:
2710                         if (optlen != TCPOLEN_SACK_PERMITTED)
2711                                 continue;
2712                         if (!is_syn)
2713                                 continue;
2714                         to->to_flags |= TOF_SACK_PERMITTED;
2715                         break;
2716                 case TCPOPT_SACK:
2717                         if ((optlen - 2) & 0x07)        /* not multiple of 8 */
2718                                 continue;
2719                         to->to_nsackblocks = (optlen - 2) / 8;
2720                         to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2721                         to->to_flags |= TOF_SACK;
2722                         for (i = 0; i < to->to_nsackblocks; i++) {
2723                                 struct raw_sackblock *r = &to->to_sackblocks[i];
2724
2725                                 r->rblk_start = ntohl(r->rblk_start);
2726                                 r->rblk_end = ntohl(r->rblk_end);
2727                         }
2728                         break;
2729 #ifdef TCP_SIGNATURE
2730                 /*
2731                  * XXX In order to reply to a host which has set the
2732                  * TCP_SIGNATURE option in its initial SYN, we have to
2733                  * record the fact that the option was observed here
2734                  * for the syncache code to perform the correct response.
2735                  */
2736                 case TCPOPT_SIGNATURE:
2737                         if (optlen != TCPOLEN_SIGNATURE)
2738                                 continue;
2739                         to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2740                         break;
2741 #endif /* TCP_SIGNATURE */
2742                 default:
2743                         continue;
2744                 }
2745         }
2746 }
2747
2748 /*
2749  * Pull out of band byte out of a segment so
2750  * it doesn't appear in the user's data queue.
2751  * It is still reflected in the segment length for
2752  * sequencing purposes.
2753  * "off" is the delayed to be dropped hdrlen.
2754  */
2755 static void
2756 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2757 {
2758         int cnt = off + th->th_urp - 1;
2759
2760         while (cnt >= 0) {
2761                 if (m->m_len > cnt) {
2762                         char *cp = mtod(m, caddr_t) + cnt;
2763                         struct tcpcb *tp = sototcpcb(so);
2764
2765                         tp->t_iobc = *cp;
2766                         tp->t_oobflags |= TCPOOB_HAVEDATA;
2767                         bcopy(cp + 1, cp, m->m_len - cnt - 1);
2768                         m->m_len--;
2769                         if (m->m_flags & M_PKTHDR)
2770                                 m->m_pkthdr.len--;
2771                         return;
2772                 }
2773                 cnt -= m->m_len;
2774                 m = m->m_next;
2775                 if (m == 0)
2776                         break;
2777         }
2778         panic("tcp_pulloutofband");
2779 }
2780
2781 /*
2782  * Collect new round-trip time estimate
2783  * and update averages and current timeout.
2784  */
2785 static void
2786 tcp_xmit_timer(struct tcpcb *tp, int rtt)
2787 {
2788         int delta;
2789
2790         tcpstat.tcps_rttupdated++;
2791         tp->t_rttupdated++;
2792         if (tp->t_srtt != 0) {
2793                 /*
2794                  * srtt is stored as fixed point with 5 bits after the
2795                  * binary point (i.e., scaled by 8).  The following magic
2796                  * is equivalent to the smoothing algorithm in rfc793 with
2797                  * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2798                  * point).  Adjust rtt to origin 0.
2799                  */
2800                 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2801                         - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2802
2803                 if ((tp->t_srtt += delta) <= 0)
2804                         tp->t_srtt = 1;
2805
2806                 /*
2807                  * We accumulate a smoothed rtt variance (actually, a
2808                  * smoothed mean difference), then set the retransmit
2809                  * timer to smoothed rtt + 4 times the smoothed variance.
2810                  * rttvar is stored as fixed point with 4 bits after the
2811                  * binary point (scaled by 16).  The following is
2812                  * equivalent to rfc793 smoothing with an alpha of .75
2813                  * (rttvar = rttvar*3/4 + |delta| / 4).  This replaces
2814                  * rfc793's wired-in beta.
2815                  */
2816                 if (delta < 0)
2817                         delta = -delta;
2818                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2819                 if ((tp->t_rttvar += delta) <= 0)
2820                         tp->t_rttvar = 1;
2821                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2822                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2823         } else {
2824                 /*
2825                  * No rtt measurement yet - use the unsmoothed rtt.
2826                  * Set the variance to half the rtt (so our first
2827                  * retransmit happens at 3*rtt).
2828                  */
2829                 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2830                 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2831                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2832         }
2833         tp->t_rtttime = 0;
2834         tp->t_rxtshift = 0;
2835
2836         /*
2837          * the retransmit should happen at rtt + 4 * rttvar.
2838          * Because of the way we do the smoothing, srtt and rttvar
2839          * will each average +1/2 tick of bias.  When we compute
2840          * the retransmit timer, we want 1/2 tick of rounding and
2841          * 1 extra tick because of +-1/2 tick uncertainty in the
2842          * firing of the timer.  The bias will give us exactly the
2843          * 1.5 tick we need.  But, because the bias is
2844          * statistical, we have to test that we don't drop below
2845          * the minimum feasible timer (which is 2 ticks).
2846          */
2847         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2848                       max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2849
2850         /*
2851          * We received an ack for a packet that wasn't retransmitted;
2852          * it is probably safe to discard any error indications we've
2853          * received recently.  This isn't quite right, but close enough
2854          * for now (a route might have failed after we sent a segment,
2855          * and the return path might not be symmetrical).
2856          */
2857         tp->t_softerror = 0;
2858 }
2859
2860 /*
2861  * Determine a reasonable value for maxseg size.
2862  * If the route is known, check route for mtu.
2863  * If none, use an mss that can be handled on the outgoing
2864  * interface without forcing IP to fragment; if bigger than
2865  * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2866  * to utilize large mbufs.  If no route is found, route has no mtu,
2867  * or the destination isn't local, use a default, hopefully conservative
2868  * size (usually 512 or the default IP max size, but no more than the mtu
2869  * of the interface), as we can't discover anything about intervening
2870  * gateways or networks.  We also initialize the congestion/slow start
2871  * window to be a single segment if the destination isn't local.
2872  * While looking at the routing entry, we also initialize other path-dependent
2873  * parameters from pre-set or cached values in the routing entry.
2874  *
2875  * Also take into account the space needed for options that we
2876  * send regularly.  Make maxseg shorter by that amount to assure
2877  * that we can send maxseg amount of data even when the options
2878  * are present.  Store the upper limit of the length of options plus
2879  * data in maxopd.
2880  *
2881  * NOTE that this routine is only called when we process an incoming
2882  * segment, for outgoing segments only tcp_mssopt is called.
2883  */
2884 void
2885 tcp_mss(struct tcpcb *tp, int offer)
2886 {
2887         struct rtentry *rt;
2888         struct ifnet *ifp;
2889         int rtt, mss;
2890         u_long bufsize;
2891         struct inpcb *inp = tp->t_inpcb;
2892         struct socket *so;
2893 #ifdef INET6
2894         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2895         size_t min_protoh = isipv6 ?
2896                             sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
2897                             sizeof(struct tcpiphdr);
2898 #else
2899         const boolean_t isipv6 = FALSE;
2900         const size_t min_protoh = sizeof(struct tcpiphdr);
2901 #endif
2902
2903         if (isipv6)
2904                 rt = tcp_rtlookup6(&inp->inp_inc);
2905         else
2906                 rt = tcp_rtlookup(&inp->inp_inc);
2907         if (rt == NULL) {
2908                 tp->t_maxopd = tp->t_maxseg =
2909                     (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
2910                 return;
2911         }
2912         ifp = rt->rt_ifp;
2913         so = inp->inp_socket;
2914
2915         /*
2916          * Offer == 0 means that there was no MSS on the SYN segment,
2917          * in this case we use either the interface mtu or tcp_mssdflt.
2918          *
2919          * An offer which is too large will be cut down later.
2920          */
2921         if (offer == 0) {
2922                 if (isipv6) {
2923                         if (in6_localaddr(&inp->in6p_faddr)) {
2924                                 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
2925                                         min_protoh;
2926                         } else {
2927                                 offer = tcp_v6mssdflt;
2928                         }
2929                 } else {
2930                         if (in_localaddr(inp->inp_faddr))
2931                                 offer = ifp->if_mtu - min_protoh;
2932                         else
2933                                 offer = tcp_mssdflt;
2934                 }
2935         }
2936
2937         /*
2938          * Prevent DoS attack with too small MSS. Round up
2939          * to at least minmss.
2940          *
2941          * Sanity check: make sure that maxopd will be large
2942          * enough to allow some data on segments even is the
2943          * all the option space is used (40bytes).  Otherwise
2944          * funny things may happen in tcp_output.
2945          */
2946         offer = max(offer, tcp_minmss);
2947         offer = max(offer, 64);
2948
2949         rt->rt_rmx.rmx_mssopt = offer;
2950
2951         /*
2952          * While we're here, check if there's an initial rtt
2953          * or rttvar.  Convert from the route-table units
2954          * to scaled multiples of the slow timeout timer.
2955          */
2956         if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
2957                 /*
2958                  * XXX the lock bit for RTT indicates that the value
2959                  * is also a minimum value; this is subject to time.
2960                  */
2961                 if (rt->rt_rmx.rmx_locks & RTV_RTT)
2962                         tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
2963                 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
2964                 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
2965                 tcpstat.tcps_usedrtt++;
2966                 if (rt->rt_rmx.rmx_rttvar) {
2967                         tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
2968                             (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
2969                         tcpstat.tcps_usedrttvar++;
2970                 } else {
2971                         /* default variation is +- 1 rtt */
2972                         tp->t_rttvar =
2973                             tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
2974                 }
2975                 TCPT_RANGESET(tp->t_rxtcur,
2976                               ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
2977                               tp->t_rttmin, TCPTV_REXMTMAX);
2978         }
2979
2980         /*
2981          * if there's an mtu associated with the route, use it
2982          * else, use the link mtu.  Take the smaller of mss or offer
2983          * as our final mss.
2984          */
2985         if (rt->rt_rmx.rmx_mtu) {
2986                 mss = rt->rt_rmx.rmx_mtu - min_protoh;
2987         } else {
2988                 if (isipv6)
2989                         mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
2990                 else
2991                         mss = ifp->if_mtu - min_protoh;
2992         }
2993         mss = min(mss, offer);
2994
2995         /*
2996          * maxopd stores the maximum length of data AND options
2997          * in a segment; maxseg is the amount of data in a normal
2998          * segment.  We need to store this value (maxopd) apart
2999          * from maxseg, because now every segment carries options
3000          * and thus we normally have somewhat less data in segments.
3001          */
3002         tp->t_maxopd = mss;
3003
3004         if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
3005             ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3006                 mss -= TCPOLEN_TSTAMP_APPA;
3007
3008 #if     (MCLBYTES & (MCLBYTES - 1)) == 0
3009                 if (mss > MCLBYTES)
3010                         mss &= ~(MCLBYTES-1);
3011 #else
3012                 if (mss > MCLBYTES)
3013                         mss = mss / MCLBYTES * MCLBYTES;
3014 #endif
3015         /*
3016          * If there's a pipesize, change the socket buffer
3017          * to that size.  Make the socket buffers an integral
3018          * number of mss units; if the mss is larger than
3019          * the socket buffer, decrease the mss.
3020          */
3021 #ifdef RTV_SPIPE
3022         if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3023 #endif
3024                 bufsize = so->so_snd.ssb_hiwat;
3025         if (bufsize < mss)
3026                 mss = bufsize;
3027         else {
3028                 bufsize = roundup(bufsize, mss);
3029                 if (bufsize > sb_max)
3030                         bufsize = sb_max;
3031                 if (bufsize > so->so_snd.ssb_hiwat)
3032                         ssb_reserve(&so->so_snd, bufsize, so, NULL);
3033         }
3034         tp->t_maxseg = mss;
3035
3036 #ifdef RTV_RPIPE
3037         if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3038 #endif
3039                 bufsize = so->so_rcv.ssb_hiwat;
3040         if (bufsize > mss) {
3041                 bufsize = roundup(bufsize, mss);
3042                 if (bufsize > sb_max)
3043                         bufsize = sb_max;
3044                 if (bufsize > so->so_rcv.ssb_hiwat)
3045                         ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3046         }
3047
3048         /*
3049          * Set the slow-start flight size depending on whether this
3050          * is a local network or not.
3051          */
3052         if (tcp_do_rfc3390)
3053                 tp->snd_cwnd = min(4 * mss, max(2 * mss, 4380));
3054         else
3055                 tp->snd_cwnd = mss;
3056
3057         if (rt->rt_rmx.rmx_ssthresh) {
3058                 /*
3059                  * There's some sort of gateway or interface
3060                  * buffer limit on the path.  Use this to set
3061                  * the slow start threshhold, but set the
3062                  * threshold to no less than 2*mss.
3063                  */
3064                 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3065                 tcpstat.tcps_usedssthresh++;
3066         }
3067 }
3068
3069 /*
3070  * Determine the MSS option to send on an outgoing SYN.
3071  */
3072 int
3073 tcp_mssopt(struct tcpcb *tp)
3074 {
3075         struct rtentry *rt;
3076 #ifdef INET6
3077         boolean_t isipv6 =
3078             ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3079         int min_protoh = isipv6 ?
3080                              sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3081                              sizeof(struct tcpiphdr);
3082 #else
3083         const boolean_t isipv6 = FALSE;
3084         const size_t min_protoh = sizeof(struct tcpiphdr);
3085 #endif
3086
3087         if (isipv6)
3088                 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3089         else
3090                 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3091         if (rt == NULL)
3092                 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3093
3094         return (rt->rt_ifp->if_mtu - min_protoh);
3095 }
3096
3097 /*
3098  * When a partial ack arrives, force the retransmission of the
3099  * next unacknowledged segment.  Do not exit Fast Recovery.
3100  *
3101  * Implement the Slow-but-Steady variant of NewReno by restarting the
3102  * the retransmission timer.  Turn it off here so it can be restarted
3103  * later in tcp_output().
3104  */
3105 static void
3106 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3107 {
3108         tcp_seq old_snd_nxt = tp->snd_nxt;
3109         u_long ocwnd = tp->snd_cwnd;
3110
3111         tcp_callout_stop(tp, tp->tt_rexmt);
3112         tp->t_rtttime = 0;
3113         tp->snd_nxt = th->th_ack;
3114         /* Set snd_cwnd to one segment beyond acknowledged offset. */
3115         tp->snd_cwnd = tp->t_maxseg;
3116         tp->t_flags |= TF_ACKNOW;
3117         tcp_output(tp);
3118         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3119                 tp->snd_nxt = old_snd_nxt;
3120         /* partial window deflation */
3121         if (ocwnd > acked)
3122                 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3123         else
3124                 tp->snd_cwnd = tp->t_maxseg;
3125 }
3126
3127 /*
3128  * In contrast to the Slow-but-Steady NewReno variant,
3129  * we do not reset the retransmission timer for SACK retransmissions,
3130  * except when retransmitting snd_una.
3131  */
3132 static void
3133 tcp_sack_rexmt(struct tcpcb *tp, struct tcphdr *th)
3134 {
3135         uint32_t pipe, seglen;
3136         tcp_seq nextrexmt;
3137         boolean_t lostdup;
3138         tcp_seq old_snd_nxt = tp->snd_nxt;
3139         u_long ocwnd = tp->snd_cwnd;
3140         int nseg = 0;           /* consecutive new segments */
3141 #define MAXBURST 4              /* limit burst of new packets on partial ack */
3142
3143         tp->t_rtttime = 0;
3144         pipe = tcp_sack_compute_pipe(tp);
3145         while ((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg &&
3146             (!tcp_do_smartsack || nseg < MAXBURST) &&
3147             tcp_sack_nextseg(tp, &nextrexmt, &seglen, &lostdup)) {
3148                 uint32_t sent;
3149                 tcp_seq old_snd_max;
3150                 int error;
3151
3152                 if (nextrexmt == tp->snd_max)
3153                         ++nseg;
3154                 tp->snd_nxt = nextrexmt;
3155                 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3156                 old_snd_max = tp->snd_max;
3157                 if (nextrexmt == tp->snd_una)
3158                         tcp_callout_stop(tp, tp->tt_rexmt);
3159                 error = tcp_output(tp);
3160                 if (error != 0)
3161                         break;
3162                 sent = tp->snd_nxt - nextrexmt;
3163                 if (sent <= 0)
3164                         break;
3165                 if (!lostdup)
3166                         pipe += sent;
3167                 tcpstat.tcps_sndsackpack++;
3168                 tcpstat.tcps_sndsackbyte += sent;
3169                 if (SEQ_LT(nextrexmt, old_snd_max) &&
3170                     SEQ_LT(tp->rexmt_high, tp->snd_nxt))
3171                         tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3172         }
3173         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3174                 tp->snd_nxt = old_snd_nxt;
3175         tp->snd_cwnd = ocwnd;
3176 }