38112c22fb7fc713ac8b08e3662e8ef0494d9d3c
[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  */
69
70 #include "opt_inet.h"
71 #include "opt_inet6.h"
72 #include "opt_ipsec.h"
73 #include "opt_tcpdebug.h"
74 #include "opt_tcp_input.h"
75
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
80 #include <sys/malloc.h>
81 #include <sys/mbuf.h>
82 #include <sys/proc.h>           /* for proc0 declaration */
83 #include <sys/protosw.h>
84 #include <sys/socket.h>
85 #include <sys/socketvar.h>
86 #include <sys/syslog.h>
87 #include <sys/in_cksum.h>
88
89 #include <sys/socketvar2.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 static int tcp_do_eifel_detect = 1;
171 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
172     &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
173
174 static int tcp_do_abc = 1;
175 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
176     &tcp_do_abc, 0,
177     "TCP Appropriate Byte Counting (RFC 3465)");
178
179 /*
180  * The following value actually takes range [25ms, 250ms],
181  * given that most modern systems use 1ms ~ 10ms as the unit
182  * of timestamp option.
183  */
184 static u_int tcp_paws_tolerance = 25;
185 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, paws_tolerance, CTLFLAG_RW,
186     &tcp_paws_tolerance, 0, "RFC1323 PAWS tolerance");
187
188 /*
189  * Define as tunable for easy testing with SACK on and off.
190  * Warning:  do not change setting in the middle of an existing active TCP flow,
191  *   else strange things might happen to that flow.
192  */
193 int tcp_do_sack = 1;
194 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
195     &tcp_do_sack, 0, "Enable SACK Algorithms");
196
197 int tcp_do_smartsack = 1;
198 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
199     &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
200
201 int tcp_do_rescuesack = 1;
202 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack, CTLFLAG_RW,
203     &tcp_do_rescuesack, 0, "Rescue retransmission for SACK");
204
205 int tcp_aggressive_rescuesack = 0;
206 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack_agg, CTLFLAG_RW,
207     &tcp_aggressive_rescuesack, 0, "Aggressive rescue retransmission for SACK");
208
209 int tcp_do_rfc3517bis = 0;
210 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3517bis, CTLFLAG_RW,
211     &tcp_do_rfc3517bis, 0, "Enable RFC3517 update");
212
213 int tcp_rfc3517bis_rxt = 0;
214 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3517bis_rxt, CTLFLAG_RW,
215     &tcp_rfc3517bis_rxt, 0, "Enable RFC3517 retransmit update");
216
217 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
218     "TCP Segment Reassembly Queue");
219
220 int tcp_reass_maxseg = 0;
221 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
222     &tcp_reass_maxseg, 0,
223     "Global maximum number of TCP Segments in Reassembly Queue");
224
225 int tcp_reass_qsize = 0;
226 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
227     &tcp_reass_qsize, 0,
228     "Global number of TCP Segments currently in Reassembly Queue");
229
230 static int tcp_reass_overflows = 0;
231 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
232     &tcp_reass_overflows, 0,
233     "Global number of TCP Segment Reassembly Queue Overflows");
234
235 int tcp_do_autorcvbuf = 1;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
237     &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
238
239 int tcp_autorcvbuf_inc = 16*1024;
240 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
241     &tcp_autorcvbuf_inc, 0,
242     "Incrementor step size of automatic receive buffer");
243
244 int tcp_autorcvbuf_max = 2*1024*1024;
245 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
246     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
247
248 int tcp_sosend_agglim = 2;
249 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_agglim, CTLFLAG_RW,
250     &tcp_sosend_agglim, 0, "TCP sosend mbuf aggregation limit");
251
252 int tcp_sosend_async = 1;
253 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_async, CTLFLAG_RW,
254     &tcp_sosend_async, 0, "TCP asynchronized pru_send");
255
256 static int tcp_ignore_redun_dsack = 1;
257 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_redun_dsack, CTLFLAG_RW,
258     &tcp_ignore_redun_dsack, 0, "Ignore redundant DSACK");
259
260 static void      tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t,
261                     tcp_seq);
262 static void      tcp_pulloutofband(struct socket *,
263                      struct tcphdr *, struct mbuf *, int);
264 static int       tcp_reass(struct tcpcb *, struct tcphdr *, int *,
265                      struct mbuf *);
266 static void      tcp_xmit_timer(struct tcpcb *, int, tcp_seq);
267 static void      tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
268 static void      tcp_sack_rexmt(struct tcpcb *);
269 static boolean_t tcp_sack_limitedxmit(struct tcpcb *);
270 static int       tcp_rmx_msl(const struct tcpcb *);
271 static void      tcp_established(struct tcpcb *);
272 static boolean_t tcp_fast_recovery(struct tcpcb *, tcp_seq,
273                      const struct tcpopt *);
274
275 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
276 #ifdef INET6
277 #define ND6_HINT(tp) \
278 do { \
279         if ((tp) && (tp)->t_inpcb && \
280             ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
281             (tp)->t_inpcb->in6p_route.ro_rt) \
282                 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
283 } while (0)
284 #else
285 #define ND6_HINT(tp)
286 #endif
287
288 /*
289  * Indicate whether this ack should be delayed.  We can delay the ack if
290  *      - delayed acks are enabled and
291  *      - there is no delayed ack timer in progress and
292  *      - our last ack wasn't a 0-sized window.  We never want to delay
293  *        the ack that opens up a 0-sized window.
294  */
295 #define DELAY_ACK(tp) \
296         (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
297         !(tp->t_flags & TF_RXWIN0SENT))
298
299 #define acceptable_window_update(tp, th, tiwin)                         \
300     (SEQ_LT(tp->snd_wl1, th->th_seq) ||                                 \
301      (tp->snd_wl1 == th->th_seq &&                                      \
302       (SEQ_LT(tp->snd_wl2, th->th_ack) ||                               \
303        (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
304
305 #define iceildiv(n, d)          (((n)+(d)-1) / (d))
306 #define need_early_retransmit(tp, ownd) \
307     (tcp_do_early_retransmit && \
308      (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) && \
309      ownd < ((tp->t_rxtthresh + 1) * tp->t_maxseg) && \
310      tp->t_dupacks + 1 >= iceildiv(ownd, tp->t_maxseg) && \
311      (!TCP_DO_SACK(tp) || ownd <= tp->t_maxseg || \
312       tcp_sack_has_sacked(&tp->scb, ownd - tp->t_maxseg)))
313
314 /*
315  * Returns TRUE, if this segment can be merged with the last
316  * pending segment in the reassemble queue and this segment
317  * does not overlap with the pending segment immediately
318  * preceeding the last pending segment.
319  */
320 static __inline boolean_t
321 tcp_paws_canreasslast(const struct tcpcb *tp, const struct tcphdr *th, int tlen)
322 {
323         const struct tseg_qent *last, *prev;
324
325         last = TAILQ_LAST(&tp->t_segq, tsegqe_head);
326         if (last == NULL)
327                 return FALSE;
328
329         /* This segment comes immediately after the last pending segment */
330         if (last->tqe_th->th_seq + last->tqe_len == th->th_seq)
331                 return TRUE;
332
333         if (th->th_seq + tlen != last->tqe_th->th_seq)
334                 return FALSE;
335         /* This segment comes immediately before the last pending segment */
336
337         prev = TAILQ_PREV(last, tsegqe_head, tqe_q);
338         if (prev == NULL) {
339                 /*
340                  * No pending preceeding segment, we assume this segment
341                  * could be reassembled.
342                  */
343                 return TRUE;
344         }
345
346         /* This segment does not overlap with the preceeding segment */
347         if (SEQ_GEQ(th->th_seq, prev->tqe_th->th_seq + prev->tqe_len))
348                 return TRUE;
349
350         return FALSE;
351 }
352
353 static __inline void
354 tcp_ncr_update_rxtthresh(struct tcpcb *tp)
355 {
356         int old_rxtthresh = tp->t_rxtthresh;
357         uint32_t ownd = tp->snd_max - tp->snd_una;
358
359         tp->t_rxtthresh = max(3, ((ownd / tp->t_maxseg) >> 1));
360         if (tp->t_rxtthresh != old_rxtthresh) {
361                 tcp_sack_update_lostseq(&tp->scb, tp->snd_una,
362                     tp->t_maxseg, tp->t_rxtthresh);
363         }
364 }
365
366 static int
367 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
368 {
369         struct tseg_qent *q;
370         struct tseg_qent *p = NULL;
371         struct tseg_qent *te;
372         struct socket *so = tp->t_inpcb->inp_socket;
373         int flags;
374
375         /*
376          * Call with th == NULL after become established to
377          * force pre-ESTABLISHED data up to user socket.
378          */
379         if (th == NULL)
380                 goto present;
381
382         /*
383          * Limit the number of segments in the reassembly queue to prevent
384          * holding on to too many segments (and thus running out of mbufs).
385          * Make sure to let the missing segment through which caused this
386          * queue.  Always keep one global queue entry spare to be able to
387          * process the missing segment.
388          */
389         if (th->th_seq != tp->rcv_nxt &&
390             tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
391                 tcp_reass_overflows++;
392                 tcpstat.tcps_rcvmemdrop++;
393                 m_freem(m);
394                 /* no SACK block to report */
395                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
396                 return (0);
397         }
398
399         /* Allocate a new queue entry. */
400         te = kmalloc(sizeof(struct tseg_qent), M_TSEGQ, M_INTWAIT | M_NULLOK);
401         if (te == NULL) {
402                 tcpstat.tcps_rcvmemdrop++;
403                 m_freem(m);
404                 /* no SACK block to report */
405                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
406                 return (0);
407         }
408         atomic_add_int(&tcp_reass_qsize, 1);
409
410         /*
411          * Find a segment which begins after this one does.
412          */
413         TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
414                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
415                         break;
416                 p = q;
417         }
418
419         /*
420          * If there is a preceding segment, it may provide some of
421          * our data already.  If so, drop the data from the incoming
422          * segment.  If it provides all of our data, drop us.
423          */
424         if (p != NULL) {
425                 tcp_seq_diff_t i;
426
427                 /* conversion to int (in i) handles seq wraparound */
428                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
429                 if (i > 0) {            /* overlaps preceding segment */
430                         tp->sack_flags |=
431                             (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
432                         /* enclosing block starts w/ preceding segment */
433                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
434                         if (i >= *tlenp) {
435                                 /* preceding encloses incoming segment */
436                                 tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
437                                     p->tqe_th->th_seq + p->tqe_len,
438                                     p->tqe_th->th_flags);
439                                 tcpstat.tcps_rcvduppack++;
440                                 tcpstat.tcps_rcvdupbyte += *tlenp;
441                                 m_freem(m);
442                                 kfree(te, M_TSEGQ);
443                                 atomic_add_int(&tcp_reass_qsize, -1);
444                                 /*
445                                  * Try to present any queued data
446                                  * at the left window edge to the user.
447                                  * This is needed after the 3-WHS
448                                  * completes.
449                                  */
450                                 goto present;   /* ??? */
451                         }
452                         m_adj(m, i);
453                         *tlenp -= i;
454                         th->th_seq += i;
455                         /* incoming segment end is enclosing block end */
456                         tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
457                             th->th_seq + *tlenp, th->th_flags);
458                         /* trim end of reported D-SACK block */
459                         tp->reportblk.rblk_end = th->th_seq;
460                 }
461         }
462         tcpstat.tcps_rcvoopack++;
463         tcpstat.tcps_rcvoobyte += *tlenp;
464
465         /*
466          * While we overlap succeeding segments trim them or,
467          * if they are completely covered, dequeue them.
468          */
469         while (q) {
470                 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
471                 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
472                 tcp_seq qend_sack = TCP_SACK_BLKEND(qend, q->tqe_th->th_flags);
473                 struct tseg_qent *nq;
474
475                 if (i <= 0)
476                         break;
477                 if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
478                         /* first time through */
479                         tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
480                         tp->encloseblk = tp->reportblk;
481                         /* report trailing duplicate D-SACK segment */
482                         tp->reportblk.rblk_start = q->tqe_th->th_seq;
483                 }
484                 if ((tp->sack_flags & TSACK_F_ENCLOSESEG) &&
485                     SEQ_GT(qend_sack, tp->encloseblk.rblk_end)) {
486                         /* extend enclosing block if one exists */
487                         tp->encloseblk.rblk_end = qend_sack;
488                 }
489                 if (i < q->tqe_len) {
490                         q->tqe_th->th_seq += i;
491                         q->tqe_len -= i;
492                         m_adj(q->tqe_m, i);
493                         break;
494                 }
495
496                 nq = TAILQ_NEXT(q, tqe_q);
497                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
498                 m_freem(q->tqe_m);
499                 kfree(q, M_TSEGQ);
500                 atomic_add_int(&tcp_reass_qsize, -1);
501                 q = nq;
502         }
503
504         /* Insert the new segment queue entry into place. */
505         te->tqe_m = m;
506         te->tqe_th = th;
507         te->tqe_len = *tlenp;
508
509         /* check if can coalesce with following segment */
510         if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
511                 tcp_seq tend = te->tqe_th->th_seq + te->tqe_len;
512                 tcp_seq tend_sack = TCP_SACK_BLKEND(tend, te->tqe_th->th_flags);
513
514                 te->tqe_len += q->tqe_len;
515                 if (q->tqe_th->th_flags & TH_FIN)
516                         te->tqe_th->th_flags |= TH_FIN;
517                 m_cat(te->tqe_m, q->tqe_m);
518                 tp->encloseblk.rblk_end = tend_sack;
519                 /*
520                  * When not reporting a duplicate segment, use
521                  * the larger enclosing block as the SACK block.
522                  */
523                 if (!(tp->sack_flags & TSACK_F_DUPSEG))
524                         tp->reportblk.rblk_end = tend_sack;
525                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
526                 kfree(q, M_TSEGQ);
527                 atomic_add_int(&tcp_reass_qsize, -1);
528         }
529
530         if (p == NULL) {
531                 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
532         } else {
533                 /* check if can coalesce with preceding segment */
534                 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
535                         p->tqe_len += te->tqe_len;
536                         m_cat(p->tqe_m, te->tqe_m);
537                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
538                         /*
539                          * When not reporting a duplicate segment, use
540                          * the larger enclosing block as the SACK block.
541                          */
542                         if (!(tp->sack_flags & TSACK_F_DUPSEG))
543                                 tp->reportblk.rblk_start = p->tqe_th->th_seq;
544                         kfree(te, M_TSEGQ);
545                         atomic_add_int(&tcp_reass_qsize, -1);
546                 } else {
547                         TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
548                 }
549         }
550
551 present:
552         /*
553          * Present data to user, advancing rcv_nxt through
554          * completed sequence space.
555          */
556         if (!TCPS_HAVEESTABLISHED(tp->t_state))
557                 return (0);
558         q = TAILQ_FIRST(&tp->t_segq);
559         if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
560                 return (0);
561         tp->rcv_nxt += q->tqe_len;
562         if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
563                 /* no SACK block to report since ACK advanced */
564                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
565         }
566         /* no enclosing block to report since ACK advanced */
567         tp->sack_flags &= ~TSACK_F_ENCLOSESEG;
568         flags = q->tqe_th->th_flags & TH_FIN;
569         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
570         KASSERT(TAILQ_EMPTY(&tp->t_segq) ||
571                 TAILQ_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
572                 ("segment not coalesced"));
573         if (so->so_state & SS_CANTRCVMORE) {
574                 m_freem(q->tqe_m);
575         } else {
576                 lwkt_gettoken(&so->so_rcv.ssb_token);
577                 ssb_appendstream(&so->so_rcv, q->tqe_m);
578                 lwkt_reltoken(&so->so_rcv.ssb_token);
579         }
580         kfree(q, M_TSEGQ);
581         atomic_add_int(&tcp_reass_qsize, -1);
582         ND6_HINT(tp);
583         sorwakeup(so);
584         return (flags);
585 }
586
587 /*
588  * TCP input routine, follows pages 65-76 of the
589  * protocol specification dated September, 1981 very closely.
590  */
591 #ifdef INET6
592 int
593 tcp6_input(struct mbuf **mp, int *offp, int proto)
594 {
595         struct mbuf *m = *mp;
596         struct in6_ifaddr *ia6;
597
598         IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
599
600         /*
601          * draft-itojun-ipv6-tcp-to-anycast
602          * better place to put this in?
603          */
604         ia6 = ip6_getdstifaddr(m);
605         if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
606                 struct ip6_hdr *ip6;
607
608                 ip6 = mtod(m, struct ip6_hdr *);
609                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
610                             offsetof(struct ip6_hdr, ip6_dst));
611                 return (IPPROTO_DONE);
612         }
613
614         tcp_input(mp, offp, proto);
615         return (IPPROTO_DONE);
616 }
617 #endif
618
619 int
620 tcp_input(struct mbuf **mp, int *offp, int proto)
621 {
622         int off0;
623         struct tcphdr *th;
624         struct ip *ip = NULL;
625         struct ipovly *ipov;
626         struct inpcb *inp = NULL;
627         u_char *optp = NULL;
628         int optlen = 0;
629         int tlen, off;
630         int len = 0;
631         int drop_hdrlen;
632         struct tcpcb *tp = NULL;
633         int thflags;
634         struct socket *so = NULL;
635         int todrop, acked;
636         boolean_t ourfinisacked, needoutput = FALSE, delayed_dupack = FALSE;
637         tcp_seq th_dupack = 0; /* XXX gcc warning */
638         u_long tiwin;
639         int recvwin;
640         struct tcpopt to;               /* options in this segment */
641         struct sockaddr_in *next_hop = NULL;
642         int rstreason; /* For badport_bandlim accounting purposes */
643         int cpu;
644         struct ip6_hdr *ip6 = NULL;
645         struct mbuf *m;
646 #ifdef INET6
647         boolean_t isipv6;
648 #else
649         const boolean_t isipv6 = FALSE;
650 #endif
651 #ifdef TCPDEBUG
652         short ostate = 0;
653 #endif
654
655         off0 = *offp;
656         m = *mp;
657         *mp = NULL;
658
659         tcpstat.tcps_rcvtotal++;
660
661         if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
662                 struct m_tag *mtag;
663
664                 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
665                 KKASSERT(mtag != NULL);
666                 next_hop = m_tag_data(mtag);
667         }
668
669 #ifdef INET6
670         isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
671 #endif
672
673         if (isipv6) {
674                 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
675                 ip6 = mtod(m, struct ip6_hdr *);
676                 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
677                 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
678                         tcpstat.tcps_rcvbadsum++;
679                         goto drop;
680                 }
681                 th = (struct tcphdr *)((caddr_t)ip6 + off0);
682
683                 /*
684                  * Be proactive about unspecified IPv6 address in source.
685                  * As we use all-zero to indicate unbounded/unconnected pcb,
686                  * unspecified IPv6 address can be used to confuse us.
687                  *
688                  * Note that packets with unspecified IPv6 destination is
689                  * already dropped in ip6_input.
690                  */
691                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
692                         /* XXX stat */
693                         goto drop;
694                 }
695         } else {
696                 /*
697                  * Get IP and TCP header together in first mbuf.
698                  * Note: IP leaves IP header in first mbuf.
699                  */
700                 if (off0 > sizeof(struct ip)) {
701                         ip_stripoptions(m);
702                         off0 = sizeof(struct ip);
703                 }
704                 /* already checked and pulled up in ip_demux() */
705                 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
706                     ("TCP header not in one mbuf: m->m_len %d", m->m_len));
707                 ip = mtod(m, struct ip *);
708                 ipov = (struct ipovly *)ip;
709                 th = (struct tcphdr *)((caddr_t)ip + off0);
710                 tlen = ip->ip_len;
711
712                 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
713                         if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
714                                 th->th_sum = m->m_pkthdr.csum_data;
715                         else
716                                 th->th_sum = in_pseudo(ip->ip_src.s_addr,
717                                                 ip->ip_dst.s_addr,
718                                                 htonl(m->m_pkthdr.csum_data +
719                                                         ip->ip_len +
720                                                         IPPROTO_TCP));
721                         th->th_sum ^= 0xffff;
722                 } else {
723                         /*
724                          * Checksum extended TCP header and data.
725                          */
726                         len = sizeof(struct ip) + tlen;
727                         bzero(ipov->ih_x1, sizeof ipov->ih_x1);
728                         ipov->ih_len = (u_short)tlen;
729                         ipov->ih_len = htons(ipov->ih_len);
730                         th->th_sum = in_cksum(m, len);
731                 }
732                 if (th->th_sum) {
733                         tcpstat.tcps_rcvbadsum++;
734                         goto drop;
735                 }
736 #ifdef INET6
737                 /* Re-initialization for later version check */
738                 ip->ip_v = IPVERSION;
739 #endif
740         }
741
742         /*
743          * Check that TCP offset makes sense,
744          * pull out TCP options and adjust length.              XXX
745          */
746         off = th->th_off << 2;
747         /* already checked and pulled up in ip_demux() */
748         KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
749             ("bad TCP data offset %d (tlen %d)", off, tlen));
750         tlen -= off;    /* tlen is used instead of ti->ti_len */
751         if (off > sizeof(struct tcphdr)) {
752                 if (isipv6) {
753                         IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
754                         ip6 = mtod(m, struct ip6_hdr *);
755                         th = (struct tcphdr *)((caddr_t)ip6 + off0);
756                 } else {
757                         /* already pulled up in ip_demux() */
758                         KASSERT(m->m_len >= sizeof(struct ip) + off,
759                             ("TCP header and options not in one mbuf: "
760                              "m_len %d, off %d", m->m_len, off));
761                 }
762                 optlen = off - sizeof(struct tcphdr);
763                 optp = (u_char *)(th + 1);
764         }
765         thflags = th->th_flags;
766
767 #ifdef TCP_DROP_SYNFIN
768         /*
769          * If the drop_synfin option is enabled, drop all packets with
770          * both the SYN and FIN bits set. This prevents e.g. nmap from
771          * identifying the TCP/IP stack.
772          *
773          * This is a violation of the TCP specification.
774          */
775         if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
776                 goto drop;
777 #endif
778
779         /*
780          * Convert TCP protocol specific fields to host format.
781          */
782         th->th_seq = ntohl(th->th_seq);
783         th->th_ack = ntohl(th->th_ack);
784         th->th_win = ntohs(th->th_win);
785         th->th_urp = ntohs(th->th_urp);
786
787         /*
788          * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
789          * until after ip6_savecontrol() is called and before other functions
790          * which don't want those proto headers.
791          * Because ip6_savecontrol() is going to parse the mbuf to
792          * search for data to be passed up to user-land, it wants mbuf
793          * parameters to be unchanged.
794          * XXX: the call of ip6_savecontrol() has been obsoleted based on
795          * latest version of the advanced API (20020110).
796          */
797         drop_hdrlen = off0 + off;
798
799         /*
800          * Locate pcb for segment.
801          */
802 findpcb:
803         /* IPFIREWALL_FORWARD section */
804         if (next_hop != NULL && !isipv6) {  /* IPv6 support is not there yet */
805                 /*
806                  * Transparently forwarded. Pretend to be the destination.
807                  * already got one like this?
808                  */
809                 cpu = mycpu->gd_cpuid;
810                 inp = in_pcblookup_hash(&tcbinfo[cpu],
811                                         ip->ip_src, th->th_sport,
812                                         ip->ip_dst, th->th_dport,
813                                         0, m->m_pkthdr.rcvif);
814                 if (!inp) {
815                         /*
816                          * It's new.  Try to find the ambushing socket.
817                          */
818
819                         /*
820                          * The rest of the ipfw code stores the port in
821                          * host order.  XXX
822                          * (The IP address is still in network order.)
823                          */
824                         in_port_t dport = next_hop->sin_port ?
825                                                 htons(next_hop->sin_port) :
826                                                 th->th_dport;
827
828                         cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
829                                           next_hop->sin_addr.s_addr, dport);
830                         inp = in_pcblookup_hash(&tcbinfo[cpu],
831                                                 ip->ip_src, th->th_sport,
832                                                 next_hop->sin_addr, dport,
833                                                 1, m->m_pkthdr.rcvif);
834                 }
835         } else {
836                 if (isipv6) {
837                         inp = in6_pcblookup_hash(&tcbinfo[0],
838                                                  &ip6->ip6_src, th->th_sport,
839                                                  &ip6->ip6_dst, th->th_dport,
840                                                  1, m->m_pkthdr.rcvif);
841                 } else {
842                         cpu = mycpu->gd_cpuid;
843                         inp = in_pcblookup_hash(&tcbinfo[cpu],
844                                                 ip->ip_src, th->th_sport,
845                                                 ip->ip_dst, th->th_dport,
846                                                 1, m->m_pkthdr.rcvif);
847                 }
848         }
849
850         /*
851          * If the state is CLOSED (i.e., TCB does not exist) then
852          * all data in the incoming segment is discarded.
853          * If the TCB exists but is in CLOSED state, it is embryonic,
854          * but should either do a listen or a connect soon.
855          */
856         if (inp == NULL) {
857                 if (log_in_vain) {
858 #ifdef INET6
859                         char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
860 #else
861                         char dbuf[sizeof "aaa.bbb.ccc.ddd"];
862                         char sbuf[sizeof "aaa.bbb.ccc.ddd"];
863 #endif
864                         if (isipv6) {
865                                 strcpy(dbuf, "[");
866                                 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
867                                 strcat(dbuf, "]");
868                                 strcpy(sbuf, "[");
869                                 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
870                                 strcat(sbuf, "]");
871                         } else {
872                                 strcpy(dbuf, inet_ntoa(ip->ip_dst));
873                                 strcpy(sbuf, inet_ntoa(ip->ip_src));
874                         }
875                         switch (log_in_vain) {
876                         case 1:
877                                 if (!(thflags & TH_SYN))
878                                         break;
879                         case 2:
880                                 log(LOG_INFO,
881                                     "Connection attempt to TCP %s:%d "
882                                     "from %s:%d flags:0x%02x\n",
883                                     dbuf, ntohs(th->th_dport), sbuf,
884                                     ntohs(th->th_sport), thflags);
885                                 break;
886                         default:
887                                 break;
888                         }
889                 }
890                 if (blackhole) {
891                         switch (blackhole) {
892                         case 1:
893                                 if (thflags & TH_SYN)
894                                         goto drop;
895                                 break;
896                         case 2:
897                                 goto drop;
898                         default:
899                                 goto drop;
900                         }
901                 }
902                 rstreason = BANDLIM_RST_CLOSEDPORT;
903                 goto dropwithreset;
904         }
905
906 #ifdef IPSEC
907         if (isipv6) {
908                 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
909                         ipsec6stat.in_polvio++;
910                         goto drop;
911                 }
912         } else {
913                 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
914                         ipsecstat.in_polvio++;
915                         goto drop;
916                 }
917         }
918 #endif
919 #ifdef FAST_IPSEC
920         if (isipv6) {
921                 if (ipsec6_in_reject(m, inp))
922                         goto drop;
923         } else {
924                 if (ipsec4_in_reject(m, inp))
925                         goto drop;
926         }
927 #endif
928         /* Check the minimum TTL for socket. */
929 #ifdef INET6
930         if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
931                 goto drop;
932 #endif
933
934         tp = intotcpcb(inp);
935         if (tp == NULL) {
936                 rstreason = BANDLIM_RST_CLOSEDPORT;
937                 goto dropwithreset;
938         }
939         if (tp->t_state <= TCPS_CLOSED)
940                 goto drop;
941
942         so = inp->inp_socket;
943
944 #ifdef TCPDEBUG
945         if (so->so_options & SO_DEBUG) {
946                 ostate = tp->t_state;
947                 if (isipv6)
948                         bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
949                 else
950                         bcopy(ip, tcp_saveipgen, sizeof(*ip));
951                 tcp_savetcp = *th;
952         }
953 #endif
954
955         bzero(&to, sizeof to);
956
957         if (so->so_options & SO_ACCEPTCONN) {
958                 struct in_conninfo inc;
959
960 #ifdef INET6
961                 inc.inc_isipv6 = (isipv6 == TRUE);
962 #endif
963                 if (isipv6) {
964                         inc.inc6_faddr = ip6->ip6_src;
965                         inc.inc6_laddr = ip6->ip6_dst;
966                         inc.inc6_route.ro_rt = NULL;            /* XXX */
967                 } else {
968                         inc.inc_faddr = ip->ip_src;
969                         inc.inc_laddr = ip->ip_dst;
970                         inc.inc_route.ro_rt = NULL;             /* XXX */
971                 }
972                 inc.inc_fport = th->th_sport;
973                 inc.inc_lport = th->th_dport;
974
975                 /*
976                  * If the state is LISTEN then ignore segment if it contains
977                  * a RST.  If the segment contains an ACK then it is bad and
978                  * send a RST.  If it does not contain a SYN then it is not
979                  * interesting; drop it.
980                  *
981                  * If the state is SYN_RECEIVED (syncache) and seg contains
982                  * an ACK, but not for our SYN/ACK, send a RST.  If the seg
983                  * contains a RST, check the sequence number to see if it
984                  * is a valid reset segment.
985                  */
986                 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
987                         if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
988                                 if (!syncache_expand(&inc, th, &so, m)) {
989                                         /*
990                                          * No syncache entry, or ACK was not
991                                          * for our SYN/ACK.  Send a RST.
992                                          */
993                                         tcpstat.tcps_badsyn++;
994                                         rstreason = BANDLIM_RST_OPENPORT;
995                                         goto dropwithreset;
996                                 }
997
998                                 /*
999                                  * Could not complete 3-way handshake,
1000                                  * connection is being closed down, and
1001                                  * syncache will free mbuf.
1002                                  */
1003                                 if (so == NULL)
1004                                         return(IPPROTO_DONE);
1005
1006                                 /*
1007                                  * We must be in the correct protocol thread
1008                                  * for this connection.
1009                                  */
1010                                 KKASSERT(so->so_port == &curthread->td_msgport);
1011
1012                                 /*
1013                                  * Socket is created in state SYN_RECEIVED.
1014                                  * Continue processing segment.
1015                                  */
1016                                 inp = so->so_pcb;
1017                                 tp = intotcpcb(inp);
1018                                 /*
1019                                  * This is what would have happened in
1020                                  * tcp_output() when the SYN,ACK was sent.
1021                                  */
1022                                 tp->snd_up = tp->snd_una;
1023                                 tp->snd_max = tp->snd_nxt = tp->iss + 1;
1024                                 tp->last_ack_sent = tp->rcv_nxt;
1025
1026                                 goto after_listen;
1027                         }
1028                         if (thflags & TH_RST) {
1029                                 syncache_chkrst(&inc, th);
1030                                 goto drop;
1031                         }
1032                         if (thflags & TH_ACK) {
1033                                 syncache_badack(&inc);
1034                                 tcpstat.tcps_badsyn++;
1035                                 rstreason = BANDLIM_RST_OPENPORT;
1036                                 goto dropwithreset;
1037                         }
1038                         goto drop;
1039                 }
1040
1041                 /*
1042                  * Segment's flags are (SYN) or (SYN | FIN).
1043                  */
1044 #ifdef INET6
1045                 /*
1046                  * If deprecated address is forbidden,
1047                  * we do not accept SYN to deprecated interface
1048                  * address to prevent any new inbound connection from
1049                  * getting established.
1050                  * When we do not accept SYN, we send a TCP RST,
1051                  * with deprecated source address (instead of dropping
1052                  * it).  We compromise it as it is much better for peer
1053                  * to send a RST, and RST will be the final packet
1054                  * for the exchange.
1055                  *
1056                  * If we do not forbid deprecated addresses, we accept
1057                  * the SYN packet.  RFC2462 does not suggest dropping
1058                  * SYN in this case.
1059                  * If we decipher RFC2462 5.5.4, it says like this:
1060                  * 1. use of deprecated addr with existing
1061                  *    communication is okay - "SHOULD continue to be
1062                  *    used"
1063                  * 2. use of it with new communication:
1064                  *   (2a) "SHOULD NOT be used if alternate address
1065                  *        with sufficient scope is available"
1066                  *   (2b) nothing mentioned otherwise.
1067                  * Here we fall into (2b) case as we have no choice in
1068                  * our source address selection - we must obey the peer.
1069                  *
1070                  * The wording in RFC2462 is confusing, and there are
1071                  * multiple description text for deprecated address
1072                  * handling - worse, they are not exactly the same.
1073                  * I believe 5.5.4 is the best one, so we follow 5.5.4.
1074                  */
1075                 if (isipv6 && !ip6_use_deprecated) {
1076                         struct in6_ifaddr *ia6;
1077
1078                         if ((ia6 = ip6_getdstifaddr(m)) &&
1079                             (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
1080                                 tp = NULL;
1081                                 rstreason = BANDLIM_RST_OPENPORT;
1082                                 goto dropwithreset;
1083                         }
1084                 }
1085 #endif
1086                 /*
1087                  * If it is from this socket, drop it, it must be forged.
1088                  * Don't bother responding if the destination was a broadcast.
1089                  */
1090                 if (th->th_dport == th->th_sport) {
1091                         if (isipv6) {
1092                                 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1093                                                        &ip6->ip6_src))
1094                                         goto drop;
1095                         } else {
1096                                 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1097                                         goto drop;
1098                         }
1099                 }
1100                 /*
1101                  * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1102                  *
1103                  * Note that it is quite possible to receive unicast
1104                  * link-layer packets with a broadcast IP address. Use
1105                  * in_broadcast() to find them.
1106                  */
1107                 if (m->m_flags & (M_BCAST | M_MCAST))
1108                         goto drop;
1109                 if (isipv6) {
1110                         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1111                             IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1112                                 goto drop;
1113                 } else {
1114                         if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1115                             IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1116                             ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1117                             in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1118                                 goto drop;
1119                 }
1120                 /*
1121                  * SYN appears to be valid; create compressed TCP state
1122                  * for syncache, or perform t/tcp connection.
1123                  */
1124                 if (so->so_qlen <= so->so_qlimit) {
1125                         tcp_dooptions(&to, optp, optlen, TRUE, th->th_ack);
1126                         if (!syncache_add(&inc, &to, th, so, m))
1127                                 goto drop;
1128
1129                         /*
1130                          * Entry added to syncache, mbuf used to
1131                          * send SYN,ACK packet.
1132                          */
1133                         return(IPPROTO_DONE);
1134                 }
1135                 goto drop;
1136         }
1137
1138 after_listen:
1139         /*
1140          * Should not happen - syncache should pick up these connections.
1141          *
1142          * Once we are past handling listen sockets we must be in the
1143          * correct protocol processing thread.
1144          */
1145         KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1146         KKASSERT(so->so_port == &curthread->td_msgport);
1147
1148         /* Unscale the window into a 32-bit value. */
1149         if (!(thflags & TH_SYN))
1150                 tiwin = th->th_win << tp->snd_scale;
1151         else
1152                 tiwin = th->th_win;
1153
1154         /*
1155          * This is the second part of the MSS DoS prevention code (after
1156          * minmss on the sending side) and it deals with too many too small
1157          * tcp packets in a too short timeframe (1 second).
1158          *
1159          * XXX Removed.  This code was crap.  It does not scale to network
1160          *     speed, and default values break NFS.  Gone.
1161          */
1162         /* REMOVED */
1163
1164         /*
1165          * Segment received on connection.
1166          *
1167          * Reset idle time and keep-alive timer.  Don't waste time if less
1168          * then a second has elapsed.
1169          */
1170         if ((int)(ticks - tp->t_rcvtime) > hz)
1171                 tcp_timer_keep_activity(tp, thflags);
1172
1173         /*
1174          * Process options.
1175          * XXX this is tradtitional behavior, may need to be cleaned up.
1176          */
1177         tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0, th->th_ack);
1178         if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1179                 if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) {
1180                         tp->t_flags |= TF_RCVD_SCALE;
1181                         tp->snd_scale = to.to_requested_s_scale;
1182                 }
1183
1184                 /*
1185                  * Initial send window; will be updated upon next ACK
1186                  */
1187                 tp->snd_wnd = th->th_win;
1188
1189                 if (to.to_flags & TOF_TS) {
1190                         tp->t_flags |= TF_RCVD_TSTMP;
1191                         tp->ts_recent = to.to_tsval;
1192                         tp->ts_recent_age = ticks;
1193                 }
1194                 if (!(to.to_flags & TOF_MSS))
1195                         to.to_mss = 0;
1196                 tcp_mss(tp, to.to_mss);
1197                 /*
1198                  * Only set the TF_SACK_PERMITTED per-connection flag
1199                  * if we got a SACK_PERMITTED option from the other side
1200                  * and the global tcp_do_sack variable is true.
1201                  */
1202                 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1203                         tp->t_flags |= TF_SACK_PERMITTED;
1204         }
1205
1206         /*
1207          * Header prediction: check for the two common cases
1208          * of a uni-directional data xfer.  If the packet has
1209          * no control flags, is in-sequence, the window didn't
1210          * change and we're not retransmitting, it's a
1211          * candidate.  If the length is zero and the ack moved
1212          * forward, we're the sender side of the xfer.  Just
1213          * free the data acked & wake any higher level process
1214          * that was blocked waiting for space.  If the length
1215          * is non-zero and the ack didn't move, we're the
1216          * receiver side.  If we're getting packets in-order
1217          * (the reassembly queue is empty), add the data to
1218          * the socket buffer and note that we need a delayed ack.
1219          * Make sure that the hidden state-flags are also off.
1220          * Since we check for TCPS_ESTABLISHED above, it can only
1221          * be TH_NEEDSYN.
1222          */
1223         if (tp->t_state == TCPS_ESTABLISHED &&
1224             (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1225             !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1226             (!(to.to_flags & TOF_TS) ||
1227              TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1228             th->th_seq == tp->rcv_nxt &&
1229             tp->snd_nxt == tp->snd_max) {
1230
1231                 /*
1232                  * If last ACK falls within this segment's sequence numbers,
1233                  * record the timestamp.
1234                  * NOTE that the test is modified according to the latest
1235                  * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1236                  */
1237                 if ((to.to_flags & TOF_TS) &&
1238                     SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1239                         tp->ts_recent_age = ticks;
1240                         tp->ts_recent = to.to_tsval;
1241                 }
1242
1243                 if (tlen == 0) {
1244                         if (SEQ_GT(th->th_ack, tp->snd_una) &&
1245                             SEQ_LEQ(th->th_ack, tp->snd_max) &&
1246                             tp->snd_cwnd >= tp->snd_wnd &&
1247                             !IN_FASTRECOVERY(tp)) {
1248                                 /*
1249                                  * This is a pure ack for outstanding data.
1250                                  */
1251                                 ++tcpstat.tcps_predack;
1252                                 /*
1253                                  * "bad retransmit" recovery
1254                                  *
1255                                  * If Eifel detection applies, then
1256                                  * it is deterministic, so use it
1257                                  * unconditionally over the old heuristic.
1258                                  * Otherwise, fall back to the old heuristic.
1259                                  */
1260                                 if (tcp_do_eifel_detect &&
1261                                     (to.to_flags & TOF_TS) && to.to_tsecr &&
1262                                     (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
1263                                         /* Eifel detection applicable. */
1264                                         if (to.to_tsecr < tp->t_rexmtTS) {
1265                                                 tcp_revert_congestion_state(tp);
1266                                                 ++tcpstat.tcps_eifeldetected;
1267                                                 if (tp->t_rxtshift != 1 ||
1268                                                     ticks >= tp->t_badrxtwin)
1269                                                         ++tcpstat.tcps_rttcantdetect;
1270                                         }
1271                                 } else if (tp->t_rxtshift == 1 &&
1272                                            ticks < tp->t_badrxtwin) {
1273                                         tcp_revert_congestion_state(tp);
1274                                         ++tcpstat.tcps_rttdetected;
1275                                 }
1276                                 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
1277                                     TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
1278                                 /*
1279                                  * Recalculate the retransmit timer / rtt.
1280                                  *
1281                                  * Some machines (certain windows boxes)
1282                                  * send broken timestamp replies during the
1283                                  * SYN+ACK phase, ignore timestamps of 0.
1284                                  */
1285                                 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1286                                         tcp_xmit_timer(tp,
1287                                             ticks - to.to_tsecr + 1,
1288                                             th->th_ack);
1289                                 } else if (tp->t_rtttime &&
1290                                            SEQ_GT(th->th_ack, tp->t_rtseq)) {
1291                                         tcp_xmit_timer(tp,
1292                                             ticks - tp->t_rtttime,
1293                                             th->th_ack);
1294                                 }
1295                                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1296                                 acked = th->th_ack - tp->snd_una;
1297                                 tcpstat.tcps_rcvackpack++;
1298                                 tcpstat.tcps_rcvackbyte += acked;
1299                                 sbdrop(&so->so_snd.sb, acked);
1300                                 tp->snd_recover = th->th_ack - 1;
1301                                 tp->snd_una = th->th_ack;
1302                                 tp->t_dupacks = 0;
1303                                 /*
1304                                  * Update window information.
1305                                  */
1306                                 if (tiwin != tp->snd_wnd &&
1307                                     acceptable_window_update(tp, th, tiwin)) {
1308                                         /* keep track of pure window updates */
1309                                         if (tp->snd_wl2 == th->th_ack &&
1310                                             tiwin > tp->snd_wnd)
1311                                                 tcpstat.tcps_rcvwinupd++;
1312                                         tp->snd_wnd = tiwin;
1313                                         tp->snd_wl1 = th->th_seq;
1314                                         tp->snd_wl2 = th->th_ack;
1315                                         if (tp->snd_wnd > tp->max_sndwnd)
1316                                                 tp->max_sndwnd = tp->snd_wnd;
1317                                 }
1318                                 m_freem(m);
1319                                 ND6_HINT(tp); /* some progress has been done */
1320                                 /*
1321                                  * If all outstanding data are acked, stop
1322                                  * retransmit timer, otherwise restart timer
1323                                  * using current (possibly backed-off) value.
1324                                  * If process is waiting for space,
1325                                  * wakeup/selwakeup/signal.  If data
1326                                  * are ready to send, let tcp_output
1327                                  * decide between more output or persist.
1328                                  */
1329                                 if (tp->snd_una == tp->snd_max) {
1330                                         tcp_callout_stop(tp, tp->tt_rexmt);
1331                                 } else if (!tcp_callout_active(tp,
1332                                             tp->tt_persist)) {
1333                                         tcp_callout_reset(tp, tp->tt_rexmt,
1334                                             tp->t_rxtcur, tcp_timer_rexmt);
1335                                 }
1336                                 sowwakeup(so);
1337                                 if (so->so_snd.ssb_cc > 0)
1338                                         tcp_output(tp);
1339                                 return(IPPROTO_DONE);
1340                         }
1341                 } else if (tiwin == tp->snd_wnd &&
1342                     th->th_ack == tp->snd_una &&
1343                     TAILQ_EMPTY(&tp->t_segq) &&
1344                     tlen <= ssb_space(&so->so_rcv)) {
1345                         u_long newsize = 0;     /* automatic sockbuf scaling */
1346                         /*
1347                          * This is a pure, in-sequence data packet
1348                          * with nothing on the reassembly queue and
1349                          * we have enough buffer space to take it.
1350                          */
1351                         ++tcpstat.tcps_preddat;
1352                         tp->rcv_nxt += tlen;
1353                         tcpstat.tcps_rcvpack++;
1354                         tcpstat.tcps_rcvbyte += tlen;
1355                         ND6_HINT(tp);   /* some progress has been done */
1356                 /*
1357                  * Automatic sizing of receive socket buffer.  Often the send
1358                  * buffer size is not optimally adjusted to the actual network
1359                  * conditions at hand (delay bandwidth product).  Setting the
1360                  * buffer size too small limits throughput on links with high
1361                  * bandwidth and high delay (eg. trans-continental/oceanic links).
1362                  *
1363                  * On the receive side the socket buffer memory is only rarely
1364                  * used to any significant extent.  This allows us to be much
1365                  * more aggressive in scaling the receive socket buffer.  For
1366                  * the case that the buffer space is actually used to a large
1367                  * extent and we run out of kernel memory we can simply drop
1368                  * the new segments; TCP on the sender will just retransmit it
1369                  * later.  Setting the buffer size too big may only consume too
1370                  * much kernel memory if the application doesn't read() from
1371                  * the socket or packet loss or reordering makes use of the
1372                  * reassembly queue.
1373                  *
1374                  * The criteria to step up the receive buffer one notch are:
1375                  *  1. the number of bytes received during the time it takes
1376                  *     one timestamp to be reflected back to us (the RTT);
1377                  *  2. received bytes per RTT is within seven eighth of the
1378                  *     current socket buffer size;
1379                  *  3. receive buffer size has not hit maximal automatic size;
1380                  *
1381                  * This algorithm does one step per RTT at most and only if
1382                  * we receive a bulk stream w/o packet losses or reorderings.
1383                  * Shrinking the buffer during idle times is not necessary as
1384                  * it doesn't consume any memory when idle.
1385                  *
1386                  * TODO: Only step up if the application is actually serving
1387                  * the buffer to better manage the socket buffer resources.
1388                  */
1389                         if (tcp_do_autorcvbuf &&
1390                             to.to_tsecr &&
1391                             (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1392                                 if (to.to_tsecr > tp->rfbuf_ts &&
1393                                     to.to_tsecr - tp->rfbuf_ts < hz) {
1394                                         if (tp->rfbuf_cnt >
1395                                             (so->so_rcv.ssb_hiwat / 8 * 7) &&
1396                                             so->so_rcv.ssb_hiwat <
1397                                             tcp_autorcvbuf_max) {
1398                                                 newsize =
1399                                                     ulmin(so->so_rcv.ssb_hiwat +
1400                                                           tcp_autorcvbuf_inc,
1401                                                           tcp_autorcvbuf_max);
1402                                         }
1403                                         /* Start over with next RTT. */
1404                                         tp->rfbuf_ts = 0;
1405                                         tp->rfbuf_cnt = 0;
1406                                 } else
1407                                         tp->rfbuf_cnt += tlen;  /* add up */
1408                         }
1409                         /*
1410                          * Add data to socket buffer.
1411                          */
1412                         if (so->so_state & SS_CANTRCVMORE) {
1413                                 m_freem(m);
1414                         } else {
1415                                 /*
1416                                  * Set new socket buffer size, give up when
1417                                  * limit is reached.
1418                                  *
1419                                  * Adjusting the size can mess up ACK
1420                                  * sequencing when pure window updates are
1421                                  * being avoided (which is the default),
1422                                  * so force an ack.
1423                                  */
1424                                 lwkt_gettoken(&so->so_rcv.ssb_token);
1425                                 if (newsize) {
1426                                         tp->t_flags |= TF_RXRESIZED;
1427                                         if (!ssb_reserve(&so->so_rcv, newsize,
1428                                                          so, NULL)) {
1429                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1430                                         }
1431                                         if (newsize >=
1432                                             (TCP_MAXWIN << tp->rcv_scale)) {
1433                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1434                                         }
1435                                 }
1436                                 m_adj(m, drop_hdrlen); /* delayed header drop */
1437                                 ssb_appendstream(&so->so_rcv, m);
1438                                 lwkt_reltoken(&so->so_rcv.ssb_token);
1439                         }
1440                         sorwakeup(so);
1441                         /*
1442                          * This code is responsible for most of the ACKs
1443                          * the TCP stack sends back after receiving a data
1444                          * packet.  Note that the DELAY_ACK check fails if
1445                          * the delack timer is already running, which results
1446                          * in an ack being sent every other packet (which is
1447                          * what we want).
1448                          *
1449                          * We then further aggregate acks by not actually
1450                          * sending one until the protocol thread has completed
1451                          * processing the current backlog of packets.  This
1452                          * does not delay the ack any further, but allows us
1453                          * to take advantage of the packet aggregation that
1454                          * high speed NICs do (usually blocks of 8-10 packets)
1455                          * to send a single ack rather then four or five acks,
1456                          * greatly reducing the ack rate, the return channel
1457                          * bandwidth, and the protocol overhead on both ends.
1458                          *
1459                          * Since this also has the effect of slowing down
1460                          * the exponential slow-start ramp-up, systems with 
1461                          * very large bandwidth-delay products might want
1462                          * to turn the feature off.
1463                          */
1464                         if (DELAY_ACK(tp)) {
1465                                 tcp_callout_reset(tp, tp->tt_delack,
1466                                     tcp_delacktime, tcp_timer_delack);
1467                         } else if (tcp_aggregate_acks) {
1468                                 tp->t_flags |= TF_ACKNOW;
1469                                 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1470                                         tp->t_flags |= TF_ONOUTPUTQ;
1471                                         tp->tt_cpu = mycpu->gd_cpuid;
1472                                         TAILQ_INSERT_TAIL(
1473                                             &tcpcbackq[tp->tt_cpu],
1474                                             tp, t_outputq);
1475                                 }
1476                         } else {
1477                                 tp->t_flags |= TF_ACKNOW;
1478                                 tcp_output(tp);
1479                         }
1480                         return(IPPROTO_DONE);
1481                 }
1482         }
1483
1484         /*
1485          * Calculate amount of space in receive window,
1486          * and then do TCP input processing.
1487          * Receive window is amount of space in rcv queue,
1488          * but not less than advertised window.
1489          */
1490         recvwin = ssb_space(&so->so_rcv);
1491         if (recvwin < 0)
1492                 recvwin = 0;
1493         tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1494
1495         /* Reset receive buffer auto scaling when not in bulk receive mode. */
1496         tp->rfbuf_ts = 0;
1497         tp->rfbuf_cnt = 0;
1498
1499         switch (tp->t_state) {
1500         /*
1501          * If the state is SYN_RECEIVED:
1502          *      if seg contains an ACK, but not for our SYN/ACK, send a RST.
1503          */
1504         case TCPS_SYN_RECEIVED:
1505                 if ((thflags & TH_ACK) &&
1506                     (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1507                      SEQ_GT(th->th_ack, tp->snd_max))) {
1508                                 rstreason = BANDLIM_RST_OPENPORT;
1509                                 goto dropwithreset;
1510                 }
1511                 break;
1512
1513         /*
1514          * If the state is SYN_SENT:
1515          *      if seg contains an ACK, but not for our SYN, drop the input.
1516          *      if seg contains a RST, then drop the connection.
1517          *      if seg does not contain SYN, then drop it.
1518          * Otherwise this is an acceptable SYN segment
1519          *      initialize tp->rcv_nxt and tp->irs
1520          *      if seg contains ack then advance tp->snd_una
1521          *      if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1522          *      arrange for segment to be acked (eventually)
1523          *      continue processing rest of data/controls, beginning with URG
1524          */
1525         case TCPS_SYN_SENT:
1526                 if ((thflags & TH_ACK) &&
1527                     (SEQ_LEQ(th->th_ack, tp->iss) ||
1528                      SEQ_GT(th->th_ack, tp->snd_max))) {
1529                         rstreason = BANDLIM_UNLIMITED;
1530                         goto dropwithreset;
1531                 }
1532                 if (thflags & TH_RST) {
1533                         if (thflags & TH_ACK)
1534                                 tp = tcp_drop(tp, ECONNREFUSED);
1535                         goto drop;
1536                 }
1537                 if (!(thflags & TH_SYN))
1538                         goto drop;
1539
1540                 tp->irs = th->th_seq;
1541                 tcp_rcvseqinit(tp);
1542                 if (thflags & TH_ACK) {
1543                         /* Our SYN was acked. */
1544                         tcpstat.tcps_connects++;
1545                         soisconnected(so);
1546                         /* Do window scaling on this connection? */
1547                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1548                             (TF_RCVD_SCALE | TF_REQ_SCALE))
1549                                 tp->rcv_scale = tp->request_r_scale;
1550                         tp->rcv_adv += tp->rcv_wnd;
1551                         tp->snd_una++;          /* SYN is acked */
1552                         tcp_callout_stop(tp, tp->tt_rexmt);
1553                         /*
1554                          * If there's data, delay ACK; if there's also a FIN
1555                          * ACKNOW will be turned on later.
1556                          */
1557                         if (DELAY_ACK(tp) && tlen != 0) {
1558                                 tcp_callout_reset(tp, tp->tt_delack,
1559                                     tcp_delacktime, tcp_timer_delack);
1560                         } else {
1561                                 tp->t_flags |= TF_ACKNOW;
1562                         }
1563                         /*
1564                          * Received <SYN,ACK> in SYN_SENT[*] state.
1565                          * Transitions:
1566                          *      SYN_SENT  --> ESTABLISHED
1567                          *      SYN_SENT* --> FIN_WAIT_1
1568                          */
1569                         tp->t_starttime = ticks;
1570                         if (tp->t_flags & TF_NEEDFIN) {
1571                                 tp->t_state = TCPS_FIN_WAIT_1;
1572                                 tp->t_flags &= ~TF_NEEDFIN;
1573                                 thflags &= ~TH_SYN;
1574                         } else {
1575                                 tcp_established(tp);
1576                         }
1577                 } else {
1578                         /*
1579                          * Received initial SYN in SYN-SENT[*] state =>
1580                          * simultaneous open.
1581                          * Do 3-way handshake:
1582                          *        SYN-SENT -> SYN-RECEIVED
1583                          *        SYN-SENT* -> SYN-RECEIVED*
1584                          */
1585                         tp->t_flags |= TF_ACKNOW;
1586                         tcp_callout_stop(tp, tp->tt_rexmt);
1587                         tp->t_state = TCPS_SYN_RECEIVED;
1588                 }
1589
1590                 /*
1591                  * Advance th->th_seq to correspond to first data byte.
1592                  * If data, trim to stay within window,
1593                  * dropping FIN if necessary.
1594                  */
1595                 th->th_seq++;
1596                 if (tlen > tp->rcv_wnd) {
1597                         todrop = tlen - tp->rcv_wnd;
1598                         m_adj(m, -todrop);
1599                         tlen = tp->rcv_wnd;
1600                         thflags &= ~TH_FIN;
1601                         tcpstat.tcps_rcvpackafterwin++;
1602                         tcpstat.tcps_rcvbyteafterwin += todrop;
1603                 }
1604                 tp->snd_wl1 = th->th_seq - 1;
1605                 tp->rcv_up = th->th_seq;
1606                 /*
1607                  * Client side of transaction: already sent SYN and data.
1608                  * If the remote host used T/TCP to validate the SYN,
1609                  * our data will be ACK'd; if so, enter normal data segment
1610                  * processing in the middle of step 5, ack processing.
1611                  * Otherwise, goto step 6.
1612                  */
1613                 if (thflags & TH_ACK)
1614                         goto process_ACK;
1615
1616                 goto step6;
1617
1618         /*
1619          * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1620          *      do normal processing (we no longer bother with T/TCP).
1621          */
1622         case TCPS_LAST_ACK:
1623         case TCPS_CLOSING:
1624         case TCPS_TIME_WAIT:
1625                 break;  /* continue normal processing */
1626         }
1627
1628         /*
1629          * States other than LISTEN or SYN_SENT.
1630          * First check the RST flag and sequence number since reset segments
1631          * are exempt from the timestamp and connection count tests.  This
1632          * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1633          * below which allowed reset segments in half the sequence space
1634          * to fall though and be processed (which gives forged reset
1635          * segments with a random sequence number a 50 percent chance of
1636          * killing a connection).
1637          * Then check timestamp, if present.
1638          * Then check the connection count, if present.
1639          * Then check that at least some bytes of segment are within
1640          * receive window.  If segment begins before rcv_nxt,
1641          * drop leading data (and SYN); if nothing left, just ack.
1642          *
1643          *
1644          * If the RST bit is set, check the sequence number to see
1645          * if this is a valid reset segment.
1646          * RFC 793 page 37:
1647          *   In all states except SYN-SENT, all reset (RST) segments
1648          *   are validated by checking their SEQ-fields.  A reset is
1649          *   valid if its sequence number is in the window.
1650          * Note: this does not take into account delayed ACKs, so
1651          *   we should test against last_ack_sent instead of rcv_nxt.
1652          *   The sequence number in the reset segment is normally an
1653          *   echo of our outgoing acknowledgement numbers, but some hosts
1654          *   send a reset with the sequence number at the rightmost edge
1655          *   of our receive window, and we have to handle this case.
1656          * If we have multiple segments in flight, the intial reset
1657          * segment sequence numbers will be to the left of last_ack_sent,
1658          * but they will eventually catch up.
1659          * In any case, it never made sense to trim reset segments to
1660          * fit the receive window since RFC 1122 says:
1661          *   4.2.2.12  RST Segment: RFC-793 Section 3.4
1662          *
1663          *    A TCP SHOULD allow a received RST segment to include data.
1664          *
1665          *    DISCUSSION
1666          *         It has been suggested that a RST segment could contain
1667          *         ASCII text that encoded and explained the cause of the
1668          *         RST.  No standard has yet been established for such
1669          *         data.
1670          *
1671          * If the reset segment passes the sequence number test examine
1672          * the state:
1673          *    SYN_RECEIVED STATE:
1674          *      If passive open, return to LISTEN state.
1675          *      If active open, inform user that connection was refused.
1676          *    ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1677          *      Inform user that connection was reset, and close tcb.
1678          *    CLOSING, LAST_ACK STATES:
1679          *      Close the tcb.
1680          *    TIME_WAIT STATE:
1681          *      Drop the segment - see Stevens, vol. 2, p. 964 and
1682          *      RFC 1337.
1683          */
1684         if (thflags & TH_RST) {
1685                 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1686                     SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1687                         switch (tp->t_state) {
1688
1689                         case TCPS_SYN_RECEIVED:
1690                                 so->so_error = ECONNREFUSED;
1691                                 goto close;
1692
1693                         case TCPS_ESTABLISHED:
1694                         case TCPS_FIN_WAIT_1:
1695                         case TCPS_FIN_WAIT_2:
1696                         case TCPS_CLOSE_WAIT:
1697                                 so->so_error = ECONNRESET;
1698                         close:
1699                                 tp->t_state = TCPS_CLOSED;
1700                                 tcpstat.tcps_drops++;
1701                                 tp = tcp_close(tp);
1702                                 break;
1703
1704                         case TCPS_CLOSING:
1705                         case TCPS_LAST_ACK:
1706                                 tp = tcp_close(tp);
1707                                 break;
1708
1709                         case TCPS_TIME_WAIT:
1710                                 break;
1711                         }
1712                 }
1713                 goto drop;
1714         }
1715
1716         /*
1717          * RFC 1323 PAWS: If we have a timestamp reply on this segment
1718          * and it's less than ts_recent, drop it.
1719          */
1720         if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1721             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1722                 /* Check to see if ts_recent is over 24 days old.  */
1723                 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1724                         /*
1725                          * Invalidate ts_recent.  If this segment updates
1726                          * ts_recent, the age will be reset later and ts_recent
1727                          * will get a valid value.  If it does not, setting
1728                          * ts_recent to zero will at least satisfy the
1729                          * requirement that zero be placed in the timestamp
1730                          * echo reply when ts_recent isn't valid.  The
1731                          * age isn't reset until we get a valid ts_recent
1732                          * because we don't want out-of-order segments to be
1733                          * dropped when ts_recent is old.
1734                          */
1735                         tp->ts_recent = 0;
1736                 } else if (tcp_paws_tolerance && tlen != 0 &&
1737                     tp->t_state == TCPS_ESTABLISHED &&
1738                     (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK&&
1739                     !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1740                     th->th_ack == tp->snd_una &&
1741                     tiwin == tp->snd_wnd &&
1742                     TSTMP_GEQ(to.to_tsval + tcp_paws_tolerance, tp->ts_recent)&&
1743                     (th->th_seq == tp->rcv_nxt ||
1744                      (SEQ_GT(th->th_seq, tp->rcv_nxt) && 
1745                       tcp_paws_canreasslast(tp, th, tlen)))) {
1746                         /*
1747                          * This tends to prevent valid new segments from being
1748                          * dropped by the reordered segments sent by the fast
1749                          * retransmission algorithm on the sending side, i.e.
1750                          * the fast retransmitted segment w/ larger timestamp
1751                          * arrives earlier than the previously sent new segments
1752                          * w/ smaller timestamp.
1753                          *
1754                          * If following conditions are met, the segment is
1755                          * accepted:
1756                          * - The segment contains data
1757                          * - The connection is established
1758                          * - The header does not contain important flags
1759                          * - SYN or FIN is not needed
1760                          * - It does not acknowledge new data
1761                          * - Receive window is not changed
1762                          * - The timestamp is within "acceptable" range
1763                          * - The new segment is what we are expecting or
1764                          *   the new segment could be merged w/ the last
1765                          *   pending segment on the reassemble queue
1766                          */
1767                         tcpstat.tcps_pawsaccept++;
1768                         tcpstat.tcps_pawsdrop++;
1769                 } else {
1770                         tcpstat.tcps_rcvduppack++;
1771                         tcpstat.tcps_rcvdupbyte += tlen;
1772                         tcpstat.tcps_pawsdrop++;
1773                         if (tlen)
1774                                 goto dropafterack;
1775                         goto drop;
1776                 }
1777         }
1778
1779         /*
1780          * In the SYN-RECEIVED state, validate that the packet belongs to
1781          * this connection before trimming the data to fit the receive
1782          * window.  Check the sequence number versus IRS since we know
1783          * the sequence numbers haven't wrapped.  This is a partial fix
1784          * for the "LAND" DoS attack.
1785          */
1786         if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1787                 rstreason = BANDLIM_RST_OPENPORT;
1788                 goto dropwithreset;
1789         }
1790
1791         todrop = tp->rcv_nxt - th->th_seq;
1792         if (todrop > 0) {
1793                 if (TCP_DO_SACK(tp)) {
1794                         /* Report duplicate segment at head of packet. */
1795                         tp->reportblk.rblk_start = th->th_seq;
1796                         tp->reportblk.rblk_end = TCP_SACK_BLKEND(
1797                             th->th_seq + tlen, thflags);
1798                         if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1799                                 tp->reportblk.rblk_end = tp->rcv_nxt;
1800                         tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_SACKLEFT);
1801                         tp->t_flags |= TF_ACKNOW;
1802                 }
1803                 if (thflags & TH_SYN) {
1804                         thflags &= ~TH_SYN;
1805                         th->th_seq++;
1806                         if (th->th_urp > 1)
1807                                 th->th_urp--;
1808                         else
1809                                 thflags &= ~TH_URG;
1810                         todrop--;
1811                 }
1812                 /*
1813                  * Following if statement from Stevens, vol. 2, p. 960.
1814                  */
1815                 if (todrop > tlen ||
1816                     (todrop == tlen && !(thflags & TH_FIN))) {
1817                         /*
1818                          * Any valid FIN must be to the left of the window.
1819                          * At this point the FIN must be a duplicate or out
1820                          * of sequence; drop it.
1821                          */
1822                         thflags &= ~TH_FIN;
1823
1824                         /*
1825                          * Send an ACK to resynchronize and drop any data.
1826                          * But keep on processing for RST or ACK.
1827                          */
1828                         tp->t_flags |= TF_ACKNOW;
1829                         todrop = tlen;
1830                         tcpstat.tcps_rcvduppack++;
1831                         tcpstat.tcps_rcvdupbyte += todrop;
1832                 } else {
1833                         tcpstat.tcps_rcvpartduppack++;
1834                         tcpstat.tcps_rcvpartdupbyte += todrop;
1835                 }
1836                 drop_hdrlen += todrop;  /* drop from the top afterwards */
1837                 th->th_seq += todrop;
1838                 tlen -= todrop;
1839                 if (th->th_urp > todrop)
1840                         th->th_urp -= todrop;
1841                 else {
1842                         thflags &= ~TH_URG;
1843                         th->th_urp = 0;
1844                 }
1845         }
1846
1847         /*
1848          * If new data are received on a connection after the
1849          * user processes are gone, then RST the other end.
1850          */
1851         if ((so->so_state & SS_NOFDREF) &&
1852             tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1853                 tp = tcp_close(tp);
1854                 tcpstat.tcps_rcvafterclose++;
1855                 rstreason = BANDLIM_UNLIMITED;
1856                 goto dropwithreset;
1857         }
1858
1859         /*
1860          * If segment ends after window, drop trailing data
1861          * (and PUSH and FIN); if nothing left, just ACK.
1862          */
1863         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1864         if (todrop > 0) {
1865                 tcpstat.tcps_rcvpackafterwin++;
1866                 if (todrop >= tlen) {
1867                         tcpstat.tcps_rcvbyteafterwin += tlen;
1868                         /*
1869                          * If a new connection request is received
1870                          * while in TIME_WAIT, drop the old connection
1871                          * and start over if the sequence numbers
1872                          * are above the previous ones.
1873                          */
1874                         if (thflags & TH_SYN &&
1875                             tp->t_state == TCPS_TIME_WAIT &&
1876                             SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1877                                 tp = tcp_close(tp);
1878                                 goto findpcb;
1879                         }
1880                         /*
1881                          * If window is closed can only take segments at
1882                          * window edge, and have to drop data and PUSH from
1883                          * incoming segments.  Continue processing, but
1884                          * remember to ack.  Otherwise, drop segment
1885                          * and ack.
1886                          */
1887                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1888                                 tp->t_flags |= TF_ACKNOW;
1889                                 tcpstat.tcps_rcvwinprobe++;
1890                         } else
1891                                 goto dropafterack;
1892                 } else
1893                         tcpstat.tcps_rcvbyteafterwin += todrop;
1894                 m_adj(m, -todrop);
1895                 tlen -= todrop;
1896                 thflags &= ~(TH_PUSH | TH_FIN);
1897         }
1898
1899         /*
1900          * If last ACK falls within this segment's sequence numbers,
1901          * record its timestamp.
1902          * NOTE:
1903          * 1) That the test incorporates suggestions from the latest
1904          *    proposal of the tcplw@cray.com list (Braden 1993/04/26).
1905          * 2) That updating only on newer timestamps interferes with
1906          *    our earlier PAWS tests, so this check should be solely
1907          *    predicated on the sequence space of this segment.
1908          * 3) That we modify the segment boundary check to be
1909          *        Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1910          *    instead of RFC1323's
1911          *        Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1912          *    This modified check allows us to overcome RFC1323's
1913          *    limitations as described in Stevens TCP/IP Illustrated
1914          *    Vol. 2 p.869. In such cases, we can still calculate the
1915          *    RTT correctly when RCV.NXT == Last.ACK.Sent.
1916          */
1917         if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1918             SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1919                                         + ((thflags & TH_SYN) != 0)
1920                                         + ((thflags & TH_FIN) != 0)))) {
1921                 tp->ts_recent_age = ticks;
1922                 tp->ts_recent = to.to_tsval;
1923         }
1924
1925         /*
1926          * If a SYN is in the window, then this is an
1927          * error and we send an RST and drop the connection.
1928          */
1929         if (thflags & TH_SYN) {
1930                 tp = tcp_drop(tp, ECONNRESET);
1931                 rstreason = BANDLIM_UNLIMITED;
1932                 goto dropwithreset;
1933         }
1934
1935         /*
1936          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN
1937          * flag is on (half-synchronized state), then queue data for
1938          * later processing; else drop segment and return.
1939          */
1940         if (!(thflags & TH_ACK)) {
1941                 if (tp->t_state == TCPS_SYN_RECEIVED ||
1942                     (tp->t_flags & TF_NEEDSYN))
1943                         goto step6;
1944                 else
1945                         goto drop;
1946         }
1947
1948         /*
1949          * Ack processing.
1950          */
1951         switch (tp->t_state) {
1952         /*
1953          * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1954          * ESTABLISHED state and continue processing.
1955          * The ACK was checked above.
1956          */
1957         case TCPS_SYN_RECEIVED:
1958
1959                 tcpstat.tcps_connects++;
1960                 soisconnected(so);
1961                 /* Do window scaling? */
1962                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1963                     (TF_RCVD_SCALE | TF_REQ_SCALE))
1964                         tp->rcv_scale = tp->request_r_scale;
1965                 /*
1966                  * Make transitions:
1967                  *      SYN-RECEIVED  -> ESTABLISHED
1968                  *      SYN-RECEIVED* -> FIN-WAIT-1
1969                  */
1970                 tp->t_starttime = ticks;
1971                 if (tp->t_flags & TF_NEEDFIN) {
1972                         tp->t_state = TCPS_FIN_WAIT_1;
1973                         tp->t_flags &= ~TF_NEEDFIN;
1974                 } else {
1975                         tcp_established(tp);
1976                 }
1977                 /*
1978                  * If segment contains data or ACK, will call tcp_reass()
1979                  * later; if not, do so now to pass queued data to user.
1980                  */
1981                 if (tlen == 0 && !(thflags & TH_FIN))
1982                         tcp_reass(tp, NULL, NULL, NULL);
1983                 /* fall into ... */
1984
1985         /*
1986          * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
1987          * ACKs.  If the ack is in the range
1988          *      tp->snd_una < th->th_ack <= tp->snd_max
1989          * then advance tp->snd_una to th->th_ack and drop
1990          * data from the retransmission queue.  If this ACK reflects
1991          * more up to date window information we update our window information.
1992          */
1993         case TCPS_ESTABLISHED:
1994         case TCPS_FIN_WAIT_1:
1995         case TCPS_FIN_WAIT_2:
1996         case TCPS_CLOSE_WAIT:
1997         case TCPS_CLOSING:
1998         case TCPS_LAST_ACK:
1999         case TCPS_TIME_WAIT:
2000
2001                 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
2002                         if (TCP_DO_SACK(tp))
2003                                 tcp_sack_update_scoreboard(tp, &to);
2004                         if (!tcp_callout_active(tp, tp->tt_rexmt) ||
2005                             th->th_ack != tp->snd_una) {
2006                                 if (tlen == 0 && tiwin == tp->snd_wnd)
2007                                         tcpstat.tcps_rcvdupack++;
2008                                 tp->t_dupacks = 0;
2009                                 break;
2010                         }
2011                         if (tlen != 0 || tiwin != tp->snd_wnd) {
2012                                 if (!tcp_do_rfc3517bis ||
2013                                     !TCP_DO_SACK(tp) ||
2014                                     (to.to_flags &
2015                                      (TOF_SACK | TOF_SACK_REDUNDANT))
2016                                      != TOF_SACK) {
2017                                         tp->t_dupacks = 0;
2018                                 } else {
2019                                         delayed_dupack = TRUE;
2020                                         th_dupack = th->th_ack;
2021                                 }
2022                                 break;
2023                         }
2024                         if (tcp_fast_recovery(tp, th->th_ack, &to))
2025                                 goto drop;
2026                         else
2027                                 break;
2028                 }
2029
2030                 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2031                 tp->t_dupacks = 0;
2032                 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2033                         /*
2034                          * Detected optimistic ACK attack.
2035                          * Force slow-start to de-synchronize attack.
2036                          */
2037                         tp->snd_cwnd = tp->t_maxseg;
2038                         tp->snd_wacked = 0;
2039
2040                         tcpstat.tcps_rcvacktoomuch++;
2041                         goto dropafterack;
2042                 }
2043                 /*
2044                  * If we reach this point, ACK is not a duplicate,
2045                  *     i.e., it ACKs something we sent.
2046                  */
2047                 if (tp->t_flags & TF_NEEDSYN) {
2048                         /*
2049                          * T/TCP: Connection was half-synchronized, and our
2050                          * SYN has been ACK'd (so connection is now fully
2051                          * synchronized).  Go to non-starred state,
2052                          * increment snd_una for ACK of SYN, and check if
2053                          * we can do window scaling.
2054                          */
2055                         tp->t_flags &= ~TF_NEEDSYN;
2056                         tp->snd_una++;
2057                         /* Do window scaling? */
2058                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2059                             (TF_RCVD_SCALE | TF_REQ_SCALE))
2060                                 tp->rcv_scale = tp->request_r_scale;
2061                 }
2062
2063 process_ACK:
2064                 acked = th->th_ack - tp->snd_una;
2065                 tcpstat.tcps_rcvackpack++;
2066                 tcpstat.tcps_rcvackbyte += acked;
2067
2068                 if (tcp_do_eifel_detect && acked > 0 &&
2069                     (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2070                     (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
2071                         /* Eifel detection applicable. */
2072                         if (to.to_tsecr < tp->t_rexmtTS) {
2073                                 ++tcpstat.tcps_eifeldetected;
2074                                 tcp_revert_congestion_state(tp);
2075                                 if (tp->t_rxtshift != 1 ||
2076                                     ticks >= tp->t_badrxtwin)
2077                                         ++tcpstat.tcps_rttcantdetect;
2078                         }
2079                 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2080                         /*
2081                          * If we just performed our first retransmit,
2082                          * and the ACK arrives within our recovery window,
2083                          * then it was a mistake to do the retransmit
2084                          * in the first place.  Recover our original cwnd
2085                          * and ssthresh, and proceed to transmit where we
2086                          * left off.
2087                          */
2088                         tcp_revert_congestion_state(tp);
2089                         ++tcpstat.tcps_rttdetected;
2090                 }
2091
2092                 /*
2093                  * If we have a timestamp reply, update smoothed
2094                  * round trip time.  If no timestamp is present but
2095                  * transmit timer is running and timed sequence
2096                  * number was acked, update smoothed round trip time.
2097                  * Since we now have an rtt measurement, cancel the
2098                  * timer backoff (cf., Phil Karn's retransmit alg.).
2099                  * Recompute the initial retransmit timer.
2100                  *
2101                  * Some machines (certain windows boxes) send broken
2102                  * timestamp replies during the SYN+ACK phase, ignore
2103                  * timestamps of 0.
2104                  */
2105                 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2106                         tcp_xmit_timer(tp, ticks - to.to_tsecr + 1, th->th_ack);
2107                 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2108                         tcp_xmit_timer(tp, ticks - tp->t_rtttime, th->th_ack);
2109                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2110
2111                 /*
2112                  * If no data (only SYN) was ACK'd,
2113                  *    skip rest of ACK processing.
2114                  */
2115                 if (acked == 0)
2116                         goto step6;
2117
2118                 /* Stop looking for an acceptable ACK since one was received. */
2119                 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
2120                     TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
2121
2122                 if (acked > so->so_snd.ssb_cc) {
2123                         tp->snd_wnd -= so->so_snd.ssb_cc;
2124                         sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2125                         ourfinisacked = TRUE;
2126                 } else {
2127                         sbdrop(&so->so_snd.sb, acked);
2128                         tp->snd_wnd -= acked;
2129                         ourfinisacked = FALSE;
2130                 }
2131                 sowwakeup(so);
2132
2133                 /*
2134                  * Update window information.
2135                  */
2136                 if (acceptable_window_update(tp, th, tiwin)) {
2137                         /* keep track of pure window updates */
2138                         if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2139                             tiwin > tp->snd_wnd)
2140                                 tcpstat.tcps_rcvwinupd++;
2141                         tp->snd_wnd = tiwin;
2142                         tp->snd_wl1 = th->th_seq;
2143                         tp->snd_wl2 = th->th_ack;
2144                         if (tp->snd_wnd > tp->max_sndwnd)
2145                                 tp->max_sndwnd = tp->snd_wnd;
2146                         needoutput = TRUE;
2147                 }
2148
2149                 tp->snd_una = th->th_ack;
2150                 if (TCP_DO_SACK(tp))
2151                         tcp_sack_update_scoreboard(tp, &to);
2152                 if (IN_FASTRECOVERY(tp)) {
2153                         if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2154                                 EXIT_FASTRECOVERY(tp);
2155                                 needoutput = TRUE;
2156                                 /*
2157                                  * If the congestion window was inflated
2158                                  * to account for the other side's
2159                                  * cached packets, retract it.
2160                                  */
2161                                 if (!TCP_DO_SACK(tp))
2162                                         tp->snd_cwnd = tp->snd_ssthresh;
2163
2164                                 /*
2165                                  * Window inflation should have left us
2166                                  * with approximately snd_ssthresh outstanding
2167                                  * data.  But, in case we would be inclined
2168                                  * to send a burst, better do it using
2169                                  * slow start.
2170                                  */
2171                                 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2172                                            tp->snd_max + 2 * tp->t_maxseg))
2173                                         tp->snd_cwnd =
2174                                             (tp->snd_max - tp->snd_una) +
2175                                             2 * tp->t_maxseg;
2176
2177                                 tp->snd_wacked = 0;
2178                         } else {
2179                                 if (TCP_DO_SACK(tp)) {
2180                                         tp->snd_max_rexmt = tp->snd_max;
2181                                         tcp_sack_rexmt(tp);
2182                                 } else {
2183                                         tcp_newreno_partial_ack(tp, th, acked);
2184                                 }
2185                                 needoutput = FALSE;
2186                         }
2187                 } else {
2188                         /*
2189                          * Open the congestion window.  When in slow-start,
2190                          * open exponentially: maxseg per packet.  Otherwise,
2191                          * open linearly: maxseg per window.
2192                          */
2193                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
2194                                 u_int abc_sslimit =
2195                                     (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2196                                      tp->t_maxseg : 2 * tp->t_maxseg);
2197
2198                                 /* slow-start */
2199                                 tp->snd_cwnd += tcp_do_abc ?
2200                                     min(acked, abc_sslimit) : tp->t_maxseg;
2201                         } else {
2202                                 /* linear increase */
2203                                 tp->snd_wacked += tcp_do_abc ? acked :
2204                                     tp->t_maxseg;
2205                                 if (tp->snd_wacked >= tp->snd_cwnd) {
2206                                         tp->snd_wacked -= tp->snd_cwnd;
2207                                         tp->snd_cwnd += tp->t_maxseg;
2208                                 }
2209                         }
2210                         tp->snd_cwnd = min(tp->snd_cwnd,
2211                                            TCP_MAXWIN << tp->snd_scale);
2212                         tp->snd_recover = th->th_ack - 1;
2213                 }
2214                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2215                         tp->snd_nxt = tp->snd_una;
2216
2217                 /*
2218                  * If all outstanding data is acked, stop retransmit
2219                  * timer and remember to restart (more output or persist).
2220                  * If there is more data to be acked, restart retransmit
2221                  * timer, using current (possibly backed-off) value.
2222                  */
2223                 if (th->th_ack == tp->snd_max) {
2224                         tcp_callout_stop(tp, tp->tt_rexmt);
2225                         needoutput = TRUE;
2226                 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2227                         tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2228                             tcp_timer_rexmt);
2229                 }
2230
2231                 switch (tp->t_state) {
2232                 /*
2233                  * In FIN_WAIT_1 STATE in addition to the processing
2234                  * for the ESTABLISHED state if our FIN is now acknowledged
2235                  * then enter FIN_WAIT_2.
2236                  */
2237                 case TCPS_FIN_WAIT_1:
2238                         if (ourfinisacked) {
2239                                 /*
2240                                  * If we can't receive any more
2241                                  * data, then closing user can proceed.
2242                                  * Starting the timer is contrary to the
2243                                  * specification, but if we don't get a FIN
2244                                  * we'll hang forever.
2245                                  */
2246                                 if (so->so_state & SS_CANTRCVMORE) {
2247                                         soisdisconnected(so);
2248                                         tcp_callout_reset(tp, tp->tt_2msl,
2249                                             tp->t_maxidle, tcp_timer_2msl);
2250                                 }
2251                                 tp->t_state = TCPS_FIN_WAIT_2;
2252                         }
2253                         break;
2254
2255                 /*
2256                  * In CLOSING STATE in addition to the processing for
2257                  * the ESTABLISHED state if the ACK acknowledges our FIN
2258                  * then enter the TIME-WAIT state, otherwise ignore
2259                  * the segment.
2260                  */
2261                 case TCPS_CLOSING:
2262                         if (ourfinisacked) {
2263                                 tp->t_state = TCPS_TIME_WAIT;
2264                                 tcp_canceltimers(tp);
2265                                 tcp_callout_reset(tp, tp->tt_2msl,
2266                                             2 * tcp_rmx_msl(tp),
2267                                             tcp_timer_2msl);
2268                                 soisdisconnected(so);
2269                         }
2270                         break;
2271
2272                 /*
2273                  * In LAST_ACK, we may still be waiting for data to drain
2274                  * and/or to be acked, as well as for the ack of our FIN.
2275                  * If our FIN is now acknowledged, delete the TCB,
2276                  * enter the closed state and return.
2277                  */
2278                 case TCPS_LAST_ACK:
2279                         if (ourfinisacked) {
2280                                 tp = tcp_close(tp);
2281                                 goto drop;
2282                         }
2283                         break;
2284
2285                 /*
2286                  * In TIME_WAIT state the only thing that should arrive
2287                  * is a retransmission of the remote FIN.  Acknowledge
2288                  * it and restart the finack timer.
2289                  */
2290                 case TCPS_TIME_WAIT:
2291                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2292                             tcp_timer_2msl);
2293                         goto dropafterack;
2294                 }
2295         }
2296
2297 step6:
2298         /*
2299          * Update window information.
2300          * Don't look at window if no ACK: TAC's send garbage on first SYN.
2301          */
2302         if ((thflags & TH_ACK) &&
2303             acceptable_window_update(tp, th, tiwin)) {
2304                 /* keep track of pure window updates */
2305                 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2306                     tiwin > tp->snd_wnd)
2307                         tcpstat.tcps_rcvwinupd++;
2308                 tp->snd_wnd = tiwin;
2309                 tp->snd_wl1 = th->th_seq;
2310                 tp->snd_wl2 = th->th_ack;
2311                 if (tp->snd_wnd > tp->max_sndwnd)
2312                         tp->max_sndwnd = tp->snd_wnd;
2313                 needoutput = TRUE;
2314         }
2315
2316         /*
2317          * Process segments with URG.
2318          */
2319         if ((thflags & TH_URG) && th->th_urp &&
2320             !TCPS_HAVERCVDFIN(tp->t_state)) {
2321                 /*
2322                  * This is a kludge, but if we receive and accept
2323                  * random urgent pointers, we'll crash in
2324                  * soreceive.  It's hard to imagine someone
2325                  * actually wanting to send this much urgent data.
2326                  */
2327                 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2328                         th->th_urp = 0;                 /* XXX */
2329                         thflags &= ~TH_URG;             /* XXX */
2330                         goto dodata;                    /* XXX */
2331                 }
2332                 /*
2333                  * If this segment advances the known urgent pointer,
2334                  * then mark the data stream.  This should not happen
2335                  * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2336                  * a FIN has been received from the remote side.
2337                  * In these states we ignore the URG.
2338                  *
2339                  * According to RFC961 (Assigned Protocols),
2340                  * the urgent pointer points to the last octet
2341                  * of urgent data.  We continue, however,
2342                  * to consider it to indicate the first octet
2343                  * of data past the urgent section as the original
2344                  * spec states (in one of two places).
2345                  */
2346                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2347                         tp->rcv_up = th->th_seq + th->th_urp;
2348                         so->so_oobmark = so->so_rcv.ssb_cc +
2349                             (tp->rcv_up - tp->rcv_nxt) - 1;
2350                         if (so->so_oobmark == 0)
2351                                 sosetstate(so, SS_RCVATMARK);
2352                         sohasoutofband(so);
2353                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2354                 }
2355                 /*
2356                  * Remove out of band data so doesn't get presented to user.
2357                  * This can happen independent of advancing the URG pointer,
2358                  * but if two URG's are pending at once, some out-of-band
2359                  * data may creep in... ick.
2360                  */
2361                 if (th->th_urp <= (u_long)tlen &&
2362                     !(so->so_options & SO_OOBINLINE)) {
2363                         /* hdr drop is delayed */
2364                         tcp_pulloutofband(so, th, m, drop_hdrlen);
2365                 }
2366         } else {
2367                 /*
2368                  * If no out of band data is expected,
2369                  * pull receive urgent pointer along
2370                  * with the receive window.
2371                  */
2372                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2373                         tp->rcv_up = tp->rcv_nxt;
2374         }
2375
2376 dodata:                                                 /* XXX */
2377         /*
2378          * Process the segment text, merging it into the TCP sequencing queue,
2379          * and arranging for acknowledgment of receipt if necessary.
2380          * This process logically involves adjusting tp->rcv_wnd as data
2381          * is presented to the user (this happens in tcp_usrreq.c,
2382          * case PRU_RCVD).  If a FIN has already been received on this
2383          * connection then we just ignore the text.
2384          */
2385         if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2386                 m_adj(m, drop_hdrlen);  /* delayed header drop */
2387                 /*
2388                  * Insert segment which includes th into TCP reassembly queue
2389                  * with control block tp.  Set thflags to whether reassembly now
2390                  * includes a segment with FIN.  This handles the common case
2391                  * inline (segment is the next to be received on an established
2392                  * connection, and the queue is empty), avoiding linkage into
2393                  * and removal from the queue and repetition of various
2394                  * conversions.
2395                  * Set DELACK for segments received in order, but ack
2396                  * immediately when segments are out of order (so
2397                  * fast retransmit can work).
2398                  */
2399                 if (th->th_seq == tp->rcv_nxt &&
2400                     TAILQ_EMPTY(&tp->t_segq) &&
2401                     TCPS_HAVEESTABLISHED(tp->t_state)) {
2402                         if (DELAY_ACK(tp)) {
2403                                 tcp_callout_reset(tp, tp->tt_delack,
2404                                     tcp_delacktime, tcp_timer_delack);
2405                         } else {
2406                                 tp->t_flags |= TF_ACKNOW;
2407                         }
2408                         tp->rcv_nxt += tlen;
2409                         thflags = th->th_flags & TH_FIN;
2410                         tcpstat.tcps_rcvpack++;
2411                         tcpstat.tcps_rcvbyte += tlen;
2412                         ND6_HINT(tp);
2413                         if (so->so_state & SS_CANTRCVMORE) {
2414                                 m_freem(m);
2415                         } else {
2416                                 lwkt_gettoken(&so->so_rcv.ssb_token);
2417                                 ssb_appendstream(&so->so_rcv, m);
2418                                 lwkt_reltoken(&so->so_rcv.ssb_token);
2419                         }
2420                         sorwakeup(so);
2421                 } else {
2422                         if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
2423                                 /* Initialize SACK report block. */
2424                                 tp->reportblk.rblk_start = th->th_seq;
2425                                 tp->reportblk.rblk_end = TCP_SACK_BLKEND(
2426                                     th->th_seq + tlen, thflags);
2427                         }
2428                         thflags = tcp_reass(tp, th, &tlen, m);
2429                         tp->t_flags |= TF_ACKNOW;
2430                 }
2431
2432                 /*
2433                  * Note the amount of data that peer has sent into
2434                  * our window, in order to estimate the sender's
2435                  * buffer size.
2436                  */
2437                 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2438         } else {
2439                 m_freem(m);
2440                 thflags &= ~TH_FIN;
2441         }
2442
2443         /*
2444          * If FIN is received ACK the FIN and let the user know
2445          * that the connection is closing.
2446          */
2447         if (thflags & TH_FIN) {
2448                 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2449                         socantrcvmore(so);
2450                         /*
2451                          * If connection is half-synchronized
2452                          * (ie NEEDSYN flag on) then delay ACK,
2453                          * so it may be piggybacked when SYN is sent.
2454                          * Otherwise, since we received a FIN then no
2455                          * more input can be expected, send ACK now.
2456                          */
2457                         if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2458                                 tcp_callout_reset(tp, tp->tt_delack,
2459                                     tcp_delacktime, tcp_timer_delack);
2460                         } else {
2461                                 tp->t_flags |= TF_ACKNOW;
2462                         }
2463                         tp->rcv_nxt++;
2464                 }
2465
2466                 switch (tp->t_state) {
2467                 /*
2468                  * In SYN_RECEIVED and ESTABLISHED STATES
2469                  * enter the CLOSE_WAIT state.
2470                  */
2471                 case TCPS_SYN_RECEIVED:
2472                         tp->t_starttime = ticks;
2473                         /*FALLTHROUGH*/
2474                 case TCPS_ESTABLISHED:
2475                         tp->t_state = TCPS_CLOSE_WAIT;
2476                         break;
2477
2478                 /*
2479                  * If still in FIN_WAIT_1 STATE FIN has not been acked so
2480                  * enter the CLOSING state.
2481                  */
2482                 case TCPS_FIN_WAIT_1:
2483                         tp->t_state = TCPS_CLOSING;
2484                         break;
2485
2486                 /*
2487                  * In FIN_WAIT_2 state enter the TIME_WAIT state,
2488                  * starting the time-wait timer, turning off the other
2489                  * standard timers.
2490                  */
2491                 case TCPS_FIN_WAIT_2:
2492                         tp->t_state = TCPS_TIME_WAIT;
2493                         tcp_canceltimers(tp);
2494                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2495                                     tcp_timer_2msl);
2496                         soisdisconnected(so);
2497                         break;
2498
2499                 /*
2500                  * In TIME_WAIT state restart the 2 MSL time_wait timer.
2501                  */
2502                 case TCPS_TIME_WAIT:
2503                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2504                             tcp_timer_2msl);
2505                         break;
2506                 }
2507         }
2508
2509 #ifdef TCPDEBUG
2510         if (so->so_options & SO_DEBUG)
2511                 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2512 #endif
2513
2514         /*
2515          * Delayed duplicated ACK processing
2516          */
2517         if (delayed_dupack && tcp_fast_recovery(tp, th_dupack, &to))
2518                 needoutput = FALSE;
2519
2520         /*
2521          * Return any desired output.
2522          */
2523         if (needoutput || (tp->t_flags & TF_ACKNOW))
2524                 tcp_output(tp);
2525         tcp_sack_report_cleanup(tp);
2526         return(IPPROTO_DONE);
2527
2528 dropafterack:
2529         /*
2530          * Generate an ACK dropping incoming segment if it occupies
2531          * sequence space, where the ACK reflects our state.
2532          *
2533          * We can now skip the test for the RST flag since all
2534          * paths to this code happen after packets containing
2535          * RST have been dropped.
2536          *
2537          * In the SYN-RECEIVED state, don't send an ACK unless the
2538          * segment we received passes the SYN-RECEIVED ACK test.
2539          * If it fails send a RST.  This breaks the loop in the
2540          * "LAND" DoS attack, and also prevents an ACK storm
2541          * between two listening ports that have been sent forged
2542          * SYN segments, each with the source address of the other.
2543          */
2544         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2545             (SEQ_GT(tp->snd_una, th->th_ack) ||
2546              SEQ_GT(th->th_ack, tp->snd_max)) ) {
2547                 rstreason = BANDLIM_RST_OPENPORT;
2548                 goto dropwithreset;
2549         }
2550 #ifdef TCPDEBUG
2551         if (so->so_options & SO_DEBUG)
2552                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2553 #endif
2554         m_freem(m);
2555         tp->t_flags |= TF_ACKNOW;
2556         tcp_output(tp);
2557         tcp_sack_report_cleanup(tp);
2558         return(IPPROTO_DONE);
2559
2560 dropwithreset:
2561         /*
2562          * Generate a RST, dropping incoming segment.
2563          * Make ACK acceptable to originator of segment.
2564          * Don't bother to respond if destination was broadcast/multicast.
2565          */
2566         if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2567                 goto drop;
2568         if (isipv6) {
2569                 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2570                     IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2571                         goto drop;
2572         } else {
2573                 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2574                     IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2575                     ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2576                     in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2577                         goto drop;
2578         }
2579         /* IPv6 anycast check is done at tcp6_input() */
2580
2581         /*
2582          * Perform bandwidth limiting.
2583          */
2584 #ifdef ICMP_BANDLIM
2585         if (badport_bandlim(rstreason) < 0)
2586                 goto drop;
2587 #endif
2588
2589 #ifdef TCPDEBUG
2590         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2591                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2592 #endif
2593         if (thflags & TH_ACK)
2594                 /* mtod() below is safe as long as hdr dropping is delayed */
2595                 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2596                             TH_RST);
2597         else {
2598                 if (thflags & TH_SYN)
2599                         tlen++;
2600                 /* mtod() below is safe as long as hdr dropping is delayed */
2601                 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2602                             (tcp_seq)0, TH_RST | TH_ACK);
2603         }
2604         if (tp != NULL)
2605                 tcp_sack_report_cleanup(tp);
2606         return(IPPROTO_DONE);
2607
2608 drop:
2609         /*
2610          * Drop space held by incoming segment and return.
2611          */
2612 #ifdef TCPDEBUG
2613         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2614                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2615 #endif
2616         m_freem(m);
2617         if (tp != NULL)
2618                 tcp_sack_report_cleanup(tp);
2619         return(IPPROTO_DONE);
2620 }
2621
2622 /*
2623  * Parse TCP options and place in tcpopt.
2624  */
2625 static void
2626 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn,
2627     tcp_seq ack)
2628 {
2629         int opt, optlen, i;
2630
2631         to->to_flags = 0;
2632         for (; cnt > 0; cnt -= optlen, cp += optlen) {
2633                 opt = cp[0];
2634                 if (opt == TCPOPT_EOL)
2635                         break;
2636                 if (opt == TCPOPT_NOP)
2637                         optlen = 1;
2638                 else {
2639                         if (cnt < 2)
2640                                 break;
2641                         optlen = cp[1];
2642                         if (optlen < 2 || optlen > cnt)
2643                                 break;
2644                 }
2645                 switch (opt) {
2646                 case TCPOPT_MAXSEG:
2647                         if (optlen != TCPOLEN_MAXSEG)
2648                                 continue;
2649                         if (!is_syn)
2650                                 continue;
2651                         to->to_flags |= TOF_MSS;
2652                         bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2653                         to->to_mss = ntohs(to->to_mss);
2654                         break;
2655                 case TCPOPT_WINDOW:
2656                         if (optlen != TCPOLEN_WINDOW)
2657                                 continue;
2658                         if (!is_syn)
2659                                 continue;
2660                         to->to_flags |= TOF_SCALE;
2661                         to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2662                         break;
2663                 case TCPOPT_TIMESTAMP:
2664                         if (optlen != TCPOLEN_TIMESTAMP)
2665                                 continue;
2666                         to->to_flags |= TOF_TS;
2667                         bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2668                         to->to_tsval = ntohl(to->to_tsval);
2669                         bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2670                         to->to_tsecr = ntohl(to->to_tsecr);
2671                         /*
2672                          * If echoed timestamp is later than the current time,
2673                          * fall back to non RFC1323 RTT calculation.
2674                          */
2675                         if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2676                                 to->to_tsecr = 0;
2677                         break;
2678                 case TCPOPT_SACK_PERMITTED:
2679                         if (optlen != TCPOLEN_SACK_PERMITTED)
2680                                 continue;
2681                         if (!is_syn)
2682                                 continue;
2683                         to->to_flags |= TOF_SACK_PERMITTED;
2684                         break;
2685                 case TCPOPT_SACK:
2686                         if ((optlen - 2) & 0x07)        /* not multiple of 8 */
2687                                 continue;
2688                         to->to_nsackblocks = (optlen - 2) / 8;
2689                         to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2690                         to->to_flags |= TOF_SACK;
2691                         for (i = 0; i < to->to_nsackblocks; i++) {
2692                                 struct raw_sackblock *r = &to->to_sackblocks[i];
2693
2694                                 r->rblk_start = ntohl(r->rblk_start);
2695                                 r->rblk_end = ntohl(r->rblk_end);
2696
2697                                 if (SEQ_LEQ(r->rblk_end, r->rblk_start)) {
2698                                         /*
2699                                          * Invalid SACK block; discard all
2700                                          * SACK blocks
2701                                          */
2702                                         tcpstat.tcps_rcvbadsackopt++;
2703                                         to->to_nsackblocks = 0;
2704                                         to->to_sackblocks = NULL;
2705                                         to->to_flags &= ~TOF_SACK;
2706                                         break;
2707                                 }
2708                         }
2709                         if ((to->to_flags & TOF_SACK) &&
2710                             tcp_sack_ndsack_blocks(to->to_sackblocks,
2711                             to->to_nsackblocks, ack))
2712                                 to->to_flags |= TOF_DSACK;
2713                         break;
2714 #ifdef TCP_SIGNATURE
2715                 /*
2716                  * XXX In order to reply to a host which has set the
2717                  * TCP_SIGNATURE option in its initial SYN, we have to
2718                  * record the fact that the option was observed here
2719                  * for the syncache code to perform the correct response.
2720                  */
2721                 case TCPOPT_SIGNATURE:
2722                         if (optlen != TCPOLEN_SIGNATURE)
2723                                 continue;
2724                         to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2725                         break;
2726 #endif /* TCP_SIGNATURE */
2727                 default:
2728                         continue;
2729                 }
2730         }
2731 }
2732
2733 /*
2734  * Pull out of band byte out of a segment so
2735  * it doesn't appear in the user's data queue.
2736  * It is still reflected in the segment length for
2737  * sequencing purposes.
2738  * "off" is the delayed to be dropped hdrlen.
2739  */
2740 static void
2741 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2742 {
2743         int cnt = off + th->th_urp - 1;
2744
2745         while (cnt >= 0) {
2746                 if (m->m_len > cnt) {
2747                         char *cp = mtod(m, caddr_t) + cnt;
2748                         struct tcpcb *tp = sototcpcb(so);
2749
2750                         tp->t_iobc = *cp;
2751                         tp->t_oobflags |= TCPOOB_HAVEDATA;
2752                         bcopy(cp + 1, cp, m->m_len - cnt - 1);
2753                         m->m_len--;
2754                         if (m->m_flags & M_PKTHDR)
2755                                 m->m_pkthdr.len--;
2756                         return;
2757                 }
2758                 cnt -= m->m_len;
2759                 m = m->m_next;
2760                 if (m == NULL)
2761                         break;
2762         }
2763         panic("tcp_pulloutofband");
2764 }
2765
2766 /*
2767  * Collect new round-trip time estimate
2768  * and update averages and current timeout.
2769  */
2770 static void
2771 tcp_xmit_timer(struct tcpcb *tp, int rtt, tcp_seq ack)
2772 {
2773         int rebaserto = 0;
2774
2775         tcpstat.tcps_rttupdated++;
2776         tp->t_rttupdated++;
2777         if ((tp->rxt_flags & TRXT_F_REBASERTO) &&
2778             SEQ_GT(ack, tp->snd_max_prev)) {
2779 #ifdef DEBUG_EIFEL_RESPONSE
2780                 kprintf("srtt/rttvar, prev %d/%d, cur %d/%d, ",
2781                     tp->t_srtt_prev, tp->t_rttvar_prev,
2782                     tp->t_srtt, tp->t_rttvar);
2783 #endif
2784
2785                 tcpstat.tcps_eifelresponse++;
2786                 rebaserto = 1;
2787                 tp->rxt_flags &= ~TRXT_F_REBASERTO;
2788                 tp->t_srtt = max(tp->t_srtt_prev, (rtt << TCP_RTT_SHIFT));
2789                 tp->t_rttvar = max(tp->t_rttvar_prev,
2790                     (rtt << (TCP_RTTVAR_SHIFT - 1)));
2791                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2792                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2793
2794 #ifdef DEBUG_EIFEL_RESPONSE
2795                 kprintf("new %d/%d ", tp->t_srtt, tp->t_rttvar);
2796 #endif
2797         } else if (tp->t_srtt != 0) {
2798                 int delta;
2799
2800                 /*
2801                  * srtt is stored as fixed point with 5 bits after the
2802                  * binary point (i.e., scaled by 8).  The following magic
2803                  * is equivalent to the smoothing algorithm in rfc793 with
2804                  * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2805                  * point).  Adjust rtt to origin 0.
2806                  */
2807                 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2808                         - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2809
2810                 if ((tp->t_srtt += delta) <= 0)
2811                         tp->t_srtt = 1;
2812
2813                 /*
2814                  * We accumulate a smoothed rtt variance (actually, a
2815                  * smoothed mean difference), then set the retransmit
2816                  * timer to smoothed rtt + 4 times the smoothed variance.
2817                  * rttvar is stored as fixed point with 4 bits after the
2818                  * binary point (scaled by 16).  The following is
2819                  * equivalent to rfc793 smoothing with an alpha of .75
2820                  * (rttvar = rttvar*3/4 + |delta| / 4).  This replaces
2821                  * rfc793's wired-in beta.
2822                  */
2823                 if (delta < 0)
2824                         delta = -delta;
2825                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2826                 if ((tp->t_rttvar += delta) <= 0)
2827                         tp->t_rttvar = 1;
2828                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2829                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2830         } else {
2831                 /*
2832                  * No rtt measurement yet - use the unsmoothed rtt.
2833                  * Set the variance to half the rtt (so our first
2834                  * retransmit happens at 3*rtt).
2835                  */
2836                 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2837                 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2838                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2839         }
2840         tp->t_rtttime = 0;
2841         tp->t_rxtshift = 0;
2842
2843 #ifdef DEBUG_EIFEL_RESPONSE
2844         if (rebaserto) {
2845                 kprintf("| rxtcur prev %d, old %d, ",
2846                     tp->t_rxtcur_prev, tp->t_rxtcur);
2847         }
2848 #endif
2849
2850         /*
2851          * the retransmit should happen at rtt + 4 * rttvar.
2852          * Because of the way we do the smoothing, srtt and rttvar
2853          * will each average +1/2 tick of bias.  When we compute
2854          * the retransmit timer, we want 1/2 tick of rounding and
2855          * 1 extra tick because of +-1/2 tick uncertainty in the
2856          * firing of the timer.  The bias will give us exactly the
2857          * 1.5 tick we need.  But, because the bias is
2858          * statistical, we have to test that we don't drop below
2859          * the minimum feasible timer (which is 2 ticks).
2860          */
2861         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2862                       max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2863
2864         if (rebaserto) {
2865                 if (tp->t_rxtcur < tp->t_rxtcur_prev + tcp_eifel_rtoinc) {
2866                         /*
2867                          * RFC4015 requires that the new RTO is at least
2868                          * 2*G (tcp_eifel_rtoinc) greater then the RTO
2869                          * (t_rxtcur_prev) when the spurious retransmit
2870                          * timeout happens.
2871                          *
2872                          * The above condition could be true, if the SRTT
2873                          * and RTTVAR used to calculate t_rxtcur_prev
2874                          * resulted in a value less than t_rttmin.  So
2875                          * simply increasing SRTT by tcp_eifel_rtoinc when
2876                          * preparing for the Eifel response in
2877                          * tcp_save_congestion_state() could not ensure
2878                          * that the new RTO will be tcp_eifel_rtoinc greater
2879                          * t_rxtcur_prev.
2880                          */
2881                         tp->t_rxtcur = tp->t_rxtcur_prev + tcp_eifel_rtoinc;
2882                 }
2883 #ifdef DEBUG_EIFEL_RESPONSE
2884                 kprintf("new %d\n", tp->t_rxtcur);
2885 #endif
2886         }
2887
2888         /*
2889          * We received an ack for a packet that wasn't retransmitted;
2890          * it is probably safe to discard any error indications we've
2891          * received recently.  This isn't quite right, but close enough
2892          * for now (a route might have failed after we sent a segment,
2893          * and the return path might not be symmetrical).
2894          */
2895         tp->t_softerror = 0;
2896 }
2897
2898 /*
2899  * Determine a reasonable value for maxseg size.
2900  * If the route is known, check route for mtu.
2901  * If none, use an mss that can be handled on the outgoing
2902  * interface without forcing IP to fragment; if bigger than
2903  * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2904  * to utilize large mbufs.  If no route is found, route has no mtu,
2905  * or the destination isn't local, use a default, hopefully conservative
2906  * size (usually 512 or the default IP max size, but no more than the mtu
2907  * of the interface), as we can't discover anything about intervening
2908  * gateways or networks.  We also initialize the congestion/slow start
2909  * window to be a single segment if the destination isn't local.
2910  * While looking at the routing entry, we also initialize other path-dependent
2911  * parameters from pre-set or cached values in the routing entry.
2912  *
2913  * Also take into account the space needed for options that we
2914  * send regularly.  Make maxseg shorter by that amount to assure
2915  * that we can send maxseg amount of data even when the options
2916  * are present.  Store the upper limit of the length of options plus
2917  * data in maxopd.
2918  *
2919  * NOTE that this routine is only called when we process an incoming
2920  * segment, for outgoing segments only tcp_mssopt is called.
2921  */
2922 void
2923 tcp_mss(struct tcpcb *tp, int offer)
2924 {
2925         struct rtentry *rt;
2926         struct ifnet *ifp;
2927         int rtt, mss;
2928         u_long bufsize;
2929         struct inpcb *inp = tp->t_inpcb;
2930         struct socket *so;
2931 #ifdef INET6
2932         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2933         size_t min_protoh = isipv6 ?
2934                             sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
2935                             sizeof(struct tcpiphdr);
2936 #else
2937         const boolean_t isipv6 = FALSE;
2938         const size_t min_protoh = sizeof(struct tcpiphdr);
2939 #endif
2940
2941         if (isipv6)
2942                 rt = tcp_rtlookup6(&inp->inp_inc);
2943         else
2944                 rt = tcp_rtlookup(&inp->inp_inc);
2945         if (rt == NULL) {
2946                 tp->t_maxopd = tp->t_maxseg =
2947                     (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
2948                 return;
2949         }
2950         ifp = rt->rt_ifp;
2951         so = inp->inp_socket;
2952
2953         /*
2954          * Offer == 0 means that there was no MSS on the SYN segment,
2955          * in this case we use either the interface mtu or tcp_mssdflt.
2956          *
2957          * An offer which is too large will be cut down later.
2958          */
2959         if (offer == 0) {
2960                 if (isipv6) {
2961                         if (in6_localaddr(&inp->in6p_faddr)) {
2962                                 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
2963                                         min_protoh;
2964                         } else {
2965                                 offer = tcp_v6mssdflt;
2966                         }
2967                 } else {
2968                         if (in_localaddr(inp->inp_faddr))
2969                                 offer = ifp->if_mtu - min_protoh;
2970                         else
2971                                 offer = tcp_mssdflt;
2972                 }
2973         }
2974
2975         /*
2976          * Prevent DoS attack with too small MSS. Round up
2977          * to at least minmss.
2978          *
2979          * Sanity check: make sure that maxopd will be large
2980          * enough to allow some data on segments even is the
2981          * all the option space is used (40bytes).  Otherwise
2982          * funny things may happen in tcp_output.
2983          */
2984         offer = max(offer, tcp_minmss);
2985         offer = max(offer, 64);
2986
2987         rt->rt_rmx.rmx_mssopt = offer;
2988
2989         /*
2990          * While we're here, check if there's an initial rtt
2991          * or rttvar.  Convert from the route-table units
2992          * to scaled multiples of the slow timeout timer.
2993          */
2994         if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
2995                 /*
2996                  * XXX the lock bit for RTT indicates that the value
2997                  * is also a minimum value; this is subject to time.
2998                  */
2999                 if (rt->rt_rmx.rmx_locks & RTV_RTT)
3000                         tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
3001                 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
3002                 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
3003                 tcpstat.tcps_usedrtt++;
3004                 if (rt->rt_rmx.rmx_rttvar) {
3005                         tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
3006                             (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
3007                         tcpstat.tcps_usedrttvar++;
3008                 } else {
3009                         /* default variation is +- 1 rtt */
3010                         tp->t_rttvar =
3011                             tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
3012                 }
3013                 TCPT_RANGESET(tp->t_rxtcur,
3014                               ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
3015                               tp->t_rttmin, TCPTV_REXMTMAX);
3016         }
3017
3018         /*
3019          * if there's an mtu associated with the route, use it
3020          * else, use the link mtu.  Take the smaller of mss or offer
3021          * as our final mss.
3022          */
3023         if (rt->rt_rmx.rmx_mtu) {
3024                 mss = rt->rt_rmx.rmx_mtu - min_protoh;
3025         } else {
3026                 if (isipv6)
3027                         mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
3028                 else
3029                         mss = ifp->if_mtu - min_protoh;
3030         }
3031         mss = min(mss, offer);
3032
3033         /*
3034          * maxopd stores the maximum length of data AND options
3035          * in a segment; maxseg is the amount of data in a normal
3036          * segment.  We need to store this value (maxopd) apart
3037          * from maxseg, because now every segment carries options
3038          * and thus we normally have somewhat less data in segments.
3039          */
3040         tp->t_maxopd = mss;
3041
3042         if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
3043             ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3044                 mss -= TCPOLEN_TSTAMP_APPA;
3045
3046 #if     (MCLBYTES & (MCLBYTES - 1)) == 0
3047         if (mss > MCLBYTES)
3048                 mss &= ~(MCLBYTES-1);
3049 #else
3050         if (mss > MCLBYTES)
3051                 mss = mss / MCLBYTES * MCLBYTES;
3052 #endif
3053         /*
3054          * If there's a pipesize, change the socket buffer
3055          * to that size.  Make the socket buffers an integral
3056          * number of mss units; if the mss is larger than
3057          * the socket buffer, decrease the mss.
3058          */
3059 #ifdef RTV_SPIPE
3060         if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3061 #endif
3062                 bufsize = so->so_snd.ssb_hiwat;
3063         if (bufsize < mss)
3064                 mss = bufsize;
3065         else {
3066                 bufsize = roundup(bufsize, mss);
3067                 if (bufsize > sb_max)
3068                         bufsize = sb_max;
3069                 if (bufsize > so->so_snd.ssb_hiwat)
3070                         ssb_reserve(&so->so_snd, bufsize, so, NULL);
3071         }
3072         tp->t_maxseg = mss;
3073
3074 #ifdef RTV_RPIPE
3075         if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3076 #endif
3077                 bufsize = so->so_rcv.ssb_hiwat;
3078         if (bufsize > mss) {
3079                 bufsize = roundup(bufsize, mss);
3080                 if (bufsize > sb_max)
3081                         bufsize = sb_max;
3082                 if (bufsize > so->so_rcv.ssb_hiwat) {
3083                         lwkt_gettoken(&so->so_rcv.ssb_token);
3084                         ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3085                         lwkt_reltoken(&so->so_rcv.ssb_token);
3086                 }
3087         }
3088
3089         /*
3090          * Set the slow-start flight size
3091          *
3092          * NOTE: t_maxseg must have been configured!
3093          */
3094         tp->snd_cwnd = tcp_initial_window(tp);
3095
3096         if (rt->rt_rmx.rmx_ssthresh) {
3097                 /*
3098                  * There's some sort of gateway or interface
3099                  * buffer limit on the path.  Use this to set
3100                  * the slow start threshhold, but set the
3101                  * threshold to no less than 2*mss.
3102                  */
3103                 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3104                 tcpstat.tcps_usedssthresh++;
3105         }
3106 }
3107
3108 /*
3109  * Determine the MSS option to send on an outgoing SYN.
3110  */
3111 int
3112 tcp_mssopt(struct tcpcb *tp)
3113 {
3114         struct rtentry *rt;
3115 #ifdef INET6
3116         boolean_t isipv6 =
3117             ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3118         int min_protoh = isipv6 ?
3119                              sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3120                              sizeof(struct tcpiphdr);
3121 #else
3122         const boolean_t isipv6 = FALSE;
3123         const size_t min_protoh = sizeof(struct tcpiphdr);
3124 #endif
3125
3126         if (isipv6)
3127                 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3128         else
3129                 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3130         if (rt == NULL)
3131                 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3132
3133         return (rt->rt_ifp->if_mtu - min_protoh);
3134 }
3135
3136 /*
3137  * When a partial ack arrives, force the retransmission of the
3138  * next unacknowledged segment.  Do not exit Fast Recovery.
3139  *
3140  * Implement the Slow-but-Steady variant of NewReno by restarting the
3141  * the retransmission timer.  Turn it off here so it can be restarted
3142  * later in tcp_output().
3143  */
3144 static void
3145 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3146 {
3147         tcp_seq old_snd_nxt = tp->snd_nxt;
3148         u_long ocwnd = tp->snd_cwnd;
3149
3150         tcp_callout_stop(tp, tp->tt_rexmt);
3151         tp->t_rtttime = 0;
3152         tp->snd_nxt = th->th_ack;
3153         /* Set snd_cwnd to one segment beyond acknowledged offset. */
3154         tp->snd_cwnd = tp->t_maxseg;
3155         tp->t_flags |= TF_ACKNOW;
3156         tcp_output(tp);
3157         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3158                 tp->snd_nxt = old_snd_nxt;
3159         /* partial window deflation */
3160         if (ocwnd > acked)
3161                 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3162         else
3163                 tp->snd_cwnd = tp->t_maxseg;
3164 }
3165
3166 /*
3167  * In contrast to the Slow-but-Steady NewReno variant,
3168  * we do not reset the retransmission timer for SACK retransmissions,
3169  * except when retransmitting snd_una.
3170  */
3171 static void
3172 tcp_sack_rexmt(struct tcpcb *tp)
3173 {
3174         tcp_seq old_snd_nxt = tp->snd_nxt;
3175         u_long ocwnd = tp->snd_cwnd;
3176         uint32_t pipe;
3177         int nseg = 0;           /* consecutive new segments */
3178         int nseg_rexmt = 0;     /* retransmitted segments */
3179 #define MAXBURST 4              /* limit burst of new packets on partial ack */
3180
3181         tp->t_rtttime = 0;
3182         pipe = tcp_sack_compute_pipe(tp);
3183         while ((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg &&
3184             (!tcp_do_smartsack || nseg < MAXBURST)) {
3185                 tcp_seq old_snd_max, old_rexmt_high, nextrexmt;
3186                 uint32_t sent, seglen;
3187                 boolean_t rescue;
3188                 int error;
3189
3190                 old_rexmt_high = tp->rexmt_high;
3191                 if (!tcp_sack_nextseg(tp, &nextrexmt, &seglen, &rescue)) {
3192                         tp->rexmt_high = old_rexmt_high;
3193                         break;
3194                 }
3195
3196                 /*
3197                  * If the next tranmission is a rescue retranmission,
3198                  * we check whether we have already sent some data
3199                  * (either new segments or retransmitted segments)
3200                  * into the the network or not.  Since the idea of rescue
3201                  * retransmission is to sustain ACK clock, as long as
3202                  * some segments are in the network, ACK clock will be
3203                  * kept ticking.
3204                  */
3205                 if (rescue && (nseg_rexmt > 0 || nseg > 0)) {
3206                         tp->rexmt_high = old_rexmt_high;
3207                         break;
3208                 }
3209
3210                 if (nextrexmt == tp->snd_max)
3211                         ++nseg;
3212                 else
3213                         ++nseg_rexmt;
3214                 tp->snd_nxt = nextrexmt;
3215                 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3216                 old_snd_max = tp->snd_max;
3217                 if (nextrexmt == tp->snd_una)
3218                         tcp_callout_stop(tp, tp->tt_rexmt);
3219                 error = tcp_output(tp);
3220                 if (error != 0) {
3221                         tp->rexmt_high = old_rexmt_high;
3222                         break;
3223                 }
3224                 sent = tp->snd_nxt - nextrexmt;
3225                 if (sent <= 0) {
3226                         tp->rexmt_high = old_rexmt_high;
3227                         break;
3228                 }
3229                 pipe += sent;
3230                 tcpstat.tcps_sndsackpack++;
3231                 tcpstat.tcps_sndsackbyte += sent;
3232
3233                 if (rescue) {
3234                         tcpstat.tcps_sackrescue++;
3235                         tp->rexmt_rescue = tp->snd_nxt;
3236                         tp->sack_flags |= TSACK_F_SACKRESCUED;
3237                         break;
3238                 }
3239                 if (SEQ_LT(nextrexmt, old_snd_max) &&
3240                     SEQ_LT(tp->rexmt_high, tp->snd_nxt)) {
3241                         tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3242                         if (tcp_aggressive_rescuesack &&
3243                             (tp->sack_flags & TSACK_F_SACKRESCUED) &&
3244                             SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
3245                                 /* Drag RescueRxt along with HighRxt */
3246                                 tp->rexmt_rescue = tp->rexmt_high;
3247                         }
3248                 }
3249         }
3250         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3251                 tp->snd_nxt = old_snd_nxt;
3252         tp->snd_cwnd = ocwnd;
3253 }
3254
3255 /*
3256  * Return TRUE, if some new segments are sent
3257  */
3258 static boolean_t
3259 tcp_sack_limitedxmit(struct tcpcb *tp)
3260 {
3261         tcp_seq oldsndnxt = tp->snd_nxt;
3262         tcp_seq oldsndmax = tp->snd_max;
3263         u_long ocwnd = tp->snd_cwnd;
3264         uint32_t pipe, sent;
3265         boolean_t ret = FALSE;
3266         tcp_seq_diff_t cwnd_left;
3267         tcp_seq next;
3268
3269         tp->rexmt_high = tp->snd_una - 1;
3270         pipe = tcp_sack_compute_pipe(tp);
3271         cwnd_left = (tcp_seq_diff_t)(ocwnd - pipe);
3272         if (cwnd_left < (tcp_seq_diff_t)tp->t_maxseg)
3273                 return FALSE;
3274
3275         next = tp->snd_nxt = tp->snd_max;
3276         tp->snd_cwnd = tp->snd_nxt - tp->snd_una +
3277             rounddown(cwnd_left, tp->t_maxseg);
3278
3279         tcp_output(tp);
3280
3281         sent = tp->snd_nxt - next;
3282         if (sent > 0) {
3283                 tcpstat.tcps_sndlimited += howmany(sent, tp->t_maxseg);
3284                 ret = TRUE;
3285         }
3286
3287         if (SEQ_LT(oldsndnxt, oldsndmax)) {
3288                 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una),
3289                     ("snd_una moved in other threads"));
3290                 tp->snd_nxt = oldsndnxt;
3291         }
3292         tp->snd_cwnd = ocwnd;
3293
3294         if (ret && TCP_DO_NCR(tp))
3295                 tcp_ncr_update_rxtthresh(tp);
3296
3297         return ret;
3298 }
3299
3300 /*
3301  * Reset idle time and keep-alive timer, typically called when a valid
3302  * tcp packet is received but may also be called when FASTKEEP is set
3303  * to prevent the previous long-timeout from calculating to a drop.
3304  *
3305  * Only update t_rcvtime for non-SYN packets.
3306  *
3307  * Handle the case where one side thinks the connection is established
3308  * but the other side has, say, rebooted without cleaning out the
3309  * connection.   The SYNs could be construed as an attack and wind
3310  * up ignored, but in case it isn't an attack we can validate the
3311  * connection by forcing a keepalive.
3312  */
3313 void
3314 tcp_timer_keep_activity(struct tcpcb *tp, int thflags)
3315 {
3316         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
3317                 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
3318                         tp->t_flags |= TF_KEEPALIVE;
3319                         tcp_callout_reset(tp, tp->tt_keep, hz / 2,
3320                                           tcp_timer_keep);
3321                 } else {
3322                         tp->t_rcvtime = ticks;
3323                         tp->t_flags &= ~TF_KEEPALIVE;
3324                         tcp_callout_reset(tp, tp->tt_keep,
3325                                           tp->t_keepidle,
3326                                           tcp_timer_keep);
3327                 }
3328         }
3329 }
3330
3331 static int
3332 tcp_rmx_msl(const struct tcpcb *tp)
3333 {
3334         struct rtentry *rt;
3335         struct inpcb *inp = tp->t_inpcb;
3336         int msl;
3337 #ifdef INET6
3338         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3339 #else
3340         const boolean_t isipv6 = FALSE;
3341 #endif
3342
3343         if (isipv6)
3344                 rt = tcp_rtlookup6(&inp->inp_inc);
3345         else
3346                 rt = tcp_rtlookup(&inp->inp_inc);
3347         if (rt == NULL || rt->rt_rmx.rmx_msl == 0)
3348                 return tcp_msl;
3349
3350         msl = (rt->rt_rmx.rmx_msl * hz) / 1000;
3351         if (msl == 0)
3352                 msl = 1;
3353
3354         return msl;
3355 }
3356
3357 static void
3358 tcp_established(struct tcpcb *tp)
3359 {
3360         tp->t_state = TCPS_ESTABLISHED;
3361         tcp_callout_reset(tp, tp->tt_keep, tp->t_keepidle, tcp_timer_keep);
3362
3363         if (tp->t_rxtsyn > 0) {
3364                 /*
3365                  * RFC6298:
3366                  * "If the timer expires awaiting the ACK of a SYN segment
3367                  *  and the TCP implementation is using an RTO less than 3
3368                  *  seconds, the RTO MUST be re-initialized to 3 seconds
3369                  *  when data transmission begins"
3370                  */
3371                 if (tp->t_rxtcur < TCPTV_RTOBASE3)
3372                         tp->t_rxtcur = TCPTV_RTOBASE3;
3373         }
3374 }
3375
3376 /*
3377  * Returns TRUE, if the ACK should be dropped
3378  */
3379 static boolean_t
3380 tcp_fast_recovery(struct tcpcb *tp, tcp_seq th_ack, const struct tcpopt *to)
3381 {
3382         tcpstat.tcps_rcvdupack++;
3383
3384         /*
3385          * We have outstanding data (other than a window probe),
3386          * this is a completely duplicate ack (ie, window info
3387          * didn't change), the ack is the biggest we've seen and
3388          * we've seen exactly our rexmt threshhold of them, so
3389          * assume a packet has been dropped and retransmit it.
3390          * Kludge snd_nxt & the congestion window so we send only
3391          * this one packet.
3392          */
3393         if (IN_FASTRECOVERY(tp)) {
3394                 if (TCP_DO_SACK(tp)) {
3395                         /* No artifical cwnd inflation. */
3396                         tcp_sack_rexmt(tp);
3397                 } else {
3398                         /*
3399                          * Dup acks mean that packets have left
3400                          * the network (they're now cached at the
3401                          * receiver) so bump cwnd by the amount in
3402                          * the receiver to keep a constant cwnd
3403                          * packets in the network.
3404                          */
3405                         tp->snd_cwnd += tp->t_maxseg;
3406                         tcp_output(tp);
3407                 }
3408                 return TRUE;
3409         } else if (SEQ_LT(th_ack, tp->snd_recover)) {
3410                 tp->t_dupacks = 0;
3411                 return FALSE;
3412         } else if (tcp_ignore_redun_dsack && TCP_DO_SACK(tp) &&
3413             (to->to_flags & (TOF_DSACK | TOF_SACK_REDUNDANT)) ==
3414             (TOF_DSACK | TOF_SACK_REDUNDANT)) {
3415                 /*
3416                  * If the ACK carries DSACK and other SACK blocks
3417                  * carry information that we have already known,
3418                  * don't count this ACK as duplicate ACK.  This
3419                  * prevents spurious early retransmit and fast
3420                  * retransmit.  This also meets the requirement of
3421                  * RFC3042 that new segments should not be sent if
3422                  * the SACK blocks do not contain new information
3423                  * (XXX we actually loosen the requirment that only
3424                  * DSACK is checked here).
3425                  *
3426                  * This kind of ACKs are usually sent after spurious
3427                  * retransmit.
3428                  */
3429                 /* Do nothing; don't change t_dupacks */
3430                 return TRUE;
3431         } else if (tp->t_dupacks == 0 && TCP_DO_NCR(tp)) {
3432                 tcp_ncr_update_rxtthresh(tp);
3433         }
3434
3435         if (++tp->t_dupacks == tp->t_rxtthresh) {
3436                 tcp_seq old_snd_nxt;
3437                 u_int win;
3438
3439 fastretransmit:
3440                 if (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) {
3441                         tcp_save_congestion_state(tp);
3442                         tp->rxt_flags |= TRXT_F_FASTREXMT;
3443                 }
3444                 /*
3445                  * We know we're losing at the current window size,
3446                  * so do congestion avoidance: set ssthresh to half
3447                  * the current window and pull our congestion window
3448                  * back to the new ssthresh.
3449                  */
3450                 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_maxseg;
3451                 if (win < 2)
3452                         win = 2;
3453                 tp->snd_ssthresh = win * tp->t_maxseg;
3454                 ENTER_FASTRECOVERY(tp);
3455                 tp->snd_recover = tp->snd_max;
3456                 tcp_callout_stop(tp, tp->tt_rexmt);
3457                 tp->t_rtttime = 0;
3458                 old_snd_nxt = tp->snd_nxt;
3459                 tp->snd_nxt = th_ack;
3460                 tp->snd_cwnd = tp->t_maxseg;
3461                 tcp_output(tp);
3462                 ++tcpstat.tcps_sndfastrexmit;
3463                 tp->snd_cwnd = tp->snd_ssthresh;
3464                 tp->rexmt_high = tp->snd_nxt;
3465                 tp->sack_flags &= ~TSACK_F_SACKRESCUED;
3466                 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3467                         tp->snd_nxt = old_snd_nxt;
3468                 KASSERT(tp->snd_limited <= 2, ("tp->snd_limited too big"));
3469                 if (TCP_DO_SACK(tp)) {
3470                         tcp_sack_rexmt(tp);
3471                 } else {
3472                         tp->snd_cwnd += tp->t_maxseg *
3473                             (tp->t_dupacks - tp->snd_limited);
3474                 }
3475         } else if ((tcp_do_rfc3517bis && TCP_DO_SACK(tp)) || TCP_DO_NCR(tp)) {
3476                 if (tcp_rfc3517bis_rxt && tcp_do_rfc3517bis &&
3477                     tcp_sack_islost(&tp->scb, tp->snd_una))
3478                         goto fastretransmit;
3479                 if (tcp_do_limitedtransmit || TCP_DO_NCR(tp)) {
3480                         /* outstanding data */
3481                         uint32_t ownd = tp->snd_max - tp->snd_una;
3482
3483                         if (!tcp_sack_limitedxmit(tp) &&
3484                             need_early_retransmit(tp, ownd)) {
3485                                 ++tcpstat.tcps_sndearlyrexmit;
3486                                 tp->rxt_flags |= TRXT_F_EARLYREXMT;
3487                                 goto fastretransmit;
3488                         }
3489                 }
3490         } else if (tcp_do_limitedtransmit) {
3491                 u_long oldcwnd = tp->snd_cwnd;
3492                 tcp_seq oldsndmax = tp->snd_max;
3493                 tcp_seq oldsndnxt = tp->snd_nxt;
3494                 /* outstanding data */
3495                 uint32_t ownd = tp->snd_max - tp->snd_una;
3496                 u_int sent;
3497
3498                 KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2,
3499                     ("dupacks not 1 or 2"));
3500                 if (tp->t_dupacks == 1)
3501                         tp->snd_limited&