tcp: Limit new segements burst to 4 during extended limited transmit
[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 /*
138  * Limit burst of new packets during SACK based fast recovery
139  * or extended limited transmit.
140  */
141 #define TCP_SACK_MAXBURST       4
142
143 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
144
145 static int log_in_vain = 0;
146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW,
147     &log_in_vain, 0, "Log all incoming TCP connections");
148
149 static int blackhole = 0;
150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW,
151     &blackhole, 0, "Do not send RST when dropping refused connections");
152
153 int tcp_delack_enabled = 1;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW,
155     &tcp_delack_enabled, 0,
156     "Delay ACK to try and piggyback it onto a data packet");
157
158 #ifdef TCP_DROP_SYNFIN
159 static int drop_synfin = 0;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW,
161     &drop_synfin, 0, "Drop TCP packets with SYN+FIN set");
162 #endif
163
164 static int tcp_do_limitedtransmit = 1;
165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, limitedtransmit, CTLFLAG_RW,
166     &tcp_do_limitedtransmit, 0, "Enable RFC 3042 (Limited Transmit)");
167
168 static int tcp_do_early_retransmit = 1;
169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, earlyretransmit, CTLFLAG_RW,
170     &tcp_do_early_retransmit, 0, "Early retransmit");
171
172 int tcp_aggregate_acks = 1;
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, aggregate_acks, CTLFLAG_RW,
174     &tcp_aggregate_acks, 0, "Aggregate built-up acks into one ack");
175
176 static int tcp_do_eifel_detect = 1;
177 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
178     &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
179
180 static int tcp_do_abc = 1;
181 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
182     &tcp_do_abc, 0,
183     "TCP Appropriate Byte Counting (RFC 3465)");
184
185 /*
186  * The following value actually takes range [25ms, 250ms],
187  * given that most modern systems use 1ms ~ 10ms as the unit
188  * of timestamp option.
189  */
190 static u_int tcp_paws_tolerance = 25;
191 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, paws_tolerance, CTLFLAG_RW,
192     &tcp_paws_tolerance, 0, "RFC1323 PAWS tolerance");
193
194 /*
195  * Define as tunable for easy testing with SACK on and off.
196  * Warning:  do not change setting in the middle of an existing active TCP flow,
197  *   else strange things might happen to that flow.
198  */
199 int tcp_do_sack = 1;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
201     &tcp_do_sack, 0, "Enable SACK Algorithms");
202
203 int tcp_do_smartsack = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
205     &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
206
207 int tcp_do_rescuesack = 1;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack, CTLFLAG_RW,
209     &tcp_do_rescuesack, 0, "Rescue retransmission for SACK");
210
211 int tcp_aggressive_rescuesack = 0;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack_agg, CTLFLAG_RW,
213     &tcp_aggressive_rescuesack, 0, "Aggressive rescue retransmission for SACK");
214
215 static int tcp_force_sackrxt = 1;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, force_sackrxt, CTLFLAG_RW,
217     &tcp_force_sackrxt, 0, "Allowed forced SACK retransmit burst");
218
219 int tcp_do_rfc6675 = 1;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675, CTLFLAG_RW,
221     &tcp_do_rfc6675, 0, "Enable RFC6675");
222
223 int tcp_rfc6675_rxt = 0;
224 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_rxt, CTLFLAG_RW,
225     &tcp_rfc6675_rxt, 0, "Enable RFC6675 retransmit");
226
227 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
228     "TCP Segment Reassembly Queue");
229
230 int tcp_reass_maxseg = 0;
231 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
232     &tcp_reass_maxseg, 0,
233     "Global maximum number of TCP Segments in Reassembly Queue");
234
235 int tcp_reass_qsize = 0;
236 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
237     &tcp_reass_qsize, 0,
238     "Global number of TCP Segments currently in Reassembly Queue");
239
240 static int tcp_reass_overflows = 0;
241 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
242     &tcp_reass_overflows, 0,
243     "Global number of TCP Segment Reassembly Queue Overflows");
244
245 int tcp_do_autorcvbuf = 1;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
247     &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
248
249 int tcp_autorcvbuf_inc = 16*1024;
250 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
251     &tcp_autorcvbuf_inc, 0,
252     "Incrementor step size of automatic receive buffer");
253
254 int tcp_autorcvbuf_max = 2*1024*1024;
255 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
256     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
257
258 int tcp_sosend_agglim = 3;
259 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_agglim, CTLFLAG_RW,
260     &tcp_sosend_agglim, 0, "TCP sosend mbuf aggregation limit");
261
262 int tcp_sosend_async = 1;
263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_async, CTLFLAG_RW,
264     &tcp_sosend_async, 0, "TCP asynchronized pru_send");
265
266 static int tcp_ignore_redun_dsack = 1;
267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_redun_dsack, CTLFLAG_RW,
268     &tcp_ignore_redun_dsack, 0, "Ignore redundant DSACK");
269
270 static void      tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t,
271                     tcp_seq);
272 static void      tcp_pulloutofband(struct socket *,
273                      struct tcphdr *, struct mbuf *, int);
274 static int       tcp_reass(struct tcpcb *, struct tcphdr *, int *,
275                      struct mbuf *);
276 static void      tcp_xmit_timer(struct tcpcb *, int, tcp_seq);
277 static void      tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
278 static void      tcp_sack_rexmt(struct tcpcb *, boolean_t);
279 static boolean_t tcp_sack_limitedxmit(struct tcpcb *);
280 static int       tcp_rmx_msl(const struct tcpcb *);
281 static void      tcp_established(struct tcpcb *);
282 static boolean_t tcp_recv_dupack(struct tcpcb *, tcp_seq, u_int);
283
284 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
285 #ifdef INET6
286 #define ND6_HINT(tp) \
287 do { \
288         if ((tp) && (tp)->t_inpcb && \
289             ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
290             (tp)->t_inpcb->in6p_route.ro_rt) \
291                 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
292 } while (0)
293 #else
294 #define ND6_HINT(tp)
295 #endif
296
297 /*
298  * Indicate whether this ack should be delayed.  We can delay the ack if
299  *      - delayed acks are enabled and
300  *      - there is no delayed ack timer in progress and
301  *      - our last ack wasn't a 0-sized window.  We never want to delay
302  *        the ack that opens up a 0-sized window.
303  */
304 #define DELAY_ACK(tp) \
305         (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
306         !(tp->t_flags & TF_RXWIN0SENT))
307
308 #define acceptable_window_update(tp, th, tiwin)                         \
309     (SEQ_LT(tp->snd_wl1, th->th_seq) ||                                 \
310      (tp->snd_wl1 == th->th_seq &&                                      \
311       (SEQ_LT(tp->snd_wl2, th->th_ack) ||                               \
312        (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
313
314 #define iceildiv(n, d)          (((n)+(d)-1) / (d))
315 #define need_early_retransmit(tp, ownd) \
316     (tcp_do_early_retransmit && \
317      (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) && \
318      ownd < ((tp->t_rxtthresh + 1) * tp->t_maxseg) && \
319      tp->t_dupacks + 1 >= iceildiv(ownd, tp->t_maxseg) && \
320      (!TCP_DO_SACK(tp) || ownd <= tp->t_maxseg || \
321       tcp_sack_has_sacked(&tp->scb, ownd - tp->t_maxseg)))
322
323 /*
324  * Returns TRUE, if this segment can be merged with the last
325  * pending segment in the reassemble queue and this segment
326  * does not overlap with the pending segment immediately
327  * preceeding the last pending segment.
328  */
329 static __inline boolean_t
330 tcp_paws_canreasslast(const struct tcpcb *tp, const struct tcphdr *th, int tlen)
331 {
332         const struct tseg_qent *last, *prev;
333
334         last = TAILQ_LAST(&tp->t_segq, tsegqe_head);
335         if (last == NULL)
336                 return FALSE;
337
338         /* This segment comes immediately after the last pending segment */
339         if (last->tqe_th->th_seq + last->tqe_len == th->th_seq) {
340                 if (last->tqe_th->th_flags & TH_FIN) {
341                         /* No segments should follow segment w/ FIN */
342                         return FALSE;
343                 }
344                 return TRUE;
345         }
346
347         if (th->th_seq + tlen != last->tqe_th->th_seq)
348                 return FALSE;
349         /* This segment comes immediately before the last pending segment */
350
351         prev = TAILQ_PREV(last, tsegqe_head, tqe_q);
352         if (prev == NULL) {
353                 /*
354                  * No pending preceeding segment, we assume this segment
355                  * could be reassembled.
356                  */
357                 return TRUE;
358         }
359
360         /* This segment does not overlap with the preceeding segment */
361         if (SEQ_GEQ(th->th_seq, prev->tqe_th->th_seq + prev->tqe_len))
362                 return TRUE;
363
364         return FALSE;
365 }
366
367 static __inline void
368 tcp_ncr_update_rxtthresh(struct tcpcb *tp)
369 {
370         int old_rxtthresh = tp->t_rxtthresh;
371         uint32_t ownd = tp->snd_max - tp->snd_una;
372
373         tp->t_rxtthresh = max(tcprexmtthresh, ((ownd / tp->t_maxseg) >> 1));
374         if (tp->t_rxtthresh != old_rxtthresh) {
375                 tcp_sack_update_lostseq(&tp->scb, tp->snd_una,
376                     tp->t_maxseg, tp->t_rxtthresh);
377         }
378 }
379
380 static int
381 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
382 {
383         struct tseg_qent *q;
384         struct tseg_qent *p = NULL;
385         struct tseg_qent *te;
386         struct socket *so = tp->t_inpcb->inp_socket;
387         int flags;
388
389         /*
390          * Call with th == NULL after become established to
391          * force pre-ESTABLISHED data up to user socket.
392          */
393         if (th == NULL)
394                 goto present;
395
396         /*
397          * Limit the number of segments in the reassembly queue to prevent
398          * holding on to too many segments (and thus running out of mbufs).
399          * Make sure to let the missing segment through which caused this
400          * queue.  Always keep one global queue entry spare to be able to
401          * process the missing segment.
402          */
403         if (th->th_seq != tp->rcv_nxt &&
404             tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
405                 tcp_reass_overflows++;
406                 tcpstat.tcps_rcvmemdrop++;
407                 m_freem(m);
408                 /* no SACK block to report */
409                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
410                 return (0);
411         }
412
413         /* Allocate a new queue entry. */
414         te = kmalloc(sizeof(struct tseg_qent), M_TSEGQ, M_INTWAIT | M_NULLOK);
415         if (te == NULL) {
416                 tcpstat.tcps_rcvmemdrop++;
417                 m_freem(m);
418                 /* no SACK block to report */
419                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
420                 return (0);
421         }
422         atomic_add_int(&tcp_reass_qsize, 1);
423
424         if (th->th_flags & TH_FIN)
425                 tp->t_flags |= TF_QUEDFIN;
426
427         /*
428          * Find a segment which begins after this one does.
429          */
430         TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
431                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
432                         break;
433                 p = q;
434         }
435
436         /*
437          * If there is a preceding segment, it may provide some of
438          * our data already.  If so, drop the data from the incoming
439          * segment.  If it provides all of our data, drop us.
440          */
441         if (p != NULL) {
442                 tcp_seq_diff_t i;
443
444                 /* conversion to int (in i) handles seq wraparound */
445                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
446                 if (i > 0) {            /* overlaps preceding segment */
447                         tp->sack_flags |=
448                             (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
449                         /* enclosing block starts w/ preceding segment */
450                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
451                         if (i >= *tlenp) {
452                                 if (th->th_flags & TH_FIN)
453                                         p->tqe_th->th_flags |= TH_FIN;
454
455                                 /* preceding encloses incoming segment */
456                                 tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
457                                     p->tqe_th->th_seq + p->tqe_len,
458                                     p->tqe_th->th_flags);
459                                 tcpstat.tcps_rcvduppack++;
460                                 tcpstat.tcps_rcvdupbyte += *tlenp;
461                                 m_freem(m);
462                                 kfree(te, M_TSEGQ);
463                                 atomic_add_int(&tcp_reass_qsize, -1);
464                                 /*
465                                  * Try to present any queued data
466                                  * at the left window edge to the user.
467                                  * This is needed after the 3-WHS
468                                  * completes.
469                                  */
470                                 goto present;   /* ??? */
471                         }
472                         m_adj(m, i);
473                         *tlenp -= i;
474                         th->th_seq += i;
475                         /* incoming segment end is enclosing block end */
476                         tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
477                             th->th_seq + *tlenp, th->th_flags);
478                         /* trim end of reported D-SACK block */
479                         tp->reportblk.rblk_end = th->th_seq;
480                 }
481         }
482         tcpstat.tcps_rcvoopack++;
483         tcpstat.tcps_rcvoobyte += *tlenp;
484
485         /*
486          * While we overlap succeeding segments trim them or,
487          * if they are completely covered, dequeue them.
488          */
489         while (q) {
490                 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
491                 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
492                 tcp_seq qend_sack = TCP_SACK_BLKEND(qend, q->tqe_th->th_flags);
493                 struct tseg_qent *nq;
494
495                 if (i <= 0)
496                         break;
497                 if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
498                         /* first time through */
499                         tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
500                         tp->encloseblk = tp->reportblk;
501                         /* report trailing duplicate D-SACK segment */
502                         tp->reportblk.rblk_start = q->tqe_th->th_seq;
503                 }
504                 if ((tp->sack_flags & TSACK_F_ENCLOSESEG) &&
505                     SEQ_GT(qend_sack, tp->encloseblk.rblk_end)) {
506                         /* extend enclosing block if one exists */
507                         tp->encloseblk.rblk_end = qend_sack;
508                 }
509                 if (i < q->tqe_len) {
510                         q->tqe_th->th_seq += i;
511                         q->tqe_len -= i;
512                         m_adj(q->tqe_m, i);
513                         break;
514                 }
515
516                 if (q->tqe_th->th_flags & TH_FIN)
517                         th->th_flags |= TH_FIN;
518
519                 nq = TAILQ_NEXT(q, tqe_q);
520                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
521                 m_freem(q->tqe_m);
522                 kfree(q, M_TSEGQ);
523                 atomic_add_int(&tcp_reass_qsize, -1);
524                 q = nq;
525         }
526
527         /* Insert the new segment queue entry into place. */
528         te->tqe_m = m;
529         te->tqe_th = th;
530         te->tqe_len = *tlenp;
531
532         /* check if can coalesce with following segment */
533         if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
534                 tcp_seq tend_sack;
535
536                 te->tqe_len += q->tqe_len;
537                 if (q->tqe_th->th_flags & TH_FIN)
538                         te->tqe_th->th_flags |= TH_FIN;
539                 tend_sack = TCP_SACK_BLKEND(te->tqe_th->th_seq + te->tqe_len,
540                     te->tqe_th->th_flags);
541
542                 m_cat(te->tqe_m, q->tqe_m);
543                 tp->encloseblk.rblk_end = tend_sack;
544                 /*
545                  * When not reporting a duplicate segment, use
546                  * the larger enclosing block as the SACK block.
547                  */
548                 if (!(tp->sack_flags & TSACK_F_DUPSEG))
549                         tp->reportblk.rblk_end = tend_sack;
550                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
551                 kfree(q, M_TSEGQ);
552                 atomic_add_int(&tcp_reass_qsize, -1);
553         }
554
555         if (p == NULL) {
556                 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
557         } else {
558                 /* check if can coalesce with preceding segment */
559                 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
560                         if (te->tqe_th->th_flags & TH_FIN)
561                                 p->tqe_th->th_flags |= TH_FIN;
562                         p->tqe_len += te->tqe_len;
563                         m_cat(p->tqe_m, te->tqe_m);
564                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
565                         /*
566                          * When not reporting a duplicate segment, use
567                          * the larger enclosing block as the SACK block.
568                          */
569                         if (!(tp->sack_flags & TSACK_F_DUPSEG))
570                                 tp->reportblk.rblk_start = p->tqe_th->th_seq;
571                         kfree(te, M_TSEGQ);
572                         atomic_add_int(&tcp_reass_qsize, -1);
573                 } else {
574                         TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
575                 }
576         }
577
578 present:
579         /*
580          * Present data to user, advancing rcv_nxt through
581          * completed sequence space.
582          */
583         if (!TCPS_HAVEESTABLISHED(tp->t_state))
584                 return (0);
585         q = TAILQ_FIRST(&tp->t_segq);
586         if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
587                 return (0);
588         tp->rcv_nxt += q->tqe_len;
589         if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
590                 /* no SACK block to report since ACK advanced */
591                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
592         }
593         /* no enclosing block to report since ACK advanced */
594         tp->sack_flags &= ~TSACK_F_ENCLOSESEG;
595         flags = q->tqe_th->th_flags & TH_FIN;
596         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
597         KASSERT(TAILQ_EMPTY(&tp->t_segq) ||
598                 TAILQ_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
599                 ("segment not coalesced"));
600         if (so->so_state & SS_CANTRCVMORE) {
601                 m_freem(q->tqe_m);
602         } else {
603                 lwkt_gettoken(&so->so_rcv.ssb_token);
604                 ssb_appendstream(&so->so_rcv, q->tqe_m);
605                 lwkt_reltoken(&so->so_rcv.ssb_token);
606         }
607         kfree(q, M_TSEGQ);
608         atomic_add_int(&tcp_reass_qsize, -1);
609         ND6_HINT(tp);
610         sorwakeup(so);
611         return (flags);
612 }
613
614 /*
615  * TCP input routine, follows pages 65-76 of the
616  * protocol specification dated September, 1981 very closely.
617  */
618 #ifdef INET6
619 int
620 tcp6_input(struct mbuf **mp, int *offp, int proto)
621 {
622         struct mbuf *m = *mp;
623         struct in6_ifaddr *ia6;
624
625         IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
626
627         /*
628          * draft-itojun-ipv6-tcp-to-anycast
629          * better place to put this in?
630          */
631         ia6 = ip6_getdstifaddr(m);
632         if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
633                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
634                             offsetof(struct ip6_hdr, ip6_dst));
635                 return (IPPROTO_DONE);
636         }
637
638         tcp_input(mp, offp, proto);
639         return (IPPROTO_DONE);
640 }
641 #endif
642
643 int
644 tcp_input(struct mbuf **mp, int *offp, int proto)
645 {
646         int off0;
647         struct tcphdr *th;
648         struct ip *ip = NULL;
649         struct ipovly *ipov;
650         struct inpcb *inp = NULL;
651         u_char *optp = NULL;
652         int optlen = 0;
653         int tlen, off;
654         int len = 0;
655         int drop_hdrlen;
656         struct tcpcb *tp = NULL;
657         int thflags;
658         struct socket *so = NULL;
659         int todrop, acked;
660         boolean_t ourfinisacked, needoutput = FALSE, delayed_dupack = FALSE;
661         tcp_seq th_dupack = 0; /* XXX gcc warning */
662         u_int to_flags = 0; /* XXX gcc warning */
663         u_long tiwin;
664         int recvwin;
665         struct tcpopt to;               /* options in this segment */
666         struct sockaddr_in *next_hop = NULL;
667         int rstreason; /* For badport_bandlim accounting purposes */
668         int cpu;
669         struct ip6_hdr *ip6 = NULL;
670         struct mbuf *m;
671 #ifdef INET6
672         boolean_t isipv6;
673 #else
674         const boolean_t isipv6 = FALSE;
675 #endif
676 #ifdef TCPDEBUG
677         short ostate = 0;
678 #endif
679
680         off0 = *offp;
681         m = *mp;
682         *mp = NULL;
683
684         tcpstat.tcps_rcvtotal++;
685
686         if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
687                 struct m_tag *mtag;
688
689                 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
690                 KKASSERT(mtag != NULL);
691                 next_hop = m_tag_data(mtag);
692         }
693
694 #ifdef INET6
695         isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
696 #endif
697
698         if (isipv6) {
699                 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
700                 ip6 = mtod(m, struct ip6_hdr *);
701                 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
702                 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
703                         tcpstat.tcps_rcvbadsum++;
704                         goto drop;
705                 }
706                 th = (struct tcphdr *)((caddr_t)ip6 + off0);
707
708                 /*
709                  * Be proactive about unspecified IPv6 address in source.
710                  * As we use all-zero to indicate unbounded/unconnected pcb,
711                  * unspecified IPv6 address can be used to confuse us.
712                  *
713                  * Note that packets with unspecified IPv6 destination is
714                  * already dropped in ip6_input.
715                  */
716                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
717                         /* XXX stat */
718                         goto drop;
719                 }
720         } else {
721                 /*
722                  * Get IP and TCP header together in first mbuf.
723                  * Note: IP leaves IP header in first mbuf.
724                  */
725                 if (off0 > sizeof(struct ip)) {
726                         ip_stripoptions(m);
727                         off0 = sizeof(struct ip);
728                 }
729                 /* already checked and pulled up in ip_demux() */
730                 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
731                     ("TCP header not in one mbuf: m->m_len %d", m->m_len));
732                 ip = mtod(m, struct ip *);
733                 ipov = (struct ipovly *)ip;
734                 th = (struct tcphdr *)((caddr_t)ip + off0);
735                 tlen = ip->ip_len;
736
737                 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
738                         if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
739                                 th->th_sum = m->m_pkthdr.csum_data;
740                         else
741                                 th->th_sum = in_pseudo(ip->ip_src.s_addr,
742                                                 ip->ip_dst.s_addr,
743                                                 htonl(m->m_pkthdr.csum_data +
744                                                         ip->ip_len +
745                                                         IPPROTO_TCP));
746                         th->th_sum ^= 0xffff;
747                 } else {
748                         /*
749                          * Checksum extended TCP header and data.
750                          */
751                         len = sizeof(struct ip) + tlen;
752                         bzero(ipov->ih_x1, sizeof ipov->ih_x1);
753                         ipov->ih_len = (u_short)tlen;
754                         ipov->ih_len = htons(ipov->ih_len);
755                         th->th_sum = in_cksum(m, len);
756                 }
757                 if (th->th_sum) {
758                         tcpstat.tcps_rcvbadsum++;
759                         goto drop;
760                 }
761 #ifdef INET6
762                 /* Re-initialization for later version check */
763                 ip->ip_v = IPVERSION;
764 #endif
765         }
766
767         /*
768          * Check that TCP offset makes sense,
769          * pull out TCP options and adjust length.              XXX
770          */
771         off = th->th_off << 2;
772         /* already checked and pulled up in ip_demux() */
773         KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
774             ("bad TCP data offset %d (tlen %d)", off, tlen));
775         tlen -= off;    /* tlen is used instead of ti->ti_len */
776         if (off > sizeof(struct tcphdr)) {
777                 if (isipv6) {
778                         IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
779                         ip6 = mtod(m, struct ip6_hdr *);
780                         th = (struct tcphdr *)((caddr_t)ip6 + off0);
781                 } else {
782                         /* already pulled up in ip_demux() */
783                         KASSERT(m->m_len >= sizeof(struct ip) + off,
784                             ("TCP header and options not in one mbuf: "
785                              "m_len %d, off %d", m->m_len, off));
786                 }
787                 optlen = off - sizeof(struct tcphdr);
788                 optp = (u_char *)(th + 1);
789         }
790         thflags = th->th_flags;
791
792 #ifdef TCP_DROP_SYNFIN
793         /*
794          * If the drop_synfin option is enabled, drop all packets with
795          * both the SYN and FIN bits set. This prevents e.g. nmap from
796          * identifying the TCP/IP stack.
797          *
798          * This is a violation of the TCP specification.
799          */
800         if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
801                 goto drop;
802 #endif
803
804         /*
805          * Convert TCP protocol specific fields to host format.
806          */
807         th->th_seq = ntohl(th->th_seq);
808         th->th_ack = ntohl(th->th_ack);
809         th->th_win = ntohs(th->th_win);
810         th->th_urp = ntohs(th->th_urp);
811
812         /*
813          * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
814          * until after ip6_savecontrol() is called and before other functions
815          * which don't want those proto headers.
816          * Because ip6_savecontrol() is going to parse the mbuf to
817          * search for data to be passed up to user-land, it wants mbuf
818          * parameters to be unchanged.
819          * XXX: the call of ip6_savecontrol() has been obsoleted based on
820          * latest version of the advanced API (20020110).
821          */
822         drop_hdrlen = off0 + off;
823
824         /*
825          * Locate pcb for segment.
826          */
827 findpcb:
828         /* IPFIREWALL_FORWARD section */
829         if (next_hop != NULL && !isipv6) {  /* IPv6 support is not there yet */
830                 /*
831                  * Transparently forwarded. Pretend to be the destination.
832                  * already got one like this?
833                  */
834                 cpu = mycpu->gd_cpuid;
835                 inp = in_pcblookup_hash(&tcbinfo[cpu],
836                                         ip->ip_src, th->th_sport,
837                                         ip->ip_dst, th->th_dport,
838                                         0, m->m_pkthdr.rcvif);
839                 if (!inp) {
840                         /*
841                          * It's new.  Try to find the ambushing socket.
842                          */
843
844                         /*
845                          * The rest of the ipfw code stores the port in
846                          * host order.  XXX
847                          * (The IP address is still in network order.)
848                          */
849                         in_port_t dport = next_hop->sin_port ?
850                                                 htons(next_hop->sin_port) :
851                                                 th->th_dport;
852
853                         cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
854                                           next_hop->sin_addr.s_addr, dport);
855                         inp = in_pcblookup_hash(&tcbinfo[cpu],
856                                                 ip->ip_src, th->th_sport,
857                                                 next_hop->sin_addr, dport,
858                                                 1, m->m_pkthdr.rcvif);
859                 }
860         } else {
861                 if (isipv6) {
862                         inp = in6_pcblookup_hash(&tcbinfo[0],
863                                                  &ip6->ip6_src, th->th_sport,
864                                                  &ip6->ip6_dst, th->th_dport,
865                                                  1, m->m_pkthdr.rcvif);
866                 } else {
867                         cpu = mycpu->gd_cpuid;
868                         inp = in_pcblookup_hash(&tcbinfo[cpu],
869                                                 ip->ip_src, th->th_sport,
870                                                 ip->ip_dst, th->th_dport,
871                                                 1, m->m_pkthdr.rcvif);
872                 }
873         }
874
875         /*
876          * If the state is CLOSED (i.e., TCB does not exist) then
877          * all data in the incoming segment is discarded.
878          * If the TCB exists but is in CLOSED state, it is embryonic,
879          * but should either do a listen or a connect soon.
880          */
881         if (inp == NULL) {
882                 if (log_in_vain) {
883 #ifdef INET6
884                         char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
885 #else
886                         char dbuf[sizeof "aaa.bbb.ccc.ddd"];
887                         char sbuf[sizeof "aaa.bbb.ccc.ddd"];
888 #endif
889                         if (isipv6) {
890                                 strcpy(dbuf, "[");
891                                 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
892                                 strcat(dbuf, "]");
893                                 strcpy(sbuf, "[");
894                                 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
895                                 strcat(sbuf, "]");
896                         } else {
897                                 strcpy(dbuf, inet_ntoa(ip->ip_dst));
898                                 strcpy(sbuf, inet_ntoa(ip->ip_src));
899                         }
900                         switch (log_in_vain) {
901                         case 1:
902                                 if (!(thflags & TH_SYN))
903                                         break;
904                         case 2:
905                                 log(LOG_INFO,
906                                     "Connection attempt to TCP %s:%d "
907                                     "from %s:%d flags:0x%02x\n",
908                                     dbuf, ntohs(th->th_dport), sbuf,
909                                     ntohs(th->th_sport), thflags);
910                                 break;
911                         default:
912                                 break;
913                         }
914                 }
915                 if (blackhole) {
916                         switch (blackhole) {
917                         case 1:
918                                 if (thflags & TH_SYN)
919                                         goto drop;
920                                 break;
921                         case 2:
922                                 goto drop;
923                         default:
924                                 goto drop;
925                         }
926                 }
927                 rstreason = BANDLIM_RST_CLOSEDPORT;
928                 goto dropwithreset;
929         }
930
931 #ifdef IPSEC
932         if (isipv6) {
933                 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
934                         ipsec6stat.in_polvio++;
935                         goto drop;
936                 }
937         } else {
938                 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
939                         ipsecstat.in_polvio++;
940                         goto drop;
941                 }
942         }
943 #endif
944 #ifdef FAST_IPSEC
945         if (isipv6) {
946                 if (ipsec6_in_reject(m, inp))
947                         goto drop;
948         } else {
949                 if (ipsec4_in_reject(m, inp))
950                         goto drop;
951         }
952 #endif
953         /* Check the minimum TTL for socket. */
954 #ifdef INET6
955         if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
956                 goto drop;
957 #endif
958
959         tp = intotcpcb(inp);
960         if (tp == NULL) {
961                 rstreason = BANDLIM_RST_CLOSEDPORT;
962                 goto dropwithreset;
963         }
964         if (tp->t_state <= TCPS_CLOSED)
965                 goto drop;
966
967         so = inp->inp_socket;
968
969 #ifdef TCPDEBUG
970         if (so->so_options & SO_DEBUG) {
971                 ostate = tp->t_state;
972                 if (isipv6)
973                         bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
974                 else
975                         bcopy(ip, tcp_saveipgen, sizeof(*ip));
976                 tcp_savetcp = *th;
977         }
978 #endif
979
980         bzero(&to, sizeof to);
981
982         if (so->so_options & SO_ACCEPTCONN) {
983                 struct in_conninfo inc;
984
985 #ifdef INET6
986                 inc.inc_isipv6 = (isipv6 == TRUE);
987 #endif
988                 if (isipv6) {
989                         inc.inc6_faddr = ip6->ip6_src;
990                         inc.inc6_laddr = ip6->ip6_dst;
991                         inc.inc6_route.ro_rt = NULL;            /* XXX */
992                 } else {
993                         inc.inc_faddr = ip->ip_src;
994                         inc.inc_laddr = ip->ip_dst;
995                         inc.inc_route.ro_rt = NULL;             /* XXX */
996                 }
997                 inc.inc_fport = th->th_sport;
998                 inc.inc_lport = th->th_dport;
999
1000                 /*
1001                  * If the state is LISTEN then ignore segment if it contains
1002                  * a RST.  If the segment contains an ACK then it is bad and
1003                  * send a RST.  If it does not contain a SYN then it is not
1004                  * interesting; drop it.
1005                  *
1006                  * If the state is SYN_RECEIVED (syncache) and seg contains
1007                  * an ACK, but not for our SYN/ACK, send a RST.  If the seg
1008                  * contains a RST, check the sequence number to see if it
1009                  * is a valid reset segment.
1010                  */
1011                 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
1012                         if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
1013                                 if (!syncache_expand(&inc, th, &so, m)) {
1014                                         /*
1015                                          * No syncache entry, or ACK was not
1016                                          * for our SYN/ACK.  Send a RST.
1017                                          */
1018                                         tcpstat.tcps_badsyn++;
1019                                         rstreason = BANDLIM_RST_OPENPORT;
1020                                         goto dropwithreset;
1021                                 }
1022
1023                                 /*
1024                                  * Could not complete 3-way handshake,
1025                                  * connection is being closed down, and
1026                                  * syncache will free mbuf.
1027                                  */
1028                                 if (so == NULL)
1029                                         return(IPPROTO_DONE);
1030
1031                                 /*
1032                                  * We must be in the correct protocol thread
1033                                  * for this connection.
1034                                  */
1035                                 KKASSERT(so->so_port == &curthread->td_msgport);
1036
1037                                 /*
1038                                  * Socket is created in state SYN_RECEIVED.
1039                                  * Continue processing segment.
1040                                  */
1041                                 inp = so->so_pcb;
1042                                 tp = intotcpcb(inp);
1043                                 /*
1044                                  * This is what would have happened in
1045                                  * tcp_output() when the SYN,ACK was sent.
1046                                  */
1047                                 tp->snd_up = tp->snd_una;
1048                                 tp->snd_max = tp->snd_nxt = tp->iss + 1;
1049                                 tp->last_ack_sent = tp->rcv_nxt;
1050
1051                                 goto after_listen;
1052                         }
1053                         if (thflags & TH_RST) {
1054                                 syncache_chkrst(&inc, th);
1055                                 goto drop;
1056                         }
1057                         if (thflags & TH_ACK) {
1058                                 syncache_badack(&inc);
1059                                 tcpstat.tcps_badsyn++;
1060                                 rstreason = BANDLIM_RST_OPENPORT;
1061                                 goto dropwithreset;
1062                         }
1063                         goto drop;
1064                 }
1065
1066                 /*
1067                  * Segment's flags are (SYN) or (SYN | FIN).
1068                  */
1069 #ifdef INET6
1070                 /*
1071                  * If deprecated address is forbidden,
1072                  * we do not accept SYN to deprecated interface
1073                  * address to prevent any new inbound connection from
1074                  * getting established.
1075                  * When we do not accept SYN, we send a TCP RST,
1076                  * with deprecated source address (instead of dropping
1077                  * it).  We compromise it as it is much better for peer
1078                  * to send a RST, and RST will be the final packet
1079                  * for the exchange.
1080                  *
1081                  * If we do not forbid deprecated addresses, we accept
1082                  * the SYN packet.  RFC2462 does not suggest dropping
1083                  * SYN in this case.
1084                  * If we decipher RFC2462 5.5.4, it says like this:
1085                  * 1. use of deprecated addr with existing
1086                  *    communication is okay - "SHOULD continue to be
1087                  *    used"
1088                  * 2. use of it with new communication:
1089                  *   (2a) "SHOULD NOT be used if alternate address
1090                  *        with sufficient scope is available"
1091                  *   (2b) nothing mentioned otherwise.
1092                  * Here we fall into (2b) case as we have no choice in
1093                  * our source address selection - we must obey the peer.
1094                  *
1095                  * The wording in RFC2462 is confusing, and there are
1096                  * multiple description text for deprecated address
1097                  * handling - worse, they are not exactly the same.
1098                  * I believe 5.5.4 is the best one, so we follow 5.5.4.
1099                  */
1100                 if (isipv6 && !ip6_use_deprecated) {
1101                         struct in6_ifaddr *ia6;
1102
1103                         if ((ia6 = ip6_getdstifaddr(m)) &&
1104                             (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
1105                                 tp = NULL;
1106                                 rstreason = BANDLIM_RST_OPENPORT;
1107                                 goto dropwithreset;
1108                         }
1109                 }
1110 #endif
1111                 /*
1112                  * If it is from this socket, drop it, it must be forged.
1113                  * Don't bother responding if the destination was a broadcast.
1114                  */
1115                 if (th->th_dport == th->th_sport) {
1116                         if (isipv6) {
1117                                 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1118                                                        &ip6->ip6_src))
1119                                         goto drop;
1120                         } else {
1121                                 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1122                                         goto drop;
1123                         }
1124                 }
1125                 /*
1126                  * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1127                  *
1128                  * Note that it is quite possible to receive unicast
1129                  * link-layer packets with a broadcast IP address. Use
1130                  * in_broadcast() to find them.
1131                  */
1132                 if (m->m_flags & (M_BCAST | M_MCAST))
1133                         goto drop;
1134                 if (isipv6) {
1135                         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1136                             IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1137                                 goto drop;
1138                 } else {
1139                         if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1140                             IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1141                             ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1142                             in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1143                                 goto drop;
1144                 }
1145                 /*
1146                  * SYN appears to be valid; create compressed TCP state
1147                  * for syncache, or perform t/tcp connection.
1148                  */
1149                 if (so->so_qlen <= so->so_qlimit) {
1150                         tcp_dooptions(&to, optp, optlen, TRUE, th->th_ack);
1151                         if (!syncache_add(&inc, &to, th, so, m))
1152                                 goto drop;
1153
1154                         /*
1155                          * Entry added to syncache, mbuf used to
1156                          * send SYN,ACK packet.
1157                          */
1158                         return(IPPROTO_DONE);
1159                 }
1160                 goto drop;
1161         }
1162
1163 after_listen:
1164         /*
1165          * Should not happen - syncache should pick up these connections.
1166          *
1167          * Once we are past handling listen sockets we must be in the
1168          * correct protocol processing thread.
1169          */
1170         KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1171         KKASSERT(so->so_port == &curthread->td_msgport);
1172
1173         /* Unscale the window into a 32-bit value. */
1174         if (!(thflags & TH_SYN))
1175                 tiwin = th->th_win << tp->snd_scale;
1176         else
1177                 tiwin = th->th_win;
1178
1179         /*
1180          * This is the second part of the MSS DoS prevention code (after
1181          * minmss on the sending side) and it deals with too many too small
1182          * tcp packets in a too short timeframe (1 second).
1183          *
1184          * XXX Removed.  This code was crap.  It does not scale to network
1185          *     speed, and default values break NFS.  Gone.
1186          */
1187         /* REMOVED */
1188
1189         /*
1190          * Segment received on connection.
1191          *
1192          * Reset idle time and keep-alive timer.  Don't waste time if less
1193          * then a second has elapsed.
1194          */
1195         if ((int)(ticks - tp->t_rcvtime) > hz)
1196                 tcp_timer_keep_activity(tp, thflags);
1197
1198         /*
1199          * Process options.
1200          * XXX this is tradtitional behavior, may need to be cleaned up.
1201          */
1202         tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0, th->th_ack);
1203         if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1204                 if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) {
1205                         tp->t_flags |= TF_RCVD_SCALE;
1206                         tp->snd_scale = to.to_requested_s_scale;
1207                 }
1208
1209                 /*
1210                  * Initial send window; will be updated upon next ACK
1211                  */
1212                 tp->snd_wnd = th->th_win;
1213
1214                 if (to.to_flags & TOF_TS) {
1215                         tp->t_flags |= TF_RCVD_TSTMP;
1216                         tp->ts_recent = to.to_tsval;
1217                         tp->ts_recent_age = ticks;
1218                 }
1219                 if (!(to.to_flags & TOF_MSS))
1220                         to.to_mss = 0;
1221                 tcp_mss(tp, to.to_mss);
1222                 /*
1223                  * Only set the TF_SACK_PERMITTED per-connection flag
1224                  * if we got a SACK_PERMITTED option from the other side
1225                  * and the global tcp_do_sack variable is true.
1226                  */
1227                 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1228                         tp->t_flags |= TF_SACK_PERMITTED;
1229         }
1230
1231         /*
1232          * Header prediction: check for the two common cases
1233          * of a uni-directional data xfer.  If the packet has
1234          * no control flags, is in-sequence, the window didn't
1235          * change and we're not retransmitting, it's a
1236          * candidate.  If the length is zero and the ack moved
1237          * forward, we're the sender side of the xfer.  Just
1238          * free the data acked & wake any higher level process
1239          * that was blocked waiting for space.  If the length
1240          * is non-zero and the ack didn't move, we're the
1241          * receiver side.  If we're getting packets in-order
1242          * (the reassembly queue is empty), add the data to
1243          * the socket buffer and note that we need a delayed ack.
1244          * Make sure that the hidden state-flags are also off.
1245          * Since we check for TCPS_ESTABLISHED above, it can only
1246          * be TH_NEEDSYN.
1247          */
1248         if (tp->t_state == TCPS_ESTABLISHED &&
1249             (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1250             !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1251             (!(to.to_flags & TOF_TS) ||
1252              TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1253             th->th_seq == tp->rcv_nxt &&
1254             tp->snd_nxt == tp->snd_max) {
1255
1256                 /*
1257                  * If last ACK falls within this segment's sequence numbers,
1258                  * record the timestamp.
1259                  * NOTE that the test is modified according to the latest
1260                  * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1261                  */
1262                 if ((to.to_flags & TOF_TS) &&
1263                     SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1264                         tp->ts_recent_age = ticks;
1265                         tp->ts_recent = to.to_tsval;
1266                 }
1267
1268                 if (tlen == 0) {
1269                         if (SEQ_GT(th->th_ack, tp->snd_una) &&
1270                             SEQ_LEQ(th->th_ack, tp->snd_max) &&
1271                             tp->snd_cwnd >= tp->snd_wnd &&
1272                             !IN_FASTRECOVERY(tp)) {
1273                                 /*
1274                                  * This is a pure ack for outstanding data.
1275                                  */
1276                                 ++tcpstat.tcps_predack;
1277                                 /*
1278                                  * "bad retransmit" recovery
1279                                  *
1280                                  * If Eifel detection applies, then
1281                                  * it is deterministic, so use it
1282                                  * unconditionally over the old heuristic.
1283                                  * Otherwise, fall back to the old heuristic.
1284                                  */
1285                                 if (tcp_do_eifel_detect &&
1286                                     (to.to_flags & TOF_TS) && to.to_tsecr &&
1287                                     (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
1288                                         /* Eifel detection applicable. */
1289                                         if (to.to_tsecr < tp->t_rexmtTS) {
1290                                                 tcp_revert_congestion_state(tp);
1291                                                 ++tcpstat.tcps_eifeldetected;
1292                                                 if (tp->t_rxtshift != 1 ||
1293                                                     ticks >= tp->t_badrxtwin)
1294                                                         ++tcpstat.tcps_rttcantdetect;
1295                                         }
1296                                 } else if (tp->t_rxtshift == 1 &&
1297                                            ticks < tp->t_badrxtwin) {
1298                                         tcp_revert_congestion_state(tp);
1299                                         ++tcpstat.tcps_rttdetected;
1300                                 }
1301                                 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
1302                                     TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
1303                                 /*
1304                                  * Recalculate the retransmit timer / rtt.
1305                                  *
1306                                  * Some machines (certain windows boxes)
1307                                  * send broken timestamp replies during the
1308                                  * SYN+ACK phase, ignore timestamps of 0.
1309                                  */
1310                                 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1311                                         tcp_xmit_timer(tp,
1312                                             ticks - to.to_tsecr + 1,
1313                                             th->th_ack);
1314                                 } else if (tp->t_rtttime &&
1315                                            SEQ_GT(th->th_ack, tp->t_rtseq)) {
1316                                         tcp_xmit_timer(tp,
1317                                             ticks - tp->t_rtttime,
1318                                             th->th_ack);
1319                                 }
1320                                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1321                                 acked = th->th_ack - tp->snd_una;
1322                                 tcpstat.tcps_rcvackpack++;
1323                                 tcpstat.tcps_rcvackbyte += acked;
1324                                 sbdrop(&so->so_snd.sb, acked);
1325                                 tp->snd_recover = th->th_ack - 1;
1326                                 tp->snd_una = th->th_ack;
1327                                 tp->t_dupacks = 0;
1328                                 /*
1329                                  * Update window information.
1330                                  */
1331                                 if (tiwin != tp->snd_wnd &&
1332                                     acceptable_window_update(tp, th, tiwin)) {
1333                                         /* keep track of pure window updates */
1334                                         if (tp->snd_wl2 == th->th_ack &&
1335                                             tiwin > tp->snd_wnd)
1336                                                 tcpstat.tcps_rcvwinupd++;
1337                                         tp->snd_wnd = tiwin;
1338                                         tp->snd_wl1 = th->th_seq;
1339                                         tp->snd_wl2 = th->th_ack;
1340                                         if (tp->snd_wnd > tp->max_sndwnd)
1341                                                 tp->max_sndwnd = tp->snd_wnd;
1342                                 }
1343                                 m_freem(m);
1344                                 ND6_HINT(tp); /* some progress has been done */
1345                                 /*
1346                                  * If all outstanding data are acked, stop
1347                                  * retransmit timer, otherwise restart timer
1348                                  * using current (possibly backed-off) value.
1349                                  * If process is waiting for space,
1350                                  * wakeup/selwakeup/signal.  If data
1351                                  * are ready to send, let tcp_output
1352                                  * decide between more output or persist.
1353                                  */
1354                                 if (tp->snd_una == tp->snd_max) {
1355                                         tcp_callout_stop(tp, tp->tt_rexmt);
1356                                 } else if (!tcp_callout_active(tp,
1357                                             tp->tt_persist)) {
1358                                         tcp_callout_reset(tp, tp->tt_rexmt,
1359                                             tp->t_rxtcur, tcp_timer_rexmt);
1360                                 }
1361                                 sowwakeup(so);
1362                                 if (so->so_snd.ssb_cc > 0 &&
1363                                     !tcp_output_pending(tp))
1364                                         tcp_output_fair(tp);
1365                                 return(IPPROTO_DONE);
1366                         }
1367                 } else if (tiwin == tp->snd_wnd &&
1368                     th->th_ack == tp->snd_una &&
1369                     TAILQ_EMPTY(&tp->t_segq) &&
1370                     tlen <= ssb_space(&so->so_rcv)) {
1371                         u_long newsize = 0;     /* automatic sockbuf scaling */
1372                         /*
1373                          * This is a pure, in-sequence data packet
1374                          * with nothing on the reassembly queue and
1375                          * we have enough buffer space to take it.
1376                          */
1377                         ++tcpstat.tcps_preddat;
1378                         tp->rcv_nxt += tlen;
1379                         tcpstat.tcps_rcvpack++;
1380                         tcpstat.tcps_rcvbyte += tlen;
1381                         ND6_HINT(tp);   /* some progress has been done */
1382                 /*
1383                  * Automatic sizing of receive socket buffer.  Often the send
1384                  * buffer size is not optimally adjusted to the actual network
1385                  * conditions at hand (delay bandwidth product).  Setting the
1386                  * buffer size too small limits throughput on links with high
1387                  * bandwidth and high delay (eg. trans-continental/oceanic links).
1388                  *
1389                  * On the receive side the socket buffer memory is only rarely
1390                  * used to any significant extent.  This allows us to be much
1391                  * more aggressive in scaling the receive socket buffer.  For
1392                  * the case that the buffer space is actually used to a large
1393                  * extent and we run out of kernel memory we can simply drop
1394                  * the new segments; TCP on the sender will just retransmit it
1395                  * later.  Setting the buffer size too big may only consume too
1396                  * much kernel memory if the application doesn't read() from
1397                  * the socket or packet loss or reordering makes use of the
1398                  * reassembly queue.
1399                  *
1400                  * The criteria to step up the receive buffer one notch are:
1401                  *  1. the number of bytes received during the time it takes
1402                  *     one timestamp to be reflected back to us (the RTT);
1403                  *  2. received bytes per RTT is within seven eighth of the
1404                  *     current socket buffer size;
1405                  *  3. receive buffer size has not hit maximal automatic size;
1406                  *
1407                  * This algorithm does one step per RTT at most and only if
1408                  * we receive a bulk stream w/o packet losses or reorderings.
1409                  * Shrinking the buffer during idle times is not necessary as
1410                  * it doesn't consume any memory when idle.
1411                  *
1412                  * TODO: Only step up if the application is actually serving
1413                  * the buffer to better manage the socket buffer resources.
1414                  */
1415                         if (tcp_do_autorcvbuf &&
1416                             to.to_tsecr &&
1417                             (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1418                                 if (to.to_tsecr > tp->rfbuf_ts &&
1419                                     to.to_tsecr - tp->rfbuf_ts < hz) {
1420                                         if (tp->rfbuf_cnt >
1421                                             (so->so_rcv.ssb_hiwat / 8 * 7) &&
1422                                             so->so_rcv.ssb_hiwat <
1423                                             tcp_autorcvbuf_max) {
1424                                                 newsize =
1425                                                     ulmin(so->so_rcv.ssb_hiwat +
1426                                                           tcp_autorcvbuf_inc,
1427                                                           tcp_autorcvbuf_max);
1428                                         }
1429                                         /* Start over with next RTT. */
1430                                         tp->rfbuf_ts = 0;
1431                                         tp->rfbuf_cnt = 0;
1432                                 } else
1433                                         tp->rfbuf_cnt += tlen;  /* add up */
1434                         }
1435                         /*
1436                          * Add data to socket buffer.
1437                          */
1438                         if (so->so_state & SS_CANTRCVMORE) {
1439                                 m_freem(m);
1440                         } else {
1441                                 /*
1442                                  * Set new socket buffer size, give up when
1443                                  * limit is reached.
1444                                  *
1445                                  * Adjusting the size can mess up ACK
1446                                  * sequencing when pure window updates are
1447                                  * being avoided (which is the default),
1448                                  * so force an ack.
1449                                  */
1450                                 lwkt_gettoken(&so->so_rcv.ssb_token);
1451                                 if (newsize) {
1452                                         tp->t_flags |= TF_RXRESIZED;
1453                                         if (!ssb_reserve(&so->so_rcv, newsize,
1454                                                          so, NULL)) {
1455                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1456                                         }
1457                                         if (newsize >=
1458                                             (TCP_MAXWIN << tp->rcv_scale)) {
1459                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1460                                         }
1461                                 }
1462                                 m_adj(m, drop_hdrlen); /* delayed header drop */
1463                                 ssb_appendstream(&so->so_rcv, m);
1464                                 lwkt_reltoken(&so->so_rcv.ssb_token);
1465                         }
1466                         sorwakeup(so);
1467                         /*
1468                          * This code is responsible for most of the ACKs
1469                          * the TCP stack sends back after receiving a data
1470                          * packet.  Note that the DELAY_ACK check fails if
1471                          * the delack timer is already running, which results
1472                          * in an ack being sent every other packet (which is
1473                          * what we want).
1474                          *
1475                          * We then further aggregate acks by not actually
1476                          * sending one until the protocol thread has completed
1477                          * processing the current backlog of packets.  This
1478                          * does not delay the ack any further, but allows us
1479                          * to take advantage of the packet aggregation that
1480                          * high speed NICs do (usually blocks of 8-10 packets)
1481                          * to send a single ack rather then four or five acks,
1482                          * greatly reducing the ack rate, the return channel
1483                          * bandwidth, and the protocol overhead on both ends.
1484                          *
1485                          * Since this also has the effect of slowing down
1486                          * the exponential slow-start ramp-up, systems with 
1487                          * very large bandwidth-delay products might want
1488                          * to turn the feature off.
1489                          */
1490                         if (DELAY_ACK(tp)) {
1491                                 tcp_callout_reset(tp, tp->tt_delack,
1492                                     tcp_delacktime, tcp_timer_delack);
1493                         } else if (tcp_aggregate_acks) {
1494                                 tp->t_flags |= TF_ACKNOW;
1495                                 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1496                                         tp->t_flags |= TF_ONOUTPUTQ;
1497                                         tp->tt_cpu = mycpu->gd_cpuid;
1498                                         TAILQ_INSERT_TAIL(
1499                                             &tcpcbackq[tp->tt_cpu],
1500                                             tp, t_outputq);
1501                                 }
1502                         } else {
1503                                 tp->t_flags |= TF_ACKNOW;
1504                                 tcp_output(tp);
1505                         }
1506                         return(IPPROTO_DONE);
1507                 }
1508         }
1509
1510         /*
1511          * Calculate amount of space in receive window,
1512          * and then do TCP input processing.
1513          * Receive window is amount of space in rcv queue,
1514          * but not less than advertised window.
1515          */
1516         recvwin = ssb_space(&so->so_rcv);
1517         if (recvwin < 0)
1518                 recvwin = 0;
1519         tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1520
1521         /* Reset receive buffer auto scaling when not in bulk receive mode. */
1522         tp->rfbuf_ts = 0;
1523         tp->rfbuf_cnt = 0;
1524
1525         switch (tp->t_state) {
1526         /*
1527          * If the state is SYN_RECEIVED:
1528          *      if seg contains an ACK, but not for our SYN/ACK, send a RST.
1529          */
1530         case TCPS_SYN_RECEIVED:
1531                 if ((thflags & TH_ACK) &&
1532                     (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1533                      SEQ_GT(th->th_ack, tp->snd_max))) {
1534                                 rstreason = BANDLIM_RST_OPENPORT;
1535                                 goto dropwithreset;
1536                 }
1537                 break;
1538
1539         /*
1540          * If the state is SYN_SENT:
1541          *      if seg contains an ACK, but not for our SYN, drop the input.
1542          *      if seg contains a RST, then drop the connection.
1543          *      if seg does not contain SYN, then drop it.
1544          * Otherwise this is an acceptable SYN segment
1545          *      initialize tp->rcv_nxt and tp->irs
1546          *      if seg contains ack then advance tp->snd_una
1547          *      if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1548          *      arrange for segment to be acked (eventually)
1549          *      continue processing rest of data/controls, beginning with URG
1550          */
1551         case TCPS_SYN_SENT:
1552                 if ((thflags & TH_ACK) &&
1553                     (SEQ_LEQ(th->th_ack, tp->iss) ||
1554                      SEQ_GT(th->th_ack, tp->snd_max))) {
1555                         rstreason = BANDLIM_UNLIMITED;
1556                         goto dropwithreset;
1557                 }
1558                 if (thflags & TH_RST) {
1559                         if (thflags & TH_ACK)
1560                                 tp = tcp_drop(tp, ECONNREFUSED);
1561                         goto drop;
1562                 }
1563                 if (!(thflags & TH_SYN))
1564                         goto drop;
1565
1566                 tp->irs = th->th_seq;
1567                 tcp_rcvseqinit(tp);
1568                 if (thflags & TH_ACK) {
1569                         /* Our SYN was acked. */
1570                         tcpstat.tcps_connects++;
1571                         soisconnected(so);
1572                         /* Do window scaling on this connection? */
1573                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1574                             (TF_RCVD_SCALE | TF_REQ_SCALE))
1575                                 tp->rcv_scale = tp->request_r_scale;
1576                         tp->rcv_adv += tp->rcv_wnd;
1577                         tp->snd_una++;          /* SYN is acked */
1578                         tcp_callout_stop(tp, tp->tt_rexmt);
1579                         /*
1580                          * If there's data, delay ACK; if there's also a FIN
1581                          * ACKNOW will be turned on later.
1582                          */
1583                         if (DELAY_ACK(tp) && tlen != 0) {
1584                                 tcp_callout_reset(tp, tp->tt_delack,
1585                                     tcp_delacktime, tcp_timer_delack);
1586                         } else {
1587                                 tp->t_flags |= TF_ACKNOW;
1588                         }
1589                         /*
1590                          * Received <SYN,ACK> in SYN_SENT[*] state.
1591                          * Transitions:
1592                          *      SYN_SENT  --> ESTABLISHED
1593                          *      SYN_SENT* --> FIN_WAIT_1
1594                          */
1595                         tp->t_starttime = ticks;
1596                         if (tp->t_flags & TF_NEEDFIN) {
1597                                 tp->t_state = TCPS_FIN_WAIT_1;
1598                                 tp->t_flags &= ~TF_NEEDFIN;
1599                                 thflags &= ~TH_SYN;
1600                         } else {
1601                                 tcp_established(tp);
1602                         }
1603                 } else {
1604                         /*
1605                          * Received initial SYN in SYN-SENT[*] state =>
1606                          * simultaneous open.
1607                          * Do 3-way handshake:
1608                          *        SYN-SENT -> SYN-RECEIVED
1609                          *        SYN-SENT* -> SYN-RECEIVED*
1610                          */
1611                         tp->t_flags |= TF_ACKNOW;
1612                         tcp_callout_stop(tp, tp->tt_rexmt);
1613                         tp->t_state = TCPS_SYN_RECEIVED;
1614                 }
1615
1616                 /*
1617                  * Advance th->th_seq to correspond to first data byte.
1618                  * If data, trim to stay within window,
1619                  * dropping FIN if necessary.
1620                  */
1621                 th->th_seq++;
1622                 if (tlen > tp->rcv_wnd) {
1623                         todrop = tlen - tp->rcv_wnd;
1624                         m_adj(m, -todrop);
1625                         tlen = tp->rcv_wnd;
1626                         thflags &= ~TH_FIN;
1627                         tcpstat.tcps_rcvpackafterwin++;
1628                         tcpstat.tcps_rcvbyteafterwin += todrop;
1629                 }
1630                 tp->snd_wl1 = th->th_seq - 1;
1631                 tp->rcv_up = th->th_seq;
1632                 /*
1633                  * Client side of transaction: already sent SYN and data.
1634                  * If the remote host used T/TCP to validate the SYN,
1635                  * our data will be ACK'd; if so, enter normal data segment
1636                  * processing in the middle of step 5, ack processing.
1637                  * Otherwise, goto step 6.
1638                  */
1639                 if (thflags & TH_ACK)
1640                         goto process_ACK;
1641
1642                 goto step6;
1643
1644         /*
1645          * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1646          *      do normal processing (we no longer bother with T/TCP).
1647          */
1648         case TCPS_LAST_ACK:
1649         case TCPS_CLOSING:
1650         case TCPS_TIME_WAIT:
1651                 break;  /* continue normal processing */
1652         }
1653
1654         /*
1655          * States other than LISTEN or SYN_SENT.
1656          * First check the RST flag and sequence number since reset segments
1657          * are exempt from the timestamp and connection count tests.  This
1658          * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1659          * below which allowed reset segments in half the sequence space
1660          * to fall though and be processed (which gives forged reset
1661          * segments with a random sequence number a 50 percent chance of
1662          * killing a connection).
1663          * Then check timestamp, if present.
1664          * Then check the connection count, if present.
1665          * Then check that at least some bytes of segment are within
1666          * receive window.  If segment begins before rcv_nxt,
1667          * drop leading data (and SYN); if nothing left, just ack.
1668          *
1669          *
1670          * If the RST bit is set, check the sequence number to see
1671          * if this is a valid reset segment.
1672          * RFC 793 page 37:
1673          *   In all states except SYN-SENT, all reset (RST) segments
1674          *   are validated by checking their SEQ-fields.  A reset is
1675          *   valid if its sequence number is in the window.
1676          * Note: this does not take into account delayed ACKs, so
1677          *   we should test against last_ack_sent instead of rcv_nxt.
1678          *   The sequence number in the reset segment is normally an
1679          *   echo of our outgoing acknowledgement numbers, but some hosts
1680          *   send a reset with the sequence number at the rightmost edge
1681          *   of our receive window, and we have to handle this case.
1682          * If we have multiple segments in flight, the intial reset
1683          * segment sequence numbers will be to the left of last_ack_sent,
1684          * but they will eventually catch up.
1685          * In any case, it never made sense to trim reset segments to
1686          * fit the receive window since RFC 1122 says:
1687          *   4.2.2.12  RST Segment: RFC-793 Section 3.4
1688          *
1689          *    A TCP SHOULD allow a received RST segment to include data.
1690          *
1691          *    DISCUSSION
1692          *         It has been suggested that a RST segment could contain
1693          *         ASCII text that encoded and explained the cause of the
1694          *         RST.  No standard has yet been established for such
1695          *         data.
1696          *
1697          * If the reset segment passes the sequence number test examine
1698          * the state:
1699          *    SYN_RECEIVED STATE:
1700          *      If passive open, return to LISTEN state.
1701          *      If active open, inform user that connection was refused.
1702          *    ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1703          *      Inform user that connection was reset, and close tcb.
1704          *    CLOSING, LAST_ACK STATES:
1705          *      Close the tcb.
1706          *    TIME_WAIT STATE:
1707          *      Drop the segment - see Stevens, vol. 2, p. 964 and
1708          *      RFC 1337.
1709          */
1710         if (thflags & TH_RST) {
1711                 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1712                     SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1713                         switch (tp->t_state) {
1714
1715                         case TCPS_SYN_RECEIVED:
1716                                 so->so_error = ECONNREFUSED;
1717                                 goto close;
1718
1719                         case TCPS_ESTABLISHED:
1720                         case TCPS_FIN_WAIT_1:
1721                         case TCPS_FIN_WAIT_2:
1722                         case TCPS_CLOSE_WAIT:
1723                                 so->so_error = ECONNRESET;
1724                         close:
1725                                 tp->t_state = TCPS_CLOSED;
1726                                 tcpstat.tcps_drops++;
1727                                 tp = tcp_close(tp);
1728                                 break;
1729
1730                         case TCPS_CLOSING:
1731                         case TCPS_LAST_ACK:
1732                                 tp = tcp_close(tp);
1733                                 break;
1734
1735                         case TCPS_TIME_WAIT:
1736                                 break;
1737                         }
1738                 }
1739                 goto drop;
1740         }
1741
1742         /*
1743          * RFC 1323 PAWS: If we have a timestamp reply on this segment
1744          * and it's less than ts_recent, drop it.
1745          */
1746         if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1747             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1748                 /* Check to see if ts_recent is over 24 days old.  */
1749                 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1750                         /*
1751                          * Invalidate ts_recent.  If this segment updates
1752                          * ts_recent, the age will be reset later and ts_recent
1753                          * will get a valid value.  If it does not, setting
1754                          * ts_recent to zero will at least satisfy the
1755                          * requirement that zero be placed in the timestamp
1756                          * echo reply when ts_recent isn't valid.  The
1757                          * age isn't reset until we get a valid ts_recent
1758                          * because we don't want out-of-order segments to be
1759                          * dropped when ts_recent is old.
1760                          */
1761                         tp->ts_recent = 0;
1762                 } else if (tcp_paws_tolerance && tlen != 0 &&
1763                     tp->t_state == TCPS_ESTABLISHED &&
1764                     (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK&&
1765                     !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1766                     th->th_ack == tp->snd_una &&
1767                     tiwin == tp->snd_wnd &&
1768                     TSTMP_GEQ(to.to_tsval + tcp_paws_tolerance, tp->ts_recent)&&
1769                     (th->th_seq == tp->rcv_nxt ||
1770                      (SEQ_GT(th->th_seq, tp->rcv_nxt) && 
1771                       tcp_paws_canreasslast(tp, th, tlen)))) {
1772                         /*
1773                          * This tends to prevent valid new segments from being
1774                          * dropped by the reordered segments sent by the fast
1775                          * retransmission algorithm on the sending side, i.e.
1776                          * the fast retransmitted segment w/ larger timestamp
1777                          * arrives earlier than the previously sent new segments
1778                          * w/ smaller timestamp.
1779                          *
1780                          * If following conditions are met, the segment is
1781                          * accepted:
1782                          * - The segment contains data
1783                          * - The connection is established
1784                          * - The header does not contain important flags
1785                          * - SYN or FIN is not needed
1786                          * - It does not acknowledge new data
1787                          * - Receive window is not changed
1788                          * - The timestamp is within "acceptable" range
1789                          * - The new segment is what we are expecting or
1790                          *   the new segment could be merged w/ the last
1791                          *   pending segment on the reassemble queue
1792                          */
1793                         tcpstat.tcps_pawsaccept++;
1794                         tcpstat.tcps_pawsdrop++;
1795                 } else {
1796                         tcpstat.tcps_rcvduppack++;
1797                         tcpstat.tcps_rcvdupbyte += tlen;
1798                         tcpstat.tcps_pawsdrop++;
1799                         if (tlen)
1800                                 goto dropafterack;
1801                         goto drop;
1802                 }
1803         }
1804
1805         /*
1806          * In the SYN-RECEIVED state, validate that the packet belongs to
1807          * this connection before trimming the data to fit the receive
1808          * window.  Check the sequence number versus IRS since we know
1809          * the sequence numbers haven't wrapped.  This is a partial fix
1810          * for the "LAND" DoS attack.
1811          */
1812         if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1813                 rstreason = BANDLIM_RST_OPENPORT;
1814                 goto dropwithreset;
1815         }
1816
1817         todrop = tp->rcv_nxt - th->th_seq;
1818         if (todrop > 0) {
1819                 if (TCP_DO_SACK(tp)) {
1820                         /* Report duplicate segment at head of packet. */
1821                         tp->reportblk.rblk_start = th->th_seq;
1822                         tp->reportblk.rblk_end = TCP_SACK_BLKEND(
1823                             th->th_seq + tlen, thflags);
1824                         if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1825                                 tp->reportblk.rblk_end = tp->rcv_nxt;
1826                         tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_SACKLEFT);
1827                         tp->t_flags |= TF_ACKNOW;
1828                 }
1829                 if (thflags & TH_SYN) {
1830                         thflags &= ~TH_SYN;
1831                         th->th_seq++;
1832                         if (th->th_urp > 1)
1833                                 th->th_urp--;
1834                         else
1835                                 thflags &= ~TH_URG;
1836                         todrop--;
1837                 }
1838                 /*
1839                  * Following if statement from Stevens, vol. 2, p. 960.
1840                  */
1841                 if (todrop > tlen ||
1842                     (todrop == tlen && !(thflags & TH_FIN))) {
1843                         /*
1844                          * Any valid FIN must be to the left of the window.
1845                          * At this point the FIN must be a duplicate or out
1846                          * of sequence; drop it.
1847                          */
1848                         thflags &= ~TH_FIN;
1849
1850                         /*
1851                          * Send an ACK to resynchronize and drop any data.
1852                          * But keep on processing for RST or ACK.
1853                          */
1854                         tp->t_flags |= TF_ACKNOW;
1855                         todrop = tlen;
1856                         tcpstat.tcps_rcvduppack++;
1857                         tcpstat.tcps_rcvdupbyte += todrop;
1858                 } else {
1859                         tcpstat.tcps_rcvpartduppack++;
1860                         tcpstat.tcps_rcvpartdupbyte += todrop;
1861                 }
1862                 drop_hdrlen += todrop;  /* drop from the top afterwards */
1863                 th->th_seq += todrop;
1864                 tlen -= todrop;
1865                 if (th->th_urp > todrop)
1866                         th->th_urp -= todrop;
1867                 else {
1868                         thflags &= ~TH_URG;
1869                         th->th_urp = 0;
1870                 }
1871         }
1872
1873         /*
1874          * If new data are received on a connection after the
1875          * user processes are gone, then RST the other end.
1876          */
1877         if ((so->so_state & SS_NOFDREF) &&
1878             tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1879                 tp = tcp_close(tp);
1880                 tcpstat.tcps_rcvafterclose++;
1881                 rstreason = BANDLIM_UNLIMITED;
1882                 goto dropwithreset;
1883         }
1884
1885         /*
1886          * If segment ends after window, drop trailing data
1887          * (and PUSH and FIN); if nothing left, just ACK.
1888          */
1889         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1890         if (todrop > 0) {
1891                 tcpstat.tcps_rcvpackafterwin++;
1892                 if (todrop >= tlen) {
1893                         tcpstat.tcps_rcvbyteafterwin += tlen;
1894                         /*
1895                          * If a new connection request is received
1896                          * while in TIME_WAIT, drop the old connection
1897                          * and start over if the sequence numbers
1898                          * are above the previous ones.
1899                          */
1900                         if (thflags & TH_SYN &&
1901                             tp->t_state == TCPS_TIME_WAIT &&
1902                             SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1903                                 tp = tcp_close(tp);
1904                                 goto findpcb;
1905                         }
1906                         /*
1907                          * If window is closed can only take segments at
1908                          * window edge, and have to drop data and PUSH from
1909                          * incoming segments.  Continue processing, but
1910                          * remember to ack.  Otherwise, drop segment
1911                          * and ack.
1912                          */
1913                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1914                                 tp->t_flags |= TF_ACKNOW;
1915                                 tcpstat.tcps_rcvwinprobe++;
1916                         } else
1917                                 goto dropafterack;
1918                 } else
1919                         tcpstat.tcps_rcvbyteafterwin += todrop;
1920                 m_adj(m, -todrop);
1921                 tlen -= todrop;
1922                 thflags &= ~(TH_PUSH | TH_FIN);
1923         }
1924
1925         /*
1926          * If last ACK falls within this segment's sequence numbers,
1927          * record its timestamp.
1928          * NOTE:
1929          * 1) That the test incorporates suggestions from the latest
1930          *    proposal of the tcplw@cray.com list (Braden 1993/04/26).
1931          * 2) That updating only on newer timestamps interferes with
1932          *    our earlier PAWS tests, so this check should be solely
1933          *    predicated on the sequence space of this segment.
1934          * 3) That we modify the segment boundary check to be
1935          *        Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1936          *    instead of RFC1323's
1937          *        Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1938          *    This modified check allows us to overcome RFC1323's
1939          *    limitations as described in Stevens TCP/IP Illustrated
1940          *    Vol. 2 p.869. In such cases, we can still calculate the
1941          *    RTT correctly when RCV.NXT == Last.ACK.Sent.
1942          */
1943         if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1944             SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1945                                         + ((thflags & TH_SYN) != 0)
1946                                         + ((thflags & TH_FIN) != 0)))) {
1947                 tp->ts_recent_age = ticks;
1948                 tp->ts_recent = to.to_tsval;
1949         }
1950
1951         /*
1952          * If a SYN is in the window, then this is an
1953          * error and we send an RST and drop the connection.
1954          */
1955         if (thflags & TH_SYN) {
1956                 tp = tcp_drop(tp, ECONNRESET);
1957                 rstreason = BANDLIM_UNLIMITED;
1958                 goto dropwithreset;
1959         }
1960
1961         /*
1962          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN
1963          * flag is on (half-synchronized state), then queue data for
1964          * later processing; else drop segment and return.
1965          */
1966         if (!(thflags & TH_ACK)) {
1967                 if (tp->t_state == TCPS_SYN_RECEIVED ||
1968                     (tp->t_flags & TF_NEEDSYN))
1969                         goto step6;
1970                 else
1971                         goto drop;
1972         }
1973
1974         /*
1975          * Ack processing.
1976          */
1977         switch (tp->t_state) {
1978         /*
1979          * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1980          * ESTABLISHED state and continue processing.
1981          * The ACK was checked above.
1982          */
1983         case TCPS_SYN_RECEIVED:
1984
1985                 tcpstat.tcps_connects++;
1986                 soisconnected(so);
1987                 /* Do window scaling? */
1988                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1989                     (TF_RCVD_SCALE | TF_REQ_SCALE))
1990                         tp->rcv_scale = tp->request_r_scale;
1991                 /*
1992                  * Make transitions:
1993                  *      SYN-RECEIVED  -> ESTABLISHED
1994                  *      SYN-RECEIVED* -> FIN-WAIT-1
1995                  */
1996                 tp->t_starttime = ticks;
1997                 if (tp->t_flags & TF_NEEDFIN) {
1998                         tp->t_state = TCPS_FIN_WAIT_1;
1999                         tp->t_flags &= ~TF_NEEDFIN;
2000                 } else {
2001                         tcp_established(tp);
2002                 }
2003                 /*
2004                  * If segment contains data or ACK, will call tcp_reass()
2005                  * later; if not, do so now to pass queued data to user.
2006                  */
2007                 if (tlen == 0 && !(thflags & TH_FIN))
2008                         tcp_reass(tp, NULL, NULL, NULL);
2009                 /* fall into ... */
2010
2011         /*
2012          * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
2013          * ACKs.  If the ack is in the range
2014          *      tp->snd_una < th->th_ack <= tp->snd_max
2015          * then advance tp->snd_una to th->th_ack and drop
2016          * data from the retransmission queue.  If this ACK reflects
2017          * more up to date window information we update our window information.
2018          */
2019         case TCPS_ESTABLISHED:
2020         case TCPS_FIN_WAIT_1:
2021         case TCPS_FIN_WAIT_2:
2022         case TCPS_CLOSE_WAIT:
2023         case TCPS_CLOSING:
2024         case TCPS_LAST_ACK:
2025         case TCPS_TIME_WAIT:
2026
2027                 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
2028                         boolean_t maynotdup = FALSE;
2029
2030                         if (TCP_DO_SACK(tp))
2031                                 tcp_sack_update_scoreboard(tp, &to);
2032
2033                         if (tlen != 0 || tiwin != tp->snd_wnd ||
2034                             ((thflags & TH_FIN) && !(tp->t_flags & TF_SAWFIN)))
2035                                 maynotdup = TRUE;
2036
2037                         if (!tcp_callout_active(tp, tp->tt_rexmt) ||
2038                             th->th_ack != tp->snd_una) {
2039                                 if (!maynotdup)
2040                                         tcpstat.tcps_rcvdupack++;
2041                                 tp->t_dupacks = 0;
2042                                 break;
2043                         }
2044
2045 #define DELAY_DUPACK \
2046 do { \
2047         delayed_dupack = TRUE; \
2048         th_dupack = th->th_ack; \
2049         to_flags = to.to_flags; \
2050 } while (0)
2051                         if (maynotdup) {
2052                                 if (!tcp_do_rfc6675 ||
2053                                     !TCP_DO_SACK(tp) ||
2054                                     (to.to_flags &
2055                                      (TOF_SACK | TOF_SACK_REDUNDANT))
2056                                      != TOF_SACK) {
2057                                         tp->t_dupacks = 0;
2058                                 } else {
2059                                         DELAY_DUPACK;
2060                                 }
2061                                 break;
2062                         }
2063                         if ((thflags & TH_FIN) && !(tp->t_flags & TF_QUEDFIN)) {
2064                                 /*
2065                                  * This could happen, if the reassemable
2066                                  * queue overflew or was drained.  Don't
2067                                  * drop this FIN here; defer the duplicated
2068                                  * ACK processing until this FIN gets queued.
2069                                  */
2070                                 DELAY_DUPACK;
2071                                 break;
2072                         }
2073 #undef DELAY_DUPACK
2074
2075                         if (tcp_recv_dupack(tp, th->th_ack, to.to_flags))
2076                                 goto drop;
2077                         else
2078                                 break;
2079                 }
2080
2081                 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2082                 tp->t_dupacks = 0;
2083                 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2084                         /*
2085                          * Detected optimistic ACK attack.
2086                          * Force slow-start to de-synchronize attack.
2087                          */
2088                         tp->snd_cwnd = tp->t_maxseg;
2089                         tp->snd_wacked = 0;
2090
2091                         tcpstat.tcps_rcvacktoomuch++;
2092                         goto dropafterack;
2093                 }
2094                 /*
2095                  * If we reach this point, ACK is not a duplicate,
2096                  *     i.e., it ACKs something we sent.
2097                  */
2098                 if (tp->t_flags & TF_NEEDSYN) {
2099                         /*
2100                          * T/TCP: Connection was half-synchronized, and our
2101                          * SYN has been ACK'd (so connection is now fully
2102                          * synchronized).  Go to non-starred state,
2103                          * increment snd_una for ACK of SYN, and check if
2104                          * we can do window scaling.
2105                          */
2106                         tp->t_flags &= ~TF_NEEDSYN;
2107                         tp->snd_una++;
2108                         /* Do window scaling? */
2109                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2110                             (TF_RCVD_SCALE | TF_REQ_SCALE))
2111                                 tp->rcv_scale = tp->request_r_scale;
2112                 }
2113
2114 process_ACK:
2115                 acked = th->th_ack - tp->snd_una;
2116                 tcpstat.tcps_rcvackpack++;
2117                 tcpstat.tcps_rcvackbyte += acked;
2118
2119                 if (tcp_do_eifel_detect && acked > 0 &&
2120                     (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2121                     (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
2122                         /* Eifel detection applicable. */
2123                         if (to.to_tsecr < tp->t_rexmtTS) {
2124                                 ++tcpstat.tcps_eifeldetected;
2125                                 tcp_revert_congestion_state(tp);
2126                                 if (tp->t_rxtshift != 1 ||
2127                                     ticks >= tp->t_badrxtwin)
2128                                         ++tcpstat.tcps_rttcantdetect;
2129                         }
2130                 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2131                         /*
2132                          * If we just performed our first retransmit,
2133                          * and the ACK arrives within our recovery window,
2134                          * then it was a mistake to do the retransmit
2135                          * in the first place.  Recover our original cwnd
2136                          * and ssthresh, and proceed to transmit where we
2137                          * left off.
2138                          */
2139                         tcp_revert_congestion_state(tp);
2140                         ++tcpstat.tcps_rttdetected;
2141                 }
2142
2143                 /*
2144                  * If we have a timestamp reply, update smoothed
2145                  * round trip time.  If no timestamp is present but
2146                  * transmit timer is running and timed sequence
2147                  * number was acked, update smoothed round trip time.
2148                  * Since we now have an rtt measurement, cancel the
2149                  * timer backoff (cf., Phil Karn's retransmit alg.).
2150                  * Recompute the initial retransmit timer.
2151                  *
2152                  * Some machines (certain windows boxes) send broken
2153                  * timestamp replies during the SYN+ACK phase, ignore
2154                  * timestamps of 0.
2155                  */
2156                 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2157                         tcp_xmit_timer(tp, ticks - to.to_tsecr + 1, th->th_ack);
2158                 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2159                         tcp_xmit_timer(tp, ticks - tp->t_rtttime, th->th_ack);
2160                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2161
2162                 /*
2163                  * If no data (only SYN) was ACK'd,
2164                  *    skip rest of ACK processing.
2165                  */
2166                 if (acked == 0)
2167                         goto step6;
2168
2169                 /* Stop looking for an acceptable ACK since one was received. */
2170                 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
2171                     TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
2172
2173                 if (acked > so->so_snd.ssb_cc) {
2174                         tp->snd_wnd -= so->so_snd.ssb_cc;
2175                         sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2176                         ourfinisacked = TRUE;
2177                 } else {
2178                         sbdrop(&so->so_snd.sb, acked);
2179                         tp->snd_wnd -= acked;
2180                         ourfinisacked = FALSE;
2181                 }
2182                 sowwakeup(so);
2183
2184                 /*
2185                  * Update window information.
2186                  */
2187                 if (acceptable_window_update(tp, th, tiwin)) {
2188                         /* keep track of pure window updates */
2189                         if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2190                             tiwin > tp->snd_wnd)
2191                                 tcpstat.tcps_rcvwinupd++;
2192                         tp->snd_wnd = tiwin;
2193                         tp->snd_wl1 = th->th_seq;
2194                         tp->snd_wl2 = th->th_ack;
2195                         if (tp->snd_wnd > tp->max_sndwnd)
2196                                 tp->max_sndwnd = tp->snd_wnd;
2197                         needoutput = TRUE;
2198                 }
2199
2200                 tp->snd_una = th->th_ack;
2201                 if (TCP_DO_SACK(tp))
2202                         tcp_sack_update_scoreboard(tp, &to);
2203                 if (IN_FASTRECOVERY(tp)) {
2204                         if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2205                                 EXIT_FASTRECOVERY(tp);
2206                                 needoutput = TRUE;
2207                                 /*
2208                                  * If the congestion window was inflated
2209                                  * to account for the other side's
2210                                  * cached packets, retract it.
2211                                  */
2212                                 if (!TCP_DO_SACK(tp))
2213                                         tp->snd_cwnd = tp->snd_ssthresh;
2214
2215                                 /*
2216                                  * Window inflation should have left us
2217                                  * with approximately snd_ssthresh outstanding
2218                                  * data.  But, in case we would be inclined
2219                                  * to send a burst, better do it using
2220                                  * slow start.
2221                                  */
2222                                 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2223                                            tp->snd_max + 2 * tp->t_maxseg))
2224                                         tp->snd_cwnd =
2225                                             (tp->snd_max - tp->snd_una) +
2226                                             2 * tp->t_maxseg;
2227
2228                                 tp->snd_wacked = 0;
2229                         } else {
2230                                 if (TCP_DO_SACK(tp)) {
2231                                         tp->snd_max_rexmt = tp->snd_max;
2232                                         tcp_sack_rexmt(tp,
2233                                             tp->snd_una == tp->rexmt_high);
2234                                 } else {
2235                                         tcp_newreno_partial_ack(tp, th, acked);
2236                                 }
2237                                 needoutput = FALSE;
2238                         }
2239                 } else {
2240                         /*
2241                          * Open the congestion window.  When in slow-start,
2242                          * open exponentially: maxseg per packet.  Otherwise,
2243                          * open linearly: maxseg per window.
2244                          */
2245                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
2246                                 u_int abc_sslimit =
2247                                     (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2248                                      tp->t_maxseg : 2 * tp->t_maxseg);
2249
2250                                 /* slow-start */
2251                                 tp->snd_cwnd += tcp_do_abc ?
2252                                     min(acked, abc_sslimit) : tp->t_maxseg;
2253                         } else {
2254                                 /* linear increase */
2255                                 tp->snd_wacked += tcp_do_abc ? acked :
2256                                     tp->t_maxseg;
2257                                 if (tp->snd_wacked >= tp->snd_cwnd) {
2258                                         tp->snd_wacked -= tp->snd_cwnd;
2259                                         tp->snd_cwnd += tp->t_maxseg;
2260                                 }
2261                         }
2262                         tp->snd_cwnd = min(tp->snd_cwnd,
2263                                            TCP_MAXWIN << tp->snd_scale);
2264                         tp->snd_recover = th->th_ack - 1;
2265                 }
2266                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2267                         tp->snd_nxt = tp->snd_una;
2268
2269                 /*
2270                  * If all outstanding data is acked, stop retransmit
2271                  * timer and remember to restart (more output or persist).
2272                  * If there is more data to be acked, restart retransmit
2273                  * timer, using current (possibly backed-off) value.
2274                  */
2275                 if (th->th_ack == tp->snd_max) {
2276                         tcp_callout_stop(tp, tp->tt_rexmt);
2277                         needoutput = TRUE;
2278                 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2279                         tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2280                             tcp_timer_rexmt);
2281                 }
2282
2283                 switch (tp->t_state) {
2284                 /*
2285                  * In FIN_WAIT_1 STATE in addition to the processing
2286                  * for the ESTABLISHED state if our FIN is now acknowledged
2287                  * then enter FIN_WAIT_2.
2288                  */
2289                 case TCPS_FIN_WAIT_1:
2290                         if (ourfinisacked) {
2291                                 /*
2292                                  * If we can't receive any more
2293                                  * data, then closing user can proceed.
2294                                  * Starting the timer is contrary to the
2295                                  * specification, but if we don't get a FIN
2296                                  * we'll hang forever.
2297                                  */
2298                                 if (so->so_state & SS_CANTRCVMORE) {
2299                                         soisdisconnected(so);
2300                                         tcp_callout_reset(tp, tp->tt_2msl,
2301                                             tp->t_maxidle, tcp_timer_2msl);
2302                                 }
2303                                 tp->t_state = TCPS_FIN_WAIT_2;
2304                         }
2305                         break;
2306
2307                 /*
2308                  * In CLOSING STATE in addition to the processing for
2309                  * the ESTABLISHED state if the ACK acknowledges our FIN
2310                  * then enter the TIME-WAIT state, otherwise ignore
2311                  * the segment.
2312                  */
2313                 case TCPS_CLOSING:
2314                         if (ourfinisacked) {
2315                                 tp->t_state = TCPS_TIME_WAIT;
2316                                 tcp_canceltimers(tp);
2317                                 tcp_callout_reset(tp, tp->tt_2msl,
2318                                             2 * tcp_rmx_msl(tp),
2319                                             tcp_timer_2msl);
2320                                 soisdisconnected(so);
2321                         }
2322                         break;
2323
2324                 /*
2325                  * In LAST_ACK, we may still be waiting for data to drain
2326                  * and/or to be acked, as well as for the ack of our FIN.
2327                  * If our FIN is now acknowledged, delete the TCB,
2328                  * enter the closed state and return.
2329                  */
2330                 case TCPS_LAST_ACK:
2331                         if (ourfinisacked) {
2332                                 tp = tcp_close(tp);
2333                                 goto drop;
2334                         }
2335                         break;
2336
2337                 /*
2338                  * In TIME_WAIT state the only thing that should arrive
2339                  * is a retransmission of the remote FIN.  Acknowledge
2340                  * it and restart the finack timer.
2341                  */
2342                 case TCPS_TIME_WAIT:
2343                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2344                             tcp_timer_2msl);
2345                         goto dropafterack;
2346                 }
2347         }
2348
2349 step6:
2350         /*
2351          * Update window information.
2352          * Don't look at window if no ACK: TAC's send garbage on first SYN.
2353          */
2354         if ((thflags & TH_ACK) &&
2355             acceptable_window_update(tp, th, tiwin)) {
2356                 /* keep track of pure window updates */
2357                 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2358                     tiwin > tp->snd_wnd)
2359                         tcpstat.tcps_rcvwinupd++;
2360                 tp->snd_wnd = tiwin;
2361                 tp->snd_wl1 = th->th_seq;
2362                 tp->snd_wl2 = th->th_ack;
2363                 if (tp->snd_wnd > tp->max_sndwnd)
2364                         tp->max_sndwnd = tp->snd_wnd;
2365                 needoutput = TRUE;
2366         }
2367
2368         /*
2369          * Process segments with URG.
2370          */
2371         if ((thflags & TH_URG) && th->th_urp &&
2372             !TCPS_HAVERCVDFIN(tp->t_state)) {
2373                 /*
2374                  * This is a kludge, but if we receive and accept
2375                  * random urgent pointers, we'll crash in
2376                  * soreceive.  It's hard to imagine someone
2377                  * actually wanting to send this much urgent data.
2378                  */
2379                 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2380                         th->th_urp = 0;                 /* XXX */
2381                         thflags &= ~TH_URG;             /* XXX */
2382                         goto dodata;                    /* XXX */
2383                 }
2384                 /*
2385                  * If this segment advances the known urgent pointer,
2386                  * then mark the data stream.  This should not happen
2387                  * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2388                  * a FIN has been received from the remote side.
2389                  * In these states we ignore the URG.
2390                  *
2391                  * According to RFC961 (Assigned Protocols),
2392                  * the urgent pointer points to the last octet
2393                  * of urgent data.  We continue, however,
2394                  * to consider it to indicate the first octet
2395                  * of data past the urgent section as the original
2396                  * spec states (in one of two places).
2397                  */
2398                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2399                         tp->rcv_up = th->th_seq + th->th_urp;
2400                         so->so_oobmark = so->so_rcv.ssb_cc +
2401                             (tp->rcv_up - tp->rcv_nxt) - 1;
2402                         if (so->so_oobmark == 0)
2403                                 sosetstate(so, SS_RCVATMARK);
2404                         sohasoutofband(so);
2405                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2406                 }
2407                 /*
2408                  * Remove out of band data so doesn't get presented to user.
2409                  * This can happen independent of advancing the URG pointer,
2410                  * but if two URG's are pending at once, some out-of-band
2411                  * data may creep in... ick.
2412                  */
2413                 if (th->th_urp <= (u_long)tlen &&
2414                     !(so->so_options & SO_OOBINLINE)) {
2415                         /* hdr drop is delayed */
2416                         tcp_pulloutofband(so, th, m, drop_hdrlen);
2417                 }
2418         } else {
2419                 /*
2420                  * If no out of band data is expected,
2421                  * pull receive urgent pointer along
2422                  * with the receive window.
2423                  */
2424                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2425                         tp->rcv_up = tp->rcv_nxt;
2426         }
2427
2428 dodata:                                                 /* XXX */
2429         /*
2430          * Process the segment text, merging it into the TCP sequencing queue,
2431          * and arranging for acknowledgment of receipt if necessary.
2432          * This process logically involves adjusting tp->rcv_wnd as data
2433          * is presented to the user (this happens in tcp_usrreq.c,
2434          * case PRU_RCVD).  If a FIN has already been received on this
2435          * connection then we just ignore the text.
2436          */
2437         if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2438                 if (thflags & TH_FIN)
2439                         tp->t_flags |= TF_SAWFIN;
2440                 m_adj(m, drop_hdrlen);  /* delayed header drop */
2441                 /*
2442                  * Insert segment which includes th into TCP reassembly queue
2443                  * with control block tp.  Set thflags to whether reassembly now
2444                  * includes a segment with FIN.  This handles the common case
2445                  * inline (segment is the next to be received on an established
2446                  * connection, and the queue is empty), avoiding linkage into
2447                  * and removal from the queue and repetition of various
2448                  * conversions.
2449                  * Set DELACK for segments received in order, but ack
2450                  * immediately when segments are out of order (so
2451                  * fast retransmit can work).
2452                  */
2453                 if (th->th_seq == tp->rcv_nxt &&
2454                     TAILQ_EMPTY(&tp->t_segq) &&
2455                     TCPS_HAVEESTABLISHED(tp->t_state)) {
2456                         if (thflags & TH_FIN)
2457                                 tp->t_flags |= TF_QUEDFIN;
2458                         if (DELAY_ACK(tp)) {
2459                                 tcp_callout_reset(tp, tp->tt_delack,
2460                                     tcp_delacktime, tcp_timer_delack);
2461                         } else {
2462                                 tp->t_flags |= TF_ACKNOW;
2463                         }
2464                         tp->rcv_nxt += tlen;
2465                         thflags = th->th_flags & TH_FIN;
2466                         tcpstat.tcps_rcvpack++;
2467                         tcpstat.tcps_rcvbyte += tlen;
2468                         ND6_HINT(tp);
2469                         if (so->so_state & SS_CANTRCVMORE) {
2470                                 m_freem(m);
2471                         } else {
2472                                 lwkt_gettoken(&so->so_rcv.ssb_token);
2473                                 ssb_appendstream(&so->so_rcv, m);
2474                                 lwkt_reltoken(&so->so_rcv.ssb_token);
2475                         }
2476                         sorwakeup(so);
2477                 } else {
2478                         if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
2479                                 /* Initialize SACK report block. */
2480                                 tp->reportblk.rblk_start = th->th_seq;
2481                                 tp->reportblk.rblk_end = TCP_SACK_BLKEND(
2482                                     th->th_seq + tlen, thflags);
2483                         }
2484                         thflags = tcp_reass(tp, th, &tlen, m);
2485                         tp->t_flags |= TF_ACKNOW;
2486                 }
2487
2488                 /*
2489                  * Note the amount of data that peer has sent into
2490                  * our window, in order to estimate the sender's
2491                  * buffer size.
2492                  */
2493                 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2494         } else {
2495                 m_freem(m);
2496                 thflags &= ~TH_FIN;
2497         }
2498
2499         /*
2500          * If FIN is received ACK the FIN and let the user know
2501          * that the connection is closing.
2502          */
2503         if (thflags & TH_FIN) {
2504                 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2505                         socantrcvmore(so);
2506                         /*
2507                          * If connection is half-synchronized
2508                          * (ie NEEDSYN flag on) then delay ACK,
2509                          * so it may be piggybacked when SYN is sent.
2510                          * Otherwise, since we received a FIN then no
2511                          * more input can be expected, send ACK now.
2512                          */
2513                         if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2514                                 tcp_callout_reset(tp, tp->tt_delack,
2515                                     tcp_delacktime, tcp_timer_delack);
2516                         } else {
2517                                 tp->t_flags |= TF_ACKNOW;
2518                         }
2519                         tp->rcv_nxt++;
2520                 }
2521
2522                 switch (tp->t_state) {
2523                 /*
2524                  * In SYN_RECEIVED and ESTABLISHED STATES
2525                  * enter the CLOSE_WAIT state.
2526                  */
2527                 case TCPS_SYN_RECEIVED:
2528                         tp->t_starttime = ticks;
2529                         /*FALLTHROUGH*/
2530                 case TCPS_ESTABLISHED:
2531                         tp->t_state = TCPS_CLOSE_WAIT;
2532                         break;
2533
2534                 /*
2535                  * If still in FIN_WAIT_1 STATE FIN has not been acked so
2536                  * enter the CLOSING state.
2537                  */
2538                 case TCPS_FIN_WAIT_1:
2539                         tp->t_state = TCPS_CLOSING;
2540                         break;
2541
2542                 /*
2543                  * In FIN_WAIT_2 state enter the TIME_WAIT state,
2544                  * starting the time-wait timer, turning off the other
2545                  * standard timers.
2546                  */
2547                 case TCPS_FIN_WAIT_2:
2548                         tp->t_state = TCPS_TIME_WAIT;
2549                         tcp_canceltimers(tp);
2550                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2551                                     tcp_timer_2msl);
2552                         soisdisconnected(so);
2553                         break;
2554
2555                 /*
2556                  * In TIME_WAIT state restart the 2 MSL time_wait timer.
2557                  */
2558                 case TCPS_TIME_WAIT:
2559                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2560                             tcp_timer_2msl);
2561                         break;
2562                 }
2563         }
2564
2565 #ifdef TCPDEBUG
2566         if (so->so_options & SO_DEBUG)
2567                 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2568 #endif
2569
2570         /*
2571          * Delayed duplicated ACK processing
2572          */
2573         if (delayed_dupack && tcp_recv_dupack(tp, th_dupack, to_flags))
2574                 needoutput = FALSE;
2575
2576         /*
2577          * Return any desired output.
2578          */
2579         if ((tp->t_flags & TF_ACKNOW) ||
2580             (needoutput && tcp_sack_report_needed(tp))) {
2581                 tcp_output_cancel(tp);
2582                 tcp_output_fair(tp);
2583         } else if (needoutput && !tcp_output_pending(tp)) {
2584                 tcp_output_fair(tp);
2585         }
2586         tcp_sack_report_cleanup(tp);
2587         return(IPPROTO_DONE);
2588
2589 dropafterack:
2590         /*
2591          * Generate an ACK dropping incoming segment if it occupies
2592          * sequence space, where the ACK reflects our state.
2593          *
2594          * We can now skip the test for the RST flag since all
2595          * paths to this code happen after packets containing
2596          * RST have been dropped.
2597          *
2598          * In the SYN-RECEIVED state, don't send an ACK unless the
2599          * segment we received passes the SYN-RECEIVED ACK test.
2600          * If it fails send a RST.  This breaks the loop in the
2601          * "LAND" DoS attack, and also prevents an ACK storm
2602          * between two listening ports that have been sent forged
2603          * SYN segments, each with the source address of the other.
2604          */
2605         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2606             (SEQ_GT(tp->snd_una, th->th_ack) ||
2607              SEQ_GT(th->th_ack, tp->snd_max)) ) {
2608                 rstreason = BANDLIM_RST_OPENPORT;
2609                 goto dropwithreset;
2610         }
2611 #ifdef TCPDEBUG
2612         if (so->so_options & SO_DEBUG)
2613                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2614 #endif
2615         m_freem(m);
2616         tp->t_flags |= TF_ACKNOW;
2617         tcp_output(tp);
2618         tcp_sack_report_cleanup(tp);
2619         return(IPPROTO_DONE);
2620
2621 dropwithreset:
2622         /*
2623          * Generate a RST, dropping incoming segment.
2624          * Make ACK acceptable to originator of segment.
2625          * Don't bother to respond if destination was broadcast/multicast.
2626          */
2627         if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2628                 goto drop;
2629         if (isipv6) {
2630                 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2631                     IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2632                         goto drop;
2633         } else {
2634                 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2635                     IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2636                     ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2637                     in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2638                         goto drop;
2639         }
2640         /* IPv6 anycast check is done at tcp6_input() */
2641
2642         /*
2643          * Perform bandwidth limiting.
2644          */
2645 #ifdef ICMP_BANDLIM
2646         if (badport_bandlim(rstreason) < 0)
2647                 goto drop;
2648 #endif
2649
2650 #ifdef TCPDEBUG
2651         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2652                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2653 #endif
2654         if (thflags & TH_ACK)
2655                 /* mtod() below is safe as long as hdr dropping is delayed */
2656                 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2657                             TH_RST);
2658         else {
2659                 if (thflags & TH_SYN)
2660                         tlen++;
2661                 /* mtod() below is safe as long as hdr dropping is delayed */
2662                 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2663                             (tcp_seq)0, TH_RST | TH_ACK);
2664         }
2665         if (tp != NULL)
2666                 tcp_sack_report_cleanup(tp);
2667         return(IPPROTO_DONE);
2668
2669 drop:
2670         /*
2671          * Drop space held by incoming segment and return.
2672          */
2673 #ifdef TCPDEBUG
2674         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2675                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2676 #endif
2677         m_freem(m);
2678         if (tp != NULL)
2679                 tcp_sack_report_cleanup(tp);
2680         return(IPPROTO_DONE);
2681 }
2682
2683 /*
2684  * Parse TCP options and place in tcpopt.
2685  */
2686 static void
2687 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn,
2688     tcp_seq ack)
2689 {
2690         int opt, optlen, i;
2691
2692         to->to_flags = 0;
2693         for (; cnt > 0; cnt -= optlen, cp += optlen) {
2694                 opt = cp[0];
2695                 if (opt == TCPOPT_EOL)
2696                         break;
2697                 if (opt == TCPOPT_NOP)
2698                         optlen = 1;
2699                 else {
2700                         if (cnt < 2)
2701                                 break;
2702                         optlen = cp[1];
2703                         if (optlen < 2 || optlen > cnt)
2704                                 break;
2705                 }
2706                 switch (opt) {
2707                 case TCPOPT_MAXSEG:
2708                         if (optlen != TCPOLEN_MAXSEG)
2709                                 continue;
2710                         if (!is_syn)
2711                                 continue;
2712                         to->to_flags |= TOF_MSS;
2713                         bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2714                         to->to_mss = ntohs(to->to_mss);
2715                         break;
2716                 case TCPOPT_WINDOW:
2717                         if (optlen != TCPOLEN_WINDOW)
2718                                 continue;
2719                         if (!is_syn)
2720                                 continue;
2721                         to->to_flags |= TOF_SCALE;
2722                         to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2723                         break;
2724                 case TCPOPT_TIMESTAMP:
2725                         if (optlen != TCPOLEN_TIMESTAMP)
2726                                 continue;
2727                         to->to_flags |= TOF_TS;
2728                         bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2729                         to->to_tsval = ntohl(to->to_tsval);
2730                         bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2731                         to->to_tsecr = ntohl(to->to_tsecr);
2732                         /*
2733                          * If echoed timestamp is later than the current time,
2734                          * fall back to non RFC1323 RTT calculation.
2735                          */
2736                         if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2737                                 to->to_tsecr = 0;
2738                         break;
2739                 case TCPOPT_SACK_PERMITTED:
2740                         if (optlen != TCPOLEN_SACK_PERMITTED)
2741                                 continue;
2742                         if (!is_syn)
2743                                 continue;
2744                         to->to_flags |= TOF_SACK_PERMITTED;
2745                         break;
2746                 case TCPOPT_SACK:
2747                         if ((optlen - 2) & 0x07)        /* not multiple of 8 */
2748                                 continue;
2749                         to->to_nsackblocks = (optlen - 2) / 8;
2750                         to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2751                         to->to_flags |= TOF_SACK;
2752                         for (i = 0; i < to->to_nsackblocks; i++) {
2753                                 struct raw_sackblock *r = &to->to_sackblocks[i];
2754
2755                                 r->rblk_start = ntohl(r->rblk_start);
2756                                 r->rblk_end = ntohl(r->rblk_end);
2757
2758                                 if (SEQ_LEQ(r->rblk_end, r->rblk_start)) {
2759                                         /*
2760                                          * Invalid SACK block; discard all
2761                                          * SACK blocks
2762                                          */
2763                                         tcpstat.tcps_rcvbadsackopt++;
2764                                         to->to_nsackblocks = 0;
2765                                         to->to_sackblocks = NULL;
2766                                         to->to_flags &= ~TOF_SACK;
2767                                         break;
2768                                 }
2769                         }
2770                         if ((to->to_flags & TOF_SACK) &&
2771                             tcp_sack_ndsack_blocks(to->to_sackblocks,
2772                             to->to_nsackblocks, ack))
2773                                 to->to_flags |= TOF_DSACK;
2774                         break;
2775 #ifdef TCP_SIGNATURE
2776                 /*
2777                  * XXX In order to reply to a host which has set the
2778                  * TCP_SIGNATURE option in its initial SYN, we have to
2779                  * record the fact that the option was observed here
2780                  * for the syncache code to perform the correct response.
2781                  */
2782                 case TCPOPT_SIGNATURE:
2783                         if (optlen != TCPOLEN_SIGNATURE)
2784                                 continue;
2785                         to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2786                         break;
2787 #endif /* TCP_SIGNATURE */
2788                 default:
2789                         continue;
2790                 }
2791         }
2792 }
2793
2794 /*
2795  * Pull out of band byte out of a segment so
2796  * it doesn't appear in the user's data queue.
2797  * It is still reflected in the segment length for
2798  * sequencing purposes.
2799  * "off" is the delayed to be dropped hdrlen.
2800  */
2801 static void
2802 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2803 {
2804         int cnt = off + th->th_urp - 1;
2805
2806         while (cnt >= 0) {
2807                 if (m->m_len > cnt) {
2808                         char *cp = mtod(m, caddr_t) + cnt;
2809                         struct tcpcb *tp = sototcpcb(so);
2810
2811                         tp->t_iobc = *cp;
2812                         tp->t_oobflags |= TCPOOB_HAVEDATA;
2813                         bcopy(cp + 1, cp, m->m_len - cnt - 1);
2814                         m->m_len--;
2815                         if (m->m_flags & M_PKTHDR)
2816                                 m->m_pkthdr.len--;
2817                         return;
2818                 }
2819                 cnt -= m->m_len;
2820                 m = m->m_next;
2821                 if (m == NULL)
2822                         break;
2823         }
2824         panic("tcp_pulloutofband");
2825 }
2826
2827 /*
2828  * Collect new round-trip time estimate
2829  * and update averages and current timeout.
2830  */
2831 static void
2832 tcp_xmit_timer(struct tcpcb *tp, int rtt, tcp_seq ack)
2833 {
2834         int rebaserto = 0;
2835
2836         tcpstat.tcps_rttupdated++;
2837         tp->t_rttupdated++;
2838         if ((tp->rxt_flags & TRXT_F_REBASERTO) &&
2839             SEQ_GT(ack, tp->snd_max_prev)) {
2840 #ifdef DEBUG_EIFEL_RESPONSE
2841                 kprintf("srtt/rttvar, prev %d/%d, cur %d/%d, ",
2842                     tp->t_srtt_prev, tp->t_rttvar_prev,
2843                     tp->t_srtt, tp->t_rttvar);
2844 #endif
2845
2846                 tcpstat.tcps_eifelresponse++;
2847                 rebaserto = 1;
2848                 tp->rxt_flags &= ~TRXT_F_REBASERTO;
2849                 tp->t_srtt = max(tp->t_srtt_prev, (rtt << TCP_RTT_SHIFT));
2850                 tp->t_rttvar = max(tp->t_rttvar_prev,
2851                     (rtt << (TCP_RTTVAR_SHIFT - 1)));
2852                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2853                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2854
2855 #ifdef DEBUG_EIFEL_RESPONSE
2856                 kprintf("new %d/%d ", tp->t_srtt, tp->t_rttvar);
2857 #endif
2858         } else if (tp->t_srtt != 0) {
2859                 int delta;
2860
2861                 /*
2862                  * srtt is stored as fixed point with 5 bits after the
2863                  * binary point (i.e., scaled by 8).  The following magic
2864                  * is equivalent to the smoothing algorithm in rfc793 with
2865                  * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2866                  * point).  Adjust rtt to origin 0.
2867                  */
2868                 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2869                         - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2870
2871                 if ((tp->t_srtt += delta) <= 0)
2872                         tp->t_srtt = 1;
2873
2874                 /*
2875                  * We accumulate a smoothed rtt variance (actually, a
2876                  * smoothed mean difference), then set the retransmit
2877                  * timer to smoothed rtt + 4 times the smoothed variance.
2878                  * rttvar is stored as fixed point with 4 bits after the
2879                  * binary point (scaled by 16).  The following is
2880                  * equivalent to rfc793 smoothing with an alpha of .75
2881                  * (rttvar = rttvar*3/4 + |delta| / 4).  This replaces
2882                  * rfc793's wired-in beta.
2883                  */
2884                 if (delta < 0)
2885                         delta = -delta;
2886                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2887                 if ((tp->t_rttvar += delta) <= 0)
2888                         tp->t_rttvar = 1;
2889                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2890                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2891         } else {
2892                 /*
2893                  * No rtt measurement yet - use the unsmoothed rtt.
2894                  * Set the variance to half the rtt (so our first
2895                  * retransmit happens at 3*rtt).
2896                  */
2897                 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2898                 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2899                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2900         }
2901         tp->t_rtttime = 0;
2902         tp->t_rxtshift = 0;
2903
2904 #ifdef DEBUG_EIFEL_RESPONSE
2905         if (rebaserto) {
2906                 kprintf("| rxtcur prev %d, old %d, ",
2907                     tp->t_rxtcur_prev, tp->t_rxtcur);
2908         }
2909 #endif
2910
2911         /*
2912          * the retransmit should happen at rtt + 4 * rttvar.
2913          * Because of the way we do the smoothing, srtt and rttvar
2914          * will each average +1/2 tick of bias.  When we compute
2915          * the retransmit timer, we want 1/2 tick of rounding and
2916          * 1 extra tick because of +-1/2 tick uncertainty in the
2917          * firing of the timer.  The bias will give us exactly the
2918          * 1.5 tick we need.  But, because the bias is
2919          * statistical, we have to test that we don't drop below
2920          * the minimum feasible timer (which is 2 ticks).
2921          */
2922         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2923                       max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2924
2925         if (rebaserto) {
2926                 if (tp->t_rxtcur < tp->t_rxtcur_prev + tcp_eifel_rtoinc) {
2927                         /*
2928                          * RFC4015 requires that the new RTO is at least
2929                          * 2*G (tcp_eifel_rtoinc) greater then the RTO
2930                          * (t_rxtcur_prev) when the spurious retransmit
2931                          * timeout happens.
2932                          *
2933                          * The above condition could be true, if the SRTT
2934                          * and RTTVAR used to calculate t_rxtcur_prev
2935                          * resulted in a value less than t_rttmin.  So
2936                          * simply increasing SRTT by tcp_eifel_rtoinc when
2937                          * preparing for the Eifel response could not ensure
2938                          * that the new RTO will be tcp_eifel_rtoinc greater
2939                          * t_rxtcur_prev.
2940                          */
2941                         tp->t_rxtcur = tp->t_rxtcur_prev + tcp_eifel_rtoinc;
2942                 }
2943 #ifdef DEBUG_EIFEL_RESPONSE
2944                 kprintf("new %d\n", tp->t_rxtcur);
2945 #endif
2946         }
2947
2948         /*
2949          * We received an ack for a packet that wasn't retransmitted;
2950          * it is probably safe to discard any error indications we've
2951          * received recently.  This isn't quite right, but close enough
2952          * for now (a route might have failed after we sent a segment,
2953          * and the return path might not be symmetrical).
2954          */
2955         tp->t_softerror = 0;
2956 }
2957
2958 /*
2959  * Determine a reasonable value for maxseg size.
2960  * If the route is known, check route for mtu.
2961  * If none, use an mss that can be handled on the outgoing
2962  * interface without forcing IP to fragment; if bigger than
2963  * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2964  * to utilize large mbufs.  If no route is found, route has no mtu,
2965  * or the destination isn't local, use a default, hopefully conservative
2966  * size (usually 512 or the default IP max size, but no more than the mtu
2967  * of the interface), as we can't discover anything about intervening
2968  * gateways or networks.  We also initialize the congestion/slow start
2969  * window to be a single segment if the destination isn't local.
2970  * While looking at the routing entry, we also initialize other path-dependent
2971  * parameters from pre-set or cached values in the routing entry.
2972  *
2973  * Also take into account the space needed for options that we
2974  * send regularly.  Make maxseg shorter by that amount to assure
2975  * that we can send maxseg amount of data even when the options
2976  * are present.  Store the upper limit of the length of options plus
2977  * data in maxopd.
2978  *
2979  * NOTE that this routine is only called when we process an incoming
2980  * segment, for outgoing segments only tcp_mssopt is called.
2981  */
2982 void
2983 tcp_mss(struct tcpcb *tp, int offer)
2984 {
2985         struct rtentry *rt;
2986         struct ifnet *ifp;
2987         int rtt, mss;
2988         u_long bufsize;
2989         struct inpcb *inp = tp->t_inpcb;
2990         struct socket *so;
2991 #ifdef INET6
2992         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2993         size_t min_protoh = isipv6 ?
2994                             sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
2995                             sizeof(struct tcpiphdr);
2996 #else
2997         const boolean_t isipv6 = FALSE;
2998         const size_t min_protoh = sizeof(struct tcpiphdr);
2999 #endif
3000
3001         if (isipv6)
3002                 rt = tcp_rtlookup6(&inp->inp_inc);
3003         else
3004                 rt = tcp_rtlookup(&inp->inp_inc);
3005         if (rt == NULL) {
3006                 tp->t_maxopd = tp->t_maxseg =
3007                     (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3008                 return;
3009         }
3010         ifp = rt->rt_ifp;
3011         so = inp->inp_socket;
3012
3013         /*
3014          * Offer == 0 means that there was no MSS on the SYN segment,
3015          * in this case we use either the interface mtu or tcp_mssdflt.
3016          *
3017          * An offer which is too large will be cut down later.
3018          */
3019         if (offer == 0) {
3020                 if (isipv6) {
3021                         if (in6_localaddr(&inp->in6p_faddr)) {
3022                                 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
3023                                         min_protoh;
3024                         } else {
3025                                 offer = tcp_v6mssdflt;
3026                         }
3027                 } else {
3028                         if (in_localaddr(inp->inp_faddr))
3029                                 offer = ifp->if_mtu - min_protoh;
3030                         else
3031                                 offer = tcp_mssdflt;
3032                 }
3033         }
3034
3035         /*
3036          * Prevent DoS attack with too small MSS. Round up
3037          * to at least minmss.
3038          *
3039          * Sanity check: make sure that maxopd will be large
3040          * enough to allow some data on segments even is the
3041          * all the option space is used (40bytes).  Otherwise
3042          * funny things may happen in tcp_output.
3043          */
3044         offer = max(offer, tcp_minmss);
3045         offer = max(offer, 64);
3046
3047         rt->rt_rmx.rmx_mssopt = offer;
3048
3049         /*
3050          * While we're here, check if there's an initial rtt
3051          * or rttvar.  Convert from the route-table units
3052          * to scaled multiples of the slow timeout timer.
3053          */
3054         if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
3055                 /*
3056                  * XXX the lock bit for RTT indicates that the value
3057                  * is also a minimum value; this is subject to time.
3058                  */
3059                 if (rt->rt_rmx.rmx_locks & RTV_RTT)
3060                         tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
3061                 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
3062                 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
3063                 tcpstat.tcps_usedrtt++;
3064                 if (rt->rt_rmx.rmx_rttvar) {
3065                         tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
3066                             (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
3067                         tcpstat.tcps_usedrttvar++;
3068                 } else {
3069                         /* default variation is +- 1 rtt */
3070                         tp->t_rttvar =
3071                             tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
3072                 }
3073                 TCPT_RANGESET(tp->t_rxtcur,
3074                               ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
3075                               tp->t_rttmin, TCPTV_REXMTMAX);
3076         }
3077
3078         /*
3079          * if there's an mtu associated with the route, use it
3080          * else, use the link mtu.  Take the smaller of mss or offer
3081          * as our final mss.
3082          */
3083         if (rt->rt_rmx.rmx_mtu) {
3084                 mss = rt->rt_rmx.rmx_mtu - min_protoh;
3085         } else {
3086                 if (isipv6)
3087                         mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
3088                 else
3089                         mss = ifp->if_mtu - min_protoh;
3090         }
3091         mss = min(mss, offer);
3092
3093         /*
3094          * maxopd stores the maximum length of data AND options
3095          * in a segment; maxseg is the amount of data in a normal
3096          * segment.  We need to store this value (maxopd) apart
3097          * from maxseg, because now every segment carries options
3098          * and thus we normally have somewhat less data in segments.
3099          */
3100         tp->t_maxopd = mss;
3101
3102         if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
3103             ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3104                 mss -= TCPOLEN_TSTAMP_APPA;
3105
3106 #if     (MCLBYTES & (MCLBYTES - 1)) == 0
3107         if (mss > MCLBYTES)
3108                 mss &= ~(MCLBYTES-1);
3109 #else
3110         if (mss > MCLBYTES)
3111                 mss = mss / MCLBYTES * MCLBYTES;
3112 #endif
3113         /*
3114          * If there's a pipesize, change the socket buffer
3115          * to that size.  Make the socket buffers an integral
3116          * number of mss units; if the mss is larger than
3117          * the socket buffer, decrease the mss.
3118          */
3119 #ifdef RTV_SPIPE
3120         if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3121 #endif
3122                 bufsize = so->so_snd.ssb_hiwat;
3123         if (bufsize < mss)
3124                 mss = bufsize;
3125         else {
3126                 bufsize = roundup(bufsize, mss);
3127                 if (bufsize > sb_max)
3128                         bufsize = sb_max;
3129                 if (bufsize > so->so_snd.ssb_hiwat)
3130                         ssb_reserve(&so->so_snd, bufsize, so, NULL);
3131         }
3132         tp->t_maxseg = mss;
3133
3134 #ifdef RTV_RPIPE
3135         if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3136 #endif
3137                 bufsize = so->so_rcv.ssb_hiwat;
3138         if (bufsize > mss) {
3139                 bufsize = roundup(bufsize, mss);
3140                 if (bufsize > sb_max)
3141                         bufsize = sb_max;
3142                 if (bufsize > so->so_rcv.ssb_hiwat) {
3143                         lwkt_gettoken(&so->so_rcv.ssb_token);
3144                         ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3145                         lwkt_reltoken(&so->so_rcv.ssb_token);
3146                 }
3147         }
3148
3149         /*
3150          * Set the slow-start flight size
3151          *
3152          * NOTE: t_maxseg must have been configured!
3153          */
3154         tp->snd_cwnd = tcp_initial_window(tp);
3155
3156         if (rt->rt_rmx.rmx_ssthresh) {
3157                 /*
3158                  * There's some sort of gateway or interface
3159                  * buffer limit on the path.  Use this to set
3160                  * the slow start threshhold, but set the
3161                  * threshold to no less than 2*mss.
3162                  */
3163                 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3164                 tcpstat.tcps_usedssthresh++;
3165         }
3166 }
3167
3168 /*
3169  * Determine the MSS option to send on an outgoing SYN.
3170  */
3171 int
3172 tcp_mssopt(struct tcpcb *tp)
3173 {
3174         struct rtentry *rt;
3175 #ifdef INET6
3176         boolean_t isipv6 =
3177             ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3178         int min_protoh = isipv6 ?
3179                              sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3180                              sizeof(struct tcpiphdr);
3181 #else
3182         const boolean_t isipv6 = FALSE;
3183         const size_t min_protoh = sizeof(struct tcpiphdr);
3184 #endif
3185
3186         if (isipv6)
3187                 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3188         else
3189                 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3190         if (rt == NULL)
3191                 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3192
3193         return (rt->rt_ifp->if_mtu - min_protoh);
3194 }
3195
3196 /*
3197  * When a partial ack arrives, force the retransmission of the
3198  * next unacknowledged segment.  Do not exit Fast Recovery.
3199  *
3200  * Implement the Slow-but-Steady variant of NewReno by restarting the
3201  * the retransmission timer.  Turn it off here so it can be restarted
3202  * later in tcp_output().
3203  */
3204 static void
3205 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3206 {
3207         tcp_seq old_snd_nxt = tp->snd_nxt;
3208         u_long ocwnd = tp->snd_cwnd;
3209
3210         tcp_callout_stop(tp, tp->tt_rexmt);
3211         tp->t_rtttime = 0;
3212         tp->snd_nxt = th->th_ack;
3213         /* Set snd_cwnd to one segment beyond acknowledged offset. */
3214         tp->snd_cwnd = tp->t_maxseg;
3215         tp->t_flags |= TF_ACKNOW;
3216         tcp_output(tp);
3217         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3218                 tp->snd_nxt = old_snd_nxt;
3219         /* partial window deflation */
3220         if (ocwnd > acked)
3221                 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3222         else
3223                 tp->snd_cwnd = tp->t_maxseg;
3224 }
3225
3226 /*
3227  * In contrast to the Slow-but-Steady NewReno variant,
3228  * we do not reset the retransmission timer for SACK retransmissions,
3229  * except when retransmitting snd_una.
3230  */
3231 static void
3232 tcp_sack_rexmt(struct tcpcb *tp, boolean_t force)
3233 {
3234         tcp_seq old_snd_nxt = tp->snd_nxt;
3235         u_long ocwnd = tp->snd_cwnd;
3236         uint32_t pipe;
3237         int nseg = 0;           /* consecutive new segments */
3238         int nseg_rexmt = 0;     /* retransmitted segments */
3239         int maxrexmt = 0;
3240
3241         if (force) {
3242                 uint32_t unsacked = tcp_sack_first_unsacked_len(tp);
3243
3244                 /*
3245                  * Try to fill the first hole in the receiver's
3246                  * reassemble queue.
3247                  */
3248                 maxrexmt = howmany(unsacked, tp->t_maxseg);
3249                 if (maxrexmt > tcp_force_sackrxt)
3250                         maxrexmt = tcp_force_sackrxt;
3251         }
3252
3253         tp->t_rtttime = 0;
3254         pipe = tcp_sack_compute_pipe(tp);
3255         while (((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg
3256                 || (force && nseg_rexmt < maxrexmt && nseg == 0)) &&
3257             (!tcp_do_smartsack || nseg < TCP_SACK_MAXBURST)) {
3258                 tcp_seq old_snd_max, old_rexmt_high, nextrexmt;
3259                 uint32_t sent, seglen;
3260                 boolean_t rescue;
3261                 int error;
3262
3263                 old_rexmt_high = tp->rexmt_high;
3264                 if (!tcp_sack_nextseg(tp, &nextrexmt, &seglen, &rescue)) {
3265                         tp->rexmt_high = old_rexmt_high;
3266                         break;
3267                 }
3268
3269                 /*
3270                  * If the next tranmission is a rescue retranmission,
3271                  * we check whether we have already sent some data
3272                  * (either new segments or retransmitted segments)
3273                  * into the the network or not.  Since the idea of rescue
3274                  * retransmission is to sustain ACK clock, as long as
3275                  * some segments are in the network, ACK clock will be
3276                  * kept ticking.
3277                  */
3278                 if (rescue && (nseg_rexmt > 0 || nseg > 0)) {
3279                         tp->rexmt_high = old_rexmt_high;
3280                         break;
3281                 }
3282
3283                 if (nextrexmt == tp->snd_max)
3284                         ++nseg;
3285                 else
3286                         ++nseg_rexmt;
3287                 tp->snd_nxt = nextrexmt;
3288                 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3289                 old_snd_max = tp->snd_max;
3290                 if (nextrexmt == tp->snd_una)
3291                         tcp_callout_stop(tp, tp->tt_rexmt);
3292                 tp->t_flags |= TF_XMITNOW;
3293                 error = tcp_output(tp);
3294                 if (error != 0) {
3295                         tp->rexmt_high = old_rexmt_high;
3296                         break;
3297                 }
3298                 sent = tp->snd_nxt - nextrexmt;
3299                 if (sent <= 0) {
3300                         tp->rexmt_high = old_rexmt_high;
3301                         break;
3302                 }
3303                 pipe += sent;
3304                 tcpstat.tcps_sndsackpack++;
3305                 tcpstat.tcps_sndsackbyte += sent;
3306
3307                 if (rescue) {
3308                         tcpstat.tcps_sackrescue++;
3309                         tp->rexmt_rescue = tp->snd_nxt;
3310                         tp->sack_flags |= TSACK_F_SACKRESCUED;
3311                         break;
3312                 }
3313                 if (SEQ_LT(nextrexmt, old_snd_max) &&
3314                     SEQ_LT(tp->rexmt_high, tp->snd_nxt)) {
3315                         tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3316                         if (tcp_aggressive_rescuesack &&
3317                             (tp->sack_flags & TSACK_F_SACKRESCUED) &&
3318                             SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
3319                                 /* Drag RescueRxt along with HighRxt */
3320                                 tp->rexmt_rescue = tp->rexmt_high;
3321                         }
3322                 }
3323         }
3324         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3325                 tp->snd_nxt = old_snd_nxt;
3326         tp->snd_cwnd = ocwnd;
3327 }
3328
3329 /*
3330  * Return TRUE, if some new segments are sent
3331  */
3332 static boolean_t
3333 tcp_sack_limitedxmit(struct tcpcb *tp)
3334 {
3335         tcp_seq oldsndnxt = tp->snd_nxt;
3336         tcp_seq oldsndmax = tp->snd_max;
3337         u_long ocwnd = tp->snd_cwnd;
3338         uint32_t pipe, sent;
3339         boolean_t ret = FALSE;
3340         tcp_seq_diff_t cwnd_left;
3341         tcp_seq next;
3342
3343         tp->rexmt_high = tp->snd_una - 1;
3344         pipe = tcp_sack_compute_pipe(tp);
3345         cwnd_left = (tcp_seq_diff_t)(ocwnd - pipe);
3346         if (cwnd_left < (tcp_seq_diff_t)tp->t_maxseg)
3347                 return FALSE;
3348
3349         if (tcp_do_smartsack)
3350                 cwnd_left = ulmin(cwnd_left, tp->t_maxseg * TCP_SACK_MAXBURST);
3351
3352         next = tp->snd_nxt = tp->snd_max;
3353         tp->snd_cwnd = tp->snd_nxt - tp->snd_una +
3354             rounddown(cwnd_left, tp->t_maxseg);
3355
3356         tp->t_flags |= TF_XMITNOW;
3357         tcp_output(tp);
3358
3359         sent = tp->snd_nxt - next;
3360         if (sent > 0) {
3361                 tcpstat.tcps_sndlimited += howmany(sent, tp->t_maxseg);
3362                 ret = TRUE;
3363         }
3364
3365         if (SEQ_LT(oldsndnxt, oldsndmax)) {
3366                 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una),
3367                     ("snd_una moved in other threads"));
3368                 tp->snd_nxt = oldsndnxt;
3369         }
3370         tp->snd_cwnd = ocwnd;
3371
3372         if (ret && TCP_DO_NCR(tp))
3373                 tcp_ncr_update_rxtthresh(tp);
3374
3375         return ret;
3376 }
3377
3378 /*
3379  * Reset idle time and keep-alive timer, typically called when a valid
3380  * tcp packet is received but may also be called when FASTKEEP is set
3381  * to prevent the previous long-timeout from calculating to a drop.
3382  *
3383  * Only update t_rcvtime for non-SYN packets.
3384  *
3385  * Handle the case where one side thinks the connection is established
3386  * but the other side has, say, rebooted without cleaning out the
3387  * connection.   The SYNs could be construed as an attack and wind
3388  * up ignored, but in case it isn't an attack we can validate the
3389  * connection by forcing a keepalive.
3390  */
3391 void
3392 tcp_timer_keep_activity(struct tcpcb *tp, int thflags)
3393 {
3394         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
3395                 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
3396                         tp->t_flags |= TF_KEEPALIVE;
3397                         tcp_callout_reset(tp, tp->tt_keep, hz / 2,
3398                                           tcp_timer_keep);
3399                 } else {
3400                         tp->t_rcvtime = ticks;
3401                         tp->t_flags &= ~TF_KEEPALIVE;
3402                         tcp_callout_reset(tp, tp->tt_keep,
3403                                           tp->t_keepidle,
3404                                           tcp_timer_keep);
3405                 }
3406         }
3407 }
3408
3409 static int
3410 tcp_rmx_msl(const struct tcpcb *tp)
3411 {
3412         struct rtentry *rt;
3413         struct inpcb *inp = tp->t_inpcb;
3414         int msl;
3415 #ifdef INET6
3416         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3417 #else
3418         const boolean_t isipv6 = FALSE;
3419 #endif
3420
3421         if (isipv6)
3422                 rt = tcp_rtlookup6(&inp->inp_inc);
3423         else
3424                 rt = tcp_rtlookup(&inp->inp_inc);
3425         if (rt == NULL || rt->rt_rmx.rmx_msl == 0)
3426                 return tcp_msl;
3427
3428         msl = (rt->rt_rmx.rmx_msl * hz) / 1000;
3429         if (msl == 0)
3430                 msl = 1;
3431
3432         return msl;
3433 }
3434
3435 static void
3436 tcp_established(struct tcpcb *tp)
3437 {
3438         tp->t_state = TCPS_ESTABLISHED;
3439         tcp_callout_reset(tp, tp->tt_keep, tp->t_keepidle, tcp_timer_keep);
3440
3441         if (tp->t_rxtsyn > 0) {
3442                 /*
3443                  * RFC6298:
3444                  * "If the timer expires awaiting the ACK of a SYN segment
3445                  *  and the TCP implementation is using an RTO less than 3
3446                  *  seconds, the RTO MUST be re-initialized to 3 seconds
3447                  *  when data transmission begins"
3448                  */
3449                 if (tp->t_rxtcur < TCPTV_RTOBASE3)
3450                         tp->t_rxtcur = TCPTV_RTOBASE3;
3451         }
3452 }
3453
3454 /*
3455  * Returns TRUE, if the ACK should be dropped
3456  */
3457 static boolean_t
3458 tcp_recv_dupack(struct tcpcb *tp, tcp_seq th_ack, u_int to_flags)
3459 {
3460         boolean_t fast_sack_rexmt = TRUE;
3461
3462         tcpstat.tcps_rcvdupack++;
3463
3464         /*
3465          * We have outstanding data (other than a window probe),
3466          * this is a completely duplicate ack (ie, window info
3467          * didn't change), the ack is the biggest we've seen and
3468          * we've seen exactly our rexmt threshhold of them, so
3469          * assume a packet has been dropped and retransmit it.
3470          * Kludge snd_nxt & the congestion window so we send only
3471          * this one packet.
3472          */
3473         if (IN_FASTRECOVERY(tp)) {
3474                 if (TCP_DO_SACK(tp)) {
3475                         boolean_t force = FALSE;
3476
3477                         if (tp->snd_una == tp->rexmt_high &&
3478                             (to_flags & (TOF_SACK | TOF_SACK_REDUNDANT)) ==
3479                             TOF_SACK) {
3480                                 /*
3481                                  * New segments got SACKed and
3482                                  * no retransmit yet.
3483                                  */
3484                                 force = TRUE;
3485                         }
3486
3487                         /* No artifical cwnd inflation. */
3488                         tcp_sack_rexmt(tp, force);
3489                 } else {
3490                         /*
3491                          * Dup acks mean that packets have left
3492                          * the network (they're now cached at the
3493                          * receiver) so bump cwnd by the amount in
3494                          * the receiver to keep a constant cwnd
3495                          * packets in the network.
3496                          */
3497                         tp->snd_cwnd += tp->t_maxseg;
3498                         tcp_output(tp);
3499                 }
3500                 return TRUE;
3501         } else if (SEQ_LT(th_ack, tp->snd_recover)) {
3502                 tp->t_dupacks = 0;
3503                 return FALSE;
3504         } else if (tcp_ignore_redun_dsack && TCP_DO_SACK(tp) &&
3505             (to_flags & (TOF_DSACK | TOF_SACK_REDUNDANT)) ==
3506             (TOF_DSACK | TOF_SACK_REDUNDANT)) {
3507                 /*
3508                  * If the ACK carries DSACK and other SACK blocks
3509                  * carry information that we have already known,
3510                  * don't count this ACK as duplicate ACK.  This
3511                  * prevents spurious early retransmit and fast
3512                  * retransmit.  This also meets the requirement of
3513                  * RFC3042 that new segments should not be sent if
3514                  * the SACK blocks do not contain new information
3515