538cc8f145f6c416b1b57287e2443a171a2eb8fc
[dragonfly.git] / sys / netinet / tcp_input.c
1 /*
2  * Copyright (c) 2002, 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2002, 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
36  *      The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *      This product includes software developed by the University of
49  *      California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_input.c,v 1.107.2.38 2003/05/21 04:46:41 cjc Exp $
68  */
69
70 #include "opt_inet.h"
71 #include "opt_inet6.h"
72 #include "opt_ipsec.h"
73 #include "opt_tcpdebug.h"
74 #include "opt_tcp_input.h"
75
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
80 #include <sys/malloc.h>
81 #include <sys/mbuf.h>
82 #include <sys/proc.h>           /* for proc0 declaration */
83 #include <sys/protosw.h>
84 #include <sys/socket.h>
85 #include <sys/socketvar.h>
86 #include <sys/syslog.h>
87 #include <sys/in_cksum.h>
88
89 #include <sys/socketvar2.h>
90
91 #include <machine/cpu.h>        /* before tcp_seq.h, for tcp_random18() */
92 #include <machine/stdarg.h>
93
94 #include <net/if.h>
95 #include <net/route.h>
96
97 #include <netinet/in.h>
98 #include <netinet/in_systm.h>
99 #include <netinet/ip.h>
100 #include <netinet/ip_icmp.h>    /* for ICMP_BANDLIM */
101 #include <netinet/in_var.h>
102 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
103 #include <netinet/in_pcb.h>
104 #include <netinet/ip_var.h>
105 #include <netinet/ip6.h>
106 #include <netinet/icmp6.h>
107 #include <netinet6/nd6.h>
108 #include <netinet6/ip6_var.h>
109 #include <netinet6/in6_pcb.h>
110 #include <netinet/tcp.h>
111 #include <netinet/tcp_fsm.h>
112 #include <netinet/tcp_seq.h>
113 #include <netinet/tcp_timer.h>
114 #include <netinet/tcp_timer2.h>
115 #include <netinet/tcp_var.h>
116 #include <netinet6/tcp6_var.h>
117 #include <netinet/tcpip.h>
118
119 #ifdef TCPDEBUG
120 #include <netinet/tcp_debug.h>
121
122 u_char tcp_saveipgen[40];    /* the size must be of max ip header, now IPv6 */
123 struct tcphdr tcp_savetcp;
124 #endif
125
126 #ifdef FAST_IPSEC
127 #include <netproto/ipsec/ipsec.h>
128 #include <netproto/ipsec/ipsec6.h>
129 #endif
130
131 #ifdef IPSEC
132 #include <netinet6/ipsec.h>
133 #include <netinet6/ipsec6.h>
134 #include <netproto/key/key.h>
135 #endif
136
137 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
138
139 static int log_in_vain = 0;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW,
141     &log_in_vain, 0, "Log all incoming TCP connections");
142
143 static int blackhole = 0;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW,
145     &blackhole, 0, "Do not send RST when dropping refused connections");
146
147 int tcp_delack_enabled = 1;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW,
149     &tcp_delack_enabled, 0,
150     "Delay ACK to try and piggyback it onto a data packet");
151
152 #ifdef TCP_DROP_SYNFIN
153 static int drop_synfin = 0;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW,
155     &drop_synfin, 0, "Drop TCP packets with SYN+FIN set");
156 #endif
157
158 static int tcp_do_limitedtransmit = 1;
159 SYSCTL_INT(_net_inet_tcp, OID_AUTO, limitedtransmit, CTLFLAG_RW,
160     &tcp_do_limitedtransmit, 0, "Enable RFC 3042 (Limited Transmit)");
161
162 static int tcp_do_early_retransmit = 1;
163 SYSCTL_INT(_net_inet_tcp, OID_AUTO, earlyretransmit, CTLFLAG_RW,
164     &tcp_do_early_retransmit, 0, "Early retransmit");
165
166 int tcp_aggregate_acks = 1;
167 SYSCTL_INT(_net_inet_tcp, OID_AUTO, aggregate_acks, CTLFLAG_RW,
168     &tcp_aggregate_acks, 0, "Aggregate built-up acks into one ack");
169
170 static int tcp_do_eifel_detect = 1;
171 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
172     &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
173
174 static int tcp_do_abc = 1;
175 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
176     &tcp_do_abc, 0,
177     "TCP Appropriate Byte Counting (RFC 3465)");
178
179 /*
180  * Define as tunable for easy testing with SACK on and off.
181  * Warning:  do not change setting in the middle of an existing active TCP flow,
182  *   else strange things might happen to that flow.
183  */
184 int tcp_do_sack = 1;
185 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
186     &tcp_do_sack, 0, "Enable SACK Algorithms");
187
188 int tcp_do_smartsack = 1;
189 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
190     &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
191
192 int tcp_do_rescuesack = 1;
193 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack, CTLFLAG_RW,
194     &tcp_do_rescuesack, 0, "Rescue retransmission for SACK");
195
196 int tcp_aggressive_rescuesack = 0;
197 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack_agg, CTLFLAG_RW,
198     &tcp_aggressive_rescuesack, 0, "Aggressive rescue retransmission for SACK");
199
200 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
201     "TCP Segment Reassembly Queue");
202
203 int tcp_reass_maxseg = 0;
204 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
205     &tcp_reass_maxseg, 0,
206     "Global maximum number of TCP Segments in Reassembly Queue");
207
208 int tcp_reass_qsize = 0;
209 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
210     &tcp_reass_qsize, 0,
211     "Global number of TCP Segments currently in Reassembly Queue");
212
213 static int tcp_reass_overflows = 0;
214 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
215     &tcp_reass_overflows, 0,
216     "Global number of TCP Segment Reassembly Queue Overflows");
217
218 int tcp_do_autorcvbuf = 1;
219 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
220     &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
221
222 int tcp_autorcvbuf_inc = 16*1024;
223 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
224     &tcp_autorcvbuf_inc, 0,
225     "Incrementor step size of automatic receive buffer");
226
227 int tcp_autorcvbuf_max = 2*1024*1024;
228 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
229     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
230
231 int tcp_sosend_agglim = 2;
232 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_agglim, CTLFLAG_RW,
233     &tcp_sosend_agglim, 0, "TCP sosend mbuf aggregation limit");
234
235 int tcp_sosend_async = 1;
236 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_async, CTLFLAG_RW,
237     &tcp_sosend_async, 0, "TCP asynchronized pru_send");
238
239 static int tcp_ignore_redun_dsack = 1;
240 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_redun_dsack, CTLFLAG_RW,
241     &tcp_ignore_redun_dsack, 0, "Ignore redundant DSACK");
242
243 static void      tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t,
244                     tcp_seq);
245 static void      tcp_pulloutofband(struct socket *,
246                      struct tcphdr *, struct mbuf *, int);
247 static int       tcp_reass(struct tcpcb *, struct tcphdr *, int *,
248                      struct mbuf *);
249 static void      tcp_xmit_timer(struct tcpcb *, int, tcp_seq);
250 static void      tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
251 static void      tcp_sack_rexmt(struct tcpcb *, struct tcphdr *);
252 static int       tcp_rmx_msl(const struct tcpcb *);
253 static void      tcp_established(struct tcpcb *);
254
255 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
256 #ifdef INET6
257 #define ND6_HINT(tp) \
258 do { \
259         if ((tp) && (tp)->t_inpcb && \
260             ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
261             (tp)->t_inpcb->in6p_route.ro_rt) \
262                 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
263 } while (0)
264 #else
265 #define ND6_HINT(tp)
266 #endif
267
268 /*
269  * Indicate whether this ack should be delayed.  We can delay the ack if
270  *      - delayed acks are enabled and
271  *      - there is no delayed ack timer in progress and
272  *      - our last ack wasn't a 0-sized window.  We never want to delay
273  *        the ack that opens up a 0-sized window.
274  */
275 #define DELAY_ACK(tp) \
276         (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
277         !(tp->t_flags & TF_RXWIN0SENT))
278
279 #define acceptable_window_update(tp, th, tiwin)                         \
280     (SEQ_LT(tp->snd_wl1, th->th_seq) ||                                 \
281      (tp->snd_wl1 == th->th_seq &&                                      \
282       (SEQ_LT(tp->snd_wl2, th->th_ack) ||                               \
283        (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
284
285 static int
286 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
287 {
288         struct tseg_qent *q;
289         struct tseg_qent *p = NULL;
290         struct tseg_qent *te;
291         struct socket *so = tp->t_inpcb->inp_socket;
292         int flags;
293
294         /*
295          * Call with th == NULL after become established to
296          * force pre-ESTABLISHED data up to user socket.
297          */
298         if (th == NULL)
299                 goto present;
300
301         /*
302          * Limit the number of segments in the reassembly queue to prevent
303          * holding on to too many segments (and thus running out of mbufs).
304          * Make sure to let the missing segment through which caused this
305          * queue.  Always keep one global queue entry spare to be able to
306          * process the missing segment.
307          */
308         if (th->th_seq != tp->rcv_nxt &&
309             tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
310                 tcp_reass_overflows++;
311                 tcpstat.tcps_rcvmemdrop++;
312                 m_freem(m);
313                 /* no SACK block to report */
314                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
315                 return (0);
316         }
317
318         /* Allocate a new queue entry. */
319         te = kmalloc(sizeof(struct tseg_qent), M_TSEGQ, M_INTWAIT | M_NULLOK);
320         if (te == NULL) {
321                 tcpstat.tcps_rcvmemdrop++;
322                 m_freem(m);
323                 /* no SACK block to report */
324                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
325                 return (0);
326         }
327         atomic_add_int(&tcp_reass_qsize, 1);
328
329         /*
330          * Find a segment which begins after this one does.
331          */
332         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
333                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
334                         break;
335                 p = q;
336         }
337
338         /*
339          * If there is a preceding segment, it may provide some of
340          * our data already.  If so, drop the data from the incoming
341          * segment.  If it provides all of our data, drop us.
342          */
343         if (p != NULL) {
344                 tcp_seq_diff_t i;
345
346                 /* conversion to int (in i) handles seq wraparound */
347                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
348                 if (i > 0) {            /* overlaps preceding segment */
349                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
350                         /* enclosing block starts w/ preceding segment */
351                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
352                         if (i >= *tlenp) {
353                                 /* preceding encloses incoming segment */
354                                 tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
355                                     p->tqe_th->th_seq + p->tqe_len,
356                                     p->tqe_th->th_flags);
357                                 tcpstat.tcps_rcvduppack++;
358                                 tcpstat.tcps_rcvdupbyte += *tlenp;
359                                 m_freem(m);
360                                 kfree(te, M_TSEGQ);
361                                 atomic_add_int(&tcp_reass_qsize, -1);
362                                 /*
363                                  * Try to present any queued data
364                                  * at the left window edge to the user.
365                                  * This is needed after the 3-WHS
366                                  * completes.
367                                  */
368                                 goto present;   /* ??? */
369                         }
370                         m_adj(m, i);
371                         *tlenp -= i;
372                         th->th_seq += i;
373                         /* incoming segment end is enclosing block end */
374                         tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
375                             th->th_seq + *tlenp, th->th_flags);
376                         /* trim end of reported D-SACK block */
377                         tp->reportblk.rblk_end = th->th_seq;
378                 }
379         }
380         tcpstat.tcps_rcvoopack++;
381         tcpstat.tcps_rcvoobyte += *tlenp;
382
383         /*
384          * While we overlap succeeding segments trim them or,
385          * if they are completely covered, dequeue them.
386          */
387         while (q) {
388                 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
389                 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
390                 tcp_seq qend_sack = TCP_SACK_BLKEND(qend, q->tqe_th->th_flags);
391                 struct tseg_qent *nq;
392
393                 if (i <= 0)
394                         break;
395                 if (!(tp->t_flags & TF_DUPSEG)) {    /* first time through */
396                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
397                         tp->encloseblk = tp->reportblk;
398                         /* report trailing duplicate D-SACK segment */
399                         tp->reportblk.rblk_start = q->tqe_th->th_seq;
400                 }
401                 if ((tp->t_flags & TF_ENCLOSESEG) &&
402                     SEQ_GT(qend_sack, tp->encloseblk.rblk_end)) {
403                         /* extend enclosing block if one exists */
404                         tp->encloseblk.rblk_end = qend_sack;
405                 }
406                 if (i < q->tqe_len) {
407                         q->tqe_th->th_seq += i;
408                         q->tqe_len -= i;
409                         m_adj(q->tqe_m, i);
410                         break;
411                 }
412
413                 nq = LIST_NEXT(q, tqe_q);
414                 LIST_REMOVE(q, tqe_q);
415                 m_freem(q->tqe_m);
416                 kfree(q, M_TSEGQ);
417                 atomic_add_int(&tcp_reass_qsize, -1);
418                 q = nq;
419         }
420
421         /* Insert the new segment queue entry into place. */
422         te->tqe_m = m;
423         te->tqe_th = th;
424         te->tqe_len = *tlenp;
425
426         /* check if can coalesce with following segment */
427         if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
428                 tcp_seq tend = te->tqe_th->th_seq + te->tqe_len;
429                 tcp_seq tend_sack = TCP_SACK_BLKEND(tend, te->tqe_th->th_flags);
430
431                 te->tqe_len += q->tqe_len;
432                 if (q->tqe_th->th_flags & TH_FIN)
433                         te->tqe_th->th_flags |= TH_FIN;
434                 m_cat(te->tqe_m, q->tqe_m);
435                 tp->encloseblk.rblk_end = tend_sack;
436                 /*
437                  * When not reporting a duplicate segment, use
438                  * the larger enclosing block as the SACK block.
439                  */
440                 if (!(tp->t_flags & TF_DUPSEG))
441                         tp->reportblk.rblk_end = tend_sack;
442                 LIST_REMOVE(q, tqe_q);
443                 kfree(q, M_TSEGQ);
444                 atomic_add_int(&tcp_reass_qsize, -1);
445         }
446
447         if (p == NULL) {
448                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
449         } else {
450                 /* check if can coalesce with preceding segment */
451                 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
452                         p->tqe_len += te->tqe_len;
453                         m_cat(p->tqe_m, te->tqe_m);
454                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
455                         /*
456                          * When not reporting a duplicate segment, use
457                          * the larger enclosing block as the SACK block.
458                          */
459                         if (!(tp->t_flags & TF_DUPSEG))
460                                 tp->reportblk.rblk_start = p->tqe_th->th_seq;
461                         kfree(te, M_TSEGQ);
462                         atomic_add_int(&tcp_reass_qsize, -1);
463                 } else {
464                         LIST_INSERT_AFTER(p, te, tqe_q);
465                 }
466         }
467
468 present:
469         /*
470          * Present data to user, advancing rcv_nxt through
471          * completed sequence space.
472          */
473         if (!TCPS_HAVEESTABLISHED(tp->t_state))
474                 return (0);
475         q = LIST_FIRST(&tp->t_segq);
476         if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
477                 return (0);
478         tp->rcv_nxt += q->tqe_len;
479         if (!(tp->t_flags & TF_DUPSEG)) {
480                 /* no SACK block to report since ACK advanced */
481                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
482         }
483         /* no enclosing block to report since ACK advanced */
484         tp->t_flags &= ~TF_ENCLOSESEG;
485         flags = q->tqe_th->th_flags & TH_FIN;
486         LIST_REMOVE(q, tqe_q);
487         KASSERT(LIST_EMPTY(&tp->t_segq) ||
488                 LIST_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
489                 ("segment not coalesced"));
490         if (so->so_state & SS_CANTRCVMORE) {
491                 m_freem(q->tqe_m);
492         } else {
493                 lwkt_gettoken(&so->so_rcv.ssb_token);
494                 ssb_appendstream(&so->so_rcv, q->tqe_m);
495                 lwkt_reltoken(&so->so_rcv.ssb_token);
496         }
497         kfree(q, M_TSEGQ);
498         atomic_add_int(&tcp_reass_qsize, -1);
499         ND6_HINT(tp);
500         sorwakeup(so);
501         return (flags);
502 }
503
504 /*
505  * TCP input routine, follows pages 65-76 of the
506  * protocol specification dated September, 1981 very closely.
507  */
508 #ifdef INET6
509 int
510 tcp6_input(struct mbuf **mp, int *offp, int proto)
511 {
512         struct mbuf *m = *mp;
513         struct in6_ifaddr *ia6;
514
515         IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
516
517         /*
518          * draft-itojun-ipv6-tcp-to-anycast
519          * better place to put this in?
520          */
521         ia6 = ip6_getdstifaddr(m);
522         if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
523                 struct ip6_hdr *ip6;
524
525                 ip6 = mtod(m, struct ip6_hdr *);
526                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
527                             offsetof(struct ip6_hdr, ip6_dst));
528                 return (IPPROTO_DONE);
529         }
530
531         tcp_input(mp, offp, proto);
532         return (IPPROTO_DONE);
533 }
534 #endif
535
536 int
537 tcp_input(struct mbuf **mp, int *offp, int proto)
538 {
539         int off0;
540         struct tcphdr *th;
541         struct ip *ip = NULL;
542         struct ipovly *ipov;
543         struct inpcb *inp = NULL;
544         u_char *optp = NULL;
545         int optlen = 0;
546         int tlen, off;
547         int len = 0;
548         int drop_hdrlen;
549         struct tcpcb *tp = NULL;
550         int thflags;
551         struct socket *so = NULL;
552         int todrop, acked;
553         boolean_t ourfinisacked, needoutput = FALSE;
554         u_long tiwin;
555         int recvwin;
556         struct tcpopt to;               /* options in this segment */
557         struct sockaddr_in *next_hop = NULL;
558         int rstreason; /* For badport_bandlim accounting purposes */
559         int cpu;
560         struct ip6_hdr *ip6 = NULL;
561         struct mbuf *m;
562 #ifdef INET6
563         boolean_t isipv6;
564 #else
565         const boolean_t isipv6 = FALSE;
566 #endif
567 #ifdef TCPDEBUG
568         short ostate = 0;
569 #endif
570
571         off0 = *offp;
572         m = *mp;
573         *mp = NULL;
574
575         tcpstat.tcps_rcvtotal++;
576
577         if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
578                 struct m_tag *mtag;
579
580                 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
581                 KKASSERT(mtag != NULL);
582                 next_hop = m_tag_data(mtag);
583         }
584
585 #ifdef INET6
586         isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
587 #endif
588
589         if (isipv6) {
590                 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
591                 ip6 = mtod(m, struct ip6_hdr *);
592                 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
593                 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
594                         tcpstat.tcps_rcvbadsum++;
595                         goto drop;
596                 }
597                 th = (struct tcphdr *)((caddr_t)ip6 + off0);
598
599                 /*
600                  * Be proactive about unspecified IPv6 address in source.
601                  * As we use all-zero to indicate unbounded/unconnected pcb,
602                  * unspecified IPv6 address can be used to confuse us.
603                  *
604                  * Note that packets with unspecified IPv6 destination is
605                  * already dropped in ip6_input.
606                  */
607                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
608                         /* XXX stat */
609                         goto drop;
610                 }
611         } else {
612                 /*
613                  * Get IP and TCP header together in first mbuf.
614                  * Note: IP leaves IP header in first mbuf.
615                  */
616                 if (off0 > sizeof(struct ip)) {
617                         ip_stripoptions(m);
618                         off0 = sizeof(struct ip);
619                 }
620                 /* already checked and pulled up in ip_demux() */
621                 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
622                     ("TCP header not in one mbuf: m->m_len %d", m->m_len));
623                 ip = mtod(m, struct ip *);
624                 ipov = (struct ipovly *)ip;
625                 th = (struct tcphdr *)((caddr_t)ip + off0);
626                 tlen = ip->ip_len;
627
628                 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
629                         if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
630                                 th->th_sum = m->m_pkthdr.csum_data;
631                         else
632                                 th->th_sum = in_pseudo(ip->ip_src.s_addr,
633                                                 ip->ip_dst.s_addr,
634                                                 htonl(m->m_pkthdr.csum_data +
635                                                         ip->ip_len +
636                                                         IPPROTO_TCP));
637                         th->th_sum ^= 0xffff;
638                 } else {
639                         /*
640                          * Checksum extended TCP header and data.
641                          */
642                         len = sizeof(struct ip) + tlen;
643                         bzero(ipov->ih_x1, sizeof ipov->ih_x1);
644                         ipov->ih_len = (u_short)tlen;
645                         ipov->ih_len = htons(ipov->ih_len);
646                         th->th_sum = in_cksum(m, len);
647                 }
648                 if (th->th_sum) {
649                         tcpstat.tcps_rcvbadsum++;
650                         goto drop;
651                 }
652 #ifdef INET6
653                 /* Re-initialization for later version check */
654                 ip->ip_v = IPVERSION;
655 #endif
656         }
657
658         /*
659          * Check that TCP offset makes sense,
660          * pull out TCP options and adjust length.              XXX
661          */
662         off = th->th_off << 2;
663         /* already checked and pulled up in ip_demux() */
664         KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
665             ("bad TCP data offset %d (tlen %d)", off, tlen));
666         tlen -= off;    /* tlen is used instead of ti->ti_len */
667         if (off > sizeof(struct tcphdr)) {
668                 if (isipv6) {
669                         IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
670                         ip6 = mtod(m, struct ip6_hdr *);
671                         th = (struct tcphdr *)((caddr_t)ip6 + off0);
672                 } else {
673                         /* already pulled up in ip_demux() */
674                         KASSERT(m->m_len >= sizeof(struct ip) + off,
675                             ("TCP header and options not in one mbuf: "
676                              "m_len %d, off %d", m->m_len, off));
677                 }
678                 optlen = off - sizeof(struct tcphdr);
679                 optp = (u_char *)(th + 1);
680         }
681         thflags = th->th_flags;
682
683 #ifdef TCP_DROP_SYNFIN
684         /*
685          * If the drop_synfin option is enabled, drop all packets with
686          * both the SYN and FIN bits set. This prevents e.g. nmap from
687          * identifying the TCP/IP stack.
688          *
689          * This is a violation of the TCP specification.
690          */
691         if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
692                 goto drop;
693 #endif
694
695         /*
696          * Convert TCP protocol specific fields to host format.
697          */
698         th->th_seq = ntohl(th->th_seq);
699         th->th_ack = ntohl(th->th_ack);
700         th->th_win = ntohs(th->th_win);
701         th->th_urp = ntohs(th->th_urp);
702
703         /*
704          * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
705          * until after ip6_savecontrol() is called and before other functions
706          * which don't want those proto headers.
707          * Because ip6_savecontrol() is going to parse the mbuf to
708          * search for data to be passed up to user-land, it wants mbuf
709          * parameters to be unchanged.
710          * XXX: the call of ip6_savecontrol() has been obsoleted based on
711          * latest version of the advanced API (20020110).
712          */
713         drop_hdrlen = off0 + off;
714
715         /*
716          * Locate pcb for segment.
717          */
718 findpcb:
719         /* IPFIREWALL_FORWARD section */
720         if (next_hop != NULL && !isipv6) {  /* IPv6 support is not there yet */
721                 /*
722                  * Transparently forwarded. Pretend to be the destination.
723                  * already got one like this?
724                  */
725                 cpu = mycpu->gd_cpuid;
726                 inp = in_pcblookup_hash(&tcbinfo[cpu],
727                                         ip->ip_src, th->th_sport,
728                                         ip->ip_dst, th->th_dport,
729                                         0, m->m_pkthdr.rcvif);
730                 if (!inp) {
731                         /*
732                          * It's new.  Try to find the ambushing socket.
733                          */
734
735                         /*
736                          * The rest of the ipfw code stores the port in
737                          * host order.  XXX
738                          * (The IP address is still in network order.)
739                          */
740                         in_port_t dport = next_hop->sin_port ?
741                                                 htons(next_hop->sin_port) :
742                                                 th->th_dport;
743
744                         cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
745                                           next_hop->sin_addr.s_addr, dport);
746                         inp = in_pcblookup_hash(&tcbinfo[cpu],
747                                                 ip->ip_src, th->th_sport,
748                                                 next_hop->sin_addr, dport,
749                                                 1, m->m_pkthdr.rcvif);
750                 }
751         } else {
752                 if (isipv6) {
753                         inp = in6_pcblookup_hash(&tcbinfo[0],
754                                                  &ip6->ip6_src, th->th_sport,
755                                                  &ip6->ip6_dst, th->th_dport,
756                                                  1, m->m_pkthdr.rcvif);
757                 } else {
758                         cpu = mycpu->gd_cpuid;
759                         inp = in_pcblookup_hash(&tcbinfo[cpu],
760                                                 ip->ip_src, th->th_sport,
761                                                 ip->ip_dst, th->th_dport,
762                                                 1, m->m_pkthdr.rcvif);
763                 }
764         }
765
766         /*
767          * If the state is CLOSED (i.e., TCB does not exist) then
768          * all data in the incoming segment is discarded.
769          * If the TCB exists but is in CLOSED state, it is embryonic,
770          * but should either do a listen or a connect soon.
771          */
772         if (inp == NULL) {
773                 if (log_in_vain) {
774 #ifdef INET6
775                         char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
776 #else
777                         char dbuf[sizeof "aaa.bbb.ccc.ddd"];
778                         char sbuf[sizeof "aaa.bbb.ccc.ddd"];
779 #endif
780                         if (isipv6) {
781                                 strcpy(dbuf, "[");
782                                 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
783                                 strcat(dbuf, "]");
784                                 strcpy(sbuf, "[");
785                                 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
786                                 strcat(sbuf, "]");
787                         } else {
788                                 strcpy(dbuf, inet_ntoa(ip->ip_dst));
789                                 strcpy(sbuf, inet_ntoa(ip->ip_src));
790                         }
791                         switch (log_in_vain) {
792                         case 1:
793                                 if (!(thflags & TH_SYN))
794                                         break;
795                         case 2:
796                                 log(LOG_INFO,
797                                     "Connection attempt to TCP %s:%d "
798                                     "from %s:%d flags:0x%02x\n",
799                                     dbuf, ntohs(th->th_dport), sbuf,
800                                     ntohs(th->th_sport), thflags);
801                                 break;
802                         default:
803                                 break;
804                         }
805                 }
806                 if (blackhole) {
807                         switch (blackhole) {
808                         case 1:
809                                 if (thflags & TH_SYN)
810                                         goto drop;
811                                 break;
812                         case 2:
813                                 goto drop;
814                         default:
815                                 goto drop;
816                         }
817                 }
818                 rstreason = BANDLIM_RST_CLOSEDPORT;
819                 goto dropwithreset;
820         }
821
822 #ifdef IPSEC
823         if (isipv6) {
824                 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
825                         ipsec6stat.in_polvio++;
826                         goto drop;
827                 }
828         } else {
829                 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
830                         ipsecstat.in_polvio++;
831                         goto drop;
832                 }
833         }
834 #endif
835 #ifdef FAST_IPSEC
836         if (isipv6) {
837                 if (ipsec6_in_reject(m, inp))
838                         goto drop;
839         } else {
840                 if (ipsec4_in_reject(m, inp))
841                         goto drop;
842         }
843 #endif
844         /* Check the minimum TTL for socket. */
845 #ifdef INET6
846         if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
847                 goto drop;
848 #endif
849
850         tp = intotcpcb(inp);
851         if (tp == NULL) {
852                 rstreason = BANDLIM_RST_CLOSEDPORT;
853                 goto dropwithreset;
854         }
855         if (tp->t_state <= TCPS_CLOSED)
856                 goto drop;
857
858         so = inp->inp_socket;
859
860 #ifdef TCPDEBUG
861         if (so->so_options & SO_DEBUG) {
862                 ostate = tp->t_state;
863                 if (isipv6)
864                         bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
865                 else
866                         bcopy(ip, tcp_saveipgen, sizeof(*ip));
867                 tcp_savetcp = *th;
868         }
869 #endif
870
871         bzero(&to, sizeof to);
872
873         if (so->so_options & SO_ACCEPTCONN) {
874                 struct in_conninfo inc;
875
876 #ifdef INET6
877                 inc.inc_isipv6 = (isipv6 == TRUE);
878 #endif
879                 if (isipv6) {
880                         inc.inc6_faddr = ip6->ip6_src;
881                         inc.inc6_laddr = ip6->ip6_dst;
882                         inc.inc6_route.ro_rt = NULL;            /* XXX */
883                 } else {
884                         inc.inc_faddr = ip->ip_src;
885                         inc.inc_laddr = ip->ip_dst;
886                         inc.inc_route.ro_rt = NULL;             /* XXX */
887                 }
888                 inc.inc_fport = th->th_sport;
889                 inc.inc_lport = th->th_dport;
890
891                 /*
892                  * If the state is LISTEN then ignore segment if it contains
893                  * a RST.  If the segment contains an ACK then it is bad and
894                  * send a RST.  If it does not contain a SYN then it is not
895                  * interesting; drop it.
896                  *
897                  * If the state is SYN_RECEIVED (syncache) and seg contains
898                  * an ACK, but not for our SYN/ACK, send a RST.  If the seg
899                  * contains a RST, check the sequence number to see if it
900                  * is a valid reset segment.
901                  */
902                 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
903                         if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
904                                 if (!syncache_expand(&inc, th, &so, m)) {
905                                         /*
906                                          * No syncache entry, or ACK was not
907                                          * for our SYN/ACK.  Send a RST.
908                                          */
909                                         tcpstat.tcps_badsyn++;
910                                         rstreason = BANDLIM_RST_OPENPORT;
911                                         goto dropwithreset;
912                                 }
913
914                                 /*
915                                  * Could not complete 3-way handshake,
916                                  * connection is being closed down, and
917                                  * syncache will free mbuf.
918                                  */
919                                 if (so == NULL)
920                                         return(IPPROTO_DONE);
921
922                                 /*
923                                  * We must be in the correct protocol thread
924                                  * for this connection.
925                                  */
926                                 KKASSERT(so->so_port == &curthread->td_msgport);
927
928                                 /*
929                                  * Socket is created in state SYN_RECEIVED.
930                                  * Continue processing segment.
931                                  */
932                                 inp = so->so_pcb;
933                                 tp = intotcpcb(inp);
934                                 /*
935                                  * This is what would have happened in
936                                  * tcp_output() when the SYN,ACK was sent.
937                                  */
938                                 tp->snd_up = tp->snd_una;
939                                 tp->snd_max = tp->snd_nxt = tp->iss + 1;
940                                 tp->last_ack_sent = tp->rcv_nxt;
941
942                                 goto after_listen;
943                         }
944                         if (thflags & TH_RST) {
945                                 syncache_chkrst(&inc, th);
946                                 goto drop;
947                         }
948                         if (thflags & TH_ACK) {
949                                 syncache_badack(&inc);
950                                 tcpstat.tcps_badsyn++;
951                                 rstreason = BANDLIM_RST_OPENPORT;
952                                 goto dropwithreset;
953                         }
954                         goto drop;
955                 }
956
957                 /*
958                  * Segment's flags are (SYN) or (SYN | FIN).
959                  */
960 #ifdef INET6
961                 /*
962                  * If deprecated address is forbidden,
963                  * we do not accept SYN to deprecated interface
964                  * address to prevent any new inbound connection from
965                  * getting established.
966                  * When we do not accept SYN, we send a TCP RST,
967                  * with deprecated source address (instead of dropping
968                  * it).  We compromise it as it is much better for peer
969                  * to send a RST, and RST will be the final packet
970                  * for the exchange.
971                  *
972                  * If we do not forbid deprecated addresses, we accept
973                  * the SYN packet.  RFC2462 does not suggest dropping
974                  * SYN in this case.
975                  * If we decipher RFC2462 5.5.4, it says like this:
976                  * 1. use of deprecated addr with existing
977                  *    communication is okay - "SHOULD continue to be
978                  *    used"
979                  * 2. use of it with new communication:
980                  *   (2a) "SHOULD NOT be used if alternate address
981                  *        with sufficient scope is available"
982                  *   (2b) nothing mentioned otherwise.
983                  * Here we fall into (2b) case as we have no choice in
984                  * our source address selection - we must obey the peer.
985                  *
986                  * The wording in RFC2462 is confusing, and there are
987                  * multiple description text for deprecated address
988                  * handling - worse, they are not exactly the same.
989                  * I believe 5.5.4 is the best one, so we follow 5.5.4.
990                  */
991                 if (isipv6 && !ip6_use_deprecated) {
992                         struct in6_ifaddr *ia6;
993
994                         if ((ia6 = ip6_getdstifaddr(m)) &&
995                             (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
996                                 tp = NULL;
997                                 rstreason = BANDLIM_RST_OPENPORT;
998                                 goto dropwithreset;
999                         }
1000                 }
1001 #endif
1002                 /*
1003                  * If it is from this socket, drop it, it must be forged.
1004                  * Don't bother responding if the destination was a broadcast.
1005                  */
1006                 if (th->th_dport == th->th_sport) {
1007                         if (isipv6) {
1008                                 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1009                                                        &ip6->ip6_src))
1010                                         goto drop;
1011                         } else {
1012                                 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1013                                         goto drop;
1014                         }
1015                 }
1016                 /*
1017                  * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1018                  *
1019                  * Note that it is quite possible to receive unicast
1020                  * link-layer packets with a broadcast IP address. Use
1021                  * in_broadcast() to find them.
1022                  */
1023                 if (m->m_flags & (M_BCAST | M_MCAST))
1024                         goto drop;
1025                 if (isipv6) {
1026                         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1027                             IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1028                                 goto drop;
1029                 } else {
1030                         if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1031                             IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1032                             ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1033                             in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1034                                 goto drop;
1035                 }
1036                 /*
1037                  * SYN appears to be valid; create compressed TCP state
1038                  * for syncache, or perform t/tcp connection.
1039                  */
1040                 if (so->so_qlen <= so->so_qlimit) {
1041                         tcp_dooptions(&to, optp, optlen, TRUE, th->th_ack);
1042                         if (!syncache_add(&inc, &to, th, so, m))
1043                                 goto drop;
1044
1045                         /*
1046                          * Entry added to syncache, mbuf used to
1047                          * send SYN,ACK packet.
1048                          */
1049                         return(IPPROTO_DONE);
1050                 }
1051                 goto drop;
1052         }
1053
1054 after_listen:
1055         /*
1056          * Should not happen - syncache should pick up these connections.
1057          *
1058          * Once we are past handling listen sockets we must be in the
1059          * correct protocol processing thread.
1060          */
1061         KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1062         KKASSERT(so->so_port == &curthread->td_msgport);
1063
1064         /* Unscale the window into a 32-bit value. */
1065         if (!(thflags & TH_SYN))
1066                 tiwin = th->th_win << tp->snd_scale;
1067         else
1068                 tiwin = th->th_win;
1069
1070         /*
1071          * This is the second part of the MSS DoS prevention code (after
1072          * minmss on the sending side) and it deals with too many too small
1073          * tcp packets in a too short timeframe (1 second).
1074          *
1075          * XXX Removed.  This code was crap.  It does not scale to network
1076          *     speed, and default values break NFS.  Gone.
1077          */
1078         /* REMOVED */
1079
1080         /*
1081          * Segment received on connection.
1082          *
1083          * Reset idle time and keep-alive timer.  Don't waste time if less
1084          * then a second has elapsed.
1085          */
1086         if ((int)(ticks - tp->t_rcvtime) > hz)
1087                 tcp_timer_keep_activity(tp, thflags);
1088
1089         /*
1090          * Process options.
1091          * XXX this is tradtitional behavior, may need to be cleaned up.
1092          */
1093         tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0, th->th_ack);
1094         if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1095                 if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) {
1096                         tp->t_flags |= TF_RCVD_SCALE;
1097                         tp->snd_scale = to.to_requested_s_scale;
1098                 }
1099
1100                 /*
1101                  * Initial send window; will be updated upon next ACK
1102                  */
1103                 tp->snd_wnd = th->th_win;
1104
1105                 if (to.to_flags & TOF_TS) {
1106                         tp->t_flags |= TF_RCVD_TSTMP;
1107                         tp->ts_recent = to.to_tsval;
1108                         tp->ts_recent_age = ticks;
1109                 }
1110                 if (!(to.to_flags & TOF_MSS))
1111                         to.to_mss = 0;
1112                 tcp_mss(tp, to.to_mss);
1113                 /*
1114                  * Only set the TF_SACK_PERMITTED per-connection flag
1115                  * if we got a SACK_PERMITTED option from the other side
1116                  * and the global tcp_do_sack variable is true.
1117                  */
1118                 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1119                         tp->t_flags |= TF_SACK_PERMITTED;
1120         }
1121
1122         /*
1123          * Header prediction: check for the two common cases
1124          * of a uni-directional data xfer.  If the packet has
1125          * no control flags, is in-sequence, the window didn't
1126          * change and we're not retransmitting, it's a
1127          * candidate.  If the length is zero and the ack moved
1128          * forward, we're the sender side of the xfer.  Just
1129          * free the data acked & wake any higher level process
1130          * that was blocked waiting for space.  If the length
1131          * is non-zero and the ack didn't move, we're the
1132          * receiver side.  If we're getting packets in-order
1133          * (the reassembly queue is empty), add the data to
1134          * the socket buffer and note that we need a delayed ack.
1135          * Make sure that the hidden state-flags are also off.
1136          * Since we check for TCPS_ESTABLISHED above, it can only
1137          * be TH_NEEDSYN.
1138          */
1139         if (tp->t_state == TCPS_ESTABLISHED &&
1140             (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1141             !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1142             (!(to.to_flags & TOF_TS) ||
1143              TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1144             th->th_seq == tp->rcv_nxt &&
1145             tp->snd_nxt == tp->snd_max) {
1146
1147                 /*
1148                  * If last ACK falls within this segment's sequence numbers,
1149                  * record the timestamp.
1150                  * NOTE that the test is modified according to the latest
1151                  * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1152                  */
1153                 if ((to.to_flags & TOF_TS) &&
1154                     SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1155                         tp->ts_recent_age = ticks;
1156                         tp->ts_recent = to.to_tsval;
1157                 }
1158
1159                 if (tlen == 0) {
1160                         if (SEQ_GT(th->th_ack, tp->snd_una) &&
1161                             SEQ_LEQ(th->th_ack, tp->snd_max) &&
1162                             tp->snd_cwnd >= tp->snd_wnd &&
1163                             !IN_FASTRECOVERY(tp)) {
1164                                 /*
1165                                  * This is a pure ack for outstanding data.
1166                                  */
1167                                 ++tcpstat.tcps_predack;
1168                                 /*
1169                                  * "bad retransmit" recovery
1170                                  *
1171                                  * If Eifel detection applies, then
1172                                  * it is deterministic, so use it
1173                                  * unconditionally over the old heuristic.
1174                                  * Otherwise, fall back to the old heuristic.
1175                                  */
1176                                 if (tcp_do_eifel_detect &&
1177                                     (to.to_flags & TOF_TS) && to.to_tsecr &&
1178                                     (tp->t_flags & TF_FIRSTACCACK)) {
1179                                         /* Eifel detection applicable. */
1180                                         if (to.to_tsecr < tp->t_rexmtTS) {
1181                                                 tcp_revert_congestion_state(tp);
1182                                                 ++tcpstat.tcps_eifeldetected;
1183                                                 if (tp->t_rxtshift != 1 ||
1184                                                     ticks >= tp->t_badrxtwin)
1185                                                         ++tcpstat.tcps_rttcantdetect;
1186                                         }
1187                                 } else if (tp->t_rxtshift == 1 &&
1188                                            ticks < tp->t_badrxtwin) {
1189                                         tcp_revert_congestion_state(tp);
1190                                         ++tcpstat.tcps_rttdetected;
1191                                 }
1192                                 tp->t_flags &= ~(TF_FIRSTACCACK |
1193                                                  TF_FASTREXMT | TF_EARLYREXMT);
1194                                 /*
1195                                  * Recalculate the retransmit timer / rtt.
1196                                  *
1197                                  * Some machines (certain windows boxes)
1198                                  * send broken timestamp replies during the
1199                                  * SYN+ACK phase, ignore timestamps of 0.
1200                                  */
1201                                 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1202                                         tcp_xmit_timer(tp,
1203                                             ticks - to.to_tsecr + 1,
1204                                             th->th_ack);
1205                                 } else if (tp->t_rtttime &&
1206                                            SEQ_GT(th->th_ack, tp->t_rtseq)) {
1207                                         tcp_xmit_timer(tp,
1208                                             ticks - tp->t_rtttime,
1209                                             th->th_ack);
1210                                 }
1211                                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1212                                 acked = th->th_ack - tp->snd_una;
1213                                 tcpstat.tcps_rcvackpack++;
1214                                 tcpstat.tcps_rcvackbyte += acked;
1215                                 sbdrop(&so->so_snd.sb, acked);
1216                                 tp->snd_recover = th->th_ack - 1;
1217                                 tp->snd_una = th->th_ack;
1218                                 tp->t_dupacks = 0;
1219                                 /*
1220                                  * Update window information.
1221                                  */
1222                                 if (tiwin != tp->snd_wnd &&
1223                                     acceptable_window_update(tp, th, tiwin)) {
1224                                         /* keep track of pure window updates */
1225                                         if (tp->snd_wl2 == th->th_ack &&
1226                                             tiwin > tp->snd_wnd)
1227                                                 tcpstat.tcps_rcvwinupd++;
1228                                         tp->snd_wnd = tiwin;
1229                                         tp->snd_wl1 = th->th_seq;
1230                                         tp->snd_wl2 = th->th_ack;
1231                                         if (tp->snd_wnd > tp->max_sndwnd)
1232                                                 tp->max_sndwnd = tp->snd_wnd;
1233                                 }
1234                                 m_freem(m);
1235                                 ND6_HINT(tp); /* some progress has been done */
1236                                 /*
1237                                  * If all outstanding data are acked, stop
1238                                  * retransmit timer, otherwise restart timer
1239                                  * using current (possibly backed-off) value.
1240                                  * If process is waiting for space,
1241                                  * wakeup/selwakeup/signal.  If data
1242                                  * are ready to send, let tcp_output
1243                                  * decide between more output or persist.
1244                                  */
1245                                 if (tp->snd_una == tp->snd_max) {
1246                                         tcp_callout_stop(tp, tp->tt_rexmt);
1247                                 } else if (!tcp_callout_active(tp,
1248                                             tp->tt_persist)) {
1249                                         tcp_callout_reset(tp, tp->tt_rexmt,
1250                                             tp->t_rxtcur, tcp_timer_rexmt);
1251                                 }
1252                                 sowwakeup(so);
1253                                 if (so->so_snd.ssb_cc > 0)
1254                                         tcp_output(tp);
1255                                 return(IPPROTO_DONE);
1256                         }
1257                 } else if (tiwin == tp->snd_wnd &&
1258                     th->th_ack == tp->snd_una &&
1259                     LIST_EMPTY(&tp->t_segq) &&
1260                     tlen <= ssb_space(&so->so_rcv)) {
1261                         u_long newsize = 0;     /* automatic sockbuf scaling */
1262                         /*
1263                          * This is a pure, in-sequence data packet
1264                          * with nothing on the reassembly queue and
1265                          * we have enough buffer space to take it.
1266                          */
1267                         ++tcpstat.tcps_preddat;
1268                         tp->rcv_nxt += tlen;
1269                         tcpstat.tcps_rcvpack++;
1270                         tcpstat.tcps_rcvbyte += tlen;
1271                         ND6_HINT(tp);   /* some progress has been done */
1272                 /*
1273                  * Automatic sizing of receive socket buffer.  Often the send
1274                  * buffer size is not optimally adjusted to the actual network
1275                  * conditions at hand (delay bandwidth product).  Setting the
1276                  * buffer size too small limits throughput on links with high
1277                  * bandwidth and high delay (eg. trans-continental/oceanic links).
1278                  *
1279                  * On the receive side the socket buffer memory is only rarely
1280                  * used to any significant extent.  This allows us to be much
1281                  * more aggressive in scaling the receive socket buffer.  For
1282                  * the case that the buffer space is actually used to a large
1283                  * extent and we run out of kernel memory we can simply drop
1284                  * the new segments; TCP on the sender will just retransmit it
1285                  * later.  Setting the buffer size too big may only consume too
1286                  * much kernel memory if the application doesn't read() from
1287                  * the socket or packet loss or reordering makes use of the
1288                  * reassembly queue.
1289                  *
1290                  * The criteria to step up the receive buffer one notch are:
1291                  *  1. the number of bytes received during the time it takes
1292                  *     one timestamp to be reflected back to us (the RTT);
1293                  *  2. received bytes per RTT is within seven eighth of the
1294                  *     current socket buffer size;
1295                  *  3. receive buffer size has not hit maximal automatic size;
1296                  *
1297                  * This algorithm does one step per RTT at most and only if
1298                  * we receive a bulk stream w/o packet losses or reorderings.
1299                  * Shrinking the buffer during idle times is not necessary as
1300                  * it doesn't consume any memory when idle.
1301                  *
1302                  * TODO: Only step up if the application is actually serving
1303                  * the buffer to better manage the socket buffer resources.
1304                  */
1305                         if (tcp_do_autorcvbuf &&
1306                             to.to_tsecr &&
1307                             (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1308                                 if (to.to_tsecr > tp->rfbuf_ts &&
1309                                     to.to_tsecr - tp->rfbuf_ts < hz) {
1310                                         if (tp->rfbuf_cnt >
1311                                             (so->so_rcv.ssb_hiwat / 8 * 7) &&
1312                                             so->so_rcv.ssb_hiwat <
1313                                             tcp_autorcvbuf_max) {
1314                                                 newsize =
1315                                                     ulmin(so->so_rcv.ssb_hiwat +
1316                                                           tcp_autorcvbuf_inc,
1317                                                           tcp_autorcvbuf_max);
1318                                         }
1319                                         /* Start over with next RTT. */
1320                                         tp->rfbuf_ts = 0;
1321                                         tp->rfbuf_cnt = 0;
1322                                 } else
1323                                         tp->rfbuf_cnt += tlen;  /* add up */
1324                         }
1325                         /*
1326                          * Add data to socket buffer.
1327                          */
1328                         if (so->so_state & SS_CANTRCVMORE) {
1329                                 m_freem(m);
1330                         } else {
1331                                 /*
1332                                  * Set new socket buffer size, give up when
1333                                  * limit is reached.
1334                                  *
1335                                  * Adjusting the size can mess up ACK
1336                                  * sequencing when pure window updates are
1337                                  * being avoided (which is the default),
1338                                  * so force an ack.
1339                                  */
1340                                 lwkt_gettoken(&so->so_rcv.ssb_token);
1341                                 if (newsize) {
1342                                         tp->t_flags |= TF_RXRESIZED;
1343                                         if (!ssb_reserve(&so->so_rcv, newsize,
1344                                                          so, NULL)) {
1345                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1346                                         }
1347                                         if (newsize >=
1348                                             (TCP_MAXWIN << tp->rcv_scale)) {
1349                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1350                                         }
1351                                 }
1352                                 m_adj(m, drop_hdrlen); /* delayed header drop */
1353                                 ssb_appendstream(&so->so_rcv, m);
1354                                 lwkt_reltoken(&so->so_rcv.ssb_token);
1355                         }
1356                         sorwakeup(so);
1357                         /*
1358                          * This code is responsible for most of the ACKs
1359                          * the TCP stack sends back after receiving a data
1360                          * packet.  Note that the DELAY_ACK check fails if
1361                          * the delack timer is already running, which results
1362                          * in an ack being sent every other packet (which is
1363                          * what we want).
1364                          *
1365                          * We then further aggregate acks by not actually
1366                          * sending one until the protocol thread has completed
1367                          * processing the current backlog of packets.  This
1368                          * does not delay the ack any further, but allows us
1369                          * to take advantage of the packet aggregation that
1370                          * high speed NICs do (usually blocks of 8-10 packets)
1371                          * to send a single ack rather then four or five acks,
1372                          * greatly reducing the ack rate, the return channel
1373                          * bandwidth, and the protocol overhead on both ends.
1374                          *
1375                          * Since this also has the effect of slowing down
1376                          * the exponential slow-start ramp-up, systems with 
1377                          * very large bandwidth-delay products might want
1378                          * to turn the feature off.
1379                          */
1380                         if (DELAY_ACK(tp)) {
1381                                 tcp_callout_reset(tp, tp->tt_delack,
1382                                     tcp_delacktime, tcp_timer_delack);
1383                         } else if (tcp_aggregate_acks) {
1384                                 tp->t_flags |= TF_ACKNOW;
1385                                 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1386                                         tp->t_flags |= TF_ONOUTPUTQ;
1387                                         tp->tt_cpu = mycpu->gd_cpuid;
1388                                         TAILQ_INSERT_TAIL(
1389                                             &tcpcbackq[tp->tt_cpu],
1390                                             tp, t_outputq);
1391                                 }
1392                         } else {
1393                                 tp->t_flags |= TF_ACKNOW;
1394                                 tcp_output(tp);
1395                         }
1396                         return(IPPROTO_DONE);
1397                 }
1398         }
1399
1400         /*
1401          * Calculate amount of space in receive window,
1402          * and then do TCP input processing.
1403          * Receive window is amount of space in rcv queue,
1404          * but not less than advertised window.
1405          */
1406         recvwin = ssb_space(&so->so_rcv);
1407         if (recvwin < 0)
1408                 recvwin = 0;
1409         tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1410
1411         /* Reset receive buffer auto scaling when not in bulk receive mode. */
1412         tp->rfbuf_ts = 0;
1413         tp->rfbuf_cnt = 0;
1414
1415         switch (tp->t_state) {
1416         /*
1417          * If the state is SYN_RECEIVED:
1418          *      if seg contains an ACK, but not for our SYN/ACK, send a RST.
1419          */
1420         case TCPS_SYN_RECEIVED:
1421                 if ((thflags & TH_ACK) &&
1422                     (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1423                      SEQ_GT(th->th_ack, tp->snd_max))) {
1424                                 rstreason = BANDLIM_RST_OPENPORT;
1425                                 goto dropwithreset;
1426                 }
1427                 break;
1428
1429         /*
1430          * If the state is SYN_SENT:
1431          *      if seg contains an ACK, but not for our SYN, drop the input.
1432          *      if seg contains a RST, then drop the connection.
1433          *      if seg does not contain SYN, then drop it.
1434          * Otherwise this is an acceptable SYN segment
1435          *      initialize tp->rcv_nxt and tp->irs
1436          *      if seg contains ack then advance tp->snd_una
1437          *      if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1438          *      arrange for segment to be acked (eventually)
1439          *      continue processing rest of data/controls, beginning with URG
1440          */
1441         case TCPS_SYN_SENT:
1442                 if ((thflags & TH_ACK) &&
1443                     (SEQ_LEQ(th->th_ack, tp->iss) ||
1444                      SEQ_GT(th->th_ack, tp->snd_max))) {
1445                         rstreason = BANDLIM_UNLIMITED;
1446                         goto dropwithreset;
1447                 }
1448                 if (thflags & TH_RST) {
1449                         if (thflags & TH_ACK)
1450                                 tp = tcp_drop(tp, ECONNREFUSED);
1451                         goto drop;
1452                 }
1453                 if (!(thflags & TH_SYN))
1454                         goto drop;
1455
1456                 tp->irs = th->th_seq;
1457                 tcp_rcvseqinit(tp);
1458                 if (thflags & TH_ACK) {
1459                         /* Our SYN was acked. */
1460                         tcpstat.tcps_connects++;
1461                         soisconnected(so);
1462                         /* Do window scaling on this connection? */
1463                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1464                             (TF_RCVD_SCALE | TF_REQ_SCALE))
1465                                 tp->rcv_scale = tp->request_r_scale;
1466                         tp->rcv_adv += tp->rcv_wnd;
1467                         tp->snd_una++;          /* SYN is acked */
1468                         tcp_callout_stop(tp, tp->tt_rexmt);
1469                         /*
1470                          * If there's data, delay ACK; if there's also a FIN
1471                          * ACKNOW will be turned on later.
1472                          */
1473                         if (DELAY_ACK(tp) && tlen != 0) {
1474                                 tcp_callout_reset(tp, tp->tt_delack,
1475                                     tcp_delacktime, tcp_timer_delack);
1476                         } else {
1477                                 tp->t_flags |= TF_ACKNOW;
1478                         }
1479                         /*
1480                          * Received <SYN,ACK> in SYN_SENT[*] state.
1481                          * Transitions:
1482                          *      SYN_SENT  --> ESTABLISHED
1483                          *      SYN_SENT* --> FIN_WAIT_1
1484                          */
1485                         tp->t_starttime = ticks;
1486                         if (tp->t_flags & TF_NEEDFIN) {
1487                                 tp->t_state = TCPS_FIN_WAIT_1;
1488                                 tp->t_flags &= ~TF_NEEDFIN;
1489                                 thflags &= ~TH_SYN;
1490                         } else {
1491                                 tcp_established(tp);
1492                         }
1493                 } else {
1494                         /*
1495                          * Received initial SYN in SYN-SENT[*] state =>
1496                          * simultaneous open.
1497                          * Do 3-way handshake:
1498                          *        SYN-SENT -> SYN-RECEIVED
1499                          *        SYN-SENT* -> SYN-RECEIVED*
1500                          */
1501                         tp->t_flags |= TF_ACKNOW;
1502                         tcp_callout_stop(tp, tp->tt_rexmt);
1503                         tp->t_state = TCPS_SYN_RECEIVED;
1504                 }
1505
1506                 /*
1507                  * Advance th->th_seq to correspond to first data byte.
1508                  * If data, trim to stay within window,
1509                  * dropping FIN if necessary.
1510                  */
1511                 th->th_seq++;
1512                 if (tlen > tp->rcv_wnd) {
1513                         todrop = tlen - tp->rcv_wnd;
1514                         m_adj(m, -todrop);
1515                         tlen = tp->rcv_wnd;
1516                         thflags &= ~TH_FIN;
1517                         tcpstat.tcps_rcvpackafterwin++;
1518                         tcpstat.tcps_rcvbyteafterwin += todrop;
1519                 }
1520                 tp->snd_wl1 = th->th_seq - 1;
1521                 tp->rcv_up = th->th_seq;
1522                 /*
1523                  * Client side of transaction: already sent SYN and data.
1524                  * If the remote host used T/TCP to validate the SYN,
1525                  * our data will be ACK'd; if so, enter normal data segment
1526                  * processing in the middle of step 5, ack processing.
1527                  * Otherwise, goto step 6.
1528                  */
1529                 if (thflags & TH_ACK)
1530                         goto process_ACK;
1531
1532                 goto step6;
1533
1534         /*
1535          * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1536          *      do normal processing (we no longer bother with T/TCP).
1537          */
1538         case TCPS_LAST_ACK:
1539         case TCPS_CLOSING:
1540         case TCPS_TIME_WAIT:
1541                 break;  /* continue normal processing */
1542         }
1543
1544         /*
1545          * States other than LISTEN or SYN_SENT.
1546          * First check the RST flag and sequence number since reset segments
1547          * are exempt from the timestamp and connection count tests.  This
1548          * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1549          * below which allowed reset segments in half the sequence space
1550          * to fall though and be processed (which gives forged reset
1551          * segments with a random sequence number a 50 percent chance of
1552          * killing a connection).
1553          * Then check timestamp, if present.
1554          * Then check the connection count, if present.
1555          * Then check that at least some bytes of segment are within
1556          * receive window.  If segment begins before rcv_nxt,
1557          * drop leading data (and SYN); if nothing left, just ack.
1558          *
1559          *
1560          * If the RST bit is set, check the sequence number to see
1561          * if this is a valid reset segment.
1562          * RFC 793 page 37:
1563          *   In all states except SYN-SENT, all reset (RST) segments
1564          *   are validated by checking their SEQ-fields.  A reset is
1565          *   valid if its sequence number is in the window.
1566          * Note: this does not take into account delayed ACKs, so
1567          *   we should test against last_ack_sent instead of rcv_nxt.
1568          *   The sequence number in the reset segment is normally an
1569          *   echo of our outgoing acknowledgement numbers, but some hosts
1570          *   send a reset with the sequence number at the rightmost edge
1571          *   of our receive window, and we have to handle this case.
1572          * If we have multiple segments in flight, the intial reset
1573          * segment sequence numbers will be to the left of last_ack_sent,
1574          * but they will eventually catch up.
1575          * In any case, it never made sense to trim reset segments to
1576          * fit the receive window since RFC 1122 says:
1577          *   4.2.2.12  RST Segment: RFC-793 Section 3.4
1578          *
1579          *    A TCP SHOULD allow a received RST segment to include data.
1580          *
1581          *    DISCUSSION
1582          *         It has been suggested that a RST segment could contain
1583          *         ASCII text that encoded and explained the cause of the
1584          *         RST.  No standard has yet been established for such
1585          *         data.
1586          *
1587          * If the reset segment passes the sequence number test examine
1588          * the state:
1589          *    SYN_RECEIVED STATE:
1590          *      If passive open, return to LISTEN state.
1591          *      If active open, inform user that connection was refused.
1592          *    ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1593          *      Inform user that connection was reset, and close tcb.
1594          *    CLOSING, LAST_ACK STATES:
1595          *      Close the tcb.
1596          *    TIME_WAIT STATE:
1597          *      Drop the segment - see Stevens, vol. 2, p. 964 and
1598          *      RFC 1337.
1599          */
1600         if (thflags & TH_RST) {
1601                 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1602                     SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1603                         switch (tp->t_state) {
1604
1605                         case TCPS_SYN_RECEIVED:
1606                                 so->so_error = ECONNREFUSED;
1607                                 goto close;
1608
1609                         case TCPS_ESTABLISHED:
1610                         case TCPS_FIN_WAIT_1:
1611                         case TCPS_FIN_WAIT_2:
1612                         case TCPS_CLOSE_WAIT:
1613                                 so->so_error = ECONNRESET;
1614                         close:
1615                                 tp->t_state = TCPS_CLOSED;
1616                                 tcpstat.tcps_drops++;
1617                                 tp = tcp_close(tp);
1618                                 break;
1619
1620                         case TCPS_CLOSING:
1621                         case TCPS_LAST_ACK:
1622                                 tp = tcp_close(tp);
1623                                 break;
1624
1625                         case TCPS_TIME_WAIT:
1626                                 break;
1627                         }
1628                 }
1629                 goto drop;
1630         }
1631
1632         /*
1633          * RFC 1323 PAWS: If we have a timestamp reply on this segment
1634          * and it's less than ts_recent, drop it.
1635          */
1636         if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1637             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1638
1639                 /* Check to see if ts_recent is over 24 days old.  */
1640                 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1641                         /*
1642                          * Invalidate ts_recent.  If this segment updates
1643                          * ts_recent, the age will be reset later and ts_recent
1644                          * will get a valid value.  If it does not, setting
1645                          * ts_recent to zero will at least satisfy the
1646                          * requirement that zero be placed in the timestamp
1647                          * echo reply when ts_recent isn't valid.  The
1648                          * age isn't reset until we get a valid ts_recent
1649                          * because we don't want out-of-order segments to be
1650                          * dropped when ts_recent is old.
1651                          */
1652                         tp->ts_recent = 0;
1653                 } else {
1654                         tcpstat.tcps_rcvduppack++;
1655                         tcpstat.tcps_rcvdupbyte += tlen;
1656                         tcpstat.tcps_pawsdrop++;
1657                         if (tlen)
1658                                 goto dropafterack;
1659                         goto drop;
1660                 }
1661         }
1662
1663         /*
1664          * In the SYN-RECEIVED state, validate that the packet belongs to
1665          * this connection before trimming the data to fit the receive
1666          * window.  Check the sequence number versus IRS since we know
1667          * the sequence numbers haven't wrapped.  This is a partial fix
1668          * for the "LAND" DoS attack.
1669          */
1670         if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1671                 rstreason = BANDLIM_RST_OPENPORT;
1672                 goto dropwithreset;
1673         }
1674
1675         todrop = tp->rcv_nxt - th->th_seq;
1676         if (todrop > 0) {
1677                 if (TCP_DO_SACK(tp)) {
1678                         /* Report duplicate segment at head of packet. */
1679                         tp->reportblk.rblk_start = th->th_seq;
1680                         tp->reportblk.rblk_end = TCP_SACK_BLKEND(
1681                             th->th_seq + tlen, thflags);
1682                         if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1683                                 tp->reportblk.rblk_end = tp->rcv_nxt;
1684                         tp->t_flags |= (TF_DUPSEG | TF_SACKLEFT | TF_ACKNOW);
1685                 }
1686                 if (thflags & TH_SYN) {
1687                         thflags &= ~TH_SYN;
1688                         th->th_seq++;
1689                         if (th->th_urp > 1)
1690                                 th->th_urp--;
1691                         else
1692                                 thflags &= ~TH_URG;
1693                         todrop--;
1694                 }
1695                 /*
1696                  * Following if statement from Stevens, vol. 2, p. 960.
1697                  */
1698                 if (todrop > tlen ||
1699                     (todrop == tlen && !(thflags & TH_FIN))) {
1700                         /*
1701                          * Any valid FIN must be to the left of the window.
1702                          * At this point the FIN must be a duplicate or out
1703                          * of sequence; drop it.
1704                          */
1705                         thflags &= ~TH_FIN;
1706
1707                         /*
1708                          * Send an ACK to resynchronize and drop any data.
1709                          * But keep on processing for RST or ACK.
1710                          */
1711                         tp->t_flags |= TF_ACKNOW;
1712                         todrop = tlen;
1713                         tcpstat.tcps_rcvduppack++;
1714                         tcpstat.tcps_rcvdupbyte += todrop;
1715                 } else {
1716                         tcpstat.tcps_rcvpartduppack++;
1717                         tcpstat.tcps_rcvpartdupbyte += todrop;
1718                 }
1719                 drop_hdrlen += todrop;  /* drop from the top afterwards */
1720                 th->th_seq += todrop;
1721                 tlen -= todrop;
1722                 if (th->th_urp > todrop)
1723                         th->th_urp -= todrop;
1724                 else {
1725                         thflags &= ~TH_URG;
1726                         th->th_urp = 0;
1727                 }
1728         }
1729
1730         /*
1731          * If new data are received on a connection after the
1732          * user processes are gone, then RST the other end.
1733          */
1734         if ((so->so_state & SS_NOFDREF) &&
1735             tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1736                 tp = tcp_close(tp);
1737                 tcpstat.tcps_rcvafterclose++;
1738                 rstreason = BANDLIM_UNLIMITED;
1739                 goto dropwithreset;
1740         }
1741
1742         /*
1743          * If segment ends after window, drop trailing data
1744          * (and PUSH and FIN); if nothing left, just ACK.
1745          */
1746         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1747         if (todrop > 0) {
1748                 tcpstat.tcps_rcvpackafterwin++;
1749                 if (todrop >= tlen) {
1750                         tcpstat.tcps_rcvbyteafterwin += tlen;
1751                         /*
1752                          * If a new connection request is received
1753                          * while in TIME_WAIT, drop the old connection
1754                          * and start over if the sequence numbers
1755                          * are above the previous ones.
1756                          */
1757                         if (thflags & TH_SYN &&
1758                             tp->t_state == TCPS_TIME_WAIT &&
1759                             SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1760                                 tp = tcp_close(tp);
1761                                 goto findpcb;
1762                         }
1763                         /*
1764                          * If window is closed can only take segments at
1765                          * window edge, and have to drop data and PUSH from
1766                          * incoming segments.  Continue processing, but
1767                          * remember to ack.  Otherwise, drop segment
1768                          * and ack.
1769                          */
1770                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1771                                 tp->t_flags |= TF_ACKNOW;
1772                                 tcpstat.tcps_rcvwinprobe++;
1773                         } else
1774                                 goto dropafterack;
1775                 } else
1776                         tcpstat.tcps_rcvbyteafterwin += todrop;
1777                 m_adj(m, -todrop);
1778                 tlen -= todrop;
1779                 thflags &= ~(TH_PUSH | TH_FIN);
1780         }
1781
1782         /*
1783          * If last ACK falls within this segment's sequence numbers,
1784          * record its timestamp.
1785          * NOTE:
1786          * 1) That the test incorporates suggestions from the latest
1787          *    proposal of the tcplw@cray.com list (Braden 1993/04/26).
1788          * 2) That updating only on newer timestamps interferes with
1789          *    our earlier PAWS tests, so this check should be solely
1790          *    predicated on the sequence space of this segment.
1791          * 3) That we modify the segment boundary check to be
1792          *        Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1793          *    instead of RFC1323's
1794          *        Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1795          *    This modified check allows us to overcome RFC1323's
1796          *    limitations as described in Stevens TCP/IP Illustrated
1797          *    Vol. 2 p.869. In such cases, we can still calculate the
1798          *    RTT correctly when RCV.NXT == Last.ACK.Sent.
1799          */
1800         if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1801             SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1802                                         + ((thflags & TH_SYN) != 0)
1803                                         + ((thflags & TH_FIN) != 0)))) {
1804                 tp->ts_recent_age = ticks;
1805                 tp->ts_recent = to.to_tsval;
1806         }
1807
1808         /*
1809          * If a SYN is in the window, then this is an
1810          * error and we send an RST and drop the connection.
1811          */
1812         if (thflags & TH_SYN) {
1813                 tp = tcp_drop(tp, ECONNRESET);
1814                 rstreason = BANDLIM_UNLIMITED;
1815                 goto dropwithreset;
1816         }
1817
1818         /*
1819          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN
1820          * flag is on (half-synchronized state), then queue data for
1821          * later processing; else drop segment and return.
1822          */
1823         if (!(thflags & TH_ACK)) {
1824                 if (tp->t_state == TCPS_SYN_RECEIVED ||
1825                     (tp->t_flags & TF_NEEDSYN))
1826                         goto step6;
1827                 else
1828                         goto drop;
1829         }
1830
1831         /*
1832          * Ack processing.
1833          */
1834         switch (tp->t_state) {
1835         /*
1836          * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1837          * ESTABLISHED state and continue processing.
1838          * The ACK was checked above.
1839          */
1840         case TCPS_SYN_RECEIVED:
1841
1842                 tcpstat.tcps_connects++;
1843                 soisconnected(so);
1844                 /* Do window scaling? */
1845                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1846                     (TF_RCVD_SCALE | TF_REQ_SCALE))
1847                         tp->rcv_scale = tp->request_r_scale;
1848                 /*
1849                  * Make transitions:
1850                  *      SYN-RECEIVED  -> ESTABLISHED
1851                  *      SYN-RECEIVED* -> FIN-WAIT-1
1852                  */
1853                 tp->t_starttime = ticks;
1854                 if (tp->t_flags & TF_NEEDFIN) {
1855                         tp->t_state = TCPS_FIN_WAIT_1;
1856                         tp->t_flags &= ~TF_NEEDFIN;
1857                 } else {
1858                         tcp_established(tp);
1859                 }
1860                 /*
1861                  * If segment contains data or ACK, will call tcp_reass()
1862                  * later; if not, do so now to pass queued data to user.
1863                  */
1864                 if (tlen == 0 && !(thflags & TH_FIN))
1865                         tcp_reass(tp, NULL, NULL, NULL);
1866                 /* fall into ... */
1867
1868         /*
1869          * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
1870          * ACKs.  If the ack is in the range
1871          *      tp->snd_una < th->th_ack <= tp->snd_max
1872          * then advance tp->snd_una to th->th_ack and drop
1873          * data from the retransmission queue.  If this ACK reflects
1874          * more up to date window information we update our window information.
1875          */
1876         case TCPS_ESTABLISHED:
1877         case TCPS_FIN_WAIT_1:
1878         case TCPS_FIN_WAIT_2:
1879         case TCPS_CLOSE_WAIT:
1880         case TCPS_CLOSING:
1881         case TCPS_LAST_ACK:
1882         case TCPS_TIME_WAIT:
1883
1884                 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
1885                         if (TCP_DO_SACK(tp))
1886                                 tcp_sack_update_scoreboard(tp, &to);
1887                         if (tlen != 0 || tiwin != tp->snd_wnd) {
1888                                 tp->t_dupacks = 0;
1889                                 break;
1890                         }
1891                         tcpstat.tcps_rcvdupack++;
1892                         if (!tcp_callout_active(tp, tp->tt_rexmt) ||
1893                             th->th_ack != tp->snd_una) {
1894                                 tp->t_dupacks = 0;
1895                                 break;
1896                         }
1897                         /*
1898                          * We have outstanding data (other than
1899                          * a window probe), this is a completely
1900                          * duplicate ack (ie, window info didn't
1901                          * change), the ack is the biggest we've
1902                          * seen and we've seen exactly our rexmt
1903                          * threshhold of them, so assume a packet
1904                          * has been dropped and retransmit it.
1905                          * Kludge snd_nxt & the congestion
1906                          * window so we send only this one
1907                          * packet.
1908                          */
1909                         if (IN_FASTRECOVERY(tp)) {
1910                                 if (TCP_DO_SACK(tp)) {
1911                                         /* No artifical cwnd inflation. */
1912                                         tcp_sack_rexmt(tp, th);
1913                                 } else {
1914                                         /*
1915                                          * Dup acks mean that packets
1916                                          * have left the network
1917                                          * (they're now cached at the
1918                                          * receiver) so bump cwnd by
1919                                          * the amount in the receiver
1920                                          * to keep a constant cwnd
1921                                          * packets in the network.
1922                                          */
1923                                         tp->snd_cwnd += tp->t_maxseg;
1924                                         tcp_output(tp);
1925                                 }
1926                         } else if (SEQ_LT(th->th_ack, tp->snd_recover)) {
1927                                 tp->t_dupacks = 0;
1928                                 break;
1929                         } else if (tcp_ignore_redun_dsack && TCP_DO_SACK(tp) &&
1930                             (to.to_flags & (TOF_DSACK | TOF_SACK_REDUNDANT)) ==
1931                             (TOF_DSACK | TOF_SACK_REDUNDANT)) {
1932                                 /*
1933                                  * If the ACK carries DSACK and other
1934                                  * SACK blocks carry information that
1935                                  * we have already known, don't count
1936                                  * this ACK as duplicate ACK.  This
1937                                  * prevents spurious early retransmit
1938                                  * and fast retransmit.  This also
1939                                  * meets the requirement of RFC3042
1940                                  * that new segments should not be sent
1941                                  * if the SACK blocks do not contain
1942                                  * new information (XXX we actually
1943                                  * loosen the requirment that only DSACK
1944                                  * is checked here).
1945                                  *
1946                                  * This kind of ACKs are usually sent
1947                                  * after spurious retransmit.
1948                                  */
1949                                 /* Do nothing; don't change t_dupacks */
1950                         } else if (++tp->t_dupacks == tp->t_rxtthresh) {
1951                                 tcp_seq old_snd_nxt;
1952                                 u_int win;
1953
1954 fastretransmit:
1955                                 if (tcp_do_eifel_detect &&
1956                                     (tp->t_flags & TF_RCVD_TSTMP)) {
1957                                         tcp_save_congestion_state(tp);
1958                                         tp->t_flags |= TF_FASTREXMT;
1959                                 }
1960                                 /*
1961                                  * We know we're losing at the current
1962                                  * window size, so do congestion avoidance:
1963                                  * set ssthresh to half the current window
1964                                  * and pull our congestion window back to the
1965                                  * new ssthresh.
1966                                  */
1967                                 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 /
1968                                     tp->t_maxseg;
1969                                 if (win < 2)
1970                                         win = 2;
1971                                 tp->snd_ssthresh = win * tp->t_maxseg;
1972                                 ENTER_FASTRECOVERY(tp);
1973                                 tp->snd_recover = tp->snd_max;
1974                                 tcp_callout_stop(tp, tp->tt_rexmt);
1975                                 tp->t_rtttime = 0;
1976                                 old_snd_nxt = tp->snd_nxt;
1977                                 tp->snd_nxt = th->th_ack;
1978                                 tp->snd_cwnd = tp->t_maxseg;
1979                                 tcp_output(tp);
1980                                 ++tcpstat.tcps_sndfastrexmit;
1981                                 tp->snd_cwnd = tp->snd_ssthresh;
1982                                 tp->rexmt_high = tp->snd_nxt;
1983                                 tp->t_flags &= ~TF_SACKRESCUED;
1984                                 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
1985                                         tp->snd_nxt = old_snd_nxt;
1986                                 KASSERT(tp->snd_limited <= 2,
1987                                     ("tp->snd_limited too big"));
1988                                 if (TCP_DO_SACK(tp))
1989                                         tcp_sack_rexmt(tp, th);
1990                                 else
1991                                         tp->snd_cwnd += tp->t_maxseg *
1992                                             (tp->t_dupacks - tp->snd_limited);
1993                         } else if (tcp_do_limitedtransmit) {
1994                                 u_long oldcwnd = tp->snd_cwnd;
1995                                 tcp_seq oldsndmax = tp->snd_max;
1996                                 tcp_seq oldsndnxt = tp->snd_nxt;
1997                                 /* outstanding data */
1998                                 uint32_t ownd = tp->snd_max - tp->snd_una;
1999                                 u_int sent;
2000
2001 #define iceildiv(n, d)          (((n)+(d)-1) / (d))
2002
2003                                 KASSERT(tp->t_dupacks == 1 ||
2004                                         tp->t_dupacks == 2,
2005                                     ("dupacks not 1 or 2"));
2006                                 if (tp->t_dupacks == 1)
2007                                         tp->snd_limited = 0;
2008                                 tp->snd_nxt = tp->snd_max;
2009                                 tp->snd_cwnd = ownd +
2010                                     (tp->t_dupacks - tp->snd_limited) *
2011                                     tp->t_maxseg;
2012                                 tcp_output(tp);
2013
2014                                 if (SEQ_LT(oldsndnxt, oldsndmax)) {
2015                                         KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una),
2016                                             ("snd_una moved in other threads"));
2017                                         tp->snd_nxt = oldsndnxt;
2018                                 }
2019                                 tp->snd_cwnd = oldcwnd;
2020                                 sent = tp->snd_max - oldsndmax;
2021                                 if (sent > tp->t_maxseg) {
2022                                         KASSERT((tp->t_dupacks == 2 &&
2023                                                  tp->snd_limited == 0) ||
2024                                                 (sent == tp->t_maxseg + 1 &&
2025                                                  tp->t_flags & TF_SENTFIN),
2026                                             ("sent too much"));
2027                                         KASSERT(sent <= tp->t_maxseg * 2,
2028                                             ("sent too many segments"));
2029                                         tp->snd_limited = 2;
2030                                         tcpstat.tcps_sndlimited += 2;
2031                                 } else if (sent > 0) {
2032                                         ++tp->snd_limited;
2033                                         ++tcpstat.tcps_sndlimited;
2034                                 } else if (tcp_do_early_retransmit &&
2035                                     (tcp_do_eifel_detect &&
2036                                      (tp->t_flags & TF_RCVD_TSTMP)) &&
2037                                     ownd < 4 * tp->t_maxseg &&
2038                                     tp->t_dupacks + 1 >=
2039                                       iceildiv(ownd, tp->t_maxseg) &&
2040                                     (!TCP_DO_SACK(tp) ||
2041                                      ownd <= tp->t_maxseg ||
2042                                      tcp_sack_has_sacked(&tp->scb,
2043                                                         ownd - tp->t_maxseg))) {
2044                                         ++tcpstat.tcps_sndearlyrexmit;
2045                                         tp->t_flags |= TF_EARLYREXMT;
2046                                         goto fastretransmit;
2047                                 }
2048                         }
2049                         goto drop;
2050                 }
2051
2052                 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2053                 tp->t_dupacks = 0;
2054                 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2055                         /*
2056                          * Detected optimistic ACK attack.
2057                          * Force slow-start to de-synchronize attack.
2058                          */
2059                         tp->snd_cwnd = tp->t_maxseg;
2060                         tp->snd_wacked = 0;
2061
2062                         tcpstat.tcps_rcvacktoomuch++;
2063                         goto dropafterack;
2064                 }
2065                 /*
2066                  * If we reach this point, ACK is not a duplicate,
2067                  *     i.e., it ACKs something we sent.
2068                  */
2069                 if (tp->t_flags & TF_NEEDSYN) {
2070                         /*
2071                          * T/TCP: Connection was half-synchronized, and our
2072                          * SYN has been ACK'd (so connection is now fully
2073                          * synchronized).  Go to non-starred state,
2074                          * increment snd_una for ACK of SYN, and check if
2075                          * we can do window scaling.
2076                          */
2077                         tp->t_flags &= ~TF_NEEDSYN;
2078                         tp->snd_una++;
2079                         /* Do window scaling? */
2080                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2081                             (TF_RCVD_SCALE | TF_REQ_SCALE))
2082                                 tp->rcv_scale = tp->request_r_scale;
2083                 }
2084
2085 process_ACK:
2086                 acked = th->th_ack - tp->snd_una;
2087                 tcpstat.tcps_rcvackpack++;
2088                 tcpstat.tcps_rcvackbyte += acked;
2089
2090                 if (tcp_do_eifel_detect && acked > 0 &&
2091                     (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2092                     (tp->t_flags & TF_FIRSTACCACK)) {
2093                         /* Eifel detection applicable. */
2094                         if (to.to_tsecr < tp->t_rexmtTS) {
2095                                 ++tcpstat.tcps_eifeldetected;
2096                                 tcp_revert_congestion_state(tp);
2097                                 if (tp->t_rxtshift != 1 ||
2098                                     ticks >= tp->t_badrxtwin)
2099                                         ++tcpstat.tcps_rttcantdetect;
2100                         }
2101                 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2102                         /*
2103                          * If we just performed our first retransmit,
2104                          * and the ACK arrives within our recovery window,
2105                          * then it was a mistake to do the retransmit
2106                          * in the first place.  Recover our original cwnd
2107                          * and ssthresh, and proceed to transmit where we
2108                          * left off.
2109                          */
2110                         tcp_revert_congestion_state(tp);
2111                         ++tcpstat.tcps_rttdetected;
2112                 }
2113
2114                 /*
2115                  * If we have a timestamp reply, update smoothed
2116                  * round trip time.  If no timestamp is present but
2117                  * transmit timer is running and timed sequence
2118                  * number was acked, update smoothed round trip time.
2119                  * Since we now have an rtt measurement, cancel the
2120                  * timer backoff (cf., Phil Karn's retransmit alg.).
2121                  * Recompute the initial retransmit timer.
2122                  *
2123                  * Some machines (certain windows boxes) send broken
2124                  * timestamp replies during the SYN+ACK phase, ignore
2125                  * timestamps of 0.
2126                  */
2127                 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2128                         tcp_xmit_timer(tp, ticks - to.to_tsecr + 1, th->th_ack);
2129                 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2130                         tcp_xmit_timer(tp, ticks - tp->t_rtttime, th->th_ack);
2131                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2132
2133                 /*
2134                  * If no data (only SYN) was ACK'd,
2135                  *    skip rest of ACK processing.
2136                  */
2137                 if (acked == 0)
2138                         goto step6;
2139
2140                 /* Stop looking for an acceptable ACK since one was received. */
2141                 tp->t_flags &= ~(TF_FIRSTACCACK | TF_FASTREXMT | TF_EARLYREXMT);
2142
2143                 if (acked > so->so_snd.ssb_cc) {
2144                         tp->snd_wnd -= so->so_snd.ssb_cc;
2145                         sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2146                         ourfinisacked = TRUE;
2147                 } else {
2148                         sbdrop(&so->so_snd.sb, acked);
2149                         tp->snd_wnd -= acked;
2150                         ourfinisacked = FALSE;
2151                 }
2152                 sowwakeup(so);
2153
2154                 /*
2155                  * Update window information.
2156                  */
2157                 if (acceptable_window_update(tp, th, tiwin)) {
2158                         /* keep track of pure window updates */
2159                         if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2160                             tiwin > tp->snd_wnd)
2161                                 tcpstat.tcps_rcvwinupd++;
2162                         tp->snd_wnd = tiwin;
2163                         tp->snd_wl1 = th->th_seq;
2164                         tp->snd_wl2 = th->th_ack;
2165                         if (tp->snd_wnd > tp->max_sndwnd)
2166                                 tp->max_sndwnd = tp->snd_wnd;
2167                         needoutput = TRUE;
2168                 }
2169
2170                 tp->snd_una = th->th_ack;
2171                 if (TCP_DO_SACK(tp))
2172                         tcp_sack_update_scoreboard(tp, &to);
2173                 if (IN_FASTRECOVERY(tp)) {
2174                         if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2175                                 EXIT_FASTRECOVERY(tp);
2176                                 needoutput = TRUE;
2177                                 /*
2178                                  * If the congestion window was inflated
2179                                  * to account for the other side's
2180                                  * cached packets, retract it.
2181                                  */
2182                                 if (!TCP_DO_SACK(tp))
2183                                         tp->snd_cwnd = tp->snd_ssthresh;
2184
2185                                 /*
2186                                  * Window inflation should have left us
2187                                  * with approximately snd_ssthresh outstanding
2188                                  * data.  But, in case we would be inclined
2189                                  * to send a burst, better do it using
2190                                  * slow start.
2191                                  */
2192                                 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2193                                            tp->snd_max + 2 * tp->t_maxseg))
2194                                         tp->snd_cwnd =
2195                                             (tp->snd_max - tp->snd_una) +
2196                                             2 * tp->t_maxseg;
2197
2198                                 tp->snd_wacked = 0;
2199                         } else {
2200                                 if (TCP_DO_SACK(tp)) {
2201                                         tp->snd_max_rexmt = tp->snd_max;
2202                                         tcp_sack_rexmt(tp, th);
2203                                 } else {
2204                                         tcp_newreno_partial_ack(tp, th, acked);
2205                                 }
2206                                 needoutput = FALSE;
2207                         }
2208                 } else {
2209                         /*
2210                          * Open the congestion window.  When in slow-start,
2211                          * open exponentially: maxseg per packet.  Otherwise,
2212                          * open linearly: maxseg per window.
2213                          */
2214                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
2215                                 u_int abc_sslimit =
2216                                     (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2217                                      tp->t_maxseg : 2 * tp->t_maxseg);
2218
2219                                 /* slow-start */
2220                                 tp->snd_cwnd += tcp_do_abc ?
2221                                     min(acked, abc_sslimit) : tp->t_maxseg;
2222                         } else {
2223                                 /* linear increase */
2224                                 tp->snd_wacked += tcp_do_abc ? acked :
2225                                     tp->t_maxseg;
2226                                 if (tp->snd_wacked >= tp->snd_cwnd) {
2227                                         tp->snd_wacked -= tp->snd_cwnd;
2228                                         tp->snd_cwnd += tp->t_maxseg;
2229                                 }
2230                         }
2231                         tp->snd_cwnd = min(tp->snd_cwnd,
2232                                            TCP_MAXWIN << tp->snd_scale);
2233                         tp->snd_recover = th->th_ack - 1;
2234                 }
2235                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2236                         tp->snd_nxt = tp->snd_una;
2237
2238                 /*
2239                  * If all outstanding data is acked, stop retransmit
2240                  * timer and remember to restart (more output or persist).
2241                  * If there is more data to be acked, restart retransmit
2242                  * timer, using current (possibly backed-off) value.
2243                  */
2244                 if (th->th_ack == tp->snd_max) {
2245                         tcp_callout_stop(tp, tp->tt_rexmt);
2246                         needoutput = TRUE;
2247                 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2248                         tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2249                             tcp_timer_rexmt);
2250                 }
2251
2252                 switch (tp->t_state) {
2253                 /*
2254                  * In FIN_WAIT_1 STATE in addition to the processing
2255                  * for the ESTABLISHED state if our FIN is now acknowledged
2256                  * then enter FIN_WAIT_2.
2257                  */
2258                 case TCPS_FIN_WAIT_1:
2259                         if (ourfinisacked) {
2260                                 /*
2261                                  * If we can't receive any more
2262                                  * data, then closing user can proceed.
2263                                  * Starting the timer is contrary to the
2264                                  * specification, but if we don't get a FIN
2265                                  * we'll hang forever.
2266                                  */
2267                                 if (so->so_state & SS_CANTRCVMORE) {
2268                                         soisdisconnected(so);
2269                                         tcp_callout_reset(tp, tp->tt_2msl,
2270                                             tp->t_maxidle, tcp_timer_2msl);
2271                                 }
2272                                 tp->t_state = TCPS_FIN_WAIT_2;
2273                         }
2274                         break;
2275
2276                 /*
2277                  * In CLOSING STATE in addition to the processing for
2278                  * the ESTABLISHED state if the ACK acknowledges our FIN
2279                  * then enter the TIME-WAIT state, otherwise ignore
2280                  * the segment.
2281                  */
2282                 case TCPS_CLOSING:
2283                         if (ourfinisacked) {
2284                                 tp->t_state = TCPS_TIME_WAIT;
2285                                 tcp_canceltimers(tp);
2286                                 tcp_callout_reset(tp, tp->tt_2msl,
2287                                             2 * tcp_rmx_msl(tp),
2288                                             tcp_timer_2msl);
2289                                 soisdisconnected(so);
2290                         }
2291                         break;
2292
2293                 /*
2294                  * In LAST_ACK, we may still be waiting for data to drain
2295                  * and/or to be acked, as well as for the ack of our FIN.
2296                  * If our FIN is now acknowledged, delete the TCB,
2297                  * enter the closed state and return.
2298                  */
2299                 case TCPS_LAST_ACK:
2300                         if (ourfinisacked) {
2301                                 tp = tcp_close(tp);
2302                                 goto drop;
2303                         }
2304                         break;
2305
2306                 /*
2307                  * In TIME_WAIT state the only thing that should arrive
2308                  * is a retransmission of the remote FIN.  Acknowledge
2309                  * it and restart the finack timer.
2310                  */
2311                 case TCPS_TIME_WAIT:
2312                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2313                             tcp_timer_2msl);
2314                         goto dropafterack;
2315                 }
2316         }
2317
2318 step6:
2319         /*
2320          * Update window information.
2321          * Don't look at window if no ACK: TAC's send garbage on first SYN.
2322          */
2323         if ((thflags & TH_ACK) &&
2324             acceptable_window_update(tp, th, tiwin)) {
2325                 /* keep track of pure window updates */
2326                 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2327                     tiwin > tp->snd_wnd)
2328                         tcpstat.tcps_rcvwinupd++;
2329                 tp->snd_wnd = tiwin;
2330                 tp->snd_wl1 = th->th_seq;
2331                 tp->snd_wl2 = th->th_ack;
2332                 if (tp->snd_wnd > tp->max_sndwnd)
2333                         tp->max_sndwnd = tp->snd_wnd;
2334                 needoutput = TRUE;
2335         }
2336
2337         /*
2338          * Process segments with URG.
2339          */
2340         if ((thflags & TH_URG) && th->th_urp &&
2341             !TCPS_HAVERCVDFIN(tp->t_state)) {
2342                 /*
2343                  * This is a kludge, but if we receive and accept
2344                  * random urgent pointers, we'll crash in
2345                  * soreceive.  It's hard to imagine someone
2346                  * actually wanting to send this much urgent data.
2347                  */
2348                 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2349                         th->th_urp = 0;                 /* XXX */
2350                         thflags &= ~TH_URG;             /* XXX */
2351                         goto dodata;                    /* XXX */
2352                 }
2353                 /*
2354                  * If this segment advances the known urgent pointer,
2355                  * then mark the data stream.  This should not happen
2356                  * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2357                  * a FIN has been received from the remote side.
2358                  * In these states we ignore the URG.
2359                  *
2360                  * According to RFC961 (Assigned Protocols),
2361                  * the urgent pointer points to the last octet
2362                  * of urgent data.  We continue, however,
2363                  * to consider it to indicate the first octet
2364                  * of data past the urgent section as the original
2365                  * spec states (in one of two places).
2366                  */
2367                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2368                         tp->rcv_up = th->th_seq + th->th_urp;
2369                         so->so_oobmark = so->so_rcv.ssb_cc +
2370                             (tp->rcv_up - tp->rcv_nxt) - 1;
2371                         if (so->so_oobmark == 0)
2372                                 sosetstate(so, SS_RCVATMARK);
2373                         sohasoutofband(so);
2374                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2375                 }
2376                 /*
2377                  * Remove out of band data so doesn't get presented to user.
2378                  * This can happen independent of advancing the URG pointer,
2379                  * but if two URG's are pending at once, some out-of-band
2380                  * data may creep in... ick.
2381                  */
2382                 if (th->th_urp <= (u_long)tlen &&
2383                     !(so->so_options & SO_OOBINLINE)) {
2384                         /* hdr drop is delayed */
2385                         tcp_pulloutofband(so, th, m, drop_hdrlen);
2386                 }
2387         } else {
2388                 /*
2389                  * If no out of band data is expected,
2390                  * pull receive urgent pointer along
2391                  * with the receive window.
2392                  */
2393                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2394                         tp->rcv_up = tp->rcv_nxt;
2395         }
2396
2397 dodata:                                                 /* XXX */
2398         /*
2399          * Process the segment text, merging it into the TCP sequencing queue,
2400          * and arranging for acknowledgment of receipt if necessary.
2401          * This process logically involves adjusting tp->rcv_wnd as data
2402          * is presented to the user (this happens in tcp_usrreq.c,
2403          * case PRU_RCVD).  If a FIN has already been received on this
2404          * connection then we just ignore the text.
2405          */
2406         if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2407                 m_adj(m, drop_hdrlen);  /* delayed header drop */
2408                 /*
2409                  * Insert segment which includes th into TCP reassembly queue
2410                  * with control block tp.  Set thflags to whether reassembly now
2411                  * includes a segment with FIN.  This handles the common case
2412                  * inline (segment is the next to be received on an established
2413                  * connection, and the queue is empty), avoiding linkage into
2414                  * and removal from the queue and repetition of various
2415                  * conversions.
2416                  * Set DELACK for segments received in order, but ack
2417                  * immediately when segments are out of order (so
2418                  * fast retransmit can work).
2419                  */
2420                 if (th->th_seq == tp->rcv_nxt &&
2421                     LIST_EMPTY(&tp->t_segq) &&
2422                     TCPS_HAVEESTABLISHED(tp->t_state)) {
2423                         if (DELAY_ACK(tp)) {
2424                                 tcp_callout_reset(tp, tp->tt_delack,
2425                                     tcp_delacktime, tcp_timer_delack);
2426                         } else {
2427                                 tp->t_flags |= TF_ACKNOW;
2428                         }
2429                         tp->rcv_nxt += tlen;
2430                         thflags = th->th_flags & TH_FIN;
2431                         tcpstat.tcps_rcvpack++;
2432                         tcpstat.tcps_rcvbyte += tlen;
2433                         ND6_HINT(tp);
2434                         if (so->so_state & SS_CANTRCVMORE) {
2435                                 m_freem(m);
2436                         } else {
2437                                 lwkt_gettoken(&so->so_rcv.ssb_token);
2438                                 ssb_appendstream(&so->so_rcv, m);
2439                                 lwkt_reltoken(&so->so_rcv.ssb_token);
2440                         }
2441                         sorwakeup(so);
2442                 } else {
2443                         if (!(tp->t_flags & TF_DUPSEG)) {
2444                                 /* Initialize SACK report block. */
2445                                 tp->reportblk.rblk_start = th->th_seq;
2446                                 tp->reportblk.rblk_end = TCP_SACK_BLKEND(
2447                                     th->th_seq + tlen, thflags);
2448                         }
2449                         thflags = tcp_reass(tp, th, &tlen, m);
2450                         tp->t_flags |= TF_ACKNOW;
2451                 }
2452
2453                 /*
2454                  * Note the amount of data that peer has sent into
2455                  * our window, in order to estimate the sender's
2456                  * buffer size.
2457                  */
2458                 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2459         } else {
2460                 m_freem(m);
2461                 thflags &= ~TH_FIN;
2462         }
2463
2464         /*
2465          * If FIN is received ACK the FIN and let the user know
2466          * that the connection is closing.
2467          */
2468         if (thflags & TH_FIN) {
2469                 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2470                         socantrcvmore(so);
2471                         /*
2472                          * If connection is half-synchronized
2473                          * (ie NEEDSYN flag on) then delay ACK,
2474                          * so it may be piggybacked when SYN is sent.
2475                          * Otherwise, since we received a FIN then no
2476                          * more input can be expected, send ACK now.
2477                          */
2478                         if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2479                                 tcp_callout_reset(tp, tp->tt_delack,
2480                                     tcp_delacktime, tcp_timer_delack);
2481                         } else {
2482                                 tp->t_flags |= TF_ACKNOW;
2483                         }
2484                         tp->rcv_nxt++;
2485                 }
2486
2487                 switch (tp->t_state) {
2488                 /*
2489                  * In SYN_RECEIVED and ESTABLISHED STATES
2490                  * enter the CLOSE_WAIT state.
2491                  */
2492                 case TCPS_SYN_RECEIVED:
2493                         tp->t_starttime = ticks;
2494                         /*FALLTHROUGH*/
2495                 case TCPS_ESTABLISHED:
2496                         tp->t_state = TCPS_CLOSE_WAIT;
2497                         break;
2498
2499                 /*
2500                  * If still in FIN_WAIT_1 STATE FIN has not been acked so
2501                  * enter the CLOSING state.
2502                  */
2503                 case TCPS_FIN_WAIT_1:
2504                         tp->t_state = TCPS_CLOSING;
2505                         break;
2506
2507                 /*
2508                  * In FIN_WAIT_2 state enter the TIME_WAIT state,
2509                  * starting the time-wait timer, turning off the other
2510                  * standard timers.
2511                  */
2512                 case TCPS_FIN_WAIT_2:
2513                         tp->t_state = TCPS_TIME_WAIT;
2514                         tcp_canceltimers(tp);
2515                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2516                                     tcp_timer_2msl);
2517                         soisdisconnected(so);
2518                         break;
2519
2520                 /*
2521                  * In TIME_WAIT state restart the 2 MSL time_wait timer.
2522                  */
2523                 case TCPS_TIME_WAIT:
2524                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2525                             tcp_timer_2msl);
2526                         break;
2527                 }
2528         }
2529
2530 #ifdef TCPDEBUG
2531         if (so->so_options & SO_DEBUG)
2532                 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2533 #endif
2534
2535         /*
2536          * Return any desired output.
2537          */
2538         if (needoutput || (tp->t_flags & TF_ACKNOW))
2539                 tcp_output(tp);
2540         tcp_sack_report_cleanup(tp);
2541         return(IPPROTO_DONE);
2542
2543 dropafterack:
2544         /*
2545          * Generate an ACK dropping incoming segment if it occupies
2546          * sequence space, where the ACK reflects our state.
2547          *
2548          * We can now skip the test for the RST flag since all
2549          * paths to this code happen after packets containing
2550          * RST have been dropped.
2551          *
2552          * In the SYN-RECEIVED state, don't send an ACK unless the
2553          * segment we received passes the SYN-RECEIVED ACK test.
2554          * If it fails send a RST.  This breaks the loop in the
2555          * "LAND" DoS attack, and also prevents an ACK storm
2556          * between two listening ports that have been sent forged
2557          * SYN segments, each with the source address of the other.
2558          */
2559         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2560             (SEQ_GT(tp->snd_una, th->th_ack) ||
2561              SEQ_GT(th->th_ack, tp->snd_max)) ) {
2562                 rstreason = BANDLIM_RST_OPENPORT;
2563                 goto dropwithreset;
2564         }
2565 #ifdef TCPDEBUG
2566         if (so->so_options & SO_DEBUG)
2567                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2568 #endif
2569         m_freem(m);
2570         tp->t_flags |= TF_ACKNOW;
2571         tcp_output(tp);
2572         tcp_sack_report_cleanup(tp);
2573         return(IPPROTO_DONE);
2574
2575 dropwithreset:
2576         /*
2577          * Generate a RST, dropping incoming segment.
2578          * Make ACK acceptable to originator of segment.
2579          * Don't bother to respond if destination was broadcast/multicast.
2580          */
2581         if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2582                 goto drop;
2583         if (isipv6) {
2584                 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2585                     IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2586                         goto drop;
2587         } else {
2588                 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2589                     IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2590                     ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2591                     in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2592                         goto drop;
2593         }
2594         /* IPv6 anycast check is done at tcp6_input() */
2595
2596         /*
2597          * Perform bandwidth limiting.
2598          */
2599 #ifdef ICMP_BANDLIM
2600         if (badport_bandlim(rstreason) < 0)
2601                 goto drop;
2602 #endif
2603
2604 #ifdef TCPDEBUG
2605         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2606                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2607 #endif
2608         if (thflags & TH_ACK)
2609                 /* mtod() below is safe as long as hdr dropping is delayed */
2610                 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2611                             TH_RST);
2612         else {
2613                 if (thflags & TH_SYN)
2614                         tlen++;
2615                 /* mtod() below is safe as long as hdr dropping is delayed */
2616                 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2617                             (tcp_seq)0, TH_RST | TH_ACK);
2618         }
2619         if (tp != NULL)
2620                 tcp_sack_report_cleanup(tp);
2621         return(IPPROTO_DONE);
2622
2623 drop:
2624         /*
2625          * Drop space held by incoming segment and return.
2626          */
2627 #ifdef TCPDEBUG
2628         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2629                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2630 #endif
2631         m_freem(m);
2632         if (tp != NULL)
2633                 tcp_sack_report_cleanup(tp);
2634         return(IPPROTO_DONE);
2635 }
2636
2637 /*
2638  * Parse TCP options and place in tcpopt.
2639  */
2640 static void
2641 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn,
2642     tcp_seq ack)
2643 {
2644         int opt, optlen, i;
2645
2646         to->to_flags = 0;
2647         for (; cnt > 0; cnt -= optlen, cp += optlen) {
2648                 opt = cp[0];
2649                 if (opt == TCPOPT_EOL)
2650                         break;
2651                 if (opt == TCPOPT_NOP)
2652                         optlen = 1;
2653                 else {
2654                         if (cnt < 2)
2655                                 break;
2656                         optlen = cp[1];
2657                         if (optlen < 2 || optlen > cnt)
2658                                 break;
2659                 }
2660                 switch (opt) {
2661                 case TCPOPT_MAXSEG:
2662                         if (optlen != TCPOLEN_MAXSEG)
2663                                 continue;
2664                         if (!is_syn)
2665                                 continue;
2666                         to->to_flags |= TOF_MSS;
2667                         bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2668                         to->to_mss = ntohs(to->to_mss);
2669                         break;
2670                 case TCPOPT_WINDOW:
2671                         if (optlen != TCPOLEN_WINDOW)
2672                                 continue;
2673                         if (!is_syn)
2674                                 continue;
2675                         to->to_flags |= TOF_SCALE;
2676                         to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2677                         break;
2678                 case TCPOPT_TIMESTAMP:
2679                         if (optlen != TCPOLEN_TIMESTAMP)
2680                                 continue;
2681                         to->to_flags |= TOF_TS;
2682                         bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2683                         to->to_tsval = ntohl(to->to_tsval);
2684                         bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2685                         to->to_tsecr = ntohl(to->to_tsecr);
2686                         /*
2687                          * If echoed timestamp is later than the current time,
2688                          * fall back to non RFC1323 RTT calculation.
2689                          */
2690                         if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2691                                 to->to_tsecr = 0;
2692                         break;
2693                 case TCPOPT_SACK_PERMITTED:
2694                         if (optlen != TCPOLEN_SACK_PERMITTED)
2695                                 continue;
2696                         if (!is_syn)
2697                                 continue;
2698                         to->to_flags |= TOF_SACK_PERMITTED;
2699                         break;
2700                 case TCPOPT_SACK:
2701                         if ((optlen - 2) & 0x07)        /* not multiple of 8 */
2702                                 continue;
2703                         to->to_nsackblocks = (optlen - 2) / 8;
2704                         to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2705                         to->to_flags |= TOF_SACK;
2706                         for (i = 0; i < to->to_nsackblocks; i++) {
2707                                 struct raw_sackblock *r = &to->to_sackblocks[i];
2708
2709                                 r->rblk_start = ntohl(r->rblk_start);
2710                                 r->rblk_end = ntohl(r->rblk_end);
2711
2712                                 if (SEQ_LEQ(r->rblk_end, r->rblk_start)) {
2713                                         /*
2714                                          * Invalid SACK block; discard all
2715                                          * SACK blocks
2716                                          */
2717                                         tcpstat.tcps_rcvbadsackopt++;
2718                                         to->to_nsackblocks = 0;
2719                                         to->to_sackblocks = NULL;
2720                                         to->to_flags &= ~TOF_SACK;
2721                                         break;
2722                                 }
2723                         }
2724                         if ((to->to_flags & TOF_SACK) &&
2725                             tcp_sack_ndsack_blocks(to->to_sackblocks,
2726                             to->to_nsackblocks, ack))
2727                                 to->to_flags |= TOF_DSACK;
2728                         break;
2729 #ifdef TCP_SIGNATURE
2730                 /*
2731                  * XXX In order to reply to a host which has set the
2732                  * TCP_SIGNATURE option in its initial SYN, we have to
2733                  * record the fact that the option was observed here
2734                  * for the syncache code to perform the correct response.
2735                  */
2736                 case TCPOPT_SIGNATURE:
2737                         if (optlen != TCPOLEN_SIGNATURE)
2738                                 continue;
2739                         to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2740                         break;
2741 #endif /* TCP_SIGNATURE */
2742                 default:
2743                         continue;
2744                 }
2745         }
2746 }
2747
2748 /*
2749  * Pull out of band byte out of a segment so
2750  * it doesn't appear in the user's data queue.
2751  * It is still reflected in the segment length for
2752  * sequencing purposes.
2753  * "off" is the delayed to be dropped hdrlen.
2754  */
2755 static void
2756 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2757 {
2758         int cnt = off + th->th_urp - 1;
2759
2760         while (cnt >= 0) {
2761                 if (m->m_len > cnt) {
2762                         char *cp = mtod(m, caddr_t) + cnt;
2763                         struct tcpcb *tp = sototcpcb(so);
2764
2765                         tp->t_iobc = *cp;
2766                         tp->t_oobflags |= TCPOOB_HAVEDATA;
2767                         bcopy(cp + 1, cp, m->m_len - cnt - 1);
2768                         m->m_len--;
2769                         if (m->m_flags & M_PKTHDR)
2770                                 m->m_pkthdr.len--;
2771                         return;
2772                 }
2773                 cnt -= m->m_len;
2774                 m = m->m_next;
2775                 if (m == NULL)
2776                         break;
2777         }
2778         panic("tcp_pulloutofband");
2779 }
2780
2781 /*
2782  * Collect new round-trip time estimate
2783  * and update averages and current timeout.
2784  */
2785 static void
2786 tcp_xmit_timer(struct tcpcb *tp, int rtt, tcp_seq ack)
2787 {
2788         int rebaserto = 0;
2789
2790         tcpstat.tcps_rttupdated++;
2791         tp->t_rttupdated++;
2792         if ((tp->t_flags & TF_REBASERTO) && SEQ_GT(ack, tp->snd_max_prev)) {
2793 #ifdef DEBUG_EIFEL_RESPONSE
2794                 kprintf("srtt/rttvar, prev %d/%d, cur %d/%d, ",
2795                     tp->t_srtt_prev, tp->t_rttvar_prev,
2796                     tp->t_srtt, tp->t_rttvar);
2797 #endif
2798
2799                 tcpstat.tcps_eifelresponse++;
2800                 rebaserto = 1;
2801                 tp->t_flags &= ~TF_REBASERTO;
2802                 tp->t_srtt = max(tp->t_srtt_prev, (rtt << TCP_RTT_SHIFT));
2803                 tp->t_rttvar = max(tp->t_rttvar_prev,
2804                     (rtt << (TCP_RTTVAR_SHIFT - 1)));
2805                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2806                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2807
2808 #ifdef DEBUG_EIFEL_RESPONSE
2809                 kprintf("new %d/%d ", tp->t_srtt, tp->t_rttvar);
2810 #endif
2811         } else if (tp->t_srtt != 0) {
2812                 int delta;
2813
2814                 /*
2815                  * srtt is stored as fixed point with 5 bits after the
2816                  * binary point (i.e., scaled by 8).  The following magic
2817                  * is equivalent to the smoothing algorithm in rfc793 with
2818                  * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2819                  * point).  Adjust rtt to origin 0.
2820                  */
2821                 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2822                         - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2823
2824                 if ((tp->t_srtt += delta) <= 0)
2825                         tp->t_srtt = 1;
2826
2827                 /*
2828                  * We accumulate a smoothed rtt variance (actually, a
2829                  * smoothed mean difference), then set the retransmit
2830                  * timer to smoothed rtt + 4 times the smoothed variance.
2831                  * rttvar is stored as fixed point with 4 bits after the
2832                  * binary point (scaled by 16).  The following is
2833                  * equivalent to rfc793 smoothing with an alpha of .75
2834                  * (rttvar = rttvar*3/4 + |delta| / 4).  This replaces
2835                  * rfc793's wired-in beta.
2836                  */
2837                 if (delta < 0)
2838                         delta = -delta;
2839                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2840                 if ((tp->t_rttvar += delta) <= 0)
2841                         tp->t_rttvar = 1;
2842                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2843                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2844         } else {
2845                 /*
2846                  * No rtt measurement yet - use the unsmoothed rtt.
2847                  * Set the variance to half the rtt (so our first
2848                  * retransmit happens at 3*rtt).
2849                  */
2850                 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2851                 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2852                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2853         }
2854         tp->t_rtttime = 0;
2855         tp->t_rxtshift = 0;
2856
2857 #ifdef DEBUG_EIFEL_RESPONSE
2858         if (rebaserto) {
2859                 kprintf("| rxtcur prev %d, old %d, ",
2860                     tp->t_rxtcur_prev, tp->t_rxtcur);
2861         }
2862 #endif
2863
2864         /*
2865          * the retransmit should happen at rtt + 4 * rttvar.
2866          * Because of the way we do the smoothing, srtt and rttvar
2867          * will each average +1/2 tick of bias.  When we compute
2868          * the retransmit timer, we want 1/2 tick of rounding and
2869          * 1 extra tick because of +-1/2 tick uncertainty in the
2870          * firing of the timer.  The bias will give us exactly the
2871          * 1.5 tick we need.  But, because the bias is
2872          * statistical, we have to test that we don't drop below
2873          * the minimum feasible timer (which is 2 ticks).
2874          */
2875         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2876                       max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2877
2878         if (rebaserto) {
2879                 if (tp->t_rxtcur < tp->t_rxtcur_prev + tcp_eifel_rtoinc) {
2880                         /*
2881                          * RFC4015 requires that the new RTO is at least
2882                          * 2*G (tcp_eifel_rtoinc) greater then the RTO
2883                          * (t_rxtcur_prev) when the spurious retransmit
2884                          * timeout happens.
2885                          *
2886                          * The above condition could be true, if the SRTT
2887                          * and RTTVAR used to calculate t_rxtcur_prev
2888                          * resulted in a value less than t_rttmin.  So
2889                          * simply increasing SRTT by tcp_eifel_rtoinc when
2890                          * preparing for the Eifel response in
2891                          * tcp_save_congestion_state() could not ensure
2892                          * that the new RTO will be tcp_eifel_rtoinc greater
2893                          * t_rxtcur_prev.
2894                          */
2895                         tp->t_rxtcur = tp->t_rxtcur_prev + tcp_eifel_rtoinc;
2896                 }
2897 #ifdef DEBUG_EIFEL_RESPONSE
2898                 kprintf("new %d\n", tp->t_rxtcur);
2899 #endif
2900         }
2901
2902         /*
2903          * We received an ack for a packet that wasn't retransmitted;
2904          * it is probably safe to discard any error indications we've
2905          * received recently.  This isn't quite right, but close enough
2906          * for now (a route might have failed after we sent a segment,
2907          * and the return path might not be symmetrical).
2908          */
2909         tp->t_softerror = 0;
2910 }
2911
2912 /*
2913  * Determine a reasonable value for maxseg size.
2914  * If the route is known, check route for mtu.
2915  * If none, use an mss that can be handled on the outgoing
2916  * interface without forcing IP to fragment; if bigger than
2917  * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2918  * to utilize large mbufs.  If no route is found, route has no mtu,
2919  * or the destination isn't local, use a default, hopefully conservative
2920  * size (usually 512 or the default IP max size, but no more than the mtu
2921  * of the interface), as we can't discover anything about intervening
2922  * gateways or networks.  We also initialize the congestion/slow start
2923  * window to be a single segment if the destination isn't local.
2924  * While looking at the routing entry, we also initialize other path-dependent
2925  * parameters from pre-set or cached values in the routing entry.
2926  *
2927  * Also take into account the space needed for options that we
2928  * send regularly.  Make maxseg shorter by that amount to assure
2929  * that we can send maxseg amount of data even when the options
2930  * are present.  Store the upper limit of the length of options plus
2931  * data in maxopd.
2932  *
2933  * NOTE that this routine is only called when we process an incoming
2934  * segment, for outgoing segments only tcp_mssopt is called.
2935  */
2936 void
2937 tcp_mss(struct tcpcb *tp, int offer)
2938 {
2939         struct rtentry *rt;
2940         struct ifnet *ifp;
2941         int rtt, mss;
2942         u_long bufsize;
2943         struct inpcb *inp = tp->t_inpcb;
2944         struct socket *so;
2945 #ifdef INET6
2946         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2947         size_t min_protoh = isipv6 ?
2948                             sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
2949                             sizeof(struct tcpiphdr);
2950 #else
2951         const boolean_t isipv6 = FALSE;
2952         const size_t min_protoh = sizeof(struct tcpiphdr);
2953 #endif
2954
2955         if (isipv6)
2956                 rt = tcp_rtlookup6(&inp->inp_inc);
2957         else
2958                 rt = tcp_rtlookup(&inp->inp_inc);
2959         if (rt == NULL) {
2960                 tp->t_maxopd = tp->t_maxseg =
2961                     (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
2962                 return;
2963         }
2964         ifp = rt->rt_ifp;
2965         so = inp->inp_socket;
2966
2967         /*
2968          * Offer == 0 means that there was no MSS on the SYN segment,
2969          * in this case we use either the interface mtu or tcp_mssdflt.
2970          *
2971          * An offer which is too large will be cut down later.
2972          */
2973         if (offer == 0) {
2974                 if (isipv6) {
2975                         if (in6_localaddr(&inp->in6p_faddr)) {
2976                                 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
2977                                         min_protoh;
2978                         } else {
2979                                 offer = tcp_v6mssdflt;
2980                         }
2981                 } else {
2982                         if (in_localaddr(inp->inp_faddr))
2983                                 offer = ifp->if_mtu - min_protoh;
2984                         else
2985                                 offer = tcp_mssdflt;
2986                 }
2987         }
2988
2989         /*
2990          * Prevent DoS attack with too small MSS. Round up
2991          * to at least minmss.
2992          *
2993          * Sanity check: make sure that maxopd will be large
2994          * enough to allow some data on segments even is the
2995          * all the option space is used (40bytes).  Otherwise
2996          * funny things may happen in tcp_output.
2997          */
2998         offer = max(offer, tcp_minmss);
2999         offer = max(offer, 64);
3000
3001         rt->rt_rmx.rmx_mssopt = offer;
3002
3003         /*
3004          * While we're here, check if there's an initial rtt
3005          * or rttvar.  Convert from the route-table units
3006          * to scaled multiples of the slow timeout timer.
3007          */
3008         if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
3009                 /*
3010                  * XXX the lock bit for RTT indicates that the value
3011                  * is also a minimum value; this is subject to time.
3012                  */
3013                 if (rt->rt_rmx.rmx_locks & RTV_RTT)
3014                         tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
3015                 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
3016                 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
3017                 tcpstat.tcps_usedrtt++;
3018                 if (rt->rt_rmx.rmx_rttvar) {
3019                         tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
3020                             (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
3021                         tcpstat.tcps_usedrttvar++;
3022                 } else {
3023                         /* default variation is +- 1 rtt */
3024                         tp->t_rttvar =
3025                             tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
3026                 }
3027                 TCPT_RANGESET(tp->t_rxtcur,
3028                               ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
3029                               tp->t_rttmin, TCPTV_REXMTMAX);
3030         }
3031
3032         /*
3033          * if there's an mtu associated with the route, use it
3034          * else, use the link mtu.  Take the smaller of mss or offer
3035          * as our final mss.
3036          */
3037         if (rt->rt_rmx.rmx_mtu) {
3038                 mss = rt->rt_rmx.rmx_mtu - min_protoh;
3039         } else {
3040                 if (isipv6)
3041                         mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
3042                 else
3043                         mss = ifp->if_mtu - min_protoh;
3044         }
3045         mss = min(mss, offer);
3046
3047         /*
3048          * maxopd stores the maximum length of data AND options
3049          * in a segment; maxseg is the amount of data in a normal
3050          * segment.  We need to store this value (maxopd) apart
3051          * from maxseg, because now every segment carries options
3052          * and thus we normally have somewhat less data in segments.
3053          */
3054         tp->t_maxopd = mss;
3055
3056         if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
3057             ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3058                 mss -= TCPOLEN_TSTAMP_APPA;
3059
3060 #if     (MCLBYTES & (MCLBYTES - 1)) == 0
3061         if (mss > MCLBYTES)
3062                 mss &= ~(MCLBYTES-1);
3063 #else
3064         if (mss > MCLBYTES)
3065                 mss = mss / MCLBYTES * MCLBYTES;
3066 #endif
3067         /*
3068          * If there's a pipesize, change the socket buffer
3069          * to that size.  Make the socket buffers an integral
3070          * number of mss units; if the mss is larger than
3071          * the socket buffer, decrease the mss.
3072          */
3073 #ifdef RTV_SPIPE
3074         if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3075 #endif
3076                 bufsize = so->so_snd.ssb_hiwat;
3077         if (bufsize < mss)
3078                 mss = bufsize;
3079         else {
3080                 bufsize = roundup(bufsize, mss);
3081                 if (bufsize > sb_max)
3082                         bufsize = sb_max;
3083                 if (bufsize > so->so_snd.ssb_hiwat)
3084                         ssb_reserve(&so->so_snd, bufsize, so, NULL);
3085         }
3086         tp->t_maxseg = mss;
3087
3088 #ifdef RTV_RPIPE
3089         if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3090 #endif
3091                 bufsize = so->so_rcv.ssb_hiwat;
3092         if (bufsize > mss) {
3093                 bufsize = roundup(bufsize, mss);
3094                 if (bufsize > sb_max)
3095                         bufsize = sb_max;
3096                 if (bufsize > so->so_rcv.ssb_hiwat) {
3097                         lwkt_gettoken(&so->so_rcv.ssb_token);
3098                         ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3099                         lwkt_reltoken(&so->so_rcv.ssb_token);
3100                 }
3101         }
3102
3103         /*
3104          * Set the slow-start flight size
3105          *
3106          * NOTE: t_maxseg must have been configured!
3107          */
3108         tp->snd_cwnd = tcp_initial_window(tp);
3109
3110         if (rt->rt_rmx.rmx_ssthresh) {
3111                 /*
3112                  * There's some sort of gateway or interface
3113                  * buffer limit on the path.  Use this to set
3114                  * the slow start threshhold, but set the
3115                  * threshold to no less than 2*mss.
3116                  */
3117                 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3118                 tcpstat.tcps_usedssthresh++;
3119         }
3120 }
3121
3122 /*
3123  * Determine the MSS option to send on an outgoing SYN.
3124  */
3125 int
3126 tcp_mssopt(struct tcpcb *tp)
3127 {
3128         struct rtentry *rt;
3129 #ifdef INET6
3130         boolean_t isipv6 =
3131             ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3132         int min_protoh = isipv6 ?
3133                              sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3134                              sizeof(struct tcpiphdr);
3135 #else
3136         const boolean_t isipv6 = FALSE;
3137         const size_t min_protoh = sizeof(struct tcpiphdr);
3138 #endif
3139
3140         if (isipv6)
3141                 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3142         else
3143                 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3144         if (rt == NULL)
3145                 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3146
3147         return (rt->rt_ifp->if_mtu - min_protoh);
3148 }
3149
3150 /*
3151  * When a partial ack arrives, force the retransmission of the
3152  * next unacknowledged segment.  Do not exit Fast Recovery.
3153  *
3154  * Implement the Slow-but-Steady variant of NewReno by restarting the
3155  * the retransmission timer.  Turn it off here so it can be restarted
3156  * later in tcp_output().
3157  */
3158 static void
3159 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3160 {
3161         tcp_seq old_snd_nxt = tp->snd_nxt;
3162         u_long ocwnd = tp->snd_cwnd;
3163
3164         tcp_callout_stop(tp, tp->tt_rexmt);
3165         tp->t_rtttime = 0;
3166         tp->snd_nxt = th->th_ack;
3167         /* Set snd_cwnd to one segment beyond acknowledged offset. */
3168         tp->snd_cwnd = tp->t_maxseg;
3169         tp->t_flags |= TF_ACKNOW;
3170         tcp_output(tp);
3171         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3172                 tp->snd_nxt = old_snd_nxt;
3173         /* partial window deflation */
3174         if (ocwnd > acked)
3175                 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3176         else
3177                 tp->snd_cwnd = tp->t_maxseg;
3178 }
3179
3180 /*
3181  * In contrast to the Slow-but-Steady NewReno variant,
3182  * we do not reset the retransmission timer for SACK retransmissions,
3183  * except when retransmitting snd_una.
3184  */
3185 static void
3186 tcp_sack_rexmt(struct tcpcb *tp, struct tcphdr *th)
3187 {
3188         tcp_seq old_snd_nxt = tp->snd_nxt;
3189         u_long ocwnd = tp->snd_cwnd;
3190         uint32_t pipe;
3191         int nseg = 0;           /* consecutive new segments */
3192         int nseg_rexmt = 0;     /* retransmitted segments */
3193 #define MAXBURST 4              /* limit burst of new packets on partial ack */
3194
3195         tp->t_rtttime = 0;
3196         pipe = tcp_sack_compute_pipe(tp);
3197         while ((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg &&
3198             (!tcp_do_smartsack || nseg < MAXBURST)) {
3199                 tcp_seq old_snd_max, old_rexmt_high, nextrexmt;
3200                 uint32_t sent, seglen;
3201                 boolean_t rescue;
3202                 int error;
3203
3204                 old_rexmt_high = tp->rexmt_high;
3205                 if (!tcp_sack_nextseg(tp, &nextrexmt, &seglen, &rescue)) {
3206                         tp->rexmt_high = old_rexmt_high;
3207                         break;
3208                 }
3209
3210                 /*
3211                  * If the next tranmission is a rescue retranmission,
3212                  * we check whether we have already sent some data
3213                  * (either new segments or retransmitted segments)
3214                  * into the the network or not.  Since the idea of rescue
3215                  * retransmission is to sustain ACK clock, as long as
3216                  * some segments are in the network, ACK clock will be
3217                  * kept ticking.
3218                  */
3219                 if (rescue && (nseg_rexmt > 0 || nseg > 0)) {
3220                         tp->rexmt_high = old_rexmt_high;
3221                         break;
3222                 }
3223
3224                 if (nextrexmt == tp->snd_max)
3225                         ++nseg;
3226                 else
3227                         ++nseg_rexmt;
3228                 tp->snd_nxt = nextrexmt;
3229                 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3230                 old_snd_max = tp->snd_max;
3231                 if (nextrexmt == tp->snd_una)
3232                         tcp_callout_stop(tp, tp->tt_rexmt);
3233                 error = tcp_output(tp);
3234                 if (error != 0) {
3235                         tp->rexmt_high = old_rexmt_high;
3236                         break;
3237                 }
3238                 sent = tp->snd_nxt - nextrexmt;
3239                 if (sent <= 0) {
3240                         tp->rexmt_high = old_rexmt_high;
3241                         break;
3242                 }
3243                 pipe += sent;
3244                 tcpstat.tcps_sndsackpack++;
3245                 tcpstat.tcps_sndsackbyte += sent;
3246
3247                 if (rescue) {
3248                         tcpstat.tcps_sackrescue++;
3249                         tp->rexmt_rescue = tp->snd_nxt;
3250                         tp->t_flags |= TF_SACKRESCUED;
3251                         break;
3252                 }
3253                 if (SEQ_LT(nextrexmt, old_snd_max) &&
3254                     SEQ_LT(tp->rexmt_high, tp->snd_nxt)) {
3255                         tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3256                         if (tcp_aggressive_rescuesack &&
3257                             (tp->t_flags & TF_SACKRESCUED) &&
3258                             SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
3259                                 /* Drag RescueRxt along with HighRxt */
3260                                 tp->rexmt_rescue = tp->rexmt_high;
3261                         }
3262                 }
3263         }
3264         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3265                 tp->snd_nxt = old_snd_nxt;
3266         tp->snd_cwnd = ocwnd;
3267 }
3268
3269 /*
3270  * Reset idle time and keep-alive timer, typically called when a valid
3271  * tcp packet is received but may also be called when FASTKEEP is set
3272  * to prevent the previous long-timeout from calculating to a drop.
3273  *
3274  * Only update t_rcvtime for non-SYN packets.
3275  *
3276  * Handle the case where one side thinks the connection is established
3277  * but the other side has, say, rebooted without cleaning out the
3278  * connection.   The SYNs could be construed as an attack and wind
3279  * up ignored, but in case it isn't an attack we can validate the
3280  * connection by forcing a keepalive.
3281  */
3282 void
3283 tcp_timer_keep_activity(struct tcpcb *tp, int thflags)
3284 {
3285         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
3286                 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
3287                         tp->t_flags |= TF_KEEPALIVE;
3288                         tcp_callout_reset(tp, tp->tt_keep, hz / 2,
3289                                           tcp_timer_keep);
3290                 } else {
3291                         tp->t_rcvtime = ticks;
3292                         tp->t_flags &= ~TF_KEEPALIVE;
3293                         tcp_callout_reset(tp, tp->tt_keep,
3294                                           tp->t_keepidle,
3295                                           tcp_timer_keep);
3296                 }
3297         }
3298 }
3299
3300 static int
3301 tcp_rmx_msl(const struct tcpcb *tp)
3302 {
3303         struct rtentry *rt;
3304         struct inpcb *inp = tp->t_inpcb;
3305         int msl;
3306 #ifdef INET6
3307         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3308 #else
3309         const boolean_t isipv6 = FALSE;
3310 #endif
3311
3312         if (isipv6)
3313                 rt = tcp_rtlookup6(&inp->inp_inc);
3314         else
3315                 rt = tcp_rtlookup(&inp->inp_inc);
3316         if (rt == NULL || rt->rt_rmx.rmx_msl == 0)
3317                 return tcp_msl;
3318
3319         msl = (rt->rt_rmx.rmx_msl * hz) / 1000;
3320         if (msl == 0)
3321                 msl = 1;
3322
3323         return msl;
3324 }
3325
3326 static void
3327 tcp_established(struct tcpcb *tp)
3328 {
3329         tp->t_state = TCPS_ESTABLISHED;
3330         tcp_callout_reset(tp, tp->tt_keep, tp->t_keepidle, tcp_timer_keep);
3331
3332         if (tp->t_rxtsyn > 0) {
3333                 /*
3334                  * RFC6298:
3335                  * "If the timer expires awaiting the ACK of a SYN segment
3336                  *  and the TCP implementation is using an RTO less than 3
3337                  *  seconds, the RTO MUST be re-initialized to 3 seconds
3338                  *  when data transmission begins"
3339                  */
3340                 if (tp->t_rxtcur < TCPTV_RTOBASE3)
3341                         tp->t_rxtcur = TCPTV_RTOBASE3;
3342         }
3343 }