tcp: Allow per-tcpcb keepintvl and keepcnt
[dragonfly.git] / sys / netinet / tcp_input.c
1 /*
2  * Copyright (c) 2002, 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2002, 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
36  *      The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *      This product includes software developed by the University of
49  *      California, Berkeley and its contributors.
50  * 4. Neither the name of the University nor the names of its contributors
51  *    may be used to endorse or promote products derived from this software
52  *    without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  *
66  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
67  * $FreeBSD: src/sys/netinet/tcp_input.c,v 1.107.2.38 2003/05/21 04:46:41 cjc Exp $
68  * $DragonFly: src/sys/netinet/tcp_input.c,v 1.68 2008/08/22 09:14:17 sephe Exp $
69  */
70
71 #include "opt_ipfw.h"           /* for ipfw_fwd         */
72 #include "opt_inet.h"
73 #include "opt_inet6.h"
74 #include "opt_ipsec.h"
75 #include "opt_tcpdebug.h"
76 #include "opt_tcp_input.h"
77
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/sysctl.h>
82 #include <sys/malloc.h>
83 #include <sys/mbuf.h>
84 #include <sys/proc.h>           /* for proc0 declaration */
85 #include <sys/protosw.h>
86 #include <sys/socket.h>
87 #include <sys/socketvar.h>
88 #include <sys/syslog.h>
89 #include <sys/in_cksum.h>
90
91 #include <sys/socketvar2.h>
92
93 #include <machine/cpu.h>        /* before tcp_seq.h, for tcp_random18() */
94 #include <machine/stdarg.h>
95
96 #include <net/if.h>
97 #include <net/route.h>
98
99 #include <netinet/in.h>
100 #include <netinet/in_systm.h>
101 #include <netinet/ip.h>
102 #include <netinet/ip_icmp.h>    /* for ICMP_BANDLIM */
103 #include <netinet/in_var.h>
104 #include <netinet/icmp_var.h>   /* for ICMP_BANDLIM */
105 #include <netinet/in_pcb.h>
106 #include <netinet/ip_var.h>
107 #include <netinet/ip6.h>
108 #include <netinet/icmp6.h>
109 #include <netinet6/nd6.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet6/in6_pcb.h>
112 #include <netinet/tcp.h>
113 #include <netinet/tcp_fsm.h>
114 #include <netinet/tcp_seq.h>
115 #include <netinet/tcp_timer.h>
116 #include <netinet/tcp_timer2.h>
117 #include <netinet/tcp_var.h>
118 #include <netinet6/tcp6_var.h>
119 #include <netinet/tcpip.h>
120
121 #ifdef TCPDEBUG
122 #include <netinet/tcp_debug.h>
123
124 u_char tcp_saveipgen[40];    /* the size must be of max ip header, now IPv6 */
125 struct tcphdr tcp_savetcp;
126 #endif
127
128 #ifdef FAST_IPSEC
129 #include <netproto/ipsec/ipsec.h>
130 #include <netproto/ipsec/ipsec6.h>
131 #endif
132
133 #ifdef IPSEC
134 #include <netinet6/ipsec.h>
135 #include <netinet6/ipsec6.h>
136 #include <netproto/key/key.h>
137 #endif
138
139 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
140
141 static int log_in_vain = 0;
142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW,
143     &log_in_vain, 0, "Log all incoming TCP connections");
144
145 static int blackhole = 0;
146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_RW,
147     &blackhole, 0, "Do not send RST when dropping refused connections");
148
149 int tcp_delack_enabled = 1;
150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_RW,
151     &tcp_delack_enabled, 0,
152     "Delay ACK to try and piggyback it onto a data packet");
153
154 #ifdef TCP_DROP_SYNFIN
155 static int drop_synfin = 0;
156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_RW,
157     &drop_synfin, 0, "Drop TCP packets with SYN+FIN set");
158 #endif
159
160 static int tcp_do_limitedtransmit = 1;
161 SYSCTL_INT(_net_inet_tcp, OID_AUTO, limitedtransmit, CTLFLAG_RW,
162     &tcp_do_limitedtransmit, 0, "Enable RFC 3042 (Limited Transmit)");
163
164 static int tcp_do_early_retransmit = 1;
165 SYSCTL_INT(_net_inet_tcp, OID_AUTO, earlyretransmit, CTLFLAG_RW,
166     &tcp_do_early_retransmit, 0, "Early retransmit");
167
168 int tcp_aggregate_acks = 1;
169 SYSCTL_INT(_net_inet_tcp, OID_AUTO, aggregate_acks, CTLFLAG_RW,
170     &tcp_aggregate_acks, 0, "Aggregate built-up acks into one ack");
171
172 int tcp_do_rfc3390 = 1;
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_RW,
174     &tcp_do_rfc3390, 0,
175     "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)");
176
177 static int tcp_do_eifel_detect = 1;
178 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
179     &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
180
181 static int tcp_do_abc = 1;
182 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
183     &tcp_do_abc, 0,
184     "TCP Appropriate Byte Counting (RFC 3465)");
185
186 /*
187  * Define as tunable for easy testing with SACK on and off.
188  * Warning:  do not change setting in the middle of an existing active TCP flow,
189  *   else strange things might happen to that flow.
190  */
191 int tcp_do_sack = 1;
192 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
193     &tcp_do_sack, 0, "Enable SACK Algorithms");
194
195 int tcp_do_smartsack = 1;
196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
197     &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
198
199 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
200     "TCP Segment Reassembly Queue");
201
202 int tcp_reass_maxseg = 0;
203 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
204     &tcp_reass_maxseg, 0,
205     "Global maximum number of TCP Segments in Reassembly Queue");
206
207 int tcp_reass_qsize = 0;
208 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
209     &tcp_reass_qsize, 0,
210     "Global number of TCP Segments currently in Reassembly Queue");
211
212 static int tcp_reass_overflows = 0;
213 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
214     &tcp_reass_overflows, 0,
215     "Global number of TCP Segment Reassembly Queue Overflows");
216
217 int tcp_do_autorcvbuf = 1;
218 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
219     &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
220
221 int tcp_autorcvbuf_inc = 16*1024;
222 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
223     &tcp_autorcvbuf_inc, 0,
224     "Incrementor step size of automatic receive buffer");
225
226 int tcp_autorcvbuf_max = 2*1024*1024;
227 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
228     &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
229
230
231 static void      tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t);
232 static void      tcp_pulloutofband(struct socket *,
233                      struct tcphdr *, struct mbuf *, int);
234 static int       tcp_reass(struct tcpcb *, struct tcphdr *, int *,
235                      struct mbuf *);
236 static void      tcp_xmit_timer(struct tcpcb *, int);
237 static void      tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
238 static void      tcp_sack_rexmt(struct tcpcb *, struct tcphdr *);
239
240 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
241 #ifdef INET6
242 #define ND6_HINT(tp) \
243 do { \
244         if ((tp) && (tp)->t_inpcb && \
245             ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
246             (tp)->t_inpcb->in6p_route.ro_rt) \
247                 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
248 } while (0)
249 #else
250 #define ND6_HINT(tp)
251 #endif
252
253 /*
254  * Indicate whether this ack should be delayed.  We can delay the ack if
255  *      - delayed acks are enabled and
256  *      - there is no delayed ack timer in progress and
257  *      - our last ack wasn't a 0-sized window.  We never want to delay
258  *        the ack that opens up a 0-sized window.
259  */
260 #define DELAY_ACK(tp) \
261         (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
262         !(tp->t_flags & TF_RXWIN0SENT))
263
264 #define acceptable_window_update(tp, th, tiwin)                         \
265     (SEQ_LT(tp->snd_wl1, th->th_seq) ||                                 \
266      (tp->snd_wl1 == th->th_seq &&                                      \
267       (SEQ_LT(tp->snd_wl2, th->th_ack) ||                               \
268        (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
269
270 static int
271 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
272 {
273         struct tseg_qent *q;
274         struct tseg_qent *p = NULL;
275         struct tseg_qent *te;
276         struct socket *so = tp->t_inpcb->inp_socket;
277         int flags;
278
279         /*
280          * Call with th == NULL after become established to
281          * force pre-ESTABLISHED data up to user socket.
282          */
283         if (th == NULL)
284                 goto present;
285
286         /*
287          * Limit the number of segments in the reassembly queue to prevent
288          * holding on to too many segments (and thus running out of mbufs).
289          * Make sure to let the missing segment through which caused this
290          * queue.  Always keep one global queue entry spare to be able to
291          * process the missing segment.
292          */
293         if (th->th_seq != tp->rcv_nxt &&
294             tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
295                 tcp_reass_overflows++;
296                 tcpstat.tcps_rcvmemdrop++;
297                 m_freem(m);
298                 /* no SACK block to report */
299                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
300                 return (0);
301         }
302
303         /* Allocate a new queue entry. */
304         MALLOC(te, struct tseg_qent *, sizeof(struct tseg_qent), M_TSEGQ,
305                M_INTWAIT | M_NULLOK);
306         if (te == NULL) {
307                 tcpstat.tcps_rcvmemdrop++;
308                 m_freem(m);
309                 /* no SACK block to report */
310                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
311                 return (0);
312         }
313         atomic_add_int(&tcp_reass_qsize, 1);
314
315         /*
316          * Find a segment which begins after this one does.
317          */
318         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
319                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
320                         break;
321                 p = q;
322         }
323
324         /*
325          * If there is a preceding segment, it may provide some of
326          * our data already.  If so, drop the data from the incoming
327          * segment.  If it provides all of our data, drop us.
328          */
329         if (p != NULL) {
330                 tcp_seq_diff_t i;
331
332                 /* conversion to int (in i) handles seq wraparound */
333                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
334                 if (i > 0) {            /* overlaps preceding segment */
335                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
336                         /* enclosing block starts w/ preceding segment */
337                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
338                         if (i >= *tlenp) {
339                                 /* preceding encloses incoming segment */
340                                 tp->encloseblk.rblk_end = p->tqe_th->th_seq +
341                                     p->tqe_len;
342                                 tcpstat.tcps_rcvduppack++;
343                                 tcpstat.tcps_rcvdupbyte += *tlenp;
344                                 m_freem(m);
345                                 kfree(te, M_TSEGQ);
346                                 atomic_add_int(&tcp_reass_qsize, -1);
347                                 /*
348                                  * Try to present any queued data
349                                  * at the left window edge to the user.
350                                  * This is needed after the 3-WHS
351                                  * completes.
352                                  */
353                                 goto present;   /* ??? */
354                         }
355                         m_adj(m, i);
356                         *tlenp -= i;
357                         th->th_seq += i;
358                         /* incoming segment end is enclosing block end */
359                         tp->encloseblk.rblk_end = th->th_seq + *tlenp +
360                             ((th->th_flags & TH_FIN) != 0);
361                         /* trim end of reported D-SACK block */
362                         tp->reportblk.rblk_end = th->th_seq;
363                 }
364         }
365         tcpstat.tcps_rcvoopack++;
366         tcpstat.tcps_rcvoobyte += *tlenp;
367
368         /*
369          * While we overlap succeeding segments trim them or,
370          * if they are completely covered, dequeue them.
371          */
372         while (q) {
373                 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
374                 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
375                 struct tseg_qent *nq;
376
377                 if (i <= 0)
378                         break;
379                 if (!(tp->t_flags & TF_DUPSEG)) {    /* first time through */
380                         tp->t_flags |= (TF_DUPSEG | TF_ENCLOSESEG);
381                         tp->encloseblk = tp->reportblk;
382                         /* report trailing duplicate D-SACK segment */
383                         tp->reportblk.rblk_start = q->tqe_th->th_seq;
384                 }
385                 if ((tp->t_flags & TF_ENCLOSESEG) &&
386                     SEQ_GT(qend, tp->encloseblk.rblk_end)) {
387                         /* extend enclosing block if one exists */
388                         tp->encloseblk.rblk_end = qend;
389                 }
390                 if (i < q->tqe_len) {
391                         q->tqe_th->th_seq += i;
392                         q->tqe_len -= i;
393                         m_adj(q->tqe_m, i);
394                         break;
395                 }
396
397                 nq = LIST_NEXT(q, tqe_q);
398                 LIST_REMOVE(q, tqe_q);
399                 m_freem(q->tqe_m);
400                 kfree(q, M_TSEGQ);
401                 atomic_add_int(&tcp_reass_qsize, -1);
402                 q = nq;
403         }
404
405         /* Insert the new segment queue entry into place. */
406         te->tqe_m = m;
407         te->tqe_th = th;
408         te->tqe_len = *tlenp;
409
410         /* check if can coalesce with following segment */
411         if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
412                 tcp_seq tend = te->tqe_th->th_seq + te->tqe_len;
413
414                 te->tqe_len += q->tqe_len;
415                 if (q->tqe_th->th_flags & TH_FIN)
416                         te->tqe_th->th_flags |= TH_FIN;
417                 m_cat(te->tqe_m, q->tqe_m);
418                 tp->encloseblk.rblk_end = tend;
419                 /*
420                  * When not reporting a duplicate segment, use
421                  * the larger enclosing block as the SACK block.
422                  */
423                 if (!(tp->t_flags & TF_DUPSEG))
424                         tp->reportblk.rblk_end = tend;
425                 LIST_REMOVE(q, tqe_q);
426                 kfree(q, M_TSEGQ);
427                 atomic_add_int(&tcp_reass_qsize, -1);
428         }
429
430         if (p == NULL) {
431                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
432         } else {
433                 /* check if can coalesce with preceding segment */
434                 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
435                         p->tqe_len += te->tqe_len;
436                         m_cat(p->tqe_m, te->tqe_m);
437                         tp->encloseblk.rblk_start = p->tqe_th->th_seq;
438                         /*
439                          * When not reporting a duplicate segment, use
440                          * the larger enclosing block as the SACK block.
441                          */
442                         if (!(tp->t_flags & TF_DUPSEG))
443                                 tp->reportblk.rblk_start = p->tqe_th->th_seq;
444                         kfree(te, M_TSEGQ);
445                         atomic_add_int(&tcp_reass_qsize, -1);
446                 } else {
447                         LIST_INSERT_AFTER(p, te, tqe_q);
448                 }
449         }
450
451 present:
452         /*
453          * Present data to user, advancing rcv_nxt through
454          * completed sequence space.
455          */
456         if (!TCPS_HAVEESTABLISHED(tp->t_state))
457                 return (0);
458         q = LIST_FIRST(&tp->t_segq);
459         if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
460                 return (0);
461         tp->rcv_nxt += q->tqe_len;
462         if (!(tp->t_flags & TF_DUPSEG)) {
463                 /* no SACK block to report since ACK advanced */
464                 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
465         }
466         /* no enclosing block to report since ACK advanced */
467         tp->t_flags &= ~TF_ENCLOSESEG;
468         flags = q->tqe_th->th_flags & TH_FIN;
469         LIST_REMOVE(q, tqe_q);
470         KASSERT(LIST_EMPTY(&tp->t_segq) ||
471                 LIST_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
472                 ("segment not coalesced"));
473         if (so->so_state & SS_CANTRCVMORE) {
474                 m_freem(q->tqe_m);
475         } else {
476                 lwkt_gettoken(&so->so_rcv.ssb_token);
477                 ssb_appendstream(&so->so_rcv, q->tqe_m);
478                 lwkt_reltoken(&so->so_rcv.ssb_token);
479         }
480         kfree(q, M_TSEGQ);
481         atomic_add_int(&tcp_reass_qsize, -1);
482         ND6_HINT(tp);
483         sorwakeup(so);
484         return (flags);
485 }
486
487 /*
488  * TCP input routine, follows pages 65-76 of the
489  * protocol specification dated September, 1981 very closely.
490  */
491 #ifdef INET6
492 int
493 tcp6_input(struct mbuf **mp, int *offp, int proto)
494 {
495         struct mbuf *m = *mp;
496         struct in6_ifaddr *ia6;
497
498         IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
499
500         /*
501          * draft-itojun-ipv6-tcp-to-anycast
502          * better place to put this in?
503          */
504         ia6 = ip6_getdstifaddr(m);
505         if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
506                 struct ip6_hdr *ip6;
507
508                 ip6 = mtod(m, struct ip6_hdr *);
509                 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
510                             offsetof(struct ip6_hdr, ip6_dst));
511                 return (IPPROTO_DONE);
512         }
513
514         tcp_input(mp, offp, proto);
515         return (IPPROTO_DONE);
516 }
517 #endif
518
519 int
520 tcp_input(struct mbuf **mp, int *offp, int proto)
521 {
522         int off0;
523         struct tcphdr *th;
524         struct ip *ip = NULL;
525         struct ipovly *ipov;
526         struct inpcb *inp = NULL;
527         u_char *optp = NULL;
528         int optlen = 0;
529         int tlen, off;
530         int len = 0;
531         int drop_hdrlen;
532         struct tcpcb *tp = NULL;
533         int thflags;
534         struct socket *so = 0;
535         int todrop, acked;
536         boolean_t ourfinisacked, needoutput = FALSE;
537         u_long tiwin;
538         int recvwin;
539         struct tcpopt to;               /* options in this segment */
540         struct sockaddr_in *next_hop = NULL;
541         int rstreason; /* For badport_bandlim accounting purposes */
542         int cpu;
543         struct ip6_hdr *ip6 = NULL;
544         struct mbuf *m;
545 #ifdef INET6
546         boolean_t isipv6;
547 #else
548         const boolean_t isipv6 = FALSE;
549 #endif
550 #ifdef TCPDEBUG
551         short ostate = 0;
552 #endif
553
554         off0 = *offp;
555         m = *mp;
556         *mp = NULL;
557
558         tcpstat.tcps_rcvtotal++;
559
560         if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
561                 struct m_tag *mtag;
562
563                 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
564                 KKASSERT(mtag != NULL);
565                 next_hop = m_tag_data(mtag);
566         }
567
568 #ifdef INET6
569         isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
570 #endif
571
572         if (isipv6) {
573                 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
574                 ip6 = mtod(m, struct ip6_hdr *);
575                 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
576                 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
577                         tcpstat.tcps_rcvbadsum++;
578                         goto drop;
579                 }
580                 th = (struct tcphdr *)((caddr_t)ip6 + off0);
581
582                 /*
583                  * Be proactive about unspecified IPv6 address in source.
584                  * As we use all-zero to indicate unbounded/unconnected pcb,
585                  * unspecified IPv6 address can be used to confuse us.
586                  *
587                  * Note that packets with unspecified IPv6 destination is
588                  * already dropped in ip6_input.
589                  */
590                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
591                         /* XXX stat */
592                         goto drop;
593                 }
594         } else {
595                 /*
596                  * Get IP and TCP header together in first mbuf.
597                  * Note: IP leaves IP header in first mbuf.
598                  */
599                 if (off0 > sizeof(struct ip)) {
600                         ip_stripoptions(m);
601                         off0 = sizeof(struct ip);
602                 }
603                 /* already checked and pulled up in ip_demux() */
604                 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
605                     ("TCP header not in one mbuf: m->m_len %d", m->m_len));
606                 ip = mtod(m, struct ip *);
607                 ipov = (struct ipovly *)ip;
608                 th = (struct tcphdr *)((caddr_t)ip + off0);
609                 tlen = ip->ip_len;
610
611                 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
612                         if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
613                                 th->th_sum = m->m_pkthdr.csum_data;
614                         else
615                                 th->th_sum = in_pseudo(ip->ip_src.s_addr,
616                                                 ip->ip_dst.s_addr,
617                                                 htonl(m->m_pkthdr.csum_data +
618                                                         ip->ip_len +
619                                                         IPPROTO_TCP));
620                         th->th_sum ^= 0xffff;
621                 } else {
622                         /*
623                          * Checksum extended TCP header and data.
624                          */
625                         len = sizeof(struct ip) + tlen;
626                         bzero(ipov->ih_x1, sizeof ipov->ih_x1);
627                         ipov->ih_len = (u_short)tlen;
628                         ipov->ih_len = htons(ipov->ih_len);
629                         th->th_sum = in_cksum(m, len);
630                 }
631                 if (th->th_sum) {
632                         tcpstat.tcps_rcvbadsum++;
633                         goto drop;
634                 }
635 #ifdef INET6
636                 /* Re-initialization for later version check */
637                 ip->ip_v = IPVERSION;
638 #endif
639         }
640
641         /*
642          * Check that TCP offset makes sense,
643          * pull out TCP options and adjust length.              XXX
644          */
645         off = th->th_off << 2;
646         /* already checked and pulled up in ip_demux() */
647         KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
648             ("bad TCP data offset %d (tlen %d)", off, tlen));
649         tlen -= off;    /* tlen is used instead of ti->ti_len */
650         if (off > sizeof(struct tcphdr)) {
651                 if (isipv6) {
652                         IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
653                         ip6 = mtod(m, struct ip6_hdr *);
654                         th = (struct tcphdr *)((caddr_t)ip6 + off0);
655                 } else {
656                         /* already pulled up in ip_demux() */
657                         KASSERT(m->m_len >= sizeof(struct ip) + off,
658                             ("TCP header and options not in one mbuf: "
659                              "m_len %d, off %d", m->m_len, off));
660                 }
661                 optlen = off - sizeof(struct tcphdr);
662                 optp = (u_char *)(th + 1);
663         }
664         thflags = th->th_flags;
665
666 #ifdef TCP_DROP_SYNFIN
667         /*
668          * If the drop_synfin option is enabled, drop all packets with
669          * both the SYN and FIN bits set. This prevents e.g. nmap from
670          * identifying the TCP/IP stack.
671          *
672          * This is a violation of the TCP specification.
673          */
674         if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
675                 goto drop;
676 #endif
677
678         /*
679          * Convert TCP protocol specific fields to host format.
680          */
681         th->th_seq = ntohl(th->th_seq);
682         th->th_ack = ntohl(th->th_ack);
683         th->th_win = ntohs(th->th_win);
684         th->th_urp = ntohs(th->th_urp);
685
686         /*
687          * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
688          * until after ip6_savecontrol() is called and before other functions
689          * which don't want those proto headers.
690          * Because ip6_savecontrol() is going to parse the mbuf to
691          * search for data to be passed up to user-land, it wants mbuf
692          * parameters to be unchanged.
693          * XXX: the call of ip6_savecontrol() has been obsoleted based on
694          * latest version of the advanced API (20020110).
695          */
696         drop_hdrlen = off0 + off;
697
698         /*
699          * Locate pcb for segment.
700          */
701 findpcb:
702         /* IPFIREWALL_FORWARD section */
703         if (next_hop != NULL && !isipv6) {  /* IPv6 support is not there yet */
704                 /*
705                  * Transparently forwarded. Pretend to be the destination.
706                  * already got one like this?
707                  */
708                 cpu = mycpu->gd_cpuid;
709                 inp = in_pcblookup_hash(&tcbinfo[cpu],
710                                         ip->ip_src, th->th_sport,
711                                         ip->ip_dst, th->th_dport,
712                                         0, m->m_pkthdr.rcvif);
713                 if (!inp) {
714                         /*
715                          * It's new.  Try to find the ambushing socket.
716                          */
717
718                         /*
719                          * The rest of the ipfw code stores the port in
720                          * host order.  XXX
721                          * (The IP address is still in network order.)
722                          */
723                         in_port_t dport = next_hop->sin_port ?
724                                                 htons(next_hop->sin_port) :
725                                                 th->th_dport;
726
727                         cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
728                                           next_hop->sin_addr.s_addr, dport);
729                         inp = in_pcblookup_hash(&tcbinfo[cpu],
730                                                 ip->ip_src, th->th_sport,
731                                                 next_hop->sin_addr, dport,
732                                                 1, m->m_pkthdr.rcvif);
733                 }
734         } else {
735                 if (isipv6) {
736                         inp = in6_pcblookup_hash(&tcbinfo[0],
737                                                  &ip6->ip6_src, th->th_sport,
738                                                  &ip6->ip6_dst, th->th_dport,
739                                                  1, m->m_pkthdr.rcvif);
740                 } else {
741                         cpu = mycpu->gd_cpuid;
742                         inp = in_pcblookup_hash(&tcbinfo[cpu],
743                                                 ip->ip_src, th->th_sport,
744                                                 ip->ip_dst, th->th_dport,
745                                                 1, m->m_pkthdr.rcvif);
746                 }
747         }
748
749         /*
750          * If the state is CLOSED (i.e., TCB does not exist) then
751          * all data in the incoming segment is discarded.
752          * If the TCB exists but is in CLOSED state, it is embryonic,
753          * but should either do a listen or a connect soon.
754          */
755         if (inp == NULL) {
756                 if (log_in_vain) {
757 #ifdef INET6
758                         char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
759 #else
760                         char dbuf[sizeof "aaa.bbb.ccc.ddd"];
761                         char sbuf[sizeof "aaa.bbb.ccc.ddd"];
762 #endif
763                         if (isipv6) {
764                                 strcpy(dbuf, "[");
765                                 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
766                                 strcat(dbuf, "]");
767                                 strcpy(sbuf, "[");
768                                 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
769                                 strcat(sbuf, "]");
770                         } else {
771                                 strcpy(dbuf, inet_ntoa(ip->ip_dst));
772                                 strcpy(sbuf, inet_ntoa(ip->ip_src));
773                         }
774                         switch (log_in_vain) {
775                         case 1:
776                                 if (!(thflags & TH_SYN))
777                                         break;
778                         case 2:
779                                 log(LOG_INFO,
780                                     "Connection attempt to TCP %s:%d "
781                                     "from %s:%d flags:0x%02x\n",
782                                     dbuf, ntohs(th->th_dport), sbuf,
783                                     ntohs(th->th_sport), thflags);
784                                 break;
785                         default:
786                                 break;
787                         }
788                 }
789                 if (blackhole) {
790                         switch (blackhole) {
791                         case 1:
792                                 if (thflags & TH_SYN)
793                                         goto drop;
794                                 break;
795                         case 2:
796                                 goto drop;
797                         default:
798                                 goto drop;
799                         }
800                 }
801                 rstreason = BANDLIM_RST_CLOSEDPORT;
802                 goto dropwithreset;
803         }
804
805 #ifdef IPSEC
806         if (isipv6) {
807                 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
808                         ipsec6stat.in_polvio++;
809                         goto drop;
810                 }
811         } else {
812                 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
813                         ipsecstat.in_polvio++;
814                         goto drop;
815                 }
816         }
817 #endif
818 #ifdef FAST_IPSEC
819         if (isipv6) {
820                 if (ipsec6_in_reject(m, inp))
821                         goto drop;
822         } else {
823                 if (ipsec4_in_reject(m, inp))
824                         goto drop;
825         }
826 #endif
827         /* Check the minimum TTL for socket. */
828 #ifdef INET6
829         if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
830                 goto drop;
831 #endif
832
833         tp = intotcpcb(inp);
834         if (tp == NULL) {
835                 rstreason = BANDLIM_RST_CLOSEDPORT;
836                 goto dropwithreset;
837         }
838         if (tp->t_state <= TCPS_CLOSED)
839                 goto drop;
840
841         /* Unscale the window into a 32-bit value. */
842         if (!(thflags & TH_SYN))
843                 tiwin = th->th_win << tp->snd_scale;
844         else
845                 tiwin = th->th_win;
846
847         so = inp->inp_socket;
848
849 #ifdef TCPDEBUG
850         if (so->so_options & SO_DEBUG) {
851                 ostate = tp->t_state;
852                 if (isipv6)
853                         bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
854                 else
855                         bcopy(ip, tcp_saveipgen, sizeof(*ip));
856                 tcp_savetcp = *th;
857         }
858 #endif
859
860         bzero(&to, sizeof to);
861
862         if (so->so_options & SO_ACCEPTCONN) {
863                 struct in_conninfo inc;
864
865 #ifdef INET6
866                 inc.inc_isipv6 = (isipv6 == TRUE);
867 #endif
868                 if (isipv6) {
869                         inc.inc6_faddr = ip6->ip6_src;
870                         inc.inc6_laddr = ip6->ip6_dst;
871                         inc.inc6_route.ro_rt = NULL;            /* XXX */
872                 } else {
873                         inc.inc_faddr = ip->ip_src;
874                         inc.inc_laddr = ip->ip_dst;
875                         inc.inc_route.ro_rt = NULL;             /* XXX */
876                 }
877                 inc.inc_fport = th->th_sport;
878                 inc.inc_lport = th->th_dport;
879
880                 /*
881                  * If the state is LISTEN then ignore segment if it contains
882                  * a RST.  If the segment contains an ACK then it is bad and
883                  * send a RST.  If it does not contain a SYN then it is not
884                  * interesting; drop it.
885                  *
886                  * If the state is SYN_RECEIVED (syncache) and seg contains
887                  * an ACK, but not for our SYN/ACK, send a RST.  If the seg
888                  * contains a RST, check the sequence number to see if it
889                  * is a valid reset segment.
890                  */
891                 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
892                         if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
893                                 if (!syncache_expand(&inc, th, &so, m)) {
894                                         /*
895                                          * No syncache entry, or ACK was not
896                                          * for our SYN/ACK.  Send a RST.
897                                          */
898                                         tcpstat.tcps_badsyn++;
899                                         rstreason = BANDLIM_RST_OPENPORT;
900                                         goto dropwithreset;
901                                 }
902
903                                 /*
904                                  * Could not complete 3-way handshake,
905                                  * connection is being closed down, and
906                                  * syncache will free mbuf.
907                                  */
908                                 if (so == NULL)
909                                         return(IPPROTO_DONE);
910
911                                 /*
912                                  * We must be in the correct protocol thread
913                                  * for this connection.
914                                  */
915                                 KKASSERT(so->so_port == &curthread->td_msgport);
916
917                                 /*
918                                  * Socket is created in state SYN_RECEIVED.
919                                  * Continue processing segment.
920                                  */
921                                 inp = so->so_pcb;
922                                 tp = intotcpcb(inp);
923                                 /*
924                                  * This is what would have happened in
925                                  * tcp_output() when the SYN,ACK was sent.
926                                  */
927                                 tp->snd_up = tp->snd_una;
928                                 tp->snd_max = tp->snd_nxt = tp->iss + 1;
929                                 tp->last_ack_sent = tp->rcv_nxt;
930 /*
931  * XXX possible bug - it doesn't appear that tp->snd_wnd is unscaled
932  * until the _second_ ACK is received:
933  *    rcv SYN (set wscale opts)  --> send SYN/ACK, set snd_wnd = window.
934  *    rcv ACK, calculate tiwin --> process SYN_RECEIVED, determine wscale,
935  *        move to ESTAB, set snd_wnd to tiwin.
936  */
937                                 tp->snd_wnd = tiwin;    /* unscaled */
938                                 goto after_listen;
939                         }
940                         if (thflags & TH_RST) {
941                                 syncache_chkrst(&inc, th);
942                                 goto drop;
943                         }
944                         if (thflags & TH_ACK) {
945                                 syncache_badack(&inc);
946                                 tcpstat.tcps_badsyn++;
947                                 rstreason = BANDLIM_RST_OPENPORT;
948                                 goto dropwithreset;
949                         }
950                         goto drop;
951                 }
952
953                 /*
954                  * Segment's flags are (SYN) or (SYN | FIN).
955                  */
956 #ifdef INET6
957                 /*
958                  * If deprecated address is forbidden,
959                  * we do not accept SYN to deprecated interface
960                  * address to prevent any new inbound connection from
961                  * getting established.
962                  * When we do not accept SYN, we send a TCP RST,
963                  * with deprecated source address (instead of dropping
964                  * it).  We compromise it as it is much better for peer
965                  * to send a RST, and RST will be the final packet
966                  * for the exchange.
967                  *
968                  * If we do not forbid deprecated addresses, we accept
969                  * the SYN packet.  RFC2462 does not suggest dropping
970                  * SYN in this case.
971                  * If we decipher RFC2462 5.5.4, it says like this:
972                  * 1. use of deprecated addr with existing
973                  *    communication is okay - "SHOULD continue to be
974                  *    used"
975                  * 2. use of it with new communication:
976                  *   (2a) "SHOULD NOT be used if alternate address
977                  *        with sufficient scope is available"
978                  *   (2b) nothing mentioned otherwise.
979                  * Here we fall into (2b) case as we have no choice in
980                  * our source address selection - we must obey the peer.
981                  *
982                  * The wording in RFC2462 is confusing, and there are
983                  * multiple description text for deprecated address
984                  * handling - worse, they are not exactly the same.
985                  * I believe 5.5.4 is the best one, so we follow 5.5.4.
986                  */
987                 if (isipv6 && !ip6_use_deprecated) {
988                         struct in6_ifaddr *ia6;
989
990                         if ((ia6 = ip6_getdstifaddr(m)) &&
991                             (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
992                                 tp = NULL;
993                                 rstreason = BANDLIM_RST_OPENPORT;
994                                 goto dropwithreset;
995                         }
996                 }
997 #endif
998                 /*
999                  * If it is from this socket, drop it, it must be forged.
1000                  * Don't bother responding if the destination was a broadcast.
1001                  */
1002                 if (th->th_dport == th->th_sport) {
1003                         if (isipv6) {
1004                                 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1005                                                        &ip6->ip6_src))
1006                                         goto drop;
1007                         } else {
1008                                 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1009                                         goto drop;
1010                         }
1011                 }
1012                 /*
1013                  * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1014                  *
1015                  * Note that it is quite possible to receive unicast
1016                  * link-layer packets with a broadcast IP address. Use
1017                  * in_broadcast() to find them.
1018                  */
1019                 if (m->m_flags & (M_BCAST | M_MCAST))
1020                         goto drop;
1021                 if (isipv6) {
1022                         if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1023                             IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1024                                 goto drop;
1025                 } else {
1026                         if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1027                             IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1028                             ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1029                             in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1030                                 goto drop;
1031                 }
1032                 /*
1033                  * SYN appears to be valid; create compressed TCP state
1034                  * for syncache, or perform t/tcp connection.
1035                  */
1036                 if (so->so_qlen <= so->so_qlimit) {
1037                         tcp_dooptions(&to, optp, optlen, TRUE);
1038                         if (!syncache_add(&inc, &to, th, &so, m))
1039                                 goto drop;
1040
1041                         /*
1042                          * Entry added to syncache, mbuf used to
1043                          * send SYN,ACK packet.
1044                          */
1045                         if (so == NULL)
1046                                 return(IPPROTO_DONE);
1047
1048                         /*
1049                          * We must be in the correct protocol thread for
1050                          * this connection.
1051                          */
1052                         KKASSERT(so->so_port == &curthread->td_msgport);
1053
1054                         inp = so->so_pcb;
1055                         tp = intotcpcb(inp);
1056                         tp->snd_wnd = tiwin;
1057                         tp->t_starttime = ticks;
1058                         tp->t_state = TCPS_ESTABLISHED;
1059
1060                         /*
1061                          * If there is a FIN, or if there is data and the
1062                          * connection is local, then delay SYN,ACK(SYN) in
1063                          * the hope of piggy-backing it on a response
1064                          * segment.  Otherwise must send ACK now in case
1065                          * the other side is slow starting.
1066                          */
1067                         if (DELAY_ACK(tp) &&
1068                             ((thflags & TH_FIN) ||
1069                              (tlen != 0 &&
1070                               ((isipv6 && in6_localaddr(&inp->in6p_faddr)) ||
1071                                (!isipv6 && in_localaddr(inp->inp_faddr)))))) {
1072                                 tcp_callout_reset(tp, tp->tt_delack,
1073                                     tcp_delacktime, tcp_timer_delack);
1074                                 tp->t_flags |= TF_NEEDSYN;
1075                         } else {
1076                                 tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN);
1077                         }
1078
1079                         tcpstat.tcps_connects++;
1080                         soisconnected(so);
1081                         goto trimthenstep6;
1082                 }
1083                 goto drop;
1084         }
1085
1086 after_listen:
1087         /*
1088          * Should not happen - syncache should pick up these connections.
1089          *
1090          * Once we are past handling listen sockets we must be in the
1091          * correct protocol processing thread.
1092          */
1093         KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1094         KKASSERT(so->so_port == &curthread->td_msgport);
1095
1096         /*
1097          * This is the second part of the MSS DoS prevention code (after
1098          * minmss on the sending side) and it deals with too many too small
1099          * tcp packets in a too short timeframe (1 second).
1100          *
1101          * XXX Removed.  This code was crap.  It does not scale to network
1102          *     speed, and default values break NFS.  Gone.
1103          */
1104         /* REMOVED */
1105
1106         /*
1107          * Segment received on connection.
1108          *
1109          * Reset idle time and keep-alive timer.  Don't waste time if less
1110          * then a second has elapsed.
1111          */
1112         if ((int)(ticks - tp->t_rcvtime) > hz)
1113                 tcp_timer_keep_activity(tp, thflags);
1114
1115         /*
1116          * Process options.
1117          * XXX this is tradtitional behavior, may need to be cleaned up.
1118          */
1119         tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0);
1120         if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1121                 if (to.to_flags & TOF_SCALE) {
1122                         tp->t_flags |= TF_RCVD_SCALE;
1123                         tp->requested_s_scale = to.to_requested_s_scale;
1124                 }
1125                 if (to.to_flags & TOF_TS) {
1126                         tp->t_flags |= TF_RCVD_TSTMP;
1127                         tp->ts_recent = to.to_tsval;
1128                         tp->ts_recent_age = ticks;
1129                 }
1130                 if (to.to_flags & TOF_MSS)
1131                         tcp_mss(tp, to.to_mss);
1132                 /*
1133                  * Only set the TF_SACK_PERMITTED per-connection flag
1134                  * if we got a SACK_PERMITTED option from the other side
1135                  * and the global tcp_do_sack variable is true.
1136                  */
1137                 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1138                         tp->t_flags |= TF_SACK_PERMITTED;
1139         }
1140
1141         /*
1142          * Header prediction: check for the two common cases
1143          * of a uni-directional data xfer.  If the packet has
1144          * no control flags, is in-sequence, the window didn't
1145          * change and we're not retransmitting, it's a
1146          * candidate.  If the length is zero and the ack moved
1147          * forward, we're the sender side of the xfer.  Just
1148          * free the data acked & wake any higher level process
1149          * that was blocked waiting for space.  If the length
1150          * is non-zero and the ack didn't move, we're the
1151          * receiver side.  If we're getting packets in-order
1152          * (the reassembly queue is empty), add the data to
1153          * the socket buffer and note that we need a delayed ack.
1154          * Make sure that the hidden state-flags are also off.
1155          * Since we check for TCPS_ESTABLISHED above, it can only
1156          * be TH_NEEDSYN.
1157          */
1158         if (tp->t_state == TCPS_ESTABLISHED &&
1159             (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1160             !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1161             (!(to.to_flags & TOF_TS) ||
1162              TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1163             th->th_seq == tp->rcv_nxt &&
1164             tp->snd_nxt == tp->snd_max) {
1165
1166                 /*
1167                  * If last ACK falls within this segment's sequence numbers,
1168                  * record the timestamp.
1169                  * NOTE that the test is modified according to the latest
1170                  * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1171                  */
1172                 if ((to.to_flags & TOF_TS) &&
1173                     SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1174                         tp->ts_recent_age = ticks;
1175                         tp->ts_recent = to.to_tsval;
1176                 }
1177
1178                 if (tlen == 0) {
1179                         if (SEQ_GT(th->th_ack, tp->snd_una) &&
1180                             SEQ_LEQ(th->th_ack, tp->snd_max) &&
1181                             tp->snd_cwnd >= tp->snd_wnd &&
1182                             !IN_FASTRECOVERY(tp)) {
1183                                 /*
1184                                  * This is a pure ack for outstanding data.
1185                                  */
1186                                 ++tcpstat.tcps_predack;
1187                                 /*
1188                                  * "bad retransmit" recovery
1189                                  *
1190                                  * If Eifel detection applies, then
1191                                  * it is deterministic, so use it
1192                                  * unconditionally over the old heuristic.
1193                                  * Otherwise, fall back to the old heuristic.
1194                                  */
1195                                 if (tcp_do_eifel_detect &&
1196                                     (to.to_flags & TOF_TS) && to.to_tsecr &&
1197                                     (tp->t_flags & TF_FIRSTACCACK)) {
1198                                         /* Eifel detection applicable. */
1199                                         if (to.to_tsecr < tp->t_rexmtTS) {
1200                                                 tcp_revert_congestion_state(tp);
1201                                                 ++tcpstat.tcps_eifeldetected;
1202                                         }
1203                                 } else if (tp->t_rxtshift == 1 &&
1204                                            ticks < tp->t_badrxtwin) {
1205                                         tcp_revert_congestion_state(tp);
1206                                         ++tcpstat.tcps_rttdetected;
1207                                 }
1208                                 tp->t_flags &= ~(TF_FIRSTACCACK |
1209                                                  TF_FASTREXMT | TF_EARLYREXMT);
1210                                 /*
1211                                  * Recalculate the retransmit timer / rtt.
1212                                  *
1213                                  * Some machines (certain windows boxes)
1214                                  * send broken timestamp replies during the
1215                                  * SYN+ACK phase, ignore timestamps of 0.
1216                                  */
1217                                 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1218                                         tcp_xmit_timer(tp,
1219                                                        ticks - to.to_tsecr + 1);
1220                                 } else if (tp->t_rtttime &&
1221                                            SEQ_GT(th->th_ack, tp->t_rtseq)) {
1222                                         tcp_xmit_timer(tp,
1223                                                        ticks - tp->t_rtttime);
1224                                 }
1225                                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1226                                 acked = th->th_ack - tp->snd_una;
1227                                 tcpstat.tcps_rcvackpack++;
1228                                 tcpstat.tcps_rcvackbyte += acked;
1229                                 sbdrop(&so->so_snd.sb, acked);
1230                                 tp->snd_recover = th->th_ack - 1;
1231                                 tp->snd_una = th->th_ack;
1232                                 tp->t_dupacks = 0;
1233                                 /*
1234                                  * Update window information.
1235                                  */
1236                                 if (tiwin != tp->snd_wnd &&
1237                                     acceptable_window_update(tp, th, tiwin)) {
1238                                         /* keep track of pure window updates */
1239                                         if (tp->snd_wl2 == th->th_ack &&
1240                                             tiwin > tp->snd_wnd)
1241                                                 tcpstat.tcps_rcvwinupd++;
1242                                         tp->snd_wnd = tiwin;
1243                                         tp->snd_wl1 = th->th_seq;
1244                                         tp->snd_wl2 = th->th_ack;
1245                                         if (tp->snd_wnd > tp->max_sndwnd)
1246                                                 tp->max_sndwnd = tp->snd_wnd;
1247                                 }
1248                                 m_freem(m);
1249                                 ND6_HINT(tp); /* some progress has been done */
1250                                 /*
1251                                  * If all outstanding data are acked, stop
1252                                  * retransmit timer, otherwise restart timer
1253                                  * using current (possibly backed-off) value.
1254                                  * If process is waiting for space,
1255                                  * wakeup/selwakeup/signal.  If data
1256                                  * are ready to send, let tcp_output
1257                                  * decide between more output or persist.
1258                                  */
1259                                 if (tp->snd_una == tp->snd_max) {
1260                                         tcp_callout_stop(tp, tp->tt_rexmt);
1261                                 } else if (!tcp_callout_active(tp,
1262                                             tp->tt_persist)) {
1263                                         tcp_callout_reset(tp, tp->tt_rexmt,
1264                                             tp->t_rxtcur, tcp_timer_rexmt);
1265                                 }
1266                                 sowwakeup(so);
1267                                 if (so->so_snd.ssb_cc > 0)
1268                                         tcp_output(tp);
1269                                 return(IPPROTO_DONE);
1270                         }
1271                 } else if (tiwin == tp->snd_wnd &&
1272                     th->th_ack == tp->snd_una &&
1273                     LIST_EMPTY(&tp->t_segq) &&
1274                     tlen <= ssb_space(&so->so_rcv)) {
1275                         u_long newsize = 0;     /* automatic sockbuf scaling */
1276                         /*
1277                          * This is a pure, in-sequence data packet
1278                          * with nothing on the reassembly queue and
1279                          * we have enough buffer space to take it.
1280                          */
1281                         ++tcpstat.tcps_preddat;
1282                         tp->rcv_nxt += tlen;
1283                         tcpstat.tcps_rcvpack++;
1284                         tcpstat.tcps_rcvbyte += tlen;
1285                         ND6_HINT(tp);   /* some progress has been done */
1286                 /*
1287                  * Automatic sizing of receive socket buffer.  Often the send
1288                  * buffer size is not optimally adjusted to the actual network
1289                  * conditions at hand (delay bandwidth product).  Setting the
1290                  * buffer size too small limits throughput on links with high
1291                  * bandwidth and high delay (eg. trans-continental/oceanic links).
1292                  *
1293                  * On the receive side the socket buffer memory is only rarely
1294                  * used to any significant extent.  This allows us to be much
1295                  * more aggressive in scaling the receive socket buffer.  For
1296                  * the case that the buffer space is actually used to a large
1297                  * extent and we run out of kernel memory we can simply drop
1298                  * the new segments; TCP on the sender will just retransmit it
1299                  * later.  Setting the buffer size too big may only consume too
1300                  * much kernel memory if the application doesn't read() from
1301                  * the socket or packet loss or reordering makes use of the
1302                  * reassembly queue.
1303                  *
1304                  * The criteria to step up the receive buffer one notch are:
1305                  *  1. the number of bytes received during the time it takes
1306                  *     one timestamp to be reflected back to us (the RTT);
1307                  *  2. received bytes per RTT is within seven eighth of the
1308                  *     current socket buffer size;
1309                  *  3. receive buffer size has not hit maximal automatic size;
1310                  *
1311                  * This algorithm does one step per RTT at most and only if
1312                  * we receive a bulk stream w/o packet losses or reorderings.
1313                  * Shrinking the buffer during idle times is not necessary as
1314                  * it doesn't consume any memory when idle.
1315                  *
1316                  * TODO: Only step up if the application is actually serving
1317                  * the buffer to better manage the socket buffer resources.
1318                  */
1319                         if (tcp_do_autorcvbuf &&
1320                             to.to_tsecr &&
1321                             (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1322                                 if (to.to_tsecr > tp->rfbuf_ts &&
1323                                     to.to_tsecr - tp->rfbuf_ts < hz) {
1324                                         if (tp->rfbuf_cnt >
1325                                             (so->so_rcv.ssb_hiwat / 8 * 7) &&
1326                                             so->so_rcv.ssb_hiwat <
1327                                             tcp_autorcvbuf_max) {
1328                                                 newsize =
1329                                                     ulmin(so->so_rcv.ssb_hiwat +
1330                                                           tcp_autorcvbuf_inc,
1331                                                           tcp_autorcvbuf_max);
1332                                         }
1333                                         /* Start over with next RTT. */
1334                                         tp->rfbuf_ts = 0;
1335                                         tp->rfbuf_cnt = 0;
1336                                 } else
1337                                         tp->rfbuf_cnt += tlen;  /* add up */
1338                         }
1339                         /*
1340                          * Add data to socket buffer.
1341                          */
1342                         if (so->so_state & SS_CANTRCVMORE) {
1343                                 m_freem(m);
1344                         } else {
1345                                 /*
1346                                  * Set new socket buffer size, give up when
1347                                  * limit is reached.
1348                                  *
1349                                  * Adjusting the size can mess up ACK
1350                                  * sequencing when pure window updates are
1351                                  * being avoided (which is the default),
1352                                  * so force an ack.
1353                                  */
1354                                 lwkt_gettoken(&so->so_rcv.ssb_token);
1355                                 if (newsize) {
1356                                         tp->t_flags |= TF_RXRESIZED;
1357                                         if (!ssb_reserve(&so->so_rcv, newsize,
1358                                                          so, NULL)) {
1359                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1360                                         }
1361                                         if (newsize >=
1362                                             (TCP_MAXWIN << tp->rcv_scale)) {
1363                                                 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1364                                         }
1365                                 }
1366                                 m_adj(m, drop_hdrlen); /* delayed header drop */
1367                                 ssb_appendstream(&so->so_rcv, m);
1368                                 lwkt_reltoken(&so->so_rcv.ssb_token);
1369                         }
1370                         sorwakeup(so);
1371                         /*
1372                          * This code is responsible for most of the ACKs
1373                          * the TCP stack sends back after receiving a data
1374                          * packet.  Note that the DELAY_ACK check fails if
1375                          * the delack timer is already running, which results
1376                          * in an ack being sent every other packet (which is
1377                          * what we want).
1378                          *
1379                          * We then further aggregate acks by not actually
1380                          * sending one until the protocol thread has completed
1381                          * processing the current backlog of packets.  This
1382                          * does not delay the ack any further, but allows us
1383                          * to take advantage of the packet aggregation that
1384                          * high speed NICs do (usually blocks of 8-10 packets)
1385                          * to send a single ack rather then four or five acks,
1386                          * greatly reducing the ack rate, the return channel
1387                          * bandwidth, and the protocol overhead on both ends.
1388                          *
1389                          * Since this also has the effect of slowing down
1390                          * the exponential slow-start ramp-up, systems with 
1391                          * very large bandwidth-delay products might want
1392                          * to turn the feature off.
1393                          */
1394                         if (DELAY_ACK(tp)) {
1395                                 tcp_callout_reset(tp, tp->tt_delack,
1396                                     tcp_delacktime, tcp_timer_delack);
1397                         } else if (tcp_aggregate_acks) {
1398                                 tp->t_flags |= TF_ACKNOW;
1399                                 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1400                                         tp->t_flags |= TF_ONOUTPUTQ;
1401                                         tp->tt_cpu = mycpu->gd_cpuid;
1402                                         TAILQ_INSERT_TAIL(
1403                                             &tcpcbackq[tp->tt_cpu],
1404                                             tp, t_outputq);
1405                                 }
1406                         } else {
1407                                 tp->t_flags |= TF_ACKNOW;
1408                                 tcp_output(tp);
1409                         }
1410                         return(IPPROTO_DONE);
1411                 }
1412         }
1413
1414         /*
1415          * Calculate amount of space in receive window,
1416          * and then do TCP input processing.
1417          * Receive window is amount of space in rcv queue,
1418          * but not less than advertised window.
1419          */
1420         recvwin = ssb_space(&so->so_rcv);
1421         if (recvwin < 0)
1422                 recvwin = 0;
1423         tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1424
1425         /* Reset receive buffer auto scaling when not in bulk receive mode. */
1426         tp->rfbuf_ts = 0;
1427         tp->rfbuf_cnt = 0;
1428
1429         switch (tp->t_state) {
1430         /*
1431          * If the state is SYN_RECEIVED:
1432          *      if seg contains an ACK, but not for our SYN/ACK, send a RST.
1433          */
1434         case TCPS_SYN_RECEIVED:
1435                 if ((thflags & TH_ACK) &&
1436                     (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1437                      SEQ_GT(th->th_ack, tp->snd_max))) {
1438                                 rstreason = BANDLIM_RST_OPENPORT;
1439                                 goto dropwithreset;
1440                 }
1441                 break;
1442
1443         /*
1444          * If the state is SYN_SENT:
1445          *      if seg contains an ACK, but not for our SYN, drop the input.
1446          *      if seg contains a RST, then drop the connection.
1447          *      if seg does not contain SYN, then drop it.
1448          * Otherwise this is an acceptable SYN segment
1449          *      initialize tp->rcv_nxt and tp->irs
1450          *      if seg contains ack then advance tp->snd_una
1451          *      if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1452          *      arrange for segment to be acked (eventually)
1453          *      continue processing rest of data/controls, beginning with URG
1454          */
1455         case TCPS_SYN_SENT:
1456                 if ((thflags & TH_ACK) &&
1457                     (SEQ_LEQ(th->th_ack, tp->iss) ||
1458                      SEQ_GT(th->th_ack, tp->snd_max))) {
1459                         rstreason = BANDLIM_UNLIMITED;
1460                         goto dropwithreset;
1461                 }
1462                 if (thflags & TH_RST) {
1463                         if (thflags & TH_ACK)
1464                                 tp = tcp_drop(tp, ECONNREFUSED);
1465                         goto drop;
1466                 }
1467                 if (!(thflags & TH_SYN))
1468                         goto drop;
1469                 tp->snd_wnd = th->th_win;       /* initial send window */
1470
1471                 tp->irs = th->th_seq;
1472                 tcp_rcvseqinit(tp);
1473                 if (thflags & TH_ACK) {
1474                         /* Our SYN was acked. */
1475                         tcpstat.tcps_connects++;
1476                         soisconnected(so);
1477                         /* Do window scaling on this connection? */
1478                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1479                             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
1480                                 tp->snd_scale = tp->requested_s_scale;
1481                                 tp->rcv_scale = tp->request_r_scale;
1482                         }
1483                         tp->rcv_adv += tp->rcv_wnd;
1484                         tp->snd_una++;          /* SYN is acked */
1485                         tcp_callout_stop(tp, tp->tt_rexmt);
1486                         /*
1487                          * If there's data, delay ACK; if there's also a FIN
1488                          * ACKNOW will be turned on later.
1489                          */
1490                         if (DELAY_ACK(tp) && tlen != 0) {
1491                                 tcp_callout_reset(tp, tp->tt_delack,
1492                                     tcp_delacktime, tcp_timer_delack);
1493                         } else {
1494                                 tp->t_flags |= TF_ACKNOW;
1495                         }
1496                         /*
1497                          * Received <SYN,ACK> in SYN_SENT[*] state.
1498                          * Transitions:
1499                          *      SYN_SENT  --> ESTABLISHED
1500                          *      SYN_SENT* --> FIN_WAIT_1
1501                          */
1502                         tp->t_starttime = ticks;
1503                         if (tp->t_flags & TF_NEEDFIN) {
1504                                 tp->t_state = TCPS_FIN_WAIT_1;
1505                                 tp->t_flags &= ~TF_NEEDFIN;
1506                                 thflags &= ~TH_SYN;
1507                         } else {
1508                                 tp->t_state = TCPS_ESTABLISHED;
1509                                 tcp_callout_reset(tp, tp->tt_keep,
1510                                                   tcp_getkeepidle(tp),
1511                                                   tcp_timer_keep);
1512                         }
1513                 } else {
1514                         /*
1515                          * Received initial SYN in SYN-SENT[*] state =>
1516                          * simultaneous open.
1517                          * Do 3-way handshake:
1518                          *        SYN-SENT -> SYN-RECEIVED
1519                          *        SYN-SENT* -> SYN-RECEIVED*
1520                          */
1521                         tp->t_flags |= TF_ACKNOW;
1522                         tcp_callout_stop(tp, tp->tt_rexmt);
1523                         tp->t_state = TCPS_SYN_RECEIVED;
1524                 }
1525
1526 trimthenstep6:
1527                 /*
1528                  * Advance th->th_seq to correspond to first data byte.
1529                  * If data, trim to stay within window,
1530                  * dropping FIN if necessary.
1531                  */
1532                 th->th_seq++;
1533                 if (tlen > tp->rcv_wnd) {
1534                         todrop = tlen - tp->rcv_wnd;
1535                         m_adj(m, -todrop);
1536                         tlen = tp->rcv_wnd;
1537                         thflags &= ~TH_FIN;
1538                         tcpstat.tcps_rcvpackafterwin++;
1539                         tcpstat.tcps_rcvbyteafterwin += todrop;
1540                 }
1541                 tp->snd_wl1 = th->th_seq - 1;
1542                 tp->rcv_up = th->th_seq;
1543                 /*
1544                  * Client side of transaction: already sent SYN and data.
1545                  * If the remote host used T/TCP to validate the SYN,
1546                  * our data will be ACK'd; if so, enter normal data segment
1547                  * processing in the middle of step 5, ack processing.
1548                  * Otherwise, goto step 6.
1549                  */
1550                 if (thflags & TH_ACK)
1551                         goto process_ACK;
1552
1553                 goto step6;
1554
1555         /*
1556          * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1557          *      do normal processing (we no longer bother with T/TCP).
1558          */
1559         case TCPS_LAST_ACK:
1560         case TCPS_CLOSING:
1561         case TCPS_TIME_WAIT:
1562                 break;  /* continue normal processing */
1563         }
1564
1565         /*
1566          * States other than LISTEN or SYN_SENT.
1567          * First check the RST flag and sequence number since reset segments
1568          * are exempt from the timestamp and connection count tests.  This
1569          * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1570          * below which allowed reset segments in half the sequence space
1571          * to fall though and be processed (which gives forged reset
1572          * segments with a random sequence number a 50 percent chance of
1573          * killing a connection).
1574          * Then check timestamp, if present.
1575          * Then check the connection count, if present.
1576          * Then check that at least some bytes of segment are within
1577          * receive window.  If segment begins before rcv_nxt,
1578          * drop leading data (and SYN); if nothing left, just ack.
1579          *
1580          *
1581          * If the RST bit is set, check the sequence number to see
1582          * if this is a valid reset segment.
1583          * RFC 793 page 37:
1584          *   In all states except SYN-SENT, all reset (RST) segments
1585          *   are validated by checking their SEQ-fields.  A reset is
1586          *   valid if its sequence number is in the window.
1587          * Note: this does not take into account delayed ACKs, so
1588          *   we should test against last_ack_sent instead of rcv_nxt.
1589          *   The sequence number in the reset segment is normally an
1590          *   echo of our outgoing acknowledgement numbers, but some hosts
1591          *   send a reset with the sequence number at the rightmost edge
1592          *   of our receive window, and we have to handle this case.
1593          * If we have multiple segments in flight, the intial reset
1594          * segment sequence numbers will be to the left of last_ack_sent,
1595          * but they will eventually catch up.
1596          * In any case, it never made sense to trim reset segments to
1597          * fit the receive window since RFC 1122 says:
1598          *   4.2.2.12  RST Segment: RFC-793 Section 3.4
1599          *
1600          *    A TCP SHOULD allow a received RST segment to include data.
1601          *
1602          *    DISCUSSION
1603          *         It has been suggested that a RST segment could contain
1604          *         ASCII text that encoded and explained the cause of the
1605          *         RST.  No standard has yet been established for such
1606          *         data.
1607          *
1608          * If the reset segment passes the sequence number test examine
1609          * the state:
1610          *    SYN_RECEIVED STATE:
1611          *      If passive open, return to LISTEN state.
1612          *      If active open, inform user that connection was refused.
1613          *    ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1614          *      Inform user that connection was reset, and close tcb.
1615          *    CLOSING, LAST_ACK STATES:
1616          *      Close the tcb.
1617          *    TIME_WAIT STATE:
1618          *      Drop the segment - see Stevens, vol. 2, p. 964 and
1619          *      RFC 1337.
1620          */
1621         if (thflags & TH_RST) {
1622                 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1623                     SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1624                         switch (tp->t_state) {
1625
1626                         case TCPS_SYN_RECEIVED:
1627                                 so->so_error = ECONNREFUSED;
1628                                 goto close;
1629
1630                         case TCPS_ESTABLISHED:
1631                         case TCPS_FIN_WAIT_1:
1632                         case TCPS_FIN_WAIT_2:
1633                         case TCPS_CLOSE_WAIT:
1634                                 so->so_error = ECONNRESET;
1635                         close:
1636                                 tp->t_state = TCPS_CLOSED;
1637                                 tcpstat.tcps_drops++;
1638                                 tp = tcp_close(tp);
1639                                 break;
1640
1641                         case TCPS_CLOSING:
1642                         case TCPS_LAST_ACK:
1643                                 tp = tcp_close(tp);
1644                                 break;
1645
1646                         case TCPS_TIME_WAIT:
1647                                 break;
1648                         }
1649                 }
1650                 goto drop;
1651         }
1652
1653         /*
1654          * RFC 1323 PAWS: If we have a timestamp reply on this segment
1655          * and it's less than ts_recent, drop it.
1656          */
1657         if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1658             TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1659
1660                 /* Check to see if ts_recent is over 24 days old.  */
1661                 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1662                         /*
1663                          * Invalidate ts_recent.  If this segment updates
1664                          * ts_recent, the age will be reset later and ts_recent
1665                          * will get a valid value.  If it does not, setting
1666                          * ts_recent to zero will at least satisfy the
1667                          * requirement that zero be placed in the timestamp
1668                          * echo reply when ts_recent isn't valid.  The
1669                          * age isn't reset until we get a valid ts_recent
1670                          * because we don't want out-of-order segments to be
1671                          * dropped when ts_recent is old.
1672                          */
1673                         tp->ts_recent = 0;
1674                 } else {
1675                         tcpstat.tcps_rcvduppack++;
1676                         tcpstat.tcps_rcvdupbyte += tlen;
1677                         tcpstat.tcps_pawsdrop++;
1678                         if (tlen)
1679                                 goto dropafterack;
1680                         goto drop;
1681                 }
1682         }
1683
1684         /*
1685          * In the SYN-RECEIVED state, validate that the packet belongs to
1686          * this connection before trimming the data to fit the receive
1687          * window.  Check the sequence number versus IRS since we know
1688          * the sequence numbers haven't wrapped.  This is a partial fix
1689          * for the "LAND" DoS attack.
1690          */
1691         if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1692                 rstreason = BANDLIM_RST_OPENPORT;
1693                 goto dropwithreset;
1694         }
1695
1696         todrop = tp->rcv_nxt - th->th_seq;
1697         if (todrop > 0) {
1698                 if (TCP_DO_SACK(tp)) {
1699                         /* Report duplicate segment at head of packet. */
1700                         tp->reportblk.rblk_start = th->th_seq;
1701                         tp->reportblk.rblk_end = th->th_seq + tlen;
1702                         if (thflags & TH_FIN)
1703                                 ++tp->reportblk.rblk_end;
1704                         if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1705                                 tp->reportblk.rblk_end = tp->rcv_nxt;
1706                         tp->t_flags |= (TF_DUPSEG | TF_SACKLEFT | TF_ACKNOW);
1707                 }
1708                 if (thflags & TH_SYN) {
1709                         thflags &= ~TH_SYN;
1710                         th->th_seq++;
1711                         if (th->th_urp > 1)
1712                                 th->th_urp--;
1713                         else
1714                                 thflags &= ~TH_URG;
1715                         todrop--;
1716                 }
1717                 /*
1718                  * Following if statement from Stevens, vol. 2, p. 960.
1719                  */
1720                 if (todrop > tlen ||
1721                     (todrop == tlen && !(thflags & TH_FIN))) {
1722                         /*
1723                          * Any valid FIN must be to the left of the window.
1724                          * At this point the FIN must be a duplicate or out
1725                          * of sequence; drop it.
1726                          */
1727                         thflags &= ~TH_FIN;
1728
1729                         /*
1730                          * Send an ACK to resynchronize and drop any data.
1731                          * But keep on processing for RST or ACK.
1732                          */
1733                         tp->t_flags |= TF_ACKNOW;
1734                         todrop = tlen;
1735                         tcpstat.tcps_rcvduppack++;
1736                         tcpstat.tcps_rcvdupbyte += todrop;
1737                 } else {
1738                         tcpstat.tcps_rcvpartduppack++;
1739                         tcpstat.tcps_rcvpartdupbyte += todrop;
1740                 }
1741                 drop_hdrlen += todrop;  /* drop from the top afterwards */
1742                 th->th_seq += todrop;
1743                 tlen -= todrop;
1744                 if (th->th_urp > todrop)
1745                         th->th_urp -= todrop;
1746                 else {
1747                         thflags &= ~TH_URG;
1748                         th->th_urp = 0;
1749                 }
1750         }
1751
1752         /*
1753          * If new data are received on a connection after the
1754          * user processes are gone, then RST the other end.
1755          */
1756         if ((so->so_state & SS_NOFDREF) &&
1757             tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1758                 tp = tcp_close(tp);
1759                 tcpstat.tcps_rcvafterclose++;
1760                 rstreason = BANDLIM_UNLIMITED;
1761                 goto dropwithreset;
1762         }
1763
1764         /*
1765          * If segment ends after window, drop trailing data
1766          * (and PUSH and FIN); if nothing left, just ACK.
1767          */
1768         todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1769         if (todrop > 0) {
1770                 tcpstat.tcps_rcvpackafterwin++;
1771                 if (todrop >= tlen) {
1772                         tcpstat.tcps_rcvbyteafterwin += tlen;
1773                         /*
1774                          * If a new connection request is received
1775                          * while in TIME_WAIT, drop the old connection
1776                          * and start over if the sequence numbers
1777                          * are above the previous ones.
1778                          */
1779                         if (thflags & TH_SYN &&
1780                             tp->t_state == TCPS_TIME_WAIT &&
1781                             SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1782                                 tp = tcp_close(tp);
1783                                 goto findpcb;
1784                         }
1785                         /*
1786                          * If window is closed can only take segments at
1787                          * window edge, and have to drop data and PUSH from
1788                          * incoming segments.  Continue processing, but
1789                          * remember to ack.  Otherwise, drop segment
1790                          * and ack.
1791                          */
1792                         if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1793                                 tp->t_flags |= TF_ACKNOW;
1794                                 tcpstat.tcps_rcvwinprobe++;
1795                         } else
1796                                 goto dropafterack;
1797                 } else
1798                         tcpstat.tcps_rcvbyteafterwin += todrop;
1799                 m_adj(m, -todrop);
1800                 tlen -= todrop;
1801                 thflags &= ~(TH_PUSH | TH_FIN);
1802         }
1803
1804         /*
1805          * If last ACK falls within this segment's sequence numbers,
1806          * record its timestamp.
1807          * NOTE:
1808          * 1) That the test incorporates suggestions from the latest
1809          *    proposal of the tcplw@cray.com list (Braden 1993/04/26).
1810          * 2) That updating only on newer timestamps interferes with
1811          *    our earlier PAWS tests, so this check should be solely
1812          *    predicated on the sequence space of this segment.
1813          * 3) That we modify the segment boundary check to be
1814          *        Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1815          *    instead of RFC1323's
1816          *        Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1817          *    This modified check allows us to overcome RFC1323's
1818          *    limitations as described in Stevens TCP/IP Illustrated
1819          *    Vol. 2 p.869. In such cases, we can still calculate the
1820          *    RTT correctly when RCV.NXT == Last.ACK.Sent.
1821          */
1822         if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1823             SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1824                                         + ((thflags & TH_SYN) != 0)
1825                                         + ((thflags & TH_FIN) != 0)))) {
1826                 tp->ts_recent_age = ticks;
1827                 tp->ts_recent = to.to_tsval;
1828         }
1829
1830         /*
1831          * If a SYN is in the window, then this is an
1832          * error and we send an RST and drop the connection.
1833          */
1834         if (thflags & TH_SYN) {
1835                 tp = tcp_drop(tp, ECONNRESET);
1836                 rstreason = BANDLIM_UNLIMITED;
1837                 goto dropwithreset;
1838         }
1839
1840         /*
1841          * If the ACK bit is off:  if in SYN-RECEIVED state or SENDSYN
1842          * flag is on (half-synchronized state), then queue data for
1843          * later processing; else drop segment and return.
1844          */
1845         if (!(thflags & TH_ACK)) {
1846                 if (tp->t_state == TCPS_SYN_RECEIVED ||
1847                     (tp->t_flags & TF_NEEDSYN))
1848                         goto step6;
1849                 else
1850                         goto drop;
1851         }
1852
1853         /*
1854          * Ack processing.
1855          */
1856         switch (tp->t_state) {
1857         /*
1858          * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1859          * ESTABLISHED state and continue processing.
1860          * The ACK was checked above.
1861          */
1862         case TCPS_SYN_RECEIVED:
1863
1864                 tcpstat.tcps_connects++;
1865                 soisconnected(so);
1866                 /* Do window scaling? */
1867                 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1868                     (TF_RCVD_SCALE | TF_REQ_SCALE)) {
1869                         tp->snd_scale = tp->requested_s_scale;
1870                         tp->rcv_scale = tp->request_r_scale;
1871                 }
1872                 /*
1873                  * Make transitions:
1874                  *      SYN-RECEIVED  -> ESTABLISHED
1875                  *      SYN-RECEIVED* -> FIN-WAIT-1
1876                  */
1877                 tp->t_starttime = ticks;
1878                 if (tp->t_flags & TF_NEEDFIN) {
1879                         tp->t_state = TCPS_FIN_WAIT_1;
1880                         tp->t_flags &= ~TF_NEEDFIN;
1881                 } else {
1882                         tp->t_state = TCPS_ESTABLISHED;
1883                         tcp_callout_reset(tp, tp->tt_keep,
1884                                           tcp_getkeepidle(tp),
1885                                           tcp_timer_keep);
1886                 }
1887                 /*
1888                  * If segment contains data or ACK, will call tcp_reass()
1889                  * later; if not, do so now to pass queued data to user.
1890                  */
1891                 if (tlen == 0 && !(thflags & TH_FIN))
1892                         tcp_reass(tp, NULL, NULL, NULL);
1893                 /* fall into ... */
1894
1895         /*
1896          * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
1897          * ACKs.  If the ack is in the range
1898          *      tp->snd_una < th->th_ack <= tp->snd_max
1899          * then advance tp->snd_una to th->th_ack and drop
1900          * data from the retransmission queue.  If this ACK reflects
1901          * more up to date window information we update our window information.
1902          */
1903         case TCPS_ESTABLISHED:
1904         case TCPS_FIN_WAIT_1:
1905         case TCPS_FIN_WAIT_2:
1906         case TCPS_CLOSE_WAIT:
1907         case TCPS_CLOSING:
1908         case TCPS_LAST_ACK:
1909         case TCPS_TIME_WAIT:
1910
1911                 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
1912                         if (TCP_DO_SACK(tp))
1913                                 tcp_sack_update_scoreboard(tp, &to);
1914                         if (tlen != 0 || tiwin != tp->snd_wnd) {
1915                                 tp->t_dupacks = 0;
1916                                 break;
1917                         }
1918                         tcpstat.tcps_rcvdupack++;
1919                         if (!tcp_callout_active(tp, tp->tt_rexmt) ||
1920                             th->th_ack != tp->snd_una) {
1921                                 tp->t_dupacks = 0;
1922                                 break;
1923                         }
1924                         /*
1925                          * We have outstanding data (other than
1926                          * a window probe), this is a completely
1927                          * duplicate ack (ie, window info didn't
1928                          * change), the ack is the biggest we've
1929                          * seen and we've seen exactly our rexmt
1930                          * threshhold of them, so assume a packet
1931                          * has been dropped and retransmit it.
1932                          * Kludge snd_nxt & the congestion
1933                          * window so we send only this one
1934                          * packet.
1935                          */
1936                         if (IN_FASTRECOVERY(tp)) {
1937                                 if (TCP_DO_SACK(tp)) {
1938                                         /* No artifical cwnd inflation. */
1939                                         tcp_sack_rexmt(tp, th);
1940                                 } else {
1941                                         /*
1942                                          * Dup acks mean that packets
1943                                          * have left the network
1944                                          * (they're now cached at the
1945                                          * receiver) so bump cwnd by
1946                                          * the amount in the receiver
1947                                          * to keep a constant cwnd
1948                                          * packets in the network.
1949                                          */
1950                                         tp->snd_cwnd += tp->t_maxseg;
1951                                         tcp_output(tp);
1952                                 }
1953                         } else if (SEQ_LT(th->th_ack, tp->snd_recover)) {
1954                                 tp->t_dupacks = 0;
1955                                 break;
1956                         } else if (++tp->t_dupacks == tcprexmtthresh) {
1957                                 tcp_seq old_snd_nxt;
1958                                 u_int win;
1959
1960 fastretransmit:
1961                                 if (tcp_do_eifel_detect &&
1962                                     (tp->t_flags & TF_RCVD_TSTMP)) {
1963                                         tcp_save_congestion_state(tp);
1964                                         tp->t_flags |= TF_FASTREXMT;
1965                                 }
1966                                 /*
1967                                  * We know we're losing at the current
1968                                  * window size, so do congestion avoidance:
1969                                  * set ssthresh to half the current window
1970                                  * and pull our congestion window back to the
1971                                  * new ssthresh.
1972                                  */
1973                                 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 /
1974                                     tp->t_maxseg;
1975                                 if (win < 2)
1976                                         win = 2;
1977                                 tp->snd_ssthresh = win * tp->t_maxseg;
1978                                 ENTER_FASTRECOVERY(tp);
1979                                 tp->snd_recover = tp->snd_max;
1980                                 tcp_callout_stop(tp, tp->tt_rexmt);
1981                                 tp->t_rtttime = 0;
1982                                 old_snd_nxt = tp->snd_nxt;
1983                                 tp->snd_nxt = th->th_ack;
1984                                 tp->snd_cwnd = tp->t_maxseg;
1985                                 tcp_output(tp);
1986                                 ++tcpstat.tcps_sndfastrexmit;
1987                                 tp->snd_cwnd = tp->snd_ssthresh;
1988                                 tp->rexmt_high = tp->snd_nxt;
1989                                 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
1990                                         tp->snd_nxt = old_snd_nxt;
1991                                 KASSERT(tp->snd_limited <= 2,
1992                                     ("tp->snd_limited too big"));
1993                                 if (TCP_DO_SACK(tp))
1994                                         tcp_sack_rexmt(tp, th);
1995                                 else
1996                                         tp->snd_cwnd += tp->t_maxseg *
1997                                             (tp->t_dupacks - tp->snd_limited);
1998                         } else if (tcp_do_limitedtransmit) {
1999                                 u_long oldcwnd = tp->snd_cwnd;
2000                                 tcp_seq oldsndmax = tp->snd_max;
2001                                 tcp_seq oldsndnxt = tp->snd_nxt;
2002                                 /* outstanding data */
2003                                 uint32_t ownd = tp->snd_max - tp->snd_una;
2004                                 u_int sent;
2005
2006 #define iceildiv(n, d)          (((n)+(d)-1) / (d))
2007
2008                                 KASSERT(tp->t_dupacks == 1 ||
2009                                         tp->t_dupacks == 2,
2010                                     ("dupacks not 1 or 2"));
2011                                 if (tp->t_dupacks == 1)
2012                                         tp->snd_limited = 0;
2013                                 tp->snd_nxt = tp->snd_max;
2014                                 tp->snd_cwnd = ownd +
2015                                     (tp->t_dupacks - tp->snd_limited) *
2016                                     tp->t_maxseg;
2017                                 tcp_output(tp);
2018
2019                                 /*
2020                                  * Other acks may have been processed,
2021                                  * snd_nxt cannot be reset to a value less
2022                                  * then snd_una.
2023                                  */
2024                                 if (SEQ_LT(oldsndnxt, oldsndmax)) {
2025                                     if (SEQ_GT(oldsndnxt, tp->snd_una))
2026                                         tp->snd_nxt = oldsndnxt;
2027                                     else
2028                                         tp->snd_nxt = tp->snd_una;
2029                                 }
2030                                 tp->snd_cwnd = oldcwnd;
2031                                 sent = tp->snd_max - oldsndmax;
2032                                 if (sent > tp->t_maxseg) {
2033                                         KASSERT((tp->t_dupacks == 2 &&
2034                                                  tp->snd_limited == 0) ||
2035                                                 (sent == tp->t_maxseg + 1 &&
2036                                                  tp->t_flags & TF_SENTFIN),
2037                                             ("sent too much"));
2038                                         KASSERT(sent <= tp->t_maxseg * 2,
2039                                             ("sent too many segments"));
2040                                         tp->snd_limited = 2;
2041                                         tcpstat.tcps_sndlimited += 2;
2042                                 } else if (sent > 0) {
2043                                         ++tp->snd_limited;
2044                                         ++tcpstat.tcps_sndlimited;
2045                                 } else if (tcp_do_early_retransmit &&
2046                                     (tcp_do_eifel_detect &&
2047                                      (tp->t_flags & TF_RCVD_TSTMP)) &&
2048                                     ownd < 4 * tp->t_maxseg &&
2049                                     tp->t_dupacks + 1 >=
2050                                       iceildiv(ownd, tp->t_maxseg) &&
2051                                     (!TCP_DO_SACK(tp) ||
2052                                      ownd <= tp->t_maxseg ||
2053                                      tcp_sack_has_sacked(&tp->scb,
2054                                                         ownd - tp->t_maxseg))) {
2055                                         ++tcpstat.tcps_sndearlyrexmit;
2056                                         tp->t_flags |= TF_EARLYREXMT;
2057                                         goto fastretransmit;
2058                                 }
2059                         }
2060                         goto drop;
2061                 }
2062
2063                 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2064                 tp->t_dupacks = 0;
2065                 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2066                         /*
2067                          * Detected optimistic ACK attack.
2068                          * Force slow-start to de-synchronize attack.
2069                          */
2070                         tp->snd_cwnd = tp->t_maxseg;
2071                         tp->snd_wacked = 0;
2072
2073                         tcpstat.tcps_rcvacktoomuch++;
2074                         goto dropafterack;
2075                 }
2076                 /*
2077                  * If we reach this point, ACK is not a duplicate,
2078                  *     i.e., it ACKs something we sent.
2079                  */
2080                 if (tp->t_flags & TF_NEEDSYN) {
2081                         /*
2082                          * T/TCP: Connection was half-synchronized, and our
2083                          * SYN has been ACK'd (so connection is now fully
2084                          * synchronized).  Go to non-starred state,
2085                          * increment snd_una for ACK of SYN, and check if
2086                          * we can do window scaling.
2087                          */
2088                         tp->t_flags &= ~TF_NEEDSYN;
2089                         tp->snd_una++;
2090                         /* Do window scaling? */
2091                         if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2092                             (TF_RCVD_SCALE | TF_REQ_SCALE)) {
2093                                 tp->snd_scale = tp->requested_s_scale;
2094                                 tp->rcv_scale = tp->request_r_scale;
2095                         }
2096                 }
2097
2098 process_ACK:
2099                 acked = th->th_ack - tp->snd_una;
2100                 tcpstat.tcps_rcvackpack++;
2101                 tcpstat.tcps_rcvackbyte += acked;
2102
2103                 if (tcp_do_eifel_detect && acked > 0 &&
2104                     (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2105                     (tp->t_flags & TF_FIRSTACCACK)) {
2106                         /* Eifel detection applicable. */
2107                         if (to.to_tsecr < tp->t_rexmtTS) {
2108                                 ++tcpstat.tcps_eifeldetected;
2109                                 tcp_revert_congestion_state(tp);
2110                                 if (tp->t_rxtshift == 1 &&
2111                                     ticks >= tp->t_badrxtwin)
2112                                         ++tcpstat.tcps_rttcantdetect;
2113                         }
2114                 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2115                         /*
2116                          * If we just performed our first retransmit,
2117                          * and the ACK arrives within our recovery window,
2118                          * then it was a mistake to do the retransmit
2119                          * in the first place.  Recover our original cwnd
2120                          * and ssthresh, and proceed to transmit where we
2121                          * left off.
2122                          */
2123                         tcp_revert_congestion_state(tp);
2124                         ++tcpstat.tcps_rttdetected;
2125                 }
2126
2127                 /*
2128                  * If we have a timestamp reply, update smoothed
2129                  * round trip time.  If no timestamp is present but
2130                  * transmit timer is running and timed sequence
2131                  * number was acked, update smoothed round trip time.
2132                  * Since we now have an rtt measurement, cancel the
2133                  * timer backoff (cf., Phil Karn's retransmit alg.).
2134                  * Recompute the initial retransmit timer.
2135                  *
2136                  * Some machines (certain windows boxes) send broken
2137                  * timestamp replies during the SYN+ACK phase, ignore
2138                  * timestamps of 0.
2139                  */
2140                 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2141                         tcp_xmit_timer(tp, ticks - to.to_tsecr + 1);
2142                 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2143                         tcp_xmit_timer(tp, ticks - tp->t_rtttime);
2144                 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2145
2146                 /*
2147                  * If no data (only SYN) was ACK'd,
2148                  *    skip rest of ACK processing.
2149                  */
2150                 if (acked == 0)
2151                         goto step6;
2152
2153                 /* Stop looking for an acceptable ACK since one was received. */
2154                 tp->t_flags &= ~(TF_FIRSTACCACK | TF_FASTREXMT | TF_EARLYREXMT);
2155
2156                 if (acked > so->so_snd.ssb_cc) {
2157                         tp->snd_wnd -= so->so_snd.ssb_cc;
2158                         sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2159                         ourfinisacked = TRUE;
2160                 } else {
2161                         sbdrop(&so->so_snd.sb, acked);
2162                         tp->snd_wnd -= acked;
2163                         ourfinisacked = FALSE;
2164                 }
2165                 sowwakeup(so);
2166
2167                 /*
2168                  * Update window information.
2169                  * Don't look at window if no ACK:
2170                  * TAC's send garbage on first SYN.
2171                  */
2172                 if (SEQ_LT(tp->snd_wl1, th->th_seq) ||
2173                     (tp->snd_wl1 == th->th_seq &&
2174                      (SEQ_LT(tp->snd_wl2, th->th_ack) ||
2175                       (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd)))) {
2176                         /* keep track of pure window updates */
2177                         if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2178                             tiwin > tp->snd_wnd)
2179                                 tcpstat.tcps_rcvwinupd++;
2180                         tp->snd_wnd = tiwin;
2181                         tp->snd_wl1 = th->th_seq;
2182                         tp->snd_wl2 = th->th_ack;
2183                         if (tp->snd_wnd > tp->max_sndwnd)
2184                                 tp->max_sndwnd = tp->snd_wnd;
2185                         needoutput = TRUE;
2186                 }
2187
2188                 tp->snd_una = th->th_ack;
2189                 if (TCP_DO_SACK(tp))
2190                         tcp_sack_update_scoreboard(tp, &to);
2191                 if (IN_FASTRECOVERY(tp)) {
2192                         if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2193                                 EXIT_FASTRECOVERY(tp);
2194                                 needoutput = TRUE;
2195                                 /*
2196                                  * If the congestion window was inflated
2197                                  * to account for the other side's
2198                                  * cached packets, retract it.
2199                                  */
2200                                 if (!TCP_DO_SACK(tp))
2201                                         tp->snd_cwnd = tp->snd_ssthresh;
2202
2203                                 /*
2204                                  * Window inflation should have left us
2205                                  * with approximately snd_ssthresh outstanding
2206                                  * data.  But, in case we would be inclined
2207                                  * to send a burst, better do it using
2208                                  * slow start.
2209                                  */
2210                                 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2211                                            tp->snd_max + 2 * tp->t_maxseg))
2212                                         tp->snd_cwnd =
2213                                             (tp->snd_max - tp->snd_una) +
2214                                             2 * tp->t_maxseg;
2215
2216                                 tp->snd_wacked = 0;
2217                         } else {
2218                                 if (TCP_DO_SACK(tp)) {
2219                                         tp->snd_max_rexmt = tp->snd_max;
2220                                         tcp_sack_rexmt(tp, th);
2221                                 } else {
2222                                         tcp_newreno_partial_ack(tp, th, acked);
2223                                 }
2224                                 needoutput = FALSE;
2225                         }
2226                 } else {
2227                         /*
2228                          * Open the congestion window.  When in slow-start,
2229                          * open exponentially: maxseg per packet.  Otherwise,
2230                          * open linearly: maxseg per window.
2231                          */
2232                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
2233                                 u_int abc_sslimit =
2234                                     (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2235                                      tp->t_maxseg : 2 * tp->t_maxseg);
2236
2237                                 /* slow-start */
2238                                 tp->snd_cwnd += tcp_do_abc ?
2239                                     min(acked, abc_sslimit) : tp->t_maxseg;
2240                         } else {
2241                                 /* linear increase */
2242                                 tp->snd_wacked += tcp_do_abc ? acked :
2243                                     tp->t_maxseg;
2244                                 if (tp->snd_wacked >= tp->snd_cwnd) {
2245                                         tp->snd_wacked -= tp->snd_cwnd;
2246                                         tp->snd_cwnd += tp->t_maxseg;
2247                                 }
2248                         }
2249                         tp->snd_cwnd = min(tp->snd_cwnd,
2250                                            TCP_MAXWIN << tp->snd_scale);
2251                         tp->snd_recover = th->th_ack - 1;
2252                 }
2253                 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2254                         tp->snd_nxt = tp->snd_una;
2255
2256                 /*
2257                  * If all outstanding data is acked, stop retransmit
2258                  * timer and remember to restart (more output or persist).
2259                  * If there is more data to be acked, restart retransmit
2260                  * timer, using current (possibly backed-off) value.
2261                  */
2262                 if (th->th_ack == tp->snd_max) {
2263                         tcp_callout_stop(tp, tp->tt_rexmt);
2264                         needoutput = TRUE;
2265                 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2266                         tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2267                             tcp_timer_rexmt);
2268                 }
2269
2270                 switch (tp->t_state) {
2271                 /*
2272                  * In FIN_WAIT_1 STATE in addition to the processing
2273                  * for the ESTABLISHED state if our FIN is now acknowledged
2274                  * then enter FIN_WAIT_2.
2275                  */
2276                 case TCPS_FIN_WAIT_1:
2277                         if (ourfinisacked) {
2278                                 /*
2279                                  * If we can't receive any more
2280                                  * data, then closing user can proceed.
2281                                  * Starting the timer is contrary to the
2282                                  * specification, but if we don't get a FIN
2283                                  * we'll hang forever.
2284                                  */
2285                                 if (so->so_state & SS_CANTRCVMORE) {
2286                                         soisdisconnected(so);
2287                                         tcp_callout_reset(tp, tp->tt_2msl,
2288                                             tp->t_maxidle, tcp_timer_2msl);
2289                                 }
2290                                 tp->t_state = TCPS_FIN_WAIT_2;
2291                         }
2292                         break;
2293
2294                 /*
2295                  * In CLOSING STATE in addition to the processing for
2296                  * the ESTABLISHED state if the ACK acknowledges our FIN
2297                  * then enter the TIME-WAIT state, otherwise ignore
2298                  * the segment.
2299                  */
2300                 case TCPS_CLOSING:
2301                         if (ourfinisacked) {
2302                                 tp->t_state = TCPS_TIME_WAIT;
2303                                 tcp_canceltimers(tp);
2304                                 tcp_callout_reset(tp, tp->tt_2msl,
2305                                             2 * tcp_msl, tcp_timer_2msl);
2306                                 soisdisconnected(so);
2307                         }
2308                         break;
2309
2310                 /*
2311                  * In LAST_ACK, we may still be waiting for data to drain
2312                  * and/or to be acked, as well as for the ack of our FIN.
2313                  * If our FIN is now acknowledged, delete the TCB,
2314                  * enter the closed state and return.
2315                  */
2316                 case TCPS_LAST_ACK:
2317                         if (ourfinisacked) {
2318                                 tp = tcp_close(tp);
2319                                 goto drop;
2320                         }
2321                         break;
2322
2323                 /*
2324                  * In TIME_WAIT state the only thing that should arrive
2325                  * is a retransmission of the remote FIN.  Acknowledge
2326                  * it and restart the finack timer.
2327                  */
2328                 case TCPS_TIME_WAIT:
2329                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2330                             tcp_timer_2msl);
2331                         goto dropafterack;
2332                 }
2333         }
2334
2335 step6:
2336         /*
2337          * Update window information.
2338          * Don't look at window if no ACK: TAC's send garbage on first SYN.
2339          */
2340         if ((thflags & TH_ACK) &&
2341             acceptable_window_update(tp, th, tiwin)) {
2342                 /* keep track of pure window updates */
2343                 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2344                     tiwin > tp->snd_wnd)
2345                         tcpstat.tcps_rcvwinupd++;
2346                 tp->snd_wnd = tiwin;
2347                 tp->snd_wl1 = th->th_seq;
2348                 tp->snd_wl2 = th->th_ack;
2349                 if (tp->snd_wnd > tp->max_sndwnd)
2350                         tp->max_sndwnd = tp->snd_wnd;
2351                 needoutput = TRUE;
2352         }
2353
2354         /*
2355          * Process segments with URG.
2356          */
2357         if ((thflags & TH_URG) && th->th_urp &&
2358             !TCPS_HAVERCVDFIN(tp->t_state)) {
2359                 /*
2360                  * This is a kludge, but if we receive and accept
2361                  * random urgent pointers, we'll crash in
2362                  * soreceive.  It's hard to imagine someone
2363                  * actually wanting to send this much urgent data.
2364                  */
2365                 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2366                         th->th_urp = 0;                 /* XXX */
2367                         thflags &= ~TH_URG;             /* XXX */
2368                         goto dodata;                    /* XXX */
2369                 }
2370                 /*
2371                  * If this segment advances the known urgent pointer,
2372                  * then mark the data stream.  This should not happen
2373                  * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2374                  * a FIN has been received from the remote side.
2375                  * In these states we ignore the URG.
2376                  *
2377                  * According to RFC961 (Assigned Protocols),
2378                  * the urgent pointer points to the last octet
2379                  * of urgent data.  We continue, however,
2380                  * to consider it to indicate the first octet
2381                  * of data past the urgent section as the original
2382                  * spec states (in one of two places).
2383                  */
2384                 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2385                         tp->rcv_up = th->th_seq + th->th_urp;
2386                         so->so_oobmark = so->so_rcv.ssb_cc +
2387                             (tp->rcv_up - tp->rcv_nxt) - 1;
2388                         if (so->so_oobmark == 0)
2389                                 sosetstate(so, SS_RCVATMARK);
2390                         sohasoutofband(so);
2391                         tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2392                 }
2393                 /*
2394                  * Remove out of band data so doesn't get presented to user.
2395                  * This can happen independent of advancing the URG pointer,
2396                  * but if two URG's are pending at once, some out-of-band
2397                  * data may creep in... ick.
2398                  */
2399                 if (th->th_urp <= (u_long)tlen &&
2400                     !(so->so_options & SO_OOBINLINE)) {
2401                         /* hdr drop is delayed */
2402                         tcp_pulloutofband(so, th, m, drop_hdrlen);
2403                 }
2404         } else {
2405                 /*
2406                  * If no out of band data is expected,
2407                  * pull receive urgent pointer along
2408                  * with the receive window.
2409                  */
2410                 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2411                         tp->rcv_up = tp->rcv_nxt;
2412         }
2413
2414 dodata:                                                 /* XXX */
2415         /*
2416          * Process the segment text, merging it into the TCP sequencing queue,
2417          * and arranging for acknowledgment of receipt if necessary.
2418          * This process logically involves adjusting tp->rcv_wnd as data
2419          * is presented to the user (this happens in tcp_usrreq.c,
2420          * case PRU_RCVD).  If a FIN has already been received on this
2421          * connection then we just ignore the text.
2422          */
2423         if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2424                 m_adj(m, drop_hdrlen);  /* delayed header drop */
2425                 /*
2426                  * Insert segment which includes th into TCP reassembly queue
2427                  * with control block tp.  Set thflags to whether reassembly now
2428                  * includes a segment with FIN.  This handles the common case
2429                  * inline (segment is the next to be received on an established
2430                  * connection, and the queue is empty), avoiding linkage into
2431                  * and removal from the queue and repetition of various
2432                  * conversions.
2433                  * Set DELACK for segments received in order, but ack
2434                  * immediately when segments are out of order (so
2435                  * fast retransmit can work).
2436                  */
2437                 if (th->th_seq == tp->rcv_nxt &&
2438                     LIST_EMPTY(&tp->t_segq) &&
2439                     TCPS_HAVEESTABLISHED(tp->t_state)) {
2440                         if (DELAY_ACK(tp)) {
2441                                 tcp_callout_reset(tp, tp->tt_delack,
2442                                     tcp_delacktime, tcp_timer_delack);
2443                         } else {
2444                                 tp->t_flags |= TF_ACKNOW;
2445                         }
2446                         tp->rcv_nxt += tlen;
2447                         thflags = th->th_flags & TH_FIN;
2448                         tcpstat.tcps_rcvpack++;
2449                         tcpstat.tcps_rcvbyte += tlen;
2450                         ND6_HINT(tp);
2451                         if (so->so_state & SS_CANTRCVMORE) {
2452                                 m_freem(m);
2453                         } else {
2454                                 lwkt_gettoken(&so->so_rcv.ssb_token);
2455                                 ssb_appendstream(&so->so_rcv, m);
2456                                 lwkt_reltoken(&so->so_rcv.ssb_token);
2457                         }
2458                         sorwakeup(so);
2459                 } else {
2460                         if (!(tp->t_flags & TF_DUPSEG)) {
2461                                 /* Initialize SACK report block. */
2462                                 tp->reportblk.rblk_start = th->th_seq;
2463                                 tp->reportblk.rblk_end = th->th_seq + tlen +
2464                                     ((thflags & TH_FIN) != 0);
2465                         }
2466                         thflags = tcp_reass(tp, th, &tlen, m);
2467                         tp->t_flags |= TF_ACKNOW;
2468                 }
2469
2470                 /*
2471                  * Note the amount of data that peer has sent into
2472                  * our window, in order to estimate the sender's
2473                  * buffer size.
2474                  */
2475                 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2476         } else {
2477                 m_freem(m);
2478                 thflags &= ~TH_FIN;
2479         }
2480
2481         /*
2482          * If FIN is received ACK the FIN and let the user know
2483          * that the connection is closing.
2484          */
2485         if (thflags & TH_FIN) {
2486                 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2487                         socantrcvmore(so);
2488                         /*
2489                          * If connection is half-synchronized
2490                          * (ie NEEDSYN flag on) then delay ACK,
2491                          * so it may be piggybacked when SYN is sent.
2492                          * Otherwise, since we received a FIN then no
2493                          * more input can be expected, send ACK now.
2494                          */
2495                         if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2496                                 tcp_callout_reset(tp, tp->tt_delack,
2497                                     tcp_delacktime, tcp_timer_delack);
2498                         } else {
2499                                 tp->t_flags |= TF_ACKNOW;
2500                         }
2501                         tp->rcv_nxt++;
2502                 }
2503
2504                 switch (tp->t_state) {
2505                 /*
2506                  * In SYN_RECEIVED and ESTABLISHED STATES
2507                  * enter the CLOSE_WAIT state.
2508                  */
2509                 case TCPS_SYN_RECEIVED:
2510                         tp->t_starttime = ticks;
2511                         /*FALLTHROUGH*/
2512                 case TCPS_ESTABLISHED:
2513                         tp->t_state = TCPS_CLOSE_WAIT;
2514                         break;
2515
2516                 /*
2517                  * If still in FIN_WAIT_1 STATE FIN has not been acked so
2518                  * enter the CLOSING state.
2519                  */
2520                 case TCPS_FIN_WAIT_1:
2521                         tp->t_state = TCPS_CLOSING;
2522                         break;
2523
2524                 /*
2525                  * In FIN_WAIT_2 state enter the TIME_WAIT state,
2526                  * starting the time-wait timer, turning off the other
2527                  * standard timers.
2528                  */
2529                 case TCPS_FIN_WAIT_2:
2530                         tp->t_state = TCPS_TIME_WAIT;
2531                         tcp_canceltimers(tp);
2532                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2533                                     tcp_timer_2msl);
2534                         soisdisconnected(so);
2535                         break;
2536
2537                 /*
2538                  * In TIME_WAIT state restart the 2 MSL time_wait timer.
2539                  */
2540                 case TCPS_TIME_WAIT:
2541                         tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_msl,
2542                             tcp_timer_2msl);
2543                         break;
2544                 }
2545         }
2546
2547 #ifdef TCPDEBUG
2548         if (so->so_options & SO_DEBUG)
2549                 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2550 #endif
2551
2552         /*
2553          * Return any desired output.
2554          */
2555         if (needoutput || (tp->t_flags & TF_ACKNOW))
2556                 tcp_output(tp);
2557         return(IPPROTO_DONE);
2558
2559 dropafterack:
2560         /*
2561          * Generate an ACK dropping incoming segment if it occupies
2562          * sequence space, where the ACK reflects our state.
2563          *
2564          * We can now skip the test for the RST flag since all
2565          * paths to this code happen after packets containing
2566          * RST have been dropped.
2567          *
2568          * In the SYN-RECEIVED state, don't send an ACK unless the
2569          * segment we received passes the SYN-RECEIVED ACK test.
2570          * If it fails send a RST.  This breaks the loop in the
2571          * "LAND" DoS attack, and also prevents an ACK storm
2572          * between two listening ports that have been sent forged
2573          * SYN segments, each with the source address of the other.
2574          */
2575         if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2576             (SEQ_GT(tp->snd_una, th->th_ack) ||
2577              SEQ_GT(th->th_ack, tp->snd_max)) ) {
2578                 rstreason = BANDLIM_RST_OPENPORT;
2579                 goto dropwithreset;
2580         }
2581 #ifdef TCPDEBUG
2582         if (so->so_options & SO_DEBUG)
2583                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2584 #endif
2585         m_freem(m);
2586         tp->t_flags |= TF_ACKNOW;
2587         tcp_output(tp);
2588         return(IPPROTO_DONE);
2589
2590 dropwithreset:
2591         /*
2592          * Generate a RST, dropping incoming segment.
2593          * Make ACK acceptable to originator of segment.
2594          * Don't bother to respond if destination was broadcast/multicast.
2595          */
2596         if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2597                 goto drop;
2598         if (isipv6) {
2599                 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2600                     IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2601                         goto drop;
2602         } else {
2603                 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2604                     IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2605                     ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2606                     in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2607                         goto drop;
2608         }
2609         /* IPv6 anycast check is done at tcp6_input() */
2610
2611         /*
2612          * Perform bandwidth limiting.
2613          */
2614 #ifdef ICMP_BANDLIM
2615         if (badport_bandlim(rstreason) < 0)
2616                 goto drop;
2617 #endif
2618
2619 #ifdef TCPDEBUG
2620         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2621                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2622 #endif
2623         if (thflags & TH_ACK)
2624                 /* mtod() below is safe as long as hdr dropping is delayed */
2625                 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2626                             TH_RST);
2627         else {
2628                 if (thflags & TH_SYN)
2629                         tlen++;
2630                 /* mtod() below is safe as long as hdr dropping is delayed */
2631                 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2632                             (tcp_seq)0, TH_RST | TH_ACK);
2633         }
2634         return(IPPROTO_DONE);
2635
2636 drop:
2637         /*
2638          * Drop space held by incoming segment and return.
2639          */
2640 #ifdef TCPDEBUG
2641         if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2642                 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2643 #endif
2644         m_freem(m);
2645         return(IPPROTO_DONE);
2646 }
2647
2648 /*
2649  * Parse TCP options and place in tcpopt.
2650  */
2651 static void
2652 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn)
2653 {
2654         int opt, optlen, i;
2655
2656         to->to_flags = 0;
2657         for (; cnt > 0; cnt -= optlen, cp += optlen) {
2658                 opt = cp[0];
2659                 if (opt == TCPOPT_EOL)
2660                         break;
2661                 if (opt == TCPOPT_NOP)
2662                         optlen = 1;
2663                 else {
2664                         if (cnt < 2)
2665                                 break;
2666                         optlen = cp[1];
2667                         if (optlen < 2 || optlen > cnt)
2668                                 break;
2669                 }
2670                 switch (opt) {
2671                 case TCPOPT_MAXSEG:
2672                         if (optlen != TCPOLEN_MAXSEG)
2673                                 continue;
2674                         if (!is_syn)
2675                                 continue;
2676                         to->to_flags |= TOF_MSS;
2677                         bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2678                         to->to_mss = ntohs(to->to_mss);
2679                         break;
2680                 case TCPOPT_WINDOW:
2681                         if (optlen != TCPOLEN_WINDOW)
2682                                 continue;
2683                         if (!is_syn)
2684                                 continue;
2685                         to->to_flags |= TOF_SCALE;
2686                         to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2687                         break;
2688                 case TCPOPT_TIMESTAMP:
2689                         if (optlen != TCPOLEN_TIMESTAMP)
2690                                 continue;
2691                         to->to_flags |= TOF_TS;
2692                         bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2693                         to->to_tsval = ntohl(to->to_tsval);
2694                         bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2695                         to->to_tsecr = ntohl(to->to_tsecr);
2696                         /*
2697                          * If echoed timestamp is later than the current time,
2698                          * fall back to non RFC1323 RTT calculation.
2699                          */
2700                         if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2701                                 to->to_tsecr = 0;
2702                         break;
2703                 case TCPOPT_SACK_PERMITTED:
2704                         if (optlen != TCPOLEN_SACK_PERMITTED)
2705                                 continue;
2706                         if (!is_syn)
2707                                 continue;
2708                         to->to_flags |= TOF_SACK_PERMITTED;
2709                         break;
2710                 case TCPOPT_SACK:
2711                         if ((optlen - 2) & 0x07)        /* not multiple of 8 */
2712                                 continue;
2713                         to->to_nsackblocks = (optlen - 2) / 8;
2714                         to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2715                         to->to_flags |= TOF_SACK;
2716                         for (i = 0; i < to->to_nsackblocks; i++) {
2717                                 struct raw_sackblock *r = &to->to_sackblocks[i];
2718
2719                                 r->rblk_start = ntohl(r->rblk_start);
2720                                 r->rblk_end = ntohl(r->rblk_end);
2721                         }
2722                         break;
2723 #ifdef TCP_SIGNATURE
2724                 /*
2725                  * XXX In order to reply to a host which has set the
2726                  * TCP_SIGNATURE option in its initial SYN, we have to
2727                  * record the fact that the option was observed here
2728                  * for the syncache code to perform the correct response.
2729                  */
2730                 case TCPOPT_SIGNATURE:
2731                         if (optlen != TCPOLEN_SIGNATURE)
2732                                 continue;
2733                         to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2734                         break;
2735 #endif /* TCP_SIGNATURE */
2736                 default:
2737                         continue;
2738                 }
2739         }
2740 }
2741
2742 /*
2743  * Pull out of band byte out of a segment so
2744  * it doesn't appear in the user's data queue.
2745  * It is still reflected in the segment length for
2746  * sequencing purposes.
2747  * "off" is the delayed to be dropped hdrlen.
2748  */
2749 static void
2750 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2751 {
2752         int cnt = off + th->th_urp - 1;
2753
2754         while (cnt >= 0) {
2755                 if (m->m_len > cnt) {
2756                         char *cp = mtod(m, caddr_t) + cnt;
2757                         struct tcpcb *tp = sototcpcb(so);
2758
2759                         tp->t_iobc = *cp;
2760                         tp->t_oobflags |= TCPOOB_HAVEDATA;
2761                         bcopy(cp + 1, cp, m->m_len - cnt - 1);
2762                         m->m_len--;
2763                         if (m->m_flags & M_PKTHDR)
2764                                 m->m_pkthdr.len--;
2765                         return;
2766                 }
2767                 cnt -= m->m_len;
2768                 m = m->m_next;
2769                 if (m == 0)
2770                         break;
2771         }
2772         panic("tcp_pulloutofband");
2773 }
2774
2775 /*
2776  * Collect new round-trip time estimate
2777  * and update averages and current timeout.
2778  */
2779 static void
2780 tcp_xmit_timer(struct tcpcb *tp, int rtt)
2781 {
2782         int delta;
2783
2784         tcpstat.tcps_rttupdated++;
2785         tp->t_rttupdated++;
2786         if (tp->t_srtt != 0) {
2787                 /*
2788                  * srtt is stored as fixed point with 5 bits after the
2789                  * binary point (i.e., scaled by 8).  The following magic
2790                  * is equivalent to the smoothing algorithm in rfc793 with
2791                  * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2792                  * point).  Adjust rtt to origin 0.
2793                  */
2794                 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2795                         - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2796
2797                 if ((tp->t_srtt += delta) <= 0)
2798                         tp->t_srtt = 1;
2799
2800                 /*
2801                  * We accumulate a smoothed rtt variance (actually, a
2802                  * smoothed mean difference), then set the retransmit
2803                  * timer to smoothed rtt + 4 times the smoothed variance.
2804                  * rttvar is stored as fixed point with 4 bits after the
2805                  * binary point (scaled by 16).  The following is
2806                  * equivalent to rfc793 smoothing with an alpha of .75
2807                  * (rttvar = rttvar*3/4 + |delta| / 4).  This replaces
2808                  * rfc793's wired-in beta.
2809                  */
2810                 if (delta < 0)
2811                         delta = -delta;
2812                 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2813                 if ((tp->t_rttvar += delta) <= 0)
2814                         tp->t_rttvar = 1;
2815                 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2816                         tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2817         } else {
2818                 /*
2819                  * No rtt measurement yet - use the unsmoothed rtt.
2820                  * Set the variance to half the rtt (so our first
2821                  * retransmit happens at 3*rtt).
2822                  */
2823                 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2824                 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2825                 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2826         }
2827         tp->t_rtttime = 0;
2828         tp->t_rxtshift = 0;
2829
2830         /*
2831          * the retransmit should happen at rtt + 4 * rttvar.
2832          * Because of the way we do the smoothing, srtt and rttvar
2833          * will each average +1/2 tick of bias.  When we compute
2834          * the retransmit timer, we want 1/2 tick of rounding and
2835          * 1 extra tick because of +-1/2 tick uncertainty in the
2836          * firing of the timer.  The bias will give us exactly the
2837          * 1.5 tick we need.  But, because the bias is
2838          * statistical, we have to test that we don't drop below
2839          * the minimum feasible timer (which is 2 ticks).
2840          */
2841         TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2842                       max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2843
2844         /*
2845          * We received an ack for a packet that wasn't retransmitted;
2846          * it is probably safe to discard any error indications we've
2847          * received recently.  This isn't quite right, but close enough
2848          * for now (a route might have failed after we sent a segment,
2849          * and the return path might not be symmetrical).
2850          */
2851         tp->t_softerror = 0;
2852 }
2853
2854 /*
2855  * Determine a reasonable value for maxseg size.
2856  * If the route is known, check route for mtu.
2857  * If none, use an mss that can be handled on the outgoing
2858  * interface without forcing IP to fragment; if bigger than
2859  * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2860  * to utilize large mbufs.  If no route is found, route has no mtu,
2861  * or the destination isn't local, use a default, hopefully conservative
2862  * size (usually 512 or the default IP max size, but no more than the mtu
2863  * of the interface), as we can't discover anything about intervening
2864  * gateways or networks.  We also initialize the congestion/slow start
2865  * window to be a single segment if the destination isn't local.
2866  * While looking at the routing entry, we also initialize other path-dependent
2867  * parameters from pre-set or cached values in the routing entry.
2868  *
2869  * Also take into account the space needed for options that we
2870  * send regularly.  Make maxseg shorter by that amount to assure
2871  * that we can send maxseg amount of data even when the options
2872  * are present.  Store the upper limit of the length of options plus
2873  * data in maxopd.
2874  *
2875  * NOTE that this routine is only called when we process an incoming
2876  * segment, for outgoing segments only tcp_mssopt is called.
2877  */
2878 void
2879 tcp_mss(struct tcpcb *tp, int offer)
2880 {
2881         struct rtentry *rt;
2882         struct ifnet *ifp;
2883         int rtt, mss;
2884         u_long bufsize;
2885         struct inpcb *inp = tp->t_inpcb;
2886         struct socket *so;
2887 #ifdef INET6
2888         boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2889         size_t min_protoh = isipv6 ?
2890                             sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
2891                             sizeof(struct tcpiphdr);
2892 #else
2893         const boolean_t isipv6 = FALSE;
2894         const size_t min_protoh = sizeof(struct tcpiphdr);
2895 #endif
2896
2897         if (isipv6)
2898                 rt = tcp_rtlookup6(&inp->inp_inc);
2899         else
2900                 rt = tcp_rtlookup(&inp->inp_inc);
2901         if (rt == NULL) {
2902                 tp->t_maxopd = tp->t_maxseg =
2903                     (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
2904                 return;
2905         }
2906         ifp = rt->rt_ifp;
2907         so = inp->inp_socket;
2908
2909         /*
2910          * Offer == 0 means that there was no MSS on the SYN segment,
2911          * in this case we use either the interface mtu or tcp_mssdflt.
2912          *
2913          * An offer which is too large will be cut down later.
2914          */
2915         if (offer == 0) {
2916                 if (isipv6) {
2917                         if (in6_localaddr(&inp->in6p_faddr)) {
2918                                 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
2919                                         min_protoh;
2920                         } else {
2921                                 offer = tcp_v6mssdflt;
2922                         }
2923                 } else {
2924                         if (in_localaddr(inp->inp_faddr))
2925                                 offer = ifp->if_mtu - min_protoh;
2926                         else
2927                                 offer = tcp_mssdflt;
2928                 }
2929         }
2930
2931         /*
2932          * Prevent DoS attack with too small MSS. Round up
2933          * to at least minmss.
2934          *
2935          * Sanity check: make sure that maxopd will be large
2936          * enough to allow some data on segments even is the
2937          * all the option space is used (40bytes).  Otherwise
2938          * funny things may happen in tcp_output.
2939          */
2940         offer = max(offer, tcp_minmss);
2941         offer = max(offer, 64);
2942
2943         rt->rt_rmx.rmx_mssopt = offer;
2944
2945         /*
2946          * While we're here, check if there's an initial rtt
2947          * or rttvar.  Convert from the route-table units
2948          * to scaled multiples of the slow timeout timer.
2949          */
2950         if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
2951                 /*
2952                  * XXX the lock bit for RTT indicates that the value
2953                  * is also a minimum value; this is subject to time.
2954                  */
2955                 if (rt->rt_rmx.rmx_locks & RTV_RTT)
2956                         tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
2957                 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
2958                 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
2959                 tcpstat.tcps_usedrtt++;
2960                 if (rt->rt_rmx.rmx_rttvar) {
2961                         tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
2962                             (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
2963                         tcpstat.tcps_usedrttvar++;
2964                 } else {
2965                         /* default variation is +- 1 rtt */
2966                         tp->t_rttvar =
2967                             tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
2968                 }
2969                 TCPT_RANGESET(tp->t_rxtcur,
2970                               ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
2971                               tp->t_rttmin, TCPTV_REXMTMAX);
2972         }
2973
2974         /*
2975          * if there's an mtu associated with the route, use it
2976          * else, use the link mtu.  Take the smaller of mss or offer
2977          * as our final mss.
2978          */
2979         if (rt->rt_rmx.rmx_mtu) {
2980                 mss = rt->rt_rmx.rmx_mtu - min_protoh;
2981         } else {
2982                 if (isipv6)
2983                         mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
2984                 else
2985                         mss = ifp->if_mtu - min_protoh;
2986         }
2987         mss = min(mss, offer);
2988
2989         /*
2990          * maxopd stores the maximum length of data AND options
2991          * in a segment; maxseg is the amount of data in a normal
2992          * segment.  We need to store this value (maxopd) apart
2993          * from maxseg, because now every segment carries options
2994          * and thus we normally have somewhat less data in segments.
2995          */
2996         tp->t_maxopd = mss;
2997
2998         if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
2999             ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3000                 mss -= TCPOLEN_TSTAMP_APPA;
3001
3002 #if     (MCLBYTES & (MCLBYTES - 1)) == 0
3003                 if (mss > MCLBYTES)
3004                         mss &= ~(MCLBYTES-1);
3005 #else
3006                 if (mss > MCLBYTES)
3007                         mss = mss / MCLBYTES * MCLBYTES;
3008 #endif
3009         /*
3010          * If there's a pipesize, change the socket buffer
3011          * to that size.  Make the socket buffers an integral
3012          * number of mss units; if the mss is larger than
3013          * the socket buffer, decrease the mss.
3014          */
3015 #ifdef RTV_SPIPE
3016         if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3017 #endif
3018                 bufsize = so->so_snd.ssb_hiwat;
3019         if (bufsize < mss)
3020                 mss = bufsize;
3021         else {
3022                 bufsize = roundup(bufsize, mss);
3023                 if (bufsize > sb_max)
3024                         bufsize = sb_max;
3025                 if (bufsize > so->so_snd.ssb_hiwat)
3026                         ssb_reserve(&so->so_snd, bufsize, so, NULL);
3027         }
3028         tp->t_maxseg = mss;
3029
3030 #ifdef RTV_RPIPE
3031         if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3032 #endif
3033                 bufsize = so->so_rcv.ssb_hiwat;
3034         if (bufsize > mss) {
3035                 bufsize = roundup(bufsize, mss);
3036                 if (bufsize > sb_max)
3037                         bufsize = sb_max;
3038                 if (bufsize > so->so_rcv.ssb_hiwat) {
3039                         lwkt_gettoken(&so->so_rcv.ssb_token);
3040                         ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3041                         lwkt_reltoken(&so->so_rcv.ssb_token);
3042                 }
3043         }
3044
3045         /*
3046          * Set the slow-start flight size depending on whether this
3047          * is a local network or not.
3048          */
3049         if (tcp_do_rfc3390)
3050                 tp->snd_cwnd = min(4 * mss, max(2 * mss, 4380));
3051         else
3052                 tp->snd_cwnd = mss;
3053
3054         if (rt->rt_rmx.rmx_ssthresh) {
3055                 /*
3056                  * There's some sort of gateway or interface
3057                  * buffer limit on the path.  Use this to set
3058                  * the slow start threshhold, but set the
3059                  * threshold to no less than 2*mss.
3060                  */
3061                 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3062                 tcpstat.tcps_usedssthresh++;
3063         }
3064 }
3065
3066 /*
3067  * Determine the MSS option to send on an outgoing SYN.
3068  */
3069 int
3070 tcp_mssopt(struct tcpcb *tp)
3071 {
3072         struct rtentry *rt;
3073 #ifdef INET6
3074         boolean_t isipv6 =
3075             ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3076         int min_protoh = isipv6 ?
3077                              sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3078                              sizeof(struct tcpiphdr);
3079 #else
3080         const boolean_t isipv6 = FALSE;
3081         const size_t min_protoh = sizeof(struct tcpiphdr);
3082 #endif
3083
3084         if (isipv6)
3085                 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3086         else
3087                 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3088         if (rt == NULL)
3089                 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3090
3091         return (rt->rt_ifp->if_mtu - min_protoh);
3092 }
3093
3094 /*
3095  * When a partial ack arrives, force the retransmission of the
3096  * next unacknowledged segment.  Do not exit Fast Recovery.
3097  *
3098  * Implement the Slow-but-Steady variant of NewReno by restarting the
3099  * the retransmission timer.  Turn it off here so it can be restarted
3100  * later in tcp_output().
3101  */
3102 static void
3103 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3104 {
3105         tcp_seq old_snd_nxt = tp->snd_nxt;
3106         u_long ocwnd = tp->snd_cwnd;
3107
3108         tcp_callout_stop(tp, tp->tt_rexmt);
3109         tp->t_rtttime = 0;
3110         tp->snd_nxt = th->th_ack;
3111         /* Set snd_cwnd to one segment beyond acknowledged offset. */
3112         tp->snd_cwnd = tp->t_maxseg;
3113         tp->t_flags |= TF_ACKNOW;
3114         tcp_output(tp);
3115         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3116                 tp->snd_nxt = old_snd_nxt;
3117         /* partial window deflation */
3118         if (ocwnd > acked)
3119                 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3120         else
3121                 tp->snd_cwnd = tp->t_maxseg;
3122 }
3123
3124 /*
3125  * In contrast to the Slow-but-Steady NewReno variant,
3126  * we do not reset the retransmission timer for SACK retransmissions,
3127  * except when retransmitting snd_una.
3128  */
3129 static void
3130 tcp_sack_rexmt(struct tcpcb *tp, struct tcphdr *th)
3131 {
3132         uint32_t pipe, seglen;
3133         tcp_seq nextrexmt;
3134         boolean_t lostdup;
3135         tcp_seq old_snd_nxt = tp->snd_nxt;
3136         u_long ocwnd = tp->snd_cwnd;
3137         int nseg = 0;           /* consecutive new segments */
3138 #define MAXBURST 4              /* limit burst of new packets on partial ack */
3139
3140         tp->t_rtttime = 0;
3141         pipe = tcp_sack_compute_pipe(tp);
3142         while ((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg &&
3143             (!tcp_do_smartsack || nseg < MAXBURST) &&
3144             tcp_sack_nextseg(tp, &nextrexmt, &seglen, &lostdup)) {
3145                 uint32_t sent;
3146                 tcp_seq old_snd_max;
3147                 int error;
3148
3149                 if (nextrexmt == tp->snd_max)
3150                         ++nseg;
3151                 tp->snd_nxt = nextrexmt;
3152                 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3153                 old_snd_max = tp->snd_max;
3154                 if (nextrexmt == tp->snd_una)
3155                         tcp_callout_stop(tp, tp->tt_rexmt);
3156                 error = tcp_output(tp);
3157                 if (error != 0)
3158                         break;
3159                 sent = tp->snd_nxt - nextrexmt;
3160                 if (sent <= 0)
3161                         break;
3162                 if (!lostdup)
3163                         pipe += sent;
3164                 tcpstat.tcps_sndsackpack++;
3165                 tcpstat.tcps_sndsackbyte += sent;
3166                 if (SEQ_LT(nextrexmt, old_snd_max) &&
3167                     SEQ_LT(tp->rexmt_high, tp->snd_nxt))
3168                         tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3169         }
3170         if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3171                 tp->snd_nxt = old_snd_nxt;
3172         tp->snd_cwnd = ocwnd;
3173 }
3174
3175 /*
3176  * Reset idle time and keep-alive timer, typically called when a valid
3177  * tcp packet is received but may also be called when FASTKEEP is set
3178  * to prevent the previous long-timeout from calculating to a drop.
3179  *
3180  * Only update t_rcvtime for non-SYN packets.
3181  *
3182  * Handle the case where one side thinks the connection is established
3183  * but the other side has, say, rebooted without cleaning out the
3184  * connection.   The SYNs could be construed as an attack and wind
3185  * up ignored, but in case it isn't an attack we can validate the
3186  * connection by forcing a keepalive.
3187  */
3188 void
3189 tcp_timer_keep_activity(struct tcpcb *tp, int thflags)
3190 {
3191         if (TCPS_HAVEESTABLISHED(tp->t_state)) {
3192                 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
3193                         tp->t_flags |= TF_KEEPALIVE;
3194                         tcp_callout_reset(tp, tp->tt_keep, hz / 2,
3195                                           tcp_timer_keep);
3196                 } else {
3197                         tp->t_rcvtime = ticks;
3198                         tp->t_flags &= ~TF_KEEPALIVE;
3199                         tcp_callout_reset(tp, tp->tt_keep,
3200                                           tcp_getkeepidle(tp),
3201                                           tcp_timer_keep);
3202                 }
3203         }
3204 }