2 * Copyright (c) 2002, 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2002, 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
36 * The Regents of the University of California. All rights reserved.
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
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. Neither the name of the University nor the names of its contributors
47 * may be used to endorse or promote products derived from this software
48 * without specific prior written permission.
50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
62 * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
63 * $FreeBSD: src/sys/netinet/tcp_input.c,v 1.107.2.38 2003/05/21 04:46:41 cjc Exp $
67 #include "opt_inet6.h"
68 #include "opt_ipsec.h"
69 #include "opt_tcpdebug.h"
70 #include "opt_tcp_input.h"
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/kernel.h>
75 #include <sys/sysctl.h>
76 #include <sys/malloc.h>
78 #include <sys/proc.h> /* for proc0 declaration */
79 #include <sys/protosw.h>
80 #include <sys/socket.h>
81 #include <sys/socketvar.h>
82 #include <sys/syslog.h>
83 #include <sys/in_cksum.h>
85 #include <sys/socketvar2.h>
87 #include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */
88 #include <machine/stdarg.h>
91 #include <net/route.h>
93 #include <netinet/in.h>
94 #include <netinet/in_systm.h>
95 #include <netinet/ip.h>
96 #include <netinet/ip_icmp.h> /* for ICMP_BANDLIM */
97 #include <netinet/in_var.h>
98 #include <netinet/icmp_var.h> /* for ICMP_BANDLIM */
99 #include <netinet/in_pcb.h>
100 #include <netinet/ip_var.h>
101 #include <netinet/ip6.h>
102 #include <netinet/icmp6.h>
103 #include <netinet6/nd6.h>
104 #include <netinet6/ip6_var.h>
105 #include <netinet6/in6_pcb.h>
106 #include <netinet/tcp.h>
107 #include <netinet/tcp_fsm.h>
108 #include <netinet/tcp_seq.h>
109 #include <netinet/tcp_timer.h>
110 #include <netinet/tcp_timer2.h>
111 #include <netinet/tcp_var.h>
112 #include <netinet6/tcp6_var.h>
113 #include <netinet/tcpip.h>
116 #include <netinet/tcp_debug.h>
118 u_char tcp_saveipgen[40]; /* the size must be of max ip header, now IPv6 */
119 struct tcphdr tcp_savetcp;
123 #include <netproto/ipsec/ipsec.h>
124 #include <netproto/ipsec/ipsec6.h>
128 #include <netinet6/ipsec.h>
129 #include <netinet6/ipsec6.h>
130 #include <netproto/key/key.h>
134 * Limit burst of new packets during SACK based fast recovery
135 * or extended limited transmit.
137 #define TCP_SACK_MAXBURST 4
139 MALLOC_DEFINE(M_TSEGQ, "tseg_qent", "TCP segment queue entry");
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");
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");
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");
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");
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)");
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");
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");
172 static int tcp_do_eifel_detect = 1;
173 SYSCTL_INT(_net_inet_tcp, OID_AUTO, eifel, CTLFLAG_RW,
174 &tcp_do_eifel_detect, 0, "Eifel detection algorithm (RFC 3522)");
176 static int tcp_do_abc = 1;
177 SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc, CTLFLAG_RW,
179 "TCP Appropriate Byte Counting (RFC 3465)");
182 * The following value actually takes range [25ms, 250ms],
183 * given that most modern systems use 1ms ~ 10ms as the unit
184 * of timestamp option.
186 static u_int tcp_paws_tolerance = 25;
187 SYSCTL_UINT(_net_inet_tcp, OID_AUTO, paws_tolerance, CTLFLAG_RW,
188 &tcp_paws_tolerance, 0, "RFC1323 PAWS tolerance");
191 * Define as tunable for easy testing with SACK on and off.
192 * Warning: do not change setting in the middle of an existing active TCP flow,
193 * else strange things might happen to that flow.
196 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW,
197 &tcp_do_sack, 0, "Enable SACK Algorithms");
199 int tcp_do_smartsack = 1;
200 SYSCTL_INT(_net_inet_tcp, OID_AUTO, smartsack, CTLFLAG_RW,
201 &tcp_do_smartsack, 0, "Enable Smart SACK Algorithms");
203 int tcp_do_rescuesack = 1;
204 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack, CTLFLAG_RW,
205 &tcp_do_rescuesack, 0, "Rescue retransmission for SACK");
207 int tcp_aggressive_rescuesack = 0;
208 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rescuesack_agg, CTLFLAG_RW,
209 &tcp_aggressive_rescuesack, 0, "Aggressive rescue retransmission for SACK");
211 static int tcp_force_sackrxt = 1;
212 SYSCTL_INT(_net_inet_tcp, OID_AUTO, force_sackrxt, CTLFLAG_RW,
213 &tcp_force_sackrxt, 0, "Allowed forced SACK retransmit burst");
215 int tcp_do_rfc6675 = 1;
216 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675, CTLFLAG_RW,
217 &tcp_do_rfc6675, 0, "Enable RFC6675");
219 int tcp_rfc6675_rxt = 0;
220 SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_rxt, CTLFLAG_RW,
221 &tcp_rfc6675_rxt, 0, "Enable RFC6675 retransmit");
223 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
224 "TCP Segment Reassembly Queue");
226 int tcp_reass_maxseg = 0;
227 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RD,
228 &tcp_reass_maxseg, 0,
229 "Global maximum number of TCP Segments in Reassembly Queue");
231 int tcp_reass_qsize = 0;
232 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
234 "Global number of TCP Segments currently in Reassembly Queue");
236 static int tcp_reass_overflows = 0;
237 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
238 &tcp_reass_overflows, 0,
239 "Global number of TCP Segment Reassembly Queue Overflows");
241 int tcp_do_autorcvbuf = 1;
242 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_RW,
243 &tcp_do_autorcvbuf, 0, "Enable automatic receive buffer sizing");
245 int tcp_autorcvbuf_inc = 16*1024;
246 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_inc, CTLFLAG_RW,
247 &tcp_autorcvbuf_inc, 0,
248 "Incrementor step size of automatic receive buffer");
250 int tcp_autorcvbuf_max = 2*1024*1024;
251 SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_RW,
252 &tcp_autorcvbuf_max, 0, "Max size of automatic receive buffer");
254 int tcp_sosend_agglim = 2;
255 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_agglim, CTLFLAG_RW,
256 &tcp_sosend_agglim, 0, "TCP sosend mbuf aggregation limit");
258 int tcp_sosend_async = 1;
259 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_async, CTLFLAG_RW,
260 &tcp_sosend_async, 0, "TCP asynchronized pru_send");
262 int tcp_sosend_jcluster = 1;
263 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sosend_jcluster, CTLFLAG_RW,
264 &tcp_sosend_jcluster, 0, "TCP output uses jcluster");
266 static int tcp_ignore_redun_dsack = 1;
267 SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_redun_dsack, CTLFLAG_RW,
268 &tcp_ignore_redun_dsack, 0, "Ignore redundant DSACK");
270 static int tcp_reuseport_ext = 1;
271 SYSCTL_INT(_net_inet_tcp, OID_AUTO, reuseport_ext, CTLFLAG_RW,
272 &tcp_reuseport_ext, 0, "SO_REUSEPORT extension");
274 static void tcp_dooptions(struct tcpopt *, u_char *, int, boolean_t,
276 static void tcp_pulloutofband(struct socket *,
277 struct tcphdr *, struct mbuf *, int);
278 static int tcp_reass(struct tcpcb *, struct tcphdr *, int *,
280 static void tcp_xmit_timer(struct tcpcb *, int, tcp_seq);
281 static void tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *, int);
282 static void tcp_sack_rexmt(struct tcpcb *, boolean_t);
283 static boolean_t tcp_sack_limitedxmit(struct tcpcb *);
284 static int tcp_rmx_msl(const struct tcpcb *);
285 static void tcp_established(struct tcpcb *);
286 static boolean_t tcp_recv_dupack(struct tcpcb *, tcp_seq, u_int);
288 /* Neighbor Discovery, Neighbor Unreachability Detection Upper layer hint. */
290 #define ND6_HINT(tp) \
292 if ((tp) && (tp)->t_inpcb && \
293 ((tp)->t_inpcb->inp_vflag & INP_IPV6) && \
294 (tp)->t_inpcb->in6p_route.ro_rt) \
295 nd6_nud_hint((tp)->t_inpcb->in6p_route.ro_rt, NULL, 0); \
302 * Indicate whether this ack should be delayed. We can delay the ack if
303 * - delayed acks are enabled and
304 * - there is no delayed ack timer in progress and
305 * - our last ack wasn't a 0-sized window. We never want to delay
306 * the ack that opens up a 0-sized window.
308 #define DELAY_ACK(tp) \
309 (tcp_delack_enabled && !tcp_callout_pending(tp, tp->tt_delack) && \
310 !(tp->t_flags & TF_RXWIN0SENT))
312 #define acceptable_window_update(tp, th, tiwin) \
313 (SEQ_LT(tp->snd_wl1, th->th_seq) || \
314 (tp->snd_wl1 == th->th_seq && \
315 (SEQ_LT(tp->snd_wl2, th->th_ack) || \
316 (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))
318 #define iceildiv(n, d) (((n)+(d)-1) / (d))
319 #define need_early_retransmit(tp, ownd) \
320 (tcp_do_early_retransmit && \
321 (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) && \
322 ownd < ((tp->t_rxtthresh + 1) * tp->t_maxseg) && \
323 tp->t_dupacks + 1 >= iceildiv(ownd, tp->t_maxseg) && \
324 (!TCP_DO_SACK(tp) || ownd <= tp->t_maxseg || \
325 tcp_sack_has_sacked(&tp->scb, ownd - tp->t_maxseg)))
328 * Returns TRUE, if this segment can be merged with the last
329 * pending segment in the reassemble queue and this segment
330 * does not overlap with the pending segment immediately
331 * preceeding the last pending segment.
333 static __inline boolean_t
334 tcp_paws_canreasslast(const struct tcpcb *tp, const struct tcphdr *th, int tlen)
336 const struct tseg_qent *last, *prev;
338 last = TAILQ_LAST(&tp->t_segq, tsegqe_head);
342 /* This segment comes immediately after the last pending segment */
343 if (last->tqe_th->th_seq + last->tqe_len == th->th_seq) {
344 if (last->tqe_th->th_flags & TH_FIN) {
345 /* No segments should follow segment w/ FIN */
351 if (th->th_seq + tlen != last->tqe_th->th_seq)
353 /* This segment comes immediately before the last pending segment */
355 prev = TAILQ_PREV(last, tsegqe_head, tqe_q);
358 * No pending preceeding segment, we assume this segment
359 * could be reassembled.
364 /* This segment does not overlap with the preceeding segment */
365 if (SEQ_GEQ(th->th_seq, prev->tqe_th->th_seq + prev->tqe_len))
372 tcp_ncr_update_rxtthresh(struct tcpcb *tp)
374 int old_rxtthresh = tp->t_rxtthresh;
375 uint32_t ownd = tp->snd_max - tp->snd_una;
377 tp->t_rxtthresh = max(tcprexmtthresh, ((ownd / tp->t_maxseg) >> 1));
378 if (tp->t_rxtthresh != old_rxtthresh) {
379 tcp_sack_update_lostseq(&tp->scb, tp->snd_una,
380 tp->t_maxseg, tp->t_rxtthresh);
385 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
388 struct tseg_qent *p = NULL;
389 struct tseg_qent *te;
390 struct socket *so = tp->t_inpcb->inp_socket;
394 * Call with th == NULL after become established to
395 * force pre-ESTABLISHED data up to user socket.
401 * Limit the number of segments in the reassembly queue to prevent
402 * holding on to too many segments (and thus running out of mbufs).
403 * Make sure to let the missing segment through which caused this
404 * queue. Always keep one global queue entry spare to be able to
405 * process the missing segment.
407 if (th->th_seq != tp->rcv_nxt &&
408 tcp_reass_qsize + 1 >= tcp_reass_maxseg) {
409 tcp_reass_overflows++;
410 tcpstat.tcps_rcvmemdrop++;
412 /* no SACK block to report */
413 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
417 /* Allocate a new queue entry. */
418 te = kmalloc(sizeof(struct tseg_qent), M_TSEGQ, M_INTWAIT | M_NULLOK);
420 tcpstat.tcps_rcvmemdrop++;
422 /* no SACK block to report */
423 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
426 atomic_add_int(&tcp_reass_qsize, 1);
428 if (th->th_flags & TH_FIN)
429 tp->t_flags |= TF_QUEDFIN;
432 * Find a segment which begins after this one does.
434 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
435 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
441 * If there is a preceding segment, it may provide some of
442 * our data already. If so, drop the data from the incoming
443 * segment. If it provides all of our data, drop us.
448 /* conversion to int (in i) handles seq wraparound */
449 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
450 if (i > 0) { /* overlaps preceding segment */
452 (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
453 /* enclosing block starts w/ preceding segment */
454 tp->encloseblk.rblk_start = p->tqe_th->th_seq;
456 if (th->th_flags & TH_FIN)
457 p->tqe_th->th_flags |= TH_FIN;
459 /* preceding encloses incoming segment */
460 tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
461 p->tqe_th->th_seq + p->tqe_len,
462 p->tqe_th->th_flags);
463 tcpstat.tcps_rcvduppack++;
464 tcpstat.tcps_rcvdupbyte += *tlenp;
467 atomic_add_int(&tcp_reass_qsize, -1);
469 * Try to present any queued data
470 * at the left window edge to the user.
471 * This is needed after the 3-WHS
474 goto present; /* ??? */
479 /* incoming segment end is enclosing block end */
480 tp->encloseblk.rblk_end = TCP_SACK_BLKEND(
481 th->th_seq + *tlenp, th->th_flags);
482 /* trim end of reported D-SACK block */
483 tp->reportblk.rblk_end = th->th_seq;
486 tcpstat.tcps_rcvoopack++;
487 tcpstat.tcps_rcvoobyte += *tlenp;
490 * While we overlap succeeding segments trim them or,
491 * if they are completely covered, dequeue them.
494 tcp_seq_diff_t i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
495 tcp_seq qend = q->tqe_th->th_seq + q->tqe_len;
496 tcp_seq qend_sack = TCP_SACK_BLKEND(qend, q->tqe_th->th_flags);
497 struct tseg_qent *nq;
501 if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
502 /* first time through */
503 tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG);
504 tp->encloseblk = tp->reportblk;
505 /* report trailing duplicate D-SACK segment */
506 tp->reportblk.rblk_start = q->tqe_th->th_seq;
508 if ((tp->sack_flags & TSACK_F_ENCLOSESEG) &&
509 SEQ_GT(qend_sack, tp->encloseblk.rblk_end)) {
510 /* extend enclosing block if one exists */
511 tp->encloseblk.rblk_end = qend_sack;
513 if (i < q->tqe_len) {
514 q->tqe_th->th_seq += i;
520 if (q->tqe_th->th_flags & TH_FIN)
521 th->th_flags |= TH_FIN;
523 nq = TAILQ_NEXT(q, tqe_q);
524 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
527 atomic_add_int(&tcp_reass_qsize, -1);
531 /* Insert the new segment queue entry into place. */
534 te->tqe_len = *tlenp;
536 /* check if can coalesce with following segment */
537 if (q != NULL && (th->th_seq + *tlenp == q->tqe_th->th_seq)) {
540 te->tqe_len += q->tqe_len;
541 if (q->tqe_th->th_flags & TH_FIN)
542 te->tqe_th->th_flags |= TH_FIN;
543 tend_sack = TCP_SACK_BLKEND(te->tqe_th->th_seq + te->tqe_len,
544 te->tqe_th->th_flags);
546 m_cat(te->tqe_m, q->tqe_m);
547 tp->encloseblk.rblk_end = tend_sack;
549 * When not reporting a duplicate segment, use
550 * the larger enclosing block as the SACK block.
552 if (!(tp->sack_flags & TSACK_F_DUPSEG))
553 tp->reportblk.rblk_end = tend_sack;
554 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
556 atomic_add_int(&tcp_reass_qsize, -1);
560 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
562 /* check if can coalesce with preceding segment */
563 if (p->tqe_th->th_seq + p->tqe_len == th->th_seq) {
564 if (te->tqe_th->th_flags & TH_FIN)
565 p->tqe_th->th_flags |= TH_FIN;
566 p->tqe_len += te->tqe_len;
567 m_cat(p->tqe_m, te->tqe_m);
568 tp->encloseblk.rblk_start = p->tqe_th->th_seq;
570 * When not reporting a duplicate segment, use
571 * the larger enclosing block as the SACK block.
573 if (!(tp->sack_flags & TSACK_F_DUPSEG))
574 tp->reportblk.rblk_start = p->tqe_th->th_seq;
576 atomic_add_int(&tcp_reass_qsize, -1);
578 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
584 * Present data to user, advancing rcv_nxt through
585 * completed sequence space.
587 if (!TCPS_HAVEESTABLISHED(tp->t_state))
589 q = TAILQ_FIRST(&tp->t_segq);
590 if (q == NULL || q->tqe_th->th_seq != tp->rcv_nxt)
592 tp->rcv_nxt += q->tqe_len;
593 if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
594 /* no SACK block to report since ACK advanced */
595 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
597 /* no enclosing block to report since ACK advanced */
598 tp->sack_flags &= ~TSACK_F_ENCLOSESEG;
599 flags = q->tqe_th->th_flags & TH_FIN;
600 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
601 KASSERT(TAILQ_EMPTY(&tp->t_segq) ||
602 TAILQ_FIRST(&tp->t_segq)->tqe_th->th_seq != tp->rcv_nxt,
603 ("segment not coalesced"));
604 if (so->so_state & SS_CANTRCVMORE) {
607 lwkt_gettoken(&so->so_rcv.ssb_token);
608 ssb_appendstream(&so->so_rcv, q->tqe_m);
609 lwkt_reltoken(&so->so_rcv.ssb_token);
612 atomic_add_int(&tcp_reass_qsize, -1);
619 * TCP input routine, follows pages 65-76 of the
620 * protocol specification dated September, 1981 very closely.
624 tcp6_input(struct mbuf **mp, int *offp, int proto)
626 struct mbuf *m = *mp;
627 struct in6_ifaddr *ia6;
629 IP6_EXTHDR_CHECK(m, *offp, sizeof(struct tcphdr), IPPROTO_DONE);
632 * draft-itojun-ipv6-tcp-to-anycast
633 * better place to put this in?
635 ia6 = ip6_getdstifaddr(m);
636 if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) {
637 icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR,
638 offsetof(struct ip6_hdr, ip6_dst));
639 return (IPPROTO_DONE);
642 tcp_input(mp, offp, proto);
643 return (IPPROTO_DONE);
648 tcp_input(struct mbuf **mp, int *offp, int proto)
652 struct ip *ip = NULL;
654 struct inpcb *inp = NULL;
660 struct tcpcb *tp = NULL;
662 struct socket *so = NULL;
664 boolean_t ourfinisacked, needoutput = FALSE, delayed_dupack = FALSE;
665 tcp_seq th_dupack = 0; /* XXX gcc warning */
666 u_int to_flags = 0; /* XXX gcc warning */
669 struct tcpopt to; /* options in this segment */
670 struct sockaddr_in *next_hop = NULL;
671 int rstreason; /* For badport_bandlim accounting purposes */
673 struct ip6_hdr *ip6 = NULL;
678 const boolean_t isipv6 = FALSE;
688 tcpstat.tcps_rcvtotal++;
690 if (m->m_pkthdr.fw_flags & IPFORWARD_MBUF_TAGGED) {
693 mtag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL);
694 KKASSERT(mtag != NULL);
695 next_hop = m_tag_data(mtag);
699 isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? TRUE : FALSE;
703 /* IP6_EXTHDR_CHECK() is already done at tcp6_input() */
704 ip6 = mtod(m, struct ip6_hdr *);
705 tlen = (sizeof *ip6) + ntohs(ip6->ip6_plen) - off0;
706 if (in6_cksum(m, IPPROTO_TCP, off0, tlen)) {
707 tcpstat.tcps_rcvbadsum++;
710 th = (struct tcphdr *)((caddr_t)ip6 + off0);
713 * Be proactive about unspecified IPv6 address in source.
714 * As we use all-zero to indicate unbounded/unconnected pcb,
715 * unspecified IPv6 address can be used to confuse us.
717 * Note that packets with unspecified IPv6 destination is
718 * already dropped in ip6_input.
720 if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) {
726 * Get IP and TCP header together in first mbuf.
727 * Note: IP leaves IP header in first mbuf.
729 if (off0 > sizeof(struct ip)) {
731 off0 = sizeof(struct ip);
733 /* already checked and pulled up in ip_demux() */
734 KASSERT(m->m_len >= sizeof(struct tcpiphdr),
735 ("TCP header not in one mbuf: m->m_len %d", m->m_len));
736 ip = mtod(m, struct ip *);
737 ipov = (struct ipovly *)ip;
738 th = (struct tcphdr *)((caddr_t)ip + off0);
741 if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) {
742 if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR)
743 th->th_sum = m->m_pkthdr.csum_data;
745 th->th_sum = in_pseudo(ip->ip_src.s_addr,
747 htonl(m->m_pkthdr.csum_data +
750 th->th_sum ^= 0xffff;
753 * Checksum extended TCP header and data.
755 len = sizeof(struct ip) + tlen;
756 bzero(ipov->ih_x1, sizeof ipov->ih_x1);
757 ipov->ih_len = (u_short)tlen;
758 ipov->ih_len = htons(ipov->ih_len);
759 th->th_sum = in_cksum(m, len);
762 tcpstat.tcps_rcvbadsum++;
766 /* Re-initialization for later version check */
767 ip->ip_v = IPVERSION;
772 * Check that TCP offset makes sense,
773 * pull out TCP options and adjust length. XXX
775 off = th->th_off << 2;
776 /* already checked and pulled up in ip_demux() */
777 KASSERT(off >= sizeof(struct tcphdr) && off <= tlen,
778 ("bad TCP data offset %d (tlen %d)", off, tlen));
779 tlen -= off; /* tlen is used instead of ti->ti_len */
780 if (off > sizeof(struct tcphdr)) {
782 IP6_EXTHDR_CHECK(m, off0, off, IPPROTO_DONE);
783 ip6 = mtod(m, struct ip6_hdr *);
784 th = (struct tcphdr *)((caddr_t)ip6 + off0);
786 /* already pulled up in ip_demux() */
787 KASSERT(m->m_len >= sizeof(struct ip) + off,
788 ("TCP header and options not in one mbuf: "
789 "m_len %d, off %d", m->m_len, off));
791 optlen = off - sizeof(struct tcphdr);
792 optp = (u_char *)(th + 1);
794 thflags = th->th_flags;
796 #ifdef TCP_DROP_SYNFIN
798 * If the drop_synfin option is enabled, drop all packets with
799 * both the SYN and FIN bits set. This prevents e.g. nmap from
800 * identifying the TCP/IP stack.
802 * This is a violation of the TCP specification.
804 if (drop_synfin && (thflags & (TH_SYN | TH_FIN)) == (TH_SYN | TH_FIN))
809 * Convert TCP protocol specific fields to host format.
811 th->th_seq = ntohl(th->th_seq);
812 th->th_ack = ntohl(th->th_ack);
813 th->th_win = ntohs(th->th_win);
814 th->th_urp = ntohs(th->th_urp);
817 * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options,
818 * until after ip6_savecontrol() is called and before other functions
819 * which don't want those proto headers.
820 * Because ip6_savecontrol() is going to parse the mbuf to
821 * search for data to be passed up to user-land, it wants mbuf
822 * parameters to be unchanged.
823 * XXX: the call of ip6_savecontrol() has been obsoleted based on
824 * latest version of the advanced API (20020110).
826 drop_hdrlen = off0 + off;
829 * Locate pcb for segment.
832 /* IPFIREWALL_FORWARD section */
833 if (next_hop != NULL && !isipv6) { /* IPv6 support is not there yet */
835 * Transparently forwarded. Pretend to be the destination.
836 * already got one like this?
838 cpu = mycpu->gd_cpuid;
839 inp = in_pcblookup_hash(&tcbinfo[cpu],
840 ip->ip_src, th->th_sport,
841 ip->ip_dst, th->th_dport,
842 0, m->m_pkthdr.rcvif);
845 * It's new. Try to find the ambushing socket.
849 * The rest of the ipfw code stores the port in
851 * (The IP address is still in network order.)
853 in_port_t dport = next_hop->sin_port ?
854 htons(next_hop->sin_port) :
857 cpu = tcp_addrcpu(ip->ip_src.s_addr, th->th_sport,
858 next_hop->sin_addr.s_addr, dport);
859 inp = in_pcblookup_hash(&tcbinfo[cpu],
860 ip->ip_src, th->th_sport,
861 next_hop->sin_addr, dport,
862 1, m->m_pkthdr.rcvif);
866 inp = in6_pcblookup_hash(&tcbinfo[0],
867 &ip6->ip6_src, th->th_sport,
868 &ip6->ip6_dst, th->th_dport,
869 1, m->m_pkthdr.rcvif);
871 cpu = mycpu->gd_cpuid;
872 inp = in_pcblookup_pkthash(&tcbinfo[cpu],
873 ip->ip_src, th->th_sport,
874 ip->ip_dst, th->th_dport,
875 1, m->m_pkthdr.rcvif,
876 tcp_reuseport_ext ? m : NULL);
881 * If the state is CLOSED (i.e., TCB does not exist) then
882 * all data in the incoming segment is discarded.
883 * If the TCB exists but is in CLOSED state, it is embryonic,
884 * but should either do a listen or a connect soon.
889 char dbuf[INET6_ADDRSTRLEN+2], sbuf[INET6_ADDRSTRLEN+2];
891 char dbuf[sizeof "aaa.bbb.ccc.ddd"];
892 char sbuf[sizeof "aaa.bbb.ccc.ddd"];
896 strcat(dbuf, ip6_sprintf(&ip6->ip6_dst));
899 strcat(sbuf, ip6_sprintf(&ip6->ip6_src));
902 strcpy(dbuf, inet_ntoa(ip->ip_dst));
903 strcpy(sbuf, inet_ntoa(ip->ip_src));
905 switch (log_in_vain) {
907 if (!(thflags & TH_SYN))
911 "Connection attempt to TCP %s:%d "
912 "from %s:%d flags:0x%02x\n",
913 dbuf, ntohs(th->th_dport), sbuf,
914 ntohs(th->th_sport), thflags);
923 if (thflags & TH_SYN)
932 rstreason = BANDLIM_RST_CLOSEDPORT;
938 if (ipsec6_in_reject_so(m, inp->inp_socket)) {
939 ipsec6stat.in_polvio++;
943 if (ipsec4_in_reject_so(m, inp->inp_socket)) {
944 ipsecstat.in_polvio++;
951 if (ipsec6_in_reject(m, inp))
954 if (ipsec4_in_reject(m, inp))
958 /* Check the minimum TTL for socket. */
960 if ((isipv6 ? ip6->ip6_hlim : ip->ip_ttl) < inp->inp_ip_minttl)
966 rstreason = BANDLIM_RST_CLOSEDPORT;
969 if (tp->t_state <= TCPS_CLOSED)
972 so = inp->inp_socket;
975 if (so->so_options & SO_DEBUG) {
976 ostate = tp->t_state;
978 bcopy(ip6, tcp_saveipgen, sizeof(*ip6));
980 bcopy(ip, tcp_saveipgen, sizeof(*ip));
985 bzero(&to, sizeof to);
987 if (so->so_options & SO_ACCEPTCONN) {
988 struct in_conninfo inc;
991 inc.inc_isipv6 = (isipv6 == TRUE);
994 inc.inc6_faddr = ip6->ip6_src;
995 inc.inc6_laddr = ip6->ip6_dst;
996 inc.inc6_route.ro_rt = NULL; /* XXX */
998 inc.inc_faddr = ip->ip_src;
999 inc.inc_laddr = ip->ip_dst;
1000 inc.inc_route.ro_rt = NULL; /* XXX */
1002 inc.inc_fport = th->th_sport;
1003 inc.inc_lport = th->th_dport;
1006 * If the state is LISTEN then ignore segment if it contains
1007 * a RST. If the segment contains an ACK then it is bad and
1008 * send a RST. If it does not contain a SYN then it is not
1009 * interesting; drop it.
1011 * If the state is SYN_RECEIVED (syncache) and seg contains
1012 * an ACK, but not for our SYN/ACK, send a RST. If the seg
1013 * contains a RST, check the sequence number to see if it
1014 * is a valid reset segment.
1016 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) != TH_SYN) {
1017 if ((thflags & (TH_RST | TH_ACK | TH_SYN)) == TH_ACK) {
1018 if (!syncache_expand(&inc, th, &so, m)) {
1020 * No syncache entry, or ACK was not
1021 * for our SYN/ACK. Send a RST.
1023 tcpstat.tcps_badsyn++;
1024 rstreason = BANDLIM_RST_OPENPORT;
1029 * Could not complete 3-way handshake,
1030 * connection is being closed down, and
1031 * syncache will free mbuf.
1034 return(IPPROTO_DONE);
1037 * We must be in the correct protocol thread
1038 * for this connection.
1040 KKASSERT(so->so_port == &curthread->td_msgport);
1043 * Socket is created in state SYN_RECEIVED.
1044 * Continue processing segment.
1047 tp = intotcpcb(inp);
1049 * This is what would have happened in
1050 * tcp_output() when the SYN,ACK was sent.
1052 tp->snd_up = tp->snd_una;
1053 tp->snd_max = tp->snd_nxt = tp->iss + 1;
1054 tp->last_ack_sent = tp->rcv_nxt;
1058 if (thflags & TH_RST) {
1059 syncache_chkrst(&inc, th);
1062 if (thflags & TH_ACK) {
1063 syncache_badack(&inc);
1064 tcpstat.tcps_badsyn++;
1065 rstreason = BANDLIM_RST_OPENPORT;
1072 * Segment's flags are (SYN) or (SYN | FIN).
1076 * If deprecated address is forbidden,
1077 * we do not accept SYN to deprecated interface
1078 * address to prevent any new inbound connection from
1079 * getting established.
1080 * When we do not accept SYN, we send a TCP RST,
1081 * with deprecated source address (instead of dropping
1082 * it). We compromise it as it is much better for peer
1083 * to send a RST, and RST will be the final packet
1086 * If we do not forbid deprecated addresses, we accept
1087 * the SYN packet. RFC2462 does not suggest dropping
1089 * If we decipher RFC2462 5.5.4, it says like this:
1090 * 1. use of deprecated addr with existing
1091 * communication is okay - "SHOULD continue to be
1093 * 2. use of it with new communication:
1094 * (2a) "SHOULD NOT be used if alternate address
1095 * with sufficient scope is available"
1096 * (2b) nothing mentioned otherwise.
1097 * Here we fall into (2b) case as we have no choice in
1098 * our source address selection - we must obey the peer.
1100 * The wording in RFC2462 is confusing, and there are
1101 * multiple description text for deprecated address
1102 * handling - worse, they are not exactly the same.
1103 * I believe 5.5.4 is the best one, so we follow 5.5.4.
1105 if (isipv6 && !ip6_use_deprecated) {
1106 struct in6_ifaddr *ia6;
1108 if ((ia6 = ip6_getdstifaddr(m)) &&
1109 (ia6->ia6_flags & IN6_IFF_DEPRECATED)) {
1111 rstreason = BANDLIM_RST_OPENPORT;
1117 * If it is from this socket, drop it, it must be forged.
1118 * Don't bother responding if the destination was a broadcast.
1120 if (th->th_dport == th->th_sport) {
1122 if (IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst,
1126 if (ip->ip_dst.s_addr == ip->ip_src.s_addr)
1131 * RFC1122 4.2.3.10, p. 104: discard bcast/mcast SYN
1133 * Note that it is quite possible to receive unicast
1134 * link-layer packets with a broadcast IP address. Use
1135 * in_broadcast() to find them.
1137 if (m->m_flags & (M_BCAST | M_MCAST))
1140 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
1141 IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
1144 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
1145 IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
1146 ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
1147 in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
1151 * SYN appears to be valid; create compressed TCP state
1152 * for syncache, or perform t/tcp connection.
1154 if (so->so_qlen <= so->so_qlimit) {
1155 tcp_dooptions(&to, optp, optlen, TRUE, th->th_ack);
1156 if (!syncache_add(&inc, &to, th, so, m))
1160 * Entry added to syncache, mbuf used to
1161 * send SYN,ACK packet.
1163 return(IPPROTO_DONE);
1170 * Should not happen - syncache should pick up these connections.
1172 * Once we are past handling listen sockets we must be in the
1173 * correct protocol processing thread.
1175 KASSERT(tp->t_state != TCPS_LISTEN, ("tcp_input: TCPS_LISTEN state"));
1176 KKASSERT(so->so_port == &curthread->td_msgport);
1178 /* Unscale the window into a 32-bit value. */
1179 if (!(thflags & TH_SYN))
1180 tiwin = th->th_win << tp->snd_scale;
1185 * This is the second part of the MSS DoS prevention code (after
1186 * minmss on the sending side) and it deals with too many too small
1187 * tcp packets in a too short timeframe (1 second).
1189 * XXX Removed. This code was crap. It does not scale to network
1190 * speed, and default values break NFS. Gone.
1195 * Segment received on connection.
1197 * Reset idle time and keep-alive timer. Don't waste time if less
1198 * then a second has elapsed.
1200 if ((int)(ticks - tp->t_rcvtime) > hz)
1201 tcp_timer_keep_activity(tp, thflags);
1205 * XXX this is tradtitional behavior, may need to be cleaned up.
1207 tcp_dooptions(&to, optp, optlen, (thflags & TH_SYN) != 0, th->th_ack);
1208 if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) {
1209 if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) {
1210 tp->t_flags |= TF_RCVD_SCALE;
1211 tp->snd_scale = to.to_requested_s_scale;
1215 * Initial send window; will be updated upon next ACK
1217 tp->snd_wnd = th->th_win;
1219 if (to.to_flags & TOF_TS) {
1220 tp->t_flags |= TF_RCVD_TSTMP;
1221 tp->ts_recent = to.to_tsval;
1222 tp->ts_recent_age = ticks;
1224 if (!(to.to_flags & TOF_MSS))
1226 tcp_mss(tp, to.to_mss);
1228 * Only set the TF_SACK_PERMITTED per-connection flag
1229 * if we got a SACK_PERMITTED option from the other side
1230 * and the global tcp_do_sack variable is true.
1232 if (tcp_do_sack && (to.to_flags & TOF_SACK_PERMITTED))
1233 tp->t_flags |= TF_SACK_PERMITTED;
1237 * Header prediction: check for the two common cases
1238 * of a uni-directional data xfer. If the packet has
1239 * no control flags, is in-sequence, the window didn't
1240 * change and we're not retransmitting, it's a
1241 * candidate. If the length is zero and the ack moved
1242 * forward, we're the sender side of the xfer. Just
1243 * free the data acked & wake any higher level process
1244 * that was blocked waiting for space. If the length
1245 * is non-zero and the ack didn't move, we're the
1246 * receiver side. If we're getting packets in-order
1247 * (the reassembly queue is empty), add the data to
1248 * the socket buffer and note that we need a delayed ack.
1249 * Make sure that the hidden state-flags are also off.
1250 * Since we check for TCPS_ESTABLISHED above, it can only
1253 if (tp->t_state == TCPS_ESTABLISHED &&
1254 (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK &&
1255 !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1256 (!(to.to_flags & TOF_TS) ||
1257 TSTMP_GEQ(to.to_tsval, tp->ts_recent)) &&
1258 th->th_seq == tp->rcv_nxt &&
1259 tp->snd_nxt == tp->snd_max) {
1262 * If last ACK falls within this segment's sequence numbers,
1263 * record the timestamp.
1264 * NOTE that the test is modified according to the latest
1265 * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1267 if ((to.to_flags & TOF_TS) &&
1268 SEQ_LEQ(th->th_seq, tp->last_ack_sent)) {
1269 tp->ts_recent_age = ticks;
1270 tp->ts_recent = to.to_tsval;
1274 if (SEQ_GT(th->th_ack, tp->snd_una) &&
1275 SEQ_LEQ(th->th_ack, tp->snd_max) &&
1276 tp->snd_cwnd >= tp->snd_wnd &&
1277 !IN_FASTRECOVERY(tp)) {
1279 * This is a pure ack for outstanding data.
1281 ++tcpstat.tcps_predack;
1283 * "bad retransmit" recovery
1285 * If Eifel detection applies, then
1286 * it is deterministic, so use it
1287 * unconditionally over the old heuristic.
1288 * Otherwise, fall back to the old heuristic.
1290 if (tcp_do_eifel_detect &&
1291 (to.to_flags & TOF_TS) && to.to_tsecr &&
1292 (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
1293 /* Eifel detection applicable. */
1294 if (to.to_tsecr < tp->t_rexmtTS) {
1295 tcp_revert_congestion_state(tp);
1296 ++tcpstat.tcps_eifeldetected;
1297 if (tp->t_rxtshift != 1 ||
1298 ticks >= tp->t_badrxtwin)
1299 ++tcpstat.tcps_rttcantdetect;
1301 } else if (tp->t_rxtshift == 1 &&
1302 ticks < tp->t_badrxtwin) {
1303 tcp_revert_congestion_state(tp);
1304 ++tcpstat.tcps_rttdetected;
1306 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
1307 TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
1309 * Recalculate the retransmit timer / rtt.
1311 * Some machines (certain windows boxes)
1312 * send broken timestamp replies during the
1313 * SYN+ACK phase, ignore timestamps of 0.
1315 if ((to.to_flags & TOF_TS) && to.to_tsecr) {
1317 ticks - to.to_tsecr + 1,
1319 } else if (tp->t_rtttime &&
1320 SEQ_GT(th->th_ack, tp->t_rtseq)) {
1322 ticks - tp->t_rtttime,
1325 tcp_xmit_bandwidth_limit(tp, th->th_ack);
1326 acked = th->th_ack - tp->snd_una;
1327 tcpstat.tcps_rcvackpack++;
1328 tcpstat.tcps_rcvackbyte += acked;
1329 sbdrop(&so->so_snd.sb, acked);
1330 tp->snd_recover = th->th_ack - 1;
1331 tp->snd_una = th->th_ack;
1334 * Update window information.
1336 if (tiwin != tp->snd_wnd &&
1337 acceptable_window_update(tp, th, tiwin)) {
1338 /* keep track of pure window updates */
1339 if (tp->snd_wl2 == th->th_ack &&
1340 tiwin > tp->snd_wnd)
1341 tcpstat.tcps_rcvwinupd++;
1342 tp->snd_wnd = tiwin;
1343 tp->snd_wl1 = th->th_seq;
1344 tp->snd_wl2 = th->th_ack;
1345 if (tp->snd_wnd > tp->max_sndwnd)
1346 tp->max_sndwnd = tp->snd_wnd;
1349 ND6_HINT(tp); /* some progress has been done */
1351 * If all outstanding data are acked, stop
1352 * retransmit timer, otherwise restart timer
1353 * using current (possibly backed-off) value.
1354 * If process is waiting for space,
1355 * wakeup/selwakeup/signal. If data
1356 * are ready to send, let tcp_output
1357 * decide between more output or persist.
1359 if (tp->snd_una == tp->snd_max) {
1360 tcp_callout_stop(tp, tp->tt_rexmt);
1361 } else if (!tcp_callout_active(tp,
1363 tcp_callout_reset(tp, tp->tt_rexmt,
1364 tp->t_rxtcur, tcp_timer_rexmt);
1367 if (so->so_snd.ssb_cc > 0 &&
1368 !tcp_output_pending(tp))
1369 tcp_output_fair(tp);
1370 return(IPPROTO_DONE);
1372 } else if (tiwin == tp->snd_wnd &&
1373 th->th_ack == tp->snd_una &&
1374 TAILQ_EMPTY(&tp->t_segq) &&
1375 tlen <= ssb_space(&so->so_rcv)) {
1376 u_long newsize = 0; /* automatic sockbuf scaling */
1378 * This is a pure, in-sequence data packet
1379 * with nothing on the reassembly queue and
1380 * we have enough buffer space to take it.
1382 ++tcpstat.tcps_preddat;
1383 tp->rcv_nxt += tlen;
1384 tcpstat.tcps_rcvpack++;
1385 tcpstat.tcps_rcvbyte += tlen;
1386 ND6_HINT(tp); /* some progress has been done */
1388 * Automatic sizing of receive socket buffer. Often the send
1389 * buffer size is not optimally adjusted to the actual network
1390 * conditions at hand (delay bandwidth product). Setting the
1391 * buffer size too small limits throughput on links with high
1392 * bandwidth and high delay (eg. trans-continental/oceanic links).
1394 * On the receive side the socket buffer memory is only rarely
1395 * used to any significant extent. This allows us to be much
1396 * more aggressive in scaling the receive socket buffer. For
1397 * the case that the buffer space is actually used to a large
1398 * extent and we run out of kernel memory we can simply drop
1399 * the new segments; TCP on the sender will just retransmit it
1400 * later. Setting the buffer size too big may only consume too
1401 * much kernel memory if the application doesn't read() from
1402 * the socket or packet loss or reordering makes use of the
1405 * The criteria to step up the receive buffer one notch are:
1406 * 1. the number of bytes received during the time it takes
1407 * one timestamp to be reflected back to us (the RTT);
1408 * 2. received bytes per RTT is within seven eighth of the
1409 * current socket buffer size;
1410 * 3. receive buffer size has not hit maximal automatic size;
1412 * This algorithm does one step per RTT at most and only if
1413 * we receive a bulk stream w/o packet losses or reorderings.
1414 * Shrinking the buffer during idle times is not necessary as
1415 * it doesn't consume any memory when idle.
1417 * TODO: Only step up if the application is actually serving
1418 * the buffer to better manage the socket buffer resources.
1420 if (tcp_do_autorcvbuf &&
1422 (so->so_rcv.ssb_flags & SSB_AUTOSIZE)) {
1423 if (to.to_tsecr > tp->rfbuf_ts &&
1424 to.to_tsecr - tp->rfbuf_ts < hz) {
1426 (so->so_rcv.ssb_hiwat / 8 * 7) &&
1427 so->so_rcv.ssb_hiwat <
1428 tcp_autorcvbuf_max) {
1430 ulmin(so->so_rcv.ssb_hiwat +
1432 tcp_autorcvbuf_max);
1434 /* Start over with next RTT. */
1438 tp->rfbuf_cnt += tlen; /* add up */
1441 * Add data to socket buffer.
1443 if (so->so_state & SS_CANTRCVMORE) {
1447 * Set new socket buffer size, give up when
1450 * Adjusting the size can mess up ACK
1451 * sequencing when pure window updates are
1452 * being avoided (which is the default),
1455 lwkt_gettoken(&so->so_rcv.ssb_token);
1457 tp->t_flags |= TF_RXRESIZED;
1458 if (!ssb_reserve(&so->so_rcv, newsize,
1460 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1463 (TCP_MAXWIN << tp->rcv_scale)) {
1464 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AUTOSIZE);
1467 m_adj(m, drop_hdrlen); /* delayed header drop */
1468 ssb_appendstream(&so->so_rcv, m);
1469 lwkt_reltoken(&so->so_rcv.ssb_token);
1473 * This code is responsible for most of the ACKs
1474 * the TCP stack sends back after receiving a data
1475 * packet. Note that the DELAY_ACK check fails if
1476 * the delack timer is already running, which results
1477 * in an ack being sent every other packet (which is
1480 * We then further aggregate acks by not actually
1481 * sending one until the protocol thread has completed
1482 * processing the current backlog of packets. This
1483 * does not delay the ack any further, but allows us
1484 * to take advantage of the packet aggregation that
1485 * high speed NICs do (usually blocks of 8-10 packets)
1486 * to send a single ack rather then four or five acks,
1487 * greatly reducing the ack rate, the return channel
1488 * bandwidth, and the protocol overhead on both ends.
1490 * Since this also has the effect of slowing down
1491 * the exponential slow-start ramp-up, systems with
1492 * very large bandwidth-delay products might want
1493 * to turn the feature off.
1495 if (DELAY_ACK(tp)) {
1496 tcp_callout_reset(tp, tp->tt_delack,
1497 tcp_delacktime, tcp_timer_delack);
1498 } else if (tcp_aggregate_acks) {
1499 tp->t_flags |= TF_ACKNOW;
1500 if (!(tp->t_flags & TF_ONOUTPUTQ)) {
1501 tp->t_flags |= TF_ONOUTPUTQ;
1502 tp->tt_cpu = mycpu->gd_cpuid;
1504 &tcpcbackq[tp->tt_cpu],
1508 tp->t_flags |= TF_ACKNOW;
1511 return(IPPROTO_DONE);
1516 * Calculate amount of space in receive window,
1517 * and then do TCP input processing.
1518 * Receive window is amount of space in rcv queue,
1519 * but not less than advertised window.
1521 recvwin = ssb_space(&so->so_rcv);
1524 tp->rcv_wnd = imax(recvwin, (int)(tp->rcv_adv - tp->rcv_nxt));
1526 /* Reset receive buffer auto scaling when not in bulk receive mode. */
1530 switch (tp->t_state) {
1532 * If the state is SYN_RECEIVED:
1533 * if seg contains an ACK, but not for our SYN/ACK, send a RST.
1535 case TCPS_SYN_RECEIVED:
1536 if ((thflags & TH_ACK) &&
1537 (SEQ_LEQ(th->th_ack, tp->snd_una) ||
1538 SEQ_GT(th->th_ack, tp->snd_max))) {
1539 rstreason = BANDLIM_RST_OPENPORT;
1545 * If the state is SYN_SENT:
1546 * if seg contains an ACK, but not for our SYN, drop the input.
1547 * if seg contains a RST, then drop the connection.
1548 * if seg does not contain SYN, then drop it.
1549 * Otherwise this is an acceptable SYN segment
1550 * initialize tp->rcv_nxt and tp->irs
1551 * if seg contains ack then advance tp->snd_una
1552 * if SYN has been acked change to ESTABLISHED else SYN_RCVD state
1553 * arrange for segment to be acked (eventually)
1554 * continue processing rest of data/controls, beginning with URG
1557 if ((thflags & TH_ACK) &&
1558 (SEQ_LEQ(th->th_ack, tp->iss) ||
1559 SEQ_GT(th->th_ack, tp->snd_max))) {
1560 rstreason = BANDLIM_UNLIMITED;
1563 if (thflags & TH_RST) {
1564 if (thflags & TH_ACK)
1565 tp = tcp_drop(tp, ECONNREFUSED);
1568 if (!(thflags & TH_SYN))
1571 tp->irs = th->th_seq;
1573 if (thflags & TH_ACK) {
1574 /* Our SYN was acked. */
1575 tcpstat.tcps_connects++;
1577 /* Do window scaling on this connection? */
1578 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1579 (TF_RCVD_SCALE | TF_REQ_SCALE))
1580 tp->rcv_scale = tp->request_r_scale;
1581 tp->rcv_adv += tp->rcv_wnd;
1582 tp->snd_una++; /* SYN is acked */
1583 tcp_callout_stop(tp, tp->tt_rexmt);
1585 * If there's data, delay ACK; if there's also a FIN
1586 * ACKNOW will be turned on later.
1588 if (DELAY_ACK(tp) && tlen != 0) {
1589 tcp_callout_reset(tp, tp->tt_delack,
1590 tcp_delacktime, tcp_timer_delack);
1592 tp->t_flags |= TF_ACKNOW;
1595 * Received <SYN,ACK> in SYN_SENT[*] state.
1597 * SYN_SENT --> ESTABLISHED
1598 * SYN_SENT* --> FIN_WAIT_1
1600 tp->t_starttime = ticks;
1601 if (tp->t_flags & TF_NEEDFIN) {
1602 tp->t_state = TCPS_FIN_WAIT_1;
1603 tp->t_flags &= ~TF_NEEDFIN;
1606 tcp_established(tp);
1610 * Received initial SYN in SYN-SENT[*] state =>
1611 * simultaneous open.
1612 * Do 3-way handshake:
1613 * SYN-SENT -> SYN-RECEIVED
1614 * SYN-SENT* -> SYN-RECEIVED*
1616 tp->t_flags |= TF_ACKNOW;
1617 tcp_callout_stop(tp, tp->tt_rexmt);
1618 tp->t_state = TCPS_SYN_RECEIVED;
1622 * Advance th->th_seq to correspond to first data byte.
1623 * If data, trim to stay within window,
1624 * dropping FIN if necessary.
1627 if (tlen > tp->rcv_wnd) {
1628 todrop = tlen - tp->rcv_wnd;
1632 tcpstat.tcps_rcvpackafterwin++;
1633 tcpstat.tcps_rcvbyteafterwin += todrop;
1635 tp->snd_wl1 = th->th_seq - 1;
1636 tp->rcv_up = th->th_seq;
1638 * Client side of transaction: already sent SYN and data.
1639 * If the remote host used T/TCP to validate the SYN,
1640 * our data will be ACK'd; if so, enter normal data segment
1641 * processing in the middle of step 5, ack processing.
1642 * Otherwise, goto step 6.
1644 if (thflags & TH_ACK)
1650 * If the state is LAST_ACK or CLOSING or TIME_WAIT:
1651 * do normal processing (we no longer bother with T/TCP).
1655 case TCPS_TIME_WAIT:
1656 break; /* continue normal processing */
1660 * States other than LISTEN or SYN_SENT.
1661 * First check the RST flag and sequence number since reset segments
1662 * are exempt from the timestamp and connection count tests. This
1663 * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix
1664 * below which allowed reset segments in half the sequence space
1665 * to fall though and be processed (which gives forged reset
1666 * segments with a random sequence number a 50 percent chance of
1667 * killing a connection).
1668 * Then check timestamp, if present.
1669 * Then check the connection count, if present.
1670 * Then check that at least some bytes of segment are within
1671 * receive window. If segment begins before rcv_nxt,
1672 * drop leading data (and SYN); if nothing left, just ack.
1675 * If the RST bit is set, check the sequence number to see
1676 * if this is a valid reset segment.
1678 * In all states except SYN-SENT, all reset (RST) segments
1679 * are validated by checking their SEQ-fields. A reset is
1680 * valid if its sequence number is in the window.
1681 * Note: this does not take into account delayed ACKs, so
1682 * we should test against last_ack_sent instead of rcv_nxt.
1683 * The sequence number in the reset segment is normally an
1684 * echo of our outgoing acknowledgement numbers, but some hosts
1685 * send a reset with the sequence number at the rightmost edge
1686 * of our receive window, and we have to handle this case.
1687 * If we have multiple segments in flight, the intial reset
1688 * segment sequence numbers will be to the left of last_ack_sent,
1689 * but they will eventually catch up.
1690 * In any case, it never made sense to trim reset segments to
1691 * fit the receive window since RFC 1122 says:
1692 * 4.2.2.12 RST Segment: RFC-793 Section 3.4
1694 * A TCP SHOULD allow a received RST segment to include data.
1697 * It has been suggested that a RST segment could contain
1698 * ASCII text that encoded and explained the cause of the
1699 * RST. No standard has yet been established for such
1702 * If the reset segment passes the sequence number test examine
1704 * SYN_RECEIVED STATE:
1705 * If passive open, return to LISTEN state.
1706 * If active open, inform user that connection was refused.
1707 * ESTABLISHED, FIN_WAIT_1, FIN_WAIT_2, CLOSE_WAIT STATES:
1708 * Inform user that connection was reset, and close tcb.
1709 * CLOSING, LAST_ACK STATES:
1712 * Drop the segment - see Stevens, vol. 2, p. 964 and
1715 if (thflags & TH_RST) {
1716 if (SEQ_GEQ(th->th_seq, tp->last_ack_sent) &&
1717 SEQ_LEQ(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) {
1718 switch (tp->t_state) {
1720 case TCPS_SYN_RECEIVED:
1721 so->so_error = ECONNREFUSED;
1724 case TCPS_ESTABLISHED:
1725 case TCPS_FIN_WAIT_1:
1726 case TCPS_FIN_WAIT_2:
1727 case TCPS_CLOSE_WAIT:
1728 so->so_error = ECONNRESET;
1730 tp->t_state = TCPS_CLOSED;
1731 tcpstat.tcps_drops++;
1740 case TCPS_TIME_WAIT:
1748 * RFC 1323 PAWS: If we have a timestamp reply on this segment
1749 * and it's less than ts_recent, drop it.
1751 if ((to.to_flags & TOF_TS) && tp->ts_recent != 0 &&
1752 TSTMP_LT(to.to_tsval, tp->ts_recent)) {
1753 /* Check to see if ts_recent is over 24 days old. */
1754 if ((int)(ticks - tp->ts_recent_age) > TCP_PAWS_IDLE) {
1756 * Invalidate ts_recent. If this segment updates
1757 * ts_recent, the age will be reset later and ts_recent
1758 * will get a valid value. If it does not, setting
1759 * ts_recent to zero will at least satisfy the
1760 * requirement that zero be placed in the timestamp
1761 * echo reply when ts_recent isn't valid. The
1762 * age isn't reset until we get a valid ts_recent
1763 * because we don't want out-of-order segments to be
1764 * dropped when ts_recent is old.
1767 } else if (tcp_paws_tolerance && tlen != 0 &&
1768 tp->t_state == TCPS_ESTABLISHED &&
1769 (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK&&
1770 !(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)) &&
1771 th->th_ack == tp->snd_una &&
1772 tiwin == tp->snd_wnd &&
1773 TSTMP_GEQ(to.to_tsval + tcp_paws_tolerance, tp->ts_recent)&&
1774 (th->th_seq == tp->rcv_nxt ||
1775 (SEQ_GT(th->th_seq, tp->rcv_nxt) &&
1776 tcp_paws_canreasslast(tp, th, tlen)))) {
1778 * This tends to prevent valid new segments from being
1779 * dropped by the reordered segments sent by the fast
1780 * retransmission algorithm on the sending side, i.e.
1781 * the fast retransmitted segment w/ larger timestamp
1782 * arrives earlier than the previously sent new segments
1783 * w/ smaller timestamp.
1785 * If following conditions are met, the segment is
1787 * - The segment contains data
1788 * - The connection is established
1789 * - The header does not contain important flags
1790 * - SYN or FIN is not needed
1791 * - It does not acknowledge new data
1792 * - Receive window is not changed
1793 * - The timestamp is within "acceptable" range
1794 * - The new segment is what we are expecting or
1795 * the new segment could be merged w/ the last
1796 * pending segment on the reassemble queue
1798 tcpstat.tcps_pawsaccept++;
1799 tcpstat.tcps_pawsdrop++;
1801 tcpstat.tcps_rcvduppack++;
1802 tcpstat.tcps_rcvdupbyte += tlen;
1803 tcpstat.tcps_pawsdrop++;
1811 * In the SYN-RECEIVED state, validate that the packet belongs to
1812 * this connection before trimming the data to fit the receive
1813 * window. Check the sequence number versus IRS since we know
1814 * the sequence numbers haven't wrapped. This is a partial fix
1815 * for the "LAND" DoS attack.
1817 if (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) {
1818 rstreason = BANDLIM_RST_OPENPORT;
1822 todrop = tp->rcv_nxt - th->th_seq;
1824 if (TCP_DO_SACK(tp)) {
1825 /* Report duplicate segment at head of packet. */
1826 tp->reportblk.rblk_start = th->th_seq;
1827 tp->reportblk.rblk_end = TCP_SACK_BLKEND(
1828 th->th_seq + tlen, thflags);
1829 if (SEQ_GT(tp->reportblk.rblk_end, tp->rcv_nxt))
1830 tp->reportblk.rblk_end = tp->rcv_nxt;
1831 tp->sack_flags |= (TSACK_F_DUPSEG | TSACK_F_SACKLEFT);
1832 tp->t_flags |= TF_ACKNOW;
1834 if (thflags & TH_SYN) {
1844 * Following if statement from Stevens, vol. 2, p. 960.
1846 if (todrop > tlen ||
1847 (todrop == tlen && !(thflags & TH_FIN))) {
1849 * Any valid FIN must be to the left of the window.
1850 * At this point the FIN must be a duplicate or out
1851 * of sequence; drop it.
1856 * Send an ACK to resynchronize and drop any data.
1857 * But keep on processing for RST or ACK.
1859 tp->t_flags |= TF_ACKNOW;
1861 tcpstat.tcps_rcvduppack++;
1862 tcpstat.tcps_rcvdupbyte += todrop;
1864 tcpstat.tcps_rcvpartduppack++;
1865 tcpstat.tcps_rcvpartdupbyte += todrop;
1867 drop_hdrlen += todrop; /* drop from the top afterwards */
1868 th->th_seq += todrop;
1870 if (th->th_urp > todrop)
1871 th->th_urp -= todrop;
1879 * If new data are received on a connection after the
1880 * user processes are gone, then RST the other end.
1882 if ((so->so_state & SS_NOFDREF) &&
1883 tp->t_state > TCPS_CLOSE_WAIT && tlen) {
1885 tcpstat.tcps_rcvafterclose++;
1886 rstreason = BANDLIM_UNLIMITED;
1891 * If segment ends after window, drop trailing data
1892 * (and PUSH and FIN); if nothing left, just ACK.
1894 todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd);
1896 tcpstat.tcps_rcvpackafterwin++;
1897 if (todrop >= tlen) {
1898 tcpstat.tcps_rcvbyteafterwin += tlen;
1900 * If a new connection request is received
1901 * while in TIME_WAIT, drop the old connection
1902 * and start over if the sequence numbers
1903 * are above the previous ones.
1905 if (thflags & TH_SYN &&
1906 tp->t_state == TCPS_TIME_WAIT &&
1907 SEQ_GT(th->th_seq, tp->rcv_nxt)) {
1912 * If window is closed can only take segments at
1913 * window edge, and have to drop data and PUSH from
1914 * incoming segments. Continue processing, but
1915 * remember to ack. Otherwise, drop segment
1918 if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) {
1919 tp->t_flags |= TF_ACKNOW;
1920 tcpstat.tcps_rcvwinprobe++;
1924 tcpstat.tcps_rcvbyteafterwin += todrop;
1927 thflags &= ~(TH_PUSH | TH_FIN);
1931 * If last ACK falls within this segment's sequence numbers,
1932 * record its timestamp.
1934 * 1) That the test incorporates suggestions from the latest
1935 * proposal of the tcplw@cray.com list (Braden 1993/04/26).
1936 * 2) That updating only on newer timestamps interferes with
1937 * our earlier PAWS tests, so this check should be solely
1938 * predicated on the sequence space of this segment.
1939 * 3) That we modify the segment boundary check to be
1940 * Last.ACK.Sent <= SEG.SEQ + SEG.LEN
1941 * instead of RFC1323's
1942 * Last.ACK.Sent < SEG.SEQ + SEG.LEN,
1943 * This modified check allows us to overcome RFC1323's
1944 * limitations as described in Stevens TCP/IP Illustrated
1945 * Vol. 2 p.869. In such cases, we can still calculate the
1946 * RTT correctly when RCV.NXT == Last.ACK.Sent.
1948 if ((to.to_flags & TOF_TS) && SEQ_LEQ(th->th_seq, tp->last_ack_sent) &&
1949 SEQ_LEQ(tp->last_ack_sent, (th->th_seq + tlen
1950 + ((thflags & TH_SYN) != 0)
1951 + ((thflags & TH_FIN) != 0)))) {
1952 tp->ts_recent_age = ticks;
1953 tp->ts_recent = to.to_tsval;
1957 * If a SYN is in the window, then this is an
1958 * error and we send an RST and drop the connection.
1960 if (thflags & TH_SYN) {
1961 tp = tcp_drop(tp, ECONNRESET);
1962 rstreason = BANDLIM_UNLIMITED;
1967 * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN
1968 * flag is on (half-synchronized state), then queue data for
1969 * later processing; else drop segment and return.
1971 if (!(thflags & TH_ACK)) {
1972 if (tp->t_state == TCPS_SYN_RECEIVED ||
1973 (tp->t_flags & TF_NEEDSYN))
1982 switch (tp->t_state) {
1984 * In SYN_RECEIVED state, the ACK acknowledges our SYN, so enter
1985 * ESTABLISHED state and continue processing.
1986 * The ACK was checked above.
1988 case TCPS_SYN_RECEIVED:
1990 tcpstat.tcps_connects++;
1992 /* Do window scaling? */
1993 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
1994 (TF_RCVD_SCALE | TF_REQ_SCALE))
1995 tp->rcv_scale = tp->request_r_scale;
1998 * SYN-RECEIVED -> ESTABLISHED
1999 * SYN-RECEIVED* -> FIN-WAIT-1
2001 tp->t_starttime = ticks;
2002 if (tp->t_flags & TF_NEEDFIN) {
2003 tp->t_state = TCPS_FIN_WAIT_1;
2004 tp->t_flags &= ~TF_NEEDFIN;
2006 tcp_established(tp);
2009 * If segment contains data or ACK, will call tcp_reass()
2010 * later; if not, do so now to pass queued data to user.
2012 if (tlen == 0 && !(thflags & TH_FIN))
2013 tcp_reass(tp, NULL, NULL, NULL);
2017 * In ESTABLISHED state: drop duplicate ACKs; ACK out of range
2018 * ACKs. If the ack is in the range
2019 * tp->snd_una < th->th_ack <= tp->snd_max
2020 * then advance tp->snd_una to th->th_ack and drop
2021 * data from the retransmission queue. If this ACK reflects
2022 * more up to date window information we update our window information.
2024 case TCPS_ESTABLISHED:
2025 case TCPS_FIN_WAIT_1:
2026 case TCPS_FIN_WAIT_2:
2027 case TCPS_CLOSE_WAIT:
2030 case TCPS_TIME_WAIT:
2032 if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
2033 boolean_t maynotdup = FALSE;
2035 if (TCP_DO_SACK(tp))
2036 tcp_sack_update_scoreboard(tp, &to);
2038 if (tlen != 0 || tiwin != tp->snd_wnd ||
2039 ((thflags & TH_FIN) && !(tp->t_flags & TF_SAWFIN)))
2042 if (!tcp_callout_active(tp, tp->tt_rexmt) ||
2043 th->th_ack != tp->snd_una) {
2045 tcpstat.tcps_rcvdupack++;
2050 #define DELAY_DUPACK \
2052 delayed_dupack = TRUE; \
2053 th_dupack = th->th_ack; \
2054 to_flags = to.to_flags; \
2057 if (!tcp_do_rfc6675 ||
2060 (TOF_SACK | TOF_SACK_REDUNDANT))
2068 if ((thflags & TH_FIN) && !(tp->t_flags & TF_QUEDFIN)) {
2070 * This could happen, if the reassemable
2071 * queue overflew or was drained. Don't
2072 * drop this FIN here; defer the duplicated
2073 * ACK processing until this FIN gets queued.
2080 if (tcp_recv_dupack(tp, th->th_ack, to.to_flags))
2086 KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("th_ack <= snd_una"));
2088 if (SEQ_GT(th->th_ack, tp->snd_max)) {
2090 * Detected optimistic ACK attack.
2091 * Force slow-start to de-synchronize attack.
2093 tp->snd_cwnd = tp->t_maxseg;
2096 tcpstat.tcps_rcvacktoomuch++;
2100 * If we reach this point, ACK is not a duplicate,
2101 * i.e., it ACKs something we sent.
2103 if (tp->t_flags & TF_NEEDSYN) {
2105 * T/TCP: Connection was half-synchronized, and our
2106 * SYN has been ACK'd (so connection is now fully
2107 * synchronized). Go to non-starred state,
2108 * increment snd_una for ACK of SYN, and check if
2109 * we can do window scaling.
2111 tp->t_flags &= ~TF_NEEDSYN;
2113 /* Do window scaling? */
2114 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
2115 (TF_RCVD_SCALE | TF_REQ_SCALE))
2116 tp->rcv_scale = tp->request_r_scale;
2120 acked = th->th_ack - tp->snd_una;
2121 tcpstat.tcps_rcvackpack++;
2122 tcpstat.tcps_rcvackbyte += acked;
2124 if (tcp_do_eifel_detect && acked > 0 &&
2125 (to.to_flags & TOF_TS) && (to.to_tsecr != 0) &&
2126 (tp->rxt_flags & TRXT_F_FIRSTACCACK)) {
2127 /* Eifel detection applicable. */
2128 if (to.to_tsecr < tp->t_rexmtTS) {
2129 ++tcpstat.tcps_eifeldetected;
2130 tcp_revert_congestion_state(tp);
2131 if (tp->t_rxtshift != 1 ||
2132 ticks >= tp->t_badrxtwin)
2133 ++tcpstat.tcps_rttcantdetect;
2135 } else if (tp->t_rxtshift == 1 && ticks < tp->t_badrxtwin) {
2137 * If we just performed our first retransmit,
2138 * and the ACK arrives within our recovery window,
2139 * then it was a mistake to do the retransmit
2140 * in the first place. Recover our original cwnd
2141 * and ssthresh, and proceed to transmit where we
2144 tcp_revert_congestion_state(tp);
2145 ++tcpstat.tcps_rttdetected;
2149 * If we have a timestamp reply, update smoothed
2150 * round trip time. If no timestamp is present but
2151 * transmit timer is running and timed sequence
2152 * number was acked, update smoothed round trip time.
2153 * Since we now have an rtt measurement, cancel the
2154 * timer backoff (cf., Phil Karn's retransmit alg.).
2155 * Recompute the initial retransmit timer.
2157 * Some machines (certain windows boxes) send broken
2158 * timestamp replies during the SYN+ACK phase, ignore
2161 if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0))
2162 tcp_xmit_timer(tp, ticks - to.to_tsecr + 1, th->th_ack);
2163 else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq))
2164 tcp_xmit_timer(tp, ticks - tp->t_rtttime, th->th_ack);
2165 tcp_xmit_bandwidth_limit(tp, th->th_ack);
2168 * If no data (only SYN) was ACK'd,
2169 * skip rest of ACK processing.
2174 /* Stop looking for an acceptable ACK since one was received. */
2175 tp->rxt_flags &= ~(TRXT_F_FIRSTACCACK |
2176 TRXT_F_FASTREXMT | TRXT_F_EARLYREXMT);
2178 if (acked > so->so_snd.ssb_cc) {
2179 tp->snd_wnd -= so->so_snd.ssb_cc;
2180 sbdrop(&so->so_snd.sb, (int)so->so_snd.ssb_cc);
2181 ourfinisacked = TRUE;
2183 sbdrop(&so->so_snd.sb, acked);
2184 tp->snd_wnd -= acked;
2185 ourfinisacked = FALSE;
2190 * Update window information.
2192 if (acceptable_window_update(tp, th, tiwin)) {
2193 /* keep track of pure window updates */
2194 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2195 tiwin > tp->snd_wnd)
2196 tcpstat.tcps_rcvwinupd++;
2197 tp->snd_wnd = tiwin;
2198 tp->snd_wl1 = th->th_seq;
2199 tp->snd_wl2 = th->th_ack;
2200 if (tp->snd_wnd > tp->max_sndwnd)
2201 tp->max_sndwnd = tp->snd_wnd;
2205 tp->snd_una = th->th_ack;
2206 if (TCP_DO_SACK(tp))
2207 tcp_sack_update_scoreboard(tp, &to);
2208 if (IN_FASTRECOVERY(tp)) {
2209 if (SEQ_GEQ(th->th_ack, tp->snd_recover)) {
2210 EXIT_FASTRECOVERY(tp);
2213 * If the congestion window was inflated
2214 * to account for the other side's
2215 * cached packets, retract it.
2217 if (!TCP_DO_SACK(tp))
2218 tp->snd_cwnd = tp->snd_ssthresh;
2221 * Window inflation should have left us
2222 * with approximately snd_ssthresh outstanding
2223 * data. But, in case we would be inclined
2224 * to send a burst, better do it using
2227 if (SEQ_GT(th->th_ack + tp->snd_cwnd,
2228 tp->snd_max + 2 * tp->t_maxseg))
2230 (tp->snd_max - tp->snd_una) +
2235 if (TCP_DO_SACK(tp)) {
2236 tp->snd_max_rexmt = tp->snd_max;
2238 tp->snd_una == tp->rexmt_high);
2240 tcp_newreno_partial_ack(tp, th, acked);
2246 * Open the congestion window. When in slow-start,
2247 * open exponentially: maxseg per packet. Otherwise,
2248 * open linearly: maxseg per window.
2250 if (tp->snd_cwnd <= tp->snd_ssthresh) {
2252 (SEQ_LT(tp->snd_nxt, tp->snd_max) ?
2253 tp->t_maxseg : 2 * tp->t_maxseg);
2256 tp->snd_cwnd += tcp_do_abc ?
2257 min(acked, abc_sslimit) : tp->t_maxseg;
2259 /* linear increase */
2260 tp->snd_wacked += tcp_do_abc ? acked :
2262 if (tp->snd_wacked >= tp->snd_cwnd) {
2263 tp->snd_wacked -= tp->snd_cwnd;
2264 tp->snd_cwnd += tp->t_maxseg;
2267 tp->snd_cwnd = min(tp->snd_cwnd,
2268 TCP_MAXWIN << tp->snd_scale);
2269 tp->snd_recover = th->th_ack - 1;
2271 if (SEQ_LT(tp->snd_nxt, tp->snd_una))
2272 tp->snd_nxt = tp->snd_una;
2275 * If all outstanding data is acked, stop retransmit
2276 * timer and remember to restart (more output or persist).
2277 * If there is more data to be acked, restart retransmit
2278 * timer, using current (possibly backed-off) value.
2280 if (th->th_ack == tp->snd_max) {
2281 tcp_callout_stop(tp, tp->tt_rexmt);
2283 } else if (!tcp_callout_active(tp, tp->tt_persist)) {
2284 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
2288 switch (tp->t_state) {
2290 * In FIN_WAIT_1 STATE in addition to the processing
2291 * for the ESTABLISHED state if our FIN is now acknowledged
2292 * then enter FIN_WAIT_2.
2294 case TCPS_FIN_WAIT_1:
2295 if (ourfinisacked) {
2297 * If we can't receive any more
2298 * data, then closing user can proceed.
2299 * Starting the timer is contrary to the
2300 * specification, but if we don't get a FIN
2301 * we'll hang forever.
2303 if (so->so_state & SS_CANTRCVMORE) {
2304 soisdisconnected(so);
2305 tcp_callout_reset(tp, tp->tt_2msl,
2306 tp->t_maxidle, tcp_timer_2msl);
2308 tp->t_state = TCPS_FIN_WAIT_2;
2313 * In CLOSING STATE in addition to the processing for
2314 * the ESTABLISHED state if the ACK acknowledges our FIN
2315 * then enter the TIME-WAIT state, otherwise ignore
2319 if (ourfinisacked) {
2320 tp->t_state = TCPS_TIME_WAIT;
2321 tcp_canceltimers(tp);
2322 tcp_callout_reset(tp, tp->tt_2msl,
2323 2 * tcp_rmx_msl(tp),
2325 soisdisconnected(so);
2330 * In LAST_ACK, we may still be waiting for data to drain
2331 * and/or to be acked, as well as for the ack of our FIN.
2332 * If our FIN is now acknowledged, delete the TCB,
2333 * enter the closed state and return.
2336 if (ourfinisacked) {
2343 * In TIME_WAIT state the only thing that should arrive
2344 * is a retransmission of the remote FIN. Acknowledge
2345 * it and restart the finack timer.
2347 case TCPS_TIME_WAIT:
2348 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2356 * Update window information.
2357 * Don't look at window if no ACK: TAC's send garbage on first SYN.
2359 if ((thflags & TH_ACK) &&
2360 acceptable_window_update(tp, th, tiwin)) {
2361 /* keep track of pure window updates */
2362 if (tlen == 0 && tp->snd_wl2 == th->th_ack &&
2363 tiwin > tp->snd_wnd)
2364 tcpstat.tcps_rcvwinupd++;
2365 tp->snd_wnd = tiwin;
2366 tp->snd_wl1 = th->th_seq;
2367 tp->snd_wl2 = th->th_ack;
2368 if (tp->snd_wnd > tp->max_sndwnd)
2369 tp->max_sndwnd = tp->snd_wnd;
2374 * Process segments with URG.
2376 if ((thflags & TH_URG) && th->th_urp &&
2377 !TCPS_HAVERCVDFIN(tp->t_state)) {
2379 * This is a kludge, but if we receive and accept
2380 * random urgent pointers, we'll crash in
2381 * soreceive. It's hard to imagine someone
2382 * actually wanting to send this much urgent data.
2384 if (th->th_urp + so->so_rcv.ssb_cc > sb_max) {
2385 th->th_urp = 0; /* XXX */
2386 thflags &= ~TH_URG; /* XXX */
2387 goto dodata; /* XXX */
2390 * If this segment advances the known urgent pointer,
2391 * then mark the data stream. This should not happen
2392 * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since
2393 * a FIN has been received from the remote side.
2394 * In these states we ignore the URG.
2396 * According to RFC961 (Assigned Protocols),
2397 * the urgent pointer points to the last octet
2398 * of urgent data. We continue, however,
2399 * to consider it to indicate the first octet
2400 * of data past the urgent section as the original
2401 * spec states (in one of two places).
2403 if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) {
2404 tp->rcv_up = th->th_seq + th->th_urp;
2405 so->so_oobmark = so->so_rcv.ssb_cc +
2406 (tp->rcv_up - tp->rcv_nxt) - 1;
2407 if (so->so_oobmark == 0)
2408 sosetstate(so, SS_RCVATMARK);
2410 tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA);
2413 * Remove out of band data so doesn't get presented to user.
2414 * This can happen independent of advancing the URG pointer,
2415 * but if two URG's are pending at once, some out-of-band
2416 * data may creep in... ick.
2418 if (th->th_urp <= (u_long)tlen &&
2419 !(so->so_options & SO_OOBINLINE)) {
2420 /* hdr drop is delayed */
2421 tcp_pulloutofband(so, th, m, drop_hdrlen);
2425 * If no out of band data is expected,
2426 * pull receive urgent pointer along
2427 * with the receive window.
2429 if (SEQ_GT(tp->rcv_nxt, tp->rcv_up))
2430 tp->rcv_up = tp->rcv_nxt;
2435 * Process the segment text, merging it into the TCP sequencing queue,
2436 * and arranging for acknowledgment of receipt if necessary.
2437 * This process logically involves adjusting tp->rcv_wnd as data
2438 * is presented to the user (this happens in tcp_usrreq.c,
2439 * case PRU_RCVD). If a FIN has already been received on this
2440 * connection then we just ignore the text.
2442 if ((tlen || (thflags & TH_FIN)) && !TCPS_HAVERCVDFIN(tp->t_state)) {
2443 if (thflags & TH_FIN)
2444 tp->t_flags |= TF_SAWFIN;
2445 m_adj(m, drop_hdrlen); /* delayed header drop */
2447 * Insert segment which includes th into TCP reassembly queue
2448 * with control block tp. Set thflags to whether reassembly now
2449 * includes a segment with FIN. This handles the common case
2450 * inline (segment is the next to be received on an established
2451 * connection, and the queue is empty), avoiding linkage into
2452 * and removal from the queue and repetition of various
2454 * Set DELACK for segments received in order, but ack
2455 * immediately when segments are out of order (so
2456 * fast retransmit can work).
2458 if (th->th_seq == tp->rcv_nxt &&
2459 TAILQ_EMPTY(&tp->t_segq) &&
2460 TCPS_HAVEESTABLISHED(tp->t_state)) {
2461 if (thflags & TH_FIN)
2462 tp->t_flags |= TF_QUEDFIN;
2463 if (DELAY_ACK(tp)) {
2464 tcp_callout_reset(tp, tp->tt_delack,
2465 tcp_delacktime, tcp_timer_delack);
2467 tp->t_flags |= TF_ACKNOW;
2469 tp->rcv_nxt += tlen;
2470 thflags = th->th_flags & TH_FIN;
2471 tcpstat.tcps_rcvpack++;
2472 tcpstat.tcps_rcvbyte += tlen;
2474 if (so->so_state & SS_CANTRCVMORE) {
2477 lwkt_gettoken(&so->so_rcv.ssb_token);
2478 ssb_appendstream(&so->so_rcv, m);
2479 lwkt_reltoken(&so->so_rcv.ssb_token);
2483 if (!(tp->sack_flags & TSACK_F_DUPSEG)) {
2484 /* Initialize SACK report block. */
2485 tp->reportblk.rblk_start = th->th_seq;
2486 tp->reportblk.rblk_end = TCP_SACK_BLKEND(
2487 th->th_seq + tlen, thflags);
2489 thflags = tcp_reass(tp, th, &tlen, m);
2490 tp->t_flags |= TF_ACKNOW;
2494 * Note the amount of data that peer has sent into
2495 * our window, in order to estimate the sender's
2498 len = so->so_rcv.ssb_hiwat - (tp->rcv_adv - tp->rcv_nxt);
2505 * If FIN is received ACK the FIN and let the user know
2506 * that the connection is closing.
2508 if (thflags & TH_FIN) {
2509 if (!TCPS_HAVERCVDFIN(tp->t_state)) {
2512 * If connection is half-synchronized
2513 * (ie NEEDSYN flag on) then delay ACK,
2514 * so it may be piggybacked when SYN is sent.
2515 * Otherwise, since we received a FIN then no
2516 * more input can be expected, send ACK now.
2518 if (DELAY_ACK(tp) && (tp->t_flags & TF_NEEDSYN)) {
2519 tcp_callout_reset(tp, tp->tt_delack,
2520 tcp_delacktime, tcp_timer_delack);
2522 tp->t_flags |= TF_ACKNOW;
2527 switch (tp->t_state) {
2529 * In SYN_RECEIVED and ESTABLISHED STATES
2530 * enter the CLOSE_WAIT state.
2532 case TCPS_SYN_RECEIVED:
2533 tp->t_starttime = ticks;
2535 case TCPS_ESTABLISHED:
2536 tp->t_state = TCPS_CLOSE_WAIT;
2540 * If still in FIN_WAIT_1 STATE FIN has not been acked so
2541 * enter the CLOSING state.
2543 case TCPS_FIN_WAIT_1:
2544 tp->t_state = TCPS_CLOSING;
2548 * In FIN_WAIT_2 state enter the TIME_WAIT state,
2549 * starting the time-wait timer, turning off the other
2552 case TCPS_FIN_WAIT_2:
2553 tp->t_state = TCPS_TIME_WAIT;
2554 tcp_canceltimers(tp);
2555 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2557 soisdisconnected(so);
2561 * In TIME_WAIT state restart the 2 MSL time_wait timer.
2563 case TCPS_TIME_WAIT:
2564 tcp_callout_reset(tp, tp->tt_2msl, 2 * tcp_rmx_msl(tp),
2571 if (so->so_options & SO_DEBUG)
2572 tcp_trace(TA_INPUT, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2576 * Delayed duplicated ACK processing
2578 if (delayed_dupack && tcp_recv_dupack(tp, th_dupack, to_flags))
2582 * Return any desired output.
2584 if ((tp->t_flags & TF_ACKNOW) ||
2585 (needoutput && tcp_sack_report_needed(tp))) {
2586 tcp_output_cancel(tp);
2587 tcp_output_fair(tp);
2588 } else if (needoutput && !tcp_output_pending(tp)) {
2589 tcp_output_fair(tp);
2591 tcp_sack_report_cleanup(tp);
2592 return(IPPROTO_DONE);
2596 * Generate an ACK dropping incoming segment if it occupies
2597 * sequence space, where the ACK reflects our state.
2599 * We can now skip the test for the RST flag since all
2600 * paths to this code happen after packets containing
2601 * RST have been dropped.
2603 * In the SYN-RECEIVED state, don't send an ACK unless the
2604 * segment we received passes the SYN-RECEIVED ACK test.
2605 * If it fails send a RST. This breaks the loop in the
2606 * "LAND" DoS attack, and also prevents an ACK storm
2607 * between two listening ports that have been sent forged
2608 * SYN segments, each with the source address of the other.
2610 if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) &&
2611 (SEQ_GT(tp->snd_una, th->th_ack) ||
2612 SEQ_GT(th->th_ack, tp->snd_max)) ) {
2613 rstreason = BANDLIM_RST_OPENPORT;
2617 if (so->so_options & SO_DEBUG)
2618 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2621 tp->t_flags |= TF_ACKNOW;
2623 tcp_sack_report_cleanup(tp);
2624 return(IPPROTO_DONE);
2628 * Generate a RST, dropping incoming segment.
2629 * Make ACK acceptable to originator of segment.
2630 * Don't bother to respond if destination was broadcast/multicast.
2632 if ((thflags & TH_RST) || m->m_flags & (M_BCAST | M_MCAST))
2635 if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) ||
2636 IN6_IS_ADDR_MULTICAST(&ip6->ip6_src))
2639 if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) ||
2640 IN_MULTICAST(ntohl(ip->ip_src.s_addr)) ||
2641 ip->ip_src.s_addr == htonl(INADDR_BROADCAST) ||
2642 in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif))
2645 /* IPv6 anycast check is done at tcp6_input() */
2648 * Perform bandwidth limiting.
2651 if (badport_bandlim(rstreason) < 0)
2656 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2657 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2659 if (thflags & TH_ACK)
2660 /* mtod() below is safe as long as hdr dropping is delayed */
2661 tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack,
2664 if (thflags & TH_SYN)
2666 /* mtod() below is safe as long as hdr dropping is delayed */
2667 tcp_respond(tp, mtod(m, void *), th, m, th->th_seq + tlen,
2668 (tcp_seq)0, TH_RST | TH_ACK);
2671 tcp_sack_report_cleanup(tp);
2672 return(IPPROTO_DONE);
2676 * Drop space held by incoming segment and return.
2679 if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG))
2680 tcp_trace(TA_DROP, ostate, tp, tcp_saveipgen, &tcp_savetcp, 0);
2684 tcp_sack_report_cleanup(tp);
2685 return(IPPROTO_DONE);
2689 * Parse TCP options and place in tcpopt.
2692 tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, boolean_t is_syn,
2698 for (; cnt > 0; cnt -= optlen, cp += optlen) {
2700 if (opt == TCPOPT_EOL)
2702 if (opt == TCPOPT_NOP)
2708 if (optlen < 2 || optlen > cnt)
2713 if (optlen != TCPOLEN_MAXSEG)
2717 to->to_flags |= TOF_MSS;
2718 bcopy(cp + 2, &to->to_mss, sizeof to->to_mss);
2719 to->to_mss = ntohs(to->to_mss);
2722 if (optlen != TCPOLEN_WINDOW)
2726 to->to_flags |= TOF_SCALE;
2727 to->to_requested_s_scale = min(cp[2], TCP_MAX_WINSHIFT);
2729 case TCPOPT_TIMESTAMP:
2730 if (optlen != TCPOLEN_TIMESTAMP)
2732 to->to_flags |= TOF_TS;
2733 bcopy(cp + 2, &to->to_tsval, sizeof to->to_tsval);
2734 to->to_tsval = ntohl(to->to_tsval);
2735 bcopy(cp + 6, &to->to_tsecr, sizeof to->to_tsecr);
2736 to->to_tsecr = ntohl(to->to_tsecr);
2738 * If echoed timestamp is later than the current time,
2739 * fall back to non RFC1323 RTT calculation.
2741 if (to->to_tsecr != 0 && TSTMP_GT(to->to_tsecr, ticks))
2744 case TCPOPT_SACK_PERMITTED:
2745 if (optlen != TCPOLEN_SACK_PERMITTED)
2749 to->to_flags |= TOF_SACK_PERMITTED;
2752 if ((optlen - 2) & 0x07) /* not multiple of 8 */
2754 to->to_nsackblocks = (optlen - 2) / 8;
2755 to->to_sackblocks = (struct raw_sackblock *) (cp + 2);
2756 to->to_flags |= TOF_SACK;
2757 for (i = 0; i < to->to_nsackblocks; i++) {
2758 struct raw_sackblock *r = &to->to_sackblocks[i];
2760 r->rblk_start = ntohl(r->rblk_start);
2761 r->rblk_end = ntohl(r->rblk_end);
2763 if (SEQ_LEQ(r->rblk_end, r->rblk_start)) {
2765 * Invalid SACK block; discard all
2768 tcpstat.tcps_rcvbadsackopt++;
2769 to->to_nsackblocks = 0;
2770 to->to_sackblocks = NULL;
2771 to->to_flags &= ~TOF_SACK;
2775 if ((to->to_flags & TOF_SACK) &&
2776 tcp_sack_ndsack_blocks(to->to_sackblocks,
2777 to->to_nsackblocks, ack))
2778 to->to_flags |= TOF_DSACK;
2780 #ifdef TCP_SIGNATURE
2782 * XXX In order to reply to a host which has set the
2783 * TCP_SIGNATURE option in its initial SYN, we have to
2784 * record the fact that the option was observed here
2785 * for the syncache code to perform the correct response.
2787 case TCPOPT_SIGNATURE:
2788 if (optlen != TCPOLEN_SIGNATURE)
2790 to->to_flags |= (TOF_SIGNATURE | TOF_SIGLEN);
2792 #endif /* TCP_SIGNATURE */
2800 * Pull out of band byte out of a segment so
2801 * it doesn't appear in the user's data queue.
2802 * It is still reflected in the segment length for
2803 * sequencing purposes.
2804 * "off" is the delayed to be dropped hdrlen.
2807 tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off)
2809 int cnt = off + th->th_urp - 1;
2812 if (m->m_len > cnt) {
2813 char *cp = mtod(m, caddr_t) + cnt;
2814 struct tcpcb *tp = sototcpcb(so);
2817 tp->t_oobflags |= TCPOOB_HAVEDATA;
2818 bcopy(cp + 1, cp, m->m_len - cnt - 1);
2820 if (m->m_flags & M_PKTHDR)
2829 panic("tcp_pulloutofband");
2833 * Collect new round-trip time estimate
2834 * and update averages and current timeout.
2837 tcp_xmit_timer(struct tcpcb *tp, int rtt, tcp_seq ack)
2841 tcpstat.tcps_rttupdated++;
2843 if ((tp->rxt_flags & TRXT_F_REBASERTO) &&
2844 SEQ_GT(ack, tp->snd_max_prev)) {
2845 #ifdef DEBUG_EIFEL_RESPONSE
2846 kprintf("srtt/rttvar, prev %d/%d, cur %d/%d, ",
2847 tp->t_srtt_prev, tp->t_rttvar_prev,
2848 tp->t_srtt, tp->t_rttvar);
2851 tcpstat.tcps_eifelresponse++;
2853 tp->rxt_flags &= ~TRXT_F_REBASERTO;
2854 tp->t_srtt = max(tp->t_srtt_prev, (rtt << TCP_RTT_SHIFT));
2855 tp->t_rttvar = max(tp->t_rttvar_prev,
2856 (rtt << (TCP_RTTVAR_SHIFT - 1)));
2857 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2858 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2860 #ifdef DEBUG_EIFEL_RESPONSE
2861 kprintf("new %d/%d ", tp->t_srtt, tp->t_rttvar);
2863 } else if (tp->t_srtt != 0) {
2867 * srtt is stored as fixed point with 5 bits after the
2868 * binary point (i.e., scaled by 8). The following magic
2869 * is equivalent to the smoothing algorithm in rfc793 with
2870 * an alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed
2871 * point). Adjust rtt to origin 0.
2873 delta = ((rtt - 1) << TCP_DELTA_SHIFT)
2874 - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT));
2876 if ((tp->t_srtt += delta) <= 0)
2880 * We accumulate a smoothed rtt variance (actually, a
2881 * smoothed mean difference), then set the retransmit
2882 * timer to smoothed rtt + 4 times the smoothed variance.
2883 * rttvar is stored as fixed point with 4 bits after the
2884 * binary point (scaled by 16). The following is
2885 * equivalent to rfc793 smoothing with an alpha of .75
2886 * (rttvar = rttvar*3/4 + |delta| / 4). This replaces
2887 * rfc793's wired-in beta.
2891 delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT);
2892 if ((tp->t_rttvar += delta) <= 0)
2894 if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar)
2895 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2898 * No rtt measurement yet - use the unsmoothed rtt.
2899 * Set the variance to half the rtt (so our first
2900 * retransmit happens at 3*rtt).
2902 tp->t_srtt = rtt << TCP_RTT_SHIFT;
2903 tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1);
2904 tp->t_rttbest = tp->t_srtt + tp->t_rttvar;
2909 #ifdef DEBUG_EIFEL_RESPONSE
2911 kprintf("| rxtcur prev %d, old %d, ",
2912 tp->t_rxtcur_prev, tp->t_rxtcur);
2917 * the retransmit should happen at rtt + 4 * rttvar.
2918 * Because of the way we do the smoothing, srtt and rttvar
2919 * will each average +1/2 tick of bias. When we compute
2920 * the retransmit timer, we want 1/2 tick of rounding and
2921 * 1 extra tick because of +-1/2 tick uncertainty in the
2922 * firing of the timer. The bias will give us exactly the
2923 * 1.5 tick we need. But, because the bias is
2924 * statistical, we have to test that we don't drop below
2925 * the minimum feasible timer (which is 2 ticks).
2927 TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp),
2928 max(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX);
2931 if (tp->t_rxtcur < tp->t_rxtcur_prev + tcp_eifel_rtoinc) {
2933 * RFC4015 requires that the new RTO is at least
2934 * 2*G (tcp_eifel_rtoinc) greater then the RTO
2935 * (t_rxtcur_prev) when the spurious retransmit
2938 * The above condition could be true, if the SRTT
2939 * and RTTVAR used to calculate t_rxtcur_prev
2940 * resulted in a value less than t_rttmin. So
2941 * simply increasing SRTT by tcp_eifel_rtoinc when
2942 * preparing for the Eifel response could not ensure
2943 * that the new RTO will be tcp_eifel_rtoinc greater
2946 tp->t_rxtcur = tp->t_rxtcur_prev + tcp_eifel_rtoinc;
2948 #ifdef DEBUG_EIFEL_RESPONSE
2949 kprintf("new %d\n", tp->t_rxtcur);
2954 * We received an ack for a packet that wasn't retransmitted;
2955 * it is probably safe to discard any error indications we've
2956 * received recently. This isn't quite right, but close enough
2957 * for now (a route might have failed after we sent a segment,
2958 * and the return path might not be symmetrical).
2960 tp->t_softerror = 0;
2964 * Determine a reasonable value for maxseg size.
2965 * If the route is known, check route for mtu.
2966 * If none, use an mss that can be handled on the outgoing
2967 * interface without forcing IP to fragment; if bigger than
2968 * an mbuf cluster (MCLBYTES), round down to nearest multiple of MCLBYTES
2969 * to utilize large mbufs. If no route is found, route has no mtu,
2970 * or the destination isn't local, use a default, hopefully conservative
2971 * size (usually 512 or the default IP max size, but no more than the mtu
2972 * of the interface), as we can't discover anything about intervening
2973 * gateways or networks. We also initialize the congestion/slow start
2974 * window to be a single segment if the destination isn't local.
2975 * While looking at the routing entry, we also initialize other path-dependent
2976 * parameters from pre-set or cached values in the routing entry.
2978 * Also take into account the space needed for options that we
2979 * send regularly. Make maxseg shorter by that amount to assure
2980 * that we can send maxseg amount of data even when the options
2981 * are present. Store the upper limit of the length of options plus
2984 * NOTE that this routine is only called when we process an incoming
2985 * segment, for outgoing segments only tcp_mssopt is called.
2988 tcp_mss(struct tcpcb *tp, int offer)
2994 struct inpcb *inp = tp->t_inpcb;
2997 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
2998 size_t min_protoh = isipv6 ?
2999 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3000 sizeof(struct tcpiphdr);
3002 const boolean_t isipv6 = FALSE;
3003 const size_t min_protoh = sizeof(struct tcpiphdr);
3007 rt = tcp_rtlookup6(&inp->inp_inc);
3009 rt = tcp_rtlookup(&inp->inp_inc);
3011 tp->t_maxopd = tp->t_maxseg =
3012 (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3016 so = inp->inp_socket;
3019 * Offer == 0 means that there was no MSS on the SYN segment,
3020 * in this case we use either the interface mtu or tcp_mssdflt.
3022 * An offer which is too large will be cut down later.
3026 if (in6_localaddr(&inp->in6p_faddr)) {
3027 offer = ND_IFINFO(rt->rt_ifp)->linkmtu -
3030 offer = tcp_v6mssdflt;
3033 if (in_localaddr(inp->inp_faddr))
3034 offer = ifp->if_mtu - min_protoh;
3036 offer = tcp_mssdflt;
3041 * Prevent DoS attack with too small MSS. Round up
3042 * to at least minmss.
3044 * Sanity check: make sure that maxopd will be large
3045 * enough to allow some data on segments even is the
3046 * all the option space is used (40bytes). Otherwise
3047 * funny things may happen in tcp_output.
3049 offer = max(offer, tcp_minmss);
3050 offer = max(offer, 64);
3052 rt->rt_rmx.rmx_mssopt = offer;
3055 * While we're here, check if there's an initial rtt
3056 * or rttvar. Convert from the route-table units
3057 * to scaled multiples of the slow timeout timer.
3059 if (tp->t_srtt == 0 && (rtt = rt->rt_rmx.rmx_rtt)) {
3061 * XXX the lock bit for RTT indicates that the value
3062 * is also a minimum value; this is subject to time.
3064 if (rt->rt_rmx.rmx_locks & RTV_RTT)
3065 tp->t_rttmin = rtt / (RTM_RTTUNIT / hz);
3066 tp->t_srtt = rtt / (RTM_RTTUNIT / (hz * TCP_RTT_SCALE));
3067 tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE;
3068 tcpstat.tcps_usedrtt++;
3069 if (rt->rt_rmx.rmx_rttvar) {
3070 tp->t_rttvar = rt->rt_rmx.rmx_rttvar /
3071 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE));
3072 tcpstat.tcps_usedrttvar++;
3074 /* default variation is +- 1 rtt */
3076 tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE;
3078 TCPT_RANGESET(tp->t_rxtcur,
3079 ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1,
3080 tp->t_rttmin, TCPTV_REXMTMAX);
3084 * if there's an mtu associated with the route, use it
3085 * else, use the link mtu. Take the smaller of mss or offer
3088 if (rt->rt_rmx.rmx_mtu) {
3089 mss = rt->rt_rmx.rmx_mtu - min_protoh;
3092 mss = ND_IFINFO(rt->rt_ifp)->linkmtu - min_protoh;
3094 mss = ifp->if_mtu - min_protoh;
3096 mss = min(mss, offer);
3099 * maxopd stores the maximum length of data AND options
3100 * in a segment; maxseg is the amount of data in a normal
3101 * segment. We need to store this value (maxopd) apart
3102 * from maxseg, because now every segment carries options
3103 * and thus we normally have somewhat less data in segments.
3107 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
3108 ((tp->t_flags & TF_RCVD_TSTMP) == TF_RCVD_TSTMP))
3109 mss -= TCPOLEN_TSTAMP_APPA;
3111 #if (MCLBYTES & (MCLBYTES - 1)) == 0
3113 mss &= ~(MCLBYTES-1);
3116 mss = mss / MCLBYTES * MCLBYTES;
3119 * If there's a pipesize, change the socket buffer
3120 * to that size. Make the socket buffers an integral
3121 * number of mss units; if the mss is larger than
3122 * the socket buffer, decrease the mss.
3125 if ((bufsize = rt->rt_rmx.rmx_sendpipe) == 0)
3127 bufsize = so->so_snd.ssb_hiwat;
3131 bufsize = roundup(bufsize, mss);
3132 if (bufsize > sb_max)
3134 if (bufsize > so->so_snd.ssb_hiwat)
3135 ssb_reserve(&so->so_snd, bufsize, so, NULL);
3140 if ((bufsize = rt->rt_rmx.rmx_recvpipe) == 0)
3142 bufsize = so->so_rcv.ssb_hiwat;
3143 if (bufsize > mss) {
3144 bufsize = roundup(bufsize, mss);
3145 if (bufsize > sb_max)
3147 if (bufsize > so->so_rcv.ssb_hiwat) {
3148 lwkt_gettoken(&so->so_rcv.ssb_token);
3149 ssb_reserve(&so->so_rcv, bufsize, so, NULL);
3150 lwkt_reltoken(&so->so_rcv.ssb_token);
3155 * Set the slow-start flight size
3157 * NOTE: t_maxseg must have been configured!
3159 tp->snd_cwnd = tcp_initial_window(tp);
3161 if (rt->rt_rmx.rmx_ssthresh) {
3163 * There's some sort of gateway or interface
3164 * buffer limit on the path. Use this to set
3165 * the slow start threshhold, but set the
3166 * threshold to no less than 2*mss.
3168 tp->snd_ssthresh = max(2 * mss, rt->rt_rmx.rmx_ssthresh);
3169 tcpstat.tcps_usedssthresh++;
3174 * Determine the MSS option to send on an outgoing SYN.
3177 tcp_mssopt(struct tcpcb *tp)
3182 ((tp->t_inpcb->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3183 int min_protoh = isipv6 ?
3184 sizeof(struct ip6_hdr) + sizeof(struct tcphdr) :
3185 sizeof(struct tcpiphdr);
3187 const boolean_t isipv6 = FALSE;
3188 const size_t min_protoh = sizeof(struct tcpiphdr);
3192 rt = tcp_rtlookup6(&tp->t_inpcb->inp_inc);
3194 rt = tcp_rtlookup(&tp->t_inpcb->inp_inc);
3196 return (isipv6 ? tcp_v6mssdflt : tcp_mssdflt);
3198 return (rt->rt_ifp->if_mtu - min_protoh);
3202 * When a partial ack arrives, force the retransmission of the
3203 * next unacknowledged segment. Do not exit Fast Recovery.
3205 * Implement the Slow-but-Steady variant of NewReno by restarting the
3206 * the retransmission timer. Turn it off here so it can be restarted
3207 * later in tcp_output().
3210 tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th, int acked)
3212 tcp_seq old_snd_nxt = tp->snd_nxt;
3213 u_long ocwnd = tp->snd_cwnd;
3215 tcp_callout_stop(tp, tp->tt_rexmt);
3217 tp->snd_nxt = th->th_ack;
3218 /* Set snd_cwnd to one segment beyond acknowledged offset. */
3219 tp->snd_cwnd = tp->t_maxseg;
3220 tp->t_flags |= TF_ACKNOW;
3222 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3223 tp->snd_nxt = old_snd_nxt;
3224 /* partial window deflation */
3226 tp->snd_cwnd = ocwnd - acked + tp->t_maxseg;
3228 tp->snd_cwnd = tp->t_maxseg;
3232 * In contrast to the Slow-but-Steady NewReno variant,
3233 * we do not reset the retransmission timer for SACK retransmissions,
3234 * except when retransmitting snd_una.
3237 tcp_sack_rexmt(struct tcpcb *tp, boolean_t force)
3239 tcp_seq old_snd_nxt = tp->snd_nxt;
3240 u_long ocwnd = tp->snd_cwnd;
3242 int nseg = 0; /* consecutive new segments */
3243 int nseg_rexmt = 0; /* retransmitted segments */
3247 uint32_t unsacked = tcp_sack_first_unsacked_len(tp);
3250 * Try to fill the first hole in the receiver's
3253 maxrexmt = howmany(unsacked, tp->t_maxseg);
3254 if (maxrexmt > tcp_force_sackrxt)
3255 maxrexmt = tcp_force_sackrxt;
3259 pipe = tcp_sack_compute_pipe(tp);
3260 while (((tcp_seq_diff_t)(ocwnd - pipe) >= (tcp_seq_diff_t)tp->t_maxseg
3261 || (force && nseg_rexmt < maxrexmt && nseg == 0)) &&
3262 (!tcp_do_smartsack || nseg < TCP_SACK_MAXBURST)) {
3263 tcp_seq old_snd_max, old_rexmt_high, nextrexmt;
3264 uint32_t sent, seglen;
3268 old_rexmt_high = tp->rexmt_high;
3269 if (!tcp_sack_nextseg(tp, &nextrexmt, &seglen, &rescue)) {
3270 tp->rexmt_high = old_rexmt_high;
3275 * If the next tranmission is a rescue retranmission,
3276 * we check whether we have already sent some data
3277 * (either new segments or retransmitted segments)
3278 * into the the network or not. Since the idea of rescue
3279 * retransmission is to sustain ACK clock, as long as
3280 * some segments are in the network, ACK clock will be
3283 if (rescue && (nseg_rexmt > 0 || nseg > 0)) {
3284 tp->rexmt_high = old_rexmt_high;
3288 if (nextrexmt == tp->snd_max)
3292 tp->snd_nxt = nextrexmt;
3293 tp->snd_cwnd = nextrexmt - tp->snd_una + seglen;
3294 old_snd_max = tp->snd_max;
3295 if (nextrexmt == tp->snd_una)
3296 tcp_callout_stop(tp, tp->tt_rexmt);
3297 tp->t_flags |= TF_XMITNOW;
3298 error = tcp_output(tp);
3300 tp->rexmt_high = old_rexmt_high;
3303 sent = tp->snd_nxt - nextrexmt;
3305 tp->rexmt_high = old_rexmt_high;
3309 tcpstat.tcps_sndsackpack++;
3310 tcpstat.tcps_sndsackbyte += sent;
3313 tcpstat.tcps_sackrescue++;
3314 tp->rexmt_rescue = tp->snd_nxt;
3315 tp->sack_flags |= TSACK_F_SACKRESCUED;
3318 if (SEQ_LT(nextrexmt, old_snd_max) &&
3319 SEQ_LT(tp->rexmt_high, tp->snd_nxt)) {
3320 tp->rexmt_high = seq_min(tp->snd_nxt, old_snd_max);
3321 if (tcp_aggressive_rescuesack &&
3322 (tp->sack_flags & TSACK_F_SACKRESCUED) &&
3323 SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
3324 /* Drag RescueRxt along with HighRxt */
3325 tp->rexmt_rescue = tp->rexmt_high;
3329 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3330 tp->snd_nxt = old_snd_nxt;
3331 tp->snd_cwnd = ocwnd;
3335 * Return TRUE, if some new segments are sent
3338 tcp_sack_limitedxmit(struct tcpcb *tp)
3340 tcp_seq oldsndnxt = tp->snd_nxt;
3341 tcp_seq oldsndmax = tp->snd_max;
3342 u_long ocwnd = tp->snd_cwnd;
3343 uint32_t pipe, sent;
3344 boolean_t ret = FALSE;
3345 tcp_seq_diff_t cwnd_left;
3348 tp->rexmt_high = tp->snd_una - 1;
3349 pipe = tcp_sack_compute_pipe(tp);
3350 cwnd_left = (tcp_seq_diff_t)(ocwnd - pipe);
3351 if (cwnd_left < (tcp_seq_diff_t)tp->t_maxseg)
3354 if (tcp_do_smartsack)
3355 cwnd_left = ulmin(cwnd_left, tp->t_maxseg * TCP_SACK_MAXBURST);
3357 next = tp->snd_nxt = tp->snd_max;
3358 tp->snd_cwnd = tp->snd_nxt - tp->snd_una +
3359 rounddown(cwnd_left, tp->t_maxseg);
3361 tp->t_flags |= TF_XMITNOW;
3364 sent = tp->snd_nxt - next;
3366 tcpstat.tcps_sndlimited += howmany(sent, tp->t_maxseg);
3370 if (SEQ_LT(oldsndnxt, oldsndmax)) {
3371 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una),
3372 ("snd_una moved in other threads"));
3373 tp->snd_nxt = oldsndnxt;
3375 tp->snd_cwnd = ocwnd;
3377 if (ret && TCP_DO_NCR(tp))
3378 tcp_ncr_update_rxtthresh(tp);
3384 * Reset idle time and keep-alive timer, typically called when a valid
3385 * tcp packet is received but may also be called when FASTKEEP is set
3386 * to prevent the previous long-timeout from calculating to a drop.
3388 * Only update t_rcvtime for non-SYN packets.
3390 * Handle the case where one side thinks the connection is established
3391 * but the other side has, say, rebooted without cleaning out the
3392 * connection. The SYNs could be construed as an attack and wind
3393 * up ignored, but in case it isn't an attack we can validate the
3394 * connection by forcing a keepalive.
3397 tcp_timer_keep_activity(struct tcpcb *tp, int thflags)
3399 if (TCPS_HAVEESTABLISHED(tp->t_state)) {
3400 if ((thflags & (TH_SYN | TH_ACK)) == TH_SYN) {
3401 tp->t_flags |= TF_KEEPALIVE;
3402 tcp_callout_reset(tp, tp->tt_keep, hz / 2,
3405 tp->t_rcvtime = ticks;
3406 tp->t_flags &= ~TF_KEEPALIVE;
3407 tcp_callout_reset(tp, tp->tt_keep,
3415 tcp_rmx_msl(const struct tcpcb *tp)
3418 struct inpcb *inp = tp->t_inpcb;
3421 boolean_t isipv6 = ((inp->inp_vflag & INP_IPV6) ? TRUE : FALSE);
3423 const boolean_t isipv6 = FALSE;
3427 rt = tcp_rtlookup6(&inp->inp_inc);
3429 rt = tcp_rtlookup(&inp->inp_inc);
3430 if (rt == NULL || rt->rt_rmx.rmx_msl == 0)
3433 msl = (rt->rt_rmx.rmx_msl * hz) / 1000;
3441 tcp_established(struct tcpcb *tp)
3443 tp->t_state = TCPS_ESTABLISHED;
3444 tcp_callout_reset(tp, tp->tt_keep, tp->t_keepidle, tcp_timer_keep);
3446 if (tp->t_rxtsyn > 0) {
3449 * "If the timer expires awaiting the ACK of a SYN segment
3450 * and the TCP implementation is using an RTO less than 3
3451 * seconds, the RTO MUST be re-initialized to 3 seconds
3452 * when data transmission begins"
3454 if (tp->t_rxtcur < TCPTV_RTOBASE3)
3455 tp->t_rxtcur = TCPTV_RTOBASE3;
3460 * Returns TRUE, if the ACK should be dropped
3463 tcp_recv_dupack(struct tcpcb *tp, tcp_seq th_ack, u_int to_flags)
3465 boolean_t fast_sack_rexmt = TRUE;
3467 tcpstat.tcps_rcvdupack++;
3470 * We have outstanding data (other than a window probe),
3471 * this is a completely duplicate ack (ie, window info
3472 * didn't change), the ack is the biggest we've seen and
3473 * we've seen exactly our rexmt threshhold of them, so
3474 * assume a packet has been dropped and retransmit it.
3475 * Kludge snd_nxt & the congestion window so we send only
3478 if (IN_FASTRECOVERY(tp)) {
3479 if (TCP_DO_SACK(tp)) {
3480 boolean_t force = FALSE;
3482 if (tp->snd_una == tp->rexmt_high &&
3483 (to_flags & (TOF_SACK | TOF_SACK_REDUNDANT)) ==
3486 * New segments got SACKed and
3487 * no retransmit yet.
3492 /* No artifical cwnd inflation. */
3493 tcp_sack_rexmt(tp, force);
3496 * Dup acks mean that packets have left
3497 * the network (they're now cached at the
3498 * receiver) so bump cwnd by the amount in
3499 * the receiver to keep a constant cwnd
3500 * packets in the network.
3502 tp->snd_cwnd += tp->t_maxseg;
3506 } else if (SEQ_LT(th_ack, tp->snd_recover)) {
3509 } else if (tcp_ignore_redun_dsack && TCP_DO_SACK(tp) &&
3510 (to_flags & (TOF_DSACK | TOF_SACK_REDUNDANT)) ==
3511 (TOF_DSACK | TOF_SACK_REDUNDANT)) {
3513 * If the ACK carries DSACK and other SACK blocks
3514 * carry information that we have already known,
3515 * don't count this ACK as duplicate ACK. This
3516 * prevents spurious early retransmit and fast
3517 * retransmit. This also meets the requirement of
3518 * RFC3042 that new segments should not be sent if
3519 * the SACK blocks do not contain new information
3520 * (XXX we actually loosen the requirment that only
3521 * DSACK is checked here).
3523 * This kind of ACKs are usually sent after spurious
3526 /* Do nothing; don't change t_dupacks */
3528 } else if (tp->t_dupacks == 0 && TCP_DO_NCR(tp)) {
3529 tcp_ncr_update_rxtthresh(tp);
3532 if (++tp->t_dupacks == tp->t_rxtthresh) {
3533 tcp_seq old_snd_nxt;
3537 if (tcp_do_eifel_detect && (tp->t_flags & TF_RCVD_TSTMP)) {
3538 tcp_save_congestion_state(tp);
3539 tp->rxt_flags |= TRXT_F_FASTREXMT;
3542 * We know we're losing at the current window size,
3543 * so do congestion avoidance: set ssthresh to half
3544 * the current window and pull our congestion window
3545 * back to the new ssthresh.
3547 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_maxseg;
3550 tp->snd_ssthresh = win * tp->t_maxseg;
3551 ENTER_FASTRECOVERY(tp);
3552 tp->snd_recover = tp->snd_max;
3553 tcp_callout_stop(tp, tp->tt_rexmt);
3555 old_snd_nxt = tp->snd_nxt;
3556 tp->snd_nxt = th_ack;
3557 if (TCP_DO_SACK(tp)) {
3560 rxtlen = tcp_sack_first_unsacked_len(tp);
3561 if (rxtlen > tp->t_maxseg)
3562 rxtlen = tp->t_maxseg;
3563 tp->snd_cwnd = rxtlen;
3565 tp->snd_cwnd = tp->t_maxseg;
3568 ++tcpstat.tcps_sndfastrexmit;
3569 tp->snd_cwnd = tp->snd_ssthresh;
3570 tp->rexmt_high = tp->snd_nxt;
3571 tp->sack_flags &= ~TSACK_F_SACKRESCUED;
3572 if (SEQ_GT(old_snd_nxt, tp->snd_nxt))
3573 tp->snd_nxt = old_snd_nxt;
3574 KASSERT(tp->snd_limited <= 2, ("tp->snd_limited too big"));
3575 if (TCP_DO_SACK(tp)) {
3576 if (fast_sack_rexmt)
3577 tcp_sack_rexmt(tp, FALSE);
3579 tp->snd_cwnd += tp->t_maxseg *
3580 (tp->t_dupacks - tp->snd_limited);
3582 } else if ((tcp_do_rfc6675 && TCP_DO_SACK(tp)) || TCP_DO_NCR(tp)) {
3584 * The RFC6675 recommends to reduce the byte threshold,
3585 * and enter fast retransmit if IsLost(snd_una). However,
3586 * if we use IsLost(snd_una) based fast retransmit here,
3587 * segments reordering will cause spurious retransmit. So
3588 * we defer the IsLost(snd_una) based fast retransmit until
3589 * the extended limited transmit can't send any segments and
3590 * early retransmit can't be done.
3592 if (tcp_rfc6675_rxt && tcp_do_rfc6675 &&
3593 tcp_sack_islost(&tp->scb, tp->snd_una))
3594 goto fastretransmit;
3596 if (tcp_do_limitedtransmit || TCP_DO_NCR(tp)) {
3597 if (!tcp_sack_limitedxmit(tp)) {
3598 /* outstanding data */
3599 uint32_t ownd = tp->snd_max - tp->snd_una;
3601 if (need_early_retransmit(tp, ownd)) {
3602 ++tcpstat.tcps_sndearlyrexmit;
3603 tp->rxt_flags |= TRXT_F_EARLYREXMT;
3604 goto fastretransmit;
3605 } else if (tcp_do_rfc6675 &&
3606 tcp_sack_islost(&tp->scb, tp->snd_una)) {
3607 fast_sack_rexmt = FALSE;
3608 goto fastretransmit;
3612 } else if (tcp_do_limitedtransmit) {
3613 u_long oldcwnd = tp->snd_cwnd;
3614 tcp_seq oldsndmax = tp->snd_max;
3615 tcp_seq oldsndnxt = tp->snd_nxt;
3616 /* outstanding data */
3617 uint32_t ownd = tp->snd_max - tp->snd_una;
3620 KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2,
3621 ("dupacks not 1 or 2"));
3622 if (tp->t_dupacks == 1)
3623 tp->snd_limited = 0;
3624 tp->snd_nxt = tp->snd_max;
3625 tp->snd_cwnd = ownd +
3626 (tp->t_dupacks - tp->snd_limited) * tp->t_maxseg;
3627 tp->t_flags |= TF_XMITNOW;
3630 if (SEQ_LT(oldsndnxt, oldsndmax)) {
3631 KASSERT(SEQ_GEQ(oldsndnxt, tp->snd_una),
3632 ("snd_una moved in other threads"));
3633 tp->snd_nxt = oldsndnxt;
3635 tp->snd_cwnd = oldcwnd;
3636 sent = tp->snd_max - oldsndmax;
3637 if (sent > tp->t_maxseg) {
3638 KASSERT((tp->t_dupacks == 2 && tp->snd_limited == 0) ||
3639 (sent == tp->t_maxseg + 1 &&
3640 (tp->t_flags & TF_SENTFIN)),
3642 KASSERT(sent <= tp->t_maxseg * 2,
3643 ("sent too many segments"));
3644 tp->snd_limited = 2;
3645 tcpstat.tcps_sndlimited += 2;
3646 } else if (sent > 0) {
3648 ++tcpstat.tcps_sndlimited;
3649 } else if (need_early_retransmit(tp, ownd)) {
3650 ++tcpstat.tcps_sndearlyrexmit;
3651 tp->rxt_flags |= TRXT_F_EARLYREXMT;
3652 goto fastretransmit;
nant's projects / dragonfly.git / blob