]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_reass.c
contrib/tzdata: import tzdata 2022f
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_reass.c
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
5  *      The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
32  */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_tcpdebug.h"
40
41 /* For debugging we want counters and BB logging */
42 /* #define TCP_REASS_COUNTERS 1 */
43 /* #define TCP_REASS_LOGGING 1 */
44
45 #include <sys/param.h>
46 #include <sys/kernel.h>
47 #include <sys/eventhandler.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 #include <sys/systm.h>
55
56 #include <vm/uma.h>
57
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/route.h>
61 #include <net/vnet.h>
62
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/in_var.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip_var.h>
69 #include <netinet/ip_options.h>
70 #include <netinet/ip6.h>
71 #include <netinet6/in6_pcb.h>
72 #include <netinet6/ip6_var.h>
73 #include <netinet6/nd6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/tcp_seq.h>
77 #include <netinet/tcp_timer.h>
78 #include <netinet/tcp_var.h>
79 #ifdef TCP_REASS_LOGGING
80 #include <netinet/tcp_log_buf.h>
81 #include <netinet/tcp_hpts.h>
82 #endif
83 #include <netinet/tcpip.h>
84 #ifdef TCPDEBUG
85 #include <netinet/tcp_debug.h>
86 #endif /* TCPDEBUG */
87
88 #define TCP_R_LOG_ADD           1
89 #define TCP_R_LOG_LIMIT_REACHED 2
90 #define TCP_R_LOG_APPEND        3
91 #define TCP_R_LOG_PREPEND       4
92 #define TCP_R_LOG_REPLACE       5
93 #define TCP_R_LOG_MERGE_INTO    6
94 #define TCP_R_LOG_NEW_ENTRY     7
95 #define TCP_R_LOG_READ          8
96 #define TCP_R_LOG_ZERO          9
97 #define TCP_R_LOG_DUMP          10
98 #define TCP_R_LOG_TRIM          11
99
100 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass,
101     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
102     "TCP Segment Reassembly Queue");
103
104 static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats,
105     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
106     "TCP Segment Reassembly stats");
107
108 static int tcp_reass_maxseg = 0;
109 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
110     &tcp_reass_maxseg, 0,
111     "Global maximum number of TCP Segments in Reassembly Queue");
112
113 static uma_zone_t tcp_reass_zone;
114 SYSCTL_UMA_CUR(_net_inet_tcp_reass, OID_AUTO, cursegments, 0,
115     &tcp_reass_zone,
116     "Global number of TCP Segments currently in Reassembly Queue");
117
118 static u_int tcp_reass_maxqueuelen = 100;
119 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
120     &tcp_reass_maxqueuelen, 0,
121     "Maximum number of TCP Segments per Reassembly Queue");
122
123 static int tcp_new_limits = 0;
124 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN,
125     &tcp_new_limits, 0,
126     "Do we use the new limit method we are discussing?");
127
128 static u_int tcp_reass_queue_guard = 16;
129 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN,
130     &tcp_reass_queue_guard, 16,
131     "Number of TCP Segments in Reassembly Queue where we flip over to guard mode");
132
133 #ifdef TCP_REASS_COUNTERS
134
135 counter_u64_t reass_entry;
136 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD,
137     &reass_entry, "A segment entered reassembly ");
138
139 counter_u64_t reass_path1;
140 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD,
141     &reass_path1, "Took path 1");
142
143 counter_u64_t reass_path2;
144 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD,
145     &reass_path2, "Took path 2");
146
147 counter_u64_t reass_path3;
148 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD,
149     &reass_path3, "Took path 3");
150
151 counter_u64_t reass_path4;
152 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD,
153     &reass_path4, "Took path 4");
154
155 counter_u64_t reass_path5;
156 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD,
157     &reass_path5, "Took path 5");
158
159 counter_u64_t reass_path6;
160 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD,
161     &reass_path6, "Took path 6");
162
163 counter_u64_t reass_path7;
164 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD,
165     &reass_path7, "Took path 7");
166
167 counter_u64_t reass_fullwalk;
168 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD,
169     &reass_fullwalk, "Took a full walk ");
170
171 counter_u64_t reass_nospace;
172 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD,
173     &reass_nospace, "Had no mbuf capacity ");
174
175 counter_u64_t merge_fwd;
176 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD,
177     &merge_fwd, "Ran merge fwd");
178
179 counter_u64_t merge_into;
180 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD,
181     &merge_into, "Ran merge into");
182
183 counter_u64_t tcp_zero_input;
184 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD,
185     &tcp_zero_input, "The reassembly buffer saw a zero len segment etc");
186
187 #endif
188
189 /* Initialize TCP reassembly queue */
190 static void
191 tcp_reass_zone_change(void *tag)
192 {
193
194         /* Set the zone limit and read back the effective value. */
195         tcp_reass_maxseg = nmbclusters / 16;
196         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
197             tcp_reass_maxseg);
198 }
199
200 #ifdef TCP_REASS_LOGGING
201
202 static void
203 tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p,
204     tcp_seq seq, int len, uint8_t action, int instance)
205 {
206         uint32_t cts;
207         struct timeval tv;
208
209         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
210                 union tcp_log_stackspecific log;
211
212                 memset(&log, 0, sizeof(log));
213                 cts = tcp_get_usecs(&tv);
214                 log.u_bbr.flex1 = seq;
215                 log.u_bbr.cur_del_rate = (uint64_t)q;
216                 log.u_bbr.delRate = (uint64_t)p;
217                 if (q != NULL) {
218                         log.u_bbr.flex2 = q->tqe_start;
219                         log.u_bbr.flex3 = q->tqe_len;
220                         log.u_bbr.flex4 = q->tqe_mbuf_cnt;
221                         log.u_bbr.hptsi_gain = q->tqe_flags;
222                 }
223                 if (p != NULL)  {
224                         log.u_bbr.flex5 = p->tqe_start;
225                         log.u_bbr.pkts_out = p->tqe_len;
226                         log.u_bbr.epoch = p->tqe_mbuf_cnt;
227                         log.u_bbr.cwnd_gain = p->tqe_flags;
228                 }
229                 log.u_bbr.flex6 = tp->t_segqmbuflen;
230                 log.u_bbr.flex7 = instance;
231                 log.u_bbr.flex8 = action;
232                 log.u_bbr.timeStamp = cts;
233                 TCP_LOG_EVENTP(tp, NULL,
234                     &tp->t_inpcb->inp_socket->so_rcv,
235                     &tp->t_inpcb->inp_socket->so_snd,
236                     TCP_LOG_REASS, 0,
237                     len, &log, false, &tv);
238         }
239 }
240
241 static void
242 tcp_reass_log_dump(struct tcpcb *tp)
243 {
244         struct tseg_qent *q;
245
246         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
247                 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
248                         tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0);
249                 }
250         };
251 }
252
253 static void
254 tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m,
255     int logval, struct tseg_qent *q)
256 {
257         int cnt;
258         struct mbuf *t;
259
260         cnt = 0;
261         t = m;
262         while (t) {
263                 cnt += t->m_len;
264                 t = t->m_next;
265         }
266         tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt);
267 }
268
269 #endif
270
271 void
272 tcp_reass_global_init(void)
273 {
274
275         tcp_reass_maxseg = nmbclusters / 16;
276         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
277             &tcp_reass_maxseg);
278         tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
279             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
280         /* Set the zone limit and read back the effective value. */
281         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
282             tcp_reass_maxseg);
283 #ifdef TCP_REASS_COUNTERS
284         reass_path1 = counter_u64_alloc(M_WAITOK);
285         reass_path2 = counter_u64_alloc(M_WAITOK);
286         reass_path3 = counter_u64_alloc(M_WAITOK);
287         reass_path4 = counter_u64_alloc(M_WAITOK);
288         reass_path5 = counter_u64_alloc(M_WAITOK);
289         reass_path6 = counter_u64_alloc(M_WAITOK);
290         reass_path7 = counter_u64_alloc(M_WAITOK);
291         reass_fullwalk = counter_u64_alloc(M_WAITOK);
292         reass_nospace = counter_u64_alloc(M_WAITOK);
293         reass_entry = counter_u64_alloc(M_WAITOK);
294         merge_fwd = counter_u64_alloc(M_WAITOK);
295         merge_into = counter_u64_alloc(M_WAITOK);
296         tcp_zero_input = counter_u64_alloc(M_WAITOK);
297 #endif
298         EVENTHANDLER_REGISTER(nmbclusters_change,
299             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
300
301 }
302
303 void
304 tcp_reass_flush(struct tcpcb *tp)
305 {
306         struct tseg_qent *qe;
307
308         INP_WLOCK_ASSERT(tp->t_inpcb);
309
310         while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) {
311                 TAILQ_REMOVE(&tp->t_segq, qe, tqe_q);
312                 m_freem(qe->tqe_m);
313                 uma_zfree(tcp_reass_zone, qe);
314                 tp->t_segqlen--;
315         }
316         tp->t_segqmbuflen = 0;
317         KASSERT((tp->t_segqlen == 0),
318             ("TCP reass queue %p segment count is %d instead of 0 after flush.",
319             tp, tp->t_segqlen));
320 }
321
322 static void
323 tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last,
324     struct mbuf *m, struct tcphdr *th, int tlen,
325     struct mbuf *mlast, int lenofoh)
326 {
327
328 #ifdef TCP_REASS_LOGGING
329         tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0);
330 #endif
331         last->tqe_len += tlen;
332         last->tqe_m->m_pkthdr.len += tlen;
333         /* Preserve the FIN bit if its there */
334         last->tqe_flags |= (tcp_get_flags(th) & TH_FIN);
335         last->tqe_last->m_next = m;
336         last->tqe_last = mlast;
337         last->tqe_mbuf_cnt += lenofoh;
338         tp->t_rcvoopack++;
339         TCPSTAT_INC(tcps_rcvoopack);
340         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
341 #ifdef TCP_REASS_LOGGING
342         tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m,
343                              TCP_R_LOG_APPEND,
344                              last);
345 #endif
346 }
347
348 static void
349 tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th,
350                   int tlen, struct mbuf *mlast, int lenofoh)
351 {
352         int i;
353
354 #ifdef TCP_REASS_LOGGING
355         tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0);
356 #endif
357         if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) {
358                 /* The new data overlaps into the old */
359                 i = (th->th_seq + tlen) - first->tqe_start;
360 #ifdef TCP_REASS_LOGGING
361                 tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1);
362 #endif
363                 m_adj(first->tqe_m, i);
364                 first->tqe_len -= i;
365                 first->tqe_start += i;
366         }
367         /* Ok now setup our chain to point to the old first */
368         mlast->m_next = first->tqe_m;
369         first->tqe_m = m;
370         first->tqe_len += tlen;
371         first->tqe_start = th->th_seq;
372         first->tqe_m->m_pkthdr.len = first->tqe_len;
373         first->tqe_mbuf_cnt += lenofoh;
374         tp->t_rcvoopack++;
375         TCPSTAT_INC(tcps_rcvoopack);
376         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
377 #ifdef TCP_REASS_LOGGING
378         tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m,
379                              TCP_R_LOG_PREPEND,
380                              first);
381 #endif
382 }
383
384 static void
385 tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m,
386     tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint16_t flags)
387 {
388         /*
389          * Free the data in q, and replace
390          * it with the new segment.
391          */
392         int len_dif;
393
394 #ifdef TCP_REASS_LOGGING
395         tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0);
396 #endif
397         m_freem(q->tqe_m);
398         KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
399                 ("Tp:%p seg queue goes negative", tp));
400         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
401         q->tqe_mbuf_cnt = mbufoh;
402         q->tqe_m = m;
403         q->tqe_last = mlast;
404         q->tqe_start = seq;
405         if (len > q->tqe_len)
406                 len_dif = len - q->tqe_len;
407         else
408                 len_dif = 0;
409         tp->t_rcvoopack++;
410         TCPSTAT_INC(tcps_rcvoopack);
411         TCPSTAT_ADD(tcps_rcvoobyte, len_dif);
412         q->tqe_len = len;
413         q->tqe_flags = (flags & TH_FIN);
414         q->tqe_m->m_pkthdr.len = q->tqe_len;
415         tp->t_segqmbuflen += mbufoh;
416
417 }
418
419 static void
420 tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent,
421     struct tseg_qent *q)
422 {
423         /*
424          * Merge q into ent and free q from the list.
425          */
426 #ifdef TCP_REASS_LOGGING
427         tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0);
428 #endif
429 #ifdef TCP_REASS_COUNTERS
430         counter_u64_add(merge_into, 1);
431 #endif
432         ent->tqe_last->m_next = q->tqe_m;
433         ent->tqe_last = q->tqe_last;
434         ent->tqe_len += q->tqe_len;
435         ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt;
436         ent->tqe_m->m_pkthdr.len += q->tqe_len;
437         ent->tqe_flags |= (q->tqe_flags & TH_FIN);
438         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
439         uma_zfree(tcp_reass_zone, q);
440         tp->t_segqlen--;
441
442 }
443
444 static void
445 tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent)
446 {
447         struct tseg_qent *q, *qtmp;
448         int i;
449         tcp_seq max;
450         /*
451          * Given an entry merge forward anyplace
452          * that ent overlaps forward.
453          */
454
455         max = ent->tqe_start + ent->tqe_len;
456         q = TAILQ_NEXT(ent, tqe_q);
457         if (q == NULL) {
458                 /* Nothing left */
459                 return;
460         }
461         TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) {
462                 if (SEQ_GT(q->tqe_start, max)) {
463                         /* Beyond q */
464                         break;
465                 }
466                 /* We have some or all that are overlapping */
467                 if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) {
468                         /* It consumes it all */
469                         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
470                         m_freem(q->tqe_m);
471                         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
472                         uma_zfree(tcp_reass_zone, q);
473                         tp->t_segqlen--;
474                         continue;
475                 }
476                 /*
477                  * Trim the q entry to dovetail to this one
478                  * and then merge q into ent updating max
479                  * in the process.
480                  */
481                 i = max - q->tqe_start;
482 #ifdef TCP_REASS_LOGGING
483                 tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2);
484 #endif
485                 m_adj(q->tqe_m, i);
486                 q->tqe_len -= i;
487                 q->tqe_start += i;
488                 tcp_reass_merge_into(tp, ent, q);
489                 max = ent->tqe_start + ent->tqe_len;
490         }
491 #ifdef TCP_REASS_COUNTERS
492         counter_u64_add(merge_fwd, 1);
493 #endif
494 }
495
496 static int
497 tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast)
498 {
499         int len = MSIZE;
500
501         if (m->m_flags & M_EXT)
502                 len += m->m_ext.ext_size;
503         while (m->m_next != NULL) {
504                 m = m->m_next;
505                 len += MSIZE;
506                 if (m->m_flags & M_EXT)
507                         len += m->m_ext.ext_size;
508         }
509         *mlast = m;
510         return (len);
511 }
512
513 /*
514  * NOTE!!! the new tcp-reassembly code *must not* use
515  * m_adj() with a negative index. That alters the chain
516  * of mbufs (by possibly chopping trailing mbufs). At
517  * the front of tcp_reass we count the mbuf overhead
518  * and setup the tail pointer. If we use m_adj(m, -5)
519  * we could corrupt the tail pointer. Currently the
520  * code only uses m_adj(m, postive-num). If this
521  * changes appropriate changes to update mlast would
522  * be needed.
523  */
524 int
525 tcp_reass(struct tcpcb *tp, struct tcphdr *th, tcp_seq *seq_start,
526           int *tlenp, struct mbuf *m)
527 {
528         struct tseg_qent *q, *last, *first;
529         struct tseg_qent *p = NULL;
530         struct tseg_qent *nq = NULL;
531         struct tseg_qent *te = NULL;
532         struct mbuf *mlast = NULL;
533         struct sockbuf *sb;
534         struct socket *so = tp->t_inpcb->inp_socket;
535         char *s = NULL;
536         int flags, i, lenofoh;
537
538         INP_WLOCK_ASSERT(tp->t_inpcb);
539         /*
540          * XXX: tcp_reass() is rather inefficient with its data structures
541          * and should be rewritten (see NetBSD for optimizations).
542          */
543
544         KASSERT(th == NULL || (seq_start != NULL && tlenp != NULL),
545                 ("tcp_reass called with illegal parameter combination "
546                  "(tp=%p, th=%p, seq_start=%p, tlenp=%p, m=%p)",
547                  tp, th, seq_start, tlenp, m));
548         /*
549          * Call with th==NULL after become established to
550          * force pre-ESTABLISHED data up to user socket.
551          */
552         if (th == NULL)
553                 goto present;
554         KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt),
555                 ("Attempt to add old entry to reassembly queue (th=%p, tp=%p)",
556                  th, tp));
557 #ifdef TCP_REASS_LOGGING
558         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL);
559 #endif
560 #ifdef TCP_REASS_COUNTERS
561         counter_u64_add(reass_entry, 1);
562 #endif
563         /*
564          * Check for zero length data.
565          */
566         if ((*tlenp == 0) && ((tcp_get_flags(th) & TH_FIN) == 0)) {
567                 /*
568                  * A zero length segment does no
569                  * one any good. We could check
570                  * the rcv_nxt <-> rcv_wnd but thats
571                  * already done for us by the caller.
572                  */
573 strip_fin:
574 #ifdef TCP_REASS_COUNTERS
575                 counter_u64_add(tcp_zero_input, 1);
576 #endif
577                 m_freem(m);
578 #ifdef TCP_REASS_LOGGING
579                 tcp_reass_log_dump(tp);
580 #endif
581                 return (0);
582         } else if ((*tlenp == 0) &&
583                    (tcp_get_flags(th) & TH_FIN) &&
584                    !TCPS_HAVEESTABLISHED(tp->t_state)) {
585                 /*
586                  * We have not established, and we
587                  * have a FIN and no data. Lets treat
588                  * this as the same as if the FIN were
589                  * not present. We don't want to save
590                  * the FIN bit in a reassembly buffer
591                  * we want to get established first before
592                  * we do that (the peer will retransmit).
593                  */
594                 goto strip_fin;
595         }
596         /*
597          * Will it fit?
598          */
599         lenofoh = tcp_reass_overhead_of_chain(m, &mlast);
600         sb = &tp->t_inpcb->inp_socket->so_rcv;
601         if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
602             (sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) {
603                 /* No room */
604                 TCPSTAT_INC(tcps_rcvreassfull);
605 #ifdef TCP_REASS_COUNTERS
606                 counter_u64_add(reass_nospace, 1);
607 #endif
608 #ifdef TCP_REASS_LOGGING
609                 tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0);
610 #endif
611                 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
612                         log(LOG_DEBUG, "%s; %s: mbuf count limit reached, "
613                             "segment dropped\n", s, __func__);
614                         free(s, M_TCPLOG);
615                 }
616                 m_freem(m);
617                 *tlenp = 0;
618 #ifdef TCP_REASS_LOGGING
619                 tcp_reass_log_dump(tp);
620 #endif
621                 return (0);
622         }
623         /*
624          * First lets deal with two common cases, the
625          * segment appends to the back of our collected
626          * segments. Or the segment is the next in line.
627          */
628         last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q);
629         if (last != NULL) {
630                 if ((tcp_get_flags(th) & TH_FIN) &&
631                     SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) {
632                         /*
633                          * Someone is trying to game us, dump
634                          * the segment.
635                          */
636                         *tlenp = 0;
637                         m_freem(m);
638                         return (0);
639                 }
640                 if ((SEQ_GEQ(th->th_seq, last->tqe_start)) &&
641                     (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) {
642                         /* Common case, trailing segment is added */
643                         /**
644                          *                                 +--last
645                          *                                 v
646                          *  reassembly buffer |---|  |---| |---|
647                          *  new segment                       |---|
648                          */
649 #ifdef TCP_REASS_COUNTERS
650                         counter_u64_add(reass_path1, 1);
651 #endif
652                         if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) {
653                                 i = (last->tqe_start + last->tqe_len) - th->th_seq;
654                                 if (i < *tlenp) {
655 #ifdef TCP_REASS_LOGGING
656                                         tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3);
657                                         th->th_seq += i;
658 #endif
659                                         m_adj(m, i);
660                                         *tlenp -= i;
661                                 } else {
662                                         /* Complete overlap */
663                                         TCPSTAT_INC(tcps_rcvduppack);
664                                         TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
665                                         m_freem(m);
666                                         *tlenp = last->tqe_len;
667                                         *seq_start = last->tqe_start;
668                                         return (0);
669                                 }
670                         }
671                         if (last->tqe_flags & TH_FIN) {
672                                 /*
673                                  * We have data after the FIN on the last?
674                                  */
675                                 *tlenp = 0;
676                                 m_freem(m);
677                                 return(0);
678                         }
679                         tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh);
680                         tp->t_segqmbuflen += lenofoh;
681                         *seq_start = last->tqe_start;
682                         *tlenp = last->tqe_len;
683                         return (0);
684                 } else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) {
685                         /*
686                          * Second common case, we missed
687                          * another one and have something more
688                          * for the end.
689                          */
690                         /**
691                          *                                 +--last
692                          *                                 v
693                          *  reassembly buffer |---|  |---| |---|
694                          *  new segment                           |---|
695                          */
696                         if (last->tqe_flags & TH_FIN) {
697                                 /*
698                                  * We have data after the FIN on the last?
699                                  */
700                                 *tlenp = 0;
701                                 m_freem(m);
702                                 return(0);
703                         }
704 #ifdef TCP_REASS_COUNTERS
705                         counter_u64_add(reass_path2, 1);
706 #endif
707                         p = last;
708                         goto new_entry;
709                 }
710         } else {
711                 /* First segment (it's NULL). */
712                 goto new_entry;
713         }
714         first = TAILQ_FIRST(&tp->t_segq);
715         if (SEQ_LT(th->th_seq, first->tqe_start) &&
716             SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) &&
717             SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) {
718                 /*
719                  * The head of the queue is prepended by this and
720                  * it may be the one I want most.
721                  */
722                 /**
723                  *       first-------+
724                  *                   v
725                  *  rea:             |---|  |---| |---|
726                  *  new:         |---|
727                  * Note the case we do not deal with here is:
728                  *   rea=     |---|   |---|   |---|
729                  *   new=  |----|
730                  * Due to the fact that it could be
731                  *   new   |--------------------|
732                  * And we might need to merge forward.
733                  */
734 #ifdef INVARIANTS
735                 struct mbuf *firstmbuf;
736 #endif
737
738 #ifdef TCP_REASS_COUNTERS
739                 counter_u64_add(reass_path3, 1);
740 #endif
741                 if (SEQ_LT(th->th_seq, tp->rcv_nxt)) {
742                         /*
743                          * The resend was even before
744                          * what we have. We need to trim it.
745                          * Note TSNH (it should be trimmed
746                          * before the call to tcp_reass()).
747                          */
748 #ifdef INVARIANTS
749                         panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed",
750                               th->th_seq, tp->rcv_nxt, tp);
751 #else
752                         i = tp->rcv_nxt - th->th_seq;
753 #ifdef TCP_REASS_LOGGING
754                         tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4);
755 #endif
756                         m_adj(m, i);
757                         th->th_seq += i;
758                         *tlenp -= i;
759 #endif
760                 }
761 #ifdef INVARIANTS
762                 firstmbuf = first->tqe_m;
763 #endif
764                 tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh);
765 #ifdef INVARIANTS
766                 if (firstmbuf == first->tqe_m) {
767                         panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
768                               m, firstmbuf, first->tqe_m, tp, first);
769                 } else if (first->tqe_m != m) {
770                         panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
771                               m, firstmbuf, first->tqe_m, tp, first);
772                 }
773 #endif
774                 tp->t_segqmbuflen += lenofoh;
775                 *seq_start = first->tqe_start;
776                 *tlenp = first->tqe_len;
777                 goto present;
778         } else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) {
779                 /* New segment is before our earliest segment. */
780                 /**
781                  *           first---->+
782                  *                      v
783                  *  rea=                |---| ....
784                  *  new"         |---|
785                  *
786                  */
787                 goto new_entry;
788         }
789         /*
790          * Find a segment which begins after this one does.
791          */
792 #ifdef TCP_REASS_COUNTERS
793         counter_u64_add(reass_fullwalk, 1);
794 #endif
795         TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
796                 if (SEQ_GT(q->tqe_start, th->th_seq))
797                         break;
798         }
799         p = TAILQ_PREV(q, tsegqe_head, tqe_q);
800         /**
801          * Now is this fit just in-between only?
802          * i.e.:
803          *      p---+        +----q
804          *          v        v
805          *     res= |--|     |--|    |--|
806          *     nee       |-|
807          */
808         if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) &&
809             ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) {
810                 /* Yep no overlap */
811                 goto new_entry;
812         }
813         /**
814          * If we reach here we have some (possibly all) overlap
815          * such as:
816          *     res=     |--|     |--|    |--|
817          *     new=  |----|
818          * or  new=  |-----------------|
819          * or  new=      |--------|
820          * or  new=            |---|
821          * or  new=            |-----------|
822          */
823         if ((p != NULL) &&
824             (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) {
825                 /* conversion to int (in i) handles seq wraparound */
826
827 #ifdef TCP_REASS_COUNTERS
828                 counter_u64_add(reass_path4, 1);
829 #endif
830                 i = p->tqe_start + p->tqe_len - th->th_seq;
831                 if (i >= 0) {
832                         if (i >= *tlenp) {
833                                 /**
834                                  *       prev seg---->+
835                                  *                    v
836                                  *  reassembly buffer |---|
837                                  *  new segment        |-|
838                                  */
839                                 TCPSTAT_INC(tcps_rcvduppack);
840                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
841                                 *tlenp = p->tqe_len;
842                                 *seq_start = p->tqe_start;
843                                 m_freem(m);
844                                 /*
845                                  * Try to present any queued data
846                                  * at the left window edge to the user.
847                                  * This is needed after the 3-WHS
848                                  * completes. Note this probably
849                                  * will not work and we will return.
850                                  */
851                                 return (0);
852                         }
853                         if (i > 0) {
854                                 /**
855                                  *       prev seg---->+
856                                  *                    v
857                                  *  reassembly buffer |---|
858                                  *  new segment         |-----|
859                                  */
860 #ifdef TCP_REASS_COUNTERS
861                                 counter_u64_add(reass_path5, 1);
862 #endif
863 #ifdef TCP_REASS_LOGGING
864                                 tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5);
865 #endif
866                                 m_adj(m, i);
867                                 *tlenp -= i;
868                                 th->th_seq += i;
869                         }
870                 }
871                 if (th->th_seq == (p->tqe_start + p->tqe_len)) {
872                         /*
873                          * If dovetails in with this one
874                          * append it.
875                          */
876                         /**
877                          *       prev seg---->+
878                          *                    v
879                          *  reassembly buffer |--|     |---|
880                          *  new segment          |--|
881                          * (note: it was trimmed above if it overlapped)
882                          */
883                         tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh);
884                         tp->t_segqmbuflen += lenofoh;
885                 } else {
886 #ifdef INVARIANTS
887                         panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p",
888                               th->th_seq, p->tqe_start, p->tqe_len,
889                               p, tp);
890 #endif
891                         *tlenp = 0;
892                         m_freem(m);
893                         return (0);
894                 }
895                 q = p;
896         } else {
897                 /*
898                  * The new data runs over the
899                  * top of previously sack'd data (in q).
900                  * It may be partially overlapping, or
901                  * it may overlap the entire segment.
902                  */
903 #ifdef TCP_REASS_COUNTERS
904                 counter_u64_add(reass_path6, 1);
905 #endif
906                 if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) {
907                         /* It consumes it all */
908                         /**
909                          *             next seg---->+
910                          *                          v
911                          *  reassembly buffer |--|     |---|
912                          *  new segment              |----------|
913                          */
914 #ifdef TCP_REASS_COUNTERS
915                         counter_u64_add(reass_path7, 1);
916 #endif
917                         tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, tcp_get_flags(th));
918                 } else {
919                         /*
920                          * We just need to prepend the data
921                          * to this. It does not overrun
922                          * the end.
923                          */
924                         /**
925                          *                next seg---->+
926                          *                             v
927                          *  reassembly buffer |--|     |---|
928                          *  new segment                   |----------|
929                          */
930                         tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh);
931                         tp->t_segqmbuflen += lenofoh;
932                 }
933         }
934         /* Now does it go further than that? */
935         tcp_reass_merge_forward(tp, q);
936         *seq_start = q->tqe_start;
937         *tlenp = q->tqe_len;
938         goto present;
939
940         /*
941          * When we reach here we can't combine it
942          * with any existing segment.
943          *
944          * Limit the number of segments that can be queued to reduce the
945          * potential for mbuf exhaustion. For best performance, we want to be
946          * able to queue a full window's worth of segments. The size of the
947          * socket receive buffer determines our advertised window and grows
948          * automatically when socket buffer autotuning is enabled. Use it as the
949          * basis for our queue limit.
950          *
951          * However, allow the user to specify a ceiling for the number of
952          * segments in each queue.
953          *
954          * Always let the missing segment through which caused this queue.
955          * NB: Access to the socket buffer is left intentionally unlocked as we
956          * can tolerate stale information here.
957          *
958          * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
959          * should work but causes packets to be dropped when they shouldn't.
960          * Investigate why and re-evaluate the below limit after the behaviour
961          * is understood.
962          */
963 new_entry:
964         if (th->th_seq == tp->rcv_nxt && TCPS_HAVEESTABLISHED(tp->t_state)) {
965                 tp->rcv_nxt += *tlenp;
966                 flags = tcp_get_flags(th) & TH_FIN;
967                 TCPSTAT_INC(tcps_rcvoopack);
968                 TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
969                 SOCKBUF_LOCK(&so->so_rcv);
970                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
971                         m_freem(m);
972                 } else {
973                         sbappendstream_locked(&so->so_rcv, m, 0);
974                 }
975                 tp->t_flags |= TF_WAKESOR;
976                 return (flags);
977         }
978         if (tcp_new_limits) {
979                 if ((tp->t_segqlen > tcp_reass_queue_guard) &&
980                     (*tlenp < MSIZE)) {
981                         /*
982                          * This is really a lie, we are not full but
983                          * are getting a segment that is above
984                          * guard threshold. If it is and its below
985                          * a mbuf size (256) we drop it if it
986                          * can't fill in some place.
987                          */
988                         TCPSTAT_INC(tcps_rcvreassfull);
989                         *tlenp = 0;
990                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
991                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
992                                     "segment dropped\n", s, __func__);
993                                 free(s, M_TCPLOG);
994                         }
995                         m_freem(m);
996 #ifdef TCP_REASS_LOGGING
997                         tcp_reass_log_dump(tp);
998 #endif
999                         return (0);
1000                 }
1001         } else {
1002                 if (tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
1003                                          tcp_reass_maxqueuelen)) {
1004                         TCPSTAT_INC(tcps_rcvreassfull);
1005                         *tlenp = 0;
1006                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
1007                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
1008                                     "segment dropped\n", s, __func__);
1009                                 free(s, M_TCPLOG);
1010                         }
1011                         m_freem(m);
1012 #ifdef TCP_REASS_LOGGING
1013                         tcp_reass_log_dump(tp);
1014 #endif
1015                         return (0);
1016                 }
1017         }
1018         /*
1019          * Allocate a new queue entry. If we can't, or hit the zone limit
1020          * just drop the pkt.
1021          */
1022         te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
1023         if (te == NULL) {
1024                 TCPSTAT_INC(tcps_rcvmemdrop);
1025                 m_freem(m);
1026                 *tlenp = 0;
1027                 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
1028                                        NULL))) {
1029                         log(LOG_DEBUG, "%s; %s: global zone limit "
1030                             "reached, segment dropped\n", s, __func__);
1031                         free(s, M_TCPLOG);
1032                 }
1033                 return (0);
1034         }
1035         tp->t_segqlen++;
1036         tp->t_rcvoopack++;
1037         TCPSTAT_INC(tcps_rcvoopack);
1038         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
1039         /* Insert the new segment queue entry into place. */
1040         te->tqe_m = m;
1041         te->tqe_flags = tcp_get_flags(th);
1042         te->tqe_len = *tlenp;
1043         te->tqe_start = th->th_seq;
1044         te->tqe_last = mlast;
1045         te->tqe_mbuf_cnt = lenofoh;
1046         tp->t_segqmbuflen += te->tqe_mbuf_cnt;
1047         if (p == NULL) {
1048                 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
1049         } else {
1050                 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
1051         }
1052 #ifdef TCP_REASS_LOGGING
1053         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_NEW_ENTRY, te);
1054 #endif
1055 present:
1056         /*
1057          * Present data to user, advancing rcv_nxt through
1058          * completed sequence space.
1059          */
1060         if (!TCPS_HAVEESTABLISHED(tp->t_state))
1061                 return (0);
1062         q = TAILQ_FIRST(&tp->t_segq);
1063         KASSERT(q == NULL || SEQ_GEQ(q->tqe_start, tp->rcv_nxt),
1064                 ("Reassembly queue for %p has stale entry at head", tp));
1065         if (!q || q->tqe_start != tp->rcv_nxt) {
1066 #ifdef TCP_REASS_LOGGING
1067                 tcp_reass_log_dump(tp);
1068 #endif
1069                 return (0);
1070         }
1071         SOCKBUF_LOCK(&so->so_rcv);
1072         do {
1073                 tp->rcv_nxt += q->tqe_len;
1074                 flags = q->tqe_flags & TH_FIN;
1075                 nq = TAILQ_NEXT(q, tqe_q);
1076                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1077                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
1078                         m_freem(q->tqe_m);
1079                 } else {
1080 #ifdef TCP_REASS_LOGGING
1081                         tcp_reass_log_new_in(tp, q->tqe_start, q->tqe_len, q->tqe_m, TCP_R_LOG_READ, q);
1082                         if (th != NULL) {
1083                                 tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 1);
1084                         } else {
1085                                 tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 1);
1086                         }
1087 #endif
1088                         sbappendstream_locked(&so->so_rcv, q->tqe_m, 0);
1089                 }
1090 #ifdef TCP_REASS_LOGGING
1091                 if (th != NULL) {
1092                         tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 2);
1093                 } else {
1094                         tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 2);
1095                 }
1096 #endif
1097                 KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
1098                         ("tp:%p seg queue goes negative", tp));
1099                 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
1100                 uma_zfree(tcp_reass_zone, q);
1101                 tp->t_segqlen--;
1102                 q = nq;
1103         } while (q && q->tqe_start == tp->rcv_nxt);
1104         if (TAILQ_EMPTY(&tp->t_segq) &&
1105             (tp->t_segqmbuflen != 0)) {
1106 #ifdef INVARIANTS
1107                 panic("tp:%p segq:%p len:%d queue empty",
1108                       tp, &tp->t_segq, tp->t_segqmbuflen);
1109 #else
1110 #ifdef TCP_REASS_LOGGING
1111                 if (th != NULL) {
1112                         tcp_log_reassm(tp, NULL, NULL, th->th_seq, *tlenp, TCP_R_LOG_ZERO, 0);
1113                 } else {
1114                         tcp_log_reassm(tp, NULL, NULL, 0, 0, TCP_R_LOG_ZERO, 0);
1115                 }
1116 #endif
1117                 tp->t_segqmbuflen = 0;
1118 #endif
1119         }
1120 #ifdef TCP_REASS_LOGGING
1121         tcp_reass_log_dump(tp);
1122 #endif
1123         tp->t_flags |= TF_WAKESOR;
1124         return (flags);
1125 }