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 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_tcpdebug.h"
38
39 /* For debugging we want counters and BB logging */
40 /* #define TCP_REASS_COUNTERS 1 */
41 /* #define TCP_REASS_LOGGING 1 */
42
43 #include <sys/param.h>
44 #include <sys/kernel.h>
45 #include <sys/eventhandler.h>
46 #include <sys/malloc.h>
47 #include <sys/mbuf.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/systm.h>
53
54 #include <vm/uma.h>
55
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/route.h>
59 #include <net/vnet.h>
60
61 #include <netinet/in.h>
62 #include <netinet/in_pcb.h>
63 #include <netinet/in_systm.h>
64 #include <netinet/in_var.h>
65 #include <netinet/ip.h>
66 #include <netinet/ip_var.h>
67 #include <netinet/ip_options.h>
68 #include <netinet/ip6.h>
69 #include <netinet6/in6_pcb.h>
70 #include <netinet6/ip6_var.h>
71 #include <netinet6/nd6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/tcp_fsm.h>
74 #include <netinet/tcp_seq.h>
75 #include <netinet/tcp_timer.h>
76 #include <netinet/tcp_var.h>
77 #ifdef TCP_REASS_LOGGING
78 #include <netinet/tcp_log_buf.h>
79 #include <netinet/tcp_hpts.h>
80 #endif
81 #include <netinet6/tcp6_var.h>
82 #include <netinet/tcpip.h>
83 #ifdef TCPDEBUG
84 #include <netinet/tcp_debug.h>
85 #endif /* TCPDEBUG */
86
87 #define TCP_R_LOG_ADD 1
88 #define TCP_R_LOG_LIMIT_REACHED 2
89 #define TCP_R_LOG_APPEND 3
90 #define TCP_R_LOG_PREPEND 4
91 #define TCP_R_LOG_REPLACE 5
92 #define TCP_R_LOG_MERGE_INTO 6
93 #define TCP_R_LOG_NEW_ENTRY 7
94 #define TCP_R_LOG_READ 8
95 #define TCP_R_LOG_ZERO 9
96 #define TCP_R_LOG_DUMP 10
97 #define TCP_R_LOG_TRIM 11
98
99 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass,
100 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
101 "TCP Segment Reassembly Queue");
102
103 static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats,
104 CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
105 "TCP Segment Reassembly stats");
106
107 static int tcp_reass_maxseg = 0;
108 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
109 &tcp_reass_maxseg, 0,
110 "Global maximum number of TCP Segments in Reassembly Queue");
111
112 static uma_zone_t tcp_reass_zone;
113 SYSCTL_UMA_CUR(_net_inet_tcp_reass, OID_AUTO, cursegments, 0,
114 &tcp_reass_zone,
115 "Global number of TCP Segments currently in Reassembly Queue");
116
117 static u_int tcp_reass_maxqueuelen = 100;
118 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
119 &tcp_reass_maxqueuelen, 0,
120 "Maximum number of TCP Segments per Reassembly Queue");
121
122 static int tcp_new_limits = 0;
123 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN,
124 &tcp_new_limits, 0,
125 "Do we use the new limit method we are discussing?");
126
127 static u_int tcp_reass_queue_guard = 16;
128 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN,
129 &tcp_reass_queue_guard, 16,
130 "Number of TCP Segments in Reassembly Queue where we flip over to guard mode");
131
132 #ifdef TCP_REASS_COUNTERS
133
134 counter_u64_t reass_entry;
135 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD,
136 &reass_entry, "A segment entered reassembly ");
137
138 counter_u64_t reass_path1;
139 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD,
140 &reass_path1, "Took path 1");
141
142 counter_u64_t reass_path2;
143 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD,
144 &reass_path2, "Took path 2");
145
146 counter_u64_t reass_path3;
147 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD,
148 &reass_path3, "Took path 3");
149
150 counter_u64_t reass_path4;
151 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD,
152 &reass_path4, "Took path 4");
153
154 counter_u64_t reass_path5;
155 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD,
156 &reass_path5, "Took path 5");
157
158 counter_u64_t reass_path6;
159 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD,
160 &reass_path6, "Took path 6");
161
162 counter_u64_t reass_path7;
163 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD,
164 &reass_path7, "Took path 7");
165
166 counter_u64_t reass_fullwalk;
167 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD,
168 &reass_fullwalk, "Took a full walk ");
169
170 counter_u64_t reass_nospace;
171 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD,
172 &reass_nospace, "Had no mbuf capacity ");
173
174 counter_u64_t merge_fwd;
175 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD,
176 &merge_fwd, "Ran merge fwd");
177
178 counter_u64_t merge_into;
179 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD,
180 &merge_into, "Ran merge into");
181
182 counter_u64_t tcp_zero_input;
183 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD,
184 &tcp_zero_input, "The reassembly buffer saw a zero len segment etc");
185
186 #endif
187
188 /* Initialize TCP reassembly queue */
189 static void
tcp_reass_zone_change(void * tag)190 tcp_reass_zone_change(void *tag)
191 {
192
193 /* Set the zone limit and read back the effective value. */
194 tcp_reass_maxseg = nmbclusters / 16;
195 tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
196 tcp_reass_maxseg);
197 }
198
199 #ifdef TCP_REASS_LOGGING
200
201 static void
tcp_log_reassm(struct tcpcb * tp,struct tseg_qent * q,struct tseg_qent * p,tcp_seq seq,int len,uint8_t action,int instance)202 tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p,
203 tcp_seq seq, int len, uint8_t action, int instance)
204 {
205 uint32_t cts;
206 struct timeval tv;
207
208 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
209 union tcp_log_stackspecific log;
210
211 memset(&log, 0, sizeof(log));
212 cts = tcp_get_usecs(&tv);
213 log.u_bbr.flex1 = seq;
214 log.u_bbr.cur_del_rate = (uint64_t)q;
215 log.u_bbr.delRate = (uint64_t)p;
216 if (q != NULL) {
217 log.u_bbr.flex2 = q->tqe_start;
218 log.u_bbr.flex3 = q->tqe_len;
219 log.u_bbr.flex4 = q->tqe_mbuf_cnt;
220 log.u_bbr.hptsi_gain = q->tqe_flags;
221 }
222 if (p != NULL) {
223 log.u_bbr.flex5 = p->tqe_start;
224 log.u_bbr.pkts_out = p->tqe_len;
225 log.u_bbr.epoch = p->tqe_mbuf_cnt;
226 log.u_bbr.cwnd_gain = p->tqe_flags;
227 }
228 log.u_bbr.flex6 = tp->t_segqmbuflen;
229 log.u_bbr.flex7 = instance;
230 log.u_bbr.flex8 = action;
231 log.u_bbr.timeStamp = cts;
232 TCP_LOG_EVENTP(tp, NULL,
233 &tp->t_inpcb->inp_socket->so_rcv,
234 &tp->t_inpcb->inp_socket->so_snd,
235 TCP_LOG_REASS, 0,
236 len, &log, false, &tv);
237 }
238 }
239
240 static void
tcp_reass_log_dump(struct tcpcb * tp)241 tcp_reass_log_dump(struct tcpcb *tp)
242 {
243 struct tseg_qent *q;
244
245 if (tp->t_logstate != TCP_LOG_STATE_OFF) {
246 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
247 tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0);
248 }
249 };
250 }
251
252 static void
tcp_reass_log_new_in(struct tcpcb * tp,tcp_seq seq,int len,struct mbuf * m,int logval,struct tseg_qent * q)253 tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m,
254 int logval, struct tseg_qent *q)
255 {
256 int cnt;
257 struct mbuf *t;
258
259 cnt = 0;
260 t = m;
261 while (t) {
262 cnt += t->m_len;
263 t = t->m_next;
264 }
265 tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt);
266 }
267
268 #endif
269
270 void
tcp_reass_global_init(void)271 tcp_reass_global_init(void)
272 {
273
274 tcp_reass_maxseg = nmbclusters / 16;
275 TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
276 &tcp_reass_maxseg);
277 tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
278 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
279 /* Set the zone limit and read back the effective value. */
280 tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
281 tcp_reass_maxseg);
282 #ifdef TCP_REASS_COUNTERS
283 reass_path1 = counter_u64_alloc(M_WAITOK);
284 reass_path2 = counter_u64_alloc(M_WAITOK);
285 reass_path3 = counter_u64_alloc(M_WAITOK);
286 reass_path4 = counter_u64_alloc(M_WAITOK);
287 reass_path5 = counter_u64_alloc(M_WAITOK);
288 reass_path6 = counter_u64_alloc(M_WAITOK);
289 reass_path7 = counter_u64_alloc(M_WAITOK);
290 reass_fullwalk = counter_u64_alloc(M_WAITOK);
291 reass_nospace = counter_u64_alloc(M_WAITOK);
292 reass_entry = counter_u64_alloc(M_WAITOK);
293 merge_fwd = counter_u64_alloc(M_WAITOK);
294 merge_into = counter_u64_alloc(M_WAITOK);
295 tcp_zero_input = counter_u64_alloc(M_WAITOK);
296 #endif
297 EVENTHANDLER_REGISTER(nmbclusters_change,
298 tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
299
300 }
301
302 void
tcp_reass_flush(struct tcpcb * tp)303 tcp_reass_flush(struct tcpcb *tp)
304 {
305 struct tseg_qent *qe;
306
307 INP_WLOCK_ASSERT(tp->t_inpcb);
308
309 while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) {
310 TAILQ_REMOVE(&tp->t_segq, qe, tqe_q);
311 m_freem(qe->tqe_m);
312 uma_zfree(tcp_reass_zone, qe);
313 tp->t_segqlen--;
314 }
315 tp->t_segqmbuflen = 0;
316 KASSERT((tp->t_segqlen == 0),
317 ("TCP reass queue %p segment count is %d instead of 0 after flush.",
318 tp, tp->t_segqlen));
319 }
320
321 static void
tcp_reass_append(struct tcpcb * tp,struct tseg_qent * last,struct mbuf * m,struct tcphdr * th,int tlen,struct mbuf * mlast,int lenofoh)322 tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last,
323 struct mbuf *m, struct tcphdr *th, int tlen,
324 struct mbuf *mlast, int lenofoh)
325 {
326
327 #ifdef TCP_REASS_LOGGING
328 tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0);
329 #endif
330 last->tqe_len += tlen;
331 last->tqe_m->m_pkthdr.len += tlen;
332 /* Preserve the FIN bit if its there */
333 last->tqe_flags |= (th->th_flags & TH_FIN);
334 last->tqe_last->m_next = m;
335 last->tqe_last = mlast;
336 last->tqe_mbuf_cnt += lenofoh;
337 tp->t_rcvoopack++;
338 TCPSTAT_INC(tcps_rcvoopack);
339 TCPSTAT_ADD(tcps_rcvoobyte, tlen);
340 #ifdef TCP_REASS_LOGGING
341 tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m,
342 TCP_R_LOG_APPEND,
343 last);
344 #endif
345 }
346
347 static void
tcp_reass_prepend(struct tcpcb * tp,struct tseg_qent * first,struct mbuf * m,struct tcphdr * th,int tlen,struct mbuf * mlast,int lenofoh)348 tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th,
349 int tlen, struct mbuf *mlast, int lenofoh)
350 {
351 int i;
352
353 #ifdef TCP_REASS_LOGGING
354 tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0);
355 #endif
356 if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) {
357 /* The new data overlaps into the old */
358 i = (th->th_seq + tlen) - first->tqe_start;
359 #ifdef TCP_REASS_LOGGING
360 tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1);
361 #endif
362 m_adj(first->tqe_m, i);
363 first->tqe_len -= i;
364 first->tqe_start += i;
365 }
366 /* Ok now setup our chain to point to the old first */
367 mlast->m_next = first->tqe_m;
368 first->tqe_m = m;
369 first->tqe_len += tlen;
370 first->tqe_start = th->th_seq;
371 first->tqe_m->m_pkthdr.len = first->tqe_len;
372 first->tqe_mbuf_cnt += lenofoh;
373 tp->t_rcvoopack++;
374 TCPSTAT_INC(tcps_rcvoopack);
375 TCPSTAT_ADD(tcps_rcvoobyte, tlen);
376 #ifdef TCP_REASS_LOGGING
377 tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m,
378 TCP_R_LOG_PREPEND,
379 first);
380 #endif
381 }
382
383 static void
tcp_reass_replace(struct tcpcb * tp,struct tseg_qent * q,struct mbuf * m,tcp_seq seq,int len,struct mbuf * mlast,int mbufoh,uint8_t flags)384 tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m,
385 tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint8_t flags)
386 {
387 /*
388 * Free the data in q, and replace
389 * it with the new segment.
390 */
391 int len_dif;
392
393 #ifdef TCP_REASS_LOGGING
394 tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0);
395 #endif
396 m_freem(q->tqe_m);
397 KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
398 ("Tp:%p seg queue goes negative", tp));
399 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
400 q->tqe_mbuf_cnt = mbufoh;
401 q->tqe_m = m;
402 q->tqe_last = mlast;
403 q->tqe_start = seq;
404 if (len > q->tqe_len)
405 len_dif = len - q->tqe_len;
406 else
407 len_dif = 0;
408 tp->t_rcvoopack++;
409 TCPSTAT_INC(tcps_rcvoopack);
410 TCPSTAT_ADD(tcps_rcvoobyte, len_dif);
411 q->tqe_len = len;
412 q->tqe_flags = (flags & TH_FIN);
413 q->tqe_m->m_pkthdr.len = q->tqe_len;
414 tp->t_segqmbuflen += mbufoh;
415
416 }
417
418 static void
tcp_reass_merge_into(struct tcpcb * tp,struct tseg_qent * ent,struct tseg_qent * q)419 tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent,
420 struct tseg_qent *q)
421 {
422 /*
423 * Merge q into ent and free q from the list.
424 */
425 #ifdef TCP_REASS_LOGGING
426 tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0);
427 #endif
428 #ifdef TCP_REASS_COUNTERS
429 counter_u64_add(merge_into, 1);
430 #endif
431 ent->tqe_last->m_next = q->tqe_m;
432 ent->tqe_last = q->tqe_last;
433 ent->tqe_len += q->tqe_len;
434 ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt;
435 ent->tqe_m->m_pkthdr.len += q->tqe_len;
436 ent->tqe_flags |= (q->tqe_flags & TH_FIN);
437 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
438 uma_zfree(tcp_reass_zone, q);
439 tp->t_segqlen--;
440
441 }
442
443 static void
tcp_reass_merge_forward(struct tcpcb * tp,struct tseg_qent * ent)444 tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent)
445 {
446 struct tseg_qent *q, *qtmp;
447 int i;
448 tcp_seq max;
449 /*
450 * Given an entry merge forward anyplace
451 * that ent overlaps forward.
452 */
453
454 max = ent->tqe_start + ent->tqe_len;
455 q = TAILQ_NEXT(ent, tqe_q);
456 if (q == NULL) {
457 /* Nothing left */
458 return;
459 }
460 TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) {
461 if (SEQ_GT(q->tqe_start, max)) {
462 /* Beyond q */
463 break;
464 }
465 /* We have some or all that are overlapping */
466 if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) {
467 /* It consumes it all */
468 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
469 m_freem(q->tqe_m);
470 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
471 uma_zfree(tcp_reass_zone, q);
472 tp->t_segqlen--;
473 continue;
474 }
475 /*
476 * Trim the q entry to dovetail to this one
477 * and then merge q into ent updating max
478 * in the process.
479 */
480 i = max - q->tqe_start;
481 #ifdef TCP_REASS_LOGGING
482 tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2);
483 #endif
484 m_adj(q->tqe_m, i);
485 q->tqe_len -= i;
486 q->tqe_start += i;
487 tcp_reass_merge_into(tp, ent, q);
488 max = ent->tqe_start + ent->tqe_len;
489 }
490 #ifdef TCP_REASS_COUNTERS
491 counter_u64_add(merge_fwd, 1);
492 #endif
493 }
494
495 static int
tcp_reass_overhead_of_chain(struct mbuf * m,struct mbuf ** mlast)496 tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast)
497 {
498 int len = MSIZE;
499
500 if (m->m_flags & M_EXT)
501 len += m->m_ext.ext_size;
502 while (m->m_next != NULL) {
503 m = m->m_next;
504 len += MSIZE;
505 if (m->m_flags & M_EXT)
506 len += m->m_ext.ext_size;
507 }
508 *mlast = m;
509 return (len);
510 }
511
512 /*
513 * NOTE!!! the new tcp-reassembly code *must not* use
514 * m_adj() with a negative index. That alters the chain
515 * of mbufs (by possibly chopping trailing mbufs). At
516 * the front of tcp_reass we count the mbuf overhead
517 * and setup the tail pointer. If we use m_adj(m, -5)
518 * we could corrupt the tail pointer. Currently the
519 * code only uses m_adj(m, postive-num). If this
520 * changes appropriate changes to update mlast would
521 * be needed.
522 */
523 int
tcp_reass(struct tcpcb * tp,struct tcphdr * th,tcp_seq * seq_start,int * tlenp,struct mbuf * m)524 tcp_reass(struct tcpcb *tp, struct tcphdr *th, tcp_seq *seq_start,
525 int *tlenp, struct mbuf *m)
526 {
527 struct tseg_qent *q, *last, *first;
528 struct tseg_qent *p = NULL;
529 struct tseg_qent *nq = NULL;
530 struct tseg_qent *te = NULL;
531 struct mbuf *mlast = NULL;
532 struct sockbuf *sb;
533 struct socket *so = tp->t_inpcb->inp_socket;
534 char *s = NULL;
535 int flags, i, lenofoh;
536
537 INP_WLOCK_ASSERT(tp->t_inpcb);
538 /*
539 * XXX: tcp_reass() is rather inefficient with its data structures
540 * and should be rewritten (see NetBSD for optimizations).
541 */
542
543 KASSERT(th == NULL || (seq_start != NULL && tlenp != NULL),
544 ("tcp_reass called with illegal parameter combination "
545 "(tp=%p, th=%p, seq_start=%p, tlenp=%p, m=%p)",
546 tp, th, seq_start, tlenp, m));
547 /*
548 * Call with th==NULL after become established to
549 * force pre-ESTABLISHED data up to user socket.
550 */
551 if (th == NULL)
552 goto present;
553 KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt),
554 ("Attempt to add old entry to reassembly queue (th=%p, tp=%p)",
555 th, tp));
556 #ifdef TCP_REASS_LOGGING
557 tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL);
558 #endif
559 #ifdef TCP_REASS_COUNTERS
560 counter_u64_add(reass_entry, 1);
561 #endif
562 /*
563 * Check for zero length data.
564 */
565 if ((*tlenp == 0) && ((th->th_flags & TH_FIN) == 0)) {
566 /*
567 * A zero length segment does no
568 * one any good. We could check
569 * the rcv_nxt <-> rcv_wnd but thats
570 * already done for us by the caller.
571 */
572 strip_fin:
573 #ifdef TCP_REASS_COUNTERS
574 counter_u64_add(tcp_zero_input, 1);
575 #endif
576 m_freem(m);
577 #ifdef TCP_REASS_LOGGING
578 tcp_reass_log_dump(tp);
579 #endif
580 return (0);
581 } else if ((*tlenp == 0) &&
582 (th->th_flags & TH_FIN) &&
583 !TCPS_HAVEESTABLISHED(tp->t_state)) {
584 /*
585 * We have not established, and we
586 * have a FIN and no data. Lets treat
587 * this as the same as if the FIN were
588 * not present. We don't want to save
589 * the FIN bit in a reassembly buffer
590 * we want to get established first before
591 * we do that (the peer will retransmit).
592 */
593 goto strip_fin;
594 }
595 /*
596 * Will it fit?
597 */
598 lenofoh = tcp_reass_overhead_of_chain(m, &mlast);
599 sb = &tp->t_inpcb->inp_socket->so_rcv;
600 if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
601 (sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) {
602 /* No room */
603 TCPSTAT_INC(tcps_rcvreassfull);
604 #ifdef TCP_REASS_COUNTERS
605 counter_u64_add(reass_nospace, 1);
606 #endif
607 #ifdef TCP_REASS_LOGGING
608 tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0);
609 #endif
610 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
611 log(LOG_DEBUG, "%s; %s: mbuf count limit reached, "
612 "segment dropped\n", s, __func__);
613 free(s, M_TCPLOG);
614 }
615 m_freem(m);
616 *tlenp = 0;
617 #ifdef TCP_REASS_LOGGING
618 tcp_reass_log_dump(tp);
619 #endif
620 return (0);
621 }
622 /*
623 * First lets deal with two common cases, the
624 * segment appends to the back of our collected
625 * segments. Or the segment is the next in line.
626 */
627 last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q);
628 if (last != NULL) {
629 if ((th->th_flags & TH_FIN) &&
630 SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) {
631 /*
632 * Someone is trying to game us, dump
633 * the segment.
634 */
635 *tlenp = 0;
636 m_freem(m);
637 return (0);
638 }
639 if ((SEQ_GEQ(th->th_seq, last->tqe_start)) &&
640 (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) {
641 /* Common case, trailing segment is added */
642 /**
643 * +--last
644 * v
645 * reassembly buffer |---| |---| |---|
646 * new segment |---|
647 */
648 #ifdef TCP_REASS_COUNTERS
649 counter_u64_add(reass_path1, 1);
650 #endif
651 if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) {
652 i = (last->tqe_start + last->tqe_len) - th->th_seq;
653 if (i < *tlenp) {
654 #ifdef TCP_REASS_LOGGING
655 tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3);
656 th->th_seq += i;
657 #endif
658 m_adj(m, i);
659 *tlenp -= i;
660 } else {
661 /* Complete overlap */
662 TCPSTAT_INC(tcps_rcvduppack);
663 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
664 m_freem(m);
665 *tlenp = last->tqe_len;
666 *seq_start = last->tqe_start;
667 return (0);
668 }
669 }
670 if (last->tqe_flags & TH_FIN) {
671 /*
672 * We have data after the FIN on the last?
673 */
674 *tlenp = 0;
675 m_freem(m);
676 return(0);
677 }
678 tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh);
679 tp->t_segqmbuflen += lenofoh;
680 *seq_start = last->tqe_start;
681 *tlenp = last->tqe_len;
682 return (0);
683 } else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) {
684 /*
685 * Second common case, we missed
686 * another one and have something more
687 * for the end.
688 */
689 /**
690 * +--last
691 * v
692 * reassembly buffer |---| |---| |---|
693 * new segment |---|
694 */
695 if (last->tqe_flags & TH_FIN) {
696 /*
697 * We have data after the FIN on the last?
698 */
699 *tlenp = 0;
700 m_freem(m);
701 return(0);
702 }
703 #ifdef TCP_REASS_COUNTERS
704 counter_u64_add(reass_path2, 1);
705 #endif
706 p = last;
707 goto new_entry;
708 }
709 } else {
710 /* First segment (it's NULL). */
711 goto new_entry;
712 }
713 first = TAILQ_FIRST(&tp->t_segq);
714 if (SEQ_LT(th->th_seq, first->tqe_start) &&
715 SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) &&
716 SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) {
717 /*
718 * The head of the queue is prepended by this and
719 * it may be the one I want most.
720 */
721 /**
722 * first-------+
723 * v
724 * rea: |---| |---| |---|
725 * new: |---|
726 * Note the case we do not deal with here is:
727 * rea= |---| |---| |---|
728 * new= |----|
729 * Due to the fact that it could be
730 * new |--------------------|
731 * And we might need to merge forward.
732 */
733 #ifdef INVARIANTS
734 struct mbuf *firstmbuf;
735 #endif
736
737 #ifdef TCP_REASS_COUNTERS
738 counter_u64_add(reass_path3, 1);
739 #endif
740 if (SEQ_LT(th->th_seq, tp->rcv_nxt)) {
741 /*
742 * The resend was even before
743 * what we have. We need to trim it.
744 * Note TSNH (it should be trimmed
745 * before the call to tcp_reass()).
746 */
747 #ifdef INVARIANTS
748 panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed",
749 th->th_seq, tp->rcv_nxt, tp);
750 #else
751 i = tp->rcv_nxt - th->th_seq;
752 #ifdef TCP_REASS_LOGGING
753 tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4);
754 #endif
755 m_adj(m, i);
756 th->th_seq += i;
757 *tlenp -= i;
758 #endif
759 }
760 #ifdef INVARIANTS
761 firstmbuf = first->tqe_m;
762 #endif
763 tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh);
764 #ifdef INVARIANTS
765 if (firstmbuf == first->tqe_m) {
766 panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
767 m, firstmbuf, first->tqe_m, tp, first);
768 } else if (first->tqe_m != m) {
769 panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
770 m, firstmbuf, first->tqe_m, tp, first);
771 }
772 #endif
773 tp->t_segqmbuflen += lenofoh;
774 *seq_start = first->tqe_start;
775 *tlenp = first->tqe_len;
776 goto present;
777 } else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) {
778 /* New segment is before our earliest segment. */
779 /**
780 * first---->+
781 * v
782 * rea= |---| ....
783 * new" |---|
784 *
785 */
786 goto new_entry;
787 }
788 /*
789 * Find a segment which begins after this one does.
790 */
791 #ifdef TCP_REASS_COUNTERS
792 counter_u64_add(reass_fullwalk, 1);
793 #endif
794 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
795 if (SEQ_GT(q->tqe_start, th->th_seq))
796 break;
797 }
798 p = TAILQ_PREV(q, tsegqe_head, tqe_q);
799 /**
800 * Now is this fit just in-between only?
801 * i.e.:
802 * p---+ +----q
803 * v v
804 * res= |--| |--| |--|
805 * nee |-|
806 */
807 if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) &&
808 ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) {
809 /* Yep no overlap */
810 goto new_entry;
811 }
812 /**
813 * If we reach here we have some (possibly all) overlap
814 * such as:
815 * res= |--| |--| |--|
816 * new= |----|
817 * or new= |-----------------|
818 * or new= |--------|
819 * or new= |---|
820 * or new= |-----------|
821 */
822 if ((p != NULL) &&
823 (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) {
824 /* conversion to int (in i) handles seq wraparound */
825
826 #ifdef TCP_REASS_COUNTERS
827 counter_u64_add(reass_path4, 1);
828 #endif
829 i = p->tqe_start + p->tqe_len - th->th_seq;
830 if (i >= 0) {
831 if (i >= *tlenp) {
832 /**
833 * prev seg---->+
834 * v
835 * reassembly buffer |---|
836 * new segment |-|
837 */
838 TCPSTAT_INC(tcps_rcvduppack);
839 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
840 *tlenp = p->tqe_len;
841 *seq_start = p->tqe_start;
842 m_freem(m);
843 /*
844 * Try to present any queued data
845 * at the left window edge to the user.
846 * This is needed after the 3-WHS
847 * completes. Note this probably
848 * will not work and we will return.
849 */
850 return (0);
851 }
852 if (i > 0) {
853 /**
854 * prev seg---->+
855 * v
856 * reassembly buffer |---|
857 * new segment |-----|
858 */
859 #ifdef TCP_REASS_COUNTERS
860 counter_u64_add(reass_path5, 1);
861 #endif
862 #ifdef TCP_REASS_LOGGING
863 tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5);
864 #endif
865 m_adj(m, i);
866 *tlenp -= i;
867 th->th_seq += i;
868 }
869 }
870 if (th->th_seq == (p->tqe_start + p->tqe_len)) {
871 /*
872 * If dovetails in with this one
873 * append it.
874 */
875 /**
876 * prev seg---->+
877 * v
878 * reassembly buffer |--| |---|
879 * new segment |--|
880 * (note: it was trimmed above if it overlapped)
881 */
882 tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh);
883 tp->t_segqmbuflen += lenofoh;
884 } else {
885 #ifdef INVARIANTS
886 panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p",
887 th->th_seq, p->tqe_start, p->tqe_len,
888 p, tp);
889 #endif
890 *tlenp = 0;
891 m_freem(m);
892 return (0);
893 }
894 q = p;
895 } else {
896 /*
897 * The new data runs over the
898 * top of previously sack'd data (in q).
899 * It may be partially overlapping, or
900 * it may overlap the entire segment.
901 */
902 #ifdef TCP_REASS_COUNTERS
903 counter_u64_add(reass_path6, 1);
904 #endif
905 if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) {
906 /* It consumes it all */
907 /**
908 * next seg---->+
909 * v
910 * reassembly buffer |--| |---|
911 * new segment |----------|
912 */
913 #ifdef TCP_REASS_COUNTERS
914 counter_u64_add(reass_path7, 1);
915 #endif
916 tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, th->th_flags);
917 } else {
918 /*
919 * We just need to prepend the data
920 * to this. It does not overrun
921 * the end.
922 */
923 /**
924 * next seg---->+
925 * v
926 * reassembly buffer |--| |---|
927 * new segment |----------|
928 */
929 tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh);
930 tp->t_segqmbuflen += lenofoh;
931 }
932 }
933 /* Now does it go further than that? */
934 tcp_reass_merge_forward(tp, q);
935 *seq_start = q->tqe_start;
936 *tlenp = q->tqe_len;
937 goto present;
938
939 /*
940 * When we reach here we can't combine it
941 * with any existing segment.
942 *
943 * Limit the number of segments that can be queued to reduce the
944 * potential for mbuf exhaustion. For best performance, we want to be
945 * able to queue a full window's worth of segments. The size of the
946 * socket receive buffer determines our advertised window and grows
947 * automatically when socket buffer autotuning is enabled. Use it as the
948 * basis for our queue limit.
949 *
950 * However, allow the user to specify a ceiling for the number of
951 * segments in each queue.
952 *
953 * Always let the missing segment through which caused this queue.
954 * NB: Access to the socket buffer is left intentionally unlocked as we
955 * can tolerate stale information here.
956 *
957 * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
958 * should work but causes packets to be dropped when they shouldn't.
959 * Investigate why and re-evaluate the below limit after the behaviour
960 * is understood.
961 */
962 new_entry:
963 if (th->th_seq == tp->rcv_nxt && TCPS_HAVEESTABLISHED(tp->t_state)) {
964 tp->rcv_nxt += *tlenp;
965 flags = th->th_flags & TH_FIN;
966 TCPSTAT_INC(tcps_rcvoopack);
967 TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
968 SOCKBUF_LOCK(&so->so_rcv);
969 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
970 m_freem(m);
971 } else {
972 sbappendstream_locked(&so->so_rcv, m, 0);
973 }
974 tp->t_flags |= TF_WAKESOR;
975 return (flags);
976 }
977 if (tcp_new_limits) {
978 if ((tp->t_segqlen > tcp_reass_queue_guard) &&
979 (*tlenp < MSIZE)) {
980 /*
981 * This is really a lie, we are not full but
982 * are getting a segment that is above
983 * guard threshold. If it is and its below
984 * a mbuf size (256) we drop it if it
985 * can't fill in some place.
986 */
987 TCPSTAT_INC(tcps_rcvreassfull);
988 *tlenp = 0;
989 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
990 log(LOG_DEBUG, "%s; %s: queue limit reached, "
991 "segment dropped\n", s, __func__);
992 free(s, M_TCPLOG);
993 }
994 m_freem(m);
995 #ifdef TCP_REASS_LOGGING
996 tcp_reass_log_dump(tp);
997 #endif
998 return (0);
999 }
1000 } else {
1001 if (tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
1002 tcp_reass_maxqueuelen)) {
1003 TCPSTAT_INC(tcps_rcvreassfull);
1004 *tlenp = 0;
1005 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
1006 log(LOG_DEBUG, "%s; %s: queue limit reached, "
1007 "segment dropped\n", s, __func__);
1008 free(s, M_TCPLOG);
1009 }
1010 m_freem(m);
1011 #ifdef TCP_REASS_LOGGING
1012 tcp_reass_log_dump(tp);
1013 #endif
1014 return (0);
1015 }
1016 }
1017 /*
1018 * Allocate a new queue entry. If we can't, or hit the zone limit
1019 * just drop the pkt.
1020 */
1021 te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
1022 if (te == NULL) {
1023 TCPSTAT_INC(tcps_rcvmemdrop);
1024 m_freem(m);
1025 *tlenp = 0;
1026 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
1027 NULL))) {
1028 log(LOG_DEBUG, "%s; %s: global zone limit "
1029 "reached, segment dropped\n", s, __func__);
1030 free(s, M_TCPLOG);
1031 }
1032 return (0);
1033 }
1034 tp->t_segqlen++;
1035 tp->t_rcvoopack++;
1036 TCPSTAT_INC(tcps_rcvoopack);
1037 TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
1038 /* Insert the new segment queue entry into place. */
1039 te->tqe_m = m;
1040 te->tqe_flags = th->th_flags;
1041 te->tqe_len = *tlenp;
1042 te->tqe_start = th->th_seq;
1043 te->tqe_last = mlast;
1044 te->tqe_mbuf_cnt = lenofoh;
1045 tp->t_segqmbuflen += te->tqe_mbuf_cnt;
1046 if (p == NULL) {
1047 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
1048 } else {
1049 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
1050 }
1051 #ifdef TCP_REASS_LOGGING
1052 tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_NEW_ENTRY, te);
1053 #endif
1054 present:
1055 /*
1056 * Present data to user, advancing rcv_nxt through
1057 * completed sequence space.
1058 */
1059 if (!TCPS_HAVEESTABLISHED(tp->t_state))
1060 return (0);
1061 q = TAILQ_FIRST(&tp->t_segq);
1062 KASSERT(q == NULL || SEQ_GEQ(q->tqe_start, tp->rcv_nxt),
1063 ("Reassembly queue for %p has stale entry at head", tp));
1064 if (!q || q->tqe_start != tp->rcv_nxt) {
1065 #ifdef TCP_REASS_LOGGING
1066 tcp_reass_log_dump(tp);
1067 #endif
1068 return (0);
1069 }
1070 SOCKBUF_LOCK(&so->so_rcv);
1071 do {
1072 tp->rcv_nxt += q->tqe_len;
1073 flags = q->tqe_flags & TH_FIN;
1074 nq = TAILQ_NEXT(q, tqe_q);
1075 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1076 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
1077 m_freem(q->tqe_m);
1078 } else {
1079 #ifdef TCP_REASS_LOGGING
1080 tcp_reass_log_new_in(tp, q->tqe_start, q->tqe_len, q->tqe_m, TCP_R_LOG_READ, q);
1081 if (th != NULL) {
1082 tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 1);
1083 } else {
1084 tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 1);
1085 }
1086 #endif
1087 sbappendstream_locked(&so->so_rcv, q->tqe_m, 0);
1088 }
1089 #ifdef TCP_REASS_LOGGING
1090 if (th != NULL) {
1091 tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 2);
1092 } else {
1093 tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 2);
1094 }
1095 #endif
1096 KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
1097 ("tp:%p seg queue goes negative", tp));
1098 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
1099 uma_zfree(tcp_reass_zone, q);
1100 tp->t_segqlen--;
1101 q = nq;
1102 } while (q && q->tqe_start == tp->rcv_nxt);
1103 if (TAILQ_EMPTY(&tp->t_segq) &&
1104 (tp->t_segqmbuflen != 0)) {
1105 #ifdef INVARIANTS
1106 panic("tp:%p segq:%p len:%d queue empty",
1107 tp, &tp->t_segq, tp->t_segqmbuflen);
1108 #else
1109 #ifdef TCP_REASS_LOGGING
1110 if (th != NULL) {
1111 tcp_log_reassm(tp, NULL, NULL, th->th_seq, *tlenp, TCP_R_LOG_ZERO, 0);
1112 } else {
1113 tcp_log_reassm(tp, NULL, NULL, 0, 0, TCP_R_LOG_ZERO, 0);
1114 }
1115 #endif
1116 tp->t_segqmbuflen = 0;
1117 #endif
1118 }
1119 #ifdef TCP_REASS_LOGGING
1120 tcp_reass_log_dump(tp);
1121 #endif
1122 tp->t_flags |= TF_WAKESOR;
1123 return (flags);
1124 }
1125