xref: /freebsd/sys/netinet/tcp_reass.c (revision d34048812292b714a0bf99967270d18fe3097c62)
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 #include <sys/param.h>
42 #include <sys/kernel.h>
43 #include <sys/eventhandler.h>
44 #include <sys/malloc.h>
45 #include <sys/mbuf.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/syslog.h>
50 #include <sys/systm.h>
51 
52 #include <vm/uma.h>
53 
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/route.h>
57 #include <net/vnet.h>
58 
59 #include <netinet/in.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/in_systm.h>
62 #include <netinet/in_var.h>
63 #include <netinet/ip.h>
64 #include <netinet/ip_var.h>
65 #include <netinet/ip_options.h>
66 #include <netinet/ip6.h>
67 #include <netinet6/in6_pcb.h>
68 #include <netinet6/ip6_var.h>
69 #include <netinet6/nd6.h>
70 #include <netinet/tcp.h>
71 #include <netinet/tcp_fsm.h>
72 #include <netinet/tcp_seq.h>
73 #include <netinet/tcp_timer.h>
74 #include <netinet/tcp_var.h>
75 #include <netinet/tcp_log_buf.h>
76 #include <netinet/tcp_hpts.h>
77 #include <netinet6/tcp6_var.h>
78 #include <netinet/tcpip.h>
79 #ifdef TCPDEBUG
80 #include <netinet/tcp_debug.h>
81 #endif /* TCPDEBUG */
82 
83 #define TCP_R_LOG_ADD		1
84 #define TCP_R_LOG_LIMIT_REACHED 2
85 #define TCP_R_LOG_APPEND	3
86 #define TCP_R_LOG_PREPEND	4
87 #define TCP_R_LOG_REPLACE	5
88 #define TCP_R_LOG_MERGE_INTO	6
89 #define TCP_R_LOG_NEW_ENTRY	7
90 #define TCP_R_LOG_READ		8
91 #define TCP_R_LOG_ZERO		9
92 #define TCP_R_LOG_DUMP		10
93 #define TCP_R_LOG_TRIM		11
94 
95 /* For debugging we want counters and BB logging */
96 /* #define TCP_REASS_COUNTERS 1 */
97 /* #define TCP_REASS_LOGGING 1 */
98 
99 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
100     "TCP Segment Reassembly Queue");
101 
102 static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats, CTLFLAG_RW, 0,
103     "TCP Segment Reassembly stats");
104 
105 
106 static int tcp_reass_maxseg = 0;
107 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
108     &tcp_reass_maxseg, 0,
109     "Global maximum number of TCP Segments in Reassembly Queue");
110 
111 static uma_zone_t tcp_reass_zone;
112 SYSCTL_UMA_CUR(_net_inet_tcp_reass, OID_AUTO, cursegments, 0,
113     &tcp_reass_zone,
114     "Global number of TCP Segments currently in Reassembly Queue");
115 
116 static u_int tcp_reass_maxqueuelen = 100;
117 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
118     &tcp_reass_maxqueuelen, 0,
119     "Maximum number of TCP Segments per Reassembly Queue");
120 
121 static int tcp_new_limits = 0;
122 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN,
123     &tcp_new_limits, 0,
124     "Do we use the new limit method we are discussing?");
125 
126 static u_int tcp_reass_queue_guard = 16;
127 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN,
128     &tcp_reass_queue_guard, 16,
129     "Number of TCP Segments in Reassembly Queue where we flip over to guard mode");
130 
131 #ifdef TCP_REASS_COUNTERS
132 
133 counter_u64_t reass_entry;
134 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD,
135     &reass_entry, "A segment entered reassembly ");
136 
137 counter_u64_t reass_path1;
138 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD,
139     &reass_path1, "Took path 1");
140 
141 counter_u64_t reass_path2;
142 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD,
143     &reass_path2, "Took path 2");
144 
145 counter_u64_t reass_path3;
146 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD,
147     &reass_path3, "Took path 3");
148 
149 counter_u64_t reass_path4;
150 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD,
151     &reass_path4, "Took path 4");
152 
153 counter_u64_t reass_path5;
154 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD,
155     &reass_path5, "Took path 5");
156 
157 counter_u64_t reass_path6;
158 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD,
159     &reass_path6, "Took path 6");
160 
161 counter_u64_t reass_path7;
162 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD,
163     &reass_path7, "Took path 7");
164 
165 counter_u64_t reass_fullwalk;
166 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD,
167     &reass_fullwalk, "Took a full walk ");
168 
169 counter_u64_t reass_nospace;
170 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD,
171     &reass_nospace, "Had no mbuf capacity ");
172 
173 counter_u64_t merge_fwd;
174 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD,
175     &merge_fwd, "Ran merge fwd");
176 
177 counter_u64_t merge_into;
178 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD,
179     &merge_into, "Ran merge into");
180 
181 counter_u64_t tcp_zero_input;
182 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD,
183     &tcp_zero_input, "The reassembly buffer saw a zero len segment etc");
184 
185 #endif
186 
187 /* Initialize TCP reassembly queue */
188 static void
189 tcp_reass_zone_change(void *tag)
190 {
191 
192 	/* Set the zone limit and read back the effective value. */
193 	tcp_reass_maxseg = nmbclusters / 16;
194 	tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
195 	    tcp_reass_maxseg);
196 }
197 
198 #ifdef TCP_REASS_LOGGING
199 
200 static void
201 tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p,
202     tcp_seq seq, int len, uint8_t action, int instance)
203 {
204 	uint32_t cts;
205 	struct timeval tv;
206 
207 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
208 		union tcp_log_stackspecific log;
209 
210 		memset(&log, 0, sizeof(log));
211 		cts = tcp_get_usecs(&tv);
212 		log.u_bbr.flex1 = seq;
213 		log.u_bbr.cur_del_rate = (uint64_t)q;
214 		log.u_bbr.delRate = (uint64_t)p;
215 		if (q != NULL) {
216 			log.u_bbr.flex2 = q->tqe_start;
217 			log.u_bbr.flex3 = q->tqe_len;
218 			log.u_bbr.flex4 = q->tqe_mbuf_cnt;
219 			log.u_bbr.hptsi_gain = q->tqe_flags;
220 		}
221 		if (p != NULL)  {
222 			log.u_bbr.flex5 = p->tqe_start;
223 			log.u_bbr.pkts_out = p->tqe_len;
224 			log.u_bbr.epoch = p->tqe_mbuf_cnt;
225 			log.u_bbr.cwnd_gain = p->tqe_flags;
226 		}
227 		log.u_bbr.flex6 = tp->t_segqmbuflen;
228 		log.u_bbr.flex7 = instance;
229 		log.u_bbr.flex8 = action;
230 		log.u_bbr.timeStamp = cts;
231 		TCP_LOG_EVENTP(tp, NULL,
232 		    &tp->t_inpcb->inp_socket->so_rcv,
233 		    &tp->t_inpcb->inp_socket->so_snd,
234 		    TCP_LOG_REASS, 0,
235 		    len, &log, false, &tv);
236 	}
237 }
238 
239 static void
240 tcp_reass_log_dump(struct tcpcb *tp)
241 {
242 	struct tseg_qent *q;
243 
244 	if (tp->t_logstate != TCP_LOG_STATE_OFF) {
245 		TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
246 			tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0);
247 		}
248 	};
249 }
250 
251 static void
252 tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m,
253     int logval, struct tseg_qent *q)
254 {
255 	int cnt;
256 	struct mbuf *t;
257 
258 	cnt = 0;
259 	t = m;
260 	while (t) {
261 		cnt += t->m_len;
262 		t = t->m_next;
263 	}
264 	tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt);
265 }
266 
267 #endif
268 
269 void
270 tcp_reass_global_init(void)
271 {
272 
273 	tcp_reass_maxseg = nmbclusters / 16;
274 	TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
275 	    &tcp_reass_maxseg);
276 	tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
277 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
278 	/* Set the zone limit and read back the effective value. */
279 	tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
280 	    tcp_reass_maxseg);
281 #ifdef TCP_REASS_COUNTERS
282 	reass_path1 = counter_u64_alloc(M_WAITOK);
283 	reass_path2 = counter_u64_alloc(M_WAITOK);
284 	reass_path3 = counter_u64_alloc(M_WAITOK);
285 	reass_path4 = counter_u64_alloc(M_WAITOK);
286 	reass_path5 = counter_u64_alloc(M_WAITOK);
287 	reass_path6 = counter_u64_alloc(M_WAITOK);
288 	reass_path7 = counter_u64_alloc(M_WAITOK);
289 	reass_fullwalk = counter_u64_alloc(M_WAITOK);
290 	reass_nospace = counter_u64_alloc(M_WAITOK);
291 	reass_entry = counter_u64_alloc(M_WAITOK);
292 	merge_fwd = counter_u64_alloc(M_WAITOK);
293 	merge_into = counter_u64_alloc(M_WAITOK);
294 	tcp_zero_input = counter_u64_alloc(M_WAITOK);
295 #endif
296 	EVENTHANDLER_REGISTER(nmbclusters_change,
297 	    tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
298 
299 }
300 
301 void
302 tcp_reass_flush(struct tcpcb *tp)
303 {
304 	struct tseg_qent *qe;
305 
306 	INP_WLOCK_ASSERT(tp->t_inpcb);
307 
308 	while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) {
309 		TAILQ_REMOVE(&tp->t_segq, qe, tqe_q);
310 		m_freem(qe->tqe_m);
311 		uma_zfree(tcp_reass_zone, qe);
312 		tp->t_segqlen--;
313 	}
314 	tp->t_segqmbuflen = 0;
315 	KASSERT((tp->t_segqlen == 0),
316 	    ("TCP reass queue %p segment count is %d instead of 0 after flush.",
317 	    tp, tp->t_segqlen));
318 }
319 
320 static void
321 tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last,
322     struct mbuf *m, struct tcphdr *th, int tlen,
323     struct mbuf *mlast, int lenofoh)
324 {
325 
326 #ifdef TCP_REASS_LOGGING
327 	tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0);
328 #endif
329 	last->tqe_len += tlen;
330 	last->tqe_m->m_pkthdr.len += tlen;
331 	/* Preserve the FIN bit if its there */
332 	last->tqe_flags |= (th->th_flags & TH_FIN);
333 	last->tqe_last->m_next = m;
334 	last->tqe_last = mlast;
335 	last->tqe_mbuf_cnt += lenofoh;
336 	tp->t_rcvoopack++;
337 	TCPSTAT_INC(tcps_rcvoopack);
338 	TCPSTAT_ADD(tcps_rcvoobyte, tlen);
339 #ifdef TCP_REASS_LOGGING
340 	tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m,
341 			     TCP_R_LOG_APPEND,
342 			     last);
343 #endif
344 }
345 
346 static void
347 tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th,
348 		  int tlen, struct mbuf *mlast, int lenofoh)
349 {
350 	int i;
351 
352 #ifdef TCP_REASS_LOGGING
353 	tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0);
354 #endif
355 	if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) {
356 		/* The new data overlaps into the old */
357 		i = (th->th_seq + tlen) - first->tqe_start;
358 #ifdef TCP_REASS_LOGGING
359 		tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1);
360 #endif
361 		m_adj(first->tqe_m, i);
362 		first->tqe_len -= i;
363 		first->tqe_start += i;
364 	}
365 	/* Ok now setup our chain to point to the old first */
366 	mlast->m_next = first->tqe_m;
367 	first->tqe_m = m;
368 	first->tqe_len += tlen;
369 	first->tqe_start = th->th_seq;
370 	first->tqe_m->m_pkthdr.len = first->tqe_len;
371 	first->tqe_mbuf_cnt += lenofoh;
372 	tp->t_rcvoopack++;
373 	TCPSTAT_INC(tcps_rcvoopack);
374 	TCPSTAT_ADD(tcps_rcvoobyte, tlen);
375 #ifdef TCP_REASS_LOGGING
376 	tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m,
377 			     TCP_R_LOG_PREPEND,
378 			     first);
379 #endif
380 }
381 
382 static void
383 tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m,
384     tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint8_t flags)
385 {
386 	/*
387 	 * Free the data in q, and replace
388 	 * it with the new segment.
389 	 */
390 	int len_dif;
391 
392 #ifdef TCP_REASS_LOGGING
393 	tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0);
394 #endif
395 	m_freem(q->tqe_m);
396 	KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
397 		("Tp:%p seg queue goes negative", tp));
398 	tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
399 	q->tqe_mbuf_cnt = mbufoh;
400 	q->tqe_m = m;
401 	q->tqe_last = mlast;
402 	q->tqe_start = seq;
403 	if (len > q->tqe_len)
404 		len_dif = len - q->tqe_len;
405 	else
406 		len_dif = 0;
407 	tp->t_rcvoopack++;
408 	TCPSTAT_INC(tcps_rcvoopack);
409 	TCPSTAT_ADD(tcps_rcvoobyte, len_dif);
410 	q->tqe_len = len;
411 	q->tqe_flags = (flags & TH_FIN);
412 	q->tqe_m->m_pkthdr.len = q->tqe_len;
413 	tp->t_segqmbuflen += mbufoh;
414 
415 }
416 
417 static void
418 tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent,
419     struct tseg_qent *q)
420 {
421 	/*
422 	 * Merge q into ent and free q from the list.
423 	 */
424 #ifdef TCP_REASS_LOGGING
425 	tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0);
426 #endif
427 #ifdef TCP_REASS_COUNTERS
428 	counter_u64_add(merge_into, 1);
429 #endif
430 	ent->tqe_last->m_next = q->tqe_m;
431 	ent->tqe_last = q->tqe_last;
432 	ent->tqe_len += q->tqe_len;
433 	ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt;
434 	ent->tqe_m->m_pkthdr.len += q->tqe_len;
435 	ent->tqe_flags |= (q->tqe_flags & TH_FIN);
436 	TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
437 	uma_zfree(tcp_reass_zone, q);
438 	tp->t_segqlen--;
439 
440 }
441 
442 static void
443 tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent)
444 {
445 	struct tseg_qent *q, *qtmp;
446 	int i;
447 	tcp_seq max;
448 	/*
449 	 * Given an entry merge forward anyplace
450 	 * that ent overlaps forward.
451 	 */
452 
453 	max = ent->tqe_start + ent->tqe_len;
454 	q = TAILQ_NEXT(ent, tqe_q);
455 	if (q == NULL) {
456 		/* Nothing left */
457 		return;
458 	}
459 	TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) {
460 		if (SEQ_GT(q->tqe_start, max)) {
461 			/* Beyond q */
462 			break;
463 		}
464 		/* We have some or all that are overlapping */
465 		if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) {
466 			/* It consumes it all */
467 			tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
468 			m_freem(q->tqe_m);
469 			TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
470 			uma_zfree(tcp_reass_zone, q);
471 			tp->t_segqlen--;
472 			continue;
473 		}
474 		/*
475 		 * Trim the q entry to dovetail to this one
476 		 * and then merge q into ent updating max
477 		 * in the process.
478 		 */
479 		i = max - q->tqe_start;
480 #ifdef TCP_REASS_LOGGING
481 		tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2);
482 #endif
483 		m_adj(q->tqe_m, i);
484 		q->tqe_len -= i;
485 		q->tqe_start += i;
486 		tcp_reass_merge_into(tp, ent, q);
487 		max = ent->tqe_start + ent->tqe_len;
488 	}
489 #ifdef TCP_REASS_COUNTERS
490 	counter_u64_add(merge_fwd, 1);
491 #endif
492 }
493 
494 static int
495 tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast)
496 {
497 	int len = MSIZE;
498 
499 	if (m->m_flags & M_EXT)
500 		len += m->m_ext.ext_size;
501 	while (m->m_next != NULL) {
502 		m = m->m_next;
503 		len += MSIZE;
504 		if (m->m_flags & M_EXT)
505 			len += m->m_ext.ext_size;
506 	}
507 	*mlast = m;
508 	return (len);
509 }
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
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 	/*
544 	 * Call with th==NULL after become established to
545 	 * force pre-ESTABLISHED data up to user socket.
546 	 */
547 	if (th == NULL)
548 		goto present;
549 	KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt),
550 		("Attempt to add old entry to reassembly queue (th=%p, tp=%p)",
551 		 th, tp));
552 #ifdef TCP_REASS_LOGGING
553 	tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL);
554 #endif
555 #ifdef TCP_REASS_COUNTERS
556 	counter_u64_add(reass_entry, 1);
557 #endif
558 	/*
559 	 * Check for zero length data.
560 	 */
561 	if ((*tlenp == 0) && ((th->th_flags & TH_FIN) == 0)) {
562 		/*
563 		 * A zero length segment does no
564 		 * one any good. We could check
565 		 * the rcv_nxt <-> rcv_wnd but thats
566 		 * already done for us by the caller.
567 		 */
568 #ifdef TCP_REASS_COUNTERS
569 		counter_u64_add(tcp_zero_input, 1);
570 #endif
571 		m_freem(m);
572 #ifdef TCP_REASS_LOGGING
573 		tcp_reass_log_dump(tp);
574 #endif
575 		return (0);
576 	}
577 	/*
578 	 * Will it fit?
579 	 */
580 	lenofoh = tcp_reass_overhead_of_chain(m, &mlast);
581 	sb = &tp->t_inpcb->inp_socket->so_rcv;
582 	if ((sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) {
583 		/* No room */
584 		TCPSTAT_INC(tcps_rcvreassfull);
585 #ifdef TCP_REASS_COUNTERS
586 		counter_u64_add(reass_nospace, 1);
587 #endif
588 #ifdef TCP_REASS_LOGGING
589 		tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0);
590 #endif
591 		m_freem(m);
592 		*tlenp = 0;
593 #ifdef TCP_REASS_LOGGING
594 		tcp_reass_log_dump(tp);
595 #endif
596 		return (0);
597 	}
598 	/*
599 	 * First lets deal with two common cases, the
600 	 * segment appends to the back of our collected
601 	 * segments. Or the segment is the next in line.
602 	 */
603 	last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q);
604 	if (last != NULL) {
605 		if ((th->th_flags & TH_FIN) &&
606 		    SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) {
607 			/*
608 			 * Someone is trying to game us, dump
609 			 * the segment.
610 			 */
611 			*tlenp = 0;
612 			m_freem(m);
613 			return (0);
614 		}
615 		if ((SEQ_GEQ(th->th_seq, last->tqe_start)) &&
616 		    (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) {
617 			/* Common case, trailing segment is added */
618 			/**
619 			 *                                 +--last
620 			 *                                 v
621 			 *  reassembly buffer |---|  |---| |---|
622 			 *  new segment                       |---|
623 			 */
624 #ifdef TCP_REASS_COUNTERS
625 			counter_u64_add(reass_path1, 1);
626 #endif
627 			if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) {
628 				i = (last->tqe_start + last->tqe_len) - th->th_seq;
629 				if (i < *tlenp) {
630 #ifdef TCP_REASS_LOGGING
631 					tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3);
632 					th->th_seq += i;
633 #endif
634 					m_adj(m, i);
635 					*tlenp -= i;
636 				} else {
637 					/* Complete overlap */
638 					TCPSTAT_INC(tcps_rcvduppack);
639 					TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
640 					m_freem(m);
641 					*tlenp = last->tqe_len;
642 					*seq_start = last->tqe_start;
643 					return (0);
644 				}
645 			}
646 			if (last->tqe_flags & TH_FIN) {
647 				/*
648 				 * We have data after the FIN on the last?
649 				 */
650 				*tlenp = 0;
651 				m_freem(m);
652 				return(0);
653 			}
654 			tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh);
655 			tp->t_segqmbuflen += lenofoh;
656 			*seq_start = last->tqe_start;
657 			*tlenp = last->tqe_len;
658 			return (0);
659 		} else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) {
660 			/*
661 			 * Second common case, we missed
662 			 * another one and have something more
663 			 * for the end.
664 			 */
665 			/**
666 			 *                                 +--last
667 			 *                                 v
668 			 *  reassembly buffer |---|  |---| |---|
669 			 *  new segment                           |---|
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 #ifdef TCP_REASS_COUNTERS
680 			counter_u64_add(reass_path2, 1);
681 #endif
682 			p = last;
683 			goto new_entry;
684 		}
685 	} else {
686 		/* First segment (it's NULL). */
687 		goto new_entry;
688 	}
689 	first = TAILQ_FIRST(&tp->t_segq);
690 	if (SEQ_LT(th->th_seq, first->tqe_start) &&
691 	    SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) &&
692 	    SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) {
693 		/*
694 		 * The head of the queue is prepended by this and
695 		 * it may be the one I want most.
696 		 */
697 		/**
698 		 *       first-------+
699 		 *                   v
700 		 *  rea:             |---|  |---| |---|
701 		 *  new:         |---|
702 		 * Note the case we do not deal with here is:
703 		 *   rea=     |---|   |---|   |---|
704 		 *   new=  |----|
705 		 * Due to the fact that it could be
706 		 *   new   |--------------------|
707 		 * And we might need to merge forward.
708 		 */
709 #ifdef INVARIANTS
710 		struct mbuf *firstmbuf;
711 #endif
712 
713 #ifdef TCP_REASS_COUNTERS
714 		counter_u64_add(reass_path3, 1);
715 #endif
716 		if (SEQ_LT(th->th_seq, tp->rcv_nxt)) {
717 			/*
718 			 * The resend was even before
719 			 * what we have. We need to trim it.
720 			 * Note TSNH (it should be trimmed
721 			 * before the call to tcp_reass()).
722 			 */
723 #ifdef INVARIANTS
724 			panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed",
725 			      th->th_seq, tp->rcv_nxt, tp);
726 #else
727 			i = tp->rcv_nxt - th->th_seq;
728 #ifdef TCP_REASS_LOGGING
729 			tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4);
730 #endif
731 			m_adj(m, i);
732 			th->th_seq += i;
733 			*tlenp -= i;
734 #endif
735 		}
736 #ifdef INVARIANTS
737 		firstmbuf = first->tqe_m;
738 #endif
739 		tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh);
740 #ifdef INVARIANTS
741 		if (firstmbuf == first->tqe_m) {
742 			panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
743 			      m, firstmbuf, first->tqe_m, tp, first);
744 		} else if (first->tqe_m != m) {
745 			panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
746 			      m, firstmbuf, first->tqe_m, tp, first);
747 		}
748 #endif
749 		tp->t_segqmbuflen += lenofoh;
750 		*seq_start = first->tqe_start;
751 		*tlenp = first->tqe_len;
752 		goto present;
753 	} else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) {
754 		/* New segment is before our earliest segment. */
755 		/**
756 		 *           first---->+
757 		 *                      v
758 		 *  rea=                |---| ....
759 		 *  new"         |---|
760 		 *
761 		 */
762 		goto new_entry;
763 	}
764 	/*
765 	 * Find a segment which begins after this one does.
766 	 */
767 #ifdef TCP_REASS_COUNTERS
768 	counter_u64_add(reass_fullwalk, 1);
769 #endif
770 	TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
771 		if (SEQ_GT(q->tqe_start, th->th_seq))
772 			break;
773 	}
774 	p = TAILQ_PREV(q, tsegqe_head, tqe_q);
775 	/**
776 	 * Now is this fit just in-between only?
777 	 * i.e.:
778 	 *      p---+        +----q
779 	 *          v        v
780 	 *     res= |--|     |--|    |--|
781 	 *     nee       |-|
782 	 */
783 	if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) &&
784 	    ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) {
785 		/* Yep no overlap */
786 		goto new_entry;
787 	}
788 	/**
789 	 * If we reach here we have some (possibly all) overlap
790 	 * such as:
791 	 *     res=     |--|     |--|    |--|
792 	 *     new=  |----|
793 	 * or  new=  |-----------------|
794 	 * or  new=      |--------|
795 	 * or  new=            |---|
796 	 * or  new=            |-----------|
797 	 */
798 	if ((p != NULL) &&
799 	    (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) {
800 		/* conversion to int (in i) handles seq wraparound */
801 
802 #ifdef TCP_REASS_COUNTERS
803 		counter_u64_add(reass_path4, 1);
804 #endif
805 		i = p->tqe_start + p->tqe_len - th->th_seq;
806 		if (i >= 0) {
807 			if (i >= *tlenp) {
808 				/**
809 				 *       prev seg---->+
810 				 *                    v
811 				 *  reassembly buffer |---|
812 				 *  new segment        |-|
813 				 */
814 				TCPSTAT_INC(tcps_rcvduppack);
815 				TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
816 				*tlenp = p->tqe_len;
817 				*seq_start = p->tqe_start;
818 				m_freem(m);
819 				/*
820 				 * Try to present any queued data
821 				 * at the left window edge to the user.
822 				 * This is needed after the 3-WHS
823 				 * completes. Note this probably
824 				 * will not work and we will return.
825 				 */
826 				return (0);
827 			}
828 			if (i > 0) {
829 				/**
830 				 *       prev seg---->+
831 				 *                    v
832 				 *  reassembly buffer |---|
833 				 *  new segment         |-----|
834 				 */
835 #ifdef TCP_REASS_COUNTERS
836 				counter_u64_add(reass_path5, 1);
837 #endif
838 #ifdef TCP_REASS_LOGGING
839 				tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5);
840 #endif
841 				m_adj(m, i);
842 				*tlenp -= i;
843 				th->th_seq += i;
844 			}
845 		}
846 		if (th->th_seq == (p->tqe_start + p->tqe_len)) {
847 			/*
848 			 * If dovetails in with this one
849 			 * append it.
850 			 */
851 			/**
852 			 *       prev seg---->+
853 			 *                    v
854 			 *  reassembly buffer |--|     |---|
855 			 *  new segment          |--|
856 			 * (note: it was trimmed above if it overlapped)
857 			 */
858 			tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh);
859 			tp->t_segqmbuflen += lenofoh;
860 		} else {
861 #ifdef INVARIANTS
862 			panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p",
863 			      th->th_seq, p->tqe_start, p->tqe_len,
864 			      p, tp);
865 #endif
866 			*tlenp = 0;
867 			m_freem(m);
868 			return (0);
869 		}
870 		q = p;
871 	} else {
872 		/*
873 		 * The new data runs over the
874 		 * top of previously sack'd data (in q).
875 		 * It may be partially overlapping, or
876 		 * it may overlap the entire segment.
877 		 */
878 #ifdef TCP_REASS_COUNTERS
879 		counter_u64_add(reass_path6, 1);
880 #endif
881 		if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) {
882 			/* It consumes it all */
883 			/**
884 			 *             next seg---->+
885 			 *                          v
886 			 *  reassembly buffer |--|     |---|
887 			 *  new segment              |----------|
888 			 */
889 #ifdef TCP_REASS_COUNTERS
890 			counter_u64_add(reass_path7, 1);
891 #endif
892 			tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, th->th_flags);
893 		} else {
894 			/*
895 			 * We just need to prepend the data
896 			 * to this. It does not overrun
897 			 * the end.
898 			 */
899 			/**
900 			 *                next seg---->+
901 			 *                             v
902 			 *  reassembly buffer |--|     |---|
903 			 *  new segment                   |----------|
904 			 */
905 			tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh);
906 			tp->t_segqmbuflen += lenofoh;
907 		}
908 	}
909 	/* Now does it go further than that? */
910 	tcp_reass_merge_forward(tp, q);
911 	*seq_start = q->tqe_start;
912 	*tlenp = q->tqe_len;
913 	goto present;
914 
915 	/*
916 	 * When we reach here we can't combine it
917 	 * with any existing segment.
918 	 *
919 	 * Limit the number of segments that can be queued to reduce the
920 	 * potential for mbuf exhaustion. For best performance, we want to be
921 	 * able to queue a full window's worth of segments. The size of the
922 	 * socket receive buffer determines our advertised window and grows
923 	 * automatically when socket buffer autotuning is enabled. Use it as the
924 	 * basis for our queue limit.
925 	 *
926 	 * However, allow the user to specify a ceiling for the number of
927 	 * segments in each queue.
928 	 *
929 	 * Always let the missing segment through which caused this queue.
930 	 * NB: Access to the socket buffer is left intentionally unlocked as we
931 	 * can tolerate stale information here.
932 	 *
933 	 * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
934 	 * should work but causes packets to be dropped when they shouldn't.
935 	 * Investigate why and re-evaluate the below limit after the behaviour
936 	 * is understood.
937 	 */
938 new_entry:
939 	if (tcp_new_limits) {
940 		if ((tp->t_segqlen > tcp_reass_queue_guard) &&
941 		    (*tlenp < MSIZE)) {
942 			/*
943 			 * This is really a lie, we are not full but
944 			 * are getting a segment that is above
945 			 * guard threshold. If it is and its below
946 			 * a mbuf size (256) we drop it if it
947 			 * can't fill in some place.
948 			 */
949 			TCPSTAT_INC(tcps_rcvreassfull);
950 			*tlenp = 0;
951 			if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
952 				log(LOG_DEBUG, "%s; %s: queue limit reached, "
953 				    "segment dropped\n", s, __func__);
954 				free(s, M_TCPLOG);
955 			}
956 			m_freem(m);
957 #ifdef TCP_REASS_LOGGING
958 			tcp_reass_log_dump(tp);
959 #endif
960 			return (0);
961 		}
962 	} else {
963 
964 		if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
965 		    tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
966 					 tcp_reass_maxqueuelen)) {
967 			TCPSTAT_INC(tcps_rcvreassfull);
968 			*tlenp = 0;
969 			if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
970 				log(LOG_DEBUG, "%s; %s: queue limit reached, "
971 				    "segment dropped\n", s, __func__);
972 				free(s, M_TCPLOG);
973 			}
974 			m_freem(m);
975 #ifdef TCP_REASS_LOGGING
976 			tcp_reass_log_dump(tp);
977 #endif
978 			return (0);
979 		}
980 	}
981 	/*
982 	 * Allocate a new queue entry. If we can't, or hit the zone limit
983 	 * just drop the pkt.
984 	 */
985 	te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
986 	if (te == NULL) {
987 		TCPSTAT_INC(tcps_rcvmemdrop);
988 		m_freem(m);
989 		*tlenp = 0;
990 		if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
991 				       NULL))) {
992 			log(LOG_DEBUG, "%s; %s: global zone limit "
993 			    "reached, segment dropped\n", s, __func__);
994 			free(s, M_TCPLOG);
995 		}
996 		return (0);
997 	}
998 	tp->t_segqlen++;
999 	tp->t_rcvoopack++;
1000 	TCPSTAT_INC(tcps_rcvoopack);
1001 	TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
1002 	/* Insert the new segment queue entry into place. */
1003 	te->tqe_m = m;
1004 	te->tqe_flags = th->th_flags;
1005 	te->tqe_len = *tlenp;
1006 	te->tqe_start = th->th_seq;
1007 	te->tqe_last = mlast;
1008 	te->tqe_mbuf_cnt = lenofoh;
1009 	tp->t_segqmbuflen += te->tqe_mbuf_cnt;
1010 	if (p == NULL) {
1011 		TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
1012 	} else {
1013 		TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
1014 	}
1015 #ifdef TCP_REASS_LOGGING
1016 	tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_NEW_ENTRY, te);
1017 #endif
1018 present:
1019 	/*
1020 	 * Present data to user, advancing rcv_nxt through
1021 	 * completed sequence space.
1022 	 */
1023 	if (!TCPS_HAVEESTABLISHED(tp->t_state))
1024 		return (0);
1025 	q = TAILQ_FIRST(&tp->t_segq);
1026 	KASSERT(q == NULL || SEQ_GEQ(q->tqe_start, tp->rcv_nxt),
1027 		("Reassembly queue for %p has stale entry at head", tp));
1028 	if (!q || q->tqe_start != tp->rcv_nxt) {
1029 #ifdef TCP_REASS_LOGGING
1030 		tcp_reass_log_dump(tp);
1031 #endif
1032 		return (0);
1033 	}
1034 	SOCKBUF_LOCK(&so->so_rcv);
1035 	do {
1036 		tp->rcv_nxt += q->tqe_len;
1037 		flags = q->tqe_flags & TH_FIN;
1038 		nq = TAILQ_NEXT(q, tqe_q);
1039 		TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1040 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
1041 			m_freem(q->tqe_m);
1042 		} else {
1043 #ifdef TCP_REASS_LOGGING
1044 			tcp_reass_log_new_in(tp, q->tqe_start, q->tqe_len, q->tqe_m, TCP_R_LOG_READ, q);
1045 			tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 1);
1046 #endif
1047 			sbappendstream_locked(&so->so_rcv, q->tqe_m, 0);
1048 		}
1049 #ifdef TCP_REASS_LOGGING
1050 		tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 2);
1051 #endif
1052 		KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
1053 			("tp:%p seg queue goes negative", tp));
1054 		tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
1055 		uma_zfree(tcp_reass_zone, q);
1056 		tp->t_segqlen--;
1057 		q = nq;
1058 	} while (q && q->tqe_start == tp->rcv_nxt);
1059 	if (TAILQ_EMPTY(&tp->t_segq) &&
1060 	    (tp->t_segqmbuflen != 0)) {
1061 #ifdef INVARIANTS
1062 		panic("tp:%p segq:%p len:%d queue empty",
1063 		      tp, &tp->t_segq, tp->t_segqmbuflen);
1064 #else
1065 #ifdef TCP_REASS_LOGGING
1066 		tcp_log_reassm(tp, NULL, NULL, th->th_seq, *tlenp, TCP_R_LOG_ZERO, 0);
1067 #endif
1068 		tp->t_segqmbuflen = 0;
1069 #endif
1070 	}
1071 #ifdef TCP_REASS_LOGGING
1072 	tcp_reass_log_dump(tp);
1073 #endif
1074 	sorwakeup_locked(so);
1075 	return (flags);
1076 }
1077