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