]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_reass.c
Merge llvm-project release/15.x llvmorg-15.0.7-0-g8dfdcc7b7bf6
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_reass.c
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
5  *      The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
32  */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39
40 /* For debugging we want counters and BB logging */
41 /* #define TCP_REASS_COUNTERS 1 */
42 /* #define TCP_REASS_LOGGING 1 */
43
44 #include <sys/param.h>
45 #include <sys/kernel.h>
46 #include <sys/eventhandler.h>
47 #include <sys/malloc.h>
48 #include <sys/mbuf.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/syslog.h>
53 #include <sys/systm.h>
54
55 #include <vm/uma.h>
56
57 #include <net/if.h>
58 #include <net/if_var.h>
59 #include <net/route.h>
60 #include <net/vnet.h>
61
62 #include <netinet/in.h>
63 #include <netinet/in_pcb.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/in_var.h>
66 #include <netinet/ip.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/in6_pcb.h>
71 #include <netinet6/ip6_var.h>
72 #include <netinet6/nd6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/tcp_fsm.h>
75 #include <netinet/tcp_seq.h>
76 #include <netinet/tcp_timer.h>
77 #include <netinet/tcp_var.h>
78 #ifdef TCP_REASS_LOGGING
79 #include <netinet/tcp_log_buf.h>
80 #include <netinet/tcp_hpts.h>
81 #endif
82 #include <netinet/tcpip.h>
83
84 #define TCP_R_LOG_ADD           1
85 #define TCP_R_LOG_LIMIT_REACHED 2
86 #define TCP_R_LOG_APPEND        3
87 #define TCP_R_LOG_PREPEND       4
88 #define TCP_R_LOG_REPLACE       5
89 #define TCP_R_LOG_MERGE_INTO    6
90 #define TCP_R_LOG_NEW_ENTRY     7
91 #define TCP_R_LOG_READ          8
92 #define TCP_R_LOG_ZERO          9
93 #define TCP_R_LOG_DUMP          10
94 #define TCP_R_LOG_TRIM          11
95
96 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass,
97     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
98     "TCP Segment Reassembly Queue");
99
100 static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats,
101     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
102     "TCP Segment Reassembly stats");
103
104 static int tcp_reass_maxseg = 0;
105 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
106     &tcp_reass_maxseg, 0,
107     "Global maximum number of TCP Segments in Reassembly Queue");
108
109 static uma_zone_t tcp_reass_zone;
110 SYSCTL_UMA_CUR(_net_inet_tcp_reass, OID_AUTO, cursegments, 0,
111     &tcp_reass_zone,
112     "Global number of TCP Segments currently in Reassembly Queue");
113
114 static u_int tcp_reass_maxqueuelen = 100;
115 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
116     &tcp_reass_maxqueuelen, 0,
117     "Maximum number of TCP Segments per Reassembly Queue");
118
119 static int tcp_new_limits = 0;
120 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN,
121     &tcp_new_limits, 0,
122     "Do we use the new limit method we are discussing?");
123
124 static u_int tcp_reass_queue_guard = 16;
125 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN,
126     &tcp_reass_queue_guard, 16,
127     "Number of TCP Segments in Reassembly Queue where we flip over to guard mode");
128
129 #ifdef TCP_REASS_COUNTERS
130
131 counter_u64_t reass_entry;
132 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD,
133     &reass_entry, "A segment entered reassembly ");
134
135 counter_u64_t reass_path1;
136 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD,
137     &reass_path1, "Took path 1");
138
139 counter_u64_t reass_path2;
140 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD,
141     &reass_path2, "Took path 2");
142
143 counter_u64_t reass_path3;
144 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD,
145     &reass_path3, "Took path 3");
146
147 counter_u64_t reass_path4;
148 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD,
149     &reass_path4, "Took path 4");
150
151 counter_u64_t reass_path5;
152 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD,
153     &reass_path5, "Took path 5");
154
155 counter_u64_t reass_path6;
156 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD,
157     &reass_path6, "Took path 6");
158
159 counter_u64_t reass_path7;
160 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD,
161     &reass_path7, "Took path 7");
162
163 counter_u64_t reass_fullwalk;
164 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD,
165     &reass_fullwalk, "Took a full walk ");
166
167 counter_u64_t reass_nospace;
168 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD,
169     &reass_nospace, "Had no mbuf capacity ");
170
171 counter_u64_t merge_fwd;
172 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD,
173     &merge_fwd, "Ran merge fwd");
174
175 counter_u64_t merge_into;
176 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD,
177     &merge_into, "Ran merge into");
178
179 counter_u64_t tcp_zero_input;
180 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD,
181     &tcp_zero_input, "The reassembly buffer saw a zero len segment etc");
182
183 #endif
184
185 /* Initialize TCP reassembly queue */
186 static void
187 tcp_reass_zone_change(void *tag)
188 {
189
190         /* Set the zone limit and read back the effective value. */
191         tcp_reass_maxseg = nmbclusters / 16;
192         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
193             tcp_reass_maxseg);
194 }
195
196 #ifdef TCP_REASS_LOGGING
197
198 static void
199 tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p,
200     tcp_seq seq, int len, uint8_t action, int instance)
201 {
202         struct socket *so = tptosocket(tp);
203         uint32_t cts;
204         struct timeval tv;
205
206         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
207                 union tcp_log_stackspecific log;
208
209                 memset(&log, 0, sizeof(log));
210                 cts = tcp_get_usecs(&tv);
211                 log.u_bbr.flex1 = seq;
212                 log.u_bbr.cur_del_rate = (uint64_t)q;
213                 log.u_bbr.delRate = (uint64_t)p;
214                 if (q != NULL) {
215                         log.u_bbr.flex2 = q->tqe_start;
216                         log.u_bbr.flex3 = q->tqe_len;
217                         log.u_bbr.flex4 = q->tqe_mbuf_cnt;
218                         log.u_bbr.hptsi_gain = q->tqe_flags;
219                 }
220                 if (p != NULL)  {
221                         log.u_bbr.flex5 = p->tqe_start;
222                         log.u_bbr.pkts_out = p->tqe_len;
223                         log.u_bbr.epoch = p->tqe_mbuf_cnt;
224                         log.u_bbr.cwnd_gain = p->tqe_flags;
225                 }
226                 log.u_bbr.flex6 = tp->t_segqmbuflen;
227                 log.u_bbr.flex7 = instance;
228                 log.u_bbr.flex8 = action;
229                 log.u_bbr.timeStamp = cts;
230                 TCP_LOG_EVENTP(tp, NULL, &so->so_rcv, &so->so_snd,
231                     TCP_LOG_REASS, 0,
232                     len, &log, false, &tv);
233         }
234 }
235
236 static void
237 tcp_reass_log_dump(struct tcpcb *tp)
238 {
239         struct tseg_qent *q;
240
241         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
242                 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
243                         tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0);
244                 }
245         };
246 }
247
248 static void
249 tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m,
250     int logval, struct tseg_qent *q)
251 {
252         int cnt;
253         struct mbuf *t;
254
255         cnt = 0;
256         t = m;
257         while (t) {
258                 cnt += t->m_len;
259                 t = t->m_next;
260         }
261         tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt);
262 }
263
264 #endif
265
266 void
267 tcp_reass_global_init(void)
268 {
269
270         tcp_reass_maxseg = nmbclusters / 16;
271         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
272             &tcp_reass_maxseg);
273         tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
274             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
275         /* Set the zone limit and read back the effective value. */
276         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
277             tcp_reass_maxseg);
278 #ifdef TCP_REASS_COUNTERS
279         reass_path1 = counter_u64_alloc(M_WAITOK);
280         reass_path2 = counter_u64_alloc(M_WAITOK);
281         reass_path3 = counter_u64_alloc(M_WAITOK);
282         reass_path4 = counter_u64_alloc(M_WAITOK);
283         reass_path5 = counter_u64_alloc(M_WAITOK);
284         reass_path6 = counter_u64_alloc(M_WAITOK);
285         reass_path7 = counter_u64_alloc(M_WAITOK);
286         reass_fullwalk = counter_u64_alloc(M_WAITOK);
287         reass_nospace = counter_u64_alloc(M_WAITOK);
288         reass_entry = counter_u64_alloc(M_WAITOK);
289         merge_fwd = counter_u64_alloc(M_WAITOK);
290         merge_into = counter_u64_alloc(M_WAITOK);
291         tcp_zero_input = counter_u64_alloc(M_WAITOK);
292 #endif
293         EVENTHANDLER_REGISTER(nmbclusters_change,
294             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
295
296 }
297
298 void
299 tcp_reass_flush(struct tcpcb *tp)
300 {
301         struct tseg_qent *qe;
302
303         INP_WLOCK_ASSERT(tptoinpcb(tp));
304
305         while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) {
306                 TAILQ_REMOVE(&tp->t_segq, qe, tqe_q);
307                 m_freem(qe->tqe_m);
308                 uma_zfree(tcp_reass_zone, qe);
309                 tp->t_segqlen--;
310         }
311         tp->t_segqmbuflen = 0;
312         KASSERT((tp->t_segqlen == 0),
313             ("TCP reass queue %p segment count is %d instead of 0 after flush.",
314             tp, tp->t_segqlen));
315 }
316
317 static void
318 tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last,
319     struct mbuf *m, struct tcphdr *th, int tlen,
320     struct mbuf *mlast, int lenofoh)
321 {
322
323 #ifdef TCP_REASS_LOGGING
324         tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0);
325 #endif
326         last->tqe_len += tlen;
327         last->tqe_m->m_pkthdr.len += tlen;
328         /* Preserve the FIN bit if its there */
329         last->tqe_flags |= (tcp_get_flags(th) & TH_FIN);
330         last->tqe_last->m_next = m;
331         last->tqe_last = mlast;
332         last->tqe_mbuf_cnt += lenofoh;
333         tp->t_rcvoopack++;
334         TCPSTAT_INC(tcps_rcvoopack);
335         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
336 #ifdef TCP_REASS_LOGGING
337         tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m,
338                              TCP_R_LOG_APPEND,
339                              last);
340 #endif
341 }
342
343 static void
344 tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th,
345                   int tlen, struct mbuf *mlast, int lenofoh)
346 {
347         int i;
348
349 #ifdef TCP_REASS_LOGGING
350         tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0);
351 #endif
352         if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) {
353                 /* The new data overlaps into the old */
354                 i = (th->th_seq + tlen) - first->tqe_start;
355 #ifdef TCP_REASS_LOGGING
356                 tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1);
357 #endif
358                 m_adj(first->tqe_m, i);
359                 first->tqe_len -= i;
360                 first->tqe_start += i;
361         }
362         /* Ok now setup our chain to point to the old first */
363         mlast->m_next = first->tqe_m;
364         first->tqe_m = m;
365         first->tqe_len += tlen;
366         first->tqe_start = th->th_seq;
367         first->tqe_m->m_pkthdr.len = first->tqe_len;
368         first->tqe_mbuf_cnt += lenofoh;
369         tp->t_rcvoopack++;
370         TCPSTAT_INC(tcps_rcvoopack);
371         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
372 #ifdef TCP_REASS_LOGGING
373         tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m,
374                              TCP_R_LOG_PREPEND,
375                              first);
376 #endif
377 }
378
379 static void
380 tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m,
381     tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint16_t flags)
382 {
383         /*
384          * Free the data in q, and replace
385          * it with the new segment.
386          */
387         int len_dif;
388
389 #ifdef TCP_REASS_LOGGING
390         tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0);
391 #endif
392         m_freem(q->tqe_m);
393         KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
394                 ("Tp:%p seg queue goes negative", tp));
395         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
396         q->tqe_mbuf_cnt = mbufoh;
397         q->tqe_m = m;
398         q->tqe_last = mlast;
399         q->tqe_start = seq;
400         if (len > q->tqe_len)
401                 len_dif = len - q->tqe_len;
402         else
403                 len_dif = 0;
404         tp->t_rcvoopack++;
405         TCPSTAT_INC(tcps_rcvoopack);
406         TCPSTAT_ADD(tcps_rcvoobyte, len_dif);
407         q->tqe_len = len;
408         q->tqe_flags = (flags & TH_FIN);
409         q->tqe_m->m_pkthdr.len = q->tqe_len;
410         tp->t_segqmbuflen += mbufoh;
411
412 }
413
414 static void
415 tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent,
416     struct tseg_qent *q)
417 {
418         /*
419          * Merge q into ent and free q from the list.
420          */
421 #ifdef TCP_REASS_LOGGING
422         tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0);
423 #endif
424 #ifdef TCP_REASS_COUNTERS
425         counter_u64_add(merge_into, 1);
426 #endif
427         ent->tqe_last->m_next = q->tqe_m;
428         ent->tqe_last = q->tqe_last;
429         ent->tqe_len += q->tqe_len;
430         ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt;
431         ent->tqe_m->m_pkthdr.len += q->tqe_len;
432         ent->tqe_flags |= (q->tqe_flags & TH_FIN);
433         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
434         uma_zfree(tcp_reass_zone, q);
435         tp->t_segqlen--;
436
437 }
438
439 static void
440 tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent)
441 {
442         struct tseg_qent *q, *qtmp;
443         int i;
444         tcp_seq max;
445         /*
446          * Given an entry merge forward anyplace
447          * that ent overlaps forward.
448          */
449
450         max = ent->tqe_start + ent->tqe_len;
451         q = TAILQ_NEXT(ent, tqe_q);
452         if (q == NULL) {
453                 /* Nothing left */
454                 return;
455         }
456         TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) {
457                 if (SEQ_GT(q->tqe_start, max)) {
458                         /* Beyond q */
459                         break;
460                 }
461                 /* We have some or all that are overlapping */
462                 if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) {
463                         /* It consumes it all */
464                         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
465                         m_freem(q->tqe_m);
466                         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
467                         uma_zfree(tcp_reass_zone, q);
468                         tp->t_segqlen--;
469                         continue;
470                 }
471                 /*
472                  * Trim the q entry to dovetail to this one
473                  * and then merge q into ent updating max
474                  * in the process.
475                  */
476                 i = max - q->tqe_start;
477 #ifdef TCP_REASS_LOGGING
478                 tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2);
479 #endif
480                 m_adj(q->tqe_m, i);
481                 q->tqe_len -= i;
482                 q->tqe_start += i;
483                 tcp_reass_merge_into(tp, ent, q);
484                 max = ent->tqe_start + ent->tqe_len;
485         }
486 #ifdef TCP_REASS_COUNTERS
487         counter_u64_add(merge_fwd, 1);
488 #endif
489 }
490
491 static int
492 tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast)
493 {
494         int len = MSIZE;
495
496         if (m->m_flags & M_EXT)
497                 len += m->m_ext.ext_size;
498         while (m->m_next != NULL) {
499                 m = m->m_next;
500                 len += MSIZE;
501                 if (m->m_flags & M_EXT)
502                         len += m->m_ext.ext_size;
503         }
504         *mlast = m;
505         return (len);
506 }
507
508 /*
509  * NOTE!!! the new tcp-reassembly code *must not* use
510  * m_adj() with a negative index. That alters the chain
511  * of mbufs (by possibly chopping trailing mbufs). At
512  * the front of tcp_reass we count the mbuf overhead
513  * and setup the tail pointer. If we use m_adj(m, -5)
514  * we could corrupt the tail pointer. Currently the
515  * code only uses m_adj(m, postive-num). If this
516  * changes appropriate changes to update mlast would
517  * be needed.
518  */
519 int
520 tcp_reass(struct tcpcb *tp, struct tcphdr *th, tcp_seq *seq_start,
521           int *tlenp, struct mbuf *m)
522 {
523         struct tseg_qent *q, *last, *first;
524         struct tseg_qent *p = NULL;
525         struct tseg_qent *nq = NULL;
526         struct tseg_qent *te = NULL;
527         struct mbuf *mlast = NULL;
528         struct inpcb *inp = tptoinpcb(tp);
529         struct socket *so = tptosocket(tp);
530         struct sockbuf *sb = &so->so_rcv;
531         char *s = NULL;
532         int flags, i, lenofoh;
533
534         INP_WLOCK_ASSERT(inp);
535         /*
536          * XXX: tcp_reass() is rather inefficient with its data structures
537          * and should be rewritten (see NetBSD for optimizations).
538          */
539
540         KASSERT(th == NULL || (seq_start != NULL && tlenp != NULL),
541                 ("tcp_reass called with illegal parameter combination "
542                  "(tp=%p, th=%p, seq_start=%p, tlenp=%p, m=%p)",
543                  tp, th, seq_start, tlenp, m));
544         /*
545          * Call with th==NULL after become established to
546          * force pre-ESTABLISHED data up to user socket.
547          */
548         if (th == NULL)
549                 goto present;
550         KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt),
551                 ("Attempt to add old entry to reassembly queue (th=%p, tp=%p)",
552                  th, tp));
553 #ifdef TCP_REASS_LOGGING
554         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL);
555 #endif
556 #ifdef TCP_REASS_COUNTERS
557         counter_u64_add(reass_entry, 1);
558 #endif
559         /*
560          * Check for zero length data.
561          */
562         if ((*tlenp == 0) && ((tcp_get_flags(th) & TH_FIN) == 0)) {
563                 /*
564                  * A zero length segment does no
565                  * one any good. We could check
566                  * the rcv_nxt <-> rcv_wnd but thats
567                  * already done for us by the caller.
568                  */
569 strip_fin:
570 #ifdef TCP_REASS_COUNTERS
571                 counter_u64_add(tcp_zero_input, 1);
572 #endif
573                 m_freem(m);
574 #ifdef TCP_REASS_LOGGING
575                 tcp_reass_log_dump(tp);
576 #endif
577                 return (0);
578         } else if ((*tlenp == 0) &&
579                    (tcp_get_flags(th) & TH_FIN) &&
580                    !TCPS_HAVEESTABLISHED(tp->t_state)) {
581                 /*
582                  * We have not established, and we
583                  * have a FIN and no data. Lets treat
584                  * this as the same as if the FIN were
585                  * not present. We don't want to save
586                  * the FIN bit in a reassembly buffer
587                  * we want to get established first before
588                  * we do that (the peer will retransmit).
589                  */
590                 goto strip_fin;
591         }
592         /*
593          * Will it fit?
594          */
595         lenofoh = tcp_reass_overhead_of_chain(m, &mlast);
596         if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
597             (sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) {
598                 /* No room */
599                 TCPSTAT_INC(tcps_rcvreassfull);
600 #ifdef TCP_REASS_COUNTERS
601                 counter_u64_add(reass_nospace, 1);
602 #endif
603 #ifdef TCP_REASS_LOGGING
604                 tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0);
605 #endif
606                 if ((s = tcp_log_addrs(&inp->inp_inc, th, NULL, NULL))) {
607                         log(LOG_DEBUG, "%s; %s: mbuf count limit reached, "
608                             "segment dropped\n", s, __func__);
609                         free(s, M_TCPLOG);
610                 }
611                 m_freem(m);
612                 *tlenp = 0;
613 #ifdef TCP_REASS_LOGGING
614                 tcp_reass_log_dump(tp);
615 #endif
616                 return (0);
617         }
618         /*
619          * First lets deal with two common cases, the
620          * segment appends to the back of our collected
621          * segments. Or the segment is the next in line.
622          */
623         last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q);
624         if (last != NULL) {
625                 if ((tcp_get_flags(th) & TH_FIN) &&
626                     SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) {
627                         /*
628                          * Someone is trying to game us, dump
629                          * the segment.
630                          */
631                         *tlenp = 0;
632                         m_freem(m);
633                         return (0);
634                 }
635                 if ((SEQ_GEQ(th->th_seq, last->tqe_start)) &&
636                     (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) {
637                         /* Common case, trailing segment is added */
638                         /**
639                          *                                 +--last
640                          *                                 v
641                          *  reassembly buffer |---|  |---| |---|
642                          *  new segment                       |---|
643                          */
644 #ifdef TCP_REASS_COUNTERS
645                         counter_u64_add(reass_path1, 1);
646 #endif
647                         if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) {
648                                 i = (last->tqe_start + last->tqe_len) - th->th_seq;
649                                 if (i < *tlenp) {
650 #ifdef TCP_REASS_LOGGING
651                                         tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3);
652                                         th->th_seq += i;
653 #endif
654                                         m_adj(m, i);
655                                         *tlenp -= i;
656                                 } else {
657                                         /* Complete overlap */
658                                         TCPSTAT_INC(tcps_rcvduppack);
659                                         TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
660                                         m_freem(m);
661                                         *tlenp = last->tqe_len;
662                                         *seq_start = last->tqe_start;
663                                         return (0);
664                                 }
665                         }
666                         if (last->tqe_flags & TH_FIN) {
667                                 /*
668                                  * We have data after the FIN on the last?
669                                  */
670                                 *tlenp = 0;
671                                 m_freem(m);
672                                 return(0);
673                         }
674                         tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh);
675                         tp->t_segqmbuflen += lenofoh;
676                         *seq_start = last->tqe_start;
677                         *tlenp = last->tqe_len;
678                         return (0);
679                 } else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) {
680                         /*
681                          * Second common case, we missed
682                          * another one and have something more
683                          * for the end.
684                          */
685                         /**
686                          *                                 +--last
687                          *                                 v
688                          *  reassembly buffer |---|  |---| |---|
689                          *  new segment                           |---|
690                          */
691                         if (last->tqe_flags & TH_FIN) {
692                                 /*
693                                  * We have data after the FIN on the last?
694                                  */
695                                 *tlenp = 0;
696                                 m_freem(m);
697                                 return(0);
698                         }
699 #ifdef TCP_REASS_COUNTERS
700                         counter_u64_add(reass_path2, 1);
701 #endif
702                         p = last;
703                         goto new_entry;
704                 }
705         } else {
706                 /* First segment (it's NULL). */
707                 goto new_entry;
708         }
709         first = TAILQ_FIRST(&tp->t_segq);
710         if (SEQ_LT(th->th_seq, first->tqe_start) &&
711             SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) &&
712             SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) {
713                 /*
714                  * The head of the queue is prepended by this and
715                  * it may be the one I want most.
716                  */
717                 /**
718                  *       first-------+
719                  *                   v
720                  *  rea:             |---|  |---| |---|
721                  *  new:         |---|
722                  * Note the case we do not deal with here is:
723                  *   rea=     |---|   |---|   |---|
724                  *   new=  |----|
725                  * Due to the fact that it could be
726                  *   new   |--------------------|
727                  * And we might need to merge forward.
728                  */
729 #ifdef INVARIANTS
730                 struct mbuf *firstmbuf;
731 #endif
732
733 #ifdef TCP_REASS_COUNTERS
734                 counter_u64_add(reass_path3, 1);
735 #endif
736                 if (SEQ_LT(th->th_seq, tp->rcv_nxt)) {
737                         /*
738                          * The resend was even before
739                          * what we have. We need to trim it.
740                          * Note TSNH (it should be trimmed
741                          * before the call to tcp_reass()).
742                          */
743 #ifdef INVARIANTS
744                         panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed",
745                               th->th_seq, tp->rcv_nxt, tp);
746 #else
747                         i = tp->rcv_nxt - th->th_seq;
748 #ifdef TCP_REASS_LOGGING
749                         tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4);
750 #endif
751                         m_adj(m, i);
752                         th->th_seq += i;
753                         *tlenp -= i;
754 #endif
755                 }
756 #ifdef INVARIANTS
757                 firstmbuf = first->tqe_m;
758 #endif
759                 tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh);
760 #ifdef INVARIANTS
761                 if (firstmbuf == first->tqe_m) {
762                         panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
763                               m, firstmbuf, first->tqe_m, tp, first);
764                 } else if (first->tqe_m != m) {
765                         panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
766                               m, firstmbuf, first->tqe_m, tp, first);
767                 }
768 #endif
769                 tp->t_segqmbuflen += lenofoh;
770                 *seq_start = first->tqe_start;
771                 *tlenp = first->tqe_len;
772                 goto present;
773         } else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) {
774                 /* New segment is before our earliest segment. */
775                 /**
776                  *           first---->+
777                  *                      v
778                  *  rea=                |---| ....
779                  *  new"         |---|
780                  *
781                  */
782                 goto new_entry;
783         }
784         /*
785          * Find a segment which begins after this one does.
786          */
787 #ifdef TCP_REASS_COUNTERS
788         counter_u64_add(reass_fullwalk, 1);
789 #endif
790         TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
791                 if (SEQ_GT(q->tqe_start, th->th_seq))
792                         break;
793         }
794         p = TAILQ_PREV(q, tsegqe_head, tqe_q);
795         /**
796          * Now is this fit just in-between only?
797          * i.e.:
798          *      p---+        +----q
799          *          v        v
800          *     res= |--|     |--|    |--|
801          *     nee       |-|
802          */
803         if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) &&
804             ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) {
805                 /* Yep no overlap */
806                 goto new_entry;
807         }
808         /**
809          * If we reach here we have some (possibly all) overlap
810          * such as:
811          *     res=     |--|     |--|    |--|
812          *     new=  |----|
813          * or  new=  |-----------------|
814          * or  new=      |--------|
815          * or  new=            |---|
816          * or  new=            |-----------|
817          */
818         if ((p != NULL) &&
819             (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) {
820                 /* conversion to int (in i) handles seq wraparound */
821
822 #ifdef TCP_REASS_COUNTERS
823                 counter_u64_add(reass_path4, 1);
824 #endif
825                 i = p->tqe_start + p->tqe_len - th->th_seq;
826                 if (i >= 0) {
827                         if (i >= *tlenp) {
828                                 /**
829                                  *       prev seg---->+
830                                  *                    v
831                                  *  reassembly buffer |---|
832                                  *  new segment        |-|
833                                  */
834                                 TCPSTAT_INC(tcps_rcvduppack);
835                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
836                                 *tlenp = p->tqe_len;
837                                 *seq_start = p->tqe_start;
838                                 m_freem(m);
839                                 /*
840                                  * Try to present any queued data
841                                  * at the left window edge to the user.
842                                  * This is needed after the 3-WHS
843                                  * completes. Note this probably
844                                  * will not work and we will return.
845                                  */
846                                 return (0);
847                         }
848                         if (i > 0) {
849                                 /**
850                                  *       prev seg---->+
851                                  *                    v
852                                  *  reassembly buffer |---|
853                                  *  new segment         |-----|
854                                  */
855 #ifdef TCP_REASS_COUNTERS
856                                 counter_u64_add(reass_path5, 1);
857 #endif
858 #ifdef TCP_REASS_LOGGING
859                                 tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5);
860 #endif
861                                 m_adj(m, i);
862                                 *tlenp -= i;
863                                 th->th_seq += i;
864                         }
865                 }
866                 if (th->th_seq == (p->tqe_start + p->tqe_len)) {
867                         /*
868                          * If dovetails in with this one
869                          * append it.
870                          */
871                         /**
872                          *       prev seg---->+
873                          *                    v
874                          *  reassembly buffer |--|     |---|
875                          *  new segment          |--|
876                          * (note: it was trimmed above if it overlapped)
877                          */
878                         tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh);
879                         tp->t_segqmbuflen += lenofoh;
880                 } else {
881 #ifdef INVARIANTS
882                         panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p",
883                               th->th_seq, p->tqe_start, p->tqe_len,
884                               p, tp);
885 #endif
886                         *tlenp = 0;
887                         m_freem(m);
888                         return (0);
889                 }
890                 q = p;
891         } else {
892                 /*
893                  * The new data runs over the
894                  * top of previously sack'd data (in q).
895                  * It may be partially overlapping, or
896                  * it may overlap the entire segment.
897                  */
898 #ifdef TCP_REASS_COUNTERS
899                 counter_u64_add(reass_path6, 1);
900 #endif
901                 if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) {
902                         /* It consumes it all */
903                         /**
904                          *             next seg---->+
905                          *                          v
906                          *  reassembly buffer |--|     |---|
907                          *  new segment              |----------|
908                          */
909 #ifdef TCP_REASS_COUNTERS
910                         counter_u64_add(reass_path7, 1);
911 #endif
912                         tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, tcp_get_flags(th));
913                 } else {
914                         /*
915                          * We just need to prepend the data
916                          * to this. It does not overrun
917                          * the end.
918                          */
919                         /**
920                          *                next seg---->+
921                          *                             v
922                          *  reassembly buffer |--|     |---|
923                          *  new segment                   |----------|
924                          */
925                         tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh);
926                         tp->t_segqmbuflen += lenofoh;
927                 }
928         }
929         /* Now does it go further than that? */
930         tcp_reass_merge_forward(tp, q);
931         *seq_start = q->tqe_start;
932         *tlenp = q->tqe_len;
933         goto present;
934
935         /*
936          * When we reach here we can't combine it
937          * with any existing segment.
938          *
939          * Limit the number of segments that can be queued to reduce the
940          * potential for mbuf exhaustion. For best performance, we want to be
941          * able to queue a full window's worth of segments. The size of the
942          * socket receive buffer determines our advertised window and grows
943          * automatically when socket buffer autotuning is enabled. Use it as the
944          * basis for our queue limit.
945          *
946          * However, allow the user to specify a ceiling for the number of
947          * segments in each queue.
948          *
949          * Always let the missing segment through which caused this queue.
950          * NB: Access to the socket buffer is left intentionally unlocked as we
951          * can tolerate stale information here.
952          *
953          * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
954          * should work but causes packets to be dropped when they shouldn't.
955          * Investigate why and re-evaluate the below limit after the behaviour
956          * is understood.
957          */
958 new_entry:
959         if (th->th_seq == tp->rcv_nxt && TCPS_HAVEESTABLISHED(tp->t_state)) {
960                 tp->rcv_nxt += *tlenp;
961                 flags = tcp_get_flags(th) & TH_FIN;
962                 TCPSTAT_INC(tcps_rcvoopack);
963                 TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
964                 SOCKBUF_LOCK(&so->so_rcv);
965                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
966                         m_freem(m);
967                 } else {
968                         sbappendstream_locked(&so->so_rcv, m, 0);
969                 }
970                 tp->t_flags |= TF_WAKESOR;
971                 return (flags);
972         }
973         if (tcp_new_limits) {
974                 if ((tp->t_segqlen > tcp_reass_queue_guard) &&
975                     (*tlenp < MSIZE)) {
976                         /*
977                          * This is really a lie, we are not full but
978                          * are getting a segment that is above
979                          * guard threshold. If it is and its below
980                          * a mbuf size (256) we drop it if it
981                          * can't fill in some place.
982                          */
983                         TCPSTAT_INC(tcps_rcvreassfull);
984                         *tlenp = 0;
985                         if ((s = tcp_log_addrs(&inp->inp_inc, th, NULL, NULL))) {
986                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
987                                     "segment dropped\n", s, __func__);
988                                 free(s, M_TCPLOG);
989                         }
990                         m_freem(m);
991 #ifdef TCP_REASS_LOGGING
992                         tcp_reass_log_dump(tp);
993 #endif
994                         return (0);
995                 }
996         } else {
997                 if (tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
998                                          tcp_reass_maxqueuelen)) {
999                         TCPSTAT_INC(tcps_rcvreassfull);
1000                         *tlenp = 0;
1001                         if ((s = tcp_log_addrs(&inp->inp_inc, th, NULL, NULL))) {
1002                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
1003                                     "segment dropped\n", s, __func__);
1004                                 free(s, M_TCPLOG);
1005                         }
1006                         m_freem(m);
1007 #ifdef TCP_REASS_LOGGING
1008                         tcp_reass_log_dump(tp);
1009 #endif
1010                         return (0);
1011                 }
1012         }
1013         /*
1014          * Allocate a new queue entry. If we can't, or hit the zone limit
1015          * just drop the pkt.
1016          */
1017         te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
1018         if (te == NULL) {
1019                 TCPSTAT_INC(tcps_rcvmemdrop);
1020                 m_freem(m);
1021                 *tlenp = 0;
1022                 if ((s = tcp_log_addrs(&inp->inp_inc, th, NULL, NULL))) {
1023                         log(LOG_DEBUG, "%s; %s: global zone limit "
1024                             "reached, segment dropped\n", s, __func__);
1025                         free(s, M_TCPLOG);
1026                 }
1027                 return (0);
1028         }
1029         tp->t_segqlen++;
1030         tp->t_rcvoopack++;
1031         TCPSTAT_INC(tcps_rcvoopack);
1032         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
1033         /* Insert the new segment queue entry into place. */
1034         te->tqe_m = m;
1035         te->tqe_flags = tcp_get_flags(th);
1036         te->tqe_len = *tlenp;
1037         te->tqe_start = th->th_seq;
1038         te->tqe_last = mlast;
1039         te->tqe_mbuf_cnt = lenofoh;
1040         tp->t_segqmbuflen += te->tqe_mbuf_cnt;
1041         if (p == NULL) {
1042                 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
1043         } else {
1044                 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
1045         }
1046 #ifdef TCP_REASS_LOGGING
1047         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_NEW_ENTRY, te);
1048 #endif
1049 present:
1050         /*
1051          * Present data to user, advancing rcv_nxt through
1052          * completed sequence space.
1053          */
1054         if (!TCPS_HAVEESTABLISHED(tp->t_state))
1055                 return (0);
1056         q = TAILQ_FIRST(&tp->t_segq);
1057         KASSERT(q == NULL || SEQ_GEQ(q->tqe_start, tp->rcv_nxt),
1058                 ("Reassembly queue for %p has stale entry at head", tp));
1059         if (!q || q->tqe_start != tp->rcv_nxt) {
1060 #ifdef TCP_REASS_LOGGING
1061                 tcp_reass_log_dump(tp);
1062 #endif
1063                 return (0);
1064         }
1065         SOCKBUF_LOCK(&so->so_rcv);
1066         do {
1067                 tp->rcv_nxt += q->tqe_len;
1068                 flags = q->tqe_flags & TH_FIN;
1069                 nq = TAILQ_NEXT(q, tqe_q);
1070                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1071                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
1072                         m_freem(q->tqe_m);
1073                 } else {
1074 #ifdef TCP_REASS_LOGGING
1075                         tcp_reass_log_new_in(tp, q->tqe_start, q->tqe_len, q->tqe_m, TCP_R_LOG_READ, q);
1076                         if (th != NULL) {
1077                                 tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 1);
1078                         } else {
1079                                 tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 1);
1080                         }
1081 #endif
1082                         sbappendstream_locked(&so->so_rcv, q->tqe_m, 0);
1083                 }
1084 #ifdef TCP_REASS_LOGGING
1085                 if (th != NULL) {
1086                         tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 2);
1087                 } else {
1088                         tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 2);
1089                 }
1090 #endif
1091                 KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
1092                         ("tp:%p seg queue goes negative", tp));
1093                 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
1094                 uma_zfree(tcp_reass_zone, q);
1095                 tp->t_segqlen--;
1096                 q = nq;
1097         } while (q && q->tqe_start == tp->rcv_nxt);
1098         if (TAILQ_EMPTY(&tp->t_segq) &&
1099             (tp->t_segqmbuflen != 0)) {
1100 #ifdef INVARIANTS
1101                 panic("tp:%p segq:%p len:%d queue empty",
1102                       tp, &tp->t_segq, tp->t_segqmbuflen);
1103 #else
1104 #ifdef TCP_REASS_LOGGING
1105                 if (th != NULL) {
1106                         tcp_log_reassm(tp, NULL, NULL, th->th_seq, *tlenp, TCP_R_LOG_ZERO, 0);
1107                 } else {
1108                         tcp_log_reassm(tp, NULL, NULL, 0, 0, TCP_R_LOG_ZERO, 0);
1109                 }
1110 #endif
1111                 tp->t_segqmbuflen = 0;
1112 #endif
1113         }
1114 #ifdef TCP_REASS_LOGGING
1115         tcp_reass_log_dump(tp);
1116 #endif
1117         tp->t_flags |= TF_WAKESOR;
1118         return (flags);
1119 }