]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/netinet/tcp_reass.c
Move SYSCTL_ADD_PROC() to unlocked context in if_ure to avoid lock order reversal.
[FreeBSD/FreeBSD.git] / sys / netinet / tcp_reass.c
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
5  *      The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
32  */
33
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36
37 #include "opt_inet.h"
38 #include "opt_inet6.h"
39 #include "opt_tcpdebug.h"
40
41 /* For debugging we want counters and BB logging */
42 /* #define TCP_REASS_COUNTERS 1 */
43 /* #define TCP_REASS_LOGGING 1 */
44
45 #include <sys/param.h>
46 #include <sys/kernel.h>
47 #include <sys/eventhandler.h>
48 #include <sys/malloc.h>
49 #include <sys/mbuf.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/syslog.h>
54 #include <sys/systm.h>
55
56 #include <vm/uma.h>
57
58 #include <net/if.h>
59 #include <net/if_var.h>
60 #include <net/route.h>
61 #include <net/vnet.h>
62
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/in_systm.h>
66 #include <netinet/in_var.h>
67 #include <netinet/ip.h>
68 #include <netinet/ip_var.h>
69 #include <netinet/ip_options.h>
70 #include <netinet/ip6.h>
71 #include <netinet6/in6_pcb.h>
72 #include <netinet6/ip6_var.h>
73 #include <netinet6/nd6.h>
74 #include <netinet/tcp.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/tcp_seq.h>
77 #include <netinet/tcp_timer.h>
78 #include <netinet/tcp_var.h>
79 #ifdef TCP_REASS_LOGGING
80 #include <netinet/tcp_log_buf.h>
81 #include <netinet/tcp_hpts.h>
82 #endif
83 #include <netinet6/tcp6_var.h>
84 #include <netinet/tcpip.h>
85 #ifdef TCPDEBUG
86 #include <netinet/tcp_debug.h>
87 #endif /* TCPDEBUG */
88
89 #define TCP_R_LOG_ADD           1
90 #define TCP_R_LOG_LIMIT_REACHED 2
91 #define TCP_R_LOG_APPEND        3
92 #define TCP_R_LOG_PREPEND       4
93 #define TCP_R_LOG_REPLACE       5
94 #define TCP_R_LOG_MERGE_INTO    6
95 #define TCP_R_LOG_NEW_ENTRY     7
96 #define TCP_R_LOG_READ          8
97 #define TCP_R_LOG_ZERO          9
98 #define TCP_R_LOG_DUMP          10
99 #define TCP_R_LOG_TRIM          11
100
101 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass,
102     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
103     "TCP Segment Reassembly Queue");
104
105 static SYSCTL_NODE(_net_inet_tcp_reass, OID_AUTO, stats,
106     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
107     "TCP Segment Reassembly stats");
108
109 static int tcp_reass_maxseg = 0;
110 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
111     &tcp_reass_maxseg, 0,
112     "Global maximum number of TCP Segments in Reassembly Queue");
113
114 static uma_zone_t tcp_reass_zone;
115 SYSCTL_UMA_CUR(_net_inet_tcp_reass, OID_AUTO, cursegments, 0,
116     &tcp_reass_zone,
117     "Global number of TCP Segments currently in Reassembly Queue");
118
119 static u_int tcp_reass_maxqueuelen = 100;
120 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, maxqueuelen, CTLFLAG_RWTUN,
121     &tcp_reass_maxqueuelen, 0,
122     "Maximum number of TCP Segments per Reassembly Queue");
123
124 static int tcp_new_limits = 0;
125 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, new_limit, CTLFLAG_RWTUN,
126     &tcp_new_limits, 0,
127     "Do we use the new limit method we are discussing?");
128
129 static u_int tcp_reass_queue_guard = 16;
130 SYSCTL_UINT(_net_inet_tcp_reass, OID_AUTO, queueguard, CTLFLAG_RWTUN,
131     &tcp_reass_queue_guard, 16,
132     "Number of TCP Segments in Reassembly Queue where we flip over to guard mode");
133
134 #ifdef TCP_REASS_COUNTERS
135
136 counter_u64_t reass_entry;
137 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, entry, CTLFLAG_RD,
138     &reass_entry, "A segment entered reassembly ");
139
140 counter_u64_t reass_path1;
141 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path1, CTLFLAG_RD,
142     &reass_path1, "Took path 1");
143
144 counter_u64_t reass_path2;
145 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path2, CTLFLAG_RD,
146     &reass_path2, "Took path 2");
147
148 counter_u64_t reass_path3;
149 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path3, CTLFLAG_RD,
150     &reass_path3, "Took path 3");
151
152 counter_u64_t reass_path4;
153 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path4, CTLFLAG_RD,
154     &reass_path4, "Took path 4");
155
156 counter_u64_t reass_path5;
157 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path5, CTLFLAG_RD,
158     &reass_path5, "Took path 5");
159
160 counter_u64_t reass_path6;
161 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path6, CTLFLAG_RD,
162     &reass_path6, "Took path 6");
163
164 counter_u64_t reass_path7;
165 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, path7, CTLFLAG_RD,
166     &reass_path7, "Took path 7");
167
168 counter_u64_t reass_fullwalk;
169 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, fullwalk, CTLFLAG_RD,
170     &reass_fullwalk, "Took a full walk ");
171
172 counter_u64_t reass_nospace;
173 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, nospace, CTLFLAG_RD,
174     &reass_nospace, "Had no mbuf capacity ");
175
176 counter_u64_t merge_fwd;
177 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_fwd, CTLFLAG_RD,
178     &merge_fwd, "Ran merge fwd");
179
180 counter_u64_t merge_into;
181 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, merge_into, CTLFLAG_RD,
182     &merge_into, "Ran merge into");
183
184 counter_u64_t tcp_zero_input;
185 SYSCTL_COUNTER_U64(_net_inet_tcp_reass_stats, OID_AUTO, zero_input, CTLFLAG_RD,
186     &tcp_zero_input, "The reassembly buffer saw a zero len segment etc");
187
188 #endif
189
190 /* Initialize TCP reassembly queue */
191 static void
192 tcp_reass_zone_change(void *tag)
193 {
194
195         /* Set the zone limit and read back the effective value. */
196         tcp_reass_maxseg = nmbclusters / 16;
197         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
198             tcp_reass_maxseg);
199 }
200
201 #ifdef TCP_REASS_LOGGING
202
203 static void
204 tcp_log_reassm(struct tcpcb *tp, struct tseg_qent *q, struct tseg_qent *p,
205     tcp_seq seq, int len, uint8_t action, int instance)
206 {
207         uint32_t cts;
208         struct timeval tv;
209
210         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
211                 union tcp_log_stackspecific log;
212
213                 memset(&log, 0, sizeof(log));
214                 cts = tcp_get_usecs(&tv);
215                 log.u_bbr.flex1 = seq;
216                 log.u_bbr.cur_del_rate = (uint64_t)q;
217                 log.u_bbr.delRate = (uint64_t)p;
218                 if (q != NULL) {
219                         log.u_bbr.flex2 = q->tqe_start;
220                         log.u_bbr.flex3 = q->tqe_len;
221                         log.u_bbr.flex4 = q->tqe_mbuf_cnt;
222                         log.u_bbr.hptsi_gain = q->tqe_flags;
223                 }
224                 if (p != NULL)  {
225                         log.u_bbr.flex5 = p->tqe_start;
226                         log.u_bbr.pkts_out = p->tqe_len;
227                         log.u_bbr.epoch = p->tqe_mbuf_cnt;
228                         log.u_bbr.cwnd_gain = p->tqe_flags;
229                 }
230                 log.u_bbr.flex6 = tp->t_segqmbuflen;
231                 log.u_bbr.flex7 = instance;
232                 log.u_bbr.flex8 = action;
233                 log.u_bbr.timeStamp = cts;
234                 TCP_LOG_EVENTP(tp, NULL,
235                     &tp->t_inpcb->inp_socket->so_rcv,
236                     &tp->t_inpcb->inp_socket->so_snd,
237                     TCP_LOG_REASS, 0,
238                     len, &log, false, &tv);
239         }
240 }
241
242 static void
243 tcp_reass_log_dump(struct tcpcb *tp)
244 {
245         struct tseg_qent *q;
246
247         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
248                 TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
249                         tcp_log_reassm(tp, q, NULL, q->tqe_start, q->tqe_len, TCP_R_LOG_DUMP, 0);
250                 }
251         };
252 }
253
254 static void
255 tcp_reass_log_new_in(struct tcpcb *tp, tcp_seq seq, int len, struct mbuf *m,
256     int logval, struct tseg_qent *q)
257 {
258         int cnt;
259         struct mbuf *t;
260
261         cnt = 0;
262         t = m;
263         while (t) {
264                 cnt += t->m_len;
265                 t = t->m_next;
266         }
267         tcp_log_reassm(tp, q, NULL, seq, len, logval, cnt);
268 }
269
270 #endif
271
272 void
273 tcp_reass_global_init(void)
274 {
275
276         tcp_reass_maxseg = nmbclusters / 16;
277         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
278             &tcp_reass_maxseg);
279         tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
280             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
281         /* Set the zone limit and read back the effective value. */
282         tcp_reass_maxseg = uma_zone_set_max(tcp_reass_zone,
283             tcp_reass_maxseg);
284 #ifdef TCP_REASS_COUNTERS
285         reass_path1 = counter_u64_alloc(M_WAITOK);
286         reass_path2 = counter_u64_alloc(M_WAITOK);
287         reass_path3 = counter_u64_alloc(M_WAITOK);
288         reass_path4 = counter_u64_alloc(M_WAITOK);
289         reass_path5 = counter_u64_alloc(M_WAITOK);
290         reass_path6 = counter_u64_alloc(M_WAITOK);
291         reass_path7 = counter_u64_alloc(M_WAITOK);
292         reass_fullwalk = counter_u64_alloc(M_WAITOK);
293         reass_nospace = counter_u64_alloc(M_WAITOK);
294         reass_entry = counter_u64_alloc(M_WAITOK);
295         merge_fwd = counter_u64_alloc(M_WAITOK);
296         merge_into = counter_u64_alloc(M_WAITOK);
297         tcp_zero_input = counter_u64_alloc(M_WAITOK);
298 #endif
299         EVENTHANDLER_REGISTER(nmbclusters_change,
300             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
301
302 }
303
304 void
305 tcp_reass_flush(struct tcpcb *tp)
306 {
307         struct tseg_qent *qe;
308
309         INP_WLOCK_ASSERT(tp->t_inpcb);
310
311         while ((qe = TAILQ_FIRST(&tp->t_segq)) != NULL) {
312                 TAILQ_REMOVE(&tp->t_segq, qe, tqe_q);
313                 m_freem(qe->tqe_m);
314                 uma_zfree(tcp_reass_zone, qe);
315                 tp->t_segqlen--;
316         }
317         tp->t_segqmbuflen = 0;
318         KASSERT((tp->t_segqlen == 0),
319             ("TCP reass queue %p segment count is %d instead of 0 after flush.",
320             tp, tp->t_segqlen));
321 }
322
323 static void
324 tcp_reass_append(struct tcpcb *tp, struct tseg_qent *last,
325     struct mbuf *m, struct tcphdr *th, int tlen,
326     struct mbuf *mlast, int lenofoh)
327 {
328
329 #ifdef TCP_REASS_LOGGING
330         tcp_log_reassm(tp, last, NULL, th->th_seq, tlen, TCP_R_LOG_APPEND, 0);
331 #endif
332         last->tqe_len += tlen;
333         last->tqe_m->m_pkthdr.len += tlen;
334         /* Preserve the FIN bit if its there */
335         last->tqe_flags |= (th->th_flags & TH_FIN);
336         last->tqe_last->m_next = m;
337         last->tqe_last = mlast;
338         last->tqe_mbuf_cnt += lenofoh;
339         tp->t_rcvoopack++;
340         TCPSTAT_INC(tcps_rcvoopack);
341         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
342 #ifdef TCP_REASS_LOGGING
343         tcp_reass_log_new_in(tp, last->tqe_start, lenofoh, last->tqe_m,
344                              TCP_R_LOG_APPEND,
345                              last);
346 #endif
347 }
348
349 static void
350 tcp_reass_prepend(struct tcpcb *tp, struct tseg_qent *first, struct mbuf *m, struct tcphdr *th,
351                   int tlen, struct mbuf *mlast, int lenofoh)
352 {
353         int i;
354
355 #ifdef TCP_REASS_LOGGING
356         tcp_log_reassm(tp, first, NULL, th->th_seq, tlen, TCP_R_LOG_PREPEND, 0);
357 #endif
358         if (SEQ_GT((th->th_seq + tlen), first->tqe_start)) {
359                 /* The new data overlaps into the old */
360                 i = (th->th_seq + tlen) - first->tqe_start;
361 #ifdef TCP_REASS_LOGGING
362                 tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 1);
363 #endif
364                 m_adj(first->tqe_m, i);
365                 first->tqe_len -= i;
366                 first->tqe_start += i;
367         }
368         /* Ok now setup our chain to point to the old first */
369         mlast->m_next = first->tqe_m;
370         first->tqe_m = m;
371         first->tqe_len += tlen;
372         first->tqe_start = th->th_seq;
373         first->tqe_m->m_pkthdr.len = first->tqe_len;
374         first->tqe_mbuf_cnt += lenofoh;
375         tp->t_rcvoopack++;
376         TCPSTAT_INC(tcps_rcvoopack);
377         TCPSTAT_ADD(tcps_rcvoobyte, tlen);
378 #ifdef TCP_REASS_LOGGING
379         tcp_reass_log_new_in(tp, first->tqe_start, lenofoh, first->tqe_m,
380                              TCP_R_LOG_PREPEND,
381                              first);
382 #endif
383 }
384
385 static void
386 tcp_reass_replace(struct tcpcb *tp, struct tseg_qent *q, struct mbuf *m,
387     tcp_seq seq, int len, struct mbuf *mlast, int mbufoh, uint8_t flags)
388 {
389         /*
390          * Free the data in q, and replace
391          * it with the new segment.
392          */
393         int len_dif;
394
395 #ifdef TCP_REASS_LOGGING
396         tcp_log_reassm(tp, q, NULL, seq, len, TCP_R_LOG_REPLACE, 0);
397 #endif
398         m_freem(q->tqe_m);
399         KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
400                 ("Tp:%p seg queue goes negative", tp));
401         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
402         q->tqe_mbuf_cnt = mbufoh;
403         q->tqe_m = m;
404         q->tqe_last = mlast;
405         q->tqe_start = seq;
406         if (len > q->tqe_len)
407                 len_dif = len - q->tqe_len;
408         else
409                 len_dif = 0;
410         tp->t_rcvoopack++;
411         TCPSTAT_INC(tcps_rcvoopack);
412         TCPSTAT_ADD(tcps_rcvoobyte, len_dif);
413         q->tqe_len = len;
414         q->tqe_flags = (flags & TH_FIN);
415         q->tqe_m->m_pkthdr.len = q->tqe_len;
416         tp->t_segqmbuflen += mbufoh;
417
418 }
419
420 static void
421 tcp_reass_merge_into(struct tcpcb *tp, struct tseg_qent *ent,
422     struct tseg_qent *q)
423 {
424         /*
425          * Merge q into ent and free q from the list.
426          */
427 #ifdef TCP_REASS_LOGGING
428         tcp_log_reassm(tp, q, ent, 0, 0, TCP_R_LOG_MERGE_INTO, 0);
429 #endif
430 #ifdef TCP_REASS_COUNTERS
431         counter_u64_add(merge_into, 1);
432 #endif
433         ent->tqe_last->m_next = q->tqe_m;
434         ent->tqe_last = q->tqe_last;
435         ent->tqe_len += q->tqe_len;
436         ent->tqe_mbuf_cnt += q->tqe_mbuf_cnt;
437         ent->tqe_m->m_pkthdr.len += q->tqe_len;
438         ent->tqe_flags |= (q->tqe_flags & TH_FIN);
439         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
440         uma_zfree(tcp_reass_zone, q);
441         tp->t_segqlen--;
442
443 }
444
445 static void
446 tcp_reass_merge_forward(struct tcpcb *tp, struct tseg_qent *ent)
447 {
448         struct tseg_qent *q, *qtmp;
449         int i;
450         tcp_seq max;
451         /*
452          * Given an entry merge forward anyplace
453          * that ent overlaps forward.
454          */
455
456         max = ent->tqe_start + ent->tqe_len;
457         q = TAILQ_NEXT(ent, tqe_q);
458         if (q == NULL) {
459                 /* Nothing left */
460                 return;
461         }
462         TAILQ_FOREACH_FROM_SAFE(q, &tp->t_segq, tqe_q, qtmp) {
463                 if (SEQ_GT(q->tqe_start, max)) {
464                         /* Beyond q */
465                         break;
466                 }
467                 /* We have some or all that are overlapping */
468                 if (SEQ_GEQ(max, (q->tqe_start + q->tqe_len))) {
469                         /* It consumes it all */
470                         tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
471                         m_freem(q->tqe_m);
472                         TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
473                         uma_zfree(tcp_reass_zone, q);
474                         tp->t_segqlen--;
475                         continue;
476                 }
477                 /*
478                  * Trim the q entry to dovetail to this one
479                  * and then merge q into ent updating max
480                  * in the process.
481                  */
482                 i = max - q->tqe_start;
483 #ifdef TCP_REASS_LOGGING
484                 tcp_log_reassm(tp, q, NULL, 0, i, TCP_R_LOG_TRIM, 2);
485 #endif
486                 m_adj(q->tqe_m, i);
487                 q->tqe_len -= i;
488                 q->tqe_start += i;
489                 tcp_reass_merge_into(tp, ent, q);
490                 max = ent->tqe_start + ent->tqe_len;
491         }
492 #ifdef TCP_REASS_COUNTERS
493         counter_u64_add(merge_fwd, 1);
494 #endif
495 }
496
497 static int
498 tcp_reass_overhead_of_chain(struct mbuf *m, struct mbuf **mlast)
499 {
500         int len = MSIZE;
501
502         if (m->m_flags & M_EXT)
503                 len += m->m_ext.ext_size;
504         while (m->m_next != NULL) {
505                 m = m->m_next;
506                 len += MSIZE;
507                 if (m->m_flags & M_EXT)
508                         len += m->m_ext.ext_size;
509         }
510         *mlast = m;
511         return (len);
512 }
513
514 /*
515  * NOTE!!! the new tcp-reassembly code *must not* use
516  * m_adj() with a negative index. That alters the chain
517  * of mbufs (by possibly chopping trailing mbufs). At
518  * the front of tcp_reass we count the mbuf overhead
519  * and setup the tail pointer. If we use m_adj(m, -5)
520  * we could corrupt the tail pointer. Currently the
521  * code only uses m_adj(m, postive-num). If this
522  * changes appropriate changes to update mlast would
523  * be needed.
524  */
525 int
526 tcp_reass(struct tcpcb *tp, struct tcphdr *th, tcp_seq *seq_start,
527           int *tlenp, struct mbuf *m)
528 {
529         struct tseg_qent *q, *last, *first;
530         struct tseg_qent *p = NULL;
531         struct tseg_qent *nq = NULL;
532         struct tseg_qent *te = NULL;
533         struct mbuf *mlast = NULL;
534         struct sockbuf *sb;
535         struct socket *so = tp->t_inpcb->inp_socket;
536         char *s = NULL;
537         int flags, i, lenofoh;
538
539         INP_WLOCK_ASSERT(tp->t_inpcb);
540         /*
541          * XXX: tcp_reass() is rather inefficient with its data structures
542          * and should be rewritten (see NetBSD for optimizations).
543          */
544
545         KASSERT(th == NULL || (seq_start != NULL && tlenp != NULL),
546                 ("tcp_reass called with illegal parameter combination "
547                  "(tp=%p, th=%p, seq_start=%p, tlenp=%p, m=%p)",
548                  tp, th, seq_start, tlenp, m));
549         /*
550          * Call with th==NULL after become established to
551          * force pre-ESTABLISHED data up to user socket.
552          */
553         if (th == NULL)
554                 goto present;
555         KASSERT(SEQ_GEQ(th->th_seq, tp->rcv_nxt),
556                 ("Attempt to add old entry to reassembly queue (th=%p, tp=%p)",
557                  th, tp));
558 #ifdef TCP_REASS_LOGGING
559         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_ADD, NULL);
560 #endif
561 #ifdef TCP_REASS_COUNTERS
562         counter_u64_add(reass_entry, 1);
563 #endif
564         /*
565          * Check for zero length data.
566          */
567         if ((*tlenp == 0) && ((th->th_flags & TH_FIN) == 0)) {
568                 /*
569                  * A zero length segment does no
570                  * one any good. We could check
571                  * the rcv_nxt <-> rcv_wnd but thats
572                  * already done for us by the caller.
573                  */
574 #ifdef TCP_REASS_COUNTERS
575                 counter_u64_add(tcp_zero_input, 1);
576 #endif
577                 m_freem(m);
578 #ifdef TCP_REASS_LOGGING
579                 tcp_reass_log_dump(tp);
580 #endif
581                 return (0);
582         }
583         /*
584          * Will it fit?
585          */
586         lenofoh = tcp_reass_overhead_of_chain(m, &mlast);
587         sb = &tp->t_inpcb->inp_socket->so_rcv;
588         if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
589             (sb->sb_mbcnt + tp->t_segqmbuflen + lenofoh) > sb->sb_mbmax) {
590                 /* No room */
591                 TCPSTAT_INC(tcps_rcvreassfull);
592 #ifdef TCP_REASS_COUNTERS
593                 counter_u64_add(reass_nospace, 1);
594 #endif
595 #ifdef TCP_REASS_LOGGING
596                 tcp_log_reassm(tp, NULL, NULL, th->th_seq, lenofoh, TCP_R_LOG_LIMIT_REACHED, 0);
597 #endif
598                 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
599                         log(LOG_DEBUG, "%s; %s: mbuf count limit reached, "
600                             "segment dropped\n", s, __func__);
601                         free(s, M_TCPLOG);
602                 }
603                 m_freem(m);
604                 *tlenp = 0;
605 #ifdef TCP_REASS_LOGGING
606                 tcp_reass_log_dump(tp);
607 #endif
608                 return (0);
609         }
610         /*
611          * First lets deal with two common cases, the
612          * segment appends to the back of our collected
613          * segments. Or the segment is the next in line.
614          */
615         last = TAILQ_LAST_FAST(&tp->t_segq, tseg_qent, tqe_q);
616         if (last != NULL) {
617                 if ((th->th_flags & TH_FIN) &&
618                     SEQ_LT((th->th_seq + *tlenp), (last->tqe_start + last->tqe_len))) {
619                         /*
620                          * Someone is trying to game us, dump
621                          * the segment.
622                          */
623                         *tlenp = 0;
624                         m_freem(m);
625                         return (0);
626                 }
627                 if ((SEQ_GEQ(th->th_seq, last->tqe_start)) &&
628                     (SEQ_GEQ((last->tqe_start + last->tqe_len), th->th_seq))) {
629                         /* Common case, trailing segment is added */
630                         /**
631                          *                                 +--last
632                          *                                 v
633                          *  reassembly buffer |---|  |---| |---|
634                          *  new segment                       |---|
635                          */
636 #ifdef TCP_REASS_COUNTERS
637                         counter_u64_add(reass_path1, 1);
638 #endif
639                         if (SEQ_GT((last->tqe_start + last->tqe_len), th->th_seq)) {
640                                 i = (last->tqe_start + last->tqe_len) - th->th_seq;
641                                 if (i < *tlenp) {
642 #ifdef TCP_REASS_LOGGING
643                                         tcp_log_reassm(tp, last, NULL, 0, i, TCP_R_LOG_TRIM, 3);
644                                         th->th_seq += i;
645 #endif
646                                         m_adj(m, i);
647                                         *tlenp -= i;
648                                 } else {
649                                         /* Complete overlap */
650                                         TCPSTAT_INC(tcps_rcvduppack);
651                                         TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
652                                         m_freem(m);
653                                         *tlenp = last->tqe_len;
654                                         *seq_start = last->tqe_start;
655                                         return (0);
656                                 }
657                         }
658                         if (last->tqe_flags & TH_FIN) {
659                                 /*
660                                  * We have data after the FIN on the last?
661                                  */
662                                 *tlenp = 0;
663                                 m_freem(m);
664                                 return(0);
665                         }
666                         tcp_reass_append(tp, last, m, th, *tlenp, mlast, lenofoh);
667                         tp->t_segqmbuflen += lenofoh;
668                         *seq_start = last->tqe_start;
669                         *tlenp = last->tqe_len;
670                         return (0);
671                 } else if (SEQ_GT(th->th_seq, (last->tqe_start + last->tqe_len))) {
672                         /*
673                          * Second common case, we missed
674                          * another one and have something more
675                          * for the end.
676                          */
677                         /**
678                          *                                 +--last
679                          *                                 v
680                          *  reassembly buffer |---|  |---| |---|
681                          *  new segment                           |---|
682                          */
683                         if (last->tqe_flags & TH_FIN) {
684                                 /*
685                                  * We have data after the FIN on the last?
686                                  */
687                                 *tlenp = 0;
688                                 m_freem(m);
689                                 return(0);
690                         }
691 #ifdef TCP_REASS_COUNTERS
692                         counter_u64_add(reass_path2, 1);
693 #endif
694                         p = last;
695                         goto new_entry;
696                 }
697         } else {
698                 /* First segment (it's NULL). */
699                 goto new_entry;
700         }
701         first = TAILQ_FIRST(&tp->t_segq);
702         if (SEQ_LT(th->th_seq, first->tqe_start) &&
703             SEQ_GEQ((th->th_seq + *tlenp),first->tqe_start) &&
704             SEQ_LT((th->th_seq + *tlenp), (first->tqe_start + first->tqe_len))) {
705                 /*
706                  * The head of the queue is prepended by this and
707                  * it may be the one I want most.
708                  */
709                 /**
710                  *       first-------+
711                  *                   v
712                  *  rea:             |---|  |---| |---|
713                  *  new:         |---|
714                  * Note the case we do not deal with here is:
715                  *   rea=     |---|   |---|   |---|
716                  *   new=  |----|
717                  * Due to the fact that it could be
718                  *   new   |--------------------|
719                  * And we might need to merge forward.
720                  */
721 #ifdef INVARIANTS
722                 struct mbuf *firstmbuf;
723 #endif
724
725 #ifdef TCP_REASS_COUNTERS
726                 counter_u64_add(reass_path3, 1);
727 #endif
728                 if (SEQ_LT(th->th_seq, tp->rcv_nxt)) {
729                         /*
730                          * The resend was even before
731                          * what we have. We need to trim it.
732                          * Note TSNH (it should be trimmed
733                          * before the call to tcp_reass()).
734                          */
735 #ifdef INVARIANTS
736                         panic("th->th_seq:%u rcv_nxt:%u tp:%p not pre-trimmed",
737                               th->th_seq, tp->rcv_nxt, tp);
738 #else
739                         i = tp->rcv_nxt - th->th_seq;
740 #ifdef TCP_REASS_LOGGING
741                         tcp_log_reassm(tp, first, NULL, 0, i, TCP_R_LOG_TRIM, 4);
742 #endif
743                         m_adj(m, i);
744                         th->th_seq += i;
745                         *tlenp -= i;
746 #endif
747                 }
748 #ifdef INVARIANTS
749                 firstmbuf = first->tqe_m;
750 #endif
751                 tcp_reass_prepend(tp, first, m, th, *tlenp, mlast, lenofoh);
752 #ifdef INVARIANTS
753                 if (firstmbuf == first->tqe_m) {
754                         panic("First stayed same m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
755                               m, firstmbuf, first->tqe_m, tp, first);
756                 } else if (first->tqe_m != m) {
757                         panic("First did not change to m:%p foobar:%p first->tqe_m:%p tp:%p first:%p",
758                               m, firstmbuf, first->tqe_m, tp, first);
759                 }
760 #endif
761                 tp->t_segqmbuflen += lenofoh;
762                 *seq_start = first->tqe_start;
763                 *tlenp = first->tqe_len;
764                 goto present;
765         } else if (SEQ_LT((th->th_seq + *tlenp), first->tqe_start)) {
766                 /* New segment is before our earliest segment. */
767                 /**
768                  *           first---->+
769                  *                      v
770                  *  rea=                |---| ....
771                  *  new"         |---|
772                  *
773                  */
774                 goto new_entry;
775         }
776         /*
777          * Find a segment which begins after this one does.
778          */
779 #ifdef TCP_REASS_COUNTERS
780         counter_u64_add(reass_fullwalk, 1);
781 #endif
782         TAILQ_FOREACH(q, &tp->t_segq, tqe_q) {
783                 if (SEQ_GT(q->tqe_start, th->th_seq))
784                         break;
785         }
786         p = TAILQ_PREV(q, tsegqe_head, tqe_q);
787         /**
788          * Now is this fit just in-between only?
789          * i.e.:
790          *      p---+        +----q
791          *          v        v
792          *     res= |--|     |--|    |--|
793          *     nee       |-|
794          */
795         if (SEQ_LT((th->th_seq + *tlenp), q->tqe_start) &&
796             ((p == NULL) || (SEQ_GT(th->th_seq, (p->tqe_start + p->tqe_len))))) {
797                 /* Yep no overlap */
798                 goto new_entry;
799         }
800         /**
801          * If we reach here we have some (possibly all) overlap
802          * such as:
803          *     res=     |--|     |--|    |--|
804          *     new=  |----|
805          * or  new=  |-----------------|
806          * or  new=      |--------|
807          * or  new=            |---|
808          * or  new=            |-----------|
809          */
810         if ((p != NULL) &&
811             (SEQ_LEQ(th->th_seq, (p->tqe_start + p->tqe_len)))) {
812                 /* conversion to int (in i) handles seq wraparound */
813
814 #ifdef TCP_REASS_COUNTERS
815                 counter_u64_add(reass_path4, 1);
816 #endif
817                 i = p->tqe_start + p->tqe_len - th->th_seq;
818                 if (i >= 0) {
819                         if (i >= *tlenp) {
820                                 /**
821                                  *       prev seg---->+
822                                  *                    v
823                                  *  reassembly buffer |---|
824                                  *  new segment        |-|
825                                  */
826                                 TCPSTAT_INC(tcps_rcvduppack);
827                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
828                                 *tlenp = p->tqe_len;
829                                 *seq_start = p->tqe_start;
830                                 m_freem(m);
831                                 /*
832                                  * Try to present any queued data
833                                  * at the left window edge to the user.
834                                  * This is needed after the 3-WHS
835                                  * completes. Note this probably
836                                  * will not work and we will return.
837                                  */
838                                 return (0);
839                         }
840                         if (i > 0) {
841                                 /**
842                                  *       prev seg---->+
843                                  *                    v
844                                  *  reassembly buffer |---|
845                                  *  new segment         |-----|
846                                  */
847 #ifdef TCP_REASS_COUNTERS
848                                 counter_u64_add(reass_path5, 1);
849 #endif
850 #ifdef TCP_REASS_LOGGING
851                                 tcp_log_reassm(tp, p, NULL, 0, i, TCP_R_LOG_TRIM, 5);
852 #endif
853                                 m_adj(m, i);
854                                 *tlenp -= i;
855                                 th->th_seq += i;
856                         }
857                 }
858                 if (th->th_seq == (p->tqe_start + p->tqe_len)) {
859                         /*
860                          * If dovetails in with this one
861                          * append it.
862                          */
863                         /**
864                          *       prev seg---->+
865                          *                    v
866                          *  reassembly buffer |--|     |---|
867                          *  new segment          |--|
868                          * (note: it was trimmed above if it overlapped)
869                          */
870                         tcp_reass_append(tp, p, m, th, *tlenp, mlast, lenofoh);
871                         tp->t_segqmbuflen += lenofoh;
872                 } else {
873 #ifdef INVARIANTS
874                         panic("Impossible cut th_seq:%u p->seq:%u(%d) p:%p tp:%p",
875                               th->th_seq, p->tqe_start, p->tqe_len,
876                               p, tp);
877 #endif
878                         *tlenp = 0;
879                         m_freem(m);
880                         return (0);
881                 }
882                 q = p;
883         } else {
884                 /*
885                  * The new data runs over the
886                  * top of previously sack'd data (in q).
887                  * It may be partially overlapping, or
888                  * it may overlap the entire segment.
889                  */
890 #ifdef TCP_REASS_COUNTERS
891                 counter_u64_add(reass_path6, 1);
892 #endif
893                 if (SEQ_GEQ((th->th_seq + *tlenp), (q->tqe_start + q->tqe_len))) {
894                         /* It consumes it all */
895                         /**
896                          *             next seg---->+
897                          *                          v
898                          *  reassembly buffer |--|     |---|
899                          *  new segment              |----------|
900                          */
901 #ifdef TCP_REASS_COUNTERS
902                         counter_u64_add(reass_path7, 1);
903 #endif
904                         tcp_reass_replace(tp, q, m, th->th_seq, *tlenp, mlast, lenofoh, th->th_flags);
905                 } else {
906                         /*
907                          * We just need to prepend the data
908                          * to this. It does not overrun
909                          * the end.
910                          */
911                         /**
912                          *                next seg---->+
913                          *                             v
914                          *  reassembly buffer |--|     |---|
915                          *  new segment                   |----------|
916                          */
917                         tcp_reass_prepend(tp, q, m, th, *tlenp, mlast, lenofoh);
918                         tp->t_segqmbuflen += lenofoh;
919                 }
920         }
921         /* Now does it go further than that? */
922         tcp_reass_merge_forward(tp, q);
923         *seq_start = q->tqe_start;
924         *tlenp = q->tqe_len;
925         goto present;
926
927         /*
928          * When we reach here we can't combine it
929          * with any existing segment.
930          *
931          * Limit the number of segments that can be queued to reduce the
932          * potential for mbuf exhaustion. For best performance, we want to be
933          * able to queue a full window's worth of segments. The size of the
934          * socket receive buffer determines our advertised window and grows
935          * automatically when socket buffer autotuning is enabled. Use it as the
936          * basis for our queue limit.
937          *
938          * However, allow the user to specify a ceiling for the number of
939          * segments in each queue.
940          *
941          * Always let the missing segment through which caused this queue.
942          * NB: Access to the socket buffer is left intentionally unlocked as we
943          * can tolerate stale information here.
944          *
945          * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
946          * should work but causes packets to be dropped when they shouldn't.
947          * Investigate why and re-evaluate the below limit after the behaviour
948          * is understood.
949          */
950 new_entry:
951         if (th->th_seq == tp->rcv_nxt && TCPS_HAVEESTABLISHED(tp->t_state)) {
952                 tp->rcv_nxt += *tlenp;
953                 flags = th->th_flags & TH_FIN;
954                 TCPSTAT_INC(tcps_rcvoopack);
955                 TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
956                 SOCKBUF_LOCK(&so->so_rcv);
957                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
958                         m_freem(m);
959                 } else {
960                         sbappendstream_locked(&so->so_rcv, m, 0);
961                 }
962                 SOCKBUF_UNLOCK(&so->so_rcv);
963                 tp->t_flags |= TF_WAKESOR;
964                 return (flags);
965         }
966         if (tcp_new_limits) {
967                 if ((tp->t_segqlen > tcp_reass_queue_guard) &&
968                     (*tlenp < MSIZE)) {
969                         /*
970                          * This is really a lie, we are not full but
971                          * are getting a segment that is above
972                          * guard threshold. If it is and its below
973                          * a mbuf size (256) we drop it if it
974                          * can't fill in some place.
975                          */
976                         TCPSTAT_INC(tcps_rcvreassfull);
977                         *tlenp = 0;
978                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
979                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
980                                     "segment dropped\n", s, __func__);
981                                 free(s, M_TCPLOG);
982                         }
983                         m_freem(m);
984 #ifdef TCP_REASS_LOGGING
985                         tcp_reass_log_dump(tp);
986 #endif
987                         return (0);
988                 }
989         } else {
990                 if (tp->t_segqlen >= min((so->so_rcv.sb_hiwat / tp->t_maxseg) + 1,
991                                          tcp_reass_maxqueuelen)) {
992                         TCPSTAT_INC(tcps_rcvreassfull);
993                         *tlenp = 0;
994                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
995                                 log(LOG_DEBUG, "%s; %s: queue limit reached, "
996                                     "segment dropped\n", s, __func__);
997                                 free(s, M_TCPLOG);
998                         }
999                         m_freem(m);
1000 #ifdef TCP_REASS_LOGGING
1001                         tcp_reass_log_dump(tp);
1002 #endif
1003                         return (0);
1004                 }
1005         }
1006         /*
1007          * Allocate a new queue entry. If we can't, or hit the zone limit
1008          * just drop the pkt.
1009          */
1010         te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
1011         if (te == NULL) {
1012                 TCPSTAT_INC(tcps_rcvmemdrop);
1013                 m_freem(m);
1014                 *tlenp = 0;
1015                 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
1016                                        NULL))) {
1017                         log(LOG_DEBUG, "%s; %s: global zone limit "
1018                             "reached, segment dropped\n", s, __func__);
1019                         free(s, M_TCPLOG);
1020                 }
1021                 return (0);
1022         }
1023         tp->t_segqlen++;
1024         tp->t_rcvoopack++;
1025         TCPSTAT_INC(tcps_rcvoopack);
1026         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
1027         /* Insert the new segment queue entry into place. */
1028         te->tqe_m = m;
1029         te->tqe_flags = th->th_flags;
1030         te->tqe_len = *tlenp;
1031         te->tqe_start = th->th_seq;
1032         te->tqe_last = mlast;
1033         te->tqe_mbuf_cnt = lenofoh;
1034         tp->t_segqmbuflen += te->tqe_mbuf_cnt;
1035         if (p == NULL) {
1036                 TAILQ_INSERT_HEAD(&tp->t_segq, te, tqe_q);
1037         } else {
1038                 TAILQ_INSERT_AFTER(&tp->t_segq, p, te, tqe_q);
1039         }
1040 #ifdef TCP_REASS_LOGGING
1041         tcp_reass_log_new_in(tp, th->th_seq, *tlenp, m, TCP_R_LOG_NEW_ENTRY, te);
1042 #endif
1043 present:
1044         /*
1045          * Present data to user, advancing rcv_nxt through
1046          * completed sequence space.
1047          */
1048         if (!TCPS_HAVEESTABLISHED(tp->t_state))
1049                 return (0);
1050         q = TAILQ_FIRST(&tp->t_segq);
1051         KASSERT(q == NULL || SEQ_GEQ(q->tqe_start, tp->rcv_nxt),
1052                 ("Reassembly queue for %p has stale entry at head", tp));
1053         if (!q || q->tqe_start != tp->rcv_nxt) {
1054 #ifdef TCP_REASS_LOGGING
1055                 tcp_reass_log_dump(tp);
1056 #endif
1057                 return (0);
1058         }
1059         SOCKBUF_LOCK(&so->so_rcv);
1060         do {
1061                 tp->rcv_nxt += q->tqe_len;
1062                 flags = q->tqe_flags & TH_FIN;
1063                 nq = TAILQ_NEXT(q, tqe_q);
1064                 TAILQ_REMOVE(&tp->t_segq, q, tqe_q);
1065                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) {
1066                         m_freem(q->tqe_m);
1067                 } else {
1068 #ifdef TCP_REASS_LOGGING
1069                         tcp_reass_log_new_in(tp, q->tqe_start, q->tqe_len, q->tqe_m, TCP_R_LOG_READ, q);
1070                         if (th != NULL) {
1071                                 tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 1);
1072                         } else {
1073                                 tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 1);
1074                         }
1075 #endif
1076                         sbappendstream_locked(&so->so_rcv, q->tqe_m, 0);
1077                 }
1078 #ifdef TCP_REASS_LOGGING
1079                 if (th != NULL) {
1080                         tcp_log_reassm(tp, q, NULL, th->th_seq, *tlenp, TCP_R_LOG_READ, 2);
1081                 } else {
1082                         tcp_log_reassm(tp, q, NULL, 0, 0, TCP_R_LOG_READ, 2);
1083                 }
1084 #endif
1085                 KASSERT(tp->t_segqmbuflen >= q->tqe_mbuf_cnt,
1086                         ("tp:%p seg queue goes negative", tp));
1087                 tp->t_segqmbuflen -= q->tqe_mbuf_cnt;
1088                 uma_zfree(tcp_reass_zone, q);
1089                 tp->t_segqlen--;
1090                 q = nq;
1091         } while (q && q->tqe_start == tp->rcv_nxt);
1092         if (TAILQ_EMPTY(&tp->t_segq) &&
1093             (tp->t_segqmbuflen != 0)) {
1094 #ifdef INVARIANTS
1095                 panic("tp:%p segq:%p len:%d queue empty",
1096                       tp, &tp->t_segq, tp->t_segqmbuflen);
1097 #else
1098 #ifdef TCP_REASS_LOGGING
1099                 if (th != NULL) {
1100                         tcp_log_reassm(tp, NULL, NULL, th->th_seq, *tlenp, TCP_R_LOG_ZERO, 0);
1101                 } else {
1102                         tcp_log_reassm(tp, NULL, NULL, 0, 0, TCP_R_LOG_ZERO, 0);
1103                 }
1104 #endif
1105                 tp->t_segqmbuflen = 0;
1106 #endif
1107         }
1108 #ifdef TCP_REASS_LOGGING
1109         tcp_reass_log_dump(tp);
1110 #endif
1111         SOCKBUF_UNLOCK(&so->so_rcv);
1112         tp->t_flags |= TF_WAKESOR;
1113         return (flags);
1114 }