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