sys/netinet/tcp_reass.c

   1 /*-
   2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
   3  *      The Regents of the University of California.  All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 4. Neither the name of the University nor the names of its contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  *
  29  *      @(#)tcp_input.c 8.12 (Berkeley) 5/24/95
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include "opt_inet.h"
  36 #include "opt_inet6.h"
  37 #include "opt_tcpdebug.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/kernel.h>
  41 #include <sys/malloc.h>
  42 #include <sys/mbuf.h>
  43 #include <sys/socket.h>
  44 #include <sys/socketvar.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/syslog.h>
  47 #include <sys/systm.h>
  48
  49 #include <vm/uma.h>
  50
  51 #include <net/if.h>
  52 #include <net/route.h>
  53 #include <net/vnet.h>
  54
  55 #include <netinet/in.h>
  56 #include <netinet/in_pcb.h>
  57 #include <netinet/in_systm.h>
  58 #include <netinet/in_var.h>
  59 #include <netinet/ip.h>
  60 #include <netinet/ip_var.h>
  61 #include <netinet/ip_options.h>
  62 #include <netinet/ip6.h>
  63 #include <netinet6/in6_pcb.h>
  64 #include <netinet6/ip6_var.h>
  65 #include <netinet6/nd6.h>
  66 #include <netinet/tcp.h>
  67 #include <netinet/tcp_fsm.h>
  68 #include <netinet/tcp_seq.h>
  69 #include <netinet/tcp_timer.h>
  70 #include <netinet/tcp_var.h>
  71 #include <netinet6/tcp6_var.h>
  72 #include <netinet/tcpip.h>
  73 #ifdef TCPDEBUG
  74 #include <netinet/tcp_debug.h>
  75 #endif /* TCPDEBUG */
  76
  77 static int tcp_reass_sysctl_maxseg(SYSCTL_HANDLER_ARGS);
  78 static int tcp_reass_sysctl_qsize(SYSCTL_HANDLER_ARGS);
  79
  80 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
  81     "TCP Segment Reassembly Queue");
  82
  83 static int tcp_reass_maxseg = 0;
  84 SYSCTL_VNET_PROC(_net_inet_tcp_reass, OID_AUTO, maxsegments,
  85     CTLTYPE_INT | CTLFLAG_RDTUN,
  86     &tcp_reass_maxseg, 0, &tcp_reass_sysctl_maxseg, "I",
  87     "Global maximum number of TCP Segments in Reassembly Queue");
  88
  89 static int tcp_reass_qsize = 0;
  90 SYSCTL_PROC(_net_inet_tcp_reass, OID_AUTO, cursegments,
  91     CTLTYPE_INT | CTLFLAG_RD,
  92     &tcp_reass_qsize, 0, &tcp_reass_sysctl_qsize, "I",
  93     "Global number of TCP Segments currently in Reassembly Queue");
  94
  95 static int tcp_reass_overflows = 0;
  96 SYSCTL_INT(_net_inet_tcp_reass, OID_AUTO, overflows,
  97     CTLTYPE_INT | CTLFLAG_RD,
  98     &tcp_reass_overflows, 0,
  99     "Global number of TCP Segment Reassembly Queue Overflows");
 100
 101 static uma_zone_t tcp_reass_zone;
 102
 103 /* Initialize TCP reassembly queue */
 104 static void
 105 tcp_reass_zone_change(void *tag)
 106 {
 107
 108         tcp_reass_maxseg = nmbclusters / 16;
 109         uma_zone_set_max(tcp_reass_zone, tcp_reass_maxseg);
 110 }
 111
 112 void
 113 tcp_reass_global_init(void)
 114 {
 115
 116         tcp_reass_maxseg = nmbclusters / 16;
 117         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
 118             &tcp_reass_maxseg);
 119         tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
 120             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 121         uma_zone_set_max(tcp_reass_zone, tcp_reass_maxseg);
 122         EVENTHANDLER_REGISTER(nmbclusters_change,
 123             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
 124 }
 125
 126 void
 127 tcp_reass_flush(struct tcpcb *tp)
 128 {
 129         struct tseg_qent *qe;
 130
 131         INP_WLOCK_ASSERT(tp->t_inpcb);
 132
 133         while ((qe = LIST_FIRST(&tp->t_segq)) != NULL) {
 134                 LIST_REMOVE(qe, tqe_q);
 135                 m_freem(qe->tqe_m);
 136                 uma_zfree(tcp_reass_zone, qe);
 137                 tp->t_segqlen--;
 138         }
 139
 140         KASSERT((tp->t_segqlen == 0),
 141             ("TCP reass queue %p segment count is %d instead of 0 after flush.",
 142             tp, tp->t_segqlen));
 143 }
 144
 145 static int
 146 tcp_reass_sysctl_maxseg(SYSCTL_HANDLER_ARGS)
 147 {
 148         tcp_reass_maxseg = uma_zone_get_max(tcp_reass_zone);
 149         return (sysctl_handle_int(oidp, arg1, arg2, req));
 150 }
 151
 152 static int
 153 tcp_reass_sysctl_qsize(SYSCTL_HANDLER_ARGS)
 154 {
 155         tcp_reass_qsize = uma_zone_get_cur(tcp_reass_zone);
 156         return (sysctl_handle_int(oidp, arg1, arg2, req));
 157 }
 158
 159 int
 160 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
 161 {
 162         struct tseg_qent *q;
 163         struct tseg_qent *p = NULL;
 164         struct tseg_qent *nq;
 165         struct tseg_qent *te = NULL;
 166         struct socket *so = tp->t_inpcb->inp_socket;
 167         char *s = NULL;
 168         int flags;
 169         struct tseg_qent tqs;
 170
 171         INP_WLOCK_ASSERT(tp->t_inpcb);
 172
 173         /*
 174          * XXX: tcp_reass() is rather inefficient with its data structures
 175          * and should be rewritten (see NetBSD for optimizations).
 176          */
 177
 178         /*
 179          * Call with th==NULL after become established to
 180          * force pre-ESTABLISHED data up to user socket.
 181          */
 182         if (th == NULL)
 183                 goto present;
 184
 185         /*
 186          * Limit the number of segments that can be queued to reduce the
 187          * potential for mbuf exhaustion. For best performance, we want to be
 188          * able to queue a full window's worth of segments. The size of the
 189          * socket receive buffer determines our advertised window and grows
 190          * automatically when socket buffer autotuning is enabled. Use it as the
 191          * basis for our queue limit.
 192          * Always let the missing segment through which caused this queue.
 193          * NB: Access to the socket buffer is left intentionally unlocked as we
 194          * can tolerate stale information here.
 195          *
 196          * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
 197          * should work but causes packets to be dropped when they shouldn't.
 198          * Investigate why and re-evaluate the below limit after the behaviour
 199          * is understood.
 200          */
 201         if ((th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) &&
 202             tp->t_segqlen >= (so->so_rcv.sb_hiwat / tp->t_maxseg) + 1) {
 203                 tcp_reass_overflows++;
 204                 TCPSTAT_INC(tcps_rcvmemdrop);
 205                 m_freem(m);
 206                 *tlenp = 0;
 207                 if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL, NULL))) {
 208                         log(LOG_DEBUG, "%s; %s: queue limit reached, "
 209                             "segment dropped\n", s, __func__);
 210                         free(s, M_TCPLOG);
 211                 }
 212                 return (0);
 213         }
 214
 215         /*
 216          * Allocate a new queue entry. If we can't, or hit the zone limit
 217          * just drop the pkt.
 218          *
 219          * Use a temporary structure on the stack for the missing segment
 220          * when the zone is exhausted. Otherwise we may get stuck.
 221          */
 222         te = uma_zalloc(tcp_reass_zone, M_NOWAIT);
 223         if (te == NULL) {
 224                 if (th->th_seq != tp->rcv_nxt || !TCPS_HAVEESTABLISHED(tp->t_state)) {
 225                         TCPSTAT_INC(tcps_rcvmemdrop);
 226                         m_freem(m);
 227                         *tlenp = 0;
 228                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
 229                             NULL))) {
 230                                 log(LOG_DEBUG, "%s; %s: global zone limit "
 231                                     "reached, segment dropped\n", s, __func__);
 232                                 free(s, M_TCPLOG);
 233                         }
 234                         return (0);
 235                 } else {
 236                         bzero(&tqs, sizeof(struct tseg_qent));
 237                         te = &tqs;
 238                         if ((s = tcp_log_addrs(&tp->t_inpcb->inp_inc, th, NULL,
 239                             NULL))) {
 240                                 log(LOG_DEBUG,
 241                                     "%s; %s: global zone limit reached, using "
 242                                     "stack for missing segment\n", s, __func__);
 243                                 free(s, M_TCPLOG);
 244                         }
 245                 }
 246         }
 247         tp->t_segqlen++;
 248
 249         /*
 250          * Find a segment which begins after this one does.
 251          */
 252         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
 253                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
 254                         break;
 255                 p = q;
 256         }
 257
 258         /*
 259          * If there is a preceding segment, it may provide some of
 260          * our data already.  If so, drop the data from the incoming
 261          * segment.  If it provides all of our data, drop us.
 262          */
 263         if (p != NULL) {
 264                 int i;
 265                 /* conversion to int (in i) handles seq wraparound */
 266                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
 267                 if (i > 0) {
 268                         if (i >= *tlenp) {
 269                                 TCPSTAT_INC(tcps_rcvduppack);
 270                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
 271                                 m_freem(m);
 272                                 if (te != &tqs)
 273                                         uma_zfree(tcp_reass_zone, te);
 274                                 tp->t_segqlen--;
 275                                 /*
 276                                  * Try to present any queued data
 277                                  * at the left window edge to the user.
 278                                  * This is needed after the 3-WHS
 279                                  * completes.
 280                                  */
 281                                 goto present;   /* ??? */
 282                         }
 283                         m_adj(m, i);
 284                         *tlenp -= i;
 285                         th->th_seq += i;
 286                 }
 287         }
 288         tp->t_rcvoopack++;
 289         TCPSTAT_INC(tcps_rcvoopack);
 290         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
 291
 292         /*
 293          * While we overlap succeeding segments trim them or,
 294          * if they are completely covered, dequeue them.
 295          */
 296         while (q) {
 297                 int i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
 298                 if (i <= 0)
 299                         break;
 300                 if (i < q->tqe_len) {
 301                         q->tqe_th->th_seq += i;
 302                         q->tqe_len -= i;
 303                         m_adj(q->tqe_m, i);
 304                         break;
 305                 }
 306
 307                 nq = LIST_NEXT(q, tqe_q);
 308                 LIST_REMOVE(q, tqe_q);
 309                 m_freem(q->tqe_m);
 310                 uma_zfree(tcp_reass_zone, q);
 311                 tp->t_segqlen--;
 312                 q = nq;
 313         }
 314
 315         /* Insert the new segment queue entry into place. */
 316         te->tqe_m = m;
 317         te->tqe_th = th;
 318         te->tqe_len = *tlenp;
 319
 320         if (p == NULL) {
 321                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
 322         } else {
 323                 KASSERT(te != &tqs, ("%s: temporary stack based entry not "
 324                     "first element in queue", __func__));
 325                 LIST_INSERT_AFTER(p, te, tqe_q);
 326         }
 327
 328 present:
 329         /*
 330          * Present data to user, advancing rcv_nxt through
 331          * completed sequence space.
 332          */
 333         if (!TCPS_HAVEESTABLISHED(tp->t_state))
 334                 return (0);
 335         q = LIST_FIRST(&tp->t_segq);
 336         if (!q || q->tqe_th->th_seq != tp->rcv_nxt)
 337                 return (0);
 338         SOCKBUF_LOCK(&so->so_rcv);
 339         do {
 340                 tp->rcv_nxt += q->tqe_len;
 341                 flags = q->tqe_th->th_flags & TH_FIN;
 342                 nq = LIST_NEXT(q, tqe_q);
 343                 LIST_REMOVE(q, tqe_q);
 344                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
 345                         m_freem(q->tqe_m);
 346                 else
 347                         sbappendstream_locked(&so->so_rcv, q->tqe_m);
 348                 if (q != &tqs)
 349                         uma_zfree(tcp_reass_zone, q);
 350                 tp->t_segqlen--;
 351                 q = nq;
 352         } while (q && q->tqe_th->th_seq == tp->rcv_nxt);
 353         ND6_HINT(tp);
 354         sorwakeup_locked(so);
 355         return (flags);
 356 }