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 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, reass, CTLFLAG_RW, 0,
  78     "TCP Segment Reassembly Queue");
  79
  80 static VNET_DEFINE(int, tcp_reass_maxseg) = 0;
  81 #define V_tcp_reass_maxseg              VNET(tcp_reass_maxseg)
  82 SYSCTL_VNET_INT(_net_inet_tcp_reass, OID_AUTO, maxsegments, CTLFLAG_RDTUN,
  83     &VNET_NAME(tcp_reass_maxseg), 0,
  84     "Global maximum number of TCP Segments in Reassembly Queue");
  85
  86 VNET_DEFINE(int, tcp_reass_qsize) = 0;
  87 SYSCTL_VNET_INT(_net_inet_tcp_reass, OID_AUTO, cursegments, CTLFLAG_RD,
  88     &VNET_NAME(tcp_reass_qsize), 0,
  89     "Global number of TCP Segments currently in Reassembly Queue");
  90
  91 static VNET_DEFINE(int, tcp_reass_maxqlen) = 48;
  92 #define V_tcp_reass_maxqlen             VNET(tcp_reass_maxqlen)
  93 SYSCTL_VNET_INT(_net_inet_tcp_reass, OID_AUTO, maxqlen, CTLFLAG_RW,
  94     &VNET_NAME(tcp_reass_maxqlen), 0,
  95     "Maximum number of TCP Segments per individual Reassembly Queue");
  96
  97 static VNET_DEFINE(int, tcp_reass_overflows) = 0;
  98 #define V_tcp_reass_overflows           VNET(tcp_reass_overflows)
  99 SYSCTL_VNET_INT(_net_inet_tcp_reass, OID_AUTO, overflows, CTLFLAG_RD,
 100     &VNET_NAME(tcp_reass_overflows), 0,
 101     "Global number of TCP Segment Reassembly Queue Overflows");
 102
 103 /* Initialize TCP reassembly queue */
 104 static void
 105 tcp_reass_zone_change(void *tag)
 106 {
 107
 108         V_tcp_reass_maxseg = nmbclusters / 16;
 109         uma_zone_set_max(V_tcp_reass_zone, V_tcp_reass_maxseg);
 110 }
 111
 112 VNET_DEFINE(uma_zone_t, tcp_reass_zone);
 113
 114 void
 115 tcp_reass_init(void)
 116 {
 117
 118         V_tcp_reass_maxseg = nmbclusters / 16;
 119         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
 120             &V_tcp_reass_maxseg);
 121         V_tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
 122             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 123         uma_zone_set_max(V_tcp_reass_zone, V_tcp_reass_maxseg);
 124         EVENTHANDLER_REGISTER(nmbclusters_change,
 125             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
 126 }
 127
 128 #ifdef VIMAGE
 129 void
 130 tcp_reass_destroy(void)
 131 {
 132
 133         uma_zdestroy(V_tcp_reass_zone);
 134 }
 135 #endif
 136
 137 int
 138 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
 139 {
 140         struct tseg_qent *q;
 141         struct tseg_qent *p = NULL;
 142         struct tseg_qent *nq;
 143         struct tseg_qent *te = NULL;
 144         struct socket *so = tp->t_inpcb->inp_socket;
 145         int flags;
 146
 147         INP_WLOCK_ASSERT(tp->t_inpcb);
 148
 149         /*
 150          * XXX: tcp_reass() is rather inefficient with its data structures
 151          * and should be rewritten (see NetBSD for optimizations).
 152          */
 153
 154         /*
 155          * Call with th==NULL after become established to
 156          * force pre-ESTABLISHED data up to user socket.
 157          */
 158         if (th == NULL)
 159                 goto present;
 160
 161         /*
 162          * Limit the number of segments in the reassembly queue to prevent
 163          * holding on to too many segments (and thus running out of mbufs).
 164          * Make sure to let the missing segment through which caused this
 165          * queue.  Always keep one global queue entry spare to be able to
 166          * process the missing segment.
 167          */
 168         if (th->th_seq != tp->rcv_nxt &&
 169             (V_tcp_reass_qsize + 1 >= V_tcp_reass_maxseg ||
 170              tp->t_segqlen >= V_tcp_reass_maxqlen)) {
 171                 V_tcp_reass_overflows++;
 172                 TCPSTAT_INC(tcps_rcvmemdrop);
 173                 m_freem(m);
 174                 *tlenp = 0;
 175                 return (0);
 176         }
 177
 178         /*
 179          * Allocate a new queue entry. If we can't, or hit the zone limit
 180          * just drop the pkt.
 181          */
 182         te = uma_zalloc(V_tcp_reass_zone, M_NOWAIT);
 183         if (te == NULL) {
 184                 TCPSTAT_INC(tcps_rcvmemdrop);
 185                 m_freem(m);
 186                 *tlenp = 0;
 187                 return (0);
 188         }
 189         tp->t_segqlen++;
 190         V_tcp_reass_qsize++;
 191
 192         /*
 193          * Find a segment which begins after this one does.
 194          */
 195         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
 196                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
 197                         break;
 198                 p = q;
 199         }
 200
 201         /*
 202          * If there is a preceding segment, it may provide some of
 203          * our data already.  If so, drop the data from the incoming
 204          * segment.  If it provides all of our data, drop us.
 205          */
 206         if (p != NULL) {
 207                 int i;
 208                 /* conversion to int (in i) handles seq wraparound */
 209                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
 210                 if (i > 0) {
 211                         if (i >= *tlenp) {
 212                                 TCPSTAT_INC(tcps_rcvduppack);
 213                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
 214                                 m_freem(m);
 215                                 uma_zfree(V_tcp_reass_zone, te);
 216                                 tp->t_segqlen--;
 217                                 V_tcp_reass_qsize--;
 218                                 /*
 219                                  * Try to present any queued data
 220                                  * at the left window edge to the user.
 221                                  * This is needed after the 3-WHS
 222                                  * completes.
 223                                  */
 224                                 goto present;   /* ??? */
 225                         }
 226                         m_adj(m, i);
 227                         *tlenp -= i;
 228                         th->th_seq += i;
 229                 }
 230         }
 231         TCPSTAT_INC(tcps_rcvoopack);
 232         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
 233
 234         /*
 235          * While we overlap succeeding segments trim them or,
 236          * if they are completely covered, dequeue them.
 237          */
 238         while (q) {
 239                 int i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
 240                 if (i <= 0)
 241                         break;
 242                 if (i < q->tqe_len) {
 243                         q->tqe_th->th_seq += i;
 244                         q->tqe_len -= i;
 245                         m_adj(q->tqe_m, i);
 246                         break;
 247                 }
 248
 249                 nq = LIST_NEXT(q, tqe_q);
 250                 LIST_REMOVE(q, tqe_q);
 251                 m_freem(q->tqe_m);
 252                 uma_zfree(V_tcp_reass_zone, q);
 253                 tp->t_segqlen--;
 254                 V_tcp_reass_qsize--;
 255                 q = nq;
 256         }
 257
 258         /* Insert the new segment queue entry into place. */
 259         te->tqe_m = m;
 260         te->tqe_th = th;
 261         te->tqe_len = *tlenp;
 262
 263         if (p == NULL) {
 264                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
 265         } else {
 266                 LIST_INSERT_AFTER(p, te, tqe_q);
 267         }
 268
 269 present:
 270         /*
 271          * Present data to user, advancing rcv_nxt through
 272          * completed sequence space.
 273          */
 274         if (!TCPS_HAVEESTABLISHED(tp->t_state))
 275                 return (0);
 276         q = LIST_FIRST(&tp->t_segq);
 277         if (!q || q->tqe_th->th_seq != tp->rcv_nxt)
 278                 return (0);
 279         SOCKBUF_LOCK(&so->so_rcv);
 280         do {
 281                 tp->rcv_nxt += q->tqe_len;
 282                 flags = q->tqe_th->th_flags & TH_FIN;
 283                 nq = LIST_NEXT(q, tqe_q);
 284                 LIST_REMOVE(q, tqe_q);
 285                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
 286                         m_freem(q->tqe_m);
 287                 else
 288                         sbappendstream_locked(&so->so_rcv, q->tqe_m);
 289                 uma_zfree(V_tcp_reass_zone, q);
 290                 tp->t_segqlen--;
 291                 V_tcp_reass_qsize--;
 292                 q = nq;
 293         } while (q && q->tqe_th->th_seq == tp->rcv_nxt);
 294         ND6_HINT(tp);
 295         sorwakeup_locked(so);
 296         return (flags);
 297 }