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 VNET_DEFINE(int, tcp_reass_maxseg) = 0;
  84 #define V_tcp_reass_maxseg              VNET(tcp_reass_maxseg)
  85 SYSCTL_VNET_PROC(_net_inet_tcp_reass, OID_AUTO, maxsegments,
  86     CTLTYPE_INT | CTLFLAG_RDTUN,
  87     &VNET_NAME(tcp_reass_maxseg), 0, &tcp_reass_sysctl_maxseg, "I",
  88     "Global maximum number of TCP Segments in Reassembly Queue");
  89
  90 static VNET_DEFINE(int, tcp_reass_qsize) = 0;
  91 #define V_tcp_reass_qsize               VNET(tcp_reass_qsize)
  92 SYSCTL_VNET_PROC(_net_inet_tcp_reass, OID_AUTO, cursegments,
  93     CTLTYPE_INT | CTLFLAG_RD,
  94     &VNET_NAME(tcp_reass_qsize), 0, &tcp_reass_sysctl_qsize, "I",
  95     "Global number of TCP Segments currently in 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,
 100     CTLTYPE_INT | CTLFLAG_RD,
 101     &VNET_NAME(tcp_reass_overflows), 0,
 102     "Global number of TCP Segment Reassembly Queue Overflows");
 103
 104 static VNET_DEFINE(uma_zone_t, tcp_reass_zone);
 105 #define V_tcp_reass_zone                VNET(tcp_reass_zone)
 106
 107 /* Initialize TCP reassembly queue */
 108 static void
 109 tcp_reass_zone_change(void *tag)
 110 {
 111
 112         V_tcp_reass_maxseg = nmbclusters / 16;
 113         uma_zone_set_max(V_tcp_reass_zone, V_tcp_reass_maxseg);
 114 }
 115
 116 void
 117 tcp_reass_init(void)
 118 {
 119
 120         V_tcp_reass_maxseg = nmbclusters / 16;
 121         TUNABLE_INT_FETCH("net.inet.tcp.reass.maxsegments",
 122             &V_tcp_reass_maxseg);
 123         V_tcp_reass_zone = uma_zcreate("tcpreass", sizeof (struct tseg_qent),
 124             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
 125         uma_zone_set_max(V_tcp_reass_zone, V_tcp_reass_maxseg);
 126         EVENTHANDLER_REGISTER(nmbclusters_change,
 127             tcp_reass_zone_change, NULL, EVENTHANDLER_PRI_ANY);
 128 }
 129
 130 #ifdef VIMAGE
 131 void
 132 tcp_reass_destroy(void)
 133 {
 134
 135         uma_zdestroy(V_tcp_reass_zone);
 136 }
 137 #endif
 138
 139 void
 140 tcp_reass_flush(struct tcpcb *tp)
 141 {
 142         struct tseg_qent *qe;
 143
 144         INP_WLOCK_ASSERT(tp->t_inpcb);
 145
 146         while ((qe = LIST_FIRST(&tp->t_segq)) != NULL) {
 147                 LIST_REMOVE(qe, tqe_q);
 148                 m_freem(qe->tqe_m);
 149                 uma_zfree(V_tcp_reass_zone, qe);
 150                 tp->t_segqlen--;
 151         }
 152
 153         KASSERT((tp->t_segqlen == 0),
 154             ("TCP reass queue %p segment count is %d instead of 0 after flush.",
 155             tp, tp->t_segqlen));
 156 }
 157
 158 static int
 159 tcp_reass_sysctl_maxseg(SYSCTL_HANDLER_ARGS)
 160 {
 161         V_tcp_reass_maxseg = uma_zone_get_max(V_tcp_reass_zone);
 162         return (sysctl_handle_int(oidp, arg1, arg2, req));
 163 }
 164
 165 static int
 166 tcp_reass_sysctl_qsize(SYSCTL_HANDLER_ARGS)
 167 {
 168         V_tcp_reass_qsize = uma_zone_get_cur(V_tcp_reass_zone);
 169         return (sysctl_handle_int(oidp, arg1, arg2, req));
 170 }
 171
 172 int
 173 tcp_reass(struct tcpcb *tp, struct tcphdr *th, int *tlenp, struct mbuf *m)
 174 {
 175         struct tseg_qent *q;
 176         struct tseg_qent *p = NULL;
 177         struct tseg_qent *nq;
 178         struct tseg_qent *te = NULL;
 179         struct socket *so = tp->t_inpcb->inp_socket;
 180         int flags;
 181
 182         INP_WLOCK_ASSERT(tp->t_inpcb);
 183
 184         /*
 185          * XXX: tcp_reass() is rather inefficient with its data structures
 186          * and should be rewritten (see NetBSD for optimizations).
 187          */
 188
 189         /*
 190          * Call with th==NULL after become established to
 191          * force pre-ESTABLISHED data up to user socket.
 192          */
 193         if (th == NULL)
 194                 goto present;
 195
 196         /*
 197          * Limit the number of segments that can be queued to reduce the
 198          * potential for mbuf exhaustion. For best performance, we want to be
 199          * able to queue a full window's worth of segments. The size of the
 200          * socket receive buffer determines our advertised window and grows
 201          * automatically when socket buffer autotuning is enabled. Use it as the
 202          * basis for our queue limit.
 203          * Always let the missing segment through which caused this queue.
 204          * NB: Access to the socket buffer is left intentionally unlocked as we
 205          * can tolerate stale information here.
 206          *
 207          * XXXLAS: Using sbspace(so->so_rcv) instead of so->so_rcv.sb_hiwat
 208          * should work but causes packets to be dropped when they shouldn't.
 209          * Investigate why and re-evaluate the below limit after the behaviour
 210          * is understood.
 211          */
 212         if (th->th_seq != tp->rcv_nxt &&
 213             tp->t_segqlen >= (so->so_rcv.sb_hiwat / tp->t_maxseg) + 1) {
 214                 V_tcp_reass_overflows++;
 215                 TCPSTAT_INC(tcps_rcvmemdrop);
 216                 m_freem(m);
 217                 *tlenp = 0;
 218                 return (0);
 219         }
 220
 221         /*
 222          * Allocate a new queue entry. If we can't, or hit the zone limit
 223          * just drop the pkt.
 224          */
 225         te = uma_zalloc(V_tcp_reass_zone, M_NOWAIT);
 226         if (te == NULL) {
 227                 TCPSTAT_INC(tcps_rcvmemdrop);
 228                 m_freem(m);
 229                 *tlenp = 0;
 230                 return (0);
 231         }
 232         tp->t_segqlen++;
 233
 234         /*
 235          * Find a segment which begins after this one does.
 236          */
 237         LIST_FOREACH(q, &tp->t_segq, tqe_q) {
 238                 if (SEQ_GT(q->tqe_th->th_seq, th->th_seq))
 239                         break;
 240                 p = q;
 241         }
 242
 243         /*
 244          * If there is a preceding segment, it may provide some of
 245          * our data already.  If so, drop the data from the incoming
 246          * segment.  If it provides all of our data, drop us.
 247          */
 248         if (p != NULL) {
 249                 int i;
 250                 /* conversion to int (in i) handles seq wraparound */
 251                 i = p->tqe_th->th_seq + p->tqe_len - th->th_seq;
 252                 if (i > 0) {
 253                         if (i >= *tlenp) {
 254                                 TCPSTAT_INC(tcps_rcvduppack);
 255                                 TCPSTAT_ADD(tcps_rcvdupbyte, *tlenp);
 256                                 m_freem(m);
 257                                 uma_zfree(V_tcp_reass_zone, te);
 258                                 tp->t_segqlen--;
 259                                 /*
 260                                  * Try to present any queued data
 261                                  * at the left window edge to the user.
 262                                  * This is needed after the 3-WHS
 263                                  * completes.
 264                                  */
 265                                 goto present;   /* ??? */
 266                         }
 267                         m_adj(m, i);
 268                         *tlenp -= i;
 269                         th->th_seq += i;
 270                 }
 271         }
 272         tp->t_rcvoopack++;
 273         TCPSTAT_INC(tcps_rcvoopack);
 274         TCPSTAT_ADD(tcps_rcvoobyte, *tlenp);
 275
 276         /*
 277          * While we overlap succeeding segments trim them or,
 278          * if they are completely covered, dequeue them.
 279          */
 280         while (q) {
 281                 int i = (th->th_seq + *tlenp) - q->tqe_th->th_seq;
 282                 if (i <= 0)
 283                         break;
 284                 if (i < q->tqe_len) {
 285                         q->tqe_th->th_seq += i;
 286                         q->tqe_len -= i;
 287                         m_adj(q->tqe_m, i);
 288                         break;
 289                 }
 290
 291                 nq = LIST_NEXT(q, tqe_q);
 292                 LIST_REMOVE(q, tqe_q);
 293                 m_freem(q->tqe_m);
 294                 uma_zfree(V_tcp_reass_zone, q);
 295                 tp->t_segqlen--;
 296                 q = nq;
 297         }
 298
 299         /* Insert the new segment queue entry into place. */
 300         te->tqe_m = m;
 301         te->tqe_th = th;
 302         te->tqe_len = *tlenp;
 303
 304         if (p == NULL) {
 305                 LIST_INSERT_HEAD(&tp->t_segq, te, tqe_q);
 306         } else {
 307                 LIST_INSERT_AFTER(p, te, tqe_q);
 308         }
 309
 310 present:
 311         /*
 312          * Present data to user, advancing rcv_nxt through
 313          * completed sequence space.
 314          */
 315         if (!TCPS_HAVEESTABLISHED(tp->t_state))
 316                 return (0);
 317         q = LIST_FIRST(&tp->t_segq);
 318         if (!q || q->tqe_th->th_seq != tp->rcv_nxt)
 319                 return (0);
 320         SOCKBUF_LOCK(&so->so_rcv);
 321         do {
 322                 tp->rcv_nxt += q->tqe_len;
 323                 flags = q->tqe_th->th_flags & TH_FIN;
 324                 nq = LIST_NEXT(q, tqe_q);
 325                 LIST_REMOVE(q, tqe_q);
 326                 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
 327                         m_freem(q->tqe_m);
 328                 else
 329                         sbappendstream_locked(&so->so_rcv, q->tqe_m);
 330                 uma_zfree(V_tcp_reass_zone, q);
 331                 tp->t_segqlen--;
 332                 q = nq;
 333         } while (q && q->tqe_th->th_seq == tp->rcv_nxt);
 334         ND6_HINT(tp);
 335         sorwakeup_locked(so);
 336         return (flags);
 337 }