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