[BACK]Return to ip_input.c CVS log [TXT][DIR] Up to [cvs.NetBSD.org] / src / sys / netinet

Please note that diffs are not public domain; they are subject to the copyright notices on the relevant files.

Diff for /src/sys/netinet/ip_input.c between version 1.193 and 1.194

version 1.193, 2003/12/12 21:17:59 version 1.194, 2003/12/14 00:09:24
Line 236  uint16_t ip_id;
Line 236  uint16_t ip_id;
 struct pfil_head inet_pfil_hook;  struct pfil_head inet_pfil_hook;
 #endif  #endif
   
   /*
    * Cached copy of nmbclusters. If nbclusters is different,
    * recalculate IP parameters derived from nmbclusters.
    */
   static int      ip_nmbclusters;                 /* copy of nmbclusters */
   static void     ip_nmbclusters_changed __P((void));     /* recalc limits */
   
   #define CHECK_NMBCLUSTER_PARAMS() \
   do { if __predict_false(ip_nmbclusters != nmbclusters)  \
           ip_nmbclusters_changed();                       \
   } while  (0)
   
   
 /* IP datagram reassembly queues (hashed) */  /* IP datagram reassembly queues (hashed) */
 #define IPREASS_NHASH_LOG2      6  #define IPREASS_NHASH_LOG2      6
 #define IPREASS_NHASH           (1 << IPREASS_NHASH_LOG2)  #define IPREASS_NHASH           (1 << IPREASS_NHASH_LOG2)
Line 244  struct pfil_head inet_pfil_hook;
Line 257  struct pfil_head inet_pfil_hook;
         (((((x) & 0xF) | ((((x) >> 8) & 0xF) << 4)) ^ (y)) & IPREASS_HMASK)          (((((x) & 0xF) | ((((x) >> 8) & 0xF) << 4)) ^ (y)) & IPREASS_HMASK)
 struct ipqhead ipq[IPREASS_NHASH];  struct ipqhead ipq[IPREASS_NHASH];
 int     ipq_locked;  int     ipq_locked;
 int     ip_nfragpackets = 0;  static int      ip_nfragpackets;        /* packets in reass queue */
 int     ip_maxfragpackets = 200;  static int      ip_nfrags;              /* total fragments in reass queues */
 int     ip_nfrags = 0;         /* total fragments in reass queues */  
   int     ip_maxfragpackets = 200;        /* limit on packets. XXX sysctl */
   int     ip_maxfrags;                    /* limit on fragments. XXX sysctl */
   
   
   /*
    * Additive-Increase/Multiplicative-Decrease (AIMD) strategy for
    * IP reassembly queue buffer managment.
    *
    * We keep a count of total IP fragments (NB: not fragmented packets!)
    * awaiting reassembly (ip_nfrags) and a limit (ip_maxfrags) on fragments.
    * If ip_nfrags exceeds ip_maxfrags the limit, we drop half the
    * total fragments in  reassembly queues.This AIMD policy avoids
    * repeatedly deleting single packets under heavy fragmentation load
    * (e.g., from lossy NFS peers).
    */
   static u_int    ip_reass_ttl_decr __P((u_int ticks));
   static void     ip_reass_drophalf __P((void));
   
   
 static __inline int ipq_lock_try __P((void));  static __inline int ipq_lock_try __P((void));
 static __inline void ipq_unlock __P((void));  static __inline void ipq_unlock __P((void));
Line 346  struct mowner ip_tx_mowner = { "internet
Line 377  struct mowner ip_tx_mowner = { "internet
 #endif  #endif
   
 /*  /*
    * Compute IP limits derived from the value of nmbclusters.
    */
   static void
   ip_nmbclusters_changed(void)
   {
           ip_maxfrags = nmbclusters / 4;
           ip_nmbclusters =  nmbclusters;
   }
   
   /*
  * IP initialization: fill in IP protocol switch table.   * IP initialization: fill in IP protocol switch table.
  * All protocols not implemented in kernel go to raw IP protocol handler.   * All protocols not implemented in kernel go to raw IP protocol handler.
  */   */
Line 375  ip_init()
Line 416  ip_init()
                 LIST_INIT(&ipq[i]);                  LIST_INIT(&ipq[i]);
   
         ip_id = time.tv_sec & 0xfffff;          ip_id = time.tv_sec & 0xfffff;
   
         ipintrq.ifq_maxlen = ipqmaxlen;          ipintrq.ifq_maxlen = ipqmaxlen;
           ip_nmbclusters_changed();
   
         TAILQ_INIT(&in_ifaddrhead);          TAILQ_INIT(&in_ifaddrhead);
         in_ifaddrhashtbl = hashinit(IN_IFADDR_HASH_SIZE, HASH_LIST, M_IFADDR,          in_ifaddrhashtbl = hashinit(IN_IFADDR_HASH_SIZE, HASH_LIST, M_IFADDR,
             M_WAITOK, &in_ifaddrhash);              M_WAITOK, &in_ifaddrhash);
Line 1023  ip_reass(ipqe, fp, ipqhead)
Line 1067  ip_reass(ipqe, fp, ipqhead)
         m->m_data += hlen;          m->m_data += hlen;
         m->m_len -= hlen;          m->m_len -= hlen;
   
   #ifdef  notyet
           /* make sure fragment limit is up-to-date */
           CHECK_NMBCLUSTER_PARAMS();
   
           /* If we have too many fragments, drop the older half. */
           if (ip_nfrags >= ip_maxfrags)
                   ip_reass_drophalf(void);
   #endif
   
         /*          /*
          * We are about to add a fragment; increment frag count.           * We are about to add a fragment; increment frag count.
          */           */
Line 1221  ip_freef(fp)
Line 1274  ip_freef(fp)
 }  }
   
 /*  /*
    * IP reassembly TTL machinery for  multiplicative drop.
    */
   static u_int    fragttl_histo[(IPFRAGTTL+1)];
   
   
   /*
    * Decrement TTL of all reasembly queue entries by `ticks'.
    * Count number of distinct fragments (as opposed to partial, fragmented
    * datagrams) in the reassembly queue.  While we  traverse the entire
    * reassembly queue, compute and return the median TTL over all fragments.
    */
   static u_int
   ip_reass_ttl_decr(u_int ticks)
   {
           u_int i, nfrags, median;
           struct ipq *fp, *nfp;
           u_int dropfraction, keepfraction;
   
           nfrags = 0;
           memset(fragttl_histo, 0, sizeof fragttl_histo);
   
           for (i = 0; i < IPREASS_NHASH; i++) {
                   for (fp = LIST_FIRST(&ipq[i]); fp != NULL; fp = nfp) {
                           fp->ipq_ttl = ((fp->ipq_ttl  <= ticks) ?
                                          0 : fp->ipq_ttl - ticks);
                           nfp = LIST_NEXT(fp, ipq_q);
                           if (fp->ipq_ttl == 0) {
                                   ipstat.ips_fragtimeout++;
                                   ip_freef(fp);
                           } else {
                                   nfrags += fp->ipq_nfrags;
                                   fragttl_histo[fp->ipq_ttl] += fp->ipq_nfrags;
                           }
                   }
           }
   
           KASSERT(ip_nfrags == nfrags);
   
           /* Find median (or other drop fraction) in histogram. */
           dropfraction = (ip_nfrags / 2);
           keepfraction = ip_nfrags - dropfraction;
           for (i = IPFRAGTTL, median = 0; i >= 0; i--) {
                   median +=  fragttl_histo[i];
                   if (median >= keepfraction)
                           break;
           }
   
           /* Return TTL of median (or other fraction). */
           return (u_int)i;
   }
   
   void
   ip_reass_drophalf(void)
   {
   
           u_int median_ticks;
           /*
            * Compute median TTL of all fragments, and count frags
            * with that TTL or lower (roughly half of all fragments).
            */
           median_ticks = ip_reass_ttl_decr(0);
   
           /* Drop half. */
           median_ticks = ip_reass_ttl_decr(median_ticks);
   
   }
   
   /*
  * IP timer processing;   * IP timer processing;
  * if a timer expires on a reassembly   * if a timer expires on a reassembly
  * queue, discard it.   * queue, discard it.
Line 1230  ip_slowtimo()
Line 1351  ip_slowtimo()
 {  {
         static u_int dropscanidx = 0;          static u_int dropscanidx = 0;
         u_int i;          u_int i;
         struct ipq *fp, *nfp;          u_int median_ttl;
         int s = splsoftnet();          int s = splsoftnet();
   
         IPQ_LOCK();          IPQ_LOCK();
         for (i = 0; i < IPREASS_NHASH; i++) {  
                 for (fp = LIST_FIRST(&ipq[i]); fp != NULL; fp = nfp) {          /* Age TTL of all fragments by 1 tick .*/
                         nfp = LIST_NEXT(fp, ipq_q);          median_ttl = ip_reass_ttl_decr(1);
                         if (--fp->ipq_ttl == 0) {  
                                 ipstat.ips_fragtimeout++;          /* make sure fragment limit is up-to-date */
                                 ip_freef(fp);          CHECK_NMBCLUSTER_PARAMS();
                         }  
                 }          /* If we have too many fragments, drop the older half. */
         }          if (ip_nfrags > ip_maxfrags)
                   ip_reass_ttl_decr(median_ttl);
   
         /*          /*
          * If we are over the maximum number of fragments           * If we are over the maximum number of fragmented packets
          * (due to the limit being lowered), drain off           * (due to the limit being lowered), drain off
          * enough to get down to the new limit. Start draining           * enough to get down to the new limit. Start draining
          * from the reassembly hashqueue most recently drained.           * from the reassembly hashqueue most recently drained.
Line 1283  ip_slowtimo()
Line 1406  ip_slowtimo()
 void  void
 ip_drain()  ip_drain()
 {  {
         int i;  
   
         /*          /*
          * We may be called from a device's interrupt context.  If           * We may be called from a device's interrupt context.  If
Line 1292  ip_drain()
Line 1414  ip_drain()
         if (ipq_lock_try() == 0)          if (ipq_lock_try() == 0)
                 return;                  return;
   
         for (i = 0; i < IPREASS_NHASH; i++) {          /*
                 struct ipqhead *ipqh = &ipq[i];           * Drop half the total fragments now. If more mbufs are needed,
                 struct ipq *fp, *nfp;           *  we will be called again soon.
                 for (fp = LIST_FIRST(ipqh); fp != NULL; fp = nfp) {           */
                         nfp = LIST_NEXT(fp, ipq_q);          ip_reass_drophalf();
                         ip_freef(fp);  
                         ipstat.ips_fragdropped++;  
                 }  
         }  
   
         IPQ_UNLOCK();          IPQ_UNLOCK();
 }  }
Legend:
Removed from v.1.193  
changed lines
  Added in v.1.194
CVSweb <webmaster@jp.NetBSD.org>