net/core/dst.c

   1 /*
   2  * net/core/dst.c       Protocol independent destination cache.
   3  *
   4  * Authors:             Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
   5  *
   6  */
   7
   8 #include <linux/bitops.h>
   9 #include <linux/errno.h>
  10 #include <linux/init.h>
  11 #include <linux/kernel.h>
  12 #include <linux/mm.h>
  13 #include <linux/module.h>
  14 #include <linux/netdevice.h>
  15 #include <linux/sched.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/string.h>
  18 #include <linux/types.h>
  19
  20 #include <net/dst.h>
  21
  22 /* Locking strategy:
  23  * 1) Garbage collection state of dead destination cache
  24  *    entries is protected by dst_lock.
  25  * 2) GC is run only from BH context, and is the only remover
  26  *    of entries.
  27  * 3) Entries are added to the garbage list from both BH
  28  *    and non-BH context, so local BH disabling is needed.
  29  * 4) All operations modify state, so a spinlock is used.
  30  */
  31 static struct dst_entry         *dst_garbage_list;
  32 #if RT_CACHE_DEBUG >= 2
  33 static atomic_t                  dst_total = ATOMIC_INIT(0);
  34 #endif
  35 static DEFINE_SPINLOCK(dst_lock);
  36
  37 static unsigned long dst_gc_timer_expires;
  38 static unsigned long dst_gc_timer_inc = DST_GC_MAX;
  39 static void dst_run_gc(unsigned long);
  40 static void ___dst_free(struct dst_entry * dst);
  41
  42 static struct timer_list dst_gc_timer =
  43         TIMER_INITIALIZER(dst_run_gc, DST_GC_MIN, 0);
  44
  45 static void dst_run_gc(unsigned long dummy)
  46 {
  47         int    delayed = 0;
  48         struct dst_entry * dst, **dstp;
  49
  50         if (!spin_trylock(&dst_lock)) {
  51                 mod_timer(&dst_gc_timer, jiffies + HZ/10);
  52                 return;
  53         }
  54
  55
  56         del_timer(&dst_gc_timer);
  57         dstp = &dst_garbage_list;
  58         while ((dst = *dstp) != NULL) {
  59                 if (atomic_read(&dst->__refcnt)) {
  60                         dstp = &dst->next;
  61                         delayed++;
  62                         continue;
  63                 }
  64                 *dstp = dst->next;
  65
  66                 dst = dst_destroy(dst);
  67                 if (dst) {
  68                         /* NOHASH and still referenced. Unless it is already
  69                          * on gc list, invalidate it and add to gc list.
  70                          *
  71                          * Note: this is temporary. Actually, NOHASH dst's
  72                          * must be obsoleted when parent is obsoleted.
  73                          * But we do not have state "obsoleted, but
  74                          * referenced by parent", so it is right.
  75                          */
  76                         if (dst->obsolete > 1)
  77                                 continue;
  78
  79                         ___dst_free(dst);
  80                         dst->next = *dstp;
  81                         *dstp = dst;
  82                         dstp = &dst->next;
  83                 }
  84         }
  85         if (!dst_garbage_list) {
  86                 dst_gc_timer_inc = DST_GC_MAX;
  87                 goto out;
  88         }
  89         if ((dst_gc_timer_expires += dst_gc_timer_inc) > DST_GC_MAX)
  90                 dst_gc_timer_expires = DST_GC_MAX;
  91         dst_gc_timer_inc += DST_GC_INC;
  92         dst_gc_timer.expires = jiffies + dst_gc_timer_expires;
  93 #if RT_CACHE_DEBUG >= 2
  94         printk("dst_total: %d/%d %ld\n",
  95                atomic_read(&dst_total), delayed,  dst_gc_timer_expires);
  96 #endif
  97         add_timer(&dst_gc_timer);
  98
  99 out:
 100         spin_unlock(&dst_lock);
 101 }
 102
 103 static int dst_discard_in(struct sk_buff *skb)
 104 {
 105         kfree_skb(skb);
 106         return 0;
 107 }
 108
 109 static int dst_discard_out(struct sk_buff *skb)
 110 {
 111         kfree_skb(skb);
 112         return 0;
 113 }
 114
 115 void * dst_alloc(struct dst_ops * ops)
 116 {
 117         struct dst_entry * dst;
 118
 119         if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
 120                 if (ops->gc())
 121                         return NULL;
 122         }
 123         dst = kmem_cache_alloc(ops->kmem_cachep, SLAB_ATOMIC);
 124         if (!dst)
 125                 return NULL;
 126         memset(dst, 0, ops->entry_size);
 127         atomic_set(&dst->__refcnt, 0);
 128         dst->ops = ops;
 129         dst->lastuse = jiffies;
 130         dst->path = dst;
 131         dst->input = dst_discard_in;
 132         dst->output = dst_discard_out;
 133 #if RT_CACHE_DEBUG >= 2
 134         atomic_inc(&dst_total);
 135 #endif
 136         atomic_inc(&ops->entries);
 137         return dst;
 138 }
 139
 140 static void ___dst_free(struct dst_entry * dst)
 141 {
 142         /* The first case (dev==NULL) is required, when
 143            protocol module is unloaded.
 144          */
 145         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 146                 dst->input = dst_discard_in;
 147                 dst->output = dst_discard_out;
 148         }
 149         dst->obsolete = 2;
 150 }
 151
 152 void __dst_free(struct dst_entry * dst)
 153 {
 154         spin_lock_bh(&dst_lock);
 155         ___dst_free(dst);
 156         dst->next = dst_garbage_list;
 157         dst_garbage_list = dst;
 158         if (dst_gc_timer_inc > DST_GC_INC) {
 159                 dst_gc_timer_inc = DST_GC_INC;
 160                 dst_gc_timer_expires = DST_GC_MIN;
 161                 mod_timer(&dst_gc_timer, jiffies + dst_gc_timer_expires);
 162         }
 163         spin_unlock_bh(&dst_lock);
 164 }
 165
 166 struct dst_entry *dst_destroy(struct dst_entry * dst)
 167 {
 168         struct dst_entry *child;
 169         struct neighbour *neigh;
 170         struct hh_cache *hh;
 171
 172         smp_rmb();
 173
 174 again:
 175         neigh = dst->neighbour;
 176         hh = dst->hh;
 177         child = dst->child;
 178
 179         dst->hh = NULL;
 180         if (hh && atomic_dec_and_test(&hh->hh_refcnt))
 181                 kfree(hh);
 182
 183         if (neigh) {
 184                 dst->neighbour = NULL;
 185                 neigh_release(neigh);
 186         }
 187
 188         atomic_dec(&dst->ops->entries);
 189
 190         if (dst->ops->destroy)
 191                 dst->ops->destroy(dst);
 192         if (dst->dev)
 193                 dev_put(dst->dev);
 194 #if RT_CACHE_DEBUG >= 2
 195         atomic_dec(&dst_total);
 196 #endif
 197         kmem_cache_free(dst->ops->kmem_cachep, dst);
 198
 199         dst = child;
 200         if (dst) {
 201                 if (atomic_dec_and_test(&dst->__refcnt)) {
 202                         /* We were real parent of this dst, so kill child. */
 203                         if (dst->flags&DST_NOHASH)
 204                                 goto again;
 205                 } else {
 206                         /* Child is still referenced, return it for freeing. */
 207                         if (dst->flags&DST_NOHASH)
 208                                 return dst;
 209                         /* Child is still in his hash table */
 210                 }
 211         }
 212         return NULL;
 213 }
 214
 215 /* Dirty hack. We did it in 2.2 (in __dst_free),
 216  * we have _very_ good reasons not to repeat
 217  * this mistake in 2.3, but we have no choice
 218  * now. _It_ _is_ _explicit_ _deliberate_
 219  * _race_ _condition_.
 220  *
 221  * Commented and originally written by Alexey.
 222  */
 223 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 224                               int unregister)
 225 {
 226         if (dst->ops->ifdown)
 227                 dst->ops->ifdown(dst, dev, unregister);
 228
 229         if (dev != dst->dev)
 230                 return;
 231
 232         if (!unregister) {
 233                 dst->input = dst_discard_in;
 234                 dst->output = dst_discard_out;
 235         } else {
 236                 dst->dev = &loopback_dev;
 237                 dev_hold(&loopback_dev);
 238                 dev_put(dev);
 239                 if (dst->neighbour && dst->neighbour->dev == dev) {
 240                         dst->neighbour->dev = &loopback_dev;
 241                         dev_put(dev);
 242                         dev_hold(&loopback_dev);
 243                 }
 244         }
 245 }
 246
 247 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
 248 {
 249         struct net_device *dev = ptr;
 250         struct dst_entry *dst;
 251
 252         switch (event) {
 253         case NETDEV_UNREGISTER:
 254         case NETDEV_DOWN:
 255                 spin_lock_bh(&dst_lock);
 256                 for (dst = dst_garbage_list; dst; dst = dst->next) {
 257                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 258                 }
 259                 spin_unlock_bh(&dst_lock);
 260                 break;
 261         }
 262         return NOTIFY_DONE;
 263 }
 264
 265 static struct notifier_block dst_dev_notifier = {
 266         .notifier_call  = dst_dev_event,
 267 };
 268
 269 void __init dst_init(void)
 270 {
 271         register_netdevice_notifier(&dst_dev_notifier);
 272 }
 273
 274 EXPORT_SYMBOL(__dst_free);
 275 EXPORT_SYMBOL(dst_alloc);
 276 EXPORT_SYMBOL(dst_destroy);