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