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         int    work_performed;
  49         struct dst_entry * dst, **dstp;
  50
  51         if (!spin_trylock(&dst_lock)) {
  52                 mod_timer(&dst_gc_timer, jiffies + HZ/10);
  53                 return;
  54         }
  55
  56         del_timer(&dst_gc_timer);
  57         dstp = &dst_garbage_list;
  58         work_performed = 0;
  59         while ((dst = *dstp) != NULL) {
  60                 if (atomic_read(&dst->__refcnt)) {
  61                         dstp = &dst->next;
  62                         delayed++;
  63                         continue;
  64                 }
  65                 *dstp = dst->next;
  66                 work_performed = 1;
  67
  68                 dst = dst_destroy(dst);
  69                 if (dst) {
  70                         /* NOHASH and still referenced. Unless it is already
  71                          * on gc list, invalidate it and add to gc list.
  72                          *
  73                          * Note: this is temporary. Actually, NOHASH dst's
  74                          * must be obsoleted when parent is obsoleted.
  75                          * But we do not have state "obsoleted, but
  76                          * referenced by parent", so it is right.
  77                          */
  78                         if (dst->obsolete > 1)
  79                                 continue;
  80
  81                         ___dst_free(dst);
  82                         dst->next = *dstp;
  83                         *dstp = dst;
  84                         dstp = &dst->next;
  85                 }
  86         }
  87         if (!dst_garbage_list) {
  88                 dst_gc_timer_inc = DST_GC_MAX;
  89                 goto out;
  90         }
  91         if (!work_performed) {
  92                 if ((dst_gc_timer_expires += dst_gc_timer_inc) > DST_GC_MAX)
  93                         dst_gc_timer_expires = DST_GC_MAX;
  94                 dst_gc_timer_inc += DST_GC_INC;
  95         } else {
  96                 dst_gc_timer_inc = DST_GC_INC;
  97                 dst_gc_timer_expires = DST_GC_MIN;
  98         }
  99         dst_gc_timer.expires = jiffies + dst_gc_timer_expires;
 100 #if RT_CACHE_DEBUG >= 2
 101         printk("dst_total: %d/%d %ld\n",
 102                atomic_read(&dst_total), delayed,  dst_gc_timer_expires);
 103 #endif
 104         add_timer(&dst_gc_timer);
 105
 106 out:
 107         spin_unlock(&dst_lock);
 108 }
 109
 110 static int dst_discard_in(struct sk_buff *skb)
 111 {
 112         kfree_skb(skb);
 113         return 0;
 114 }
 115
 116 static int dst_discard_out(struct sk_buff *skb)
 117 {
 118         kfree_skb(skb);
 119         return 0;
 120 }
 121
 122 void * dst_alloc(struct dst_ops * ops)
 123 {
 124         struct dst_entry * dst;
 125
 126         if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
 127                 if (ops->gc())
 128                         return NULL;
 129         }
 130         dst = kmem_cache_alloc(ops->kmem_cachep, SLAB_ATOMIC);
 131         if (!dst)
 132                 return NULL;
 133         memset(dst, 0, ops->entry_size);
 134         atomic_set(&dst->__refcnt, 0);
 135         dst->ops = ops;
 136         dst->lastuse = jiffies;
 137         dst->path = dst;
 138         dst->input = dst_discard_in;
 139         dst->output = dst_discard_out;
 140 #if RT_CACHE_DEBUG >= 2
 141         atomic_inc(&dst_total);
 142 #endif
 143         atomic_inc(&ops->entries);
 144         return dst;
 145 }
 146
 147 static void ___dst_free(struct dst_entry * dst)
 148 {
 149         /* The first case (dev==NULL) is required, when
 150            protocol module is unloaded.
 151          */
 152         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 153                 dst->input = dst_discard_in;
 154                 dst->output = dst_discard_out;
 155         }
 156         dst->obsolete = 2;
 157 }
 158
 159 void __dst_free(struct dst_entry * dst)
 160 {
 161         spin_lock_bh(&dst_lock);
 162         ___dst_free(dst);
 163         dst->next = dst_garbage_list;
 164         dst_garbage_list = dst;
 165         if (dst_gc_timer_inc > DST_GC_INC) {
 166                 dst_gc_timer_inc = DST_GC_INC;
 167                 dst_gc_timer_expires = DST_GC_MIN;
 168                 mod_timer(&dst_gc_timer, jiffies + dst_gc_timer_expires);
 169         }
 170         spin_unlock_bh(&dst_lock);
 171 }
 172
 173 struct dst_entry *dst_destroy(struct dst_entry * dst)
 174 {
 175         struct dst_entry *child;
 176         struct neighbour *neigh;
 177         struct hh_cache *hh;
 178
 179         smp_rmb();
 180
 181 again:
 182         neigh = dst->neighbour;
 183         hh = dst->hh;
 184         child = dst->child;
 185
 186         dst->hh = NULL;
 187         if (hh && atomic_dec_and_test(&hh->hh_refcnt))
 188                 kfree(hh);
 189
 190         if (neigh) {
 191                 dst->neighbour = NULL;
 192                 neigh_release(neigh);
 193         }
 194
 195         atomic_dec(&dst->ops->entries);
 196
 197         if (dst->ops->destroy)
 198                 dst->ops->destroy(dst);
 199         if (dst->dev)
 200                 dev_put(dst->dev);
 201 #if RT_CACHE_DEBUG >= 2
 202         atomic_dec(&dst_total);
 203 #endif
 204         kmem_cache_free(dst->ops->kmem_cachep, dst);
 205
 206         dst = child;
 207         if (dst) {
 208                 int nohash = dst->flags & DST_NOHASH;
 209
 210                 if (atomic_dec_and_test(&dst->__refcnt)) {
 211                         /* We were real parent of this dst, so kill child. */
 212                         if (nohash)
 213                                 goto again;
 214                 } else {
 215                         /* Child is still referenced, return it for freeing. */
 216                         if (nohash)
 217                                 return dst;
 218                         /* Child is still in his hash table */
 219                 }
 220         }
 221         return NULL;
 222 }
 223
 224 /* Dirty hack. We did it in 2.2 (in __dst_free),
 225  * we have _very_ good reasons not to repeat
 226  * this mistake in 2.3, but we have no choice
 227  * now. _It_ _is_ _explicit_ _deliberate_
 228  * _race_ _condition_.
 229  *
 230  * Commented and originally written by Alexey.
 231  */
 232 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 233                               int unregister)
 234 {
 235         if (dst->ops->ifdown)
 236                 dst->ops->ifdown(dst, dev, unregister);
 237
 238         if (dev != dst->dev)
 239                 return;
 240
 241         if (!unregister) {
 242                 dst->input = dst_discard_in;
 243                 dst->output = dst_discard_out;
 244         } else {
 245                 dst->dev = &loopback_dev;
 246                 dev_hold(&loopback_dev);
 247                 dev_put(dev);
 248                 if (dst->neighbour && dst->neighbour->dev == dev) {
 249                         dst->neighbour->dev = &loopback_dev;
 250                         dev_put(dev);
 251                         dev_hold(&loopback_dev);
 252                 }
 253         }
 254 }
 255
 256 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
 257 {
 258         struct net_device *dev = ptr;
 259         struct dst_entry *dst;
 260
 261         switch (event) {
 262         case NETDEV_UNREGISTER:
 263         case NETDEV_DOWN:
 264                 spin_lock_bh(&dst_lock);
 265                 for (dst = dst_garbage_list; dst; dst = dst->next) {
 266                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 267                 }
 268                 spin_unlock_bh(&dst_lock);
 269                 break;
 270         }
 271         return NOTIFY_DONE;
 272 }
 273
 274 static struct notifier_block dst_dev_notifier = {
 275         .notifier_call  = dst_dev_event,
 276 };
 277
 278 void __init dst_init(void)
 279 {
 280         register_netdevice_notifier(&dst_dev_notifier);
 281 }
 282
 283 EXPORT_SYMBOL(__dst_free);
 284 EXPORT_SYMBOL(dst_alloc);
 285 EXPORT_SYMBOL(dst_destroy);