net/core/dst.c

   1 /*
   2  * net/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/types.h>
  10 #include <linux/kernel.h>
  11 #include <linux/sched.h>
  12 #include <linux/mm.h>
  13 #include <linux/string.h>
  14 #include <linux/errno.h>
  15 #include <linux/netdevice.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/init.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 spinlock_t                dst_lock = SPIN_LOCK_UNLOCKED;
  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
  40 static struct timer_list dst_gc_timer =
  41         { data: DST_GC_MIN, function: dst_run_gc };
  42
  43
  44 static void dst_run_gc(unsigned long dummy)
  45 {
  46         int    delayed = 0;
  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
  55         del_timer(&dst_gc_timer);
  56         dstp = &dst_garbage_list;
  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                 dst_destroy(dst);
  65         }
  66         if (!dst_garbage_list) {
  67                 dst_gc_timer_inc = DST_GC_MAX;
  68                 goto out;
  69         }
  70         if ((dst_gc_timer_expires += dst_gc_timer_inc) > DST_GC_MAX)
  71                 dst_gc_timer_expires = DST_GC_MAX;
  72         dst_gc_timer_inc += DST_GC_INC;
  73         dst_gc_timer.expires = jiffies + dst_gc_timer_expires;
  74 #if RT_CACHE_DEBUG >= 2
  75         printk("dst_total: %d/%d %ld\n",
  76                atomic_read(&dst_total), delayed,  dst_gc_timer_expires);
  77 #endif
  78         add_timer(&dst_gc_timer);
  79
  80 out:
  81         spin_unlock(&dst_lock);
  82 }
  83
  84 static int dst_discard(struct sk_buff *skb)
  85 {
  86         kfree_skb(skb);
  87         return 0;
  88 }
  89
  90 static int dst_blackhole(struct sk_buff *skb)
  91 {
  92         kfree_skb(skb);
  93         return 0;
  94 }
  95
  96 void * dst_alloc(struct dst_ops * ops)
  97 {
  98         struct dst_entry * dst;
  99
 100         if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
 101                 if (ops->gc())
 102                         return NULL;
 103         }
 104         dst = kmem_cache_alloc(ops->kmem_cachep, SLAB_ATOMIC);
 105         if (!dst)
 106                 return NULL;
 107         memset(dst, 0, ops->entry_size);
 108         atomic_set(&dst->__refcnt, 0);
 109         dst->ops = ops;
 110         dst->lastuse = jiffies;
 111         dst->input = dst_discard;
 112         dst->output = dst_blackhole;
 113 #if RT_CACHE_DEBUG >= 2
 114         atomic_inc(&dst_total);
 115 #endif
 116         atomic_inc(&ops->entries);
 117         return dst;
 118 }
 119
 120 void __dst_free(struct dst_entry * dst)
 121 {
 122         spin_lock_bh(&dst_lock);
 123
 124         /* The first case (dev==NULL) is required, when
 125            protocol module is unloaded.
 126          */
 127         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 128                 dst->input = dst_discard;
 129                 dst->output = dst_blackhole;
 130         }
 131         dst->obsolete = 2;
 132         dst->next = dst_garbage_list;
 133         dst_garbage_list = dst;
 134         if (dst_gc_timer_inc > DST_GC_INC) {
 135                 dst_gc_timer_inc = DST_GC_INC;
 136                 dst_gc_timer_expires = DST_GC_MIN;
 137                 mod_timer(&dst_gc_timer, jiffies + dst_gc_timer_expires);
 138         }
 139
 140         spin_unlock_bh(&dst_lock);
 141 }
 142
 143 void dst_destroy(struct dst_entry * dst)
 144 {
 145         struct neighbour *neigh;
 146         struct hh_cache *hh;
 147
 148         smp_rmb();
 149
 150         neigh = dst->neighbour;
 151         hh = dst->hh;
 152
 153         dst->hh = NULL;
 154         if (hh && atomic_dec_and_test(&hh->hh_refcnt))
 155                 kfree(hh);
 156
 157         if (neigh) {
 158                 dst->neighbour = NULL;
 159                 neigh_release(neigh);
 160         }
 161
 162         atomic_dec(&dst->ops->entries);
 163
 164         if (dst->ops->destroy)
 165                 dst->ops->destroy(dst);
 166         if (dst->dev)
 167                 dev_put(dst->dev);
 168 #if RT_CACHE_DEBUG >= 2
 169         atomic_dec(&dst_total);
 170 #endif
 171         kmem_cache_free(dst->ops->kmem_cachep, dst);
 172 }
 173
 174 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
 175 {
 176         struct net_device *dev = ptr;
 177         struct dst_entry *dst;
 178
 179         switch (event) {
 180         case NETDEV_UNREGISTER:
 181         case NETDEV_DOWN:
 182                 spin_lock_bh(&dst_lock);
 183                 for (dst = dst_garbage_list; dst; dst = dst->next) {
 184                         if (dst->dev == dev) {
 185                                 /* Dirty hack. We did it in 2.2 (in __dst_free),
 186                                    we have _very_ good reasons not to repeat
 187                                    this mistake in 2.3, but we have no choice
 188                                    now. _It_ _is_ _explicit_ _deliberate_
 189                                    _race_ _condition_.
 190                                  */
 191                                 if (event!=NETDEV_DOWN &&
 192                                     !(dev->features & NETIF_F_DYNALLOC) &&
 193                                     dst->output == dst_blackhole) {
 194                                         dst->dev = &loopback_dev;
 195                                         dev_put(dev);
 196                                         dev_hold(&loopback_dev);
 197                                         dst->output = dst_discard;
 198                                         if (dst->neighbour && dst->neighbour->dev == dev) {
 199                                                 dst->neighbour->dev = &loopback_dev;
 200                                                 dev_put(dev);
 201                                                 dev_hold(&loopback_dev);
 202                                         }
 203                                 } else {
 204                                         dst->input = dst_discard;
 205                                         dst->output = dst_blackhole;
 206                                 }
 207                         }
 208                 }
 209                 spin_unlock_bh(&dst_lock);
 210                 break;
 211         }
 212         return NOTIFY_DONE;
 213 }
 214
 215 struct notifier_block dst_dev_notifier = {
 216         dst_dev_event,
 217         NULL,
 218         0
 219 };
 220
 221 void __init dst_init(void)
 222 {
 223         register_netdevice_notifier(&dst_dev_notifier);
 224 }