release/src-rt-6.x/linux/linux-2.6/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/workqueue.h>
  13 #include <linux/mm.h>
  14 #include <linux/module.h>
  15 #include <linux/netdevice.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/string.h>
  18 #include <linux/types.h>
  19
  20 #include <net/dst.h>
  21
  22 /*
  23  * Theory of operations:
  24  * 1) We use a list, protected by a spinlock, to add
  25  *    new entries from both BH and non-BH context.
  26  * 2) In order to keep spinlock held for a small delay,
  27  *    we use a second list where are stored long lived
  28  *    entries, that are handled by the garbage collect thread
  29  *    fired by a workqueue.
  30  * 3) This list is guarded by a mutex,
  31  *    so that the gc_task and dst_dev_event() can be synchronized.
  32  */
  33 #if RT_CACHE_DEBUG >= 2
  34 static atomic_t                  dst_total = ATOMIC_INIT(0);
  35 #endif
  36
  37 /*
  38  * We want to keep lock & list close together
  39  * to dirty as few cache lines as possible in __dst_free().
  40  * As this is not a very strong hint, we dont force an alignment on SMP.
  41  */
  42 static struct {
  43         spinlock_t              lock;
  44         struct dst_entry        *list;
  45         unsigned long           timer_inc;
  46         unsigned long           timer_expires;
  47 } dst_garbage = {
  48         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  49         .timer_inc = DST_GC_MAX,
  50 };
  51 static void dst_gc_task(struct work_struct *work);
  52 static void ___dst_free(struct dst_entry * dst);
  53
  54 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  55
  56 static DEFINE_MUTEX(dst_gc_mutex);
  57 /*
  58  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  59  */
  60 static struct dst_entry         *dst_busy_list;
  61
  62 static void dst_gc_task(struct work_struct *work)
  63 {
  64         int    delayed = 0;
  65         int    work_performed = 0;
  66         unsigned long expires = ~0L;
  67         struct dst_entry *dst, *next, head;
  68         struct dst_entry *last = &head;
  69 #if RT_CACHE_DEBUG >= 2
  70         ktime_t time_start = ktime_get();
  71         struct timespec elapsed;
  72 #endif
  73
  74         mutex_lock(&dst_gc_mutex);
  75         next = dst_busy_list;
  76
  77 loop:
  78         while ((dst = next) != NULL) {
  79                 next = dst->next;
  80                 prefetch(&next->next);
  81                 if (likely(atomic_read(&dst->__refcnt))) {
  82                         last->next = dst;
  83                         last = dst;
  84                         delayed++;
  85                         continue;
  86                 }
  87                 work_performed++;
  88
  89                 dst = dst_destroy(dst);
  90                 if (dst) {
  91                         /* NOHASH and still referenced. Unless it is already
  92                          * on gc list, invalidate it and add to gc list.
  93                          *
  94                          * Note: this is temporary. Actually, NOHASH dst's
  95                          * must be obsoleted when parent is obsoleted.
  96                          * But we do not have state "obsoleted, but
  97                          * referenced by parent", so it is right.
  98                          */
  99                         if (dst->obsolete > 1)
 100                                 continue;
 101
 102                         ___dst_free(dst);
 103                         dst->next = next;
 104                         next = dst;
 105                 }
 106         }
 107
 108         spin_lock_bh(&dst_garbage.lock);
 109         next = dst_garbage.list;
 110         if (next) {
 111                 dst_garbage.list = NULL;
 112                 spin_unlock_bh(&dst_garbage.lock);
 113                 goto loop;
 114         }
 115         last->next = NULL;
 116         dst_busy_list = head.next;
 117         if (!dst_busy_list)
 118                 dst_garbage.timer_inc = DST_GC_MAX;
 119         else {
 120                 /*
 121                  * if we freed less than 1/10 of delayed entries,
 122                  * we can sleep longer.
 123                  */
 124                 if (work_performed <= delayed/10) {
 125                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 126                         if (dst_garbage.timer_expires > DST_GC_MAX)
 127                                 dst_garbage.timer_expires = DST_GC_MAX;
 128                         dst_garbage.timer_inc += DST_GC_INC;
 129                 } else {
 130                         dst_garbage.timer_inc = DST_GC_INC;
 131                         dst_garbage.timer_expires = DST_GC_MIN;
 132                 }
 133                 expires = dst_garbage.timer_expires;
 134                 /*
 135                  * if the next desired timer is more than 4 seconds in the future
 136                  * then round the timer to whole seconds
 137                  */
 138                 if (expires > 4*HZ)
 139                         expires = round_jiffies_relative(expires);
 140                 schedule_delayed_work(&dst_gc_work, expires);
 141         }
 142
 143         spin_unlock_bh(&dst_garbage.lock);
 144         mutex_unlock(&dst_gc_mutex);
 145 #if RT_CACHE_DEBUG >= 2
 146         elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
 147         printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
 148                 " expires: %lu elapsed: %lu us\n",
 149                 atomic_read(&dst_total), delayed, work_performed,
 150                 expires,
 151                 elapsed.tv_sec * USEC_PER_SEC + elapsed.tv_nsec / NSEC_PER_USEC);
 152 #endif
 153 }
 154
 155 static int dst_discard(struct sk_buff *skb)
 156 {
 157         kfree_skb(skb);
 158         return 0;
 159 }
 160
 161 void * dst_alloc(struct dst_ops * ops)
 162 {
 163         struct dst_entry * dst;
 164
 165         if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
 166                 if (ops->gc())
 167                         return NULL;
 168         }
 169         dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
 170         if (!dst)
 171                 return NULL;
 172         atomic_set(&dst->__refcnt, 0);
 173         dst->ops = ops;
 174         dst->lastuse = jiffies;
 175         dst->path = dst;
 176         dst->input = dst->output = dst_discard;
 177 #if RT_CACHE_DEBUG >= 2
 178         atomic_inc(&dst_total);
 179 #endif
 180         atomic_inc(&ops->entries);
 181         return dst;
 182 }
 183
 184 static void ___dst_free(struct dst_entry * dst)
 185 {
 186         /* The first case (dev==NULL) is required, when
 187            protocol module is unloaded.
 188          */
 189         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 190                 dst->input = dst->output = dst_discard;
 191         }
 192         dst->obsolete = 2;
 193 }
 194
 195 void __dst_free(struct dst_entry * dst)
 196 {
 197         spin_lock_bh(&dst_garbage.lock);
 198         ___dst_free(dst);
 199         dst->next = dst_garbage.list;
 200         dst_garbage.list = dst;
 201         if (dst_garbage.timer_inc > DST_GC_INC) {
 202                 dst_garbage.timer_inc = DST_GC_INC;
 203                 dst_garbage.timer_expires = DST_GC_MIN;
 204                 cancel_delayed_work(&dst_gc_work);
 205                 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 206         }
 207         spin_unlock_bh(&dst_garbage.lock);
 208 }
 209
 210 struct dst_entry *dst_destroy(struct dst_entry * dst)
 211 {
 212         struct dst_entry *child;
 213         struct neighbour *neigh;
 214         struct hh_cache *hh;
 215
 216         smp_rmb();
 217
 218 again:
 219         neigh = dst->neighbour;
 220         hh = dst->hh;
 221         child = dst->child;
 222
 223         dst->hh = NULL;
 224         if (hh && atomic_dec_and_test(&hh->hh_refcnt))
 225                 kfree(hh);
 226
 227         if (neigh) {
 228                 dst->neighbour = NULL;
 229                 neigh_release(neigh);
 230         }
 231
 232         atomic_dec(&dst->ops->entries);
 233
 234         if (dst->ops->destroy)
 235                 dst->ops->destroy(dst);
 236         if (dst->dev)
 237                 dev_put(dst->dev);
 238 #if RT_CACHE_DEBUG >= 2
 239         atomic_dec(&dst_total);
 240 #endif
 241         kmem_cache_free(dst->ops->kmem_cachep, dst);
 242
 243         dst = child;
 244         if (dst) {
 245                 int nohash = dst->flags & DST_NOHASH;
 246
 247                 if (atomic_dec_and_test(&dst->__refcnt)) {
 248                         /* We were real parent of this dst, so kill child. */
 249                         if (nohash)
 250                                 goto again;
 251                 } else {
 252                         /* Child is still referenced, return it for freeing. */
 253                         if (nohash)
 254                                 return dst;
 255                         /* Child is still in his hash table */
 256                 }
 257         }
 258         return NULL;
 259 }
 260
 261 /* Dirty hack. We did it in 2.2 (in __dst_free),
 262  * we have _very_ good reasons not to repeat
 263  * this mistake in 2.3, but we have no choice
 264  * now. _It_ _is_ _explicit_ _deliberate_
 265  * _race_ _condition_.
 266  *
 267  * Commented and originally written by Alexey.
 268  */
 269 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 270                               int unregister)
 271 {
 272         if (dst->ops->ifdown)
 273                 dst->ops->ifdown(dst, dev, unregister);
 274
 275         if (dev != dst->dev)
 276                 return;
 277
 278         if (!unregister) {
 279                 dst->input = dst->output = dst_discard;
 280         } else {
 281                 dst->dev = &loopback_dev;
 282                 dev_hold(&loopback_dev);
 283                 dev_put(dev);
 284                 if (dst->neighbour && dst->neighbour->dev == dev) {
 285                         dst->neighbour->dev = &loopback_dev;
 286                         dev_put(dev);
 287                         dev_hold(&loopback_dev);
 288                 }
 289         }
 290 }
 291
 292 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
 293 {
 294         struct net_device *dev = ptr;
 295         struct dst_entry *dst, *last = NULL;
 296
 297         switch (event) {
 298         case NETDEV_UNREGISTER:
 299         case NETDEV_DOWN:
 300                 mutex_lock(&dst_gc_mutex);
 301                 for (dst = dst_busy_list; dst; dst = dst->next) {
 302                         last = dst;
 303                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 304                 }
 305
 306                 spin_lock_bh(&dst_garbage.lock);
 307                 dst = dst_garbage.list;
 308                 dst_garbage.list = NULL;
 309                 spin_unlock_bh(&dst_garbage.lock);
 310
 311                 if (last)
 312                         last->next = dst;
 313                 else
 314                         dst_busy_list = dst;
 315                 for (; dst; dst = dst->next) {
 316                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 317                 }
 318                 mutex_unlock(&dst_gc_mutex);
 319                 break;
 320         }
 321         return NOTIFY_DONE;
 322 }
 323
 324 static struct notifier_block dst_dev_notifier = {
 325         .notifier_call  = dst_dev_event,
 326 };
 327
 328 void __init dst_init(void)
 329 {
 330         register_netdevice_notifier(&dst_dev_notifier);
 331 }
 332
 333 EXPORT_SYMBOL(__dst_free);
 334 EXPORT_SYMBOL(dst_alloc);
 335 EXPORT_SYMBOL(dst_destroy);