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 #include <net/net_namespace.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/dst.h>
  23
  24 /*
  25  * Theory of operations:
  26  * 1) We use a list, protected by a spinlock, to add
  27  *    new entries from both BH and non-BH context.
  28  * 2) In order to keep spinlock held for a small delay,
  29  *    we use a second list where are stored long lived
  30  *    entries, that are handled by the garbage collect thread
  31  *    fired by a workqueue.
  32  * 3) This list is guarded by a mutex,
  33  *    so that the gc_task and dst_dev_event() can be synchronized.
  34  */
  35 #if RT_CACHE_DEBUG >= 2
  36 static atomic_t                  dst_total = ATOMIC_INIT(0);
  37 #endif
  38
  39 /*
  40  * We want to keep lock & list close together
  41  * to dirty as few cache lines as possible in __dst_free().
  42  * As this is not a very strong hint, we dont force an alignment on SMP.
  43  */
  44 static struct {
  45         spinlock_t              lock;
  46         struct dst_entry        *list;
  47         unsigned long           timer_inc;
  48         unsigned long           timer_expires;
  49 } dst_garbage = {
  50         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  51         .timer_inc = DST_GC_MAX,
  52 };
  53 static void dst_gc_task(struct work_struct *work);
  54 static void ___dst_free(struct dst_entry * dst);
  55
  56 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  57
  58 static DEFINE_MUTEX(dst_gc_mutex);
  59 /*
  60  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  61  */
  62 static struct dst_entry         *dst_busy_list;
  63
  64 static void dst_gc_task(struct work_struct *work)
  65 {
  66         int    delayed = 0;
  67         int    work_performed = 0;
  68         unsigned long expires = ~0L;
  69         struct dst_entry *dst, *next, head;
  70         struct dst_entry *last = &head;
  71 #if RT_CACHE_DEBUG >= 2
  72         ktime_t time_start = ktime_get();
  73         struct timespec elapsed;
  74 #endif
  75
  76         mutex_lock(&dst_gc_mutex);
  77         next = dst_busy_list;
  78
  79 loop:
  80         while ((dst = next) != NULL) {
  81                 next = dst->next;
  82                 prefetch(&next->next);
  83                 if (likely(atomic_read(&dst->__refcnt))) {
  84                         last->next = dst;
  85                         last = dst;
  86                         delayed++;
  87                         continue;
  88                 }
  89                 work_performed++;
  90
  91                 dst = dst_destroy(dst);
  92                 if (dst) {
  93                         /* NOHASH and still referenced. Unless it is already
  94                          * on gc list, invalidate it and add to gc list.
  95                          *
  96                          * Note: this is temporary. Actually, NOHASH dst's
  97                          * must be obsoleted when parent is obsoleted.
  98                          * But we do not have state "obsoleted, but
  99                          * referenced by parent", so it is right.
 100                          */
 101                         if (dst->obsolete > 1)
 102                                 continue;
 103
 104                         ___dst_free(dst);
 105                         dst->next = next;
 106                         next = dst;
 107                 }
 108         }
 109
 110         spin_lock_bh(&dst_garbage.lock);
 111         next = dst_garbage.list;
 112         if (next) {
 113                 dst_garbage.list = NULL;
 114                 spin_unlock_bh(&dst_garbage.lock);
 115                 goto loop;
 116         }
 117         last->next = NULL;
 118         dst_busy_list = head.next;
 119         if (!dst_busy_list)
 120                 dst_garbage.timer_inc = DST_GC_MAX;
 121         else {
 122                 /*
 123                  * if we freed less than 1/10 of delayed entries,
 124                  * we can sleep longer.
 125                  */
 126                 if (work_performed <= delayed/10) {
 127                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 128                         if (dst_garbage.timer_expires > DST_GC_MAX)
 129                                 dst_garbage.timer_expires = DST_GC_MAX;
 130                         dst_garbage.timer_inc += DST_GC_INC;
 131                 } else {
 132                         dst_garbage.timer_inc = DST_GC_INC;
 133                         dst_garbage.timer_expires = DST_GC_MIN;
 134                 }
 135                 expires = dst_garbage.timer_expires;
 136                 /*
 137                  * if the next desired timer is more than 4 seconds in the future
 138                  * then round the timer to whole seconds
 139                  */
 140                 if (expires > 4*HZ)
 141                         expires = round_jiffies_relative(expires);
 142                 schedule_delayed_work(&dst_gc_work, expires);
 143         }
 144
 145         spin_unlock_bh(&dst_garbage.lock);
 146         mutex_unlock(&dst_gc_mutex);
 147 #if RT_CACHE_DEBUG >= 2
 148         elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
 149         printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
 150                 " expires: %lu elapsed: %lu us\n",
 151                 atomic_read(&dst_total), delayed, work_performed,
 152                 expires,
 153                 elapsed.tv_sec * USEC_PER_SEC + elapsed.tv_nsec / NSEC_PER_USEC);
 154 #endif
 155 }
 156
 157 static int dst_discard(struct sk_buff *skb)
 158 {
 159         kfree_skb(skb);
 160         return 0;
 161 }
 162
 163 void * dst_alloc(struct dst_ops * ops)
 164 {
 165         struct dst_entry * dst;
 166
 167         if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
 168                 if (ops->gc())
 169                         return NULL;
 170         }
 171         dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
 172         if (!dst)
 173                 return NULL;
 174         atomic_set(&dst->__refcnt, 0);
 175         dst->ops = ops;
 176         dst->lastuse = jiffies;
 177         dst->path = dst;
 178         dst->input = dst->output = dst_discard;
 179 #if RT_CACHE_DEBUG >= 2
 180         atomic_inc(&dst_total);
 181 #endif
 182         atomic_inc(&ops->entries);
 183         return dst;
 184 }
 185
 186 static void ___dst_free(struct dst_entry * dst)
 187 {
 188         /* The first case (dev==NULL) is required, when
 189            protocol module is unloaded.
 190          */
 191         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
 192                 dst->input = dst->output = dst_discard;
 193         }
 194         dst->obsolete = 2;
 195 }
 196
 197 void __dst_free(struct dst_entry * dst)
 198 {
 199         spin_lock_bh(&dst_garbage.lock);
 200         ___dst_free(dst);
 201         dst->next = dst_garbage.list;
 202         dst_garbage.list = dst;
 203         if (dst_garbage.timer_inc > DST_GC_INC) {
 204                 dst_garbage.timer_inc = DST_GC_INC;
 205                 dst_garbage.timer_expires = DST_GC_MIN;
 206                 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 207         }
 208         spin_unlock_bh(&dst_garbage.lock);
 209 }
 210
 211 struct dst_entry *dst_destroy(struct dst_entry * dst)
 212 {
 213         struct dst_entry *child;
 214         struct neighbour *neigh;
 215         struct hh_cache *hh;
 216
 217         smp_rmb();
 218
 219 again:
 220         neigh = dst->neighbour;
 221         hh = dst->hh;
 222         child = dst->child;
 223
 224         dst->hh = NULL;
 225         if (hh && atomic_dec_and_test(&hh->hh_refcnt))
 226                 kfree(hh);
 227
 228         if (neigh) {
 229                 dst->neighbour = NULL;
 230                 neigh_release(neigh);
 231         }
 232
 233         atomic_dec(&dst->ops->entries);
 234
 235         if (dst->ops->destroy)
 236                 dst->ops->destroy(dst);
 237         if (dst->dev)
 238                 dev_put(dst->dev);
 239 #if RT_CACHE_DEBUG >= 2
 240         atomic_dec(&dst_total);
 241 #endif
 242         kmem_cache_free(dst->ops->kmem_cachep, dst);
 243
 244         dst = child;
 245         if (dst) {
 246                 int nohash = dst->flags & DST_NOHASH;
 247
 248                 if (atomic_dec_and_test(&dst->__refcnt)) {
 249                         /* We were real parent of this dst, so kill child. */
 250                         if (nohash)
 251                                 goto again;
 252                 } else {
 253                         /* Child is still referenced, return it for freeing. */
 254                         if (nohash)
 255                                 return dst;
 256                         /* Child is still in his hash table */
 257                 }
 258         }
 259         return NULL;
 260 }
 261
 262 /* Dirty hack. We did it in 2.2 (in __dst_free),
 263  * we have _very_ good reasons not to repeat
 264  * this mistake in 2.3, but we have no choice
 265  * now. _It_ _is_ _explicit_ _deliberate_
 266  * _race_ _condition_.
 267  *
 268  * Commented and originally written by Alexey.
 269  */
 270 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 271                               int unregister)
 272 {
 273         if (dst->ops->ifdown)
 274                 dst->ops->ifdown(dst, dev, unregister);
 275
 276         if (dev != dst->dev)
 277                 return;
 278
 279         if (!unregister) {
 280                 dst->input = dst->output = dst_discard;
 281         } else {
 282                 dst->dev = init_net.loopback_dev;
 283                 dev_hold(dst->dev);
 284                 dev_put(dev);
 285                 if (dst->neighbour && dst->neighbour->dev == dev) {
 286                         dst->neighbour->dev = init_net.loopback_dev;
 287                         dev_put(dev);
 288                         dev_hold(dst->neighbour->dev);
 289                 }
 290         }
 291 }
 292
 293 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
 294 {
 295         struct net_device *dev = ptr;
 296         struct dst_entry *dst, *last = NULL;
 297
 298         if (dev->nd_net != &init_net)
 299                 return NOTIFY_DONE;
 300
 301         switch (event) {
 302         case NETDEV_UNREGISTER:
 303         case NETDEV_DOWN:
 304                 mutex_lock(&dst_gc_mutex);
 305                 for (dst = dst_busy_list; dst; dst = dst->next) {
 306                         last = dst;
 307                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 308                 }
 309
 310                 spin_lock_bh(&dst_garbage.lock);
 311                 dst = dst_garbage.list;
 312                 dst_garbage.list = NULL;
 313                 spin_unlock_bh(&dst_garbage.lock);
 314
 315                 if (last)
 316                         last->next = dst;
 317                 else
 318                         dst_busy_list = dst;
 319                 for (; dst; dst = dst->next) {
 320                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 321                 }
 322                 mutex_unlock(&dst_gc_mutex);
 323                 break;
 324         }
 325         return NOTIFY_DONE;
 326 }
 327
 328 static struct notifier_block dst_dev_notifier = {
 329         .notifier_call  = dst_dev_event,
 330 };
 331
 332 void __init dst_init(void)
 333 {
 334         register_netdevice_notifier(&dst_dev_notifier);
 335 }
 336
 337 EXPORT_SYMBOL(__dst_free);
 338 EXPORT_SYMBOL(dst_alloc);
 339 EXPORT_SYMBOL(dst_destroy);