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/slab.h>
  16 #include <linux/netdevice.h>
  17 #include <linux/skbuff.h>
  18 #include <linux/string.h>
  19 #include <linux/types.h>
  20 #include <net/net_namespace.h>
  21 #include <linux/sched.h>
  22
  23 #include <net/dst.h>
  24
  25 /*
  26  * Theory of operations:
  27  * 1) We use a list, protected by a spinlock, to add
  28  *    new entries from both BH and non-BH context.
  29  * 2) In order to keep spinlock held for a small delay,
  30  *    we use a second list where are stored long lived
  31  *    entries, that are handled by the garbage collect thread
  32  *    fired by a workqueue.
  33  * 3) This list is guarded by a mutex,
  34  *    so that the gc_task and dst_dev_event() can be synchronized.
  35  */
  36 #if RT_CACHE_DEBUG >= 2
  37 static atomic_t                  dst_total = ATOMIC_INIT(0);
  38 #endif
  39
  40 /*
  41  * We want to keep lock & list close together
  42  * to dirty as few cache lines as possible in __dst_free().
  43  * As this is not a very strong hint, we dont force an alignment on SMP.
  44  */
  45 static struct {
  46         spinlock_t              lock;
  47         struct dst_entry        *list;
  48         unsigned long           timer_inc;
  49         unsigned long           timer_expires;
  50 } dst_garbage = {
  51         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  52         .timer_inc = DST_GC_MAX,
  53 };
  54 static void dst_gc_task(struct work_struct *work);
  55 static void ___dst_free(struct dst_entry *dst);
  56
  57 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  58
  59 static DEFINE_MUTEX(dst_gc_mutex);
  60 /*
  61  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  62  */
  63 static struct dst_entry         *dst_busy_list;
  64
  65 static void dst_gc_task(struct work_struct *work)
  66 {
  67         int    delayed = 0;
  68         int    work_performed = 0;
  69         unsigned long expires = ~0L;
  70         struct dst_entry *dst, *next, head;
  71         struct dst_entry *last = &head;
  72 #if RT_CACHE_DEBUG >= 2
  73         ktime_t time_start = ktime_get();
  74         struct timespec elapsed;
  75 #endif
  76
  77         mutex_lock(&dst_gc_mutex);
  78         next = dst_busy_list;
  79
  80 loop:
  81         while ((dst = next) != NULL) {
  82                 next = dst->next;
  83                 prefetch(&next->next);
  84                 cond_resched();
  85                 if (likely(atomic_read(&dst->__refcnt))) {
  86                         last->next = dst;
  87                         last = dst;
  88                         delayed++;
  89                         continue;
  90                 }
  91                 work_performed++;
  92
  93                 dst = dst_destroy(dst);
  94                 if (dst) {
  95                         /* NOHASH and still referenced. Unless it is already
  96                          * on gc list, invalidate it and add to gc list.
  97                          *
  98                          * Note: this is temporary. Actually, NOHASH dst's
  99                          * must be obsoleted when parent is obsoleted.
 100                          * But we do not have state "obsoleted, but
 101                          * referenced by parent", so it is right.
 102                          */
 103                         if (dst->obsolete > 1)
 104                                 continue;
 105
 106                         ___dst_free(dst);
 107                         dst->next = next;
 108                         next = dst;
 109                 }
 110         }
 111
 112         spin_lock_bh(&dst_garbage.lock);
 113         next = dst_garbage.list;
 114         if (next) {
 115                 dst_garbage.list = NULL;
 116                 spin_unlock_bh(&dst_garbage.lock);
 117                 goto loop;
 118         }
 119         last->next = NULL;
 120         dst_busy_list = head.next;
 121         if (!dst_busy_list)
 122                 dst_garbage.timer_inc = DST_GC_MAX;
 123         else {
 124                 /*
 125                  * if we freed less than 1/10 of delayed entries,
 126                  * we can sleep longer.
 127                  */
 128                 if (work_performed <= delayed/10) {
 129                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 130                         if (dst_garbage.timer_expires > DST_GC_MAX)
 131                                 dst_garbage.timer_expires = DST_GC_MAX;
 132                         dst_garbage.timer_inc += DST_GC_INC;
 133                 } else {
 134                         dst_garbage.timer_inc = DST_GC_INC;
 135                         dst_garbage.timer_expires = DST_GC_MIN;
 136                 }
 137                 expires = dst_garbage.timer_expires;
 138                 /*
 139                  * if the next desired timer is more than 4 seconds in the
 140                  * future then round the timer to whole seconds
 141                  */
 142                 if (expires > 4*HZ)
 143                         expires = round_jiffies_relative(expires);
 144                 schedule_delayed_work(&dst_gc_work, expires);
 145         }
 146
 147         spin_unlock_bh(&dst_garbage.lock);
 148         mutex_unlock(&dst_gc_mutex);
 149 #if RT_CACHE_DEBUG >= 2
 150         elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
 151         printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
 152                 " expires: %lu elapsed: %lu us\n",
 153                 atomic_read(&dst_total), delayed, work_performed,
 154                 expires,
 155                 elapsed.tv_sec * USEC_PER_SEC +
 156                   elapsed.tv_nsec / NSEC_PER_USEC);
 157 #endif
 158 }
 159
 160 int dst_discard(struct sk_buff *skb)
 161 {
 162         kfree_skb(skb);
 163         return 0;
 164 }
 165 EXPORT_SYMBOL(dst_discard);
 166
 167 void *dst_alloc(struct dst_ops *ops)
 168 {
 169         struct dst_entry *dst;
 170
 171         if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
 172                 if (ops->gc(ops))
 173                         return NULL;
 174         }
 175         dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
 176         if (!dst)
 177                 return NULL;
 178         atomic_set(&dst->__refcnt, 0);
 179         dst->ops = ops;
 180         dst->lastuse = jiffies;
 181         dst->path = dst;
 182         dst->input = dst->output = dst_discard;
 183 #if RT_CACHE_DEBUG >= 2
 184         atomic_inc(&dst_total);
 185 #endif
 186         dst_entries_add(ops, 1);
 187         return dst;
 188 }
 189 EXPORT_SYMBOL(dst_alloc);
 190
 191 static void ___dst_free(struct dst_entry *dst)
 192 {
 193         /* The first case (dev==NULL) is required, when
 194            protocol module is unloaded.
 195          */
 196         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
 197                 dst->input = dst->output = dst_discard;
 198         dst->obsolete = 2;
 199 }
 200
 201 void __dst_free(struct dst_entry *dst)
 202 {
 203         spin_lock_bh(&dst_garbage.lock);
 204         ___dst_free(dst);
 205         dst->next = dst_garbage.list;
 206         dst_garbage.list = dst;
 207         if (dst_garbage.timer_inc > DST_GC_INC) {
 208                 dst_garbage.timer_inc = DST_GC_INC;
 209                 dst_garbage.timer_expires = DST_GC_MIN;
 210                 cancel_delayed_work(&dst_gc_work);
 211                 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 212         }
 213         spin_unlock_bh(&dst_garbage.lock);
 214 }
 215 EXPORT_SYMBOL(__dst_free);
 216
 217 struct dst_entry *dst_destroy(struct dst_entry * dst)
 218 {
 219         struct dst_entry *child;
 220         struct neighbour *neigh;
 221         struct hh_cache *hh;
 222
 223         smp_rmb();
 224
 225 again:
 226         neigh = dst->neighbour;
 227         hh = dst->hh;
 228         child = dst->child;
 229
 230         dst->hh = NULL;
 231         if (hh)
 232                 hh_cache_put(hh);
 233
 234         if (neigh) {
 235                 dst->neighbour = NULL;
 236                 neigh_release(neigh);
 237         }
 238
 239         dst_entries_add(dst->ops, -1);
 240
 241         if (dst->ops->destroy)
 242                 dst->ops->destroy(dst);
 243         if (dst->dev)
 244                 dev_put(dst->dev);
 245 #if RT_CACHE_DEBUG >= 2
 246         atomic_dec(&dst_total);
 247 #endif
 248         kmem_cache_free(dst->ops->kmem_cachep, dst);
 249
 250         dst = child;
 251         if (dst) {
 252                 int nohash = dst->flags & DST_NOHASH;
 253
 254                 if (atomic_dec_and_test(&dst->__refcnt)) {
 255                         /* We were real parent of this dst, so kill child. */
 256                         if (nohash)
 257                                 goto again;
 258                 } else {
 259                         /* Child is still referenced, return it for freeing. */
 260                         if (nohash)
 261                                 return dst;
 262                         /* Child is still in his hash table */
 263                 }
 264         }
 265         return NULL;
 266 }
 267 EXPORT_SYMBOL(dst_destroy);
 268
 269 void dst_release(struct dst_entry *dst)
 270 {
 271         if (dst) {
 272                 int newrefcnt;
 273
 274                 newrefcnt = atomic_dec_return(&dst->__refcnt);
 275                 WARN_ON(newrefcnt < 0);
 276                 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
 277                         dst = dst_destroy(dst);
 278                         if (dst)
 279                                 __dst_free(dst);
 280                 }
 281         }
 282 }
 283 EXPORT_SYMBOL(dst_release);
 284
 285 /**
 286  * skb_dst_set_noref - sets skb dst, without a reference
 287  * @skb: buffer
 288  * @dst: dst entry
 289  *
 290  * Sets skb dst, assuming a reference was not taken on dst
 291  * skb_dst_drop() should not dst_release() this dst
 292  */
 293 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
 294 {
 295         WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
 296         /* If dst not in cache, we must take a reference, because
 297          * dst_release() will destroy dst as soon as its refcount becomes zero
 298          */
 299         if (unlikely(dst->flags & DST_NOCACHE)) {
 300                 dst_hold(dst);
 301                 skb_dst_set(skb, dst);
 302         } else {
 303                 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
 304         }
 305 }
 306 EXPORT_SYMBOL(skb_dst_set_noref);
 307
 308 /* Dirty hack. We did it in 2.2 (in __dst_free),
 309  * we have _very_ good reasons not to repeat
 310  * this mistake in 2.3, but we have no choice
 311  * now. _It_ _is_ _explicit_ _deliberate_
 312  * _race_ _condition_.
 313  *
 314  * Commented and originally written by Alexey.
 315  */
 316 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 317                        int unregister)
 318 {
 319         if (dst->ops->ifdown)
 320                 dst->ops->ifdown(dst, dev, unregister);
 321
 322         if (dev != dst->dev)
 323                 return;
 324
 325         if (!unregister) {
 326                 dst->input = dst->output = dst_discard;
 327         } else {
 328                 dst->dev = dev_net(dst->dev)->loopback_dev;
 329                 dev_hold(dst->dev);
 330                 dev_put(dev);
 331                 if (dst->neighbour && dst->neighbour->dev == dev) {
 332                         dst->neighbour->dev = dst->dev;
 333                         dev_hold(dst->dev);
 334                         dev_put(dev);
 335                 }
 336         }
 337 }
 338
 339 static int dst_dev_event(struct notifier_block *this, unsigned long event,
 340                          void *ptr)
 341 {
 342         struct net_device *dev = ptr;
 343         struct dst_entry *dst, *last = NULL;
 344
 345         switch (event) {
 346         case NETDEV_UNREGISTER:
 347         case NETDEV_DOWN:
 348                 mutex_lock(&dst_gc_mutex);
 349                 for (dst = dst_busy_list; dst; dst = dst->next) {
 350                         last = dst;
 351                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 352                 }
 353
 354                 spin_lock_bh(&dst_garbage.lock);
 355                 dst = dst_garbage.list;
 356                 dst_garbage.list = NULL;
 357                 spin_unlock_bh(&dst_garbage.lock);
 358
 359                 if (last)
 360                         last->next = dst;
 361                 else
 362                         dst_busy_list = dst;
 363                 for (; dst; dst = dst->next)
 364                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 365                 mutex_unlock(&dst_gc_mutex);
 366                 break;
 367         }
 368         return NOTIFY_DONE;
 369 }
 370
 371 static struct notifier_block dst_dev_notifier = {
 372         .notifier_call  = dst_dev_event,
 373         .priority = -10, /* must be called after other network notifiers */
 374 };
 375
 376 void __init dst_init(void)
 377 {
 378         register_netdevice_notifier(&dst_dev_notifier);
 379 }