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 #include <linux/prefetch.h>
  23
  24 #include <net/dst.h>
  25
  26 /*
  27  * Theory of operations:
  28  * 1) We use a list, protected by a spinlock, to add
  29  *    new entries from both BH and non-BH context.
  30  * 2) In order to keep spinlock held for a small delay,
  31  *    we use a second list where are stored long lived
  32  *    entries, that are handled by the garbage collect thread
  33  *    fired by a workqueue.
  34  * 3) This list is guarded by a mutex,
  35  *    so that the gc_task and dst_dev_event() can be synchronized.
  36  */
  37
  38 /*
  39  * We want to keep lock & list close together
  40  * to dirty as few cache lines as possible in __dst_free().
  41  * As this is not a very strong hint, we dont force an alignment on SMP.
  42  */
  43 static struct {
  44         spinlock_t              lock;
  45         struct dst_entry        *list;
  46         unsigned long           timer_inc;
  47         unsigned long           timer_expires;
  48 } dst_garbage = {
  49         .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
  50         .timer_inc = DST_GC_MAX,
  51 };
  52 static void dst_gc_task(struct work_struct *work);
  53 static void ___dst_free(struct dst_entry *dst);
  54
  55 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
  56
  57 static DEFINE_MUTEX(dst_gc_mutex);
  58 /*
  59  * long lived entries are maintained in this list, guarded by dst_gc_mutex
  60  */
  61 static struct dst_entry         *dst_busy_list;
  62
  63 static void dst_gc_task(struct work_struct *work)
  64 {
  65         int    delayed = 0;
  66         int    work_performed = 0;
  67         unsigned long expires = ~0L;
  68         struct dst_entry *dst, *next, head;
  69         struct dst_entry *last = &head;
  70
  71         mutex_lock(&dst_gc_mutex);
  72         next = dst_busy_list;
  73
  74 loop:
  75         while ((dst = next) != NULL) {
  76                 next = dst->next;
  77                 prefetch(&next->next);
  78                 cond_resched();
  79                 if (likely(atomic_read(&dst->__refcnt))) {
  80                         last->next = dst;
  81                         last = dst;
  82                         delayed++;
  83                         continue;
  84                 }
  85                 work_performed++;
  86
  87                 dst = dst_destroy(dst);
  88                 if (dst) {
  89                         /* NOHASH and still referenced. Unless it is already
  90                          * on gc list, invalidate it and add to gc list.
  91                          *
  92                          * Note: this is temporary. Actually, NOHASH dst's
  93                          * must be obsoleted when parent is obsoleted.
  94                          * But we do not have state "obsoleted, but
  95                          * referenced by parent", so it is right.
  96                          */
  97                         if (dst->obsolete > 0)
  98                                 continue;
  99
 100                         ___dst_free(dst);
 101                         dst->next = next;
 102                         next = dst;
 103                 }
 104         }
 105
 106         spin_lock_bh(&dst_garbage.lock);
 107         next = dst_garbage.list;
 108         if (next) {
 109                 dst_garbage.list = NULL;
 110                 spin_unlock_bh(&dst_garbage.lock);
 111                 goto loop;
 112         }
 113         last->next = NULL;
 114         dst_busy_list = head.next;
 115         if (!dst_busy_list)
 116                 dst_garbage.timer_inc = DST_GC_MAX;
 117         else {
 118                 /*
 119                  * if we freed less than 1/10 of delayed entries,
 120                  * we can sleep longer.
 121                  */
 122                 if (work_performed <= delayed/10) {
 123                         dst_garbage.timer_expires += dst_garbage.timer_inc;
 124                         if (dst_garbage.timer_expires > DST_GC_MAX)
 125                                 dst_garbage.timer_expires = DST_GC_MAX;
 126                         dst_garbage.timer_inc += DST_GC_INC;
 127                 } else {
 128                         dst_garbage.timer_inc = DST_GC_INC;
 129                         dst_garbage.timer_expires = DST_GC_MIN;
 130                 }
 131                 expires = dst_garbage.timer_expires;
 132                 /*
 133                  * if the next desired timer is more than 4 seconds in the
 134                  * future then round the timer to whole seconds
 135                  */
 136                 if (expires > 4*HZ)
 137                         expires = round_jiffies_relative(expires);
 138                 schedule_delayed_work(&dst_gc_work, expires);
 139         }
 140
 141         spin_unlock_bh(&dst_garbage.lock);
 142         mutex_unlock(&dst_gc_mutex);
 143 }
 144
 145 int dst_discard(struct sk_buff *skb)
 146 {
 147         kfree_skb(skb);
 148         return 0;
 149 }
 150 EXPORT_SYMBOL(dst_discard);
 151
 152 const u32 dst_default_metrics[RTAX_MAX];
 153
 154 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
 155                 int initial_ref, int initial_obsolete, unsigned short flags)
 156 {
 157         struct dst_entry *dst;
 158
 159         if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
 160                 if (ops->gc(ops))
 161                         return NULL;
 162         }
 163         dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
 164         if (!dst)
 165                 return NULL;
 166         dst->child = NULL;
 167         dst->dev = dev;
 168         if (dev)
 169                 dev_hold(dev);
 170         dst->ops = ops;
 171         dst_init_metrics(dst, dst_default_metrics, true);
 172         dst->expires = 0UL;
 173         dst->path = dst;
 174 #ifdef CONFIG_XFRM
 175         dst->xfrm = NULL;
 176 #endif
 177         dst->input = dst_discard;
 178         dst->output = dst_discard;
 179         dst->error = 0;
 180         dst->obsolete = initial_obsolete;
 181         dst->header_len = 0;
 182         dst->trailer_len = 0;
 183 #ifdef CONFIG_IP_ROUTE_CLASSID
 184         dst->tclassid = 0;
 185 #endif
 186         atomic_set(&dst->__refcnt, initial_ref);
 187         dst->__use = 0;
 188         dst->lastuse = jiffies;
 189         dst->flags = flags;
 190         dst->pending_confirm = 0;
 191         dst->next = NULL;
 192         if (!(flags & DST_NOCOUNT))
 193                 dst_entries_add(ops, 1);
 194         return dst;
 195 }
 196 EXPORT_SYMBOL(dst_alloc);
 197
 198 static void ___dst_free(struct dst_entry *dst)
 199 {
 200         /* The first case (dev==NULL) is required, when
 201            protocol module is unloaded.
 202          */
 203         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
 204                 dst->input = dst->output = dst_discard;
 205         dst->obsolete = DST_OBSOLETE_DEAD;
 206 }
 207
 208 void __dst_free(struct dst_entry *dst)
 209 {
 210         spin_lock_bh(&dst_garbage.lock);
 211         ___dst_free(dst);
 212         dst->next = dst_garbage.list;
 213         dst_garbage.list = dst;
 214         if (dst_garbage.timer_inc > DST_GC_INC) {
 215                 dst_garbage.timer_inc = DST_GC_INC;
 216                 dst_garbage.timer_expires = DST_GC_MIN;
 217                 cancel_delayed_work(&dst_gc_work);
 218                 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
 219         }
 220         spin_unlock_bh(&dst_garbage.lock);
 221 }
 222 EXPORT_SYMBOL(__dst_free);
 223
 224 struct dst_entry *dst_destroy(struct dst_entry * dst)
 225 {
 226         struct dst_entry *child;
 227
 228         smp_rmb();
 229
 230 again:
 231         child = dst->child;
 232
 233         if (!(dst->flags & DST_NOCOUNT))
 234                 dst_entries_add(dst->ops, -1);
 235
 236         if (dst->ops->destroy)
 237                 dst->ops->destroy(dst);
 238         if (dst->dev)
 239                 dev_put(dst->dev);
 240         kmem_cache_free(dst->ops->kmem_cachep, dst);
 241
 242         dst = child;
 243         if (dst) {
 244                 int nohash = dst->flags & DST_NOHASH;
 245
 246                 if (atomic_dec_and_test(&dst->__refcnt)) {
 247                         /* We were real parent of this dst, so kill child. */
 248                         if (nohash)
 249                                 goto again;
 250                 } else {
 251                         /* Child is still referenced, return it for freeing. */
 252                         if (nohash)
 253                                 return dst;
 254                         /* Child is still in his hash table */
 255                 }
 256         }
 257         return NULL;
 258 }
 259 EXPORT_SYMBOL(dst_destroy);
 260
 261 void dst_release(struct dst_entry *dst)
 262 {
 263         if (dst) {
 264                 int newrefcnt;
 265
 266                 newrefcnt = atomic_dec_return(&dst->__refcnt);
 267                 WARN_ON(newrefcnt < 0);
 268                 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
 269                         dst = dst_destroy(dst);
 270                         if (dst)
 271                                 __dst_free(dst);
 272                 }
 273         }
 274 }
 275 EXPORT_SYMBOL(dst_release);
 276
 277 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
 278 {
 279         u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
 280
 281         if (p) {
 282                 u32 *old_p = __DST_METRICS_PTR(old);
 283                 unsigned long prev, new;
 284
 285                 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
 286
 287                 new = (unsigned long) p;
 288                 prev = cmpxchg(&dst->_metrics, old, new);
 289
 290                 if (prev != old) {
 291                         kfree(p);
 292                         p = __DST_METRICS_PTR(prev);
 293                         if (prev & DST_METRICS_READ_ONLY)
 294                                 p = NULL;
 295                 }
 296         }
 297         return p;
 298 }
 299 EXPORT_SYMBOL(dst_cow_metrics_generic);
 300
 301 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
 302 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
 303 {
 304         unsigned long prev, new;
 305
 306         new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
 307         prev = cmpxchg(&dst->_metrics, old, new);
 308         if (prev == old)
 309                 kfree(__DST_METRICS_PTR(old));
 310 }
 311 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
 312
 313 /**
 314  * skb_dst_set_noref - sets skb dst, without a reference
 315  * @skb: buffer
 316  * @dst: dst entry
 317  *
 318  * Sets skb dst, assuming a reference was not taken on dst
 319  * skb_dst_drop() should not dst_release() this dst
 320  */
 321 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
 322 {
 323         WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
 324         /* If dst not in cache, we must take a reference, because
 325          * dst_release() will destroy dst as soon as its refcount becomes zero
 326          */
 327         if (unlikely(dst->flags & DST_NOCACHE)) {
 328                 dst_hold(dst);
 329                 skb_dst_set(skb, dst);
 330         } else {
 331                 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
 332         }
 333 }
 334 EXPORT_SYMBOL(skb_dst_set_noref);
 335
 336 /* Dirty hack. We did it in 2.2 (in __dst_free),
 337  * we have _very_ good reasons not to repeat
 338  * this mistake in 2.3, but we have no choice
 339  * now. _It_ _is_ _explicit_ _deliberate_
 340  * _race_ _condition_.
 341  *
 342  * Commented and originally written by Alexey.
 343  */
 344 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
 345                        int unregister)
 346 {
 347         if (dst->ops->ifdown)
 348                 dst->ops->ifdown(dst, dev, unregister);
 349
 350         if (dev != dst->dev)
 351                 return;
 352
 353         if (!unregister) {
 354                 dst->input = dst->output = dst_discard;
 355         } else {
 356                 dst->dev = dev_net(dst->dev)->loopback_dev;
 357                 dev_hold(dst->dev);
 358                 dev_put(dev);
 359         }
 360 }
 361
 362 static int dst_dev_event(struct notifier_block *this, unsigned long event,
 363                          void *ptr)
 364 {
 365         struct net_device *dev = ptr;
 366         struct dst_entry *dst, *last = NULL;
 367
 368         switch (event) {
 369         case NETDEV_UNREGISTER:
 370         case NETDEV_DOWN:
 371                 mutex_lock(&dst_gc_mutex);
 372                 for (dst = dst_busy_list; dst; dst = dst->next) {
 373                         last = dst;
 374                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 375                 }
 376
 377                 spin_lock_bh(&dst_garbage.lock);
 378                 dst = dst_garbage.list;
 379                 dst_garbage.list = NULL;
 380                 spin_unlock_bh(&dst_garbage.lock);
 381
 382                 if (last)
 383                         last->next = dst;
 384                 else
 385                         dst_busy_list = dst;
 386                 for (; dst; dst = dst->next)
 387                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
 388                 mutex_unlock(&dst_gc_mutex);
 389                 break;
 390         }
 391         return NOTIFY_DONE;
 392 }
 393
 394 static struct notifier_block dst_dev_notifier = {
 395         .notifier_call  = dst_dev_event,
 396         .priority = -10, /* must be called after other network notifiers */
 397 };
 398
 399 void __init dst_init(void)
 400 {
 401         register_netdevice_notifier(&dst_dev_notifier);
 402 }