2 * net/core/dst.c Protocol independent destination cache.
4 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
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>
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>
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.
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.
45 struct dst_entry *list;
46 unsigned long timer_inc;
47 unsigned long timer_expires;
49 .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
50 .timer_inc = DST_GC_MAX,
52 static void dst_gc_task(struct work_struct *work);
53 static void ___dst_free(struct dst_entry *dst);
55 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
57 static DEFINE_MUTEX(dst_gc_mutex);
59 * long lived entries are maintained in this list, guarded by dst_gc_mutex
61 static struct dst_entry *dst_busy_list;
63 static void dst_gc_task(struct work_struct *work)
66 int work_performed = 0;
67 unsigned long expires = ~0L;
68 struct dst_entry *dst, *next, head;
69 struct dst_entry *last = &head;
71 mutex_lock(&dst_gc_mutex);
75 while ((dst = next) != NULL) {
77 prefetch(&next->next);
79 if (likely(atomic_read(&dst->__refcnt))) {
87 dst = dst_destroy(dst);
89 /* NOHASH and still referenced. Unless it is already
90 * on gc list, invalidate it and add to gc list.
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.
97 if (dst->obsolete > 0)
106 spin_lock_bh(&dst_garbage.lock);
107 next = dst_garbage.list;
109 dst_garbage.list = NULL;
110 spin_unlock_bh(&dst_garbage.lock);
114 dst_busy_list = head.next;
116 dst_garbage.timer_inc = DST_GC_MAX;
119 * if we freed less than 1/10 of delayed entries,
120 * we can sleep longer.
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;
128 dst_garbage.timer_inc = DST_GC_INC;
129 dst_garbage.timer_expires = DST_GC_MIN;
131 expires = dst_garbage.timer_expires;
133 * if the next desired timer is more than 4 seconds in the
134 * future then round the timer to whole seconds
137 expires = round_jiffies_relative(expires);
138 schedule_delayed_work(&dst_gc_work, expires);
141 spin_unlock_bh(&dst_garbage.lock);
142 mutex_unlock(&dst_gc_mutex);
145 int dst_discard(struct sk_buff *skb)
150 EXPORT_SYMBOL(dst_discard);
152 const u32 dst_default_metrics[RTAX_MAX + 1] = {
153 /* This initializer is needed to force linker to place this variable
154 * into const section. Otherwise it might end into bss section.
155 * We really want to avoid false sharing on this variable, and catch
158 [RTAX_MAX] = 0xdeadbeef,
162 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
163 int initial_ref, int initial_obsolete, unsigned short flags)
165 struct dst_entry *dst;
167 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
171 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
179 dst_init_metrics(dst, dst_default_metrics, true);
185 dst->input = dst_discard;
186 dst->output = dst_discard;
188 dst->obsolete = initial_obsolete;
190 dst->trailer_len = 0;
191 #ifdef CONFIG_IP_ROUTE_CLASSID
194 atomic_set(&dst->__refcnt, initial_ref);
196 dst->lastuse = jiffies;
198 dst->pending_confirm = 0;
200 if (!(flags & DST_NOCOUNT))
201 dst_entries_add(ops, 1);
204 EXPORT_SYMBOL(dst_alloc);
206 static void ___dst_free(struct dst_entry *dst)
208 /* The first case (dev==NULL) is required, when
209 protocol module is unloaded.
211 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
212 dst->input = dst->output = dst_discard;
213 dst->obsolete = DST_OBSOLETE_DEAD;
216 void __dst_free(struct dst_entry *dst)
218 spin_lock_bh(&dst_garbage.lock);
220 dst->next = dst_garbage.list;
221 dst_garbage.list = dst;
222 if (dst_garbage.timer_inc > DST_GC_INC) {
223 dst_garbage.timer_inc = DST_GC_INC;
224 dst_garbage.timer_expires = DST_GC_MIN;
225 mod_delayed_work(system_wq, &dst_gc_work,
226 dst_garbage.timer_expires);
228 spin_unlock_bh(&dst_garbage.lock);
230 EXPORT_SYMBOL(__dst_free);
232 struct dst_entry *dst_destroy(struct dst_entry * dst)
234 struct dst_entry *child;
241 if (!(dst->flags & DST_NOCOUNT))
242 dst_entries_add(dst->ops, -1);
244 if (dst->ops->destroy)
245 dst->ops->destroy(dst);
248 kmem_cache_free(dst->ops->kmem_cachep, dst);
252 int nohash = dst->flags & DST_NOHASH;
254 if (atomic_dec_and_test(&dst->__refcnt)) {
255 /* We were real parent of this dst, so kill child. */
259 /* Child is still referenced, return it for freeing. */
262 /* Child is still in his hash table */
267 EXPORT_SYMBOL(dst_destroy);
269 void dst_release(struct dst_entry *dst)
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);
283 EXPORT_SYMBOL(dst_release);
285 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
287 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
290 u32 *old_p = __DST_METRICS_PTR(old);
291 unsigned long prev, new;
293 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
295 new = (unsigned long) p;
296 prev = cmpxchg(&dst->_metrics, old, new);
300 p = __DST_METRICS_PTR(prev);
301 if (prev & DST_METRICS_READ_ONLY)
307 EXPORT_SYMBOL(dst_cow_metrics_generic);
309 /* Caller asserts that dst_metrics_read_only(dst) is false. */
310 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
312 unsigned long prev, new;
314 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
315 prev = cmpxchg(&dst->_metrics, old, new);
317 kfree(__DST_METRICS_PTR(old));
319 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
322 * skb_dst_set_noref - sets skb dst, without a reference
326 * Sets skb dst, assuming a reference was not taken on dst
327 * skb_dst_drop() should not dst_release() this dst
329 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
331 WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
332 /* If dst not in cache, we must take a reference, because
333 * dst_release() will destroy dst as soon as its refcount becomes zero
335 if (unlikely(dst->flags & DST_NOCACHE)) {
337 skb_dst_set(skb, dst);
339 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
342 EXPORT_SYMBOL(skb_dst_set_noref);
344 /* Dirty hack. We did it in 2.2 (in __dst_free),
345 * we have _very_ good reasons not to repeat
346 * this mistake in 2.3, but we have no choice
347 * now. _It_ _is_ _explicit_ _deliberate_
348 * _race_ _condition_.
350 * Commented and originally written by Alexey.
352 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
355 if (dst->ops->ifdown)
356 dst->ops->ifdown(dst, dev, unregister);
362 dst->input = dst->output = dst_discard;
364 dst->dev = dev_net(dst->dev)->loopback_dev;
370 static int dst_dev_event(struct notifier_block *this, unsigned long event,
373 struct net_device *dev = ptr;
374 struct dst_entry *dst, *last = NULL;
377 case NETDEV_UNREGISTER_FINAL:
379 mutex_lock(&dst_gc_mutex);
380 for (dst = dst_busy_list; dst; dst = dst->next) {
382 dst_ifdown(dst, dev, event != NETDEV_DOWN);
385 spin_lock_bh(&dst_garbage.lock);
386 dst = dst_garbage.list;
387 dst_garbage.list = NULL;
388 spin_unlock_bh(&dst_garbage.lock);
394 for (; dst; dst = dst->next)
395 dst_ifdown(dst, dev, event != NETDEV_DOWN);
396 mutex_unlock(&dst_gc_mutex);
402 static struct notifier_block dst_dev_notifier = {
403 .notifier_call = dst_dev_event,
404 .priority = -10, /* must be called after other network notifiers */
407 void __init dst_init(void)
409 register_netdevice_notifier(&dst_dev_notifier);