Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next-2.6
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / net / core / dst.c
blob81a4fa1c95ed32a266ef503b796ee998f4599a1d
1 /*
2 * net/core/dst.c Protocol independent destination cache.
4 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
6 */
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>
24 #include <net/dst.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.
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,
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)
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;
71 mutex_lock(&dst_gc_mutex);
72 next = dst_busy_list;
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;
85 work_performed++;
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.
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 > 1)
98 continue;
100 ___dst_free(dst);
101 dst->next = next;
102 next = dst;
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;
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 {
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;
127 } else {
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
136 if (expires > 4*HZ)
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)
147 kfree_skb(skb);
148 return 0;
150 EXPORT_SYMBOL(dst_discard);
152 const u32 dst_default_metrics[RTAX_MAX];
154 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
155 int initial_ref, int initial_obsolete, int flags)
157 struct dst_entry *dst;
159 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
160 if (ops->gc(ops))
161 return NULL;
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 dst->neighbour = NULL;
175 dst->hh = NULL;
176 #ifdef CONFIG_XFRM
177 dst->xfrm = NULL;
178 #endif
179 dst->input = dst_discard;
180 dst->output = dst_discard;
181 dst->error = 0;
182 dst->obsolete = initial_obsolete;
183 dst->header_len = 0;
184 dst->trailer_len = 0;
185 #ifdef CONFIG_IP_ROUTE_CLASSID
186 dst->tclassid = 0;
187 #endif
188 atomic_set(&dst->__refcnt, initial_ref);
189 dst->__use = 0;
190 dst->lastuse = jiffies;
191 dst->flags = flags;
192 dst->next = NULL;
193 dst_entries_add(ops, 1);
194 return dst;
196 EXPORT_SYMBOL(dst_alloc);
198 static void ___dst_free(struct dst_entry *dst)
200 /* The first case (dev==NULL) is required, when
201 protocol module is unloaded.
203 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
204 dst->input = dst->output = dst_discard;
205 dst->obsolete = 2;
208 void __dst_free(struct dst_entry *dst)
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);
220 spin_unlock_bh(&dst_garbage.lock);
222 EXPORT_SYMBOL(__dst_free);
224 struct dst_entry *dst_destroy(struct dst_entry * dst)
226 struct dst_entry *child;
227 struct neighbour *neigh;
228 struct hh_cache *hh;
230 smp_rmb();
232 again:
233 neigh = dst->neighbour;
234 hh = dst->hh;
235 child = dst->child;
237 dst->hh = NULL;
238 if (hh)
239 hh_cache_put(hh);
241 if (neigh) {
242 dst->neighbour = NULL;
243 neigh_release(neigh);
246 dst_entries_add(dst->ops, -1);
248 if (dst->ops->destroy)
249 dst->ops->destroy(dst);
250 if (dst->dev)
251 dev_put(dst->dev);
252 kmem_cache_free(dst->ops->kmem_cachep, dst);
254 dst = child;
255 if (dst) {
256 int nohash = dst->flags & DST_NOHASH;
258 if (atomic_dec_and_test(&dst->__refcnt)) {
259 /* We were real parent of this dst, so kill child. */
260 if (nohash)
261 goto again;
262 } else {
263 /* Child is still referenced, return it for freeing. */
264 if (nohash)
265 return dst;
266 /* Child is still in his hash table */
269 return NULL;
271 EXPORT_SYMBOL(dst_destroy);
273 void dst_release(struct dst_entry *dst)
275 if (dst) {
276 int newrefcnt;
278 newrefcnt = atomic_dec_return(&dst->__refcnt);
279 WARN_ON(newrefcnt < 0);
280 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
281 dst = dst_destroy(dst);
282 if (dst)
283 __dst_free(dst);
287 EXPORT_SYMBOL(dst_release);
289 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
291 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
293 if (p) {
294 u32 *old_p = __DST_METRICS_PTR(old);
295 unsigned long prev, new;
297 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
299 new = (unsigned long) p;
300 prev = cmpxchg(&dst->_metrics, old, new);
302 if (prev != old) {
303 kfree(p);
304 p = __DST_METRICS_PTR(prev);
305 if (prev & DST_METRICS_READ_ONLY)
306 p = NULL;
309 return p;
311 EXPORT_SYMBOL(dst_cow_metrics_generic);
313 /* Caller asserts that dst_metrics_read_only(dst) is false. */
314 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
316 unsigned long prev, new;
318 new = (unsigned long) dst_default_metrics;
319 prev = cmpxchg(&dst->_metrics, old, new);
320 if (prev == old)
321 kfree(__DST_METRICS_PTR(old));
323 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
326 * skb_dst_set_noref - sets skb dst, without a reference
327 * @skb: buffer
328 * @dst: dst entry
330 * Sets skb dst, assuming a reference was not taken on dst
331 * skb_dst_drop() should not dst_release() this dst
333 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
335 WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
336 /* If dst not in cache, we must take a reference, because
337 * dst_release() will destroy dst as soon as its refcount becomes zero
339 if (unlikely(dst->flags & DST_NOCACHE)) {
340 dst_hold(dst);
341 skb_dst_set(skb, dst);
342 } else {
343 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
346 EXPORT_SYMBOL(skb_dst_set_noref);
348 /* Dirty hack. We did it in 2.2 (in __dst_free),
349 * we have _very_ good reasons not to repeat
350 * this mistake in 2.3, but we have no choice
351 * now. _It_ _is_ _explicit_ _deliberate_
352 * _race_ _condition_.
354 * Commented and originally written by Alexey.
356 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
357 int unregister)
359 if (dst->ops->ifdown)
360 dst->ops->ifdown(dst, dev, unregister);
362 if (dev != dst->dev)
363 return;
365 if (!unregister) {
366 dst->input = dst->output = dst_discard;
367 } else {
368 dst->dev = dev_net(dst->dev)->loopback_dev;
369 dev_hold(dst->dev);
370 dev_put(dev);
371 if (dst->neighbour && dst->neighbour->dev == dev) {
372 dst->neighbour->dev = dst->dev;
373 dev_hold(dst->dev);
374 dev_put(dev);
379 static int dst_dev_event(struct notifier_block *this, unsigned long event,
380 void *ptr)
382 struct net_device *dev = ptr;
383 struct dst_entry *dst, *last = NULL;
385 switch (event) {
386 case NETDEV_UNREGISTER:
387 case NETDEV_DOWN:
388 mutex_lock(&dst_gc_mutex);
389 for (dst = dst_busy_list; dst; dst = dst->next) {
390 last = dst;
391 dst_ifdown(dst, dev, event != NETDEV_DOWN);
394 spin_lock_bh(&dst_garbage.lock);
395 dst = dst_garbage.list;
396 dst_garbage.list = NULL;
397 spin_unlock_bh(&dst_garbage.lock);
399 if (last)
400 last->next = dst;
401 else
402 dst_busy_list = dst;
403 for (; dst; dst = dst->next)
404 dst_ifdown(dst, dev, event != NETDEV_DOWN);
405 mutex_unlock(&dst_gc_mutex);
406 break;
408 return NOTIFY_DONE;
411 static struct notifier_block dst_dev_notifier = {
412 .notifier_call = dst_dev_event,
413 .priority = -10, /* must be called after other network notifiers */
416 void __init dst_init(void)
418 register_netdevice_notifier(&dst_dev_notifier);