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>
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.
38 * We want to keep lock & list close together
39 * to dirty as few cache lines as possible in __dst_free().
40 * As this is not a very strong hint, we dont force an alignment on SMP.
44 struct dst_entry
*list
;
45 unsigned long timer_inc
;
46 unsigned long timer_expires
;
48 .lock
= __SPIN_LOCK_UNLOCKED(dst_garbage
.lock
),
49 .timer_inc
= DST_GC_MAX
,
51 static void dst_gc_task(struct work_struct
*work
);
52 static void ___dst_free(struct dst_entry
*dst
);
54 static DECLARE_DELAYED_WORK(dst_gc_work
, dst_gc_task
);
56 static DEFINE_MUTEX(dst_gc_mutex
);
58 * long lived entries are maintained in this list, guarded by dst_gc_mutex
60 static struct dst_entry
*dst_busy_list
;
62 static void dst_gc_task(struct work_struct
*work
)
65 int work_performed
= 0;
66 unsigned long expires
= ~0L;
67 struct dst_entry
*dst
, *next
, head
;
68 struct dst_entry
*last
= &head
;
70 mutex_lock(&dst_gc_mutex
);
74 while ((dst
= next
) != NULL
) {
76 prefetch(&next
->next
);
78 if (likely(atomic_read(&dst
->__refcnt
))) {
86 dst
= dst_destroy(dst
);
88 /* NOHASH and still referenced. Unless it is already
89 * on gc list, invalidate it and add to gc list.
91 * Note: this is temporary. Actually, NOHASH dst's
92 * must be obsoleted when parent is obsoleted.
93 * But we do not have state "obsoleted, but
94 * referenced by parent", so it is right.
96 if (dst
->obsolete
> 1)
105 spin_lock_bh(&dst_garbage
.lock
);
106 next
= dst_garbage
.list
;
108 dst_garbage
.list
= NULL
;
109 spin_unlock_bh(&dst_garbage
.lock
);
113 dst_busy_list
= head
.next
;
115 dst_garbage
.timer_inc
= DST_GC_MAX
;
118 * if we freed less than 1/10 of delayed entries,
119 * we can sleep longer.
121 if (work_performed
<= delayed
/10) {
122 dst_garbage
.timer_expires
+= dst_garbage
.timer_inc
;
123 if (dst_garbage
.timer_expires
> DST_GC_MAX
)
124 dst_garbage
.timer_expires
= DST_GC_MAX
;
125 dst_garbage
.timer_inc
+= DST_GC_INC
;
127 dst_garbage
.timer_inc
= DST_GC_INC
;
128 dst_garbage
.timer_expires
= DST_GC_MIN
;
130 expires
= dst_garbage
.timer_expires
;
132 * if the next desired timer is more than 4 seconds in the
133 * future then round the timer to whole seconds
136 expires
= round_jiffies_relative(expires
);
137 schedule_delayed_work(&dst_gc_work
, expires
);
140 spin_unlock_bh(&dst_garbage
.lock
);
141 mutex_unlock(&dst_gc_mutex
);
144 int dst_discard(struct sk_buff
*skb
)
149 EXPORT_SYMBOL(dst_discard
);
151 const u32 dst_default_metrics
[RTAX_MAX
];
153 void *dst_alloc(struct dst_ops
*ops
, struct net_device
*dev
,
154 int initial_ref
, int initial_obsolete
, int flags
)
156 struct dst_entry
*dst
;
158 if (ops
->gc
&& dst_entries_get_fast(ops
) > ops
->gc_thresh
) {
162 dst
= kmem_cache_alloc(ops
->kmem_cachep
, GFP_ATOMIC
);
170 dst_init_metrics(dst
, dst_default_metrics
, true);
173 dst
->neighbour
= NULL
;
178 dst
->input
= dst_discard
;
179 dst
->output
= dst_discard
;
181 dst
->obsolete
= initial_obsolete
;
183 dst
->trailer_len
= 0;
184 #ifdef CONFIG_IP_ROUTE_CLASSID
187 atomic_set(&dst
->__refcnt
, initial_ref
);
189 dst
->lastuse
= jiffies
;
192 dst_entries_add(ops
, 1);
195 EXPORT_SYMBOL(dst_alloc
);
197 static void ___dst_free(struct dst_entry
*dst
)
199 /* The first case (dev==NULL) is required, when
200 protocol module is unloaded.
202 if (dst
->dev
== NULL
|| !(dst
->dev
->flags
&IFF_UP
))
203 dst
->input
= dst
->output
= dst_discard
;
207 void __dst_free(struct dst_entry
*dst
)
209 spin_lock_bh(&dst_garbage
.lock
);
211 dst
->next
= dst_garbage
.list
;
212 dst_garbage
.list
= dst
;
213 if (dst_garbage
.timer_inc
> DST_GC_INC
) {
214 dst_garbage
.timer_inc
= DST_GC_INC
;
215 dst_garbage
.timer_expires
= DST_GC_MIN
;
216 cancel_delayed_work(&dst_gc_work
);
217 schedule_delayed_work(&dst_gc_work
, dst_garbage
.timer_expires
);
219 spin_unlock_bh(&dst_garbage
.lock
);
221 EXPORT_SYMBOL(__dst_free
);
223 struct dst_entry
*dst_destroy(struct dst_entry
* dst
)
225 struct dst_entry
*child
;
226 struct neighbour
*neigh
;
232 neigh
= dst
->neighbour
;
241 dst
->neighbour
= NULL
;
242 neigh_release(neigh
);
245 dst_entries_add(dst
->ops
, -1);
247 if (dst
->ops
->destroy
)
248 dst
->ops
->destroy(dst
);
251 kmem_cache_free(dst
->ops
->kmem_cachep
, dst
);
255 int nohash
= dst
->flags
& DST_NOHASH
;
257 if (atomic_dec_and_test(&dst
->__refcnt
)) {
258 /* We were real parent of this dst, so kill child. */
262 /* Child is still referenced, return it for freeing. */
265 /* Child is still in his hash table */
270 EXPORT_SYMBOL(dst_destroy
);
272 void dst_release(struct dst_entry
*dst
)
277 newrefcnt
= atomic_dec_return(&dst
->__refcnt
);
278 WARN_ON(newrefcnt
< 0);
279 if (unlikely(dst
->flags
& DST_NOCACHE
) && !newrefcnt
) {
280 dst
= dst_destroy(dst
);
286 EXPORT_SYMBOL(dst_release
);
288 u32
*dst_cow_metrics_generic(struct dst_entry
*dst
, unsigned long old
)
290 u32
*p
= kmalloc(sizeof(u32
) * RTAX_MAX
, GFP_ATOMIC
);
293 u32
*old_p
= __DST_METRICS_PTR(old
);
294 unsigned long prev
, new;
296 memcpy(p
, old_p
, sizeof(u32
) * RTAX_MAX
);
298 new = (unsigned long) p
;
299 prev
= cmpxchg(&dst
->_metrics
, old
, new);
303 p
= __DST_METRICS_PTR(prev
);
304 if (prev
& DST_METRICS_READ_ONLY
)
310 EXPORT_SYMBOL(dst_cow_metrics_generic
);
312 /* Caller asserts that dst_metrics_read_only(dst) is false. */
313 void __dst_destroy_metrics_generic(struct dst_entry
*dst
, unsigned long old
)
315 unsigned long prev
, new;
317 new = (unsigned long) dst_default_metrics
;
318 prev
= cmpxchg(&dst
->_metrics
, old
, new);
320 kfree(__DST_METRICS_PTR(old
));
322 EXPORT_SYMBOL(__dst_destroy_metrics_generic
);
325 * skb_dst_set_noref - sets skb dst, without a reference
329 * Sets skb dst, assuming a reference was not taken on dst
330 * skb_dst_drop() should not dst_release() this dst
332 void skb_dst_set_noref(struct sk_buff
*skb
, struct dst_entry
*dst
)
334 WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
335 /* If dst not in cache, we must take a reference, because
336 * dst_release() will destroy dst as soon as its refcount becomes zero
338 if (unlikely(dst
->flags
& DST_NOCACHE
)) {
340 skb_dst_set(skb
, dst
);
342 skb
->_skb_refdst
= (unsigned long)dst
| SKB_DST_NOREF
;
345 EXPORT_SYMBOL(skb_dst_set_noref
);
347 /* Dirty hack. We did it in 2.2 (in __dst_free),
348 * we have _very_ good reasons not to repeat
349 * this mistake in 2.3, but we have no choice
350 * now. _It_ _is_ _explicit_ _deliberate_
351 * _race_ _condition_.
353 * Commented and originally written by Alexey.
355 static void dst_ifdown(struct dst_entry
*dst
, struct net_device
*dev
,
358 if (dst
->ops
->ifdown
)
359 dst
->ops
->ifdown(dst
, dev
, unregister
);
365 dst
->input
= dst
->output
= dst_discard
;
367 dst
->dev
= dev_net(dst
->dev
)->loopback_dev
;
370 if (dst
->neighbour
&& dst
->neighbour
->dev
== dev
) {
371 dst
->neighbour
->dev
= dst
->dev
;
378 static int dst_dev_event(struct notifier_block
*this, unsigned long event
,
381 struct net_device
*dev
= ptr
;
382 struct dst_entry
*dst
, *last
= NULL
;
385 case NETDEV_UNREGISTER
:
387 mutex_lock(&dst_gc_mutex
);
388 for (dst
= dst_busy_list
; dst
; dst
= dst
->next
) {
390 dst_ifdown(dst
, dev
, event
!= NETDEV_DOWN
);
393 spin_lock_bh(&dst_garbage
.lock
);
394 dst
= dst_garbage
.list
;
395 dst_garbage
.list
= NULL
;
396 spin_unlock_bh(&dst_garbage
.lock
);
402 for (; dst
; dst
= dst
->next
)
403 dst_ifdown(dst
, dev
, event
!= NETDEV_DOWN
);
404 mutex_unlock(&dst_gc_mutex
);
410 static struct notifier_block dst_dev_notifier
= {
411 .notifier_call
= dst_dev_event
,
412 .priority
= -10, /* must be called after other network notifiers */
415 void __init
dst_init(void)
417 register_netdevice_notifier(&dst_dev_notifier
);