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/netdevice.h>
16 #include <linux/skbuff.h>
17 #include <linux/string.h>
18 #include <linux/types.h>
23 * Theory of operations:
24 * 1) We use a list, protected by a spinlock, to add
25 * new entries from both BH and non-BH context.
26 * 2) In order to keep spinlock held for a small delay,
27 * we use a second list where are stored long lived
28 * entries, that are handled by the garbage collect thread
29 * fired by a workqueue.
30 * 3) This list is guarded by a mutex,
31 * so that the gc_task and dst_dev_event() can be synchronized.
33 #if RT_CACHE_DEBUG >= 2
34 static atomic_t dst_total
= ATOMIC_INIT(0);
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
;
69 #if RT_CACHE_DEBUG >= 2
70 ktime_t time_start
= ktime_get();
71 struct timespec elapsed
;
74 mutex_lock(&dst_gc_mutex
);
78 while ((dst
= next
) != NULL
) {
80 prefetch(&next
->next
);
81 if (likely(atomic_read(&dst
->__refcnt
))) {
89 dst
= dst_destroy(dst
);
91 /* NOHASH and still referenced. Unless it is already
92 * on gc list, invalidate it and add to gc list.
94 * Note: this is temporary. Actually, NOHASH dst's
95 * must be obsoleted when parent is obsoleted.
96 * But we do not have state "obsoleted, but
97 * referenced by parent", so it is right.
99 if (dst
->obsolete
> 1)
108 spin_lock_bh(&dst_garbage
.lock
);
109 next
= dst_garbage
.list
;
111 dst_garbage
.list
= NULL
;
112 spin_unlock_bh(&dst_garbage
.lock
);
116 dst_busy_list
= head
.next
;
118 dst_garbage
.timer_inc
= DST_GC_MAX
;
121 * if we freed less than 1/10 of delayed entries,
122 * we can sleep longer.
124 if (work_performed
<= delayed
/10) {
125 dst_garbage
.timer_expires
+= dst_garbage
.timer_inc
;
126 if (dst_garbage
.timer_expires
> DST_GC_MAX
)
127 dst_garbage
.timer_expires
= DST_GC_MAX
;
128 dst_garbage
.timer_inc
+= DST_GC_INC
;
130 dst_garbage
.timer_inc
= DST_GC_INC
;
131 dst_garbage
.timer_expires
= DST_GC_MIN
;
133 expires
= dst_garbage
.timer_expires
;
135 * if the next desired timer is more than 4 seconds in the future
136 * then round the timer to whole seconds
139 expires
= round_jiffies_relative(expires
);
140 schedule_delayed_work(&dst_gc_work
, expires
);
143 spin_unlock_bh(&dst_garbage
.lock
);
144 mutex_unlock(&dst_gc_mutex
);
145 #if RT_CACHE_DEBUG >= 2
146 elapsed
= ktime_to_timespec(ktime_sub(ktime_get(), time_start
));
147 printk(KERN_DEBUG
"dst_total: %d delayed: %d work_perf: %d"
148 " expires: %lu elapsed: %lu us\n",
149 atomic_read(&dst_total
), delayed
, work_performed
,
151 elapsed
.tv_sec
* USEC_PER_SEC
+ elapsed
.tv_nsec
/ NSEC_PER_USEC
);
155 static int dst_discard(struct sk_buff
*skb
)
161 void * dst_alloc(struct dst_ops
* ops
)
163 struct dst_entry
* dst
;
165 if (ops
->gc
&& atomic_read(&ops
->entries
) > ops
->gc_thresh
) {
169 dst
= kmem_cache_zalloc(ops
->kmem_cachep
, GFP_ATOMIC
);
172 atomic_set(&dst
->__refcnt
, 0);
174 dst
->lastuse
= jiffies
;
176 dst
->input
= dst
->output
= dst_discard
;
177 #if RT_CACHE_DEBUG >= 2
178 atomic_inc(&dst_total
);
180 atomic_inc(&ops
->entries
);
184 static void ___dst_free(struct dst_entry
* dst
)
186 /* The first case (dev==NULL) is required, when
187 protocol module is unloaded.
189 if (dst
->dev
== NULL
|| !(dst
->dev
->flags
&IFF_UP
)) {
190 dst
->input
= dst
->output
= dst_discard
;
195 void __dst_free(struct dst_entry
* dst
)
197 spin_lock_bh(&dst_garbage
.lock
);
199 dst
->next
= dst_garbage
.list
;
200 dst_garbage
.list
= dst
;
201 if (dst_garbage
.timer_inc
> DST_GC_INC
) {
202 dst_garbage
.timer_inc
= DST_GC_INC
;
203 dst_garbage
.timer_expires
= DST_GC_MIN
;
204 cancel_delayed_work(&dst_gc_work
);
205 schedule_delayed_work(&dst_gc_work
, dst_garbage
.timer_expires
);
207 spin_unlock_bh(&dst_garbage
.lock
);
210 struct dst_entry
*dst_destroy(struct dst_entry
* dst
)
212 struct dst_entry
*child
;
213 struct neighbour
*neigh
;
219 neigh
= dst
->neighbour
;
224 if (hh
&& atomic_dec_and_test(&hh
->hh_refcnt
))
228 dst
->neighbour
= NULL
;
229 neigh_release(neigh
);
232 atomic_dec(&dst
->ops
->entries
);
234 if (dst
->ops
->destroy
)
235 dst
->ops
->destroy(dst
);
238 #if RT_CACHE_DEBUG >= 2
239 atomic_dec(&dst_total
);
241 kmem_cache_free(dst
->ops
->kmem_cachep
, dst
);
245 int nohash
= dst
->flags
& DST_NOHASH
;
247 if (atomic_dec_and_test(&dst
->__refcnt
)) {
248 /* We were real parent of this dst, so kill child. */
252 /* Child is still referenced, return it for freeing. */
255 /* Child is still in his hash table */
261 /* Dirty hack. We did it in 2.2 (in __dst_free),
262 * we have _very_ good reasons not to repeat
263 * this mistake in 2.3, but we have no choice
264 * now. _It_ _is_ _explicit_ _deliberate_
265 * _race_ _condition_.
267 * Commented and originally written by Alexey.
269 static inline void dst_ifdown(struct dst_entry
*dst
, struct net_device
*dev
,
272 if (dst
->ops
->ifdown
)
273 dst
->ops
->ifdown(dst
, dev
, unregister
);
279 dst
->input
= dst
->output
= dst_discard
;
281 dst
->dev
= &loopback_dev
;
282 dev_hold(&loopback_dev
);
284 if (dst
->neighbour
&& dst
->neighbour
->dev
== dev
) {
285 dst
->neighbour
->dev
= &loopback_dev
;
287 dev_hold(&loopback_dev
);
292 static int dst_dev_event(struct notifier_block
*this, unsigned long event
, void *ptr
)
294 struct net_device
*dev
= ptr
;
295 struct dst_entry
*dst
, *last
= NULL
;
298 case NETDEV_UNREGISTER
:
300 mutex_lock(&dst_gc_mutex
);
301 for (dst
= dst_busy_list
; dst
; dst
= dst
->next
) {
303 dst_ifdown(dst
, dev
, event
!= NETDEV_DOWN
);
306 spin_lock_bh(&dst_garbage
.lock
);
307 dst
= dst_garbage
.list
;
308 dst_garbage
.list
= NULL
;
309 spin_unlock_bh(&dst_garbage
.lock
);
315 for (; dst
; dst
= dst
->next
) {
316 dst_ifdown(dst
, dev
, event
!= NETDEV_DOWN
);
318 mutex_unlock(&dst_gc_mutex
);
324 static struct notifier_block dst_dev_notifier
= {
325 .notifier_call
= dst_dev_event
,
328 void __init
dst_init(void)
330 register_netdevice_notifier(&dst_dev_notifier
);
333 EXPORT_SYMBOL(__dst_free
);
334 EXPORT_SYMBOL(dst_alloc
);
335 EXPORT_SYMBOL(dst_destroy
);