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>
19 #include <net/net_namespace.h>
24 * Theory of operations:
25 * 1) We use a list, protected by a spinlock, to add
26 * new entries from both BH and non-BH context.
27 * 2) In order to keep spinlock held for a small delay,
28 * we use a second list where are stored long lived
29 * entries, that are handled by the garbage collect thread
30 * fired by a workqueue.
31 * 3) This list is guarded by a mutex,
32 * so that the gc_task and dst_dev_event() can be synchronized.
34 #if RT_CACHE_DEBUG >= 2
35 static atomic_t dst_total = ATOMIC_INIT(0);
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;
70 #if RT_CACHE_DEBUG >= 2
71 ktime_t time_start = ktime_get();
72 struct timespec elapsed;
75 mutex_lock(&dst_gc_mutex);
79 while ((dst = next) != NULL) {
81 prefetch(&next->next);
82 if (likely(atomic_read(&dst->__refcnt))) {
90 dst = dst_destroy(dst);
92 /* NOHASH and still referenced. Unless it is already
93 * on gc list, invalidate it and add to gc list.
95 * Note: this is temporary. Actually, NOHASH dst's
96 * must be obsoleted when parent is obsoleted.
97 * But we do not have state "obsoleted, but
98 * referenced by parent", so it is right.
100 if (dst->obsolete > 1)
109 spin_lock_bh(&dst_garbage.lock);
110 next = dst_garbage.list;
112 dst_garbage.list = NULL;
113 spin_unlock_bh(&dst_garbage.lock);
117 dst_busy_list = head.next;
119 dst_garbage.timer_inc = DST_GC_MAX;
122 * if we freed less than 1/10 of delayed entries,
123 * we can sleep longer.
125 if (work_performed <= delayed/10) {
126 dst_garbage.timer_expires += dst_garbage.timer_inc;
127 if (dst_garbage.timer_expires > DST_GC_MAX)
128 dst_garbage.timer_expires = DST_GC_MAX;
129 dst_garbage.timer_inc += DST_GC_INC;
131 dst_garbage.timer_inc = DST_GC_INC;
132 dst_garbage.timer_expires = DST_GC_MIN;
134 expires = dst_garbage.timer_expires;
136 * if the next desired timer is more than 4 seconds in the future
137 * then round the timer to whole seconds
140 expires = round_jiffies_relative(expires);
141 schedule_delayed_work(&dst_gc_work, expires);
144 spin_unlock_bh(&dst_garbage.lock);
145 mutex_unlock(&dst_gc_mutex);
146 #if RT_CACHE_DEBUG >= 2
147 elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
148 printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
149 " expires: %lu elapsed: %lu us\n",
150 atomic_read(&dst_total), delayed, work_performed,
152 elapsed.tv_sec * USEC_PER_SEC + elapsed.tv_nsec / NSEC_PER_USEC);
156 int dst_discard(struct sk_buff *skb)
161 EXPORT_SYMBOL(dst_discard);
163 void * dst_alloc(struct dst_ops * ops)
165 struct dst_entry * dst;
167 if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
171 dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
174 atomic_set(&dst->__refcnt, 0);
176 dst->lastuse = jiffies;
178 dst->input = dst->output = dst_discard;
179 #if RT_CACHE_DEBUG >= 2
180 atomic_inc(&dst_total);
182 atomic_inc(&ops->entries);
186 static void ___dst_free(struct dst_entry * dst)
188 /* The first case (dev==NULL) is required, when
189 protocol module is unloaded.
191 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
192 dst->input = dst->output = dst_discard;
197 void __dst_free(struct dst_entry * dst)
199 spin_lock_bh(&dst_garbage.lock);
201 dst->next = dst_garbage.list;
202 dst_garbage.list = dst;
203 if (dst_garbage.timer_inc > DST_GC_INC) {
204 dst_garbage.timer_inc = DST_GC_INC;
205 dst_garbage.timer_expires = DST_GC_MIN;
206 cancel_delayed_work(&dst_gc_work);
207 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
209 spin_unlock_bh(&dst_garbage.lock);
212 struct dst_entry *dst_destroy(struct dst_entry * dst)
214 struct dst_entry *child;
215 struct neighbour *neigh;
221 neigh = dst->neighbour;
226 if (hh && atomic_dec_and_test(&hh->hh_refcnt))
230 dst->neighbour = NULL;
231 neigh_release(neigh);
234 atomic_dec(&dst->ops->entries);
236 if (dst->ops->destroy)
237 dst->ops->destroy(dst);
240 #if RT_CACHE_DEBUG >= 2
241 atomic_dec(&dst_total);
243 kmem_cache_free(dst->ops->kmem_cachep, dst);
247 int nohash = dst->flags & DST_NOHASH;
249 if (atomic_dec_and_test(&dst->__refcnt)) {
250 /* We were real parent of this dst, so kill child. */
254 /* Child is still referenced, return it for freeing. */
257 /* Child is still in his hash table */
263 void dst_release(struct dst_entry *dst)
266 WARN_ON(atomic_read(&dst->__refcnt) < 1);
267 smp_mb__before_atomic_dec();
268 atomic_dec(&dst->__refcnt);
271 EXPORT_SYMBOL(dst_release);
273 /* Dirty hack. We did it in 2.2 (in __dst_free),
274 * we have _very_ good reasons not to repeat
275 * this mistake in 2.3, but we have no choice
276 * now. _It_ _is_ _explicit_ _deliberate_
277 * _race_ _condition_.
279 * Commented and originally written by Alexey.
281 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
284 if (dst->ops->ifdown)
285 dst->ops->ifdown(dst, dev, unregister);
291 dst->input = dst->output = dst_discard;
293 dst->dev = dev_net(dst->dev)->loopback_dev;
296 if (dst->neighbour && dst->neighbour->dev == dev) {
297 dst->neighbour->dev = dst->dev;
304 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
306 struct net_device *dev = ptr;
307 struct dst_entry *dst, *last = NULL;
310 case NETDEV_UNREGISTER:
312 mutex_lock(&dst_gc_mutex);
313 for (dst = dst_busy_list; dst; dst = dst->next) {
315 dst_ifdown(dst, dev, event != NETDEV_DOWN);
318 spin_lock_bh(&dst_garbage.lock);
319 dst = dst_garbage.list;
320 dst_garbage.list = NULL;
321 spin_unlock_bh(&dst_garbage.lock);
327 for (; dst; dst = dst->next) {
328 dst_ifdown(dst, dev, event != NETDEV_DOWN);
330 mutex_unlock(&dst_gc_mutex);
336 static struct notifier_block dst_dev_notifier = {
337 .notifier_call = dst_dev_event,
340 void __init dst_init(void)
342 register_netdevice_notifier(&dst_dev_notifier);
345 EXPORT_SYMBOL(__dst_free);
346 EXPORT_SYMBOL(dst_alloc);
347 EXPORT_SYMBOL(dst_destroy);