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>
21 #include <net/net_namespace.h>
25 * Theory of operations:
26 * 1) We use a list, protected by a spinlock, to add
27 * new entries from both BH and non-BH context.
28 * 2) In order to keep spinlock held for a small delay,
29 * we use a second list where are stored long lived
30 * entries, that are handled by the garbage collect thread
31 * fired by a workqueue.
32 * 3) This list is guarded by a mutex,
33 * so that the gc_task and dst_dev_event() can be synchronized.
35 #if RT_CACHE_DEBUG >= 2
36 static atomic_t dst_total = ATOMIC_INIT(0);
40 * We want to keep lock & list close together
41 * to dirty as few cache lines as possible in __dst_free().
42 * As this is not a very strong hint, we dont force an alignment on SMP.
46 struct dst_entry *list;
47 unsigned long timer_inc;
48 unsigned long timer_expires;
50 .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
51 .timer_inc = DST_GC_MAX,
53 static void dst_gc_task(struct work_struct *work);
54 static void ___dst_free(struct dst_entry * dst);
56 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
58 static DEFINE_MUTEX(dst_gc_mutex);
60 * long lived entries are maintained in this list, guarded by dst_gc_mutex
62 static struct dst_entry *dst_busy_list;
64 static void dst_gc_task(struct work_struct *work)
67 int work_performed = 0;
68 unsigned long expires = ~0L;
69 struct dst_entry *dst, *next, head;
70 struct dst_entry *last = &head;
71 #if RT_CACHE_DEBUG >= 2
72 ktime_t time_start = ktime_get();
73 struct timespec elapsed;
76 mutex_lock(&dst_gc_mutex);
80 while ((dst = next) != NULL) {
82 prefetch(&next->next);
83 if (likely(atomic_read(&dst->__refcnt))) {
91 dst = dst_destroy(dst);
93 /* NOHASH and still referenced. Unless it is already
94 * on gc list, invalidate it and add to gc list.
96 * Note: this is temporary. Actually, NOHASH dst's
97 * must be obsoleted when parent is obsoleted.
98 * But we do not have state "obsoleted, but
99 * referenced by parent", so it is right.
101 if (dst->obsolete > 1)
110 spin_lock_bh(&dst_garbage.lock);
111 next = dst_garbage.list;
113 dst_garbage.list = NULL;
114 spin_unlock_bh(&dst_garbage.lock);
118 dst_busy_list = head.next;
120 dst_garbage.timer_inc = DST_GC_MAX;
123 * if we freed less than 1/10 of delayed entries,
124 * we can sleep longer.
126 if (work_performed <= delayed/10) {
127 dst_garbage.timer_expires += dst_garbage.timer_inc;
128 if (dst_garbage.timer_expires > DST_GC_MAX)
129 dst_garbage.timer_expires = DST_GC_MAX;
130 dst_garbage.timer_inc += DST_GC_INC;
132 dst_garbage.timer_inc = DST_GC_INC;
133 dst_garbage.timer_expires = DST_GC_MIN;
135 expires = dst_garbage.timer_expires;
137 * if the next desired timer is more than 4 seconds in the future
138 * then round the timer to whole seconds
141 expires = round_jiffies_relative(expires);
142 schedule_delayed_work(&dst_gc_work, expires);
145 spin_unlock_bh(&dst_garbage.lock);
146 mutex_unlock(&dst_gc_mutex);
147 #if RT_CACHE_DEBUG >= 2
148 elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
149 printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
150 " expires: %lu elapsed: %lu us\n",
151 atomic_read(&dst_total), delayed, work_performed,
153 elapsed.tv_sec * USEC_PER_SEC + elapsed.tv_nsec / NSEC_PER_USEC);
157 static int dst_discard(struct sk_buff *skb)
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 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
208 spin_unlock_bh(&dst_garbage.lock);
211 struct dst_entry *dst_destroy(struct dst_entry * dst)
213 struct dst_entry *child;
214 struct neighbour *neigh;
220 neigh = dst->neighbour;
225 if (hh && atomic_dec_and_test(&hh->hh_refcnt))
229 dst->neighbour = NULL;
230 neigh_release(neigh);
233 atomic_dec(&dst->ops->entries);
235 if (dst->ops->destroy)
236 dst->ops->destroy(dst);
239 #if RT_CACHE_DEBUG >= 2
240 atomic_dec(&dst_total);
242 kmem_cache_free(dst->ops->kmem_cachep, dst);
246 int nohash = dst->flags & DST_NOHASH;
248 if (atomic_dec_and_test(&dst->__refcnt)) {
249 /* We were real parent of this dst, so kill child. */
253 /* Child is still referenced, return it for freeing. */
256 /* Child is still in his hash table */
262 /* Dirty hack. We did it in 2.2 (in __dst_free),
263 * we have _very_ good reasons not to repeat
264 * this mistake in 2.3, but we have no choice
265 * now. _It_ _is_ _explicit_ _deliberate_
266 * _race_ _condition_.
268 * Commented and originally written by Alexey.
270 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
273 if (dst->ops->ifdown)
274 dst->ops->ifdown(dst, dev, unregister);
280 dst->input = dst->output = dst_discard;
282 dst->dev = init_net.loopback_dev;
285 if (dst->neighbour && dst->neighbour->dev == dev) {
286 dst->neighbour->dev = init_net.loopback_dev;
288 dev_hold(dst->neighbour->dev);
293 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
295 struct net_device *dev = ptr;
296 struct dst_entry *dst, *last = NULL;
298 if (dev->nd_net != &init_net)
302 case NETDEV_UNREGISTER:
304 mutex_lock(&dst_gc_mutex);
305 for (dst = dst_busy_list; dst; dst = dst->next) {
307 dst_ifdown(dst, dev, event != NETDEV_DOWN);
310 spin_lock_bh(&dst_garbage.lock);
311 dst = dst_garbage.list;
312 dst_garbage.list = NULL;
313 spin_unlock_bh(&dst_garbage.lock);
319 for (; dst; dst = dst->next) {
320 dst_ifdown(dst, dev, event != NETDEV_DOWN);
322 mutex_unlock(&dst_gc_mutex);
328 static struct notifier_block dst_dev_notifier = {
329 .notifier_call = dst_dev_event,
332 void __init dst_init(void)
334 register_netdevice_notifier(&dst_dev_notifier);
337 EXPORT_SYMBOL(__dst_free);
338 EXPORT_SYMBOL(dst_alloc);
339 EXPORT_SYMBOL(dst_destroy);