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>
22 #include <linux/prefetch.h>
25 #include <net/dst_metadata.h>
28 * Theory of operations:
29 * 1) We use a list, protected by a spinlock, to add
30 * new entries from both BH and non-BH context.
31 * 2) In order to keep spinlock held for a small delay,
32 * we use a second list where are stored long lived
33 * entries, that are handled by the garbage collect thread
34 * fired by a workqueue.
35 * 3) This list is guarded by a mutex,
36 * so that the gc_task and dst_dev_event() can be synchronized.
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;
72 mutex_lock(&dst_gc_mutex);
76 while ((dst = next) != NULL) {
78 prefetch(&next->next);
80 if (likely(atomic_read(&dst->__refcnt))) {
88 dst = dst_destroy(dst);
90 /* NOHASH and still referenced. Unless it is already
91 * on gc list, invalidate it and add to gc list.
93 * Note: this is temporary. Actually, NOHASH dst's
94 * must be obsoleted when parent is obsoleted.
95 * But we do not have state "obsoleted, but
96 * referenced by parent", so it is right.
98 if (dst->obsolete > 0)
107 spin_lock_bh(&dst_garbage.lock);
108 next = dst_garbage.list;
110 dst_garbage.list = NULL;
111 spin_unlock_bh(&dst_garbage.lock);
115 dst_busy_list = head.next;
117 dst_garbage.timer_inc = DST_GC_MAX;
120 * if we freed less than 1/10 of delayed entries,
121 * we can sleep longer.
123 if (work_performed <= delayed/10) {
124 dst_garbage.timer_expires += dst_garbage.timer_inc;
125 if (dst_garbage.timer_expires > DST_GC_MAX)
126 dst_garbage.timer_expires = DST_GC_MAX;
127 dst_garbage.timer_inc += DST_GC_INC;
129 dst_garbage.timer_inc = DST_GC_INC;
130 dst_garbage.timer_expires = DST_GC_MIN;
132 expires = dst_garbage.timer_expires;
134 * if the next desired timer is more than 4 seconds in the
135 * future then round the timer to whole seconds
138 expires = round_jiffies_relative(expires);
139 schedule_delayed_work(&dst_gc_work, expires);
142 spin_unlock_bh(&dst_garbage.lock);
143 mutex_unlock(&dst_gc_mutex);
146 int dst_discard_sk(struct sock *sk, struct sk_buff *skb)
151 EXPORT_SYMBOL(dst_discard_sk);
153 const u32 dst_default_metrics[RTAX_MAX + 1] = {
154 /* This initializer is needed to force linker to place this variable
155 * into const section. Otherwise it might end into bss section.
156 * We really want to avoid false sharing on this variable, and catch
159 [RTAX_MAX] = 0xdeadbeef,
162 void dst_init(struct dst_entry *dst, struct dst_ops *ops,
163 struct net_device *dev, int initial_ref, int initial_obsolete,
164 unsigned short flags)
171 dst_init_metrics(dst, dst_default_metrics, true);
178 dst->input = dst_discard;
179 dst->output = dst_discard_sk;
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;
191 dst->pending_confirm = 0;
193 if (!(flags & DST_NOCOUNT))
194 dst_entries_add(ops, 1);
196 EXPORT_SYMBOL(dst_init);
198 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
199 int initial_ref, int initial_obsolete, unsigned short flags)
201 struct dst_entry *dst;
203 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
208 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
212 dst_init(dst, ops, dev, initial_ref, initial_obsolete, flags);
216 EXPORT_SYMBOL(dst_alloc);
218 static void ___dst_free(struct dst_entry *dst)
220 /* The first case (dev==NULL) is required, when
221 protocol module is unloaded.
223 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
224 dst->input = dst_discard;
225 dst->output = dst_discard_sk;
227 dst->obsolete = DST_OBSOLETE_DEAD;
230 void __dst_free(struct dst_entry *dst)
232 spin_lock_bh(&dst_garbage.lock);
234 dst->next = dst_garbage.list;
235 dst_garbage.list = dst;
236 if (dst_garbage.timer_inc > DST_GC_INC) {
237 dst_garbage.timer_inc = DST_GC_INC;
238 dst_garbage.timer_expires = DST_GC_MIN;
239 mod_delayed_work(system_wq, &dst_gc_work,
240 dst_garbage.timer_expires);
242 spin_unlock_bh(&dst_garbage.lock);
244 EXPORT_SYMBOL(__dst_free);
246 struct dst_entry *dst_destroy(struct dst_entry * dst)
248 struct dst_entry *child;
255 if (!(dst->flags & DST_NOCOUNT))
256 dst_entries_add(dst->ops, -1);
258 if (dst->ops->destroy)
259 dst->ops->destroy(dst);
263 if (dst->flags & DST_METADATA)
266 kmem_cache_free(dst->ops->kmem_cachep, dst);
270 int nohash = dst->flags & DST_NOHASH;
272 if (atomic_dec_and_test(&dst->__refcnt)) {
273 /* We were real parent of this dst, so kill child. */
277 /* Child is still referenced, return it for freeing. */
280 /* Child is still in his hash table */
285 EXPORT_SYMBOL(dst_destroy);
287 static void dst_destroy_rcu(struct rcu_head *head)
289 struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head);
291 dst = dst_destroy(dst);
296 void dst_release(struct dst_entry *dst)
301 newrefcnt = atomic_dec_return(&dst->__refcnt);
302 if (unlikely(newrefcnt < 0))
303 net_warn_ratelimited("%s: dst:%p refcnt:%d\n",
304 __func__, dst, newrefcnt);
305 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt)
306 call_rcu(&dst->rcu_head, dst_destroy_rcu);
309 EXPORT_SYMBOL(dst_release);
311 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
313 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
316 u32 *old_p = __DST_METRICS_PTR(old);
317 unsigned long prev, new;
319 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
321 new = (unsigned long) p;
322 prev = cmpxchg(&dst->_metrics, old, new);
326 p = __DST_METRICS_PTR(prev);
327 if (prev & DST_METRICS_READ_ONLY)
333 EXPORT_SYMBOL(dst_cow_metrics_generic);
335 /* Caller asserts that dst_metrics_read_only(dst) is false. */
336 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
338 unsigned long prev, new;
340 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
341 prev = cmpxchg(&dst->_metrics, old, new);
343 kfree(__DST_METRICS_PTR(old));
345 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
347 static struct dst_ops md_dst_ops = {
351 static int dst_md_discard_sk(struct sock *sk, struct sk_buff *skb)
353 WARN_ONCE(1, "Attempting to call output on metadata dst\n");
358 static int dst_md_discard(struct sk_buff *skb)
360 WARN_ONCE(1, "Attempting to call input on metadata dst\n");
365 static void __metadata_dst_init(struct metadata_dst *md_dst, u8 optslen)
367 struct dst_entry *dst;
370 dst_init(dst, &md_dst_ops, NULL, 1, DST_OBSOLETE_NONE,
371 DST_METADATA | DST_NOCACHE | DST_NOCOUNT);
373 dst->input = dst_md_discard;
374 dst->output = dst_md_discard_sk;
376 memset(dst + 1, 0, sizeof(*md_dst) + optslen - sizeof(*dst));
377 md_dst->opts_len = optslen;
380 struct metadata_dst *metadata_dst_alloc(u8 optslen, gfp_t flags)
382 struct metadata_dst *md_dst;
384 md_dst = kmalloc(sizeof(*md_dst) + optslen, flags);
388 __metadata_dst_init(md_dst, optslen);
392 EXPORT_SYMBOL_GPL(metadata_dst_alloc);
394 struct metadata_dst __percpu *metadata_dst_alloc_percpu(u8 optslen, gfp_t flags)
397 struct metadata_dst __percpu *md_dst;
399 md_dst = __alloc_percpu_gfp(sizeof(struct metadata_dst) + optslen,
400 __alignof__(struct metadata_dst), flags);
404 for_each_possible_cpu(cpu)
405 __metadata_dst_init(per_cpu_ptr(md_dst, cpu), optslen);
409 EXPORT_SYMBOL_GPL(metadata_dst_alloc_percpu);
411 /* Dirty hack. We did it in 2.2 (in __dst_free),
412 * we have _very_ good reasons not to repeat
413 * this mistake in 2.3, but we have no choice
414 * now. _It_ _is_ _explicit_ _deliberate_
415 * _race_ _condition_.
417 * Commented and originally written by Alexey.
419 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
422 if (dst->ops->ifdown)
423 dst->ops->ifdown(dst, dev, unregister);
429 dst->input = dst_discard;
430 dst->output = dst_discard_sk;
432 dst->dev = dev_net(dst->dev)->loopback_dev;
438 static int dst_dev_event(struct notifier_block *this, unsigned long event,
441 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
442 struct dst_entry *dst, *last = NULL;
445 case NETDEV_UNREGISTER_FINAL:
447 mutex_lock(&dst_gc_mutex);
448 for (dst = dst_busy_list; dst; dst = dst->next) {
450 dst_ifdown(dst, dev, event != NETDEV_DOWN);
453 spin_lock_bh(&dst_garbage.lock);
454 dst = dst_garbage.list;
455 dst_garbage.list = NULL;
456 spin_unlock_bh(&dst_garbage.lock);
462 for (; dst; dst = dst->next)
463 dst_ifdown(dst, dev, event != NETDEV_DOWN);
464 mutex_unlock(&dst_gc_mutex);
470 static struct notifier_block dst_dev_notifier = {
471 .notifier_call = dst_dev_event,
472 .priority = -10, /* must be called after other network notifiers */
475 void __init dst_subsys_init(void)
477 register_netdevice_notifier(&dst_dev_notifier);