blob: 43d94cedbf7c075e5c38dd124d9e91aec8143739 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * net/core/dst.c Protocol independent destination cache.
3 *
4 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
5 *
6 */
7
8#include <linux/bitops.h>
9#include <linux/errno.h>
10#include <linux/init.h>
11#include <linux/kernel.h>
Eric Dumazet86bba262007-09-12 14:29:01 +020012#include <linux/workqueue.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070013#include <linux/mm.h>
14#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090015#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/netdevice.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/skbuff.h>
18#include <linux/string.h>
19#include <linux/types.h>
Eric W. Biedermane9dc8652007-09-12 13:02:17 +020020#include <net/net_namespace.h>
Eric Dumazet2fc1b5d2010-02-08 15:00:39 -080021#include <linux/sched.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070022#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023
24#include <net/dst.h>
25
Eric Dumazet86bba262007-09-12 14:29:01 +020026/*
27 * Theory of operations:
28 * 1) We use a list, protected by a spinlock, to add
29 * new entries from both BH and non-BH context.
30 * 2) In order to keep spinlock held for a small delay,
31 * we use a second list where are stored long lived
32 * entries, that are handled by the garbage collect thread
33 * fired by a workqueue.
34 * 3) This list is guarded by a mutex,
35 * so that the gc_task and dst_dev_event() can be synchronized.
Linus Torvalds1da177e2005-04-16 15:20:36 -070036 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070037
Eric Dumazet86bba262007-09-12 14:29:01 +020038/*
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.
42 */
43static struct {
44 spinlock_t lock;
laurent chavey598ed932010-03-29 10:41:36 +000045 struct dst_entry *list;
Eric Dumazet86bba262007-09-12 14:29:01 +020046 unsigned long timer_inc;
47 unsigned long timer_expires;
48} dst_garbage = {
49 .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
50 .timer_inc = DST_GC_MAX,
51};
52static void dst_gc_task(struct work_struct *work);
laurent chavey598ed932010-03-29 10:41:36 +000053static void ___dst_free(struct dst_entry *dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -070054
Eric Dumazet86bba262007-09-12 14:29:01 +020055static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
Eric Dumazet86bba262007-09-12 14:29:01 +020057static DEFINE_MUTEX(dst_gc_mutex);
58/*
59 * long lived entries are maintained in this list, guarded by dst_gc_mutex
60 */
61static struct dst_entry *dst_busy_list;
62
63static void dst_gc_task(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -070064{
65 int delayed = 0;
Eric Dumazet86bba262007-09-12 14:29:01 +020066 int work_performed = 0;
67 unsigned long expires = ~0L;
68 struct dst_entry *dst, *next, head;
69 struct dst_entry *last = &head;
Linus Torvalds1da177e2005-04-16 15:20:36 -070070
Eric Dumazet86bba262007-09-12 14:29:01 +020071 mutex_lock(&dst_gc_mutex);
72 next = dst_busy_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -070073
Eric Dumazet86bba262007-09-12 14:29:01 +020074loop:
75 while ((dst = next) != NULL) {
76 next = dst->next;
77 prefetch(&next->next);
Eric Dumazet2fc1b5d2010-02-08 15:00:39 -080078 cond_resched();
Eric Dumazet86bba262007-09-12 14:29:01 +020079 if (likely(atomic_read(&dst->__refcnt))) {
80 last->next = dst;
81 last = dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 delayed++;
83 continue;
84 }
Eric Dumazet86bba262007-09-12 14:29:01 +020085 work_performed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87 dst = dst_destroy(dst);
88 if (dst) {
89 /* NOHASH and still referenced. Unless it is already
90 * on gc list, invalidate it and add to gc list.
91 *
92 * Note: this is temporary. Actually, NOHASH dst's
93 * must be obsoleted when parent is obsoleted.
94 * But we do not have state "obsoleted, but
95 * referenced by parent", so it is right.
96 */
97 if (dst->obsolete > 1)
98 continue;
99
100 ___dst_free(dst);
Eric Dumazet86bba262007-09-12 14:29:01 +0200101 dst->next = next;
102 next = dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 }
104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Eric Dumazet86bba262007-09-12 14:29:01 +0200106 spin_lock_bh(&dst_garbage.lock);
107 next = dst_garbage.list;
108 if (next) {
109 dst_garbage.list = NULL;
110 spin_unlock_bh(&dst_garbage.lock);
111 goto loop;
112 }
113 last->next = NULL;
114 dst_busy_list = head.next;
115 if (!dst_busy_list)
116 dst_garbage.timer_inc = DST_GC_MAX;
117 else {
118 /*
119 * if we freed less than 1/10 of delayed entries,
120 * we can sleep longer.
121 */
122 if (work_performed <= delayed/10) {
123 dst_garbage.timer_expires += dst_garbage.timer_inc;
124 if (dst_garbage.timer_expires > DST_GC_MAX)
125 dst_garbage.timer_expires = DST_GC_MAX;
126 dst_garbage.timer_inc += DST_GC_INC;
127 } else {
128 dst_garbage.timer_inc = DST_GC_INC;
129 dst_garbage.timer_expires = DST_GC_MIN;
130 }
131 expires = dst_garbage.timer_expires;
132 /*
laurent chavey598ed932010-03-29 10:41:36 +0000133 * if the next desired timer is more than 4 seconds in the
134 * future then round the timer to whole seconds
Eric Dumazet86bba262007-09-12 14:29:01 +0200135 */
136 if (expires > 4*HZ)
137 expires = round_jiffies_relative(expires);
138 schedule_delayed_work(&dst_gc_work, expires);
139 }
140
141 spin_unlock_bh(&dst_garbage.lock);
142 mutex_unlock(&dst_gc_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143}
144
Herbert Xu352e5122007-11-13 21:34:06 -0800145int dst_discard(struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146{
147 kfree_skb(skb);
148 return 0;
149}
Herbert Xu352e5122007-11-13 21:34:06 -0800150EXPORT_SYMBOL(dst_discard);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
David S. Miller725d1e12011-01-28 14:05:05 -0800152const u32 dst_default_metrics[RTAX_MAX];
David S. Miller62fa8a82011-01-26 20:51:05 -0800153
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700154void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
155 int initial_ref, int initial_obsolete, int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156{
laurent chavey598ed932010-03-29 10:41:36 +0000157 struct dst_entry *dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158
Eric Dumazetfc66f952010-10-08 06:37:34 +0000159 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
Daniel Lezcano569d3642008-01-18 03:56:57 -0800160 if (ops->gc(ops))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 return NULL;
162 }
David S. Millercf911662011-04-28 14:31:47 -0700163 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 if (!dst)
165 return NULL;
David S. Millercf911662011-04-28 14:31:47 -0700166 dst->child = NULL;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700167 dst->dev = dev;
168 if (dev)
169 dev_hold(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 dst->ops = ops;
David S. Miller62fa8a82011-01-26 20:51:05 -0800171 dst_init_metrics(dst, dst_default_metrics, true);
David S. Millercf911662011-04-28 14:31:47 -0700172 dst->expires = 0UL;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700173 dst->path = dst;
Eric Dumazet9de79c12011-08-08 20:56:14 +0000174 RCU_INIT_POINTER(dst->_neighbour, NULL);
David S. Millercf911662011-04-28 14:31:47 -0700175#ifdef CONFIG_XFRM
176 dst->xfrm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177#endif
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700178 dst->input = dst_discard;
179 dst->output = dst_discard;
David S. Millercf911662011-04-28 14:31:47 -0700180 dst->error = 0;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700181 dst->obsolete = initial_obsolete;
David S. Millercf911662011-04-28 14:31:47 -0700182 dst->header_len = 0;
183 dst->trailer_len = 0;
184#ifdef CONFIG_IP_ROUTE_CLASSID
185 dst->tclassid = 0;
186#endif
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700187 atomic_set(&dst->__refcnt, initial_ref);
David S. Millercf911662011-04-28 14:31:47 -0700188 dst->__use = 0;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700189 dst->lastuse = jiffies;
190 dst->flags = flags;
David S. Millercf911662011-04-28 14:31:47 -0700191 dst->next = NULL;
David S. Miller957c6652011-06-24 15:25:00 -0700192 if (!(flags & DST_NOCOUNT))
193 dst_entries_add(ops, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 return dst;
195}
laurent chavey598ed932010-03-29 10:41:36 +0000196EXPORT_SYMBOL(dst_alloc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
laurent chavey598ed932010-03-29 10:41:36 +0000198static void ___dst_free(struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
200 /* The first case (dev==NULL) is required, when
201 protocol module is unloaded.
202 */
laurent chavey598ed932010-03-29 10:41:36 +0000203 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
Denis Chengc4b10102007-06-05 00:06:57 -0700204 dst->input = dst->output = dst_discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 dst->obsolete = 2;
206}
207
laurent chavey598ed932010-03-29 10:41:36 +0000208void __dst_free(struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209{
Eric Dumazet86bba262007-09-12 14:29:01 +0200210 spin_lock_bh(&dst_garbage.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 ___dst_free(dst);
Eric Dumazet86bba262007-09-12 14:29:01 +0200212 dst->next = dst_garbage.list;
213 dst_garbage.list = dst;
214 if (dst_garbage.timer_inc > DST_GC_INC) {
215 dst_garbage.timer_inc = DST_GC_INC;
216 dst_garbage.timer_expires = DST_GC_MIN;
Benjamin Theryf262b592008-09-12 16:16:37 -0700217 cancel_delayed_work(&dst_gc_work);
Eric Dumazet86bba262007-09-12 14:29:01 +0200218 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 }
Eric Dumazet86bba262007-09-12 14:29:01 +0200220 spin_unlock_bh(&dst_garbage.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221}
Nicolas Dichteld79d9912010-07-19 23:51:38 +0000222EXPORT_SYMBOL(__dst_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223
224struct dst_entry *dst_destroy(struct dst_entry * dst)
225{
226 struct dst_entry *child;
227 struct neighbour *neigh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228
229 smp_rmb();
230
231again:
Eric Dumazet9de79c12011-08-08 20:56:14 +0000232 neigh = rcu_dereference_protected(dst->_neighbour, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 child = dst->child;
234
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235 if (neigh) {
Eric Dumazet9de79c12011-08-08 20:56:14 +0000236 RCU_INIT_POINTER(dst->_neighbour, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 neigh_release(neigh);
238 }
239
David S. Miller957c6652011-06-24 15:25:00 -0700240 if (!(dst->flags & DST_NOCOUNT))
241 dst_entries_add(dst->ops, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
243 if (dst->ops->destroy)
244 dst->ops->destroy(dst);
245 if (dst->dev)
246 dev_put(dst->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 kmem_cache_free(dst->ops->kmem_cachep, dst);
248
249 dst = child;
250 if (dst) {
Herbert Xu6775cab2005-04-16 15:24:10 -0700251 int nohash = dst->flags & DST_NOHASH;
252
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 if (atomic_dec_and_test(&dst->__refcnt)) {
254 /* We were real parent of this dst, so kill child. */
Herbert Xu6775cab2005-04-16 15:24:10 -0700255 if (nohash)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 goto again;
257 } else {
258 /* Child is still referenced, return it for freeing. */
Herbert Xu6775cab2005-04-16 15:24:10 -0700259 if (nohash)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 return dst;
261 /* Child is still in his hash table */
262 }
263 }
264 return NULL;
265}
laurent chavey598ed932010-03-29 10:41:36 +0000266EXPORT_SYMBOL(dst_destroy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
Ilpo Järvinen8d330862008-03-27 17:53:31 -0700268void dst_release(struct dst_entry *dst)
269{
270 if (dst) {
laurent chavey598ed932010-03-29 10:41:36 +0000271 int newrefcnt;
Eric Dumazetef711cf2008-11-14 00:53:54 -0800272
laurent chavey598ed932010-03-29 10:41:36 +0000273 newrefcnt = atomic_dec_return(&dst->__refcnt);
274 WARN_ON(newrefcnt < 0);
Eric Dumazet27b75c92010-10-15 05:44:11 +0000275 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
276 dst = dst_destroy(dst);
277 if (dst)
278 __dst_free(dst);
279 }
Ilpo Järvinen8d330862008-03-27 17:53:31 -0700280 }
281}
282EXPORT_SYMBOL(dst_release);
283
David S. Miller62fa8a82011-01-26 20:51:05 -0800284u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
285{
286 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
287
288 if (p) {
289 u32 *old_p = __DST_METRICS_PTR(old);
290 unsigned long prev, new;
291
292 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
293
294 new = (unsigned long) p;
295 prev = cmpxchg(&dst->_metrics, old, new);
296
297 if (prev != old) {
298 kfree(p);
299 p = __DST_METRICS_PTR(prev);
300 if (prev & DST_METRICS_READ_ONLY)
301 p = NULL;
302 }
303 }
304 return p;
305}
306EXPORT_SYMBOL(dst_cow_metrics_generic);
307
308/* Caller asserts that dst_metrics_read_only(dst) is false. */
309void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
310{
311 unsigned long prev, new;
312
Eric Dumazetb30c5162011-05-24 13:29:50 -0400313 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
David S. Miller62fa8a82011-01-26 20:51:05 -0800314 prev = cmpxchg(&dst->_metrics, old, new);
315 if (prev == old)
316 kfree(__DST_METRICS_PTR(old));
317}
318EXPORT_SYMBOL(__dst_destroy_metrics_generic);
319
Eric Dumazet27b75c92010-10-15 05:44:11 +0000320/**
321 * skb_dst_set_noref - sets skb dst, without a reference
322 * @skb: buffer
323 * @dst: dst entry
324 *
325 * Sets skb dst, assuming a reference was not taken on dst
326 * skb_dst_drop() should not dst_release() this dst
327 */
328void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
329{
330 WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
331 /* If dst not in cache, we must take a reference, because
332 * dst_release() will destroy dst as soon as its refcount becomes zero
333 */
334 if (unlikely(dst->flags & DST_NOCACHE)) {
335 dst_hold(dst);
336 skb_dst_set(skb, dst);
337 } else {
338 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
339 }
340}
341EXPORT_SYMBOL(skb_dst_set_noref);
342
Linus Torvalds1da177e2005-04-16 15:20:36 -0700343/* Dirty hack. We did it in 2.2 (in __dst_free),
344 * we have _very_ good reasons not to repeat
345 * this mistake in 2.3, but we have no choice
346 * now. _It_ _is_ _explicit_ _deliberate_
347 * _race_ _condition_.
348 *
349 * Commented and originally written by Alexey.
350 */
stephen hemminger56115512010-04-12 07:38:05 +0000351static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
352 int unregister)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 if (dst->ops->ifdown)
355 dst->ops->ifdown(dst, dev, unregister);
356
357 if (dev != dst->dev)
358 return;
359
360 if (!unregister) {
Denis Chengc4b10102007-06-05 00:06:57 -0700361 dst->input = dst->output = dst_discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362 } else {
Eric Dumazet9de79c12011-08-08 20:56:14 +0000363 struct neighbour *neigh;
364
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900365 dst->dev = dev_net(dst->dev)->loopback_dev;
Daniel Lezcanode3cb742007-09-25 19:16:28 -0700366 dev_hold(dst->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 dev_put(dev);
Eric Dumazet9de79c12011-08-08 20:56:14 +0000368 rcu_read_lock();
David Miller27217452011-12-02 16:52:08 +0000369 neigh = dst_get_neighbour_noref(dst);
Eric Dumazet9de79c12011-08-08 20:56:14 +0000370 if (neigh && neigh->dev == dev) {
371 neigh->dev = dst->dev;
Eric Dumazet64b7d962007-12-11 02:00:30 -0800372 dev_hold(dst->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 dev_put(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 }
Eric Dumazet9de79c12011-08-08 20:56:14 +0000375 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 }
377}
378
laurent chavey598ed932010-03-29 10:41:36 +0000379static int dst_dev_event(struct notifier_block *this, unsigned long event,
380 void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381{
382 struct net_device *dev = ptr;
Eric Dumazet86bba262007-09-12 14:29:01 +0200383 struct dst_entry *dst, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 switch (event) {
386 case NETDEV_UNREGISTER:
387 case NETDEV_DOWN:
Eric Dumazet86bba262007-09-12 14:29:01 +0200388 mutex_lock(&dst_gc_mutex);
389 for (dst = dst_busy_list; dst; dst = dst->next) {
390 last = dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 dst_ifdown(dst, dev, event != NETDEV_DOWN);
392 }
Eric Dumazet86bba262007-09-12 14:29:01 +0200393
394 spin_lock_bh(&dst_garbage.lock);
395 dst = dst_garbage.list;
396 dst_garbage.list = NULL;
397 spin_unlock_bh(&dst_garbage.lock);
398
399 if (last)
400 last->next = dst;
401 else
402 dst_busy_list = dst;
laurent chavey598ed932010-03-29 10:41:36 +0000403 for (; dst; dst = dst->next)
Eric Dumazet86bba262007-09-12 14:29:01 +0200404 dst_ifdown(dst, dev, event != NETDEV_DOWN);
Eric Dumazet86bba262007-09-12 14:29:01 +0200405 mutex_unlock(&dst_gc_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 break;
407 }
408 return NOTIFY_DONE;
409}
410
411static struct notifier_block dst_dev_notifier = {
412 .notifier_call = dst_dev_event,
Eric Dumazet332dd962010-11-09 11:46:33 -0800413 .priority = -10, /* must be called after other network notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414};
415
416void __init dst_init(void)
417{
418 register_netdevice_notifier(&dst_dev_notifier);
419}