blob: ee6153e2cf43b83abb51e96620d482e820560ca9 [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 */
David S. Millerf5b0a872012-07-19 12:31:33 -070097 if (dst->obsolete > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -070098 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
Eric Dumazeta37e6e32012-08-07 10:55:45 +0000152const u32 dst_default_metrics[RTAX_MAX + 1] = {
153 /* This initializer is needed to force linker to place this variable
154 * into const section. Otherwise it might end into bss section.
155 * We really want to avoid false sharing on this variable, and catch
156 * any writes on it.
157 */
158 [RTAX_MAX] = 0xdeadbeef,
159};
160
David S. Miller62fa8a82011-01-26 20:51:05 -0800161
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700162void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
David S. Miller5110effe2012-07-02 02:21:03 -0700163 int initial_ref, int initial_obsolete, unsigned short flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164{
laurent chavey598ed932010-03-29 10:41:36 +0000165 struct dst_entry *dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
Eric Dumazetfc66f952010-10-08 06:37:34 +0000167 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
Daniel Lezcano569d3642008-01-18 03:56:57 -0800168 if (ops->gc(ops))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169 return NULL;
170 }
David S. Millercf911662011-04-28 14:31:47 -0700171 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 if (!dst)
173 return NULL;
David S. Millercf911662011-04-28 14:31:47 -0700174 dst->child = NULL;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700175 dst->dev = dev;
176 if (dev)
177 dev_hold(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 dst->ops = ops;
David S. Miller62fa8a82011-01-26 20:51:05 -0800179 dst_init_metrics(dst, dst_default_metrics, true);
David S. Millercf911662011-04-28 14:31:47 -0700180 dst->expires = 0UL;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700181 dst->path = dst;
David S. Millercf911662011-04-28 14:31:47 -0700182#ifdef CONFIG_XFRM
183 dst->xfrm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#endif
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700185 dst->input = dst_discard;
186 dst->output = dst_discard;
David S. Millercf911662011-04-28 14:31:47 -0700187 dst->error = 0;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700188 dst->obsolete = initial_obsolete;
David S. Millercf911662011-04-28 14:31:47 -0700189 dst->header_len = 0;
190 dst->trailer_len = 0;
191#ifdef CONFIG_IP_ROUTE_CLASSID
192 dst->tclassid = 0;
193#endif
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700194 atomic_set(&dst->__refcnt, initial_ref);
David S. Millercf911662011-04-28 14:31:47 -0700195 dst->__use = 0;
David S. Miller5c1e6aa2011-04-28 14:13:38 -0700196 dst->lastuse = jiffies;
197 dst->flags = flags;
David S. Miller5110effe2012-07-02 02:21:03 -0700198 dst->pending_confirm = 0;
David S. Millercf911662011-04-28 14:31:47 -0700199 dst->next = NULL;
David S. Miller957c6652011-06-24 15:25:00 -0700200 if (!(flags & DST_NOCOUNT))
201 dst_entries_add(ops, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 return dst;
203}
laurent chavey598ed932010-03-29 10:41:36 +0000204EXPORT_SYMBOL(dst_alloc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
laurent chavey598ed932010-03-29 10:41:36 +0000206static void ___dst_free(struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207{
208 /* The first case (dev==NULL) is required, when
209 protocol module is unloaded.
210 */
laurent chavey598ed932010-03-29 10:41:36 +0000211 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
Denis Chengc4b10102007-06-05 00:06:57 -0700212 dst->input = dst->output = dst_discard;
David S. Millerf5b0a872012-07-19 12:31:33 -0700213 dst->obsolete = DST_OBSOLETE_DEAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214}
215
laurent chavey598ed932010-03-29 10:41:36 +0000216void __dst_free(struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217{
Eric Dumazet86bba262007-09-12 14:29:01 +0200218 spin_lock_bh(&dst_garbage.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 ___dst_free(dst);
Eric Dumazet86bba262007-09-12 14:29:01 +0200220 dst->next = dst_garbage.list;
221 dst_garbage.list = dst;
222 if (dst_garbage.timer_inc > DST_GC_INC) {
223 dst_garbage.timer_inc = DST_GC_INC;
224 dst_garbage.timer_expires = DST_GC_MIN;
Tejun Heo41f63c52012-08-03 10:30:47 -0700225 mod_delayed_work(system_wq, &dst_gc_work,
226 dst_garbage.timer_expires);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 }
Eric Dumazet86bba262007-09-12 14:29:01 +0200228 spin_unlock_bh(&dst_garbage.lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229}
Nicolas Dichteld79d9912010-07-19 23:51:38 +0000230EXPORT_SYMBOL(__dst_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231
232struct dst_entry *dst_destroy(struct dst_entry * dst)
233{
234 struct dst_entry *child;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235
236 smp_rmb();
237
238again:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 child = dst->child;
240
David S. Miller957c6652011-06-24 15:25:00 -0700241 if (!(dst->flags & DST_NOCOUNT))
242 dst_entries_add(dst->ops, -1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243
244 if (dst->ops->destroy)
245 dst->ops->destroy(dst);
246 if (dst->dev)
247 dev_put(dst->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 kmem_cache_free(dst->ops->kmem_cachep, dst);
249
250 dst = child;
251 if (dst) {
Herbert Xu6775cab2005-04-16 15:24:10 -0700252 int nohash = dst->flags & DST_NOHASH;
253
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 if (atomic_dec_and_test(&dst->__refcnt)) {
255 /* We were real parent of this dst, so kill child. */
Herbert Xu6775cab2005-04-16 15:24:10 -0700256 if (nohash)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 goto again;
258 } else {
259 /* Child is still referenced, return it for freeing. */
Herbert Xu6775cab2005-04-16 15:24:10 -0700260 if (nohash)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 return dst;
262 /* Child is still in his hash table */
263 }
264 }
265 return NULL;
266}
laurent chavey598ed932010-03-29 10:41:36 +0000267EXPORT_SYMBOL(dst_destroy);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268
Ilpo Järvinen8d330862008-03-27 17:53:31 -0700269void dst_release(struct dst_entry *dst)
270{
271 if (dst) {
laurent chavey598ed932010-03-29 10:41:36 +0000272 int newrefcnt;
Eric Dumazetef711cf2008-11-14 00:53:54 -0800273
laurent chavey598ed932010-03-29 10:41:36 +0000274 newrefcnt = atomic_dec_return(&dst->__refcnt);
275 WARN_ON(newrefcnt < 0);
Eric Dumazet27b75c92010-10-15 05:44:11 +0000276 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
277 dst = dst_destroy(dst);
278 if (dst)
279 __dst_free(dst);
280 }
Ilpo Järvinen8d330862008-03-27 17:53:31 -0700281 }
282}
283EXPORT_SYMBOL(dst_release);
284
David S. Miller62fa8a82011-01-26 20:51:05 -0800285u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
286{
287 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
288
289 if (p) {
290 u32 *old_p = __DST_METRICS_PTR(old);
291 unsigned long prev, new;
292
293 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
294
295 new = (unsigned long) p;
296 prev = cmpxchg(&dst->_metrics, old, new);
297
298 if (prev != old) {
299 kfree(p);
300 p = __DST_METRICS_PTR(prev);
301 if (prev & DST_METRICS_READ_ONLY)
302 p = NULL;
303 }
304 }
305 return p;
306}
307EXPORT_SYMBOL(dst_cow_metrics_generic);
308
309/* Caller asserts that dst_metrics_read_only(dst) is false. */
310void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
311{
312 unsigned long prev, new;
313
Eric Dumazetb30c5162011-05-24 13:29:50 -0400314 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
David S. Miller62fa8a82011-01-26 20:51:05 -0800315 prev = cmpxchg(&dst->_metrics, old, new);
316 if (prev == old)
317 kfree(__DST_METRICS_PTR(old));
318}
319EXPORT_SYMBOL(__dst_destroy_metrics_generic);
320
Eric Dumazet27b75c92010-10-15 05:44:11 +0000321/**
322 * skb_dst_set_noref - sets skb dst, without a reference
323 * @skb: buffer
324 * @dst: dst entry
325 *
326 * Sets skb dst, assuming a reference was not taken on dst
327 * skb_dst_drop() should not dst_release() this dst
328 */
329void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
330{
331 WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
332 /* If dst not in cache, we must take a reference, because
333 * dst_release() will destroy dst as soon as its refcount becomes zero
334 */
335 if (unlikely(dst->flags & DST_NOCACHE)) {
336 dst_hold(dst);
337 skb_dst_set(skb, dst);
338 } else {
339 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
340 }
341}
342EXPORT_SYMBOL(skb_dst_set_noref);
343
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344/* Dirty hack. We did it in 2.2 (in __dst_free),
345 * we have _very_ good reasons not to repeat
346 * this mistake in 2.3, but we have no choice
347 * now. _It_ _is_ _explicit_ _deliberate_
348 * _race_ _condition_.
349 *
350 * Commented and originally written by Alexey.
351 */
stephen hemminger56115512010-04-12 07:38:05 +0000352static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
353 int unregister)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354{
355 if (dst->ops->ifdown)
356 dst->ops->ifdown(dst, dev, unregister);
357
358 if (dev != dst->dev)
359 return;
360
361 if (!unregister) {
Denis Chengc4b10102007-06-05 00:06:57 -0700362 dst->input = dst->output = dst_discard;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363 } else {
YOSHIFUJI Hideakic346dca2008-03-25 21:47:49 +0900364 dst->dev = dev_net(dst->dev)->loopback_dev;
Daniel Lezcanode3cb742007-09-25 19:16:28 -0700365 dev_hold(dst->dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 dev_put(dev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 }
368}
369
laurent chavey598ed932010-03-29 10:41:36 +0000370static int dst_dev_event(struct notifier_block *this, unsigned long event,
371 void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372{
373 struct net_device *dev = ptr;
Eric Dumazet86bba262007-09-12 14:29:01 +0200374 struct dst_entry *dst, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 switch (event) {
Eric Dumazet0115e8e2012-08-22 17:19:46 +0000377 case NETDEV_UNREGISTER_FINAL:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 case NETDEV_DOWN:
Eric Dumazet86bba262007-09-12 14:29:01 +0200379 mutex_lock(&dst_gc_mutex);
380 for (dst = dst_busy_list; dst; dst = dst->next) {
381 last = dst;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 dst_ifdown(dst, dev, event != NETDEV_DOWN);
383 }
Eric Dumazet86bba262007-09-12 14:29:01 +0200384
385 spin_lock_bh(&dst_garbage.lock);
386 dst = dst_garbage.list;
387 dst_garbage.list = NULL;
388 spin_unlock_bh(&dst_garbage.lock);
389
390 if (last)
391 last->next = dst;
392 else
393 dst_busy_list = dst;
laurent chavey598ed932010-03-29 10:41:36 +0000394 for (; dst; dst = dst->next)
Eric Dumazet86bba262007-09-12 14:29:01 +0200395 dst_ifdown(dst, dev, event != NETDEV_DOWN);
Eric Dumazet86bba262007-09-12 14:29:01 +0200396 mutex_unlock(&dst_gc_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 break;
398 }
399 return NOTIFY_DONE;
400}
401
402static struct notifier_block dst_dev_notifier = {
403 .notifier_call = dst_dev_event,
Eric Dumazet332dd962010-11-09 11:46:33 -0800404 .priority = -10, /* must be called after other network notifiers */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405};
406
407void __init dst_init(void)
408{
409 register_netdevice_notifier(&dst_dev_notifier);
410}