Loading...
1/*
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>
12#include <linux/workqueue.h>
13#include <linux/mm.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>
23
24#include <net/dst.h>
25
26/*
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.
36 */
37
38/*
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;
45 struct dst_entry *list;
46 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);
53static void ___dst_free(struct dst_entry *dst);
54
55static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
56
57static 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)
64{
65 int delayed = 0;
66 int work_performed = 0;
67 unsigned long expires = ~0L;
68 struct dst_entry *dst, *next, head;
69 struct dst_entry *last = &head;
70
71 mutex_lock(&dst_gc_mutex);
72 next = dst_busy_list;
73
74loop:
75 while ((dst = next) != NULL) {
76 next = dst->next;
77 prefetch(&next->next);
78 cond_resched();
79 if (likely(atomic_read(&dst->__refcnt))) {
80 last->next = dst;
81 last = dst;
82 delayed++;
83 continue;
84 }
85 work_performed++;
86
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);
101 dst->next = next;
102 next = dst;
103 }
104 }
105
106 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 /*
133 * if the next desired timer is more than 4 seconds in the
134 * future then round the timer to whole seconds
135 */
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);
143}
144
145int dst_discard(struct sk_buff *skb)
146{
147 kfree_skb(skb);
148 return 0;
149}
150EXPORT_SYMBOL(dst_discard);
151
152const u32 dst_default_metrics[RTAX_MAX];
153
154void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
155 int initial_ref, int initial_obsolete, int flags)
156{
157 struct dst_entry *dst;
158
159 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
160 if (ops->gc(ops))
161 return NULL;
162 }
163 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
164 if (!dst)
165 return NULL;
166 dst->child = NULL;
167 dst->dev = dev;
168 if (dev)
169 dev_hold(dev);
170 dst->ops = ops;
171 dst_init_metrics(dst, dst_default_metrics, true);
172 dst->expires = 0UL;
173 dst->path = dst;
174 dst->_neighbour = NULL;
175#ifdef CONFIG_XFRM
176 dst->xfrm = NULL;
177#endif
178 dst->input = dst_discard;
179 dst->output = dst_discard;
180 dst->error = 0;
181 dst->obsolete = initial_obsolete;
182 dst->header_len = 0;
183 dst->trailer_len = 0;
184#ifdef CONFIG_IP_ROUTE_CLASSID
185 dst->tclassid = 0;
186#endif
187 atomic_set(&dst->__refcnt, initial_ref);
188 dst->__use = 0;
189 dst->lastuse = jiffies;
190 dst->flags = flags;
191 dst->next = NULL;
192 if (!(flags & DST_NOCOUNT))
193 dst_entries_add(ops, 1);
194 return dst;
195}
196EXPORT_SYMBOL(dst_alloc);
197
198static void ___dst_free(struct dst_entry *dst)
199{
200 /* The first case (dev==NULL) is required, when
201 protocol module is unloaded.
202 */
203 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
204 dst->input = dst->output = dst_discard;
205 dst->obsolete = 2;
206}
207
208void __dst_free(struct dst_entry *dst)
209{
210 spin_lock_bh(&dst_garbage.lock);
211 ___dst_free(dst);
212 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;
217 cancel_delayed_work(&dst_gc_work);
218 schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
219 }
220 spin_unlock_bh(&dst_garbage.lock);
221}
222EXPORT_SYMBOL(__dst_free);
223
224struct dst_entry *dst_destroy(struct dst_entry * dst)
225{
226 struct dst_entry *child;
227 struct neighbour *neigh;
228
229 smp_rmb();
230
231again:
232 neigh = dst->_neighbour;
233 child = dst->child;
234
235 if (neigh) {
236 dst->_neighbour = NULL;
237 neigh_release(neigh);
238 }
239
240 if (!(dst->flags & DST_NOCOUNT))
241 dst_entries_add(dst->ops, -1);
242
243 if (dst->ops->destroy)
244 dst->ops->destroy(dst);
245 if (dst->dev)
246 dev_put(dst->dev);
247 kmem_cache_free(dst->ops->kmem_cachep, dst);
248
249 dst = child;
250 if (dst) {
251 int nohash = dst->flags & DST_NOHASH;
252
253 if (atomic_dec_and_test(&dst->__refcnt)) {
254 /* We were real parent of this dst, so kill child. */
255 if (nohash)
256 goto again;
257 } else {
258 /* Child is still referenced, return it for freeing. */
259 if (nohash)
260 return dst;
261 /* Child is still in his hash table */
262 }
263 }
264 return NULL;
265}
266EXPORT_SYMBOL(dst_destroy);
267
268void dst_release(struct dst_entry *dst)
269{
270 if (dst) {
271 int newrefcnt;
272
273 newrefcnt = atomic_dec_return(&dst->__refcnt);
274 WARN_ON(newrefcnt < 0);
275 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
276 dst = dst_destroy(dst);
277 if (dst)
278 __dst_free(dst);
279 }
280 }
281}
282EXPORT_SYMBOL(dst_release);
283
284u32 *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
313 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
314 prev = cmpxchg(&dst->_metrics, old, new);
315 if (prev == old)
316 kfree(__DST_METRICS_PTR(old));
317}
318EXPORT_SYMBOL(__dst_destroy_metrics_generic);
319
320/**
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
343/* 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 */
351static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
352 int unregister)
353{
354 if (dst->ops->ifdown)
355 dst->ops->ifdown(dst, dev, unregister);
356
357 if (dev != dst->dev)
358 return;
359
360 if (!unregister) {
361 dst->input = dst->output = dst_discard;
362 } else {
363 dst->dev = dev_net(dst->dev)->loopback_dev;
364 dev_hold(dst->dev);
365 dev_put(dev);
366 if (dst->_neighbour && dst->_neighbour->dev == dev) {
367 dst->_neighbour->dev = dst->dev;
368 dev_hold(dst->dev);
369 dev_put(dev);
370 }
371 }
372}
373
374static int dst_dev_event(struct notifier_block *this, unsigned long event,
375 void *ptr)
376{
377 struct net_device *dev = ptr;
378 struct dst_entry *dst, *last = NULL;
379
380 switch (event) {
381 case NETDEV_UNREGISTER:
382 case NETDEV_DOWN:
383 mutex_lock(&dst_gc_mutex);
384 for (dst = dst_busy_list; dst; dst = dst->next) {
385 last = dst;
386 dst_ifdown(dst, dev, event != NETDEV_DOWN);
387 }
388
389 spin_lock_bh(&dst_garbage.lock);
390 dst = dst_garbage.list;
391 dst_garbage.list = NULL;
392 spin_unlock_bh(&dst_garbage.lock);
393
394 if (last)
395 last->next = dst;
396 else
397 dst_busy_list = dst;
398 for (; dst; dst = dst->next)
399 dst_ifdown(dst, dev, event != NETDEV_DOWN);
400 mutex_unlock(&dst_gc_mutex);
401 break;
402 }
403 return NOTIFY_DONE;
404}
405
406static struct notifier_block dst_dev_notifier = {
407 .notifier_call = dst_dev_event,
408 .priority = -10, /* must be called after other network notifiers */
409};
410
411void __init dst_init(void)
412{
413 register_netdevice_notifier(&dst_dev_notifier);
414}
1/*
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>
12#include <linux/workqueue.h>
13#include <linux/mm.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>
23#include <net/lwtunnel.h>
24
25#include <net/dst.h>
26#include <net/dst_metadata.h>
27
28/*
29 * Theory of operations:
30 * 1) We use a list, protected by a spinlock, to add
31 * new entries from both BH and non-BH context.
32 * 2) In order to keep spinlock held for a small delay,
33 * we use a second list where are stored long lived
34 * entries, that are handled by the garbage collect thread
35 * fired by a workqueue.
36 * 3) This list is guarded by a mutex,
37 * so that the gc_task and dst_dev_event() can be synchronized.
38 */
39
40/*
41 * We want to keep lock & list close together
42 * to dirty as few cache lines as possible in __dst_free().
43 * As this is not a very strong hint, we dont force an alignment on SMP.
44 */
45static struct {
46 spinlock_t lock;
47 struct dst_entry *list;
48 unsigned long timer_inc;
49 unsigned long timer_expires;
50} dst_garbage = {
51 .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
52 .timer_inc = DST_GC_MAX,
53};
54static void dst_gc_task(struct work_struct *work);
55static void ___dst_free(struct dst_entry *dst);
56
57static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
58
59static DEFINE_MUTEX(dst_gc_mutex);
60/*
61 * long lived entries are maintained in this list, guarded by dst_gc_mutex
62 */
63static struct dst_entry *dst_busy_list;
64
65static void dst_gc_task(struct work_struct *work)
66{
67 int delayed = 0;
68 int work_performed = 0;
69 unsigned long expires = ~0L;
70 struct dst_entry *dst, *next, head;
71 struct dst_entry *last = &head;
72
73 mutex_lock(&dst_gc_mutex);
74 next = dst_busy_list;
75
76loop:
77 while ((dst = next) != NULL) {
78 next = dst->next;
79 prefetch(&next->next);
80 cond_resched();
81 if (likely(atomic_read(&dst->__refcnt))) {
82 last->next = dst;
83 last = dst;
84 delayed++;
85 continue;
86 }
87 work_performed++;
88
89 dst = dst_destroy(dst);
90 if (dst) {
91 /* NOHASH and still referenced. Unless it is already
92 * on gc list, invalidate it and add to gc list.
93 *
94 * Note: this is temporary. Actually, NOHASH dst's
95 * must be obsoleted when parent is obsoleted.
96 * But we do not have state "obsoleted, but
97 * referenced by parent", so it is right.
98 */
99 if (dst->obsolete > 0)
100 continue;
101
102 ___dst_free(dst);
103 dst->next = next;
104 next = dst;
105 }
106 }
107
108 spin_lock_bh(&dst_garbage.lock);
109 next = dst_garbage.list;
110 if (next) {
111 dst_garbage.list = NULL;
112 spin_unlock_bh(&dst_garbage.lock);
113 goto loop;
114 }
115 last->next = NULL;
116 dst_busy_list = head.next;
117 if (!dst_busy_list)
118 dst_garbage.timer_inc = DST_GC_MAX;
119 else {
120 /*
121 * if we freed less than 1/10 of delayed entries,
122 * we can sleep longer.
123 */
124 if (work_performed <= delayed/10) {
125 dst_garbage.timer_expires += dst_garbage.timer_inc;
126 if (dst_garbage.timer_expires > DST_GC_MAX)
127 dst_garbage.timer_expires = DST_GC_MAX;
128 dst_garbage.timer_inc += DST_GC_INC;
129 } else {
130 dst_garbage.timer_inc = DST_GC_INC;
131 dst_garbage.timer_expires = DST_GC_MIN;
132 }
133 expires = dst_garbage.timer_expires;
134 /*
135 * if the next desired timer is more than 4 seconds in the
136 * future then round the timer to whole seconds
137 */
138 if (expires > 4*HZ)
139 expires = round_jiffies_relative(expires);
140 schedule_delayed_work(&dst_gc_work, expires);
141 }
142
143 spin_unlock_bh(&dst_garbage.lock);
144 mutex_unlock(&dst_gc_mutex);
145}
146
147int dst_discard_out(struct net *net, struct sock *sk, struct sk_buff *skb)
148{
149 kfree_skb(skb);
150 return 0;
151}
152EXPORT_SYMBOL(dst_discard_out);
153
154const u32 dst_default_metrics[RTAX_MAX + 1] = {
155 /* This initializer is needed to force linker to place this variable
156 * into const section. Otherwise it might end into bss section.
157 * We really want to avoid false sharing on this variable, and catch
158 * any writes on it.
159 */
160 [RTAX_MAX] = 0xdeadbeef,
161};
162
163void dst_init(struct dst_entry *dst, struct dst_ops *ops,
164 struct net_device *dev, int initial_ref, int initial_obsolete,
165 unsigned short flags)
166{
167 dst->child = NULL;
168 dst->dev = dev;
169 if (dev)
170 dev_hold(dev);
171 dst->ops = ops;
172 dst_init_metrics(dst, dst_default_metrics, true);
173 dst->expires = 0UL;
174 dst->path = dst;
175 dst->from = NULL;
176#ifdef CONFIG_XFRM
177 dst->xfrm = NULL;
178#endif
179 dst->input = dst_discard;
180 dst->output = dst_discard_out;
181 dst->error = 0;
182 dst->obsolete = initial_obsolete;
183 dst->header_len = 0;
184 dst->trailer_len = 0;
185#ifdef CONFIG_IP_ROUTE_CLASSID
186 dst->tclassid = 0;
187#endif
188 dst->lwtstate = NULL;
189 atomic_set(&dst->__refcnt, initial_ref);
190 dst->__use = 0;
191 dst->lastuse = jiffies;
192 dst->flags = flags;
193 dst->pending_confirm = 0;
194 dst->next = NULL;
195 if (!(flags & DST_NOCOUNT))
196 dst_entries_add(ops, 1);
197}
198EXPORT_SYMBOL(dst_init);
199
200void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
201 int initial_ref, int initial_obsolete, unsigned short flags)
202{
203 struct dst_entry *dst;
204
205 if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
206 if (ops->gc(ops))
207 return NULL;
208 }
209
210 dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
211 if (!dst)
212 return NULL;
213
214 dst_init(dst, ops, dev, initial_ref, initial_obsolete, flags);
215
216 return dst;
217}
218EXPORT_SYMBOL(dst_alloc);
219
220static void ___dst_free(struct dst_entry *dst)
221{
222 /* The first case (dev==NULL) is required, when
223 protocol module is unloaded.
224 */
225 if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
226 dst->input = dst_discard;
227 dst->output = dst_discard_out;
228 }
229 dst->obsolete = DST_OBSOLETE_DEAD;
230}
231
232void __dst_free(struct dst_entry *dst)
233{
234 spin_lock_bh(&dst_garbage.lock);
235 ___dst_free(dst);
236 dst->next = dst_garbage.list;
237 dst_garbage.list = dst;
238 if (dst_garbage.timer_inc > DST_GC_INC) {
239 dst_garbage.timer_inc = DST_GC_INC;
240 dst_garbage.timer_expires = DST_GC_MIN;
241 mod_delayed_work(system_wq, &dst_gc_work,
242 dst_garbage.timer_expires);
243 }
244 spin_unlock_bh(&dst_garbage.lock);
245}
246EXPORT_SYMBOL(__dst_free);
247
248struct dst_entry *dst_destroy(struct dst_entry * dst)
249{
250 struct dst_entry *child;
251
252 smp_rmb();
253
254again:
255 child = dst->child;
256
257 if (!(dst->flags & DST_NOCOUNT))
258 dst_entries_add(dst->ops, -1);
259
260 if (dst->ops->destroy)
261 dst->ops->destroy(dst);
262 if (dst->dev)
263 dev_put(dst->dev);
264
265 lwtstate_put(dst->lwtstate);
266
267 if (dst->flags & DST_METADATA)
268 metadata_dst_free((struct metadata_dst *)dst);
269 else
270 kmem_cache_free(dst->ops->kmem_cachep, dst);
271
272 dst = child;
273 if (dst) {
274 int nohash = dst->flags & DST_NOHASH;
275
276 if (atomic_dec_and_test(&dst->__refcnt)) {
277 /* We were real parent of this dst, so kill child. */
278 if (nohash)
279 goto again;
280 } else {
281 /* Child is still referenced, return it for freeing. */
282 if (nohash)
283 return dst;
284 /* Child is still in his hash table */
285 }
286 }
287 return NULL;
288}
289EXPORT_SYMBOL(dst_destroy);
290
291static void dst_destroy_rcu(struct rcu_head *head)
292{
293 struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head);
294
295 dst = dst_destroy(dst);
296 if (dst)
297 __dst_free(dst);
298}
299
300void dst_release(struct dst_entry *dst)
301{
302 if (dst) {
303 int newrefcnt;
304 unsigned short nocache = dst->flags & DST_NOCACHE;
305
306 newrefcnt = atomic_dec_return(&dst->__refcnt);
307 if (unlikely(newrefcnt < 0))
308 net_warn_ratelimited("%s: dst:%p refcnt:%d\n",
309 __func__, dst, newrefcnt);
310 if (!newrefcnt && unlikely(nocache))
311 call_rcu(&dst->rcu_head, dst_destroy_rcu);
312 }
313}
314EXPORT_SYMBOL(dst_release);
315
316u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
317{
318 u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
319
320 if (p) {
321 u32 *old_p = __DST_METRICS_PTR(old);
322 unsigned long prev, new;
323
324 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
325
326 new = (unsigned long) p;
327 prev = cmpxchg(&dst->_metrics, old, new);
328
329 if (prev != old) {
330 kfree(p);
331 p = __DST_METRICS_PTR(prev);
332 if (prev & DST_METRICS_READ_ONLY)
333 p = NULL;
334 }
335 }
336 return p;
337}
338EXPORT_SYMBOL(dst_cow_metrics_generic);
339
340/* Caller asserts that dst_metrics_read_only(dst) is false. */
341void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
342{
343 unsigned long prev, new;
344
345 new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
346 prev = cmpxchg(&dst->_metrics, old, new);
347 if (prev == old)
348 kfree(__DST_METRICS_PTR(old));
349}
350EXPORT_SYMBOL(__dst_destroy_metrics_generic);
351
352static struct dst_ops md_dst_ops = {
353 .family = AF_UNSPEC,
354};
355
356static int dst_md_discard_out(struct net *net, struct sock *sk, struct sk_buff *skb)
357{
358 WARN_ONCE(1, "Attempting to call output on metadata dst\n");
359 kfree_skb(skb);
360 return 0;
361}
362
363static int dst_md_discard(struct sk_buff *skb)
364{
365 WARN_ONCE(1, "Attempting to call input on metadata dst\n");
366 kfree_skb(skb);
367 return 0;
368}
369
370static void __metadata_dst_init(struct metadata_dst *md_dst, u8 optslen)
371{
372 struct dst_entry *dst;
373
374 dst = &md_dst->dst;
375 dst_init(dst, &md_dst_ops, NULL, 1, DST_OBSOLETE_NONE,
376 DST_METADATA | DST_NOCACHE | DST_NOCOUNT);
377
378 dst->input = dst_md_discard;
379 dst->output = dst_md_discard_out;
380
381 memset(dst + 1, 0, sizeof(*md_dst) + optslen - sizeof(*dst));
382}
383
384struct metadata_dst *metadata_dst_alloc(u8 optslen, gfp_t flags)
385{
386 struct metadata_dst *md_dst;
387
388 md_dst = kmalloc(sizeof(*md_dst) + optslen, flags);
389 if (!md_dst)
390 return NULL;
391
392 __metadata_dst_init(md_dst, optslen);
393
394 return md_dst;
395}
396EXPORT_SYMBOL_GPL(metadata_dst_alloc);
397
398void metadata_dst_free(struct metadata_dst *md_dst)
399{
400#ifdef CONFIG_DST_CACHE
401 dst_cache_destroy(&md_dst->u.tun_info.dst_cache);
402#endif
403 kfree(md_dst);
404}
405
406struct metadata_dst __percpu *metadata_dst_alloc_percpu(u8 optslen, gfp_t flags)
407{
408 int cpu;
409 struct metadata_dst __percpu *md_dst;
410
411 md_dst = __alloc_percpu_gfp(sizeof(struct metadata_dst) + optslen,
412 __alignof__(struct metadata_dst), flags);
413 if (!md_dst)
414 return NULL;
415
416 for_each_possible_cpu(cpu)
417 __metadata_dst_init(per_cpu_ptr(md_dst, cpu), optslen);
418
419 return md_dst;
420}
421EXPORT_SYMBOL_GPL(metadata_dst_alloc_percpu);
422
423/* Dirty hack. We did it in 2.2 (in __dst_free),
424 * we have _very_ good reasons not to repeat
425 * this mistake in 2.3, but we have no choice
426 * now. _It_ _is_ _explicit_ _deliberate_
427 * _race_ _condition_.
428 *
429 * Commented and originally written by Alexey.
430 */
431static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
432 int unregister)
433{
434 if (dst->ops->ifdown)
435 dst->ops->ifdown(dst, dev, unregister);
436
437 if (dev != dst->dev)
438 return;
439
440 if (!unregister) {
441 dst->input = dst_discard;
442 dst->output = dst_discard_out;
443 } else {
444 dst->dev = dev_net(dst->dev)->loopback_dev;
445 dev_hold(dst->dev);
446 dev_put(dev);
447 }
448}
449
450static int dst_dev_event(struct notifier_block *this, unsigned long event,
451 void *ptr)
452{
453 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
454 struct dst_entry *dst, *last = NULL;
455
456 switch (event) {
457 case NETDEV_UNREGISTER_FINAL:
458 case NETDEV_DOWN:
459 mutex_lock(&dst_gc_mutex);
460 for (dst = dst_busy_list; dst; dst = dst->next) {
461 last = dst;
462 dst_ifdown(dst, dev, event != NETDEV_DOWN);
463 }
464
465 spin_lock_bh(&dst_garbage.lock);
466 dst = dst_garbage.list;
467 dst_garbage.list = NULL;
468 spin_unlock_bh(&dst_garbage.lock);
469
470 if (last)
471 last->next = dst;
472 else
473 dst_busy_list = dst;
474 for (; dst; dst = dst->next)
475 dst_ifdown(dst, dev, event != NETDEV_DOWN);
476 mutex_unlock(&dst_gc_mutex);
477 break;
478 }
479 return NOTIFY_DONE;
480}
481
482static struct notifier_block dst_dev_notifier = {
483 .notifier_call = dst_dev_event,
484 .priority = -10, /* must be called after other network notifiers */
485};
486
487void __init dst_subsys_init(void)
488{
489 register_netdevice_notifier(&dst_dev_notifier);
490}