Version:  2.6.34 2.6.35 2.6.36 2.6.37 2.6.38 2.6.39 3.0 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 3.10 3.11 3.12 3.13 3.14

Linux/net/core/dst.c

  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  */
 43 static 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 };
 52 static void dst_gc_task(struct work_struct *work);
 53 static void ___dst_free(struct dst_entry *dst);
 54 
 55 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
 56 
 57 static DEFINE_MUTEX(dst_gc_mutex);
 58 /*
 59  * long lived entries are maintained in this list, guarded by dst_gc_mutex
 60  */
 61 static struct dst_entry         *dst_busy_list;
 62 
 63 static 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 
 74 loop:
 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 > 0)
 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 
145 int dst_discard(struct sk_buff *skb)
146 {
147         kfree_skb(skb);
148         return 0;
149 }
150 EXPORT_SYMBOL(dst_discard);
151 
152 const 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 
161 
162 void *dst_alloc(struct dst_ops *ops, struct net_device *dev,
163                 int initial_ref, int initial_obsolete, unsigned short flags)
164 {
165         struct dst_entry *dst;
166 
167         if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
168                 if (ops->gc(ops))
169                         return NULL;
170         }
171         dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC);
172         if (!dst)
173                 return NULL;
174         dst->child = NULL;
175         dst->dev = dev;
176         if (dev)
177                 dev_hold(dev);
178         dst->ops = ops;
179         dst_init_metrics(dst, dst_default_metrics, true);
180         dst->expires = 0UL;
181         dst->path = dst;
182         dst->from = NULL;
183 #ifdef CONFIG_XFRM
184         dst->xfrm = NULL;
185 #endif
186         dst->input = dst_discard;
187         dst->output = dst_discard;
188         dst->error = 0;
189         dst->obsolete = initial_obsolete;
190         dst->header_len = 0;
191         dst->trailer_len = 0;
192 #ifdef CONFIG_IP_ROUTE_CLASSID
193         dst->tclassid = 0;
194 #endif
195         atomic_set(&dst->__refcnt, initial_ref);
196         dst->__use = 0;
197         dst->lastuse = jiffies;
198         dst->flags = flags;
199         dst->pending_confirm = 0;
200         dst->next = NULL;
201         if (!(flags & DST_NOCOUNT))
202                 dst_entries_add(ops, 1);
203         return dst;
204 }
205 EXPORT_SYMBOL(dst_alloc);
206 
207 static void ___dst_free(struct dst_entry *dst)
208 {
209         /* The first case (dev==NULL) is required, when
210            protocol module is unloaded.
211          */
212         if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
213                 dst->input = dst->output = dst_discard;
214         dst->obsolete = DST_OBSOLETE_DEAD;
215 }
216 
217 void __dst_free(struct dst_entry *dst)
218 {
219         spin_lock_bh(&dst_garbage.lock);
220         ___dst_free(dst);
221         dst->next = dst_garbage.list;
222         dst_garbage.list = dst;
223         if (dst_garbage.timer_inc > DST_GC_INC) {
224                 dst_garbage.timer_inc = DST_GC_INC;
225                 dst_garbage.timer_expires = DST_GC_MIN;
226                 mod_delayed_work(system_wq, &dst_gc_work,
227                                  dst_garbage.timer_expires);
228         }
229         spin_unlock_bh(&dst_garbage.lock);
230 }
231 EXPORT_SYMBOL(__dst_free);
232 
233 struct dst_entry *dst_destroy(struct dst_entry * dst)
234 {
235         struct dst_entry *child;
236 
237         smp_rmb();
238 
239 again:
240         child = dst->child;
241 
242         if (!(dst->flags & DST_NOCOUNT))
243                 dst_entries_add(dst->ops, -1);
244 
245         if (dst->ops->destroy)
246                 dst->ops->destroy(dst);
247         if (dst->dev)
248                 dev_put(dst->dev);
249         kmem_cache_free(dst->ops->kmem_cachep, dst);
250 
251         dst = child;
252         if (dst) {
253                 int nohash = dst->flags & DST_NOHASH;
254 
255                 if (atomic_dec_and_test(&dst->__refcnt)) {
256                         /* We were real parent of this dst, so kill child. */
257                         if (nohash)
258                                 goto again;
259                 } else {
260                         /* Child is still referenced, return it for freeing. */
261                         if (nohash)
262                                 return dst;
263                         /* Child is still in his hash table */
264                 }
265         }
266         return NULL;
267 }
268 EXPORT_SYMBOL(dst_destroy);
269 
270 void dst_release(struct dst_entry *dst)
271 {
272         if (dst) {
273                 int newrefcnt;
274 
275                 newrefcnt = atomic_dec_return(&dst->__refcnt);
276                 WARN_ON(newrefcnt < 0);
277                 if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
278                         dst = dst_destroy(dst);
279                         if (dst)
280                                 __dst_free(dst);
281                 }
282         }
283 }
284 EXPORT_SYMBOL(dst_release);
285 
286 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
287 {
288         u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
289 
290         if (p) {
291                 u32 *old_p = __DST_METRICS_PTR(old);
292                 unsigned long prev, new;
293 
294                 memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
295 
296                 new = (unsigned long) p;
297                 prev = cmpxchg(&dst->_metrics, old, new);
298 
299                 if (prev != old) {
300                         kfree(p);
301                         p = __DST_METRICS_PTR(prev);
302                         if (prev & DST_METRICS_READ_ONLY)
303                                 p = NULL;
304                 }
305         }
306         return p;
307 }
308 EXPORT_SYMBOL(dst_cow_metrics_generic);
309 
310 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
311 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
312 {
313         unsigned long prev, new;
314 
315         new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
316         prev = cmpxchg(&dst->_metrics, old, new);
317         if (prev == old)
318                 kfree(__DST_METRICS_PTR(old));
319 }
320 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
321 
322 /**
323  * __skb_dst_set_noref - sets skb dst, without a reference
324  * @skb: buffer
325  * @dst: dst entry
326  * @force: if force is set, use noref version even for DST_NOCACHE entries
327  *
328  * Sets skb dst, assuming a reference was not taken on dst
329  * skb_dst_drop() should not dst_release() this dst
330  */
331 void __skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst, bool force)
332 {
333         WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
334         /* If dst not in cache, we must take a reference, because
335          * dst_release() will destroy dst as soon as its refcount becomes zero
336          */
337         if (unlikely((dst->flags & DST_NOCACHE) && !force)) {
338                 dst_hold(dst);
339                 skb_dst_set(skb, dst);
340         } else {
341                 skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
342         }
343 }
344 EXPORT_SYMBOL(__skb_dst_set_noref);
345 
346 /* Dirty hack. We did it in 2.2 (in __dst_free),
347  * we have _very_ good reasons not to repeat
348  * this mistake in 2.3, but we have no choice
349  * now. _It_ _is_ _explicit_ _deliberate_
350  * _race_ _condition_.
351  *
352  * Commented and originally written by Alexey.
353  */
354 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
355                        int unregister)
356 {
357         if (dst->ops->ifdown)
358                 dst->ops->ifdown(dst, dev, unregister);
359 
360         if (dev != dst->dev)
361                 return;
362 
363         if (!unregister) {
364                 dst->input = dst->output = dst_discard;
365         } else {
366                 dst->dev = dev_net(dst->dev)->loopback_dev;
367                 dev_hold(dst->dev);
368                 dev_put(dev);
369         }
370 }
371 
372 static int dst_dev_event(struct notifier_block *this, unsigned long event,
373                          void *ptr)
374 {
375         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
376         struct dst_entry *dst, *last = NULL;
377 
378         switch (event) {
379         case NETDEV_UNREGISTER_FINAL:
380         case NETDEV_DOWN:
381                 mutex_lock(&dst_gc_mutex);
382                 for (dst = dst_busy_list; dst; dst = dst->next) {
383                         last = dst;
384                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
385                 }
386 
387                 spin_lock_bh(&dst_garbage.lock);
388                 dst = dst_garbage.list;
389                 dst_garbage.list = NULL;
390                 spin_unlock_bh(&dst_garbage.lock);
391 
392                 if (last)
393                         last->next = dst;
394                 else
395                         dst_busy_list = dst;
396                 for (; dst; dst = dst->next)
397                         dst_ifdown(dst, dev, event != NETDEV_DOWN);
398                 mutex_unlock(&dst_gc_mutex);
399                 break;
400         }
401         return NOTIFY_DONE;
402 }
403 
404 static struct notifier_block dst_dev_notifier = {
405         .notifier_call  = dst_dev_event,
406         .priority = -10, /* must be called after other network notifiers */
407 };
408 
409 void __init dst_init(void)
410 {
411         register_netdevice_notifier(&dst_dev_notifier);
412 }
413 

This page was automatically generated by LXR 0.3.1 (source).  •  Linux is a registered trademark of Linus Torvalds  •  Contact us