Commit | Line | Data |
---|---|---|
1da177e4 LT |
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> | |
86bba269 | 12 | #include <linux/workqueue.h> |
1da177e4 LT |
13 | #include <linux/mm.h> |
14 | #include <linux/module.h> | |
15 | #include <linux/netdevice.h> | |
1da177e4 LT |
16 | #include <linux/skbuff.h> |
17 | #include <linux/string.h> | |
18 | #include <linux/types.h> | |
e9dc8653 | 19 | #include <net/net_namespace.h> |
1da177e4 LT |
20 | |
21 | #include <net/dst.h> | |
22 | ||
86bba269 ED |
23 | /* |
24 | * Theory of operations: | |
25 | * 1) We use a list, protected by a spinlock, to add | |
26 | * new entries from both BH and non-BH context. | |
27 | * 2) In order to keep spinlock held for a small delay, | |
28 | * we use a second list where are stored long lived | |
29 | * entries, that are handled by the garbage collect thread | |
30 | * fired by a workqueue. | |
31 | * 3) This list is guarded by a mutex, | |
32 | * so that the gc_task and dst_dev_event() can be synchronized. | |
1da177e4 | 33 | */ |
4ec93edb | 34 | #if RT_CACHE_DEBUG >= 2 |
1da177e4 LT |
35 | static atomic_t dst_total = ATOMIC_INIT(0); |
36 | #endif | |
1da177e4 | 37 | |
86bba269 ED |
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); | |
1da177e4 LT |
53 | static void ___dst_free(struct dst_entry * dst); |
54 | ||
86bba269 | 55 | static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task); |
1da177e4 | 56 | |
86bba269 ED |
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) | |
1da177e4 LT |
64 | { |
65 | int delayed = 0; | |
86bba269 ED |
66 | int work_performed = 0; |
67 | unsigned long expires = ~0L; | |
68 | struct dst_entry *dst, *next, head; | |
69 | struct dst_entry *last = &head; | |
70 | #if RT_CACHE_DEBUG >= 2 | |
71 | ktime_t time_start = ktime_get(); | |
72 | struct timespec elapsed; | |
73 | #endif | |
1da177e4 | 74 | |
86bba269 ED |
75 | mutex_lock(&dst_gc_mutex); |
76 | next = dst_busy_list; | |
1da177e4 | 77 | |
86bba269 ED |
78 | loop: |
79 | while ((dst = next) != NULL) { | |
80 | next = dst->next; | |
81 | prefetch(&next->next); | |
82 | if (likely(atomic_read(&dst->__refcnt))) { | |
83 | last->next = dst; | |
84 | last = dst; | |
1da177e4 LT |
85 | delayed++; |
86 | continue; | |
87 | } | |
86bba269 | 88 | work_performed++; |
1da177e4 LT |
89 | |
90 | dst = dst_destroy(dst); | |
91 | if (dst) { | |
92 | /* NOHASH and still referenced. Unless it is already | |
93 | * on gc list, invalidate it and add to gc list. | |
94 | * | |
95 | * Note: this is temporary. Actually, NOHASH dst's | |
96 | * must be obsoleted when parent is obsoleted. | |
97 | * But we do not have state "obsoleted, but | |
98 | * referenced by parent", so it is right. | |
99 | */ | |
100 | if (dst->obsolete > 1) | |
101 | continue; | |
102 | ||
103 | ___dst_free(dst); | |
86bba269 ED |
104 | dst->next = next; |
105 | next = dst; | |
1da177e4 LT |
106 | } |
107 | } | |
86bba269 ED |
108 | |
109 | spin_lock_bh(&dst_garbage.lock); | |
110 | next = dst_garbage.list; | |
111 | if (next) { | |
112 | dst_garbage.list = NULL; | |
113 | spin_unlock_bh(&dst_garbage.lock); | |
114 | goto loop; | |
1da177e4 | 115 | } |
86bba269 ED |
116 | last->next = NULL; |
117 | dst_busy_list = head.next; | |
118 | if (!dst_busy_list) | |
119 | dst_garbage.timer_inc = DST_GC_MAX; | |
120 | else { | |
121 | /* | |
122 | * if we freed less than 1/10 of delayed entries, | |
123 | * we can sleep longer. | |
124 | */ | |
125 | if (work_performed <= delayed/10) { | |
126 | dst_garbage.timer_expires += dst_garbage.timer_inc; | |
127 | if (dst_garbage.timer_expires > DST_GC_MAX) | |
128 | dst_garbage.timer_expires = DST_GC_MAX; | |
129 | dst_garbage.timer_inc += DST_GC_INC; | |
130 | } else { | |
131 | dst_garbage.timer_inc = DST_GC_INC; | |
132 | dst_garbage.timer_expires = DST_GC_MIN; | |
133 | } | |
134 | expires = dst_garbage.timer_expires; | |
135 | /* | |
136 | * if the next desired timer is more than 4 seconds in the future | |
137 | * then round the timer to whole seconds | |
138 | */ | |
139 | if (expires > 4*HZ) | |
140 | expires = round_jiffies_relative(expires); | |
141 | schedule_delayed_work(&dst_gc_work, expires); | |
f0098f78 | 142 | } |
86bba269 ED |
143 | |
144 | spin_unlock_bh(&dst_garbage.lock); | |
145 | mutex_unlock(&dst_gc_mutex); | |
1da177e4 | 146 | #if RT_CACHE_DEBUG >= 2 |
86bba269 ED |
147 | elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start)); |
148 | printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d" | |
149 | " expires: %lu elapsed: %lu us\n", | |
150 | atomic_read(&dst_total), delayed, work_performed, | |
151 | expires, | |
152 | elapsed.tv_sec * USEC_PER_SEC + elapsed.tv_nsec / NSEC_PER_USEC); | |
1da177e4 | 153 | #endif |
1da177e4 LT |
154 | } |
155 | ||
352e512c | 156 | int dst_discard(struct sk_buff *skb) |
1da177e4 LT |
157 | { |
158 | kfree_skb(skb); | |
159 | return 0; | |
160 | } | |
352e512c | 161 | EXPORT_SYMBOL(dst_discard); |
1da177e4 LT |
162 | |
163 | void * dst_alloc(struct dst_ops * ops) | |
164 | { | |
165 | struct dst_entry * dst; | |
166 | ||
167 | if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) { | |
569d3645 | 168 | if (ops->gc(ops)) |
1da177e4 LT |
169 | return NULL; |
170 | } | |
c3762229 | 171 | dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC); |
1da177e4 LT |
172 | if (!dst) |
173 | return NULL; | |
1da177e4 LT |
174 | atomic_set(&dst->__refcnt, 0); |
175 | dst->ops = ops; | |
176 | dst->lastuse = jiffies; | |
177 | dst->path = dst; | |
c4b1010f | 178 | dst->input = dst->output = dst_discard; |
4ec93edb | 179 | #if RT_CACHE_DEBUG >= 2 |
1da177e4 LT |
180 | atomic_inc(&dst_total); |
181 | #endif | |
182 | atomic_inc(&ops->entries); | |
183 | return dst; | |
184 | } | |
185 | ||
186 | static void ___dst_free(struct dst_entry * dst) | |
187 | { | |
188 | /* The first case (dev==NULL) is required, when | |
189 | protocol module is unloaded. | |
190 | */ | |
191 | if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) { | |
c4b1010f | 192 | dst->input = dst->output = dst_discard; |
1da177e4 LT |
193 | } |
194 | dst->obsolete = 2; | |
195 | } | |
196 | ||
197 | void __dst_free(struct dst_entry * dst) | |
198 | { | |
86bba269 | 199 | spin_lock_bh(&dst_garbage.lock); |
1da177e4 | 200 | ___dst_free(dst); |
86bba269 ED |
201 | dst->next = dst_garbage.list; |
202 | dst_garbage.list = dst; | |
203 | if (dst_garbage.timer_inc > DST_GC_INC) { | |
204 | dst_garbage.timer_inc = DST_GC_INC; | |
205 | dst_garbage.timer_expires = DST_GC_MIN; | |
f262b59b | 206 | cancel_delayed_work(&dst_gc_work); |
86bba269 | 207 | schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires); |
1da177e4 | 208 | } |
86bba269 | 209 | spin_unlock_bh(&dst_garbage.lock); |
1da177e4 LT |
210 | } |
211 | ||
212 | struct dst_entry *dst_destroy(struct dst_entry * dst) | |
213 | { | |
214 | struct dst_entry *child; | |
215 | struct neighbour *neigh; | |
216 | struct hh_cache *hh; | |
217 | ||
218 | smp_rmb(); | |
219 | ||
220 | again: | |
221 | neigh = dst->neighbour; | |
222 | hh = dst->hh; | |
223 | child = dst->child; | |
224 | ||
225 | dst->hh = NULL; | |
226 | if (hh && atomic_dec_and_test(&hh->hh_refcnt)) | |
227 | kfree(hh); | |
228 | ||
229 | if (neigh) { | |
230 | dst->neighbour = NULL; | |
231 | neigh_release(neigh); | |
232 | } | |
233 | ||
234 | atomic_dec(&dst->ops->entries); | |
235 | ||
236 | if (dst->ops->destroy) | |
237 | dst->ops->destroy(dst); | |
238 | if (dst->dev) | |
239 | dev_put(dst->dev); | |
4ec93edb | 240 | #if RT_CACHE_DEBUG >= 2 |
1da177e4 LT |
241 | atomic_dec(&dst_total); |
242 | #endif | |
243 | kmem_cache_free(dst->ops->kmem_cachep, dst); | |
244 | ||
245 | dst = child; | |
246 | if (dst) { | |
6775cab9 HX |
247 | int nohash = dst->flags & DST_NOHASH; |
248 | ||
1da177e4 LT |
249 | if (atomic_dec_and_test(&dst->__refcnt)) { |
250 | /* We were real parent of this dst, so kill child. */ | |
6775cab9 | 251 | if (nohash) |
1da177e4 LT |
252 | goto again; |
253 | } else { | |
254 | /* Child is still referenced, return it for freeing. */ | |
6775cab9 | 255 | if (nohash) |
1da177e4 LT |
256 | return dst; |
257 | /* Child is still in his hash table */ | |
258 | } | |
259 | } | |
260 | return NULL; | |
261 | } | |
262 | ||
8d330868 IJ |
263 | void dst_release(struct dst_entry *dst) |
264 | { | |
265 | if (dst) { | |
ef711cf1 ED |
266 | int newrefcnt; |
267 | ||
8d330868 | 268 | smp_mb__before_atomic_dec(); |
ef711cf1 ED |
269 | newrefcnt = atomic_dec_return(&dst->__refcnt); |
270 | WARN_ON(newrefcnt < 0); | |
8d330868 IJ |
271 | } |
272 | } | |
273 | EXPORT_SYMBOL(dst_release); | |
274 | ||
1da177e4 LT |
275 | /* Dirty hack. We did it in 2.2 (in __dst_free), |
276 | * we have _very_ good reasons not to repeat | |
277 | * this mistake in 2.3, but we have no choice | |
278 | * now. _It_ _is_ _explicit_ _deliberate_ | |
279 | * _race_ _condition_. | |
280 | * | |
281 | * Commented and originally written by Alexey. | |
282 | */ | |
283 | static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev, | |
284 | int unregister) | |
285 | { | |
286 | if (dst->ops->ifdown) | |
287 | dst->ops->ifdown(dst, dev, unregister); | |
288 | ||
289 | if (dev != dst->dev) | |
290 | return; | |
291 | ||
292 | if (!unregister) { | |
c4b1010f | 293 | dst->input = dst->output = dst_discard; |
1da177e4 | 294 | } else { |
c346dca1 | 295 | dst->dev = dev_net(dst->dev)->loopback_dev; |
de3cb747 | 296 | dev_hold(dst->dev); |
1da177e4 LT |
297 | dev_put(dev); |
298 | if (dst->neighbour && dst->neighbour->dev == dev) { | |
5a3e55d6 | 299 | dst->neighbour->dev = dst->dev; |
64b7d961 | 300 | dev_hold(dst->dev); |
1da177e4 | 301 | dev_put(dev); |
1da177e4 LT |
302 | } |
303 | } | |
304 | } | |
305 | ||
306 | static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr) | |
307 | { | |
308 | struct net_device *dev = ptr; | |
86bba269 | 309 | struct dst_entry *dst, *last = NULL; |
1da177e4 LT |
310 | |
311 | switch (event) { | |
312 | case NETDEV_UNREGISTER: | |
313 | case NETDEV_DOWN: | |
86bba269 ED |
314 | mutex_lock(&dst_gc_mutex); |
315 | for (dst = dst_busy_list; dst; dst = dst->next) { | |
316 | last = dst; | |
317 | dst_ifdown(dst, dev, event != NETDEV_DOWN); | |
318 | } | |
319 | ||
320 | spin_lock_bh(&dst_garbage.lock); | |
321 | dst = dst_garbage.list; | |
322 | dst_garbage.list = NULL; | |
323 | spin_unlock_bh(&dst_garbage.lock); | |
324 | ||
325 | if (last) | |
326 | last->next = dst; | |
327 | else | |
328 | dst_busy_list = dst; | |
329 | for (; dst; dst = dst->next) { | |
1da177e4 LT |
330 | dst_ifdown(dst, dev, event != NETDEV_DOWN); |
331 | } | |
86bba269 | 332 | mutex_unlock(&dst_gc_mutex); |
1da177e4 LT |
333 | break; |
334 | } | |
335 | return NOTIFY_DONE; | |
336 | } | |
337 | ||
338 | static struct notifier_block dst_dev_notifier = { | |
339 | .notifier_call = dst_dev_event, | |
340 | }; | |
341 | ||
342 | void __init dst_init(void) | |
343 | { | |
344 | register_netdevice_notifier(&dst_dev_notifier); | |
345 | } | |
346 | ||
347 | EXPORT_SYMBOL(__dst_free); | |
348 | EXPORT_SYMBOL(dst_alloc); | |
349 | EXPORT_SYMBOL(dst_destroy); |