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> | |
5a0e3ad6 | 15 | #include <linux/slab.h> |
1da177e4 | 16 | #include <linux/netdevice.h> |
1da177e4 LT |
17 | #include <linux/skbuff.h> |
18 | #include <linux/string.h> | |
19 | #include <linux/types.h> | |
e9dc8653 | 20 | #include <net/net_namespace.h> |
2fc1b5dd | 21 | #include <linux/sched.h> |
268bb0ce | 22 | #include <linux/prefetch.h> |
1da177e4 LT |
23 | |
24 | #include <net/dst.h> | |
f38a9eb1 | 25 | #include <net/dst_metadata.h> |
1da177e4 | 26 | |
86bba269 ED |
27 | /* |
28 | * Theory of operations: | |
29 | * 1) We use a list, protected by a spinlock, to add | |
30 | * new entries from both BH and non-BH context. | |
31 | * 2) In order to keep spinlock held for a small delay, | |
32 | * we use a second list where are stored long lived | |
33 | * entries, that are handled by the garbage collect thread | |
34 | * fired by a workqueue. | |
35 | * 3) This list is guarded by a mutex, | |
36 | * so that the gc_task and dst_dev_event() can be synchronized. | |
1da177e4 | 37 | */ |
1da177e4 | 38 | |
86bba269 ED |
39 | /* |
40 | * We want to keep lock & list close together | |
41 | * to dirty as few cache lines as possible in __dst_free(). | |
42 | * As this is not a very strong hint, we dont force an alignment on SMP. | |
43 | */ | |
44 | static struct { | |
45 | spinlock_t lock; | |
598ed936 | 46 | struct dst_entry *list; |
86bba269 ED |
47 | unsigned long timer_inc; |
48 | unsigned long timer_expires; | |
49 | } dst_garbage = { | |
50 | .lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock), | |
51 | .timer_inc = DST_GC_MAX, | |
52 | }; | |
53 | static void dst_gc_task(struct work_struct *work); | |
598ed936 | 54 | static void ___dst_free(struct dst_entry *dst); |
1da177e4 | 55 | |
86bba269 | 56 | static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task); |
1da177e4 | 57 | |
86bba269 ED |
58 | static DEFINE_MUTEX(dst_gc_mutex); |
59 | /* | |
60 | * long lived entries are maintained in this list, guarded by dst_gc_mutex | |
61 | */ | |
62 | static struct dst_entry *dst_busy_list; | |
63 | ||
64 | static void dst_gc_task(struct work_struct *work) | |
1da177e4 LT |
65 | { |
66 | int delayed = 0; | |
86bba269 ED |
67 | int work_performed = 0; |
68 | unsigned long expires = ~0L; | |
69 | struct dst_entry *dst, *next, head; | |
70 | struct dst_entry *last = &head; | |
1da177e4 | 71 | |
86bba269 ED |
72 | mutex_lock(&dst_gc_mutex); |
73 | next = dst_busy_list; | |
1da177e4 | 74 | |
86bba269 ED |
75 | loop: |
76 | while ((dst = next) != NULL) { | |
77 | next = dst->next; | |
78 | prefetch(&next->next); | |
2fc1b5dd | 79 | cond_resched(); |
86bba269 ED |
80 | if (likely(atomic_read(&dst->__refcnt))) { |
81 | last->next = dst; | |
82 | last = dst; | |
1da177e4 LT |
83 | delayed++; |
84 | continue; | |
85 | } | |
86bba269 | 86 | work_performed++; |
1da177e4 LT |
87 | |
88 | dst = dst_destroy(dst); | |
89 | if (dst) { | |
90 | /* NOHASH and still referenced. Unless it is already | |
91 | * on gc list, invalidate it and add to gc list. | |
92 | * | |
93 | * Note: this is temporary. Actually, NOHASH dst's | |
94 | * must be obsoleted when parent is obsoleted. | |
95 | * But we do not have state "obsoleted, but | |
96 | * referenced by parent", so it is right. | |
97 | */ | |
f5b0a874 | 98 | if (dst->obsolete > 0) |
1da177e4 LT |
99 | continue; |
100 | ||
101 | ___dst_free(dst); | |
86bba269 ED |
102 | dst->next = next; |
103 | next = dst; | |
1da177e4 LT |
104 | } |
105 | } | |
86bba269 ED |
106 | |
107 | spin_lock_bh(&dst_garbage.lock); | |
108 | next = dst_garbage.list; | |
109 | if (next) { | |
110 | dst_garbage.list = NULL; | |
111 | spin_unlock_bh(&dst_garbage.lock); | |
112 | goto loop; | |
1da177e4 | 113 | } |
86bba269 ED |
114 | last->next = NULL; |
115 | dst_busy_list = head.next; | |
116 | if (!dst_busy_list) | |
117 | dst_garbage.timer_inc = DST_GC_MAX; | |
118 | else { | |
119 | /* | |
120 | * if we freed less than 1/10 of delayed entries, | |
121 | * we can sleep longer. | |
122 | */ | |
123 | if (work_performed <= delayed/10) { | |
124 | dst_garbage.timer_expires += dst_garbage.timer_inc; | |
125 | if (dst_garbage.timer_expires > DST_GC_MAX) | |
126 | dst_garbage.timer_expires = DST_GC_MAX; | |
127 | dst_garbage.timer_inc += DST_GC_INC; | |
128 | } else { | |
129 | dst_garbage.timer_inc = DST_GC_INC; | |
130 | dst_garbage.timer_expires = DST_GC_MIN; | |
131 | } | |
132 | expires = dst_garbage.timer_expires; | |
133 | /* | |
598ed936 | 134 | * if the next desired timer is more than 4 seconds in the |
135 | * future then round the timer to whole seconds | |
86bba269 ED |
136 | */ |
137 | if (expires > 4*HZ) | |
138 | expires = round_jiffies_relative(expires); | |
139 | schedule_delayed_work(&dst_gc_work, expires); | |
f0098f78 | 140 | } |
86bba269 ED |
141 | |
142 | spin_unlock_bh(&dst_garbage.lock); | |
143 | mutex_unlock(&dst_gc_mutex); | |
1da177e4 LT |
144 | } |
145 | ||
aad88724 | 146 | int dst_discard_sk(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
147 | { |
148 | kfree_skb(skb); | |
149 | return 0; | |
150 | } | |
aad88724 | 151 | EXPORT_SYMBOL(dst_discard_sk); |
1da177e4 | 152 | |
a37e6e34 ED |
153 | const u32 dst_default_metrics[RTAX_MAX + 1] = { |
154 | /* This initializer is needed to force linker to place this variable | |
155 | * into const section. Otherwise it might end into bss section. | |
156 | * We really want to avoid false sharing on this variable, and catch | |
157 | * any writes on it. | |
158 | */ | |
159 | [RTAX_MAX] = 0xdeadbeef, | |
160 | }; | |
161 | ||
f38a9eb1 TG |
162 | void dst_init(struct dst_entry *dst, struct dst_ops *ops, |
163 | struct net_device *dev, int initial_ref, int initial_obsolete, | |
164 | unsigned short flags) | |
1da177e4 | 165 | { |
cf911662 | 166 | dst->child = NULL; |
5c1e6aa3 DM |
167 | dst->dev = dev; |
168 | if (dev) | |
169 | dev_hold(dev); | |
1da177e4 | 170 | dst->ops = ops; |
62fa8a84 | 171 | dst_init_metrics(dst, dst_default_metrics, true); |
cf911662 | 172 | dst->expires = 0UL; |
5c1e6aa3 | 173 | dst->path = dst; |
ecd98837 | 174 | dst->from = NULL; |
cf911662 DM |
175 | #ifdef CONFIG_XFRM |
176 | dst->xfrm = NULL; | |
177 | #endif | |
5c1e6aa3 | 178 | dst->input = dst_discard; |
aad88724 | 179 | dst->output = dst_discard_sk; |
cf911662 | 180 | dst->error = 0; |
5c1e6aa3 | 181 | dst->obsolete = initial_obsolete; |
cf911662 DM |
182 | dst->header_len = 0; |
183 | dst->trailer_len = 0; | |
184 | #ifdef CONFIG_IP_ROUTE_CLASSID | |
185 | dst->tclassid = 0; | |
1da177e4 | 186 | #endif |
5c1e6aa3 | 187 | atomic_set(&dst->__refcnt, initial_ref); |
cf911662 | 188 | dst->__use = 0; |
5c1e6aa3 DM |
189 | dst->lastuse = jiffies; |
190 | dst->flags = flags; | |
5110effe | 191 | dst->pending_confirm = 0; |
cf911662 | 192 | dst->next = NULL; |
957c665f DM |
193 | if (!(flags & DST_NOCOUNT)) |
194 | dst_entries_add(ops, 1); | |
f38a9eb1 TG |
195 | } |
196 | EXPORT_SYMBOL(dst_init); | |
197 | ||
198 | void *dst_alloc(struct dst_ops *ops, struct net_device *dev, | |
199 | int initial_ref, int initial_obsolete, unsigned short flags) | |
200 | { | |
201 | struct dst_entry *dst; | |
202 | ||
203 | if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) { | |
204 | if (ops->gc(ops)) | |
205 | return NULL; | |
206 | } | |
207 | ||
208 | dst = kmem_cache_alloc(ops->kmem_cachep, GFP_ATOMIC); | |
209 | if (!dst) | |
210 | return NULL; | |
211 | ||
212 | dst_init(dst, ops, dev, initial_ref, initial_obsolete, flags); | |
213 | ||
1da177e4 LT |
214 | return dst; |
215 | } | |
598ed936 | 216 | EXPORT_SYMBOL(dst_alloc); |
1da177e4 | 217 | |
598ed936 | 218 | static void ___dst_free(struct dst_entry *dst) |
1da177e4 LT |
219 | { |
220 | /* The first case (dev==NULL) is required, when | |
221 | protocol module is unloaded. | |
222 | */ | |
aad88724 ED |
223 | if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) { |
224 | dst->input = dst_discard; | |
225 | dst->output = dst_discard_sk; | |
226 | } | |
f5b0a874 | 227 | dst->obsolete = DST_OBSOLETE_DEAD; |
1da177e4 LT |
228 | } |
229 | ||
598ed936 | 230 | void __dst_free(struct dst_entry *dst) |
1da177e4 | 231 | { |
86bba269 | 232 | spin_lock_bh(&dst_garbage.lock); |
1da177e4 | 233 | ___dst_free(dst); |
86bba269 ED |
234 | dst->next = dst_garbage.list; |
235 | dst_garbage.list = dst; | |
236 | if (dst_garbage.timer_inc > DST_GC_INC) { | |
237 | dst_garbage.timer_inc = DST_GC_INC; | |
238 | dst_garbage.timer_expires = DST_GC_MIN; | |
41f63c53 TH |
239 | mod_delayed_work(system_wq, &dst_gc_work, |
240 | dst_garbage.timer_expires); | |
1da177e4 | 241 | } |
86bba269 | 242 | spin_unlock_bh(&dst_garbage.lock); |
1da177e4 | 243 | } |
d79d9913 | 244 | EXPORT_SYMBOL(__dst_free); |
1da177e4 LT |
245 | |
246 | struct dst_entry *dst_destroy(struct dst_entry * dst) | |
247 | { | |
248 | struct dst_entry *child; | |
1da177e4 LT |
249 | |
250 | smp_rmb(); | |
251 | ||
252 | again: | |
1da177e4 LT |
253 | child = dst->child; |
254 | ||
957c665f DM |
255 | if (!(dst->flags & DST_NOCOUNT)) |
256 | dst_entries_add(dst->ops, -1); | |
1da177e4 LT |
257 | |
258 | if (dst->ops->destroy) | |
259 | dst->ops->destroy(dst); | |
260 | if (dst->dev) | |
261 | dev_put(dst->dev); | |
f38a9eb1 TG |
262 | |
263 | if (dst->flags & DST_METADATA) | |
264 | kfree(dst); | |
265 | else | |
266 | kmem_cache_free(dst->ops->kmem_cachep, dst); | |
1da177e4 LT |
267 | |
268 | dst = child; | |
269 | if (dst) { | |
6775cab9 HX |
270 | int nohash = dst->flags & DST_NOHASH; |
271 | ||
1da177e4 LT |
272 | if (atomic_dec_and_test(&dst->__refcnt)) { |
273 | /* We were real parent of this dst, so kill child. */ | |
6775cab9 | 274 | if (nohash) |
1da177e4 LT |
275 | goto again; |
276 | } else { | |
277 | /* Child is still referenced, return it for freeing. */ | |
6775cab9 | 278 | if (nohash) |
1da177e4 LT |
279 | return dst; |
280 | /* Child is still in his hash table */ | |
281 | } | |
282 | } | |
283 | return NULL; | |
284 | } | |
598ed936 | 285 | EXPORT_SYMBOL(dst_destroy); |
1da177e4 | 286 | |
f8864972 ED |
287 | static void dst_destroy_rcu(struct rcu_head *head) |
288 | { | |
289 | struct dst_entry *dst = container_of(head, struct dst_entry, rcu_head); | |
290 | ||
291 | dst = dst_destroy(dst); | |
292 | if (dst) | |
293 | __dst_free(dst); | |
294 | } | |
295 | ||
8d330868 IJ |
296 | void dst_release(struct dst_entry *dst) |
297 | { | |
298 | if (dst) { | |
598ed936 | 299 | int newrefcnt; |
ef711cf1 | 300 | |
598ed936 | 301 | newrefcnt = atomic_dec_return(&dst->__refcnt); |
8bf4ada2 KK |
302 | if (unlikely(newrefcnt < 0)) |
303 | net_warn_ratelimited("%s: dst:%p refcnt:%d\n", | |
304 | __func__, dst, newrefcnt); | |
f8864972 ED |
305 | if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) |
306 | call_rcu(&dst->rcu_head, dst_destroy_rcu); | |
8d330868 IJ |
307 | } |
308 | } | |
309 | EXPORT_SYMBOL(dst_release); | |
310 | ||
62fa8a84 DM |
311 | u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old) |
312 | { | |
313 | u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC); | |
314 | ||
315 | if (p) { | |
316 | u32 *old_p = __DST_METRICS_PTR(old); | |
317 | unsigned long prev, new; | |
318 | ||
319 | memcpy(p, old_p, sizeof(u32) * RTAX_MAX); | |
320 | ||
321 | new = (unsigned long) p; | |
322 | prev = cmpxchg(&dst->_metrics, old, new); | |
323 | ||
324 | if (prev != old) { | |
325 | kfree(p); | |
326 | p = __DST_METRICS_PTR(prev); | |
327 | if (prev & DST_METRICS_READ_ONLY) | |
328 | p = NULL; | |
329 | } | |
330 | } | |
331 | return p; | |
332 | } | |
333 | EXPORT_SYMBOL(dst_cow_metrics_generic); | |
334 | ||
335 | /* Caller asserts that dst_metrics_read_only(dst) is false. */ | |
336 | void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old) | |
337 | { | |
338 | unsigned long prev, new; | |
339 | ||
b30c516f | 340 | new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY; |
62fa8a84 DM |
341 | prev = cmpxchg(&dst->_metrics, old, new); |
342 | if (prev == old) | |
343 | kfree(__DST_METRICS_PTR(old)); | |
344 | } | |
345 | EXPORT_SYMBOL(__dst_destroy_metrics_generic); | |
346 | ||
f38a9eb1 TG |
347 | static struct dst_ops md_dst_ops = { |
348 | .family = AF_UNSPEC, | |
349 | }; | |
350 | ||
351 | static int dst_md_discard_sk(struct sock *sk, struct sk_buff *skb) | |
352 | { | |
353 | WARN_ONCE(1, "Attempting to call output on metadata dst\n"); | |
354 | kfree_skb(skb); | |
355 | return 0; | |
356 | } | |
357 | ||
358 | static int dst_md_discard(struct sk_buff *skb) | |
359 | { | |
360 | WARN_ONCE(1, "Attempting to call input on metadata dst\n"); | |
361 | kfree_skb(skb); | |
362 | return 0; | |
363 | } | |
364 | ||
d3aa45ce | 365 | static void __metadata_dst_init(struct metadata_dst *md_dst, u8 optslen) |
f38a9eb1 | 366 | { |
f38a9eb1 TG |
367 | struct dst_entry *dst; |
368 | ||
f38a9eb1 TG |
369 | dst = &md_dst->dst; |
370 | dst_init(dst, &md_dst_ops, NULL, 1, DST_OBSOLETE_NONE, | |
371 | DST_METADATA | DST_NOCACHE | DST_NOCOUNT); | |
372 | ||
373 | dst->input = dst_md_discard; | |
374 | dst->output = dst_md_discard_sk; | |
375 | ||
376 | memset(dst + 1, 0, sizeof(*md_dst) + optslen - sizeof(*dst)); | |
377 | md_dst->opts_len = optslen; | |
d3aa45ce AS |
378 | } |
379 | ||
380 | struct metadata_dst *metadata_dst_alloc(u8 optslen, gfp_t flags) | |
381 | { | |
382 | struct metadata_dst *md_dst; | |
383 | ||
384 | md_dst = kmalloc(sizeof(*md_dst) + optslen, flags); | |
385 | if (!md_dst) | |
386 | return NULL; | |
387 | ||
388 | __metadata_dst_init(md_dst, optslen); | |
f38a9eb1 TG |
389 | |
390 | return md_dst; | |
391 | } | |
392 | EXPORT_SYMBOL_GPL(metadata_dst_alloc); | |
393 | ||
d3aa45ce AS |
394 | struct metadata_dst __percpu *metadata_dst_alloc_percpu(u8 optslen, gfp_t flags) |
395 | { | |
396 | int cpu; | |
397 | struct metadata_dst __percpu *md_dst; | |
398 | ||
399 | md_dst = __alloc_percpu_gfp(sizeof(struct metadata_dst) + optslen, | |
400 | __alignof__(struct metadata_dst), flags); | |
401 | if (!md_dst) | |
402 | return NULL; | |
403 | ||
404 | for_each_possible_cpu(cpu) | |
405 | __metadata_dst_init(per_cpu_ptr(md_dst, cpu), optslen); | |
406 | ||
407 | return md_dst; | |
408 | } | |
409 | EXPORT_SYMBOL_GPL(metadata_dst_alloc_percpu); | |
410 | ||
1da177e4 LT |
411 | /* Dirty hack. We did it in 2.2 (in __dst_free), |
412 | * we have _very_ good reasons not to repeat | |
413 | * this mistake in 2.3, but we have no choice | |
414 | * now. _It_ _is_ _explicit_ _deliberate_ | |
415 | * _race_ _condition_. | |
416 | * | |
417 | * Commented and originally written by Alexey. | |
418 | */ | |
56115511 | 419 | static void dst_ifdown(struct dst_entry *dst, struct net_device *dev, |
420 | int unregister) | |
1da177e4 LT |
421 | { |
422 | if (dst->ops->ifdown) | |
423 | dst->ops->ifdown(dst, dev, unregister); | |
424 | ||
425 | if (dev != dst->dev) | |
426 | return; | |
427 | ||
428 | if (!unregister) { | |
aad88724 ED |
429 | dst->input = dst_discard; |
430 | dst->output = dst_discard_sk; | |
1da177e4 | 431 | } else { |
c346dca1 | 432 | dst->dev = dev_net(dst->dev)->loopback_dev; |
de3cb747 | 433 | dev_hold(dst->dev); |
1da177e4 | 434 | dev_put(dev); |
1da177e4 LT |
435 | } |
436 | } | |
437 | ||
598ed936 | 438 | static int dst_dev_event(struct notifier_block *this, unsigned long event, |
439 | void *ptr) | |
1da177e4 | 440 | { |
351638e7 | 441 | struct net_device *dev = netdev_notifier_info_to_dev(ptr); |
86bba269 | 442 | struct dst_entry *dst, *last = NULL; |
1da177e4 LT |
443 | |
444 | switch (event) { | |
0115e8e3 | 445 | case NETDEV_UNREGISTER_FINAL: |
1da177e4 | 446 | case NETDEV_DOWN: |
86bba269 ED |
447 | mutex_lock(&dst_gc_mutex); |
448 | for (dst = dst_busy_list; dst; dst = dst->next) { | |
449 | last = dst; | |
450 | dst_ifdown(dst, dev, event != NETDEV_DOWN); | |
451 | } | |
452 | ||
453 | spin_lock_bh(&dst_garbage.lock); | |
454 | dst = dst_garbage.list; | |
455 | dst_garbage.list = NULL; | |
456 | spin_unlock_bh(&dst_garbage.lock); | |
457 | ||
458 | if (last) | |
459 | last->next = dst; | |
460 | else | |
461 | dst_busy_list = dst; | |
598ed936 | 462 | for (; dst; dst = dst->next) |
1da177e4 | 463 | dst_ifdown(dst, dev, event != NETDEV_DOWN); |
86bba269 | 464 | mutex_unlock(&dst_gc_mutex); |
1da177e4 LT |
465 | break; |
466 | } | |
467 | return NOTIFY_DONE; | |
468 | } | |
469 | ||
470 | static struct notifier_block dst_dev_notifier = { | |
471 | .notifier_call = dst_dev_event, | |
332dd96f | 472 | .priority = -10, /* must be called after other network notifiers */ |
1da177e4 LT |
473 | }; |
474 | ||
f38a9eb1 | 475 | void __init dst_subsys_init(void) |
1da177e4 LT |
476 | { |
477 | register_netdevice_notifier(&dst_dev_notifier); | |
478 | } |