xref: /openbmc/linux/net/core/dst.c (revision 7fe2f639) 
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 > 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 
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];
153 
154 void *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 	dst->hh = NULL;
176 #ifdef CONFIG_XFRM
177 	dst->xfrm = NULL;
178 #endif
179 	dst->input = dst_discard;
180 	dst->output = dst_discard;
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 	atomic_set(&dst->__refcnt, initial_ref);
189 	dst->__use = 0;
190 	dst->lastuse = jiffies;
191 	dst->flags = flags;
192 	dst->next = NULL;
193 	if (!(flags & DST_NOCOUNT))
194 		dst_entries_add(ops, 1);
195 	return dst;
196 }
197 EXPORT_SYMBOL(dst_alloc);
198 
199 static void ___dst_free(struct dst_entry *dst)
200 {
201 	/* The first case (dev==NULL) is required, when
202 	   protocol module is unloaded.
203 	 */
204 	if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
205 		dst->input = dst->output = dst_discard;
206 	dst->obsolete = 2;
207 }
208 
209 void __dst_free(struct dst_entry *dst)
210 {
211 	spin_lock_bh(&dst_garbage.lock);
212 	___dst_free(dst);
213 	dst->next = dst_garbage.list;
214 	dst_garbage.list = dst;
215 	if (dst_garbage.timer_inc > DST_GC_INC) {
216 		dst_garbage.timer_inc = DST_GC_INC;
217 		dst_garbage.timer_expires = DST_GC_MIN;
218 		cancel_delayed_work(&dst_gc_work);
219 		schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
220 	}
221 	spin_unlock_bh(&dst_garbage.lock);
222 }
223 EXPORT_SYMBOL(__dst_free);
224 
225 struct dst_entry *dst_destroy(struct dst_entry * dst)
226 {
227 	struct dst_entry *child;
228 	struct neighbour *neigh;
229 	struct hh_cache *hh;
230 
231 	smp_rmb();
232 
233 again:
234 	neigh = dst->neighbour;
235 	hh = dst->hh;
236 	child = dst->child;
237 
238 	dst->hh = NULL;
239 	if (hh)
240 		hh_cache_put(hh);
241 
242 	if (neigh) {
243 		dst->neighbour = NULL;
244 		neigh_release(neigh);
245 	}
246 
247 	if (!(dst->flags & DST_NOCOUNT))
248 		dst_entries_add(dst->ops, -1);
249 
250 	if (dst->ops->destroy)
251 		dst->ops->destroy(dst);
252 	if (dst->dev)
253 		dev_put(dst->dev);
254 	kmem_cache_free(dst->ops->kmem_cachep, dst);
255 
256 	dst = child;
257 	if (dst) {
258 		int nohash = dst->flags & DST_NOHASH;
259 
260 		if (atomic_dec_and_test(&dst->__refcnt)) {
261 			/* We were real parent of this dst, so kill child. */
262 			if (nohash)
263 				goto again;
264 		} else {
265 			/* Child is still referenced, return it for freeing. */
266 			if (nohash)
267 				return dst;
268 			/* Child is still in his hash table */
269 		}
270 	}
271 	return NULL;
272 }
273 EXPORT_SYMBOL(dst_destroy);
274 
275 void dst_release(struct dst_entry *dst)
276 {
277 	if (dst) {
278 		int newrefcnt;
279 
280 		newrefcnt = atomic_dec_return(&dst->__refcnt);
281 		WARN_ON(newrefcnt < 0);
282 		if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
283 			dst = dst_destroy(dst);
284 			if (dst)
285 				__dst_free(dst);
286 		}
287 	}
288 }
289 EXPORT_SYMBOL(dst_release);
290 
291 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
292 {
293 	u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
294 
295 	if (p) {
296 		u32 *old_p = __DST_METRICS_PTR(old);
297 		unsigned long prev, new;
298 
299 		memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
300 
301 		new = (unsigned long) p;
302 		prev = cmpxchg(&dst->_metrics, old, new);
303 
304 		if (prev != old) {
305 			kfree(p);
306 			p = __DST_METRICS_PTR(prev);
307 			if (prev & DST_METRICS_READ_ONLY)
308 				p = NULL;
309 		}
310 	}
311 	return p;
312 }
313 EXPORT_SYMBOL(dst_cow_metrics_generic);
314 
315 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
316 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
317 {
318 	unsigned long prev, new;
319 
320 	new = ((unsigned long) dst_default_metrics) | DST_METRICS_READ_ONLY;
321 	prev = cmpxchg(&dst->_metrics, old, new);
322 	if (prev == old)
323 		kfree(__DST_METRICS_PTR(old));
324 }
325 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
326 
327 /**
328  * skb_dst_set_noref - sets skb dst, without a reference
329  * @skb: buffer
330  * @dst: dst entry
331  *
332  * Sets skb dst, assuming a reference was not taken on dst
333  * skb_dst_drop() should not dst_release() this dst
334  */
335 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
336 {
337 	WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
338 	/* If dst not in cache, we must take a reference, because
339 	 * dst_release() will destroy dst as soon as its refcount becomes zero
340 	 */
341 	if (unlikely(dst->flags & DST_NOCACHE)) {
342 		dst_hold(dst);
343 		skb_dst_set(skb, dst);
344 	} else {
345 		skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
346 	}
347 }
348 EXPORT_SYMBOL(skb_dst_set_noref);
349 
350 /* Dirty hack. We did it in 2.2 (in __dst_free),
351  * we have _very_ good reasons not to repeat
352  * this mistake in 2.3, but we have no choice
353  * now. _It_ _is_ _explicit_ _deliberate_
354  * _race_ _condition_.
355  *
356  * Commented and originally written by Alexey.
357  */
358 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
359 		       int unregister)
360 {
361 	if (dst->ops->ifdown)
362 		dst->ops->ifdown(dst, dev, unregister);
363 
364 	if (dev != dst->dev)
365 		return;
366 
367 	if (!unregister) {
368 		dst->input = dst->output = dst_discard;
369 	} else {
370 		dst->dev = dev_net(dst->dev)->loopback_dev;
371 		dev_hold(dst->dev);
372 		dev_put(dev);
373 		if (dst->neighbour && dst->neighbour->dev == dev) {
374 			dst->neighbour->dev = dst->dev;
375 			dev_hold(dst->dev);
376 			dev_put(dev);
377 		}
378 	}
379 }
380 
381 static int dst_dev_event(struct notifier_block *this, unsigned long event,
382 			 void *ptr)
383 {
384 	struct net_device *dev = ptr;
385 	struct dst_entry *dst, *last = NULL;
386 
387 	switch (event) {
388 	case NETDEV_UNREGISTER:
389 	case NETDEV_DOWN:
390 		mutex_lock(&dst_gc_mutex);
391 		for (dst = dst_busy_list; dst; dst = dst->next) {
392 			last = dst;
393 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
394 		}
395 
396 		spin_lock_bh(&dst_garbage.lock);
397 		dst = dst_garbage.list;
398 		dst_garbage.list = NULL;
399 		spin_unlock_bh(&dst_garbage.lock);
400 
401 		if (last)
402 			last->next = dst;
403 		else
404 			dst_busy_list = dst;
405 		for (; dst; dst = dst->next)
406 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
407 		mutex_unlock(&dst_gc_mutex);
408 		break;
409 	}
410 	return NOTIFY_DONE;
411 }
412 
413 static struct notifier_block dst_dev_notifier = {
414 	.notifier_call	= dst_dev_event,
415 	.priority = -10, /* must be called after other network notifiers */
416 };
417 
418 void __init dst_init(void)
419 {
420 	register_netdevice_notifier(&dst_dev_notifier);
421 }
422