xref: /openbmc/linux/net/core/dst.c (revision 615c36f5) 
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 	RCU_INIT_POINTER(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 }
196 EXPORT_SYMBOL(dst_alloc);
197 
198 static 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 
208 void __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 }
222 EXPORT_SYMBOL(__dst_free);
223 
224 struct dst_entry *dst_destroy(struct dst_entry * dst)
225 {
226 	struct dst_entry *child;
227 	struct neighbour *neigh;
228 
229 	smp_rmb();
230 
231 again:
232 	neigh = rcu_dereference_protected(dst->_neighbour, 1);
233 	child = dst->child;
234 
235 	if (neigh) {
236 		RCU_INIT_POINTER(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 }
266 EXPORT_SYMBOL(dst_destroy);
267 
268 void 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 }
282 EXPORT_SYMBOL(dst_release);
283 
284 u32 *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 }
306 EXPORT_SYMBOL(dst_cow_metrics_generic);
307 
308 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
309 void __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 }
318 EXPORT_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  */
328 void 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 }
341 EXPORT_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  */
351 static 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 		struct neighbour *neigh;
364 
365 		dst->dev = dev_net(dst->dev)->loopback_dev;
366 		dev_hold(dst->dev);
367 		dev_put(dev);
368 		rcu_read_lock();
369 		neigh = dst_get_neighbour(dst);
370 		if (neigh && neigh->dev == dev) {
371 			neigh->dev = dst->dev;
372 			dev_hold(dst->dev);
373 			dev_put(dev);
374 		}
375 		rcu_read_unlock();
376 	}
377 }
378 
379 static int dst_dev_event(struct notifier_block *this, unsigned long event,
380 			 void *ptr)
381 {
382 	struct net_device *dev = ptr;
383 	struct dst_entry *dst, *last = NULL;
384 
385 	switch (event) {
386 	case NETDEV_UNREGISTER:
387 	case NETDEV_DOWN:
388 		mutex_lock(&dst_gc_mutex);
389 		for (dst = dst_busy_list; dst; dst = dst->next) {
390 			last = dst;
391 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
392 		}
393 
394 		spin_lock_bh(&dst_garbage.lock);
395 		dst = dst_garbage.list;
396 		dst_garbage.list = NULL;
397 		spin_unlock_bh(&dst_garbage.lock);
398 
399 		if (last)
400 			last->next = dst;
401 		else
402 			dst_busy_list = dst;
403 		for (; dst; dst = dst->next)
404 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
405 		mutex_unlock(&dst_gc_mutex);
406 		break;
407 	}
408 	return NOTIFY_DONE;
409 }
410 
411 static struct notifier_block dst_dev_notifier = {
412 	.notifier_call	= dst_dev_event,
413 	.priority = -10, /* must be called after other network notifiers */
414 };
415 
416 void __init dst_init(void)
417 {
418 	register_netdevice_notifier(&dst_dev_notifier);
419 }
420