xref: /openbmc/linux/net/core/dst.c (revision 725d1e1b) 
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 
23 #include <net/dst.h>
24 
25 /*
26  * Theory of operations:
27  * 1) We use a list, protected by a spinlock, to add
28  *    new entries from both BH and non-BH context.
29  * 2) In order to keep spinlock held for a small delay,
30  *    we use a second list where are stored long lived
31  *    entries, that are handled by the garbage collect thread
32  *    fired by a workqueue.
33  * 3) This list is guarded by a mutex,
34  *    so that the gc_task and dst_dev_event() can be synchronized.
35  */
36 #if RT_CACHE_DEBUG >= 2
37 static atomic_t			 dst_total = ATOMIC_INIT(0);
38 #endif
39 
40 /*
41  * We want to keep lock & list close together
42  * to dirty as few cache lines as possible in __dst_free().
43  * As this is not a very strong hint, we dont force an alignment on SMP.
44  */
45 static struct {
46 	spinlock_t		lock;
47 	struct dst_entry	*list;
48 	unsigned long		timer_inc;
49 	unsigned long		timer_expires;
50 } dst_garbage = {
51 	.lock = __SPIN_LOCK_UNLOCKED(dst_garbage.lock),
52 	.timer_inc = DST_GC_MAX,
53 };
54 static void dst_gc_task(struct work_struct *work);
55 static void ___dst_free(struct dst_entry *dst);
56 
57 static DECLARE_DELAYED_WORK(dst_gc_work, dst_gc_task);
58 
59 static DEFINE_MUTEX(dst_gc_mutex);
60 /*
61  * long lived entries are maintained in this list, guarded by dst_gc_mutex
62  */
63 static struct dst_entry         *dst_busy_list;
64 
65 static void dst_gc_task(struct work_struct *work)
66 {
67 	int    delayed = 0;
68 	int    work_performed = 0;
69 	unsigned long expires = ~0L;
70 	struct dst_entry *dst, *next, head;
71 	struct dst_entry *last = &head;
72 #if RT_CACHE_DEBUG >= 2
73 	ktime_t time_start = ktime_get();
74 	struct timespec elapsed;
75 #endif
76 
77 	mutex_lock(&dst_gc_mutex);
78 	next = dst_busy_list;
79 
80 loop:
81 	while ((dst = next) != NULL) {
82 		next = dst->next;
83 		prefetch(&next->next);
84 		cond_resched();
85 		if (likely(atomic_read(&dst->__refcnt))) {
86 			last->next = dst;
87 			last = dst;
88 			delayed++;
89 			continue;
90 		}
91 		work_performed++;
92 
93 		dst = dst_destroy(dst);
94 		if (dst) {
95 			/* NOHASH and still referenced. Unless it is already
96 			 * on gc list, invalidate it and add to gc list.
97 			 *
98 			 * Note: this is temporary. Actually, NOHASH dst's
99 			 * must be obsoleted when parent is obsoleted.
100 			 * But we do not have state "obsoleted, but
101 			 * referenced by parent", so it is right.
102 			 */
103 			if (dst->obsolete > 1)
104 				continue;
105 
106 			___dst_free(dst);
107 			dst->next = next;
108 			next = dst;
109 		}
110 	}
111 
112 	spin_lock_bh(&dst_garbage.lock);
113 	next = dst_garbage.list;
114 	if (next) {
115 		dst_garbage.list = NULL;
116 		spin_unlock_bh(&dst_garbage.lock);
117 		goto loop;
118 	}
119 	last->next = NULL;
120 	dst_busy_list = head.next;
121 	if (!dst_busy_list)
122 		dst_garbage.timer_inc = DST_GC_MAX;
123 	else {
124 		/*
125 		 * if we freed less than 1/10 of delayed entries,
126 		 * we can sleep longer.
127 		 */
128 		if (work_performed <= delayed/10) {
129 			dst_garbage.timer_expires += dst_garbage.timer_inc;
130 			if (dst_garbage.timer_expires > DST_GC_MAX)
131 				dst_garbage.timer_expires = DST_GC_MAX;
132 			dst_garbage.timer_inc += DST_GC_INC;
133 		} else {
134 			dst_garbage.timer_inc = DST_GC_INC;
135 			dst_garbage.timer_expires = DST_GC_MIN;
136 		}
137 		expires = dst_garbage.timer_expires;
138 		/*
139 		 * if the next desired timer is more than 4 seconds in the
140 		 * future then round the timer to whole seconds
141 		 */
142 		if (expires > 4*HZ)
143 			expires = round_jiffies_relative(expires);
144 		schedule_delayed_work(&dst_gc_work, expires);
145 	}
146 
147 	spin_unlock_bh(&dst_garbage.lock);
148 	mutex_unlock(&dst_gc_mutex);
149 #if RT_CACHE_DEBUG >= 2
150 	elapsed = ktime_to_timespec(ktime_sub(ktime_get(), time_start));
151 	printk(KERN_DEBUG "dst_total: %d delayed: %d work_perf: %d"
152 		" expires: %lu elapsed: %lu us\n",
153 		atomic_read(&dst_total), delayed, work_performed,
154 		expires,
155 		elapsed.tv_sec * USEC_PER_SEC +
156 		  elapsed.tv_nsec / NSEC_PER_USEC);
157 #endif
158 }
159 
160 int dst_discard(struct sk_buff *skb)
161 {
162 	kfree_skb(skb);
163 	return 0;
164 }
165 EXPORT_SYMBOL(dst_discard);
166 
167 const u32 dst_default_metrics[RTAX_MAX];
168 
169 void *dst_alloc(struct dst_ops *ops)
170 {
171 	struct dst_entry *dst;
172 
173 	if (ops->gc && dst_entries_get_fast(ops) > ops->gc_thresh) {
174 		if (ops->gc(ops))
175 			return NULL;
176 	}
177 	dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
178 	if (!dst)
179 		return NULL;
180 	atomic_set(&dst->__refcnt, 0);
181 	dst->ops = ops;
182 	dst->lastuse = jiffies;
183 	dst->path = dst;
184 	dst->input = dst->output = dst_discard;
185 	dst_init_metrics(dst, dst_default_metrics, true);
186 #if RT_CACHE_DEBUG >= 2
187 	atomic_inc(&dst_total);
188 #endif
189 	dst_entries_add(ops, 1);
190 	return dst;
191 }
192 EXPORT_SYMBOL(dst_alloc);
193 
194 static void ___dst_free(struct dst_entry *dst)
195 {
196 	/* The first case (dev==NULL) is required, when
197 	   protocol module is unloaded.
198 	 */
199 	if (dst->dev == NULL || !(dst->dev->flags&IFF_UP))
200 		dst->input = dst->output = dst_discard;
201 	dst->obsolete = 2;
202 }
203 
204 void __dst_free(struct dst_entry *dst)
205 {
206 	spin_lock_bh(&dst_garbage.lock);
207 	___dst_free(dst);
208 	dst->next = dst_garbage.list;
209 	dst_garbage.list = dst;
210 	if (dst_garbage.timer_inc > DST_GC_INC) {
211 		dst_garbage.timer_inc = DST_GC_INC;
212 		dst_garbage.timer_expires = DST_GC_MIN;
213 		cancel_delayed_work(&dst_gc_work);
214 		schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
215 	}
216 	spin_unlock_bh(&dst_garbage.lock);
217 }
218 EXPORT_SYMBOL(__dst_free);
219 
220 struct dst_entry *dst_destroy(struct dst_entry * dst)
221 {
222 	struct dst_entry *child;
223 	struct neighbour *neigh;
224 	struct hh_cache *hh;
225 
226 	smp_rmb();
227 
228 again:
229 	neigh = dst->neighbour;
230 	hh = dst->hh;
231 	child = dst->child;
232 
233 	dst->hh = NULL;
234 	if (hh)
235 		hh_cache_put(hh);
236 
237 	if (neigh) {
238 		dst->neighbour = NULL;
239 		neigh_release(neigh);
240 	}
241 
242 	dst_entries_add(dst->ops, -1);
243 
244 	if (dst->ops->destroy)
245 		dst->ops->destroy(dst);
246 	if (dst->dev)
247 		dev_put(dst->dev);
248 #if RT_CACHE_DEBUG >= 2
249 	atomic_dec(&dst_total);
250 #endif
251 	kmem_cache_free(dst->ops->kmem_cachep, dst);
252 
253 	dst = child;
254 	if (dst) {
255 		int nohash = dst->flags & DST_NOHASH;
256 
257 		if (atomic_dec_and_test(&dst->__refcnt)) {
258 			/* We were real parent of this dst, so kill child. */
259 			if (nohash)
260 				goto again;
261 		} else {
262 			/* Child is still referenced, return it for freeing. */
263 			if (nohash)
264 				return dst;
265 			/* Child is still in his hash table */
266 		}
267 	}
268 	return NULL;
269 }
270 EXPORT_SYMBOL(dst_destroy);
271 
272 void dst_release(struct dst_entry *dst)
273 {
274 	if (dst) {
275 		int newrefcnt;
276 
277 		newrefcnt = atomic_dec_return(&dst->__refcnt);
278 		WARN_ON(newrefcnt < 0);
279 		if (unlikely(dst->flags & DST_NOCACHE) && !newrefcnt) {
280 			dst = dst_destroy(dst);
281 			if (dst)
282 				__dst_free(dst);
283 		}
284 	}
285 }
286 EXPORT_SYMBOL(dst_release);
287 
288 u32 *dst_cow_metrics_generic(struct dst_entry *dst, unsigned long old)
289 {
290 	u32 *p = kmalloc(sizeof(u32) * RTAX_MAX, GFP_ATOMIC);
291 
292 	if (p) {
293 		u32 *old_p = __DST_METRICS_PTR(old);
294 		unsigned long prev, new;
295 
296 		memcpy(p, old_p, sizeof(u32) * RTAX_MAX);
297 
298 		new = (unsigned long) p;
299 		prev = cmpxchg(&dst->_metrics, old, new);
300 
301 		if (prev != old) {
302 			kfree(p);
303 			p = __DST_METRICS_PTR(prev);
304 			if (prev & DST_METRICS_READ_ONLY)
305 				p = NULL;
306 		}
307 	}
308 	return p;
309 }
310 EXPORT_SYMBOL(dst_cow_metrics_generic);
311 
312 /* Caller asserts that dst_metrics_read_only(dst) is false.  */
313 void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
314 {
315 	unsigned long prev, new;
316 
317 	new = (unsigned long) dst_default_metrics;
318 	prev = cmpxchg(&dst->_metrics, old, new);
319 	if (prev == old)
320 		kfree(__DST_METRICS_PTR(old));
321 }
322 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
323 
324 /**
325  * skb_dst_set_noref - sets skb dst, without a reference
326  * @skb: buffer
327  * @dst: dst entry
328  *
329  * Sets skb dst, assuming a reference was not taken on dst
330  * skb_dst_drop() should not dst_release() this dst
331  */
332 void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
333 {
334 	WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
335 	/* If dst not in cache, we must take a reference, because
336 	 * dst_release() will destroy dst as soon as its refcount becomes zero
337 	 */
338 	if (unlikely(dst->flags & DST_NOCACHE)) {
339 		dst_hold(dst);
340 		skb_dst_set(skb, dst);
341 	} else {
342 		skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
343 	}
344 }
345 EXPORT_SYMBOL(skb_dst_set_noref);
346 
347 /* Dirty hack. We did it in 2.2 (in __dst_free),
348  * we have _very_ good reasons not to repeat
349  * this mistake in 2.3, but we have no choice
350  * now. _It_ _is_ _explicit_ _deliberate_
351  * _race_ _condition_.
352  *
353  * Commented and originally written by Alexey.
354  */
355 static void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
356 		       int unregister)
357 {
358 	if (dst->ops->ifdown)
359 		dst->ops->ifdown(dst, dev, unregister);
360 
361 	if (dev != dst->dev)
362 		return;
363 
364 	if (!unregister) {
365 		dst->input = dst->output = dst_discard;
366 	} else {
367 		dst->dev = dev_net(dst->dev)->loopback_dev;
368 		dev_hold(dst->dev);
369 		dev_put(dev);
370 		if (dst->neighbour && dst->neighbour->dev == dev) {
371 			dst->neighbour->dev = dst->dev;
372 			dev_hold(dst->dev);
373 			dev_put(dev);
374 		}
375 	}
376 }
377 
378 static int dst_dev_event(struct notifier_block *this, unsigned long event,
379 			 void *ptr)
380 {
381 	struct net_device *dev = ptr;
382 	struct dst_entry *dst, *last = NULL;
383 
384 	switch (event) {
385 	case NETDEV_UNREGISTER:
386 	case NETDEV_DOWN:
387 		mutex_lock(&dst_gc_mutex);
388 		for (dst = dst_busy_list; dst; dst = dst->next) {
389 			last = dst;
390 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
391 		}
392 
393 		spin_lock_bh(&dst_garbage.lock);
394 		dst = dst_garbage.list;
395 		dst_garbage.list = NULL;
396 		spin_unlock_bh(&dst_garbage.lock);
397 
398 		if (last)
399 			last->next = dst;
400 		else
401 			dst_busy_list = dst;
402 		for (; dst; dst = dst->next)
403 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
404 		mutex_unlock(&dst_gc_mutex);
405 		break;
406 	}
407 	return NOTIFY_DONE;
408 }
409 
410 static struct notifier_block dst_dev_notifier = {
411 	.notifier_call	= dst_dev_event,
412 	.priority = -10, /* must be called after other network notifiers */
413 };
414 
415 void __init dst_init(void)
416 {
417 	register_netdevice_notifier(&dst_dev_notifier);
418 }
419