xref: /openbmc/linux/net/core/dst.c (revision 86bba269) 
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/netdevice.h>
16 #include <linux/skbuff.h>
17 #include <linux/string.h>
18 #include <linux/types.h>
19 #include <net/net_namespace.h>
20 
21 #include <net/dst.h>
22 
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.
33  */
34 #if RT_CACHE_DEBUG >= 2
35 static atomic_t			 dst_total = ATOMIC_INIT(0);
36 #endif
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 #if RT_CACHE_DEBUG >= 2
71 	ktime_t time_start = ktime_get();
72 	struct timespec elapsed;
73 #endif
74 
75 	mutex_lock(&dst_gc_mutex);
76 	next = dst_busy_list;
77 
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;
85 			delayed++;
86 			continue;
87 		}
88 		work_performed++;
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);
104 			dst->next = next;
105 			next = dst;
106 		}
107 	}
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;
115 	}
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);
142 	}
143 
144 	spin_unlock_bh(&dst_garbage.lock);
145 	mutex_unlock(&dst_gc_mutex);
146 #if RT_CACHE_DEBUG >= 2
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);
153 #endif
154 }
155 
156 static int dst_discard(struct sk_buff *skb)
157 {
158 	kfree_skb(skb);
159 	return 0;
160 }
161 
162 void * dst_alloc(struct dst_ops * ops)
163 {
164 	struct dst_entry * dst;
165 
166 	if (ops->gc && atomic_read(&ops->entries) > ops->gc_thresh) {
167 		if (ops->gc())
168 			return NULL;
169 	}
170 	dst = kmem_cache_zalloc(ops->kmem_cachep, GFP_ATOMIC);
171 	if (!dst)
172 		return NULL;
173 	atomic_set(&dst->__refcnt, 0);
174 	dst->ops = ops;
175 	dst->lastuse = jiffies;
176 	dst->path = dst;
177 	dst->input = dst->output = dst_discard;
178 #if RT_CACHE_DEBUG >= 2
179 	atomic_inc(&dst_total);
180 #endif
181 	atomic_inc(&ops->entries);
182 	return dst;
183 }
184 
185 static void ___dst_free(struct dst_entry * dst)
186 {
187 	/* The first case (dev==NULL) is required, when
188 	   protocol module is unloaded.
189 	 */
190 	if (dst->dev == NULL || !(dst->dev->flags&IFF_UP)) {
191 		dst->input = dst->output = dst_discard;
192 	}
193 	dst->obsolete = 2;
194 }
195 
196 void __dst_free(struct dst_entry * dst)
197 {
198 	spin_lock_bh(&dst_garbage.lock);
199 	___dst_free(dst);
200 	dst->next = dst_garbage.list;
201 	dst_garbage.list = dst;
202 	if (dst_garbage.timer_inc > DST_GC_INC) {
203 		dst_garbage.timer_inc = DST_GC_INC;
204 		dst_garbage.timer_expires = DST_GC_MIN;
205 		schedule_delayed_work(&dst_gc_work, dst_garbage.timer_expires);
206 	}
207 	spin_unlock_bh(&dst_garbage.lock);
208 }
209 
210 struct dst_entry *dst_destroy(struct dst_entry * dst)
211 {
212 	struct dst_entry *child;
213 	struct neighbour *neigh;
214 	struct hh_cache *hh;
215 
216 	smp_rmb();
217 
218 again:
219 	neigh = dst->neighbour;
220 	hh = dst->hh;
221 	child = dst->child;
222 
223 	dst->hh = NULL;
224 	if (hh && atomic_dec_and_test(&hh->hh_refcnt))
225 		kfree(hh);
226 
227 	if (neigh) {
228 		dst->neighbour = NULL;
229 		neigh_release(neigh);
230 	}
231 
232 	atomic_dec(&dst->ops->entries);
233 
234 	if (dst->ops->destroy)
235 		dst->ops->destroy(dst);
236 	if (dst->dev)
237 		dev_put(dst->dev);
238 #if RT_CACHE_DEBUG >= 2
239 	atomic_dec(&dst_total);
240 #endif
241 	kmem_cache_free(dst->ops->kmem_cachep, dst);
242 
243 	dst = child;
244 	if (dst) {
245 		int nohash = dst->flags & DST_NOHASH;
246 
247 		if (atomic_dec_and_test(&dst->__refcnt)) {
248 			/* We were real parent of this dst, so kill child. */
249 			if (nohash)
250 				goto again;
251 		} else {
252 			/* Child is still referenced, return it for freeing. */
253 			if (nohash)
254 				return dst;
255 			/* Child is still in his hash table */
256 		}
257 	}
258 	return NULL;
259 }
260 
261 /* Dirty hack. We did it in 2.2 (in __dst_free),
262  * we have _very_ good reasons not to repeat
263  * this mistake in 2.3, but we have no choice
264  * now. _It_ _is_ _explicit_ _deliberate_
265  * _race_ _condition_.
266  *
267  * Commented and originally written by Alexey.
268  */
269 static inline void dst_ifdown(struct dst_entry *dst, struct net_device *dev,
270 			      int unregister)
271 {
272 	if (dst->ops->ifdown)
273 		dst->ops->ifdown(dst, dev, unregister);
274 
275 	if (dev != dst->dev)
276 		return;
277 
278 	if (!unregister) {
279 		dst->input = dst->output = dst_discard;
280 	} else {
281 		dst->dev = &loopback_dev;
282 		dev_hold(&loopback_dev);
283 		dev_put(dev);
284 		if (dst->neighbour && dst->neighbour->dev == dev) {
285 			dst->neighbour->dev = &loopback_dev;
286 			dev_put(dev);
287 			dev_hold(&loopback_dev);
288 		}
289 	}
290 }
291 
292 static int dst_dev_event(struct notifier_block *this, unsigned long event, void *ptr)
293 {
294 	struct net_device *dev = ptr;
295 	struct dst_entry *dst, *last = NULL;
296 
297 	if (dev->nd_net != &init_net)
298 		return NOTIFY_DONE;
299 
300 	switch (event) {
301 	case NETDEV_UNREGISTER:
302 	case NETDEV_DOWN:
303 		mutex_lock(&dst_gc_mutex);
304 		for (dst = dst_busy_list; dst; dst = dst->next) {
305 			last = dst;
306 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
307 		}
308 
309 		spin_lock_bh(&dst_garbage.lock);
310 		dst = dst_garbage.list;
311 		dst_garbage.list = NULL;
312 		spin_unlock_bh(&dst_garbage.lock);
313 
314 		if (last)
315 			last->next = dst;
316 		else
317 			dst_busy_list = dst;
318 		for (; dst; dst = dst->next) {
319 			dst_ifdown(dst, dev, event != NETDEV_DOWN);
320 		}
321 		mutex_unlock(&dst_gc_mutex);
322 		break;
323 	}
324 	return NOTIFY_DONE;
325 }
326 
327 static struct notifier_block dst_dev_notifier = {
328 	.notifier_call	= dst_dev_event,
329 };
330 
331 void __init dst_init(void)
332 {
333 	register_netdevice_notifier(&dst_dev_notifier);
334 }
335 
336 EXPORT_SYMBOL(__dst_free);
337 EXPORT_SYMBOL(dst_alloc);
338 EXPORT_SYMBOL(dst_destroy);
339