xref: /openbmc/linux/kernel/bpf/devmap.c (revision 82e6fdd6)
1 /* Copyright (c) 2017 Covalent IO, Inc. http://covalent.io
2  *
3  * This program is free software; you can redistribute it and/or
4  * modify it under the terms of version 2 of the GNU General Public
5  * License as published by the Free Software Foundation.
6  *
7  * This program is distributed in the hope that it will be useful, but
8  * WITHOUT ANY WARRANTY; without even the implied warranty of
9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
10  * General Public License for more details.
11  */
12 
13 /* Devmaps primary use is as a backend map for XDP BPF helper call
14  * bpf_redirect_map(). Because XDP is mostly concerned with performance we
15  * spent some effort to ensure the datapath with redirect maps does not use
16  * any locking. This is a quick note on the details.
17  *
18  * We have three possible paths to get into the devmap control plane bpf
19  * syscalls, bpf programs, and driver side xmit/flush operations. A bpf syscall
20  * will invoke an update, delete, or lookup operation. To ensure updates and
21  * deletes appear atomic from the datapath side xchg() is used to modify the
22  * netdev_map array. Then because the datapath does a lookup into the netdev_map
23  * array (read-only) from an RCU critical section we use call_rcu() to wait for
24  * an rcu grace period before free'ing the old data structures. This ensures the
25  * datapath always has a valid copy. However, the datapath does a "flush"
26  * operation that pushes any pending packets in the driver outside the RCU
27  * critical section. Each bpf_dtab_netdev tracks these pending operations using
28  * an atomic per-cpu bitmap. The bpf_dtab_netdev object will not be destroyed
29  * until all bits are cleared indicating outstanding flush operations have
30  * completed.
31  *
32  * BPF syscalls may race with BPF program calls on any of the update, delete
33  * or lookup operations. As noted above the xchg() operation also keep the
34  * netdev_map consistent in this case. From the devmap side BPF programs
35  * calling into these operations are the same as multiple user space threads
36  * making system calls.
37  *
38  * Finally, any of the above may race with a netdev_unregister notifier. The
39  * unregister notifier must search for net devices in the map structure that
40  * contain a reference to the net device and remove them. This is a two step
41  * process (a) dereference the bpf_dtab_netdev object in netdev_map and (b)
42  * check to see if the ifindex is the same as the net_device being removed.
43  * When removing the dev a cmpxchg() is used to ensure the correct dev is
44  * removed, in the case of a concurrent update or delete operation it is
45  * possible that the initially referenced dev is no longer in the map. As the
46  * notifier hook walks the map we know that new dev references can not be
47  * added by the user because core infrastructure ensures dev_get_by_index()
48  * calls will fail at this point.
49  */
50 #include <linux/bpf.h>
51 #include <linux/filter.h>
52 
53 #define DEV_CREATE_FLAG_MASK \
54 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
55 
56 struct bpf_dtab_netdev {
57 	struct net_device *dev;
58 	struct bpf_dtab *dtab;
59 	unsigned int bit;
60 	struct rcu_head rcu;
61 };
62 
63 struct bpf_dtab {
64 	struct bpf_map map;
65 	struct bpf_dtab_netdev **netdev_map;
66 	unsigned long __percpu *flush_needed;
67 	struct list_head list;
68 };
69 
70 static DEFINE_SPINLOCK(dev_map_lock);
71 static LIST_HEAD(dev_map_list);
72 
73 static u64 dev_map_bitmap_size(const union bpf_attr *attr)
74 {
75 	return BITS_TO_LONGS((u64) attr->max_entries) * sizeof(unsigned long);
76 }
77 
78 static struct bpf_map *dev_map_alloc(union bpf_attr *attr)
79 {
80 	struct bpf_dtab *dtab;
81 	int err = -EINVAL;
82 	u64 cost;
83 
84 	if (!capable(CAP_NET_ADMIN))
85 		return ERR_PTR(-EPERM);
86 
87 	/* check sanity of attributes */
88 	if (attr->max_entries == 0 || attr->key_size != 4 ||
89 	    attr->value_size != 4 || attr->map_flags & ~DEV_CREATE_FLAG_MASK)
90 		return ERR_PTR(-EINVAL);
91 
92 	dtab = kzalloc(sizeof(*dtab), GFP_USER);
93 	if (!dtab)
94 		return ERR_PTR(-ENOMEM);
95 
96 	bpf_map_init_from_attr(&dtab->map, attr);
97 
98 	/* make sure page count doesn't overflow */
99 	cost = (u64) dtab->map.max_entries * sizeof(struct bpf_dtab_netdev *);
100 	cost += dev_map_bitmap_size(attr) * num_possible_cpus();
101 	if (cost >= U32_MAX - PAGE_SIZE)
102 		goto free_dtab;
103 
104 	dtab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
105 
106 	/* if map size is larger than memlock limit, reject it early */
107 	err = bpf_map_precharge_memlock(dtab->map.pages);
108 	if (err)
109 		goto free_dtab;
110 
111 	err = -ENOMEM;
112 
113 	/* A per cpu bitfield with a bit per possible net device */
114 	dtab->flush_needed = __alloc_percpu_gfp(dev_map_bitmap_size(attr),
115 						__alignof__(unsigned long),
116 						GFP_KERNEL | __GFP_NOWARN);
117 	if (!dtab->flush_needed)
118 		goto free_dtab;
119 
120 	dtab->netdev_map = bpf_map_area_alloc(dtab->map.max_entries *
121 					      sizeof(struct bpf_dtab_netdev *),
122 					      dtab->map.numa_node);
123 	if (!dtab->netdev_map)
124 		goto free_dtab;
125 
126 	spin_lock(&dev_map_lock);
127 	list_add_tail_rcu(&dtab->list, &dev_map_list);
128 	spin_unlock(&dev_map_lock);
129 
130 	return &dtab->map;
131 free_dtab:
132 	free_percpu(dtab->flush_needed);
133 	kfree(dtab);
134 	return ERR_PTR(err);
135 }
136 
137 static void dev_map_free(struct bpf_map *map)
138 {
139 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
140 	int i, cpu;
141 
142 	/* At this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
143 	 * so the programs (can be more than one that used this map) were
144 	 * disconnected from events. Wait for outstanding critical sections in
145 	 * these programs to complete. The rcu critical section only guarantees
146 	 * no further reads against netdev_map. It does __not__ ensure pending
147 	 * flush operations (if any) are complete.
148 	 */
149 
150 	spin_lock(&dev_map_lock);
151 	list_del_rcu(&dtab->list);
152 	spin_unlock(&dev_map_lock);
153 
154 	synchronize_rcu();
155 
156 	/* To ensure all pending flush operations have completed wait for flush
157 	 * bitmap to indicate all flush_needed bits to be zero on _all_ cpus.
158 	 * Because the above synchronize_rcu() ensures the map is disconnected
159 	 * from the program we can assume no new bits will be set.
160 	 */
161 	for_each_online_cpu(cpu) {
162 		unsigned long *bitmap = per_cpu_ptr(dtab->flush_needed, cpu);
163 
164 		while (!bitmap_empty(bitmap, dtab->map.max_entries))
165 			cond_resched();
166 	}
167 
168 	for (i = 0; i < dtab->map.max_entries; i++) {
169 		struct bpf_dtab_netdev *dev;
170 
171 		dev = dtab->netdev_map[i];
172 		if (!dev)
173 			continue;
174 
175 		dev_put(dev->dev);
176 		kfree(dev);
177 	}
178 
179 	free_percpu(dtab->flush_needed);
180 	bpf_map_area_free(dtab->netdev_map);
181 	kfree(dtab);
182 }
183 
184 static int dev_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
185 {
186 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
187 	u32 index = key ? *(u32 *)key : U32_MAX;
188 	u32 *next = next_key;
189 
190 	if (index >= dtab->map.max_entries) {
191 		*next = 0;
192 		return 0;
193 	}
194 
195 	if (index == dtab->map.max_entries - 1)
196 		return -ENOENT;
197 	*next = index + 1;
198 	return 0;
199 }
200 
201 void __dev_map_insert_ctx(struct bpf_map *map, u32 bit)
202 {
203 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
204 	unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
205 
206 	__set_bit(bit, bitmap);
207 }
208 
209 /* __dev_map_flush is called from xdp_do_flush_map() which _must_ be signaled
210  * from the driver before returning from its napi->poll() routine. The poll()
211  * routine is called either from busy_poll context or net_rx_action signaled
212  * from NET_RX_SOFTIRQ. Either way the poll routine must complete before the
213  * net device can be torn down. On devmap tear down we ensure the ctx bitmap
214  * is zeroed before completing to ensure all flush operations have completed.
215  */
216 void __dev_map_flush(struct bpf_map *map)
217 {
218 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
219 	unsigned long *bitmap = this_cpu_ptr(dtab->flush_needed);
220 	u32 bit;
221 
222 	for_each_set_bit(bit, bitmap, map->max_entries) {
223 		struct bpf_dtab_netdev *dev = READ_ONCE(dtab->netdev_map[bit]);
224 		struct net_device *netdev;
225 
226 		/* This is possible if the dev entry is removed by user space
227 		 * between xdp redirect and flush op.
228 		 */
229 		if (unlikely(!dev))
230 			continue;
231 
232 		__clear_bit(bit, bitmap);
233 		netdev = dev->dev;
234 		if (likely(netdev->netdev_ops->ndo_xdp_flush))
235 			netdev->netdev_ops->ndo_xdp_flush(netdev);
236 	}
237 }
238 
239 /* rcu_read_lock (from syscall and BPF contexts) ensures that if a delete and/or
240  * update happens in parallel here a dev_put wont happen until after reading the
241  * ifindex.
242  */
243 struct net_device  *__dev_map_lookup_elem(struct bpf_map *map, u32 key)
244 {
245 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
246 	struct bpf_dtab_netdev *dev;
247 
248 	if (key >= map->max_entries)
249 		return NULL;
250 
251 	dev = READ_ONCE(dtab->netdev_map[key]);
252 	return dev ? dev->dev : NULL;
253 }
254 
255 static void *dev_map_lookup_elem(struct bpf_map *map, void *key)
256 {
257 	struct net_device *dev = __dev_map_lookup_elem(map, *(u32 *)key);
258 
259 	return dev ? &dev->ifindex : NULL;
260 }
261 
262 static void dev_map_flush_old(struct bpf_dtab_netdev *dev)
263 {
264 	if (dev->dev->netdev_ops->ndo_xdp_flush) {
265 		struct net_device *fl = dev->dev;
266 		unsigned long *bitmap;
267 		int cpu;
268 
269 		for_each_online_cpu(cpu) {
270 			bitmap = per_cpu_ptr(dev->dtab->flush_needed, cpu);
271 			__clear_bit(dev->bit, bitmap);
272 
273 			fl->netdev_ops->ndo_xdp_flush(dev->dev);
274 		}
275 	}
276 }
277 
278 static void __dev_map_entry_free(struct rcu_head *rcu)
279 {
280 	struct bpf_dtab_netdev *dev;
281 
282 	dev = container_of(rcu, struct bpf_dtab_netdev, rcu);
283 	dev_map_flush_old(dev);
284 	dev_put(dev->dev);
285 	kfree(dev);
286 }
287 
288 static int dev_map_delete_elem(struct bpf_map *map, void *key)
289 {
290 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
291 	struct bpf_dtab_netdev *old_dev;
292 	int k = *(u32 *)key;
293 
294 	if (k >= map->max_entries)
295 		return -EINVAL;
296 
297 	/* Use call_rcu() here to ensure any rcu critical sections have
298 	 * completed, but this does not guarantee a flush has happened
299 	 * yet. Because driver side rcu_read_lock/unlock only protects the
300 	 * running XDP program. However, for pending flush operations the
301 	 * dev and ctx are stored in another per cpu map. And additionally,
302 	 * the driver tear down ensures all soft irqs are complete before
303 	 * removing the net device in the case of dev_put equals zero.
304 	 */
305 	old_dev = xchg(&dtab->netdev_map[k], NULL);
306 	if (old_dev)
307 		call_rcu(&old_dev->rcu, __dev_map_entry_free);
308 	return 0;
309 }
310 
311 static int dev_map_update_elem(struct bpf_map *map, void *key, void *value,
312 				u64 map_flags)
313 {
314 	struct bpf_dtab *dtab = container_of(map, struct bpf_dtab, map);
315 	struct net *net = current->nsproxy->net_ns;
316 	struct bpf_dtab_netdev *dev, *old_dev;
317 	u32 i = *(u32 *)key;
318 	u32 ifindex = *(u32 *)value;
319 
320 	if (unlikely(map_flags > BPF_EXIST))
321 		return -EINVAL;
322 	if (unlikely(i >= dtab->map.max_entries))
323 		return -E2BIG;
324 	if (unlikely(map_flags == BPF_NOEXIST))
325 		return -EEXIST;
326 
327 	if (!ifindex) {
328 		dev = NULL;
329 	} else {
330 		dev = kmalloc_node(sizeof(*dev), GFP_ATOMIC | __GFP_NOWARN,
331 				   map->numa_node);
332 		if (!dev)
333 			return -ENOMEM;
334 
335 		dev->dev = dev_get_by_index(net, ifindex);
336 		if (!dev->dev) {
337 			kfree(dev);
338 			return -EINVAL;
339 		}
340 
341 		dev->bit = i;
342 		dev->dtab = dtab;
343 	}
344 
345 	/* Use call_rcu() here to ensure rcu critical sections have completed
346 	 * Remembering the driver side flush operation will happen before the
347 	 * net device is removed.
348 	 */
349 	old_dev = xchg(&dtab->netdev_map[i], dev);
350 	if (old_dev)
351 		call_rcu(&old_dev->rcu, __dev_map_entry_free);
352 
353 	return 0;
354 }
355 
356 const struct bpf_map_ops dev_map_ops = {
357 	.map_alloc = dev_map_alloc,
358 	.map_free = dev_map_free,
359 	.map_get_next_key = dev_map_get_next_key,
360 	.map_lookup_elem = dev_map_lookup_elem,
361 	.map_update_elem = dev_map_update_elem,
362 	.map_delete_elem = dev_map_delete_elem,
363 };
364 
365 static int dev_map_notification(struct notifier_block *notifier,
366 				ulong event, void *ptr)
367 {
368 	struct net_device *netdev = netdev_notifier_info_to_dev(ptr);
369 	struct bpf_dtab *dtab;
370 	int i;
371 
372 	switch (event) {
373 	case NETDEV_UNREGISTER:
374 		/* This rcu_read_lock/unlock pair is needed because
375 		 * dev_map_list is an RCU list AND to ensure a delete
376 		 * operation does not free a netdev_map entry while we
377 		 * are comparing it against the netdev being unregistered.
378 		 */
379 		rcu_read_lock();
380 		list_for_each_entry_rcu(dtab, &dev_map_list, list) {
381 			for (i = 0; i < dtab->map.max_entries; i++) {
382 				struct bpf_dtab_netdev *dev, *odev;
383 
384 				dev = READ_ONCE(dtab->netdev_map[i]);
385 				if (!dev ||
386 				    dev->dev->ifindex != netdev->ifindex)
387 					continue;
388 				odev = cmpxchg(&dtab->netdev_map[i], dev, NULL);
389 				if (dev == odev)
390 					call_rcu(&dev->rcu,
391 						 __dev_map_entry_free);
392 			}
393 		}
394 		rcu_read_unlock();
395 		break;
396 	default:
397 		break;
398 	}
399 	return NOTIFY_OK;
400 }
401 
402 static struct notifier_block dev_map_notifier = {
403 	.notifier_call = dev_map_notification,
404 };
405 
406 static int __init dev_map_init(void)
407 {
408 	register_netdevice_notifier(&dev_map_notifier);
409 	return 0;
410 }
411 
412 subsys_initcall(dev_map_init);
413