xref: /linux/kernel/bpf/local_storage.c (revision 02ff58dcf70ad7d11b01523dc404166ed11021da)
1 //SPDX-License-Identifier: GPL-2.0
2 #include <linux/bpf-cgroup.h>
3 #include <linux/bpf.h>
4 #include <linux/btf.h>
5 #include <linux/bug.h>
6 #include <linux/filter.h>
7 #include <linux/mm.h>
8 #include <linux/rbtree.h>
9 #include <linux/slab.h>
10 #include <uapi/linux/btf.h>
11 
12 DEFINE_PER_CPU(struct bpf_cgroup_storage*, bpf_cgroup_storage[MAX_BPF_CGROUP_STORAGE_TYPE]);
13 
14 #ifdef CONFIG_CGROUP_BPF
15 
16 #define LOCAL_STORAGE_CREATE_FLAG_MASK					\
17 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
18 
19 struct bpf_cgroup_storage_map {
20 	struct bpf_map map;
21 
22 	spinlock_t lock;
23 	struct bpf_prog *prog;
24 	struct rb_root root;
25 	struct list_head list;
26 };
27 
28 static struct bpf_cgroup_storage_map *map_to_storage(struct bpf_map *map)
29 {
30 	return container_of(map, struct bpf_cgroup_storage_map, map);
31 }
32 
33 static int bpf_cgroup_storage_key_cmp(
34 	const struct bpf_cgroup_storage_key *key1,
35 	const struct bpf_cgroup_storage_key *key2)
36 {
37 	if (key1->cgroup_inode_id < key2->cgroup_inode_id)
38 		return -1;
39 	else if (key1->cgroup_inode_id > key2->cgroup_inode_id)
40 		return 1;
41 	else if (key1->attach_type < key2->attach_type)
42 		return -1;
43 	else if (key1->attach_type > key2->attach_type)
44 		return 1;
45 	return 0;
46 }
47 
48 static struct bpf_cgroup_storage *cgroup_storage_lookup(
49 	struct bpf_cgroup_storage_map *map, struct bpf_cgroup_storage_key *key,
50 	bool locked)
51 {
52 	struct rb_root *root = &map->root;
53 	struct rb_node *node;
54 
55 	if (!locked)
56 		spin_lock_bh(&map->lock);
57 
58 	node = root->rb_node;
59 	while (node) {
60 		struct bpf_cgroup_storage *storage;
61 
62 		storage = container_of(node, struct bpf_cgroup_storage, node);
63 
64 		switch (bpf_cgroup_storage_key_cmp(key, &storage->key)) {
65 		case -1:
66 			node = node->rb_left;
67 			break;
68 		case 1:
69 			node = node->rb_right;
70 			break;
71 		default:
72 			if (!locked)
73 				spin_unlock_bh(&map->lock);
74 			return storage;
75 		}
76 	}
77 
78 	if (!locked)
79 		spin_unlock_bh(&map->lock);
80 
81 	return NULL;
82 }
83 
84 static int cgroup_storage_insert(struct bpf_cgroup_storage_map *map,
85 				 struct bpf_cgroup_storage *storage)
86 {
87 	struct rb_root *root = &map->root;
88 	struct rb_node **new = &(root->rb_node), *parent = NULL;
89 
90 	while (*new) {
91 		struct bpf_cgroup_storage *this;
92 
93 		this = container_of(*new, struct bpf_cgroup_storage, node);
94 
95 		parent = *new;
96 		switch (bpf_cgroup_storage_key_cmp(&storage->key, &this->key)) {
97 		case -1:
98 			new = &((*new)->rb_left);
99 			break;
100 		case 1:
101 			new = &((*new)->rb_right);
102 			break;
103 		default:
104 			return -EEXIST;
105 		}
106 	}
107 
108 	rb_link_node(&storage->node, parent, new);
109 	rb_insert_color(&storage->node, root);
110 
111 	return 0;
112 }
113 
114 static void *cgroup_storage_lookup_elem(struct bpf_map *_map, void *_key)
115 {
116 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
117 	struct bpf_cgroup_storage_key *key = _key;
118 	struct bpf_cgroup_storage *storage;
119 
120 	storage = cgroup_storage_lookup(map, key, false);
121 	if (!storage)
122 		return NULL;
123 
124 	return &READ_ONCE(storage->buf)->data[0];
125 }
126 
127 static int cgroup_storage_update_elem(struct bpf_map *map, void *_key,
128 				      void *value, u64 flags)
129 {
130 	struct bpf_cgroup_storage_key *key = _key;
131 	struct bpf_cgroup_storage *storage;
132 	struct bpf_storage_buffer *new;
133 
134 	if (flags != BPF_ANY && flags != BPF_EXIST)
135 		return -EINVAL;
136 
137 	storage = cgroup_storage_lookup((struct bpf_cgroup_storage_map *)map,
138 					key, false);
139 	if (!storage)
140 		return -ENOENT;
141 
142 	new = kmalloc_node(sizeof(struct bpf_storage_buffer) +
143 			   map->value_size,
144 			   __GFP_ZERO | GFP_ATOMIC | __GFP_NOWARN,
145 			   map->numa_node);
146 	if (!new)
147 		return -ENOMEM;
148 
149 	memcpy(&new->data[0], value, map->value_size);
150 
151 	new = xchg(&storage->buf, new);
152 	kfree_rcu(new, rcu);
153 
154 	return 0;
155 }
156 
157 int bpf_percpu_cgroup_storage_copy(struct bpf_map *_map, void *_key,
158 				   void *value)
159 {
160 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
161 	struct bpf_cgroup_storage_key *key = _key;
162 	struct bpf_cgroup_storage *storage;
163 	int cpu, off = 0;
164 	u32 size;
165 
166 	rcu_read_lock();
167 	storage = cgroup_storage_lookup(map, key, false);
168 	if (!storage) {
169 		rcu_read_unlock();
170 		return -ENOENT;
171 	}
172 
173 	/* per_cpu areas are zero-filled and bpf programs can only
174 	 * access 'value_size' of them, so copying rounded areas
175 	 * will not leak any kernel data
176 	 */
177 	size = round_up(_map->value_size, 8);
178 	for_each_possible_cpu(cpu) {
179 		bpf_long_memcpy(value + off,
180 				per_cpu_ptr(storage->percpu_buf, cpu), size);
181 		off += size;
182 	}
183 	rcu_read_unlock();
184 	return 0;
185 }
186 
187 int bpf_percpu_cgroup_storage_update(struct bpf_map *_map, void *_key,
188 				     void *value, u64 map_flags)
189 {
190 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
191 	struct bpf_cgroup_storage_key *key = _key;
192 	struct bpf_cgroup_storage *storage;
193 	int cpu, off = 0;
194 	u32 size;
195 
196 	if (map_flags != BPF_ANY && map_flags != BPF_EXIST)
197 		return -EINVAL;
198 
199 	rcu_read_lock();
200 	storage = cgroup_storage_lookup(map, key, false);
201 	if (!storage) {
202 		rcu_read_unlock();
203 		return -ENOENT;
204 	}
205 
206 	/* the user space will provide round_up(value_size, 8) bytes that
207 	 * will be copied into per-cpu area. bpf programs can only access
208 	 * value_size of it. During lookup the same extra bytes will be
209 	 * returned or zeros which were zero-filled by percpu_alloc,
210 	 * so no kernel data leaks possible
211 	 */
212 	size = round_up(_map->value_size, 8);
213 	for_each_possible_cpu(cpu) {
214 		bpf_long_memcpy(per_cpu_ptr(storage->percpu_buf, cpu),
215 				value + off, size);
216 		off += size;
217 	}
218 	rcu_read_unlock();
219 	return 0;
220 }
221 
222 static int cgroup_storage_get_next_key(struct bpf_map *_map, void *_key,
223 				       void *_next_key)
224 {
225 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
226 	struct bpf_cgroup_storage_key *key = _key;
227 	struct bpf_cgroup_storage_key *next = _next_key;
228 	struct bpf_cgroup_storage *storage;
229 
230 	spin_lock_bh(&map->lock);
231 
232 	if (list_empty(&map->list))
233 		goto enoent;
234 
235 	if (key) {
236 		storage = cgroup_storage_lookup(map, key, true);
237 		if (!storage)
238 			goto enoent;
239 
240 		storage = list_next_entry(storage, list);
241 		if (!storage)
242 			goto enoent;
243 	} else {
244 		storage = list_first_entry(&map->list,
245 					 struct bpf_cgroup_storage, list);
246 	}
247 
248 	spin_unlock_bh(&map->lock);
249 	next->attach_type = storage->key.attach_type;
250 	next->cgroup_inode_id = storage->key.cgroup_inode_id;
251 	return 0;
252 
253 enoent:
254 	spin_unlock_bh(&map->lock);
255 	return -ENOENT;
256 }
257 
258 static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
259 {
260 	int numa_node = bpf_map_attr_numa_node(attr);
261 	struct bpf_cgroup_storage_map *map;
262 
263 	if (attr->key_size != sizeof(struct bpf_cgroup_storage_key))
264 		return ERR_PTR(-EINVAL);
265 
266 	if (attr->value_size == 0)
267 		return ERR_PTR(-EINVAL);
268 
269 	if (attr->value_size > PAGE_SIZE)
270 		return ERR_PTR(-E2BIG);
271 
272 	if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK)
273 		/* reserved bits should not be used */
274 		return ERR_PTR(-EINVAL);
275 
276 	if (attr->max_entries)
277 		/* max_entries is not used and enforced to be 0 */
278 		return ERR_PTR(-EINVAL);
279 
280 	map = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
281 			   __GFP_ZERO | GFP_USER, numa_node);
282 	if (!map)
283 		return ERR_PTR(-ENOMEM);
284 
285 	map->map.pages = round_up(sizeof(struct bpf_cgroup_storage_map),
286 				  PAGE_SIZE) >> PAGE_SHIFT;
287 
288 	/* copy mandatory map attributes */
289 	bpf_map_init_from_attr(&map->map, attr);
290 
291 	spin_lock_init(&map->lock);
292 	map->root = RB_ROOT;
293 	INIT_LIST_HEAD(&map->list);
294 
295 	return &map->map;
296 }
297 
298 static void cgroup_storage_map_free(struct bpf_map *_map)
299 {
300 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
301 
302 	WARN_ON(!RB_EMPTY_ROOT(&map->root));
303 	WARN_ON(!list_empty(&map->list));
304 
305 	kfree(map);
306 }
307 
308 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
309 {
310 	return -EINVAL;
311 }
312 
313 static int cgroup_storage_check_btf(const struct bpf_map *map,
314 				    const struct btf *btf,
315 				    const struct btf_type *key_type,
316 				    const struct btf_type *value_type)
317 {
318 	const struct btf_type *t;
319 	struct btf_member *m;
320 	u32 id, size;
321 
322 	/* Key is expected to be of struct bpf_cgroup_storage_key type,
323 	 * which is:
324 	 * struct bpf_cgroup_storage_key {
325 	 *	__u64	cgroup_inode_id;
326 	 *	__u32	attach_type;
327 	 * };
328 	 */
329 
330 	/*
331 	 * Key_type must be a structure with two fields.
332 	 */
333 	if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
334 	    BTF_INFO_VLEN(key_type->info) != 2)
335 		return -EINVAL;
336 
337 	/*
338 	 * The first field must be a 64 bit integer at 0 offset.
339 	 */
340 	m = (struct btf_member *)(key_type + 1);
341 	if (m->offset)
342 		return -EINVAL;
343 	id = m->type;
344 	t = btf_type_id_size(btf, &id, NULL);
345 	size = FIELD_SIZEOF(struct bpf_cgroup_storage_key, cgroup_inode_id);
346 	if (!t || !btf_type_is_reg_int(t, size))
347 		return -EINVAL;
348 
349 	/*
350 	 * The second field must be a 32 bit integer at 64 bit offset.
351 	 */
352 	m++;
353 	if (m->offset != offsetof(struct bpf_cgroup_storage_key, attach_type) *
354 	    BITS_PER_BYTE)
355 		return -EINVAL;
356 	id = m->type;
357 	t = btf_type_id_size(btf, &id, NULL);
358 	size = FIELD_SIZEOF(struct bpf_cgroup_storage_key, attach_type);
359 	if (!t || !btf_type_is_reg_int(t, size))
360 		return -EINVAL;
361 
362 	return 0;
363 }
364 
365 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *_key,
366 					 struct seq_file *m)
367 {
368 	enum bpf_cgroup_storage_type stype = cgroup_storage_type(map);
369 	struct bpf_cgroup_storage_key *key = _key;
370 	struct bpf_cgroup_storage *storage;
371 	int cpu;
372 
373 	rcu_read_lock();
374 	storage = cgroup_storage_lookup(map_to_storage(map), key, false);
375 	if (!storage) {
376 		rcu_read_unlock();
377 		return;
378 	}
379 
380 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
381 	stype = cgroup_storage_type(map);
382 	if (stype == BPF_CGROUP_STORAGE_SHARED) {
383 		seq_puts(m, ": ");
384 		btf_type_seq_show(map->btf, map->btf_value_type_id,
385 				  &READ_ONCE(storage->buf)->data[0], m);
386 		seq_puts(m, "\n");
387 	} else {
388 		seq_puts(m, ": {\n");
389 		for_each_possible_cpu(cpu) {
390 			seq_printf(m, "\tcpu%d: ", cpu);
391 			btf_type_seq_show(map->btf, map->btf_value_type_id,
392 					  per_cpu_ptr(storage->percpu_buf, cpu),
393 					  m);
394 			seq_puts(m, "\n");
395 		}
396 		seq_puts(m, "}\n");
397 	}
398 	rcu_read_unlock();
399 }
400 
401 const struct bpf_map_ops cgroup_storage_map_ops = {
402 	.map_alloc = cgroup_storage_map_alloc,
403 	.map_free = cgroup_storage_map_free,
404 	.map_get_next_key = cgroup_storage_get_next_key,
405 	.map_lookup_elem = cgroup_storage_lookup_elem,
406 	.map_update_elem = cgroup_storage_update_elem,
407 	.map_delete_elem = cgroup_storage_delete_elem,
408 	.map_check_btf = cgroup_storage_check_btf,
409 	.map_seq_show_elem = cgroup_storage_seq_show_elem,
410 };
411 
412 int bpf_cgroup_storage_assign(struct bpf_prog *prog, struct bpf_map *_map)
413 {
414 	enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
415 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
416 	int ret = -EBUSY;
417 
418 	spin_lock_bh(&map->lock);
419 
420 	if (map->prog && map->prog != prog)
421 		goto unlock;
422 	if (prog->aux->cgroup_storage[stype] &&
423 	    prog->aux->cgroup_storage[stype] != _map)
424 		goto unlock;
425 
426 	map->prog = prog;
427 	prog->aux->cgroup_storage[stype] = _map;
428 	ret = 0;
429 unlock:
430 	spin_unlock_bh(&map->lock);
431 
432 	return ret;
433 }
434 
435 void bpf_cgroup_storage_release(struct bpf_prog *prog, struct bpf_map *_map)
436 {
437 	enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
438 	struct bpf_cgroup_storage_map *map = map_to_storage(_map);
439 
440 	spin_lock_bh(&map->lock);
441 	if (map->prog == prog) {
442 		WARN_ON(prog->aux->cgroup_storage[stype] != _map);
443 		map->prog = NULL;
444 		prog->aux->cgroup_storage[stype] = NULL;
445 	}
446 	spin_unlock_bh(&map->lock);
447 }
448 
449 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
450 {
451 	size_t size;
452 
453 	if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
454 		size = sizeof(struct bpf_storage_buffer) + map->value_size;
455 		*pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
456 				  PAGE_SIZE) >> PAGE_SHIFT;
457 	} else {
458 		size = map->value_size;
459 		*pages = round_up(round_up(size, 8) * num_possible_cpus(),
460 				  PAGE_SIZE) >> PAGE_SHIFT;
461 	}
462 
463 	return size;
464 }
465 
466 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
467 					enum bpf_cgroup_storage_type stype)
468 {
469 	struct bpf_cgroup_storage *storage;
470 	struct bpf_map *map;
471 	gfp_t flags;
472 	size_t size;
473 	u32 pages;
474 
475 	map = prog->aux->cgroup_storage[stype];
476 	if (!map)
477 		return NULL;
478 
479 	size = bpf_cgroup_storage_calculate_size(map, &pages);
480 
481 	if (bpf_map_charge_memlock(map, pages))
482 		return ERR_PTR(-EPERM);
483 
484 	storage = kmalloc_node(sizeof(struct bpf_cgroup_storage),
485 			       __GFP_ZERO | GFP_USER, map->numa_node);
486 	if (!storage)
487 		goto enomem;
488 
489 	flags = __GFP_ZERO | GFP_USER;
490 
491 	if (stype == BPF_CGROUP_STORAGE_SHARED) {
492 		storage->buf = kmalloc_node(size, flags, map->numa_node);
493 		if (!storage->buf)
494 			goto enomem;
495 	} else {
496 		storage->percpu_buf = __alloc_percpu_gfp(size, 8, flags);
497 		if (!storage->percpu_buf)
498 			goto enomem;
499 	}
500 
501 	storage->map = (struct bpf_cgroup_storage_map *)map;
502 
503 	return storage;
504 
505 enomem:
506 	bpf_map_uncharge_memlock(map, pages);
507 	kfree(storage);
508 	return ERR_PTR(-ENOMEM);
509 }
510 
511 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
512 {
513 	struct bpf_cgroup_storage *storage =
514 		container_of(rcu, struct bpf_cgroup_storage, rcu);
515 
516 	kfree(storage->buf);
517 	kfree(storage);
518 }
519 
520 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
521 {
522 	struct bpf_cgroup_storage *storage =
523 		container_of(rcu, struct bpf_cgroup_storage, rcu);
524 
525 	free_percpu(storage->percpu_buf);
526 	kfree(storage);
527 }
528 
529 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
530 {
531 	enum bpf_cgroup_storage_type stype;
532 	struct bpf_map *map;
533 	u32 pages;
534 
535 	if (!storage)
536 		return;
537 
538 	map = &storage->map->map;
539 
540 	bpf_cgroup_storage_calculate_size(map, &pages);
541 	bpf_map_uncharge_memlock(map, pages);
542 
543 	stype = cgroup_storage_type(map);
544 	if (stype == BPF_CGROUP_STORAGE_SHARED)
545 		call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
546 	else
547 		call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
548 }
549 
550 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
551 			     struct cgroup *cgroup,
552 			     enum bpf_attach_type type)
553 {
554 	struct bpf_cgroup_storage_map *map;
555 
556 	if (!storage)
557 		return;
558 
559 	storage->key.attach_type = type;
560 	storage->key.cgroup_inode_id = cgroup->kn->id.id;
561 
562 	map = storage->map;
563 
564 	spin_lock_bh(&map->lock);
565 	WARN_ON(cgroup_storage_insert(map, storage));
566 	list_add(&storage->list, &map->list);
567 	spin_unlock_bh(&map->lock);
568 }
569 
570 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
571 {
572 	struct bpf_cgroup_storage_map *map;
573 	struct rb_root *root;
574 
575 	if (!storage)
576 		return;
577 
578 	map = storage->map;
579 
580 	spin_lock_bh(&map->lock);
581 	root = &map->root;
582 	rb_erase(&storage->node, root);
583 
584 	list_del(&storage->list);
585 	spin_unlock_bh(&map->lock);
586 }
587 
588 #endif
589