1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/bpf-cgroup.h>
3 #include <linux/bpf.h>
4 #include <linux/bpf_local_storage.h>
5 #include <linux/btf.h>
6 #include <linux/bug.h>
7 #include <linux/filter.h>
8 #include <linux/mm.h>
9 #include <linux/rbtree.h>
10 #include <linux/slab.h>
11 #include <uapi/linux/btf.h>
12 #include <linux/btf_ids.h>
13
14 #ifdef CONFIG_CGROUP_BPF
15
16 #include "../cgroup/cgroup-internal.h"
17
18 #define LOCAL_STORAGE_CREATE_FLAG_MASK \
19 (BPF_F_NUMA_NODE | BPF_F_ACCESS_MASK)
20
21 struct bpf_cgroup_storage_map {
22 struct bpf_map map;
23
24 spinlock_t lock;
25 struct rb_root root;
26 struct list_head list;
27 };
28
29 static struct bpf_cgroup_storage_map *map_to_storage(struct bpf_map *map)
30 {
31 return container_of(map, struct bpf_cgroup_storage_map, map);
32 }
33
34 static bool attach_type_isolated(const struct bpf_map *map)
35 {
36 return map->key_size == sizeof(struct bpf_cgroup_storage_key);
37 }
38
39 static int bpf_cgroup_storage_key_cmp(const struct bpf_cgroup_storage_map *map,
40 const void *_key1, const void *_key2)
41 {
42 if (attach_type_isolated(&map->map)) {
43 const struct bpf_cgroup_storage_key *key1 = _key1;
44 const struct bpf_cgroup_storage_key *key2 = _key2;
45
46 if (key1->cgroup_inode_id < key2->cgroup_inode_id)
47 return -1;
48 else if (key1->cgroup_inode_id > key2->cgroup_inode_id)
49 return 1;
50 else if (key1->attach_type < key2->attach_type)
51 return -1;
52 else if (key1->attach_type > key2->attach_type)
53 return 1;
54 } else {
55 const __u64 *cgroup_inode_id1 = _key1;
56 const __u64 *cgroup_inode_id2 = _key2;
57
58 if (*cgroup_inode_id1 < *cgroup_inode_id2)
59 return -1;
60 else if (*cgroup_inode_id1 > *cgroup_inode_id2)
61 return 1;
62 }
63 return 0;
64 }
65
66 struct bpf_cgroup_storage *
67 cgroup_storage_lookup(struct bpf_cgroup_storage_map *map,
68 void *key, bool locked)
69 {
70 struct rb_root *root = &map->root;
71 struct rb_node *node;
72
73 if (!locked)
74 spin_lock_bh(&map->lock);
75
76 node = root->rb_node;
77 while (node) {
78 struct bpf_cgroup_storage *storage;
79
80 storage = container_of(node, struct bpf_cgroup_storage, node);
81
82 switch (bpf_cgroup_storage_key_cmp(map, key, &storage->key)) {
83 case -1:
84 node = node->rb_left;
85 break;
86 case 1:
87 node = node->rb_right;
88 break;
89 default:
90 if (!locked)
91 spin_unlock_bh(&map->lock);
92 return storage;
93 }
94 }
95
96 if (!locked)
97 spin_unlock_bh(&map->lock);
98
99 return NULL;
100 }
101
102 static int cgroup_storage_insert(struct bpf_cgroup_storage_map *map,
103 struct bpf_cgroup_storage *storage)
104 {
105 struct rb_root *root = &map->root;
106 struct rb_node **new = &(root->rb_node), *parent = NULL;
107
108 while (*new) {
109 struct bpf_cgroup_storage *this;
110
111 this = container_of(*new, struct bpf_cgroup_storage, node);
112
113 parent = *new;
114 switch (bpf_cgroup_storage_key_cmp(map, &storage->key, &this->key)) {
115 case -1:
116 new = &((*new)->rb_left);
117 break;
118 case 1:
119 new = &((*new)->rb_right);
120 break;
121 default:
122 return -EEXIST;
123 }
124 }
125
126 rb_link_node(&storage->node, parent, new);
127 rb_insert_color(&storage->node, root);
128
129 return 0;
130 }
131
132 static void *cgroup_storage_lookup_elem(struct bpf_map *_map, void *key)
133 {
134 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
135 struct bpf_cgroup_storage *storage;
136
137 storage = cgroup_storage_lookup(map, key, false);
138 if (!storage)
139 return NULL;
140
141 return &READ_ONCE(storage->buf)->data[0];
142 }
143
144 static long cgroup_storage_update_elem(struct bpf_map *map, void *key,
145 void *value, u64 flags)
146 {
147 struct bpf_cgroup_storage *storage;
148 struct bpf_storage_buffer *new;
149
150 if (unlikely(flags & ~(BPF_F_LOCK | BPF_EXIST)))
151 return -EINVAL;
152
153 if (unlikely((flags & BPF_F_LOCK) &&
154 !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
155 return -EINVAL;
156
157 storage = cgroup_storage_lookup((struct bpf_cgroup_storage_map *)map,
158 key, false);
159 if (!storage)
160 return -ENOENT;
161
162 if (flags & BPF_F_LOCK) {
163 copy_map_value_locked(map, storage->buf->data, value, false);
164 return 0;
165 }
166
167 new = bpf_map_kmalloc_node(map, struct_size(new, data, map->value_size),
168 __GFP_ZERO | GFP_NOWAIT | __GFP_NOWARN,
169 map->numa_node);
170 if (!new)
171 return -ENOMEM;
172
173 memcpy(&new->data[0], value, map->value_size);
174 check_and_init_map_value(map, new->data);
175
176 new = xchg(&storage->buf, new);
177 kfree_rcu(new, rcu);
178
179 return 0;
180 }
181
182 int bpf_percpu_cgroup_storage_copy(struct bpf_map *_map, void *key,
183 void *value)
184 {
185 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
186 struct bpf_cgroup_storage *storage;
187 int cpu, off = 0;
188 u32 size;
189
190 rcu_read_lock();
191 storage = cgroup_storage_lookup(map, key, false);
192 if (!storage) {
193 rcu_read_unlock();
194 return -ENOENT;
195 }
196
197 /* per_cpu areas are zero-filled and bpf programs can only
198 * access 'value_size' of them, so copying rounded areas
199 * will not leak any kernel data
200 */
201 size = round_up(_map->value_size, 8);
202 for_each_possible_cpu(cpu) {
203 bpf_long_memcpy(value + off,
204 per_cpu_ptr(storage->percpu_buf, cpu), size);
205 off += size;
206 }
207 rcu_read_unlock();
208 return 0;
209 }
210
211 int bpf_percpu_cgroup_storage_update(struct bpf_map *_map, void *key,
212 void *value, u64 map_flags)
213 {
214 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
215 struct bpf_cgroup_storage *storage;
216 int cpu, off = 0;
217 u32 size;
218
219 if (map_flags != BPF_ANY && map_flags != BPF_EXIST)
220 return -EINVAL;
221
222 rcu_read_lock();
223 storage = cgroup_storage_lookup(map, key, false);
224 if (!storage) {
225 rcu_read_unlock();
226 return -ENOENT;
227 }
228
229 /* the user space will provide round_up(value_size, 8) bytes that
230 * will be copied into per-cpu area. bpf programs can only access
231 * value_size of it. During lookup the same extra bytes will be
232 * returned or zeros which were zero-filled by percpu_alloc,
233 * so no kernel data leaks possible
234 */
235 size = round_up(_map->value_size, 8);
236 for_each_possible_cpu(cpu) {
237 bpf_long_memcpy(per_cpu_ptr(storage->percpu_buf, cpu),
238 value + off, size);
239 off += size;
240 }
241 rcu_read_unlock();
242 return 0;
243 }
244
245 static int cgroup_storage_get_next_key(struct bpf_map *_map, void *key,
246 void *_next_key)
247 {
248 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
249 struct bpf_cgroup_storage *storage;
250
251 spin_lock_bh(&map->lock);
252
253 if (list_empty(&map->list))
254 goto enoent;
255
256 if (key) {
257 storage = cgroup_storage_lookup(map, key, true);
258 if (!storage)
259 goto enoent;
260
261 storage = list_next_entry(storage, list_map);
262 if (!storage)
263 goto enoent;
264 } else {
265 storage = list_first_entry(&map->list,
266 struct bpf_cgroup_storage, list_map);
267 }
268
269 spin_unlock_bh(&map->lock);
270
271 if (attach_type_isolated(&map->map)) {
272 struct bpf_cgroup_storage_key *next = _next_key;
273 *next = storage->key;
274 } else {
275 __u64 *next = _next_key;
276 *next = storage->key.cgroup_inode_id;
277 }
278 return 0;
279
280 enoent:
281 spin_unlock_bh(&map->lock);
282 return -ENOENT;
283 }
284
285 static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
286 {
287 __u32 max_value_size = BPF_LOCAL_STORAGE_MAX_VALUE_SIZE;
288 int numa_node = bpf_map_attr_numa_node(attr);
289 struct bpf_cgroup_storage_map *map;
290
291 /* percpu is bound by PCPU_MIN_UNIT_SIZE, non-percu
292 * is the same as other local storages.
293 */
294 if (attr->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
295 max_value_size = min_t(__u32, max_value_size,
296 PCPU_MIN_UNIT_SIZE);
297
298 if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
299 attr->key_size != sizeof(__u64))
300 return ERR_PTR(-EINVAL);
301
302 if (attr->value_size == 0)
303 return ERR_PTR(-EINVAL);
304
305 if (attr->value_size > max_value_size)
306 return ERR_PTR(-E2BIG);
307
308 if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
309 !bpf_map_flags_access_ok(attr->map_flags))
310 return ERR_PTR(-EINVAL);
311
312 if (attr->max_entries)
313 /* max_entries is not used and enforced to be 0 */
314 return ERR_PTR(-EINVAL);
315
316 map = bpf_map_area_alloc(sizeof(struct bpf_cgroup_storage_map), numa_node);
317 if (!map)
318 return ERR_PTR(-ENOMEM);
319
320 /* copy mandatory map attributes */
321 bpf_map_init_from_attr(&map->map, attr);
322
323 spin_lock_init(&map->lock);
324 map->root = RB_ROOT;
325 INIT_LIST_HEAD(&map->list);
326
327 return &map->map;
328 }
329
330 static void cgroup_storage_map_free(struct bpf_map *_map)
331 {
332 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
333 struct list_head *storages = &map->list;
334 struct bpf_cgroup_storage *storage, *stmp;
335
336 cgroup_lock();
337
338 list_for_each_entry_safe(storage, stmp, storages, list_map) {
339 bpf_cgroup_storage_unlink(storage);
340 bpf_cgroup_storage_free(storage);
341 }
342
343 cgroup_unlock();
344
345 WARN_ON(!RB_EMPTY_ROOT(&map->root));
346 WARN_ON(!list_empty(&map->list));
347
348 bpf_map_area_free(map);
349 }
350
351 static long cgroup_storage_delete_elem(struct bpf_map *map, void *key)
352 {
353 return -EINVAL;
354 }
355
356 static int cgroup_storage_check_btf(const struct bpf_map *map,
357 const struct btf *btf,
358 const struct btf_type *key_type,
359 const struct btf_type *value_type)
360 {
361 if (attach_type_isolated(map)) {
362 struct btf_member *m;
363 u32 offset, size;
364
365 /* Key is expected to be of struct bpf_cgroup_storage_key type,
366 * which is:
367 * struct bpf_cgroup_storage_key {
368 * __u64 cgroup_inode_id;
369 * __u32 attach_type;
370 * };
371 */
372
373 /*
374 * Key_type must be a structure with two fields.
375 */
376 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
377 BTF_INFO_VLEN(key_type->info) != 2)
378 return -EINVAL;
379
380 /*
381 * The first field must be a 64 bit integer at 0 offset.
382 */
383 m = (struct btf_member *)(key_type + 1);
384 size = sizeof_field(struct bpf_cgroup_storage_key, cgroup_inode_id);
385 if (!btf_member_is_reg_int(btf, key_type, m, 0, size))
386 return -EINVAL;
387
388 /*
389 * The second field must be a 32 bit integer at 64 bit offset.
390 */
391 m++;
392 offset = offsetof(struct bpf_cgroup_storage_key, attach_type);
393 size = sizeof_field(struct bpf_cgroup_storage_key, attach_type);
394 if (!btf_member_is_reg_int(btf, key_type, m, offset, size))
395 return -EINVAL;
396 } else {
397 u32 int_data;
398
399 /*
400 * Key is expected to be u64, which stores the cgroup_inode_id
401 */
402
403 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
404 return -EINVAL;
405
406 int_data = *(u32 *)(key_type + 1);
407 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
408 return -EINVAL;
409 }
410
411 return 0;
412 }
413
414 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
415 struct seq_file *m)
416 {
417 enum bpf_cgroup_storage_type stype;
418 struct bpf_cgroup_storage *storage;
419 int cpu;
420
421 rcu_read_lock();
422 storage = cgroup_storage_lookup(map_to_storage(map), key, false);
423 if (!storage) {
424 rcu_read_unlock();
425 return;
426 }
427
428 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
429 stype = cgroup_storage_type(map);
430 if (stype == BPF_CGROUP_STORAGE_SHARED) {
431 seq_puts(m, ": ");
432 btf_type_seq_show(map->btf, map->btf_value_type_id,
433 &READ_ONCE(storage->buf)->data[0], m);
434 seq_puts(m, "\n");
435 } else {
436 seq_puts(m, ": {\n");
437 for_each_possible_cpu(cpu) {
438 seq_printf(m, "\tcpu%d: ", cpu);
439 btf_type_seq_show(map->btf, map->btf_value_type_id,
440 per_cpu_ptr(storage->percpu_buf, cpu),
441 m);
442 seq_puts(m, "\n");
443 }
444 seq_puts(m, "}\n");
445 }
446 rcu_read_unlock();
447 }
448
449 static u64 cgroup_storage_map_usage(const struct bpf_map *map)
450 {
451 /* Currently the dynamically allocated elements are not counted. */
452 return sizeof(struct bpf_cgroup_storage_map);
453 }
454
455 BTF_ID_LIST_SINGLE(cgroup_storage_map_btf_ids, struct,
456 bpf_cgroup_storage_map)
457 const struct bpf_map_ops cgroup_storage_map_ops = {
458 .map_alloc = cgroup_storage_map_alloc,
459 .map_free = cgroup_storage_map_free,
460 .map_get_next_key = cgroup_storage_get_next_key,
461 .map_lookup_elem = cgroup_storage_lookup_elem,
462 .map_update_elem = cgroup_storage_update_elem,
463 .map_delete_elem = cgroup_storage_delete_elem,
464 .map_check_btf = cgroup_storage_check_btf,
465 .map_seq_show_elem = cgroup_storage_seq_show_elem,
466 .map_mem_usage = cgroup_storage_map_usage,
467 .map_btf_id = &cgroup_storage_map_btf_ids[0],
468 };
469
470 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
471 {
472 enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
473
474 if (aux->cgroup_storage[stype] &&
475 aux->cgroup_storage[stype] != _map)
476 return -EBUSY;
477
478 aux->cgroup_storage[stype] = _map;
479 return 0;
480 }
481
482 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
483 {
484 size_t size;
485
486 if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
487 size = sizeof(struct bpf_storage_buffer) + map->value_size;
488 *pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
489 PAGE_SIZE) >> PAGE_SHIFT;
490 } else {
491 size = map->value_size;
492 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
493 PAGE_SIZE) >> PAGE_SHIFT;
494 }
495
496 return size;
497 }
498
499 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
500 enum bpf_cgroup_storage_type stype)
501 {
502 const gfp_t gfp = __GFP_ZERO | GFP_USER;
503 struct bpf_cgroup_storage *storage;
504 struct bpf_map *map;
505 size_t size;
506 u32 pages;
507
508 map = prog->aux->cgroup_storage[stype];
509 if (!map)
510 return NULL;
511
512 size = bpf_cgroup_storage_calculate_size(map, &pages);
513
514 storage = bpf_map_kmalloc_node(map, sizeof(struct bpf_cgroup_storage),
515 gfp, map->numa_node);
516 if (!storage)
517 goto enomem;
518
519 if (stype == BPF_CGROUP_STORAGE_SHARED) {
520 storage->buf = bpf_map_kmalloc_node(map, size, gfp,
521 map->numa_node);
522 if (!storage->buf)
523 goto enomem;
524 check_and_init_map_value(map, storage->buf->data);
525 } else {
526 storage->percpu_buf = bpf_map_alloc_percpu(map, size, 8, gfp);
527 if (!storage->percpu_buf)
528 goto enomem;
529 }
530
531 storage->map = (struct bpf_cgroup_storage_map *)map;
532
533 return storage;
534
535 enomem:
536 kfree(storage);
537 return ERR_PTR(-ENOMEM);
538 }
539
540 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
541 {
542 struct bpf_cgroup_storage *storage =
543 container_of(rcu, struct bpf_cgroup_storage, rcu);
544
545 kfree(storage->buf);
546 kfree(storage);
547 }
548
549 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
550 {
551 struct bpf_cgroup_storage *storage =
552 container_of(rcu, struct bpf_cgroup_storage, rcu);
553
554 free_percpu(storage->percpu_buf);
555 kfree(storage);
556 }
557
558 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
559 {
560 enum bpf_cgroup_storage_type stype;
561 struct bpf_map *map;
562
563 if (!storage)
564 return;
565
566 map = &storage->map->map;
567 stype = cgroup_storage_type(map);
568 if (stype == BPF_CGROUP_STORAGE_SHARED)
569 call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
570 else
571 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
572 }
573
574 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
575 struct cgroup *cgroup,
576 enum bpf_attach_type type)
577 {
578 struct bpf_cgroup_storage_map *map;
579
580 if (!storage)
581 return;
582
583 storage->key.attach_type = type;
584 storage->key.cgroup_inode_id = cgroup_id(cgroup);
585
586 map = storage->map;
587
588 spin_lock_bh(&map->lock);
589 WARN_ON(cgroup_storage_insert(map, storage));
590 list_add(&storage->list_map, &map->list);
591 list_add(&storage->list_cg, &cgroup->bpf.storages);
592 spin_unlock_bh(&map->lock);
593 }
594
595 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
596 {
597 struct bpf_cgroup_storage_map *map;
598 struct rb_root *root;
599
600 if (!storage)
601 return;
602
603 map = storage->map;
604
605 spin_lock_bh(&map->lock);
606 root = &map->root;
607 rb_erase(&storage->node, root);
608
609 list_del(&storage->list_map);
610 list_del(&storage->list_cg);
611 spin_unlock_bh(&map->lock);
612 }
613
614 #endif
615
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.