1 //SPDX-License-Identifier: GPL-2.0
2 #include <linux/bpf-cgroup.h>
6 #include <linux/filter.h>
8 #include <linux/rbtree.h>
9 #include <linux/slab.h>
10 #include <uapi/linux/btf.h>
12 DEFINE_PER_CPU(struct bpf_cgroup_storage*, bpf_cgroup_storage[MAX_BPF_CGROUP_STORAGE_TYPE]);
14 #ifdef CONFIG_CGROUP_BPF
16 #include "../cgroup/cgroup-internal.h"
18 #define LOCAL_STORAGE_CREATE_FLAG_MASK \
19 (BPF_F_NUMA_NODE | BPF_F_ACCESS_MASK)
21 struct bpf_cgroup_storage_map {
26 struct list_head list;
29 static struct bpf_cgroup_storage_map *map_to_storage(struct bpf_map *map)
31 return container_of(map, struct bpf_cgroup_storage_map, map);
34 static bool attach_type_isolated(const struct bpf_map *map)
36 return map->key_size == sizeof(struct bpf_cgroup_storage_key);
39 static int bpf_cgroup_storage_key_cmp(const struct bpf_cgroup_storage_map *map,
40 const void *_key1, const void *_key2)
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;
46 if (key1->cgroup_inode_id < key2->cgroup_inode_id)
48 else if (key1->cgroup_inode_id > key2->cgroup_inode_id)
50 else if (key1->attach_type < key2->attach_type)
52 else if (key1->attach_type > key2->attach_type)
55 const __u64 *cgroup_inode_id1 = _key1;
56 const __u64 *cgroup_inode_id2 = _key2;
58 if (*cgroup_inode_id1 < *cgroup_inode_id2)
60 else if (*cgroup_inode_id1 > *cgroup_inode_id2)
66 struct bpf_cgroup_storage *
67 cgroup_storage_lookup(struct bpf_cgroup_storage_map *map,
68 void *key, bool locked)
70 struct rb_root *root = &map->root;
74 spin_lock_bh(&map->lock);
78 struct bpf_cgroup_storage *storage;
80 storage = container_of(node, struct bpf_cgroup_storage, node);
82 switch (bpf_cgroup_storage_key_cmp(map, key, &storage->key)) {
87 node = node->rb_right;
91 spin_unlock_bh(&map->lock);
97 spin_unlock_bh(&map->lock);
102 static int cgroup_storage_insert(struct bpf_cgroup_storage_map *map,
103 struct bpf_cgroup_storage *storage)
105 struct rb_root *root = &map->root;
106 struct rb_node **new = &(root->rb_node), *parent = NULL;
109 struct bpf_cgroup_storage *this;
111 this = container_of(*new, struct bpf_cgroup_storage, node);
114 switch (bpf_cgroup_storage_key_cmp(map, &storage->key, &this->key)) {
116 new = &((*new)->rb_left);
119 new = &((*new)->rb_right);
126 rb_link_node(&storage->node, parent, new);
127 rb_insert_color(&storage->node, root);
132 static void *cgroup_storage_lookup_elem(struct bpf_map *_map, void *key)
134 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
135 struct bpf_cgroup_storage *storage;
137 storage = cgroup_storage_lookup(map, key, false);
141 return &READ_ONCE(storage->buf)->data[0];
144 static int cgroup_storage_update_elem(struct bpf_map *map, void *key,
145 void *value, u64 flags)
147 struct bpf_cgroup_storage *storage;
148 struct bpf_storage_buffer *new;
150 if (unlikely(flags & ~(BPF_F_LOCK | BPF_EXIST)))
153 if (unlikely((flags & BPF_F_LOCK) &&
154 !map_value_has_spin_lock(map)))
157 storage = cgroup_storage_lookup((struct bpf_cgroup_storage_map *)map,
162 if (flags & BPF_F_LOCK) {
163 copy_map_value_locked(map, storage->buf->data, value, false);
167 new = bpf_map_kmalloc_node(map, sizeof(struct bpf_storage_buffer) +
169 __GFP_ZERO | GFP_ATOMIC | __GFP_NOWARN,
174 memcpy(&new->data[0], value, map->value_size);
175 check_and_init_map_lock(map, new->data);
177 new = xchg(&storage->buf, new);
183 int bpf_percpu_cgroup_storage_copy(struct bpf_map *_map, void *key,
186 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
187 struct bpf_cgroup_storage *storage;
192 storage = cgroup_storage_lookup(map, key, false);
198 /* per_cpu areas are zero-filled and bpf programs can only
199 * access 'value_size' of them, so copying rounded areas
200 * will not leak any kernel data
202 size = round_up(_map->value_size, 8);
203 for_each_possible_cpu(cpu) {
204 bpf_long_memcpy(value + off,
205 per_cpu_ptr(storage->percpu_buf, cpu), size);
212 int bpf_percpu_cgroup_storage_update(struct bpf_map *_map, void *key,
213 void *value, u64 map_flags)
215 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
216 struct bpf_cgroup_storage *storage;
220 if (map_flags != BPF_ANY && map_flags != BPF_EXIST)
224 storage = cgroup_storage_lookup(map, key, false);
230 /* the user space will provide round_up(value_size, 8) bytes that
231 * will be copied into per-cpu area. bpf programs can only access
232 * value_size of it. During lookup the same extra bytes will be
233 * returned or zeros which were zero-filled by percpu_alloc,
234 * so no kernel data leaks possible
236 size = round_up(_map->value_size, 8);
237 for_each_possible_cpu(cpu) {
238 bpf_long_memcpy(per_cpu_ptr(storage->percpu_buf, cpu),
246 static int cgroup_storage_get_next_key(struct bpf_map *_map, void *key,
249 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
250 struct bpf_cgroup_storage *storage;
252 spin_lock_bh(&map->lock);
254 if (list_empty(&map->list))
258 storage = cgroup_storage_lookup(map, key, true);
262 storage = list_next_entry(storage, list_map);
266 storage = list_first_entry(&map->list,
267 struct bpf_cgroup_storage, list_map);
270 spin_unlock_bh(&map->lock);
272 if (attach_type_isolated(&map->map)) {
273 struct bpf_cgroup_storage_key *next = _next_key;
274 *next = storage->key;
276 __u64 *next = _next_key;
277 *next = storage->key.cgroup_inode_id;
282 spin_unlock_bh(&map->lock);
286 static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
288 int numa_node = bpf_map_attr_numa_node(attr);
289 struct bpf_cgroup_storage_map *map;
291 if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
292 attr->key_size != sizeof(__u64))
293 return ERR_PTR(-EINVAL);
295 if (attr->value_size == 0)
296 return ERR_PTR(-EINVAL);
298 if (attr->value_size > PAGE_SIZE)
299 return ERR_PTR(-E2BIG);
301 if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
302 !bpf_map_flags_access_ok(attr->map_flags))
303 return ERR_PTR(-EINVAL);
305 if (attr->max_entries)
306 /* max_entries is not used and enforced to be 0 */
307 return ERR_PTR(-EINVAL);
309 map = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
310 __GFP_ZERO | GFP_USER | __GFP_ACCOUNT, numa_node);
312 return ERR_PTR(-ENOMEM);
314 /* copy mandatory map attributes */
315 bpf_map_init_from_attr(&map->map, attr);
317 spin_lock_init(&map->lock);
319 INIT_LIST_HEAD(&map->list);
324 static void cgroup_storage_map_free(struct bpf_map *_map)
326 struct bpf_cgroup_storage_map *map = map_to_storage(_map);
327 struct list_head *storages = &map->list;
328 struct bpf_cgroup_storage *storage, *stmp;
330 mutex_lock(&cgroup_mutex);
332 list_for_each_entry_safe(storage, stmp, storages, list_map) {
333 bpf_cgroup_storage_unlink(storage);
334 bpf_cgroup_storage_free(storage);
337 mutex_unlock(&cgroup_mutex);
339 WARN_ON(!RB_EMPTY_ROOT(&map->root));
340 WARN_ON(!list_empty(&map->list));
345 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
350 static int cgroup_storage_check_btf(const struct bpf_map *map,
351 const struct btf *btf,
352 const struct btf_type *key_type,
353 const struct btf_type *value_type)
355 if (attach_type_isolated(map)) {
356 struct btf_member *m;
359 /* Key is expected to be of struct bpf_cgroup_storage_key type,
361 * struct bpf_cgroup_storage_key {
362 * __u64 cgroup_inode_id;
368 * Key_type must be a structure with two fields.
370 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
371 BTF_INFO_VLEN(key_type->info) != 2)
375 * The first field must be a 64 bit integer at 0 offset.
377 m = (struct btf_member *)(key_type + 1);
378 size = sizeof_field(struct bpf_cgroup_storage_key, cgroup_inode_id);
379 if (!btf_member_is_reg_int(btf, key_type, m, 0, size))
383 * The second field must be a 32 bit integer at 64 bit offset.
386 offset = offsetof(struct bpf_cgroup_storage_key, attach_type);
387 size = sizeof_field(struct bpf_cgroup_storage_key, attach_type);
388 if (!btf_member_is_reg_int(btf, key_type, m, offset, size))
394 * Key is expected to be u64, which stores the cgroup_inode_id
397 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
400 int_data = *(u32 *)(key_type + 1);
401 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
408 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
411 enum bpf_cgroup_storage_type stype = cgroup_storage_type(map);
412 struct bpf_cgroup_storage *storage;
416 storage = cgroup_storage_lookup(map_to_storage(map), key, false);
422 btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
423 stype = cgroup_storage_type(map);
424 if (stype == BPF_CGROUP_STORAGE_SHARED) {
426 btf_type_seq_show(map->btf, map->btf_value_type_id,
427 &READ_ONCE(storage->buf)->data[0], m);
430 seq_puts(m, ": {\n");
431 for_each_possible_cpu(cpu) {
432 seq_printf(m, "\tcpu%d: ", cpu);
433 btf_type_seq_show(map->btf, map->btf_value_type_id,
434 per_cpu_ptr(storage->percpu_buf, cpu),
443 static int cgroup_storage_map_btf_id;
444 const struct bpf_map_ops cgroup_storage_map_ops = {
445 .map_alloc = cgroup_storage_map_alloc,
446 .map_free = cgroup_storage_map_free,
447 .map_get_next_key = cgroup_storage_get_next_key,
448 .map_lookup_elem = cgroup_storage_lookup_elem,
449 .map_update_elem = cgroup_storage_update_elem,
450 .map_delete_elem = cgroup_storage_delete_elem,
451 .map_check_btf = cgroup_storage_check_btf,
452 .map_seq_show_elem = cgroup_storage_seq_show_elem,
453 .map_btf_name = "bpf_cgroup_storage_map",
454 .map_btf_id = &cgroup_storage_map_btf_id,
457 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
459 enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
461 if (aux->cgroup_storage[stype] &&
462 aux->cgroup_storage[stype] != _map)
465 aux->cgroup_storage[stype] = _map;
469 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
473 if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
474 size = sizeof(struct bpf_storage_buffer) + map->value_size;
475 *pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
476 PAGE_SIZE) >> PAGE_SHIFT;
478 size = map->value_size;
479 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
480 PAGE_SIZE) >> PAGE_SHIFT;
486 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
487 enum bpf_cgroup_storage_type stype)
489 const gfp_t gfp = __GFP_ZERO | GFP_USER;
490 struct bpf_cgroup_storage *storage;
495 map = prog->aux->cgroup_storage[stype];
499 size = bpf_cgroup_storage_calculate_size(map, &pages);
501 storage = bpf_map_kmalloc_node(map, sizeof(struct bpf_cgroup_storage),
502 gfp, map->numa_node);
506 if (stype == BPF_CGROUP_STORAGE_SHARED) {
507 storage->buf = bpf_map_kmalloc_node(map, size, gfp,
511 check_and_init_map_lock(map, storage->buf->data);
513 storage->percpu_buf = bpf_map_alloc_percpu(map, size, 8, gfp);
514 if (!storage->percpu_buf)
518 storage->map = (struct bpf_cgroup_storage_map *)map;
524 return ERR_PTR(-ENOMEM);
527 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
529 struct bpf_cgroup_storage *storage =
530 container_of(rcu, struct bpf_cgroup_storage, rcu);
536 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
538 struct bpf_cgroup_storage *storage =
539 container_of(rcu, struct bpf_cgroup_storage, rcu);
541 free_percpu(storage->percpu_buf);
545 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
547 enum bpf_cgroup_storage_type stype;
553 map = &storage->map->map;
554 stype = cgroup_storage_type(map);
555 if (stype == BPF_CGROUP_STORAGE_SHARED)
556 call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
558 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
561 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
562 struct cgroup *cgroup,
563 enum bpf_attach_type type)
565 struct bpf_cgroup_storage_map *map;
570 storage->key.attach_type = type;
571 storage->key.cgroup_inode_id = cgroup_id(cgroup);
575 spin_lock_bh(&map->lock);
576 WARN_ON(cgroup_storage_insert(map, storage));
577 list_add(&storage->list_map, &map->list);
578 list_add(&storage->list_cg, &cgroup->bpf.storages);
579 spin_unlock_bh(&map->lock);
582 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
584 struct bpf_cgroup_storage_map *map;
585 struct rb_root *root;
592 spin_lock_bh(&map->lock);
594 rb_erase(&storage->node, root);
596 list_del(&storage->list_map);
597 list_del(&storage->list_cg);
598 spin_unlock_bh(&map->lock);