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