kernel/bpf/local_storage.c

   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 #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 int 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                      !map_value_has_spin_lock(map)))
 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 = kmalloc_node(sizeof(struct bpf_storage_buffer) +
 168                            map->value_size,
 169                            __GFP_ZERO | GFP_ATOMIC | __GFP_NOWARN,
 170                            map->numa_node);
 171         if (!new)
 172                 return -ENOMEM;
 173
 174         memcpy(&new->data[0], value, map->value_size);
 175         check_and_init_map_lock(map, new->data);
 176
 177         new = xchg(&storage->buf, new);
 178         kfree_rcu(new, rcu);
 179
 180         return 0;
 181 }
 182
 183 int bpf_percpu_cgroup_storage_copy(struct bpf_map *_map, void *key,
 184                                    void *value)
 185 {
 186         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 187         struct bpf_cgroup_storage *storage;
 188         int cpu, off = 0;
 189         u32 size;
 190
 191         rcu_read_lock();
 192         storage = cgroup_storage_lookup(map, key, false);
 193         if (!storage) {
 194                 rcu_read_unlock();
 195                 return -ENOENT;
 196         }
 197
 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
 201          */
 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);
 206                 off += size;
 207         }
 208         rcu_read_unlock();
 209         return 0;
 210 }
 211
 212 int bpf_percpu_cgroup_storage_update(struct bpf_map *_map, void *key,
 213                                      void *value, u64 map_flags)
 214 {
 215         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 216         struct bpf_cgroup_storage *storage;
 217         int cpu, off = 0;
 218         u32 size;
 219
 220         if (map_flags != BPF_ANY && map_flags != BPF_EXIST)
 221                 return -EINVAL;
 222
 223         rcu_read_lock();
 224         storage = cgroup_storage_lookup(map, key, false);
 225         if (!storage) {
 226                 rcu_read_unlock();
 227                 return -ENOENT;
 228         }
 229
 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
 235          */
 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),
 239                                 value + off, size);
 240                 off += size;
 241         }
 242         rcu_read_unlock();
 243         return 0;
 244 }
 245
 246 static int cgroup_storage_get_next_key(struct bpf_map *_map, void *key,
 247                                        void *_next_key)
 248 {
 249         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 250         struct bpf_cgroup_storage *storage;
 251
 252         spin_lock_bh(&map->lock);
 253
 254         if (list_empty(&map->list))
 255                 goto enoent;
 256
 257         if (key) {
 258                 storage = cgroup_storage_lookup(map, key, true);
 259                 if (!storage)
 260                         goto enoent;
 261
 262                 storage = list_next_entry(storage, list_map);
 263                 if (!storage)
 264                         goto enoent;
 265         } else {
 266                 storage = list_first_entry(&map->list,
 267                                          struct bpf_cgroup_storage, list_map);
 268         }
 269
 270         spin_unlock_bh(&map->lock);
 271
 272         if (attach_type_isolated(&map->map)) {
 273                 struct bpf_cgroup_storage_key *next = _next_key;
 274                 *next = storage->key;
 275         } else {
 276                 __u64 *next = _next_key;
 277                 *next = storage->key.cgroup_inode_id;
 278         }
 279         return 0;
 280
 281 enoent:
 282         spin_unlock_bh(&map->lock);
 283         return -ENOENT;
 284 }
 285
 286 static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
 287 {
 288         int numa_node = bpf_map_attr_numa_node(attr);
 289         struct bpf_cgroup_storage_map *map;
 290         struct bpf_map_memory mem;
 291         int ret;
 292
 293         if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
 294             attr->key_size != sizeof(__u64))
 295                 return ERR_PTR(-EINVAL);
 296
 297         if (attr->value_size == 0)
 298                 return ERR_PTR(-EINVAL);
 299
 300         if (attr->value_size > PAGE_SIZE)
 301                 return ERR_PTR(-E2BIG);
 302
 303         if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
 304             !bpf_map_flags_access_ok(attr->map_flags))
 305                 return ERR_PTR(-EINVAL);
 306
 307         if (attr->max_entries)
 308                 /* max_entries is not used and enforced to be 0 */
 309                 return ERR_PTR(-EINVAL);
 310
 311         ret = bpf_map_charge_init(&mem, sizeof(struct bpf_cgroup_storage_map));
 312         if (ret < 0)
 313                 return ERR_PTR(ret);
 314
 315         map = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
 316                            __GFP_ZERO | GFP_USER, numa_node);
 317         if (!map) {
 318                 bpf_map_charge_finish(&mem);
 319                 return ERR_PTR(-ENOMEM);
 320         }
 321
 322         bpf_map_charge_move(&map->map.memory, &mem);
 323
 324         /* copy mandatory map attributes */
 325         bpf_map_init_from_attr(&map->map, attr);
 326
 327         spin_lock_init(&map->lock);
 328         map->root = RB_ROOT;
 329         INIT_LIST_HEAD(&map->list);
 330
 331         return &map->map;
 332 }
 333
 334 static void cgroup_storage_map_free(struct bpf_map *_map)
 335 {
 336         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 337         struct list_head *storages = &map->list;
 338         struct bpf_cgroup_storage *storage, *stmp;
 339
 340         mutex_lock(&cgroup_mutex);
 341
 342         list_for_each_entry_safe(storage, stmp, storages, list_map) {
 343                 bpf_cgroup_storage_unlink(storage);
 344                 bpf_cgroup_storage_free(storage);
 345         }
 346
 347         mutex_unlock(&cgroup_mutex);
 348
 349         WARN_ON(!RB_EMPTY_ROOT(&map->root));
 350         WARN_ON(!list_empty(&map->list));
 351
 352         kfree(map);
 353 }
 354
 355 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
 356 {
 357         return -EINVAL;
 358 }
 359
 360 static int cgroup_storage_check_btf(const struct bpf_map *map,
 361                                     const struct btf *btf,
 362                                     const struct btf_type *key_type,
 363                                     const struct btf_type *value_type)
 364 {
 365         if (attach_type_isolated(map)) {
 366                 struct btf_member *m;
 367                 u32 offset, size;
 368
 369                 /* Key is expected to be of struct bpf_cgroup_storage_key type,
 370                  * which is:
 371                  * struct bpf_cgroup_storage_key {
 372                  *      __u64   cgroup_inode_id;
 373                  *      __u32   attach_type;
 374                  * };
 375                  */
 376
 377                 /*
 378                  * Key_type must be a structure with two fields.
 379                  */
 380                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
 381                     BTF_INFO_VLEN(key_type->info) != 2)
 382                         return -EINVAL;
 383
 384                 /*
 385                  * The first field must be a 64 bit integer at 0 offset.
 386                  */
 387                 m = (struct btf_member *)(key_type + 1);
 388                 size = sizeof_field(struct bpf_cgroup_storage_key, cgroup_inode_id);
 389                 if (!btf_member_is_reg_int(btf, key_type, m, 0, size))
 390                         return -EINVAL;
 391
 392                 /*
 393                  * The second field must be a 32 bit integer at 64 bit offset.
 394                  */
 395                 m++;
 396                 offset = offsetof(struct bpf_cgroup_storage_key, attach_type);
 397                 size = sizeof_field(struct bpf_cgroup_storage_key, attach_type);
 398                 if (!btf_member_is_reg_int(btf, key_type, m, offset, size))
 399                         return -EINVAL;
 400         } else {
 401                 u32 int_data;
 402
 403                 /*
 404                  * Key is expected to be u64, which stores the cgroup_inode_id
 405                  */
 406
 407                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 408                         return -EINVAL;
 409
 410                 int_data = *(u32 *)(key_type + 1);
 411                 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
 412                         return -EINVAL;
 413         }
 414
 415         return 0;
 416 }
 417
 418 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
 419                                          struct seq_file *m)
 420 {
 421         enum bpf_cgroup_storage_type stype = cgroup_storage_type(map);
 422         struct bpf_cgroup_storage *storage;
 423         int cpu;
 424
 425         rcu_read_lock();
 426         storage = cgroup_storage_lookup(map_to_storage(map), key, false);
 427         if (!storage) {
 428                 rcu_read_unlock();
 429                 return;
 430         }
 431
 432         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
 433         stype = cgroup_storage_type(map);
 434         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 435                 seq_puts(m, ": ");
 436                 btf_type_seq_show(map->btf, map->btf_value_type_id,
 437                                   &READ_ONCE(storage->buf)->data[0], m);
 438                 seq_puts(m, "\n");
 439         } else {
 440                 seq_puts(m, ": {\n");
 441                 for_each_possible_cpu(cpu) {
 442                         seq_printf(m, "\tcpu%d: ", cpu);
 443                         btf_type_seq_show(map->btf, map->btf_value_type_id,
 444                                           per_cpu_ptr(storage->percpu_buf, cpu),
 445                                           m);
 446                         seq_puts(m, "\n");
 447                 }
 448                 seq_puts(m, "}\n");
 449         }
 450         rcu_read_unlock();
 451 }
 452
 453 static int cgroup_storage_map_btf_id;
 454 const struct bpf_map_ops cgroup_storage_map_ops = {
 455         .map_alloc = cgroup_storage_map_alloc,
 456         .map_free = cgroup_storage_map_free,
 457         .map_get_next_key = cgroup_storage_get_next_key,
 458         .map_lookup_elem = cgroup_storage_lookup_elem,
 459         .map_update_elem = cgroup_storage_update_elem,
 460         .map_delete_elem = cgroup_storage_delete_elem,
 461         .map_check_btf = cgroup_storage_check_btf,
 462         .map_seq_show_elem = cgroup_storage_seq_show_elem,
 463         .map_btf_name = "bpf_cgroup_storage_map",
 464         .map_btf_id = &cgroup_storage_map_btf_id,
 465 };
 466
 467 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
 468 {
 469         enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
 470
 471         if (aux->cgroup_storage[stype] &&
 472             aux->cgroup_storage[stype] != _map)
 473                 return -EBUSY;
 474
 475         aux->cgroup_storage[stype] = _map;
 476         return 0;
 477 }
 478
 479 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
 480 {
 481         size_t size;
 482
 483         if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
 484                 size = sizeof(struct bpf_storage_buffer) + map->value_size;
 485                 *pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
 486                                   PAGE_SIZE) >> PAGE_SHIFT;
 487         } else {
 488                 size = map->value_size;
 489                 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
 490                                   PAGE_SIZE) >> PAGE_SHIFT;
 491         }
 492
 493         return size;
 494 }
 495
 496 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
 497                                         enum bpf_cgroup_storage_type stype)
 498 {
 499         struct bpf_cgroup_storage *storage;
 500         struct bpf_map *map;
 501         gfp_t flags;
 502         size_t size;
 503         u32 pages;
 504
 505         map = prog->aux->cgroup_storage[stype];
 506         if (!map)
 507                 return NULL;
 508
 509         size = bpf_cgroup_storage_calculate_size(map, &pages);
 510
 511         if (bpf_map_charge_memlock(map, pages))
 512                 return ERR_PTR(-EPERM);
 513
 514         storage = kmalloc_node(sizeof(struct bpf_cgroup_storage),
 515                                __GFP_ZERO | GFP_USER, map->numa_node);
 516         if (!storage)
 517                 goto enomem;
 518
 519         flags = __GFP_ZERO | GFP_USER;
 520
 521         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 522                 storage->buf = kmalloc_node(size, flags, map->numa_node);
 523                 if (!storage->buf)
 524                         goto enomem;
 525                 check_and_init_map_lock(map, storage->buf->data);
 526         } else {
 527                 storage->percpu_buf = __alloc_percpu_gfp(size, 8, flags);
 528                 if (!storage->percpu_buf)
 529                         goto enomem;
 530         }
 531
 532         storage->map = (struct bpf_cgroup_storage_map *)map;
 533
 534         return storage;
 535
 536 enomem:
 537         bpf_map_uncharge_memlock(map, pages);
 538         kfree(storage);
 539         return ERR_PTR(-ENOMEM);
 540 }
 541
 542 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
 543 {
 544         struct bpf_cgroup_storage *storage =
 545                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 546
 547         kfree(storage->buf);
 548         kfree(storage);
 549 }
 550
 551 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
 552 {
 553         struct bpf_cgroup_storage *storage =
 554                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 555
 556         free_percpu(storage->percpu_buf);
 557         kfree(storage);
 558 }
 559
 560 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
 561 {
 562         enum bpf_cgroup_storage_type stype;
 563         struct bpf_map *map;
 564         u32 pages;
 565
 566         if (!storage)
 567                 return;
 568
 569         map = &storage->map->map;
 570
 571         bpf_cgroup_storage_calculate_size(map, &pages);
 572         bpf_map_uncharge_memlock(map, pages);
 573
 574         stype = cgroup_storage_type(map);
 575         if (stype == BPF_CGROUP_STORAGE_SHARED)
 576                 call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
 577         else
 578                 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
 579 }
 580
 581 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
 582                              struct cgroup *cgroup,
 583                              enum bpf_attach_type type)
 584 {
 585         struct bpf_cgroup_storage_map *map;
 586
 587         if (!storage)
 588                 return;
 589
 590         storage->key.attach_type = type;
 591         storage->key.cgroup_inode_id = cgroup_id(cgroup);
 592
 593         map = storage->map;
 594
 595         spin_lock_bh(&map->lock);
 596         WARN_ON(cgroup_storage_insert(map, storage));
 597         list_add(&storage->list_map, &map->list);
 598         list_add(&storage->list_cg, &cgroup->bpf.storages);
 599         spin_unlock_bh(&map->lock);
 600 }
 601
 602 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
 603 {
 604         struct bpf_cgroup_storage_map *map;
 605         struct rb_root *root;
 606
 607         if (!storage)
 608                 return;
 609
 610         map = storage->map;
 611
 612         spin_lock_bh(&map->lock);
 613         root = &map->root;
 614         rb_erase(&storage->node, root);
 615
 616         list_del(&storage->list_map);
 617         list_del(&storage->list_cg);
 618         spin_unlock_bh(&map->lock);
 619 }
 620
 621 #endif