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 = bpf_map_kmalloc_node(map, 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
 291         if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
 292             attr->key_size != sizeof(__u64))
 293                 return ERR_PTR(-EINVAL);
 294
 295         if (attr->value_size == 0)
 296                 return ERR_PTR(-EINVAL);
 297
 298         if (attr->value_size > PAGE_SIZE)
 299                 return ERR_PTR(-E2BIG);
 300
 301         if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
 302             !bpf_map_flags_access_ok(attr->map_flags))
 303                 return ERR_PTR(-EINVAL);
 304
 305         if (attr->max_entries)
 306                 /* max_entries is not used and enforced to be 0 */
 307                 return ERR_PTR(-EINVAL);
 308
 309         map = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
 310                            __GFP_ZERO | GFP_USER | __GFP_ACCOUNT, numa_node);
 311         if (!map)
 312                 return ERR_PTR(-ENOMEM);
 313
 314         /* copy mandatory map attributes */
 315         bpf_map_init_from_attr(&map->map, attr);
 316
 317         spin_lock_init(&map->lock);
 318         map->root = RB_ROOT;
 319         INIT_LIST_HEAD(&map->list);
 320
 321         return &map->map;
 322 }
 323
 324 static void cgroup_storage_map_free(struct bpf_map *_map)
 325 {
 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;
 329
 330         mutex_lock(&cgroup_mutex);
 331
 332         list_for_each_entry_safe(storage, stmp, storages, list_map) {
 333                 bpf_cgroup_storage_unlink(storage);
 334                 bpf_cgroup_storage_free(storage);
 335         }
 336
 337         mutex_unlock(&cgroup_mutex);
 338
 339         WARN_ON(!RB_EMPTY_ROOT(&map->root));
 340         WARN_ON(!list_empty(&map->list));
 341
 342         kfree(map);
 343 }
 344
 345 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
 346 {
 347         return -EINVAL;
 348 }
 349
 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)
 354 {
 355         if (attach_type_isolated(map)) {
 356                 struct btf_member *m;
 357                 u32 offset, size;
 358
 359                 /* Key is expected to be of struct bpf_cgroup_storage_key type,
 360                  * which is:
 361                  * struct bpf_cgroup_storage_key {
 362                  *      __u64   cgroup_inode_id;
 363                  *      __u32   attach_type;
 364                  * };
 365                  */
 366
 367                 /*
 368                  * Key_type must be a structure with two fields.
 369                  */
 370                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
 371                     BTF_INFO_VLEN(key_type->info) != 2)
 372                         return -EINVAL;
 373
 374                 /*
 375                  * The first field must be a 64 bit integer at 0 offset.
 376                  */
 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))
 380                         return -EINVAL;
 381
 382                 /*
 383                  * The second field must be a 32 bit integer at 64 bit offset.
 384                  */
 385                 m++;
 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))
 389                         return -EINVAL;
 390         } else {
 391                 u32 int_data;
 392
 393                 /*
 394                  * Key is expected to be u64, which stores the cgroup_inode_id
 395                  */
 396
 397                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 398                         return -EINVAL;
 399
 400                 int_data = *(u32 *)(key_type + 1);
 401                 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
 402                         return -EINVAL;
 403         }
 404
 405         return 0;
 406 }
 407
 408 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
 409                                          struct seq_file *m)
 410 {
 411         enum bpf_cgroup_storage_type stype = cgroup_storage_type(map);
 412         struct bpf_cgroup_storage *storage;
 413         int cpu;
 414
 415         rcu_read_lock();
 416         storage = cgroup_storage_lookup(map_to_storage(map), key, false);
 417         if (!storage) {
 418                 rcu_read_unlock();
 419                 return;
 420         }
 421
 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) {
 425                 seq_puts(m, ": ");
 426                 btf_type_seq_show(map->btf, map->btf_value_type_id,
 427                                   &READ_ONCE(storage->buf)->data[0], m);
 428                 seq_puts(m, "\n");
 429         } else {
 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),
 435                                           m);
 436                         seq_puts(m, "\n");
 437                 }
 438                 seq_puts(m, "}\n");
 439         }
 440         rcu_read_unlock();
 441 }
 442
 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,
 455 };
 456
 457 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
 458 {
 459         enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
 460
 461         if (aux->cgroup_storage[stype] &&
 462             aux->cgroup_storage[stype] != _map)
 463                 return -EBUSY;
 464
 465         aux->cgroup_storage[stype] = _map;
 466         return 0;
 467 }
 468
 469 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
 470 {
 471         size_t size;
 472
 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;
 477         } else {
 478                 size = map->value_size;
 479                 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
 480                                   PAGE_SIZE) >> PAGE_SHIFT;
 481         }
 482
 483         return size;
 484 }
 485
 486 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
 487                                         enum bpf_cgroup_storage_type stype)
 488 {
 489         const gfp_t gfp = __GFP_ZERO | GFP_USER;
 490         struct bpf_cgroup_storage *storage;
 491         struct bpf_map *map;
 492         size_t size;
 493         u32 pages;
 494
 495         map = prog->aux->cgroup_storage[stype];
 496         if (!map)
 497                 return NULL;
 498
 499         size = bpf_cgroup_storage_calculate_size(map, &pages);
 500
 501         storage = bpf_map_kmalloc_node(map, sizeof(struct bpf_cgroup_storage),
 502                                        gfp, map->numa_node);
 503         if (!storage)
 504                 goto enomem;
 505
 506         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 507                 storage->buf = bpf_map_kmalloc_node(map, size, gfp,
 508                                                     map->numa_node);
 509                 if (!storage->buf)
 510                         goto enomem;
 511                 check_and_init_map_lock(map, storage->buf->data);
 512         } else {
 513                 storage->percpu_buf = bpf_map_alloc_percpu(map, size, 8, gfp);
 514                 if (!storage->percpu_buf)
 515                         goto enomem;
 516         }
 517
 518         storage->map = (struct bpf_cgroup_storage_map *)map;
 519
 520         return storage;
 521
 522 enomem:
 523         kfree(storage);
 524         return ERR_PTR(-ENOMEM);
 525 }
 526
 527 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
 528 {
 529         struct bpf_cgroup_storage *storage =
 530                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 531
 532         kfree(storage->buf);
 533         kfree(storage);
 534 }
 535
 536 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
 537 {
 538         struct bpf_cgroup_storage *storage =
 539                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 540
 541         free_percpu(storage->percpu_buf);
 542         kfree(storage);
 543 }
 544
 545 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
 546 {
 547         enum bpf_cgroup_storage_type stype;
 548         struct bpf_map *map;
 549
 550         if (!storage)
 551                 return;
 552
 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);
 557         else
 558                 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
 559 }
 560
 561 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
 562                              struct cgroup *cgroup,
 563                              enum bpf_attach_type type)
 564 {
 565         struct bpf_cgroup_storage_map *map;
 566
 567         if (!storage)
 568                 return;
 569
 570         storage->key.attach_type = type;
 571         storage->key.cgroup_inode_id = cgroup_id(cgroup);
 572
 573         map = storage->map;
 574
 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);
 580 }
 581
 582 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
 583 {
 584         struct bpf_cgroup_storage_map *map;
 585         struct rb_root *root;
 586
 587         if (!storage)
 588                 return;
 589
 590         map = storage->map;
 591
 592         spin_lock_bh(&map->lock);
 593         root = &map->root;
 594         rb_erase(&storage->node, root);
 595
 596         list_del(&storage->list_map);
 597         list_del(&storage->list_cg);
 598         spin_unlock_bh(&map->lock);
 599 }
 600
 601 #endif