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/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
  13 #ifdef CONFIG_CGROUP_BPF
  14
  15 #include "../cgroup/cgroup-internal.h"
  16
  17 #define LOCAL_STORAGE_CREATE_FLAG_MASK                                  \
  18         (BPF_F_NUMA_NODE | BPF_F_ACCESS_MASK)
  19
  20 struct bpf_cgroup_storage_map {
  21         struct bpf_map map;
  22
  23         spinlock_t lock;
  24         struct rb_root root;
  25         struct list_head list;
  26 };
  27
  28 static struct bpf_cgroup_storage_map *map_to_storage(struct bpf_map *map)
  29 {
  30         return container_of(map, struct bpf_cgroup_storage_map, map);
  31 }
  32
  33 static bool attach_type_isolated(const struct bpf_map *map)
  34 {
  35         return map->key_size == sizeof(struct bpf_cgroup_storage_key);
  36 }
  37
  38 static int bpf_cgroup_storage_key_cmp(const struct bpf_cgroup_storage_map *map,
  39                                       const void *_key1, const void *_key2)
  40 {
  41         if (attach_type_isolated(&map->map)) {
  42                 const struct bpf_cgroup_storage_key *key1 = _key1;
  43                 const struct bpf_cgroup_storage_key *key2 = _key2;
  44
  45                 if (key1->cgroup_inode_id < key2->cgroup_inode_id)
  46                         return -1;
  47                 else if (key1->cgroup_inode_id > key2->cgroup_inode_id)
  48                         return 1;
  49                 else if (key1->attach_type < key2->attach_type)
  50                         return -1;
  51                 else if (key1->attach_type > key2->attach_type)
  52                         return 1;
  53         } else {
  54                 const __u64 *cgroup_inode_id1 = _key1;
  55                 const __u64 *cgroup_inode_id2 = _key2;
  56
  57                 if (*cgroup_inode_id1 < *cgroup_inode_id2)
  58                         return -1;
  59                 else if (*cgroup_inode_id1 > *cgroup_inode_id2)
  60                         return 1;
  61         }
  62         return 0;
  63 }
  64
  65 struct bpf_cgroup_storage *
  66 cgroup_storage_lookup(struct bpf_cgroup_storage_map *map,
  67                       void *key, bool locked)
  68 {
  69         struct rb_root *root = &map->root;
  70         struct rb_node *node;
  71
  72         if (!locked)
  73                 spin_lock_bh(&map->lock);
  74
  75         node = root->rb_node;
  76         while (node) {
  77                 struct bpf_cgroup_storage *storage;
  78
  79                 storage = container_of(node, struct bpf_cgroup_storage, node);
  80
  81                 switch (bpf_cgroup_storage_key_cmp(map, key, &storage->key)) {
  82                 case -1:
  83                         node = node->rb_left;
  84                         break;
  85                 case 1:
  86                         node = node->rb_right;
  87                         break;
  88                 default:
  89                         if (!locked)
  90                                 spin_unlock_bh(&map->lock);
  91                         return storage;
  92                 }
  93         }
  94
  95         if (!locked)
  96                 spin_unlock_bh(&map->lock);
  97
  98         return NULL;
  99 }
 100
 101 static int cgroup_storage_insert(struct bpf_cgroup_storage_map *map,
 102                                  struct bpf_cgroup_storage *storage)
 103 {
 104         struct rb_root *root = &map->root;
 105         struct rb_node **new = &(root->rb_node), *parent = NULL;
 106
 107         while (*new) {
 108                 struct bpf_cgroup_storage *this;
 109
 110                 this = container_of(*new, struct bpf_cgroup_storage, node);
 111
 112                 parent = *new;
 113                 switch (bpf_cgroup_storage_key_cmp(map, &storage->key, &this->key)) {
 114                 case -1:
 115                         new = &((*new)->rb_left);
 116                         break;
 117                 case 1:
 118                         new = &((*new)->rb_right);
 119                         break;
 120                 default:
 121                         return -EEXIST;
 122                 }
 123         }
 124
 125         rb_link_node(&storage->node, parent, new);
 126         rb_insert_color(&storage->node, root);
 127
 128         return 0;
 129 }
 130
 131 static void *cgroup_storage_lookup_elem(struct bpf_map *_map, void *key)
 132 {
 133         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 134         struct bpf_cgroup_storage *storage;
 135
 136         storage = cgroup_storage_lookup(map, key, false);
 137         if (!storage)
 138                 return NULL;
 139
 140         return &READ_ONCE(storage->buf)->data[0];
 141 }
 142
 143 static int cgroup_storage_update_elem(struct bpf_map *map, void *key,
 144                                       void *value, u64 flags)
 145 {
 146         struct bpf_cgroup_storage *storage;
 147         struct bpf_storage_buffer *new;
 148
 149         if (unlikely(flags & ~(BPF_F_LOCK | BPF_EXIST)))
 150                 return -EINVAL;
 151
 152         if (unlikely((flags & BPF_F_LOCK) &&
 153                      !map_value_has_spin_lock(map)))
 154                 return -EINVAL;
 155
 156         storage = cgroup_storage_lookup((struct bpf_cgroup_storage_map *)map,
 157                                         key, false);
 158         if (!storage)
 159                 return -ENOENT;
 160
 161         if (flags & BPF_F_LOCK) {
 162                 copy_map_value_locked(map, storage->buf->data, value, false);
 163                 return 0;
 164         }
 165
 166         new = bpf_map_kmalloc_node(map, sizeof(struct bpf_storage_buffer) +
 167                                    map->value_size,
 168                                    __GFP_ZERO | GFP_ATOMIC | __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 = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
 317                            __GFP_ZERO | GFP_USER | __GFP_ACCOUNT, numa_node);
 318         if (!map)
 319                 return ERR_PTR(-ENOMEM);
 320
 321         /* copy mandatory map attributes */
 322         bpf_map_init_from_attr(&map->map, attr);
 323
 324         spin_lock_init(&map->lock);
 325         map->root = RB_ROOT;
 326         INIT_LIST_HEAD(&map->list);
 327
 328         return &map->map;
 329 }
 330
 331 static void cgroup_storage_map_free(struct bpf_map *_map)
 332 {
 333         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 334         struct list_head *storages = &map->list;
 335         struct bpf_cgroup_storage *storage, *stmp;
 336
 337         mutex_lock(&cgroup_mutex);
 338
 339         list_for_each_entry_safe(storage, stmp, storages, list_map) {
 340                 bpf_cgroup_storage_unlink(storage);
 341                 bpf_cgroup_storage_free(storage);
 342         }
 343
 344         mutex_unlock(&cgroup_mutex);
 345
 346         WARN_ON(!RB_EMPTY_ROOT(&map->root));
 347         WARN_ON(!list_empty(&map->list));
 348
 349         kfree(map);
 350 }
 351
 352 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
 353 {
 354         return -EINVAL;
 355 }
 356
 357 static int cgroup_storage_check_btf(const struct bpf_map *map,
 358                                     const struct btf *btf,
 359                                     const struct btf_type *key_type,
 360                                     const struct btf_type *value_type)
 361 {
 362         if (attach_type_isolated(map)) {
 363                 struct btf_member *m;
 364                 u32 offset, size;
 365
 366                 /* Key is expected to be of struct bpf_cgroup_storage_key type,
 367                  * which is:
 368                  * struct bpf_cgroup_storage_key {
 369                  *      __u64   cgroup_inode_id;
 370                  *      __u32   attach_type;
 371                  * };
 372                  */
 373
 374                 /*
 375                  * Key_type must be a structure with two fields.
 376                  */
 377                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
 378                     BTF_INFO_VLEN(key_type->info) != 2)
 379                         return -EINVAL;
 380
 381                 /*
 382                  * The first field must be a 64 bit integer at 0 offset.
 383                  */
 384                 m = (struct btf_member *)(key_type + 1);
 385                 size = sizeof_field(struct bpf_cgroup_storage_key, cgroup_inode_id);
 386                 if (!btf_member_is_reg_int(btf, key_type, m, 0, size))
 387                         return -EINVAL;
 388
 389                 /*
 390                  * The second field must be a 32 bit integer at 64 bit offset.
 391                  */
 392                 m++;
 393                 offset = offsetof(struct bpf_cgroup_storage_key, attach_type);
 394                 size = sizeof_field(struct bpf_cgroup_storage_key, attach_type);
 395                 if (!btf_member_is_reg_int(btf, key_type, m, offset, size))
 396                         return -EINVAL;
 397         } else {
 398                 u32 int_data;
 399
 400                 /*
 401                  * Key is expected to be u64, which stores the cgroup_inode_id
 402                  */
 403
 404                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 405                         return -EINVAL;
 406
 407                 int_data = *(u32 *)(key_type + 1);
 408                 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
 409                         return -EINVAL;
 410         }
 411
 412         return 0;
 413 }
 414
 415 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
 416                                          struct seq_file *m)
 417 {
 418         enum bpf_cgroup_storage_type stype;
 419         struct bpf_cgroup_storage *storage;
 420         int cpu;
 421
 422         rcu_read_lock();
 423         storage = cgroup_storage_lookup(map_to_storage(map), key, false);
 424         if (!storage) {
 425                 rcu_read_unlock();
 426                 return;
 427         }
 428
 429         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
 430         stype = cgroup_storage_type(map);
 431         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 432                 seq_puts(m, ": ");
 433                 btf_type_seq_show(map->btf, map->btf_value_type_id,
 434                                   &READ_ONCE(storage->buf)->data[0], m);
 435                 seq_puts(m, "\n");
 436         } else {
 437                 seq_puts(m, ": {\n");
 438                 for_each_possible_cpu(cpu) {
 439                         seq_printf(m, "\tcpu%d: ", cpu);
 440                         btf_type_seq_show(map->btf, map->btf_value_type_id,
 441                                           per_cpu_ptr(storage->percpu_buf, cpu),
 442                                           m);
 443                         seq_puts(m, "\n");
 444                 }
 445                 seq_puts(m, "}\n");
 446         }
 447         rcu_read_unlock();
 448 }
 449
 450 static int cgroup_storage_map_btf_id;
 451 const struct bpf_map_ops cgroup_storage_map_ops = {
 452         .map_alloc = cgroup_storage_map_alloc,
 453         .map_free = cgroup_storage_map_free,
 454         .map_get_next_key = cgroup_storage_get_next_key,
 455         .map_lookup_elem = cgroup_storage_lookup_elem,
 456         .map_update_elem = cgroup_storage_update_elem,
 457         .map_delete_elem = cgroup_storage_delete_elem,
 458         .map_check_btf = cgroup_storage_check_btf,
 459         .map_seq_show_elem = cgroup_storage_seq_show_elem,
 460         .map_btf_name = "bpf_cgroup_storage_map",
 461         .map_btf_id = &cgroup_storage_map_btf_id,
 462 };
 463
 464 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
 465 {
 466         enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
 467
 468         if (aux->cgroup_storage[stype] &&
 469             aux->cgroup_storage[stype] != _map)
 470                 return -EBUSY;
 471
 472         aux->cgroup_storage[stype] = _map;
 473         return 0;
 474 }
 475
 476 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
 477 {
 478         size_t size;
 479
 480         if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
 481                 size = sizeof(struct bpf_storage_buffer) + map->value_size;
 482                 *pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
 483                                   PAGE_SIZE) >> PAGE_SHIFT;
 484         } else {
 485                 size = map->value_size;
 486                 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
 487                                   PAGE_SIZE) >> PAGE_SHIFT;
 488         }
 489
 490         return size;
 491 }
 492
 493 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
 494                                         enum bpf_cgroup_storage_type stype)
 495 {
 496         const gfp_t gfp = __GFP_ZERO | GFP_USER;
 497         struct bpf_cgroup_storage *storage;
 498         struct bpf_map *map;
 499         size_t size;
 500         u32 pages;
 501
 502         map = prog->aux->cgroup_storage[stype];
 503         if (!map)
 504                 return NULL;
 505
 506         size = bpf_cgroup_storage_calculate_size(map, &pages);
 507
 508         storage = bpf_map_kmalloc_node(map, sizeof(struct bpf_cgroup_storage),
 509                                        gfp, map->numa_node);
 510         if (!storage)
 511                 goto enomem;
 512
 513         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 514                 storage->buf = bpf_map_kmalloc_node(map, size, gfp,
 515                                                     map->numa_node);
 516                 if (!storage->buf)
 517                         goto enomem;
 518                 check_and_init_map_value(map, storage->buf->data);
 519         } else {
 520                 storage->percpu_buf = bpf_map_alloc_percpu(map, size, 8, gfp);
 521                 if (!storage->percpu_buf)
 522                         goto enomem;
 523         }
 524
 525         storage->map = (struct bpf_cgroup_storage_map *)map;
 526
 527         return storage;
 528
 529 enomem:
 530         kfree(storage);
 531         return ERR_PTR(-ENOMEM);
 532 }
 533
 534 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
 535 {
 536         struct bpf_cgroup_storage *storage =
 537                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 538
 539         kfree(storage->buf);
 540         kfree(storage);
 541 }
 542
 543 static void free_percpu_cgroup_storage_rcu(struct rcu_head *rcu)
 544 {
 545         struct bpf_cgroup_storage *storage =
 546                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 547
 548         free_percpu(storage->percpu_buf);
 549         kfree(storage);
 550 }
 551
 552 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
 553 {
 554         enum bpf_cgroup_storage_type stype;
 555         struct bpf_map *map;
 556
 557         if (!storage)
 558                 return;
 559
 560         map = &storage->map->map;
 561         stype = cgroup_storage_type(map);
 562         if (stype == BPF_CGROUP_STORAGE_SHARED)
 563                 call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
 564         else
 565                 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
 566 }
 567
 568 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
 569                              struct cgroup *cgroup,
 570                              enum bpf_attach_type type)
 571 {
 572         struct bpf_cgroup_storage_map *map;
 573
 574         if (!storage)
 575                 return;
 576
 577         storage->key.attach_type = type;
 578         storage->key.cgroup_inode_id = cgroup_id(cgroup);
 579
 580         map = storage->map;
 581
 582         spin_lock_bh(&map->lock);
 583         WARN_ON(cgroup_storage_insert(map, storage));
 584         list_add(&storage->list_map, &map->list);
 585         list_add(&storage->list_cg, &cgroup->bpf.storages);
 586         spin_unlock_bh(&map->lock);
 587 }
 588
 589 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
 590 {
 591         struct bpf_cgroup_storage_map *map;
 592         struct rb_root *root;
 593
 594         if (!storage)
 595                 return;
 596
 597         map = storage->map;
 598
 599         spin_lock_bh(&map->lock);
 600         root = &map->root;
 601         rb_erase(&storage->node, root);
 602
 603         list_del(&storage->list_map);
 604         list_del(&storage->list_cg);
 605         spin_unlock_bh(&map->lock);
 606 }
 607
 608 #endif