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