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, struct_size(new, data, map->value_size),
 167                                    __GFP_ZERO | GFP_ATOMIC | __GFP_NOWARN,
 168                                    map->numa_node);
 169         if (!new)
 170                 return -ENOMEM;
 171
 172         memcpy(&new->data[0], value, map->value_size);
 173         check_and_init_map_value(map, new->data);
 174
 175         new = xchg(&storage->buf, new);
 176         kfree_rcu(new, rcu);
 177
 178         return 0;
 179 }
 180
 181 int bpf_percpu_cgroup_storage_copy(struct bpf_map *_map, void *key,
 182                                    void *value)
 183 {
 184         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 185         struct bpf_cgroup_storage *storage;
 186         int cpu, off = 0;
 187         u32 size;
 188
 189         rcu_read_lock();
 190         storage = cgroup_storage_lookup(map, key, false);
 191         if (!storage) {
 192                 rcu_read_unlock();
 193                 return -ENOENT;
 194         }
 195
 196         /* per_cpu areas are zero-filled and bpf programs can only
 197          * access 'value_size' of them, so copying rounded areas
 198          * will not leak any kernel data
 199          */
 200         size = round_up(_map->value_size, 8);
 201         for_each_possible_cpu(cpu) {
 202                 bpf_long_memcpy(value + off,
 203                                 per_cpu_ptr(storage->percpu_buf, cpu), size);
 204                 off += size;
 205         }
 206         rcu_read_unlock();
 207         return 0;
 208 }
 209
 210 int bpf_percpu_cgroup_storage_update(struct bpf_map *_map, void *key,
 211                                      void *value, u64 map_flags)
 212 {
 213         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 214         struct bpf_cgroup_storage *storage;
 215         int cpu, off = 0;
 216         u32 size;
 217
 218         if (map_flags != BPF_ANY && map_flags != BPF_EXIST)
 219                 return -EINVAL;
 220
 221         rcu_read_lock();
 222         storage = cgroup_storage_lookup(map, key, false);
 223         if (!storage) {
 224                 rcu_read_unlock();
 225                 return -ENOENT;
 226         }
 227
 228         /* the user space will provide round_up(value_size, 8) bytes that
 229          * will be copied into per-cpu area. bpf programs can only access
 230          * value_size of it. During lookup the same extra bytes will be
 231          * returned or zeros which were zero-filled by percpu_alloc,
 232          * so no kernel data leaks possible
 233          */
 234         size = round_up(_map->value_size, 8);
 235         for_each_possible_cpu(cpu) {
 236                 bpf_long_memcpy(per_cpu_ptr(storage->percpu_buf, cpu),
 237                                 value + off, size);
 238                 off += size;
 239         }
 240         rcu_read_unlock();
 241         return 0;
 242 }
 243
 244 static int cgroup_storage_get_next_key(struct bpf_map *_map, void *key,
 245                                        void *_next_key)
 246 {
 247         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 248         struct bpf_cgroup_storage *storage;
 249
 250         spin_lock_bh(&map->lock);
 251
 252         if (list_empty(&map->list))
 253                 goto enoent;
 254
 255         if (key) {
 256                 storage = cgroup_storage_lookup(map, key, true);
 257                 if (!storage)
 258                         goto enoent;
 259
 260                 storage = list_next_entry(storage, list_map);
 261                 if (!storage)
 262                         goto enoent;
 263         } else {
 264                 storage = list_first_entry(&map->list,
 265                                          struct bpf_cgroup_storage, list_map);
 266         }
 267
 268         spin_unlock_bh(&map->lock);
 269
 270         if (attach_type_isolated(&map->map)) {
 271                 struct bpf_cgroup_storage_key *next = _next_key;
 272                 *next = storage->key;
 273         } else {
 274                 __u64 *next = _next_key;
 275                 *next = storage->key.cgroup_inode_id;
 276         }
 277         return 0;
 278
 279 enoent:
 280         spin_unlock_bh(&map->lock);
 281         return -ENOENT;
 282 }
 283
 284 static struct bpf_map *cgroup_storage_map_alloc(union bpf_attr *attr)
 285 {
 286         __u32 max_value_size = BPF_LOCAL_STORAGE_MAX_VALUE_SIZE;
 287         int numa_node = bpf_map_attr_numa_node(attr);
 288         struct bpf_cgroup_storage_map *map;
 289
 290         /* percpu is bound by PCPU_MIN_UNIT_SIZE, non-percu
 291          * is the same as other local storages.
 292          */
 293         if (attr->map_type == BPF_MAP_TYPE_PERCPU_CGROUP_STORAGE)
 294                 max_value_size = min_t(__u32, max_value_size,
 295                                        PCPU_MIN_UNIT_SIZE);
 296
 297         if (attr->key_size != sizeof(struct bpf_cgroup_storage_key) &&
 298             attr->key_size != sizeof(__u64))
 299                 return ERR_PTR(-EINVAL);
 300
 301         if (attr->value_size == 0)
 302                 return ERR_PTR(-EINVAL);
 303
 304         if (attr->value_size > max_value_size)
 305                 return ERR_PTR(-E2BIG);
 306
 307         if (attr->map_flags & ~LOCAL_STORAGE_CREATE_FLAG_MASK ||
 308             !bpf_map_flags_access_ok(attr->map_flags))
 309                 return ERR_PTR(-EINVAL);
 310
 311         if (attr->max_entries)
 312                 /* max_entries is not used and enforced to be 0 */
 313                 return ERR_PTR(-EINVAL);
 314
 315         map = kmalloc_node(sizeof(struct bpf_cgroup_storage_map),
 316                            __GFP_ZERO | GFP_USER | __GFP_ACCOUNT, numa_node);
 317         if (!map)
 318                 return ERR_PTR(-ENOMEM);
 319
 320         /* copy mandatory map attributes */
 321         bpf_map_init_from_attr(&map->map, attr);
 322
 323         spin_lock_init(&map->lock);
 324         map->root = RB_ROOT;
 325         INIT_LIST_HEAD(&map->list);
 326
 327         return &map->map;
 328 }
 329
 330 static void cgroup_storage_map_free(struct bpf_map *_map)
 331 {
 332         struct bpf_cgroup_storage_map *map = map_to_storage(_map);
 333         struct list_head *storages = &map->list;
 334         struct bpf_cgroup_storage *storage, *stmp;
 335
 336         mutex_lock(&cgroup_mutex);
 337
 338         list_for_each_entry_safe(storage, stmp, storages, list_map) {
 339                 bpf_cgroup_storage_unlink(storage);
 340                 bpf_cgroup_storage_free(storage);
 341         }
 342
 343         mutex_unlock(&cgroup_mutex);
 344
 345         WARN_ON(!RB_EMPTY_ROOT(&map->root));
 346         WARN_ON(!list_empty(&map->list));
 347
 348         kfree(map);
 349 }
 350
 351 static int cgroup_storage_delete_elem(struct bpf_map *map, void *key)
 352 {
 353         return -EINVAL;
 354 }
 355
 356 static int cgroup_storage_check_btf(const struct bpf_map *map,
 357                                     const struct btf *btf,
 358                                     const struct btf_type *key_type,
 359                                     const struct btf_type *value_type)
 360 {
 361         if (attach_type_isolated(map)) {
 362                 struct btf_member *m;
 363                 u32 offset, size;
 364
 365                 /* Key is expected to be of struct bpf_cgroup_storage_key type,
 366                  * which is:
 367                  * struct bpf_cgroup_storage_key {
 368                  *      __u64   cgroup_inode_id;
 369                  *      __u32   attach_type;
 370                  * };
 371                  */
 372
 373                 /*
 374                  * Key_type must be a structure with two fields.
 375                  */
 376                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_STRUCT ||
 377                     BTF_INFO_VLEN(key_type->info) != 2)
 378                         return -EINVAL;
 379
 380                 /*
 381                  * The first field must be a 64 bit integer at 0 offset.
 382                  */
 383                 m = (struct btf_member *)(key_type + 1);
 384                 size = sizeof_field(struct bpf_cgroup_storage_key, cgroup_inode_id);
 385                 if (!btf_member_is_reg_int(btf, key_type, m, 0, size))
 386                         return -EINVAL;
 387
 388                 /*
 389                  * The second field must be a 32 bit integer at 64 bit offset.
 390                  */
 391                 m++;
 392                 offset = offsetof(struct bpf_cgroup_storage_key, attach_type);
 393                 size = sizeof_field(struct bpf_cgroup_storage_key, attach_type);
 394                 if (!btf_member_is_reg_int(btf, key_type, m, offset, size))
 395                         return -EINVAL;
 396         } else {
 397                 u32 int_data;
 398
 399                 /*
 400                  * Key is expected to be u64, which stores the cgroup_inode_id
 401                  */
 402
 403                 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 404                         return -EINVAL;
 405
 406                 int_data = *(u32 *)(key_type + 1);
 407                 if (BTF_INT_BITS(int_data) != 64 || BTF_INT_OFFSET(int_data))
 408                         return -EINVAL;
 409         }
 410
 411         return 0;
 412 }
 413
 414 static void cgroup_storage_seq_show_elem(struct bpf_map *map, void *key,
 415                                          struct seq_file *m)
 416 {
 417         enum bpf_cgroup_storage_type stype;
 418         struct bpf_cgroup_storage *storage;
 419         int cpu;
 420
 421         rcu_read_lock();
 422         storage = cgroup_storage_lookup(map_to_storage(map), key, false);
 423         if (!storage) {
 424                 rcu_read_unlock();
 425                 return;
 426         }
 427
 428         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
 429         stype = cgroup_storage_type(map);
 430         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 431                 seq_puts(m, ": ");
 432                 btf_type_seq_show(map->btf, map->btf_value_type_id,
 433                                   &READ_ONCE(storage->buf)->data[0], m);
 434                 seq_puts(m, "\n");
 435         } else {
 436                 seq_puts(m, ": {\n");
 437                 for_each_possible_cpu(cpu) {
 438                         seq_printf(m, "\tcpu%d: ", cpu);
 439                         btf_type_seq_show(map->btf, map->btf_value_type_id,
 440                                           per_cpu_ptr(storage->percpu_buf, cpu),
 441                                           m);
 442                         seq_puts(m, "\n");
 443                 }
 444                 seq_puts(m, "}\n");
 445         }
 446         rcu_read_unlock();
 447 }
 448
 449 static int cgroup_storage_map_btf_id;
 450 const struct bpf_map_ops cgroup_storage_map_ops = {
 451         .map_alloc = cgroup_storage_map_alloc,
 452         .map_free = cgroup_storage_map_free,
 453         .map_get_next_key = cgroup_storage_get_next_key,
 454         .map_lookup_elem = cgroup_storage_lookup_elem,
 455         .map_update_elem = cgroup_storage_update_elem,
 456         .map_delete_elem = cgroup_storage_delete_elem,
 457         .map_check_btf = cgroup_storage_check_btf,
 458         .map_seq_show_elem = cgroup_storage_seq_show_elem,
 459         .map_btf_name = "bpf_cgroup_storage_map",
 460         .map_btf_id = &cgroup_storage_map_btf_id,
 461 };
 462
 463 int bpf_cgroup_storage_assign(struct bpf_prog_aux *aux, struct bpf_map *_map)
 464 {
 465         enum bpf_cgroup_storage_type stype = cgroup_storage_type(_map);
 466
 467         if (aux->cgroup_storage[stype] &&
 468             aux->cgroup_storage[stype] != _map)
 469                 return -EBUSY;
 470
 471         aux->cgroup_storage[stype] = _map;
 472         return 0;
 473 }
 474
 475 static size_t bpf_cgroup_storage_calculate_size(struct bpf_map *map, u32 *pages)
 476 {
 477         size_t size;
 478
 479         if (cgroup_storage_type(map) == BPF_CGROUP_STORAGE_SHARED) {
 480                 size = sizeof(struct bpf_storage_buffer) + map->value_size;
 481                 *pages = round_up(sizeof(struct bpf_cgroup_storage) + size,
 482                                   PAGE_SIZE) >> PAGE_SHIFT;
 483         } else {
 484                 size = map->value_size;
 485                 *pages = round_up(round_up(size, 8) * num_possible_cpus(),
 486                                   PAGE_SIZE) >> PAGE_SHIFT;
 487         }
 488
 489         return size;
 490 }
 491
 492 struct bpf_cgroup_storage *bpf_cgroup_storage_alloc(struct bpf_prog *prog,
 493                                         enum bpf_cgroup_storage_type stype)
 494 {
 495         const gfp_t gfp = __GFP_ZERO | GFP_USER;
 496         struct bpf_cgroup_storage *storage;
 497         struct bpf_map *map;
 498         size_t size;
 499         u32 pages;
 500
 501         map = prog->aux->cgroup_storage[stype];
 502         if (!map)
 503                 return NULL;
 504
 505         size = bpf_cgroup_storage_calculate_size(map, &pages);
 506
 507         storage = bpf_map_kmalloc_node(map, sizeof(struct bpf_cgroup_storage),
 508                                        gfp, map->numa_node);
 509         if (!storage)
 510                 goto enomem;
 511
 512         if (stype == BPF_CGROUP_STORAGE_SHARED) {
 513                 storage->buf = bpf_map_kmalloc_node(map, size, gfp,
 514                                                     map->numa_node);
 515                 if (!storage->buf)
 516                         goto enomem;
 517                 check_and_init_map_value(map, storage->buf->data);
 518         } else {
 519                 storage->percpu_buf = bpf_map_alloc_percpu(map, size, 8, gfp);
 520                 if (!storage->percpu_buf)
 521                         goto enomem;
 522         }
 523
 524         storage->map = (struct bpf_cgroup_storage_map *)map;
 525
 526         return storage;
 527
 528 enomem:
 529         kfree(storage);
 530         return ERR_PTR(-ENOMEM);
 531 }
 532
 533 static void free_shared_cgroup_storage_rcu(struct rcu_head *rcu)
 534 {
 535         struct bpf_cgroup_storage *storage =
 536                 container_of(rcu, struct bpf_cgroup_storage, rcu);
 537
 538         kfree(storage->buf);
 539         kfree(storage);
 540 }
 541
 542 static void free_percpu_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         free_percpu(storage->percpu_buf);
 548         kfree(storage);
 549 }
 550
 551 void bpf_cgroup_storage_free(struct bpf_cgroup_storage *storage)
 552 {
 553         enum bpf_cgroup_storage_type stype;
 554         struct bpf_map *map;
 555
 556         if (!storage)
 557                 return;
 558
 559         map = &storage->map->map;
 560         stype = cgroup_storage_type(map);
 561         if (stype == BPF_CGROUP_STORAGE_SHARED)
 562                 call_rcu(&storage->rcu, free_shared_cgroup_storage_rcu);
 563         else
 564                 call_rcu(&storage->rcu, free_percpu_cgroup_storage_rcu);
 565 }
 566
 567 void bpf_cgroup_storage_link(struct bpf_cgroup_storage *storage,
 568                              struct cgroup *cgroup,
 569                              enum bpf_attach_type type)
 570 {
 571         struct bpf_cgroup_storage_map *map;
 572
 573         if (!storage)
 574                 return;
 575
 576         storage->key.attach_type = type;
 577         storage->key.cgroup_inode_id = cgroup_id(cgroup);
 578
 579         map = storage->map;
 580
 581         spin_lock_bh(&map->lock);
 582         WARN_ON(cgroup_storage_insert(map, storage));
 583         list_add(&storage->list_map, &map->list);
 584         list_add(&storage->list_cg, &cgroup->bpf.storages);
 585         spin_unlock_bh(&map->lock);
 586 }
 587
 588 void bpf_cgroup_storage_unlink(struct bpf_cgroup_storage *storage)
 589 {
 590         struct bpf_cgroup_storage_map *map;
 591         struct rb_root *root;
 592
 593         if (!storage)
 594                 return;
 595
 596         map = storage->map;
 597
 598         spin_lock_bh(&map->lock);
 599         root = &map->root;
 600         rb_erase(&storage->node, root);
 601
 602         list_del(&storage->list_map);
 603         list_del(&storage->list_cg);
 604         spin_unlock_bh(&map->lock);
 605 }
 606
 607 #endif