kernel/bpf/bpf_local_storage.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* Copyright (c) 2019 Facebook  */
   3 #include <linux/rculist.h>
   4 #include <linux/list.h>
   5 #include <linux/hash.h>
   6 #include <linux/types.h>
   7 #include <linux/spinlock.h>
   8 #include <linux/bpf.h>
   9 #include <linux/btf_ids.h>
  10 #include <linux/bpf_local_storage.h>
  11 #include <net/sock.h>
  12 #include <uapi/linux/sock_diag.h>
  13 #include <uapi/linux/btf.h>
  14
  15 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE)
  16
  17 static struct bpf_local_storage_map_bucket *
  18 select_bucket(struct bpf_local_storage_map *smap,
  19               struct bpf_local_storage_elem *selem)
  20 {
  21         return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
  22 }
  23
  24 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size)
  25 {
  26         struct bpf_map *map = &smap->map;
  27
  28         if (!map->ops->map_local_storage_charge)
  29                 return 0;
  30
  31         return map->ops->map_local_storage_charge(smap, owner, size);
  32 }
  33
  34 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner,
  35                          u32 size)
  36 {
  37         struct bpf_map *map = &smap->map;
  38
  39         if (map->ops->map_local_storage_uncharge)
  40                 map->ops->map_local_storage_uncharge(smap, owner, size);
  41 }
  42
  43 static struct bpf_local_storage __rcu **
  44 owner_storage(struct bpf_local_storage_map *smap, void *owner)
  45 {
  46         struct bpf_map *map = &smap->map;
  47
  48         return map->ops->map_owner_storage_ptr(owner);
  49 }
  50
  51 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
  52 {
  53         return !hlist_unhashed(&selem->snode);
  54 }
  55
  56 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
  57 {
  58         return !hlist_unhashed(&selem->map_node);
  59 }
  60
  61 struct bpf_local_storage_elem *
  62 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
  63                 void *value, bool charge_mem)
  64 {
  65         struct bpf_local_storage_elem *selem;
  66
  67         if (charge_mem && mem_charge(smap, owner, smap->elem_size))
  68                 return NULL;
  69
  70         selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
  71                                 GFP_ATOMIC | __GFP_NOWARN);
  72         if (selem) {
  73                 if (value)
  74                         memcpy(SDATA(selem)->data, value, smap->map.value_size);
  75                 return selem;
  76         }
  77
  78         if (charge_mem)
  79                 mem_uncharge(smap, owner, smap->elem_size);
  80
  81         return NULL;
  82 }
  83
  84 /* local_storage->lock must be held and selem->local_storage == local_storage.
  85  * The caller must ensure selem->smap is still valid to be
  86  * dereferenced for its smap->elem_size and smap->cache_idx.
  87  */
  88 bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
  89                                      struct bpf_local_storage_elem *selem,
  90                                      bool uncharge_mem)
  91 {
  92         struct bpf_local_storage_map *smap;
  93         bool free_local_storage;
  94         void *owner;
  95
  96         smap = rcu_dereference(SDATA(selem)->smap);
  97         owner = local_storage->owner;
  98
  99         /* All uncharging on the owner must be done first.
 100          * The owner may be freed once the last selem is unlinked
 101          * from local_storage.
 102          */
 103         if (uncharge_mem)
 104                 mem_uncharge(smap, owner, smap->elem_size);
 105
 106         free_local_storage = hlist_is_singular_node(&selem->snode,
 107                                                     &local_storage->list);
 108         if (free_local_storage) {
 109                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 110                 local_storage->owner = NULL;
 111
 112                 /* After this RCU_INIT, owner may be freed and cannot be used */
 113                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 114
 115                 /* local_storage is not freed now.  local_storage->lock is
 116                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 117                  * will be done by the caller.
 118                  *
 119                  * Although the unlock will be done under
 120                  * rcu_read_lock(),  it is more intutivie to
 121                  * read if kfree_rcu(local_storage, rcu) is done
 122                  * after the raw_spin_unlock_bh(&local_storage->lock).
 123                  *
 124                  * Hence, a "bool free_local_storage" is returned
 125                  * to the caller which then calls the kfree_rcu()
 126                  * after unlock.
 127                  */
 128         }
 129         hlist_del_init_rcu(&selem->snode);
 130         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 131             SDATA(selem))
 132                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 133
 134         kfree_rcu(selem, rcu);
 135
 136         return free_local_storage;
 137 }
 138
 139 static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem)
 140 {
 141         struct bpf_local_storage *local_storage;
 142         bool free_local_storage = false;
 143
 144         if (unlikely(!selem_linked_to_storage(selem)))
 145                 /* selem has already been unlinked from sk */
 146                 return;
 147
 148         local_storage = rcu_dereference(selem->local_storage);
 149         raw_spin_lock_bh(&local_storage->lock);
 150         if (likely(selem_linked_to_storage(selem)))
 151                 free_local_storage = bpf_selem_unlink_storage_nolock(
 152                         local_storage, selem, true);
 153         raw_spin_unlock_bh(&local_storage->lock);
 154
 155         if (free_local_storage)
 156                 kfree_rcu(local_storage, rcu);
 157 }
 158
 159 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 160                                    struct bpf_local_storage_elem *selem)
 161 {
 162         RCU_INIT_POINTER(selem->local_storage, local_storage);
 163         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 164 }
 165
 166 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 167 {
 168         struct bpf_local_storage_map *smap;
 169         struct bpf_local_storage_map_bucket *b;
 170
 171         if (unlikely(!selem_linked_to_map(selem)))
 172                 /* selem has already be unlinked from smap */
 173                 return;
 174
 175         smap = rcu_dereference(SDATA(selem)->smap);
 176         b = select_bucket(smap, selem);
 177         raw_spin_lock_bh(&b->lock);
 178         if (likely(selem_linked_to_map(selem)))
 179                 hlist_del_init_rcu(&selem->map_node);
 180         raw_spin_unlock_bh(&b->lock);
 181 }
 182
 183 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 184                         struct bpf_local_storage_elem *selem)
 185 {
 186         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 187
 188         raw_spin_lock_bh(&b->lock);
 189         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 190         hlist_add_head_rcu(&selem->map_node, &b->list);
 191         raw_spin_unlock_bh(&b->lock);
 192 }
 193
 194 void bpf_selem_unlink(struct bpf_local_storage_elem *selem)
 195 {
 196         /* Always unlink from map before unlinking from local_storage
 197          * because selem will be freed after successfully unlinked from
 198          * the local_storage.
 199          */
 200         bpf_selem_unlink_map(selem);
 201         __bpf_selem_unlink_storage(selem);
 202 }
 203
 204 struct bpf_local_storage_data *
 205 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 206                          struct bpf_local_storage_map *smap,
 207                          bool cacheit_lockit)
 208 {
 209         struct bpf_local_storage_data *sdata;
 210         struct bpf_local_storage_elem *selem;
 211
 212         /* Fast path (cache hit) */
 213         sdata = rcu_dereference(local_storage->cache[smap->cache_idx]);
 214         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 215                 return sdata;
 216
 217         /* Slow path (cache miss) */
 218         hlist_for_each_entry_rcu(selem, &local_storage->list, snode)
 219                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 220                         break;
 221
 222         if (!selem)
 223                 return NULL;
 224
 225         sdata = SDATA(selem);
 226         if (cacheit_lockit) {
 227                 /* spinlock is needed to avoid racing with the
 228                  * parallel delete.  Otherwise, publishing an already
 229                  * deleted sdata to the cache will become a use-after-free
 230                  * problem in the next bpf_local_storage_lookup().
 231                  */
 232                 raw_spin_lock_bh(&local_storage->lock);
 233                 if (selem_linked_to_storage(selem))
 234                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 235                                            sdata);
 236                 raw_spin_unlock_bh(&local_storage->lock);
 237         }
 238
 239         return sdata;
 240 }
 241
 242 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 243                        u64 map_flags)
 244 {
 245         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 246                 /* elem already exists */
 247                 return -EEXIST;
 248
 249         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 250                 /* elem doesn't exist, cannot update it */
 251                 return -ENOENT;
 252
 253         return 0;
 254 }
 255
 256 int bpf_local_storage_alloc(void *owner,
 257                             struct bpf_local_storage_map *smap,
 258                             struct bpf_local_storage_elem *first_selem)
 259 {
 260         struct bpf_local_storage *prev_storage, *storage;
 261         struct bpf_local_storage **owner_storage_ptr;
 262         int err;
 263
 264         err = mem_charge(smap, owner, sizeof(*storage));
 265         if (err)
 266                 return err;
 267
 268         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 269                                   GFP_ATOMIC | __GFP_NOWARN);
 270         if (!storage) {
 271                 err = -ENOMEM;
 272                 goto uncharge;
 273         }
 274
 275         INIT_HLIST_HEAD(&storage->list);
 276         raw_spin_lock_init(&storage->lock);
 277         storage->owner = owner;
 278
 279         bpf_selem_link_storage_nolock(storage, first_selem);
 280         bpf_selem_link_map(smap, first_selem);
 281
 282         owner_storage_ptr =
 283                 (struct bpf_local_storage **)owner_storage(smap, owner);
 284         /* Publish storage to the owner.
 285          * Instead of using any lock of the kernel object (i.e. owner),
 286          * cmpxchg will work with any kernel object regardless what
 287          * the running context is, bh, irq...etc.
 288          *
 289          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 290          * is protected by the storage->lock.  Hence, when freeing
 291          * the owner->storage, the storage->lock must be held before
 292          * setting owner->storage ptr to NULL.
 293          */
 294         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 295         if (unlikely(prev_storage)) {
 296                 bpf_selem_unlink_map(first_selem);
 297                 err = -EAGAIN;
 298                 goto uncharge;
 299
 300                 /* Note that even first_selem was linked to smap's
 301                  * bucket->list, first_selem can be freed immediately
 302                  * (instead of kfree_rcu) because
 303                  * bpf_local_storage_map_free() does a
 304                  * synchronize_rcu() before walking the bucket->list.
 305                  * Hence, no one is accessing selem from the
 306                  * bucket->list under rcu_read_lock().
 307                  */
 308         }
 309
 310         return 0;
 311
 312 uncharge:
 313         kfree(storage);
 314         mem_uncharge(smap, owner, sizeof(*storage));
 315         return err;
 316 }
 317
 318 /* sk cannot be going away because it is linking new elem
 319  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 320  * Otherwise, it will become a leak (and other memory issues
 321  * during map destruction).
 322  */
 323 struct bpf_local_storage_data *
 324 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 325                          void *value, u64 map_flags)
 326 {
 327         struct bpf_local_storage_data *old_sdata = NULL;
 328         struct bpf_local_storage_elem *selem;
 329         struct bpf_local_storage *local_storage;
 330         int err;
 331
 332         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 333         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 334             /* BPF_F_LOCK can only be used in a value with spin_lock */
 335             unlikely((map_flags & BPF_F_LOCK) &&
 336                      !map_value_has_spin_lock(&smap->map)))
 337                 return ERR_PTR(-EINVAL);
 338
 339         local_storage = rcu_dereference(*owner_storage(smap, owner));
 340         if (!local_storage || hlist_empty(&local_storage->list)) {
 341                 /* Very first elem for the owner */
 342                 err = check_flags(NULL, map_flags);
 343                 if (err)
 344                         return ERR_PTR(err);
 345
 346                 selem = bpf_selem_alloc(smap, owner, value, true);
 347                 if (!selem)
 348                         return ERR_PTR(-ENOMEM);
 349
 350                 err = bpf_local_storage_alloc(owner, smap, selem);
 351                 if (err) {
 352                         kfree(selem);
 353                         mem_uncharge(smap, owner, smap->elem_size);
 354                         return ERR_PTR(err);
 355                 }
 356
 357                 return SDATA(selem);
 358         }
 359
 360         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 361                 /* Hoping to find an old_sdata to do inline update
 362                  * such that it can avoid taking the local_storage->lock
 363                  * and changing the lists.
 364                  */
 365                 old_sdata =
 366                         bpf_local_storage_lookup(local_storage, smap, false);
 367                 err = check_flags(old_sdata, map_flags);
 368                 if (err)
 369                         return ERR_PTR(err);
 370                 if (old_sdata && selem_linked_to_storage(SELEM(old_sdata))) {
 371                         copy_map_value_locked(&smap->map, old_sdata->data,
 372                                               value, false);
 373                         return old_sdata;
 374                 }
 375         }
 376
 377         raw_spin_lock_bh(&local_storage->lock);
 378
 379         /* Recheck local_storage->list under local_storage->lock */
 380         if (unlikely(hlist_empty(&local_storage->list))) {
 381                 /* A parallel del is happening and local_storage is going
 382                  * away.  It has just been checked before, so very
 383                  * unlikely.  Return instead of retry to keep things
 384                  * simple.
 385                  */
 386                 err = -EAGAIN;
 387                 goto unlock_err;
 388         }
 389
 390         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 391         err = check_flags(old_sdata, map_flags);
 392         if (err)
 393                 goto unlock_err;
 394
 395         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 396                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 397                                       false);
 398                 selem = SELEM(old_sdata);
 399                 goto unlock;
 400         }
 401
 402         /* local_storage->lock is held.  Hence, we are sure
 403          * we can unlink and uncharge the old_sdata successfully
 404          * later.  Hence, instead of charging the new selem now
 405          * and then uncharge the old selem later (which may cause
 406          * a potential but unnecessary charge failure),  avoid taking
 407          * a charge at all here (the "!old_sdata" check) and the
 408          * old_sdata will not be uncharged later during
 409          * bpf_selem_unlink_storage_nolock().
 410          */
 411         selem = bpf_selem_alloc(smap, owner, value, !old_sdata);
 412         if (!selem) {
 413                 err = -ENOMEM;
 414                 goto unlock_err;
 415         }
 416
 417         /* First, link the new selem to the map */
 418         bpf_selem_link_map(smap, selem);
 419
 420         /* Second, link (and publish) the new selem to local_storage */
 421         bpf_selem_link_storage_nolock(local_storage, selem);
 422
 423         /* Third, remove old selem, SELEM(old_sdata) */
 424         if (old_sdata) {
 425                 bpf_selem_unlink_map(SELEM(old_sdata));
 426                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 427                                                 false);
 428         }
 429
 430 unlock:
 431         raw_spin_unlock_bh(&local_storage->lock);
 432         return SDATA(selem);
 433
 434 unlock_err:
 435         raw_spin_unlock_bh(&local_storage->lock);
 436         return ERR_PTR(err);
 437 }
 438
 439 u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 440 {
 441         u64 min_usage = U64_MAX;
 442         u16 i, res = 0;
 443
 444         spin_lock(&cache->idx_lock);
 445
 446         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 447                 if (cache->idx_usage_counts[i] < min_usage) {
 448                         min_usage = cache->idx_usage_counts[i];
 449                         res = i;
 450
 451                         /* Found a free cache_idx */
 452                         if (!min_usage)
 453                                 break;
 454                 }
 455         }
 456         cache->idx_usage_counts[res]++;
 457
 458         spin_unlock(&cache->idx_lock);
 459
 460         return res;
 461 }
 462
 463 void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 464                                       u16 idx)
 465 {
 466         spin_lock(&cache->idx_lock);
 467         cache->idx_usage_counts[idx]--;
 468         spin_unlock(&cache->idx_lock);
 469 }
 470
 471 void bpf_local_storage_map_free(struct bpf_local_storage_map *smap)
 472 {
 473         struct bpf_local_storage_elem *selem;
 474         struct bpf_local_storage_map_bucket *b;
 475         unsigned int i;
 476
 477         /* Note that this map might be concurrently cloned from
 478          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 479          * RCU read section to finish before proceeding. New RCU
 480          * read sections should be prevented via bpf_map_inc_not_zero.
 481          */
 482         synchronize_rcu();
 483
 484         /* bpf prog and the userspace can no longer access this map
 485          * now.  No new selem (of this map) can be added
 486          * to the owner->storage or to the map bucket's list.
 487          *
 488          * The elem of this map can be cleaned up here
 489          * or when the storage is freed e.g.
 490          * by bpf_sk_storage_free() during __sk_destruct().
 491          */
 492         for (i = 0; i < (1U << smap->bucket_log); i++) {
 493                 b = &smap->buckets[i];
 494
 495                 rcu_read_lock();
 496                 /* No one is adding to b->list now */
 497                 while ((selem = hlist_entry_safe(
 498                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 499                                 struct bpf_local_storage_elem, map_node))) {
 500                         bpf_selem_unlink(selem);
 501                         cond_resched_rcu();
 502                 }
 503                 rcu_read_unlock();
 504         }
 505
 506         /* While freeing the storage we may still need to access the map.
 507          *
 508          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 509          * which then made the above while((selem = ...)) loop
 510          * exit immediately.
 511          *
 512          * However, while freeing the storage one still needs to access the
 513          * smap->elem_size to do the uncharging in
 514          * bpf_selem_unlink_storage_nolock().
 515          *
 516          * Hence, wait another rcu grace period for the storage to be freed.
 517          */
 518         synchronize_rcu();
 519
 520         kvfree(smap->buckets);
 521         kfree(smap);
 522 }
 523
 524 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 525 {
 526         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 527             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 528             attr->max_entries ||
 529             attr->key_size != sizeof(int) || !attr->value_size ||
 530             /* Enforce BTF for userspace sk dumping */
 531             !attr->btf_key_type_id || !attr->btf_value_type_id)
 532                 return -EINVAL;
 533
 534         if (!bpf_capable())
 535                 return -EPERM;
 536
 537         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 538                 return -E2BIG;
 539
 540         return 0;
 541 }
 542
 543 struct bpf_local_storage_map *bpf_local_storage_map_alloc(union bpf_attr *attr)
 544 {
 545         struct bpf_local_storage_map *smap;
 546         unsigned int i;
 547         u32 nbuckets;
 548
 549         smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 550         if (!smap)
 551                 return ERR_PTR(-ENOMEM);
 552         bpf_map_init_from_attr(&smap->map, attr);
 553
 554         nbuckets = roundup_pow_of_two(num_possible_cpus());
 555         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 556         nbuckets = max_t(u32, 2, nbuckets);
 557         smap->bucket_log = ilog2(nbuckets);
 558
 559         smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets,
 560                                  GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 561         if (!smap->buckets) {
 562                 kfree(smap);
 563                 return ERR_PTR(-ENOMEM);
 564         }
 565
 566         for (i = 0; i < nbuckets; i++) {
 567                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 568                 raw_spin_lock_init(&smap->buckets[i].lock);
 569         }
 570
 571         smap->elem_size =
 572                 sizeof(struct bpf_local_storage_elem) + attr->value_size;
 573
 574         return smap;
 575 }
 576
 577 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 578                                     const struct btf *btf,
 579                                     const struct btf_type *key_type,
 580                                     const struct btf_type *value_type)
 581 {
 582         u32 int_data;
 583
 584         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 585                 return -EINVAL;
 586
 587         int_data = *(u32 *)(key_type + 1);
 588         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 589                 return -EINVAL;
 590
 591         return 0;
 592 }