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         unsigned long flags;
 144
 145         if (unlikely(!selem_linked_to_storage(selem)))
 146                 /* selem has already been unlinked from sk */
 147                 return;
 148
 149         local_storage = rcu_dereference(selem->local_storage);
 150         raw_spin_lock_irqsave(&local_storage->lock, flags);
 151         if (likely(selem_linked_to_storage(selem)))
 152                 free_local_storage = bpf_selem_unlink_storage_nolock(
 153                         local_storage, selem, true);
 154         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 155
 156         if (free_local_storage)
 157                 kfree_rcu(local_storage, rcu);
 158 }
 159
 160 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 161                                    struct bpf_local_storage_elem *selem)
 162 {
 163         RCU_INIT_POINTER(selem->local_storage, local_storage);
 164         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 165 }
 166
 167 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 168 {
 169         struct bpf_local_storage_map *smap;
 170         struct bpf_local_storage_map_bucket *b;
 171         unsigned long flags;
 172
 173         if (unlikely(!selem_linked_to_map(selem)))
 174                 /* selem has already be unlinked from smap */
 175                 return;
 176
 177         smap = rcu_dereference(SDATA(selem)->smap);
 178         b = select_bucket(smap, selem);
 179         raw_spin_lock_irqsave(&b->lock, flags);
 180         if (likely(selem_linked_to_map(selem)))
 181                 hlist_del_init_rcu(&selem->map_node);
 182         raw_spin_unlock_irqrestore(&b->lock, flags);
 183 }
 184
 185 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 186                         struct bpf_local_storage_elem *selem)
 187 {
 188         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 189         unsigned long flags;
 190
 191         raw_spin_lock_irqsave(&b->lock, flags);
 192         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 193         hlist_add_head_rcu(&selem->map_node, &b->list);
 194         raw_spin_unlock_irqrestore(&b->lock, flags);
 195 }
 196
 197 void bpf_selem_unlink(struct bpf_local_storage_elem *selem)
 198 {
 199         /* Always unlink from map before unlinking from local_storage
 200          * because selem will be freed after successfully unlinked from
 201          * the local_storage.
 202          */
 203         bpf_selem_unlink_map(selem);
 204         __bpf_selem_unlink_storage(selem);
 205 }
 206
 207 struct bpf_local_storage_data *
 208 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 209                          struct bpf_local_storage_map *smap,
 210                          bool cacheit_lockit)
 211 {
 212         struct bpf_local_storage_data *sdata;
 213         struct bpf_local_storage_elem *selem;
 214
 215         /* Fast path (cache hit) */
 216         sdata = rcu_dereference(local_storage->cache[smap->cache_idx]);
 217         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 218                 return sdata;
 219
 220         /* Slow path (cache miss) */
 221         hlist_for_each_entry_rcu(selem, &local_storage->list, snode)
 222                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 223                         break;
 224
 225         if (!selem)
 226                 return NULL;
 227
 228         sdata = SDATA(selem);
 229         if (cacheit_lockit) {
 230                 unsigned long flags;
 231
 232                 /* spinlock is needed to avoid racing with the
 233                  * parallel delete.  Otherwise, publishing an already
 234                  * deleted sdata to the cache will become a use-after-free
 235                  * problem in the next bpf_local_storage_lookup().
 236                  */
 237                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 238                 if (selem_linked_to_storage(selem))
 239                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 240                                            sdata);
 241                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 242         }
 243
 244         return sdata;
 245 }
 246
 247 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 248                        u64 map_flags)
 249 {
 250         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 251                 /* elem already exists */
 252                 return -EEXIST;
 253
 254         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 255                 /* elem doesn't exist, cannot update it */
 256                 return -ENOENT;
 257
 258         return 0;
 259 }
 260
 261 int bpf_local_storage_alloc(void *owner,
 262                             struct bpf_local_storage_map *smap,
 263                             struct bpf_local_storage_elem *first_selem)
 264 {
 265         struct bpf_local_storage *prev_storage, *storage;
 266         struct bpf_local_storage **owner_storage_ptr;
 267         int err;
 268
 269         err = mem_charge(smap, owner, sizeof(*storage));
 270         if (err)
 271                 return err;
 272
 273         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 274                                   GFP_ATOMIC | __GFP_NOWARN);
 275         if (!storage) {
 276                 err = -ENOMEM;
 277                 goto uncharge;
 278         }
 279
 280         INIT_HLIST_HEAD(&storage->list);
 281         raw_spin_lock_init(&storage->lock);
 282         storage->owner = owner;
 283
 284         bpf_selem_link_storage_nolock(storage, first_selem);
 285         bpf_selem_link_map(smap, first_selem);
 286
 287         owner_storage_ptr =
 288                 (struct bpf_local_storage **)owner_storage(smap, owner);
 289         /* Publish storage to the owner.
 290          * Instead of using any lock of the kernel object (i.e. owner),
 291          * cmpxchg will work with any kernel object regardless what
 292          * the running context is, bh, irq...etc.
 293          *
 294          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 295          * is protected by the storage->lock.  Hence, when freeing
 296          * the owner->storage, the storage->lock must be held before
 297          * setting owner->storage ptr to NULL.
 298          */
 299         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 300         if (unlikely(prev_storage)) {
 301                 bpf_selem_unlink_map(first_selem);
 302                 err = -EAGAIN;
 303                 goto uncharge;
 304
 305                 /* Note that even first_selem was linked to smap's
 306                  * bucket->list, first_selem can be freed immediately
 307                  * (instead of kfree_rcu) because
 308                  * bpf_local_storage_map_free() does a
 309                  * synchronize_rcu() before walking the bucket->list.
 310                  * Hence, no one is accessing selem from the
 311                  * bucket->list under rcu_read_lock().
 312                  */
 313         }
 314
 315         return 0;
 316
 317 uncharge:
 318         kfree(storage);
 319         mem_uncharge(smap, owner, sizeof(*storage));
 320         return err;
 321 }
 322
 323 /* sk cannot be going away because it is linking new elem
 324  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 325  * Otherwise, it will become a leak (and other memory issues
 326  * during map destruction).
 327  */
 328 struct bpf_local_storage_data *
 329 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 330                          void *value, u64 map_flags)
 331 {
 332         struct bpf_local_storage_data *old_sdata = NULL;
 333         struct bpf_local_storage_elem *selem;
 334         struct bpf_local_storage *local_storage;
 335         unsigned long flags;
 336         int err;
 337
 338         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 339         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 340             /* BPF_F_LOCK can only be used in a value with spin_lock */
 341             unlikely((map_flags & BPF_F_LOCK) &&
 342                      !map_value_has_spin_lock(&smap->map)))
 343                 return ERR_PTR(-EINVAL);
 344
 345         local_storage = rcu_dereference(*owner_storage(smap, owner));
 346         if (!local_storage || hlist_empty(&local_storage->list)) {
 347                 /* Very first elem for the owner */
 348                 err = check_flags(NULL, map_flags);
 349                 if (err)
 350                         return ERR_PTR(err);
 351
 352                 selem = bpf_selem_alloc(smap, owner, value, true);
 353                 if (!selem)
 354                         return ERR_PTR(-ENOMEM);
 355
 356                 err = bpf_local_storage_alloc(owner, smap, selem);
 357                 if (err) {
 358                         kfree(selem);
 359                         mem_uncharge(smap, owner, smap->elem_size);
 360                         return ERR_PTR(err);
 361                 }
 362
 363                 return SDATA(selem);
 364         }
 365
 366         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 367                 /* Hoping to find an old_sdata to do inline update
 368                  * such that it can avoid taking the local_storage->lock
 369                  * and changing the lists.
 370                  */
 371                 old_sdata =
 372                         bpf_local_storage_lookup(local_storage, smap, false);
 373                 err = check_flags(old_sdata, map_flags);
 374                 if (err)
 375                         return ERR_PTR(err);
 376                 if (old_sdata && selem_linked_to_storage(SELEM(old_sdata))) {
 377                         copy_map_value_locked(&smap->map, old_sdata->data,
 378                                               value, false);
 379                         return old_sdata;
 380                 }
 381         }
 382
 383         raw_spin_lock_irqsave(&local_storage->lock, flags);
 384
 385         /* Recheck local_storage->list under local_storage->lock */
 386         if (unlikely(hlist_empty(&local_storage->list))) {
 387                 /* A parallel del is happening and local_storage is going
 388                  * away.  It has just been checked before, so very
 389                  * unlikely.  Return instead of retry to keep things
 390                  * simple.
 391                  */
 392                 err = -EAGAIN;
 393                 goto unlock_err;
 394         }
 395
 396         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 397         err = check_flags(old_sdata, map_flags);
 398         if (err)
 399                 goto unlock_err;
 400
 401         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 402                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 403                                       false);
 404                 selem = SELEM(old_sdata);
 405                 goto unlock;
 406         }
 407
 408         /* local_storage->lock is held.  Hence, we are sure
 409          * we can unlink and uncharge the old_sdata successfully
 410          * later.  Hence, instead of charging the new selem now
 411          * and then uncharge the old selem later (which may cause
 412          * a potential but unnecessary charge failure),  avoid taking
 413          * a charge at all here (the "!old_sdata" check) and the
 414          * old_sdata will not be uncharged later during
 415          * bpf_selem_unlink_storage_nolock().
 416          */
 417         selem = bpf_selem_alloc(smap, owner, value, !old_sdata);
 418         if (!selem) {
 419                 err = -ENOMEM;
 420                 goto unlock_err;
 421         }
 422
 423         /* First, link the new selem to the map */
 424         bpf_selem_link_map(smap, selem);
 425
 426         /* Second, link (and publish) the new selem to local_storage */
 427         bpf_selem_link_storage_nolock(local_storage, selem);
 428
 429         /* Third, remove old selem, SELEM(old_sdata) */
 430         if (old_sdata) {
 431                 bpf_selem_unlink_map(SELEM(old_sdata));
 432                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 433                                                 false);
 434         }
 435
 436 unlock:
 437         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 438         return SDATA(selem);
 439
 440 unlock_err:
 441         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 442         return ERR_PTR(err);
 443 }
 444
 445 u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 446 {
 447         u64 min_usage = U64_MAX;
 448         u16 i, res = 0;
 449
 450         spin_lock(&cache->idx_lock);
 451
 452         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 453                 if (cache->idx_usage_counts[i] < min_usage) {
 454                         min_usage = cache->idx_usage_counts[i];
 455                         res = i;
 456
 457                         /* Found a free cache_idx */
 458                         if (!min_usage)
 459                                 break;
 460                 }
 461         }
 462         cache->idx_usage_counts[res]++;
 463
 464         spin_unlock(&cache->idx_lock);
 465
 466         return res;
 467 }
 468
 469 void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 470                                       u16 idx)
 471 {
 472         spin_lock(&cache->idx_lock);
 473         cache->idx_usage_counts[idx]--;
 474         spin_unlock(&cache->idx_lock);
 475 }
 476
 477 void bpf_local_storage_map_free(struct bpf_local_storage_map *smap,
 478                                 int __percpu *busy_counter)
 479 {
 480         struct bpf_local_storage_elem *selem;
 481         struct bpf_local_storage_map_bucket *b;
 482         unsigned int i;
 483
 484         /* Note that this map might be concurrently cloned from
 485          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 486          * RCU read section to finish before proceeding. New RCU
 487          * read sections should be prevented via bpf_map_inc_not_zero.
 488          */
 489         synchronize_rcu();
 490
 491         /* bpf prog and the userspace can no longer access this map
 492          * now.  No new selem (of this map) can be added
 493          * to the owner->storage or to the map bucket's list.
 494          *
 495          * The elem of this map can be cleaned up here
 496          * or when the storage is freed e.g.
 497          * by bpf_sk_storage_free() during __sk_destruct().
 498          */
 499         for (i = 0; i < (1U << smap->bucket_log); i++) {
 500                 b = &smap->buckets[i];
 501
 502                 rcu_read_lock();
 503                 /* No one is adding to b->list now */
 504                 while ((selem = hlist_entry_safe(
 505                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 506                                 struct bpf_local_storage_elem, map_node))) {
 507                         if (busy_counter) {
 508                                 migrate_disable();
 509                                 __this_cpu_inc(*busy_counter);
 510                         }
 511                         bpf_selem_unlink(selem);
 512                         if (busy_counter) {
 513                                 __this_cpu_dec(*busy_counter);
 514                                 migrate_enable();
 515                         }
 516                         cond_resched_rcu();
 517                 }
 518                 rcu_read_unlock();
 519         }
 520
 521         /* While freeing the storage we may still need to access the map.
 522          *
 523          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 524          * which then made the above while((selem = ...)) loop
 525          * exit immediately.
 526          *
 527          * However, while freeing the storage one still needs to access the
 528          * smap->elem_size to do the uncharging in
 529          * bpf_selem_unlink_storage_nolock().
 530          *
 531          * Hence, wait another rcu grace period for the storage to be freed.
 532          */
 533         synchronize_rcu();
 534
 535         kvfree(smap->buckets);
 536         kfree(smap);
 537 }
 538
 539 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 540 {
 541         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 542             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 543             attr->max_entries ||
 544             attr->key_size != sizeof(int) || !attr->value_size ||
 545             /* Enforce BTF for userspace sk dumping */
 546             !attr->btf_key_type_id || !attr->btf_value_type_id)
 547                 return -EINVAL;
 548
 549         if (!bpf_capable())
 550                 return -EPERM;
 551
 552         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 553                 return -E2BIG;
 554
 555         return 0;
 556 }
 557
 558 struct bpf_local_storage_map *bpf_local_storage_map_alloc(union bpf_attr *attr)
 559 {
 560         struct bpf_local_storage_map *smap;
 561         unsigned int i;
 562         u32 nbuckets;
 563
 564         smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 565         if (!smap)
 566                 return ERR_PTR(-ENOMEM);
 567         bpf_map_init_from_attr(&smap->map, attr);
 568
 569         nbuckets = roundup_pow_of_two(num_possible_cpus());
 570         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 571         nbuckets = max_t(u32, 2, nbuckets);
 572         smap->bucket_log = ilog2(nbuckets);
 573
 574         smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets,
 575                                  GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 576         if (!smap->buckets) {
 577                 kfree(smap);
 578                 return ERR_PTR(-ENOMEM);
 579         }
 580
 581         for (i = 0; i < nbuckets; i++) {
 582                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 583                 raw_spin_lock_init(&smap->buckets[i].lock);
 584         }
 585
 586         smap->elem_size =
 587                 sizeof(struct bpf_local_storage_elem) + attr->value_size;
 588
 589         return smap;
 590 }
 591
 592 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 593                                     const struct btf *btf,
 594                                     const struct btf_type *key_type,
 595                                     const struct btf_type *value_type)
 596 {
 597         u32 int_data;
 598
 599         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 600                 return -EINVAL;
 601
 602         int_data = *(u32 *)(key_type + 1);
 603         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 604                 return -EINVAL;
 605
 606         return 0;
 607 }