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_lockless(const struct bpf_local_storage_elem *selem)
  52 {
  53         return !hlist_unhashed_lockless(&selem->snode);
  54 }
  55
  56 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
  57 {
  58         return !hlist_unhashed(&selem->snode);
  59 }
  60
  61 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
  62 {
  63         return !hlist_unhashed_lockless(&selem->map_node);
  64 }
  65
  66 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
  67 {
  68         return !hlist_unhashed(&selem->map_node);
  69 }
  70
  71 struct bpf_local_storage_elem *
  72 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
  73                 void *value, bool charge_mem)
  74 {
  75         struct bpf_local_storage_elem *selem;
  76
  77         if (charge_mem && mem_charge(smap, owner, smap->elem_size))
  78                 return NULL;
  79
  80         selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
  81                                 GFP_ATOMIC | __GFP_NOWARN);
  82         if (selem) {
  83                 if (value)
  84                         copy_map_value(&smap->map, SDATA(selem)->data, value);
  85                 return selem;
  86         }
  87
  88         if (charge_mem)
  89                 mem_uncharge(smap, owner, smap->elem_size);
  90
  91         return NULL;
  92 }
  93
  94 /* local_storage->lock must be held and selem->local_storage == local_storage.
  95  * The caller must ensure selem->smap is still valid to be
  96  * dereferenced for its smap->elem_size and smap->cache_idx.
  97  */
  98 bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
  99                                      struct bpf_local_storage_elem *selem,
 100                                      bool uncharge_mem)
 101 {
 102         struct bpf_local_storage_map *smap;
 103         bool free_local_storage;
 104         void *owner;
 105
 106         smap = rcu_dereference(SDATA(selem)->smap);
 107         owner = local_storage->owner;
 108
 109         /* All uncharging on the owner must be done first.
 110          * The owner may be freed once the last selem is unlinked
 111          * from local_storage.
 112          */
 113         if (uncharge_mem)
 114                 mem_uncharge(smap, owner, smap->elem_size);
 115
 116         free_local_storage = hlist_is_singular_node(&selem->snode,
 117                                                     &local_storage->list);
 118         if (free_local_storage) {
 119                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 120                 local_storage->owner = NULL;
 121
 122                 /* After this RCU_INIT, owner may be freed and cannot be used */
 123                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 124
 125                 /* local_storage is not freed now.  local_storage->lock is
 126                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 127                  * will be done by the caller.
 128                  *
 129                  * Although the unlock will be done under
 130                  * rcu_read_lock(),  it is more intutivie to
 131                  * read if kfree_rcu(local_storage, rcu) is done
 132                  * after the raw_spin_unlock_bh(&local_storage->lock).
 133                  *
 134                  * Hence, a "bool free_local_storage" is returned
 135                  * to the caller which then calls the kfree_rcu()
 136                  * after unlock.
 137                  */
 138         }
 139         hlist_del_init_rcu(&selem->snode);
 140         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 141             SDATA(selem))
 142                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 143
 144         kfree_rcu(selem, rcu);
 145
 146         return free_local_storage;
 147 }
 148
 149 static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem)
 150 {
 151         struct bpf_local_storage *local_storage;
 152         bool free_local_storage = false;
 153         unsigned long flags;
 154
 155         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 156                 /* selem has already been unlinked from sk */
 157                 return;
 158
 159         local_storage = rcu_dereference(selem->local_storage);
 160         raw_spin_lock_irqsave(&local_storage->lock, flags);
 161         if (likely(selem_linked_to_storage(selem)))
 162                 free_local_storage = bpf_selem_unlink_storage_nolock(
 163                         local_storage, selem, true);
 164         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 165
 166         if (free_local_storage)
 167                 kfree_rcu(local_storage, rcu);
 168 }
 169
 170 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 171                                    struct bpf_local_storage_elem *selem)
 172 {
 173         RCU_INIT_POINTER(selem->local_storage, local_storage);
 174         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 175 }
 176
 177 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 178 {
 179         struct bpf_local_storage_map *smap;
 180         struct bpf_local_storage_map_bucket *b;
 181         unsigned long flags;
 182
 183         if (unlikely(!selem_linked_to_map_lockless(selem)))
 184                 /* selem has already be unlinked from smap */
 185                 return;
 186
 187         smap = rcu_dereference(SDATA(selem)->smap);
 188         b = select_bucket(smap, selem);
 189         raw_spin_lock_irqsave(&b->lock, flags);
 190         if (likely(selem_linked_to_map(selem)))
 191                 hlist_del_init_rcu(&selem->map_node);
 192         raw_spin_unlock_irqrestore(&b->lock, flags);
 193 }
 194
 195 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 196                         struct bpf_local_storage_elem *selem)
 197 {
 198         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 199         unsigned long flags;
 200
 201         raw_spin_lock_irqsave(&b->lock, flags);
 202         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 203         hlist_add_head_rcu(&selem->map_node, &b->list);
 204         raw_spin_unlock_irqrestore(&b->lock, flags);
 205 }
 206
 207 void bpf_selem_unlink(struct bpf_local_storage_elem *selem)
 208 {
 209         /* Always unlink from map before unlinking from local_storage
 210          * because selem will be freed after successfully unlinked from
 211          * the local_storage.
 212          */
 213         bpf_selem_unlink_map(selem);
 214         __bpf_selem_unlink_storage(selem);
 215 }
 216
 217 struct bpf_local_storage_data *
 218 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 219                          struct bpf_local_storage_map *smap,
 220                          bool cacheit_lockit)
 221 {
 222         struct bpf_local_storage_data *sdata;
 223         struct bpf_local_storage_elem *selem;
 224
 225         /* Fast path (cache hit) */
 226         sdata = rcu_dereference(local_storage->cache[smap->cache_idx]);
 227         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 228                 return sdata;
 229
 230         /* Slow path (cache miss) */
 231         hlist_for_each_entry_rcu(selem, &local_storage->list, snode)
 232                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 233                         break;
 234
 235         if (!selem)
 236                 return NULL;
 237
 238         sdata = SDATA(selem);
 239         if (cacheit_lockit) {
 240                 unsigned long flags;
 241
 242                 /* spinlock is needed to avoid racing with the
 243                  * parallel delete.  Otherwise, publishing an already
 244                  * deleted sdata to the cache will become a use-after-free
 245                  * problem in the next bpf_local_storage_lookup().
 246                  */
 247                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 248                 if (selem_linked_to_storage(selem))
 249                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 250                                            sdata);
 251                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 252         }
 253
 254         return sdata;
 255 }
 256
 257 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 258                        u64 map_flags)
 259 {
 260         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 261                 /* elem already exists */
 262                 return -EEXIST;
 263
 264         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 265                 /* elem doesn't exist, cannot update it */
 266                 return -ENOENT;
 267
 268         return 0;
 269 }
 270
 271 int bpf_local_storage_alloc(void *owner,
 272                             struct bpf_local_storage_map *smap,
 273                             struct bpf_local_storage_elem *first_selem)
 274 {
 275         struct bpf_local_storage *prev_storage, *storage;
 276         struct bpf_local_storage **owner_storage_ptr;
 277         int err;
 278
 279         err = mem_charge(smap, owner, sizeof(*storage));
 280         if (err)
 281                 return err;
 282
 283         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 284                                   GFP_ATOMIC | __GFP_NOWARN);
 285         if (!storage) {
 286                 err = -ENOMEM;
 287                 goto uncharge;
 288         }
 289
 290         INIT_HLIST_HEAD(&storage->list);
 291         raw_spin_lock_init(&storage->lock);
 292         storage->owner = owner;
 293
 294         bpf_selem_link_storage_nolock(storage, first_selem);
 295         bpf_selem_link_map(smap, first_selem);
 296
 297         owner_storage_ptr =
 298                 (struct bpf_local_storage **)owner_storage(smap, owner);
 299         /* Publish storage to the owner.
 300          * Instead of using any lock of the kernel object (i.e. owner),
 301          * cmpxchg will work with any kernel object regardless what
 302          * the running context is, bh, irq...etc.
 303          *
 304          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 305          * is protected by the storage->lock.  Hence, when freeing
 306          * the owner->storage, the storage->lock must be held before
 307          * setting owner->storage ptr to NULL.
 308          */
 309         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 310         if (unlikely(prev_storage)) {
 311                 bpf_selem_unlink_map(first_selem);
 312                 err = -EAGAIN;
 313                 goto uncharge;
 314
 315                 /* Note that even first_selem was linked to smap's
 316                  * bucket->list, first_selem can be freed immediately
 317                  * (instead of kfree_rcu) because
 318                  * bpf_local_storage_map_free() does a
 319                  * synchronize_rcu() before walking the bucket->list.
 320                  * Hence, no one is accessing selem from the
 321                  * bucket->list under rcu_read_lock().
 322                  */
 323         }
 324
 325         return 0;
 326
 327 uncharge:
 328         kfree(storage);
 329         mem_uncharge(smap, owner, sizeof(*storage));
 330         return err;
 331 }
 332
 333 /* sk cannot be going away because it is linking new elem
 334  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 335  * Otherwise, it will become a leak (and other memory issues
 336  * during map destruction).
 337  */
 338 struct bpf_local_storage_data *
 339 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 340                          void *value, u64 map_flags)
 341 {
 342         struct bpf_local_storage_data *old_sdata = NULL;
 343         struct bpf_local_storage_elem *selem;
 344         struct bpf_local_storage *local_storage;
 345         unsigned long flags;
 346         int err;
 347
 348         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 349         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 350             /* BPF_F_LOCK can only be used in a value with spin_lock */
 351             unlikely((map_flags & BPF_F_LOCK) &&
 352                      !map_value_has_spin_lock(&smap->map)))
 353                 return ERR_PTR(-EINVAL);
 354
 355         local_storage = rcu_dereference(*owner_storage(smap, owner));
 356         if (!local_storage || hlist_empty(&local_storage->list)) {
 357                 /* Very first elem for the owner */
 358                 err = check_flags(NULL, map_flags);
 359                 if (err)
 360                         return ERR_PTR(err);
 361
 362                 selem = bpf_selem_alloc(smap, owner, value, true);
 363                 if (!selem)
 364                         return ERR_PTR(-ENOMEM);
 365
 366                 err = bpf_local_storage_alloc(owner, smap, selem);
 367                 if (err) {
 368                         kfree(selem);
 369                         mem_uncharge(smap, owner, smap->elem_size);
 370                         return ERR_PTR(err);
 371                 }
 372
 373                 return SDATA(selem);
 374         }
 375
 376         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 377                 /* Hoping to find an old_sdata to do inline update
 378                  * such that it can avoid taking the local_storage->lock
 379                  * and changing the lists.
 380                  */
 381                 old_sdata =
 382                         bpf_local_storage_lookup(local_storage, smap, false);
 383                 err = check_flags(old_sdata, map_flags);
 384                 if (err)
 385                         return ERR_PTR(err);
 386                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 387                         copy_map_value_locked(&smap->map, old_sdata->data,
 388                                               value, false);
 389                         return old_sdata;
 390                 }
 391         }
 392
 393         raw_spin_lock_irqsave(&local_storage->lock, flags);
 394
 395         /* Recheck local_storage->list under local_storage->lock */
 396         if (unlikely(hlist_empty(&local_storage->list))) {
 397                 /* A parallel del is happening and local_storage is going
 398                  * away.  It has just been checked before, so very
 399                  * unlikely.  Return instead of retry to keep things
 400                  * simple.
 401                  */
 402                 err = -EAGAIN;
 403                 goto unlock_err;
 404         }
 405
 406         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 407         err = check_flags(old_sdata, map_flags);
 408         if (err)
 409                 goto unlock_err;
 410
 411         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 412                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 413                                       false);
 414                 selem = SELEM(old_sdata);
 415                 goto unlock;
 416         }
 417
 418         /* local_storage->lock is held.  Hence, we are sure
 419          * we can unlink and uncharge the old_sdata successfully
 420          * later.  Hence, instead of charging the new selem now
 421          * and then uncharge the old selem later (which may cause
 422          * a potential but unnecessary charge failure),  avoid taking
 423          * a charge at all here (the "!old_sdata" check) and the
 424          * old_sdata will not be uncharged later during
 425          * bpf_selem_unlink_storage_nolock().
 426          */
 427         selem = bpf_selem_alloc(smap, owner, value, !old_sdata);
 428         if (!selem) {
 429                 err = -ENOMEM;
 430                 goto unlock_err;
 431         }
 432
 433         /* First, link the new selem to the map */
 434         bpf_selem_link_map(smap, selem);
 435
 436         /* Second, link (and publish) the new selem to local_storage */
 437         bpf_selem_link_storage_nolock(local_storage, selem);
 438
 439         /* Third, remove old selem, SELEM(old_sdata) */
 440         if (old_sdata) {
 441                 bpf_selem_unlink_map(SELEM(old_sdata));
 442                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 443                                                 false);
 444         }
 445
 446 unlock:
 447         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 448         return SDATA(selem);
 449
 450 unlock_err:
 451         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 452         return ERR_PTR(err);
 453 }
 454
 455 u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 456 {
 457         u64 min_usage = U64_MAX;
 458         u16 i, res = 0;
 459
 460         spin_lock(&cache->idx_lock);
 461
 462         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 463                 if (cache->idx_usage_counts[i] < min_usage) {
 464                         min_usage = cache->idx_usage_counts[i];
 465                         res = i;
 466
 467                         /* Found a free cache_idx */
 468                         if (!min_usage)
 469                                 break;
 470                 }
 471         }
 472         cache->idx_usage_counts[res]++;
 473
 474         spin_unlock(&cache->idx_lock);
 475
 476         return res;
 477 }
 478
 479 void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 480                                       u16 idx)
 481 {
 482         spin_lock(&cache->idx_lock);
 483         cache->idx_usage_counts[idx]--;
 484         spin_unlock(&cache->idx_lock);
 485 }
 486
 487 void bpf_local_storage_map_free(struct bpf_local_storage_map *smap,
 488                                 int __percpu *busy_counter)
 489 {
 490         struct bpf_local_storage_elem *selem;
 491         struct bpf_local_storage_map_bucket *b;
 492         unsigned int i;
 493
 494         /* Note that this map might be concurrently cloned from
 495          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 496          * RCU read section to finish before proceeding. New RCU
 497          * read sections should be prevented via bpf_map_inc_not_zero.
 498          */
 499         synchronize_rcu();
 500
 501         /* bpf prog and the userspace can no longer access this map
 502          * now.  No new selem (of this map) can be added
 503          * to the owner->storage or to the map bucket's list.
 504          *
 505          * The elem of this map can be cleaned up here
 506          * or when the storage is freed e.g.
 507          * by bpf_sk_storage_free() during __sk_destruct().
 508          */
 509         for (i = 0; i < (1U << smap->bucket_log); i++) {
 510                 b = &smap->buckets[i];
 511
 512                 rcu_read_lock();
 513                 /* No one is adding to b->list now */
 514                 while ((selem = hlist_entry_safe(
 515                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 516                                 struct bpf_local_storage_elem, map_node))) {
 517                         if (busy_counter) {
 518                                 migrate_disable();
 519                                 this_cpu_inc(*busy_counter);
 520                         }
 521                         bpf_selem_unlink(selem);
 522                         if (busy_counter) {
 523                                 this_cpu_dec(*busy_counter);
 524                                 migrate_enable();
 525                         }
 526                         cond_resched_rcu();
 527                 }
 528                 rcu_read_unlock();
 529         }
 530
 531         /* While freeing the storage we may still need to access the map.
 532          *
 533          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 534          * which then made the above while((selem = ...)) loop
 535          * exit immediately.
 536          *
 537          * However, while freeing the storage one still needs to access the
 538          * smap->elem_size to do the uncharging in
 539          * bpf_selem_unlink_storage_nolock().
 540          *
 541          * Hence, wait another rcu grace period for the storage to be freed.
 542          */
 543         synchronize_rcu();
 544
 545         kvfree(smap->buckets);
 546         kfree(smap);
 547 }
 548
 549 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 550 {
 551         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 552             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 553             attr->max_entries ||
 554             attr->key_size != sizeof(int) || !attr->value_size ||
 555             /* Enforce BTF for userspace sk dumping */
 556             !attr->btf_key_type_id || !attr->btf_value_type_id)
 557                 return -EINVAL;
 558
 559         if (!bpf_capable())
 560                 return -EPERM;
 561
 562         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 563                 return -E2BIG;
 564
 565         return 0;
 566 }
 567
 568 struct bpf_local_storage_map *bpf_local_storage_map_alloc(union bpf_attr *attr)
 569 {
 570         struct bpf_local_storage_map *smap;
 571         unsigned int i;
 572         u32 nbuckets;
 573
 574         smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 575         if (!smap)
 576                 return ERR_PTR(-ENOMEM);
 577         bpf_map_init_from_attr(&smap->map, attr);
 578
 579         nbuckets = roundup_pow_of_two(num_possible_cpus());
 580         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 581         nbuckets = max_t(u32, 2, nbuckets);
 582         smap->bucket_log = ilog2(nbuckets);
 583
 584         smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets,
 585                                  GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT);
 586         if (!smap->buckets) {
 587                 kfree(smap);
 588                 return ERR_PTR(-ENOMEM);
 589         }
 590
 591         for (i = 0; i < nbuckets; i++) {
 592                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 593                 raw_spin_lock_init(&smap->buckets[i].lock);
 594         }
 595
 596         smap->elem_size =
 597                 sizeof(struct bpf_local_storage_elem) + attr->value_size;
 598
 599         return smap;
 600 }
 601
 602 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 603                                     const struct btf *btf,
 604                                     const struct btf_type *key_type,
 605                                     const struct btf_type *value_type)
 606 {
 607         u32 int_data;
 608
 609         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 610                 return -EINVAL;
 611
 612         int_data = *(u32 *)(key_type + 1);
 613         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 614                 return -EINVAL;
 615
 616         return 0;
 617 }