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