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 #include <linux/rcupdate.h>
  15 #include <linux/rcupdate_trace.h>
  16 #include <linux/rcupdate_wait.h>
  17
  18 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE)
  19
  20 static struct bpf_local_storage_map_bucket *
  21 select_bucket(struct bpf_local_storage_map *smap,
  22               struct bpf_local_storage_elem *selem)
  23 {
  24         return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
  25 }
  26
  27 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size)
  28 {
  29         struct bpf_map *map = &smap->map;
  30
  31         if (!map->ops->map_local_storage_charge)
  32                 return 0;
  33
  34         return map->ops->map_local_storage_charge(smap, owner, size);
  35 }
  36
  37 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner,
  38                          u32 size)
  39 {
  40         struct bpf_map *map = &smap->map;
  41
  42         if (map->ops->map_local_storage_uncharge)
  43                 map->ops->map_local_storage_uncharge(smap, owner, size);
  44 }
  45
  46 static struct bpf_local_storage __rcu **
  47 owner_storage(struct bpf_local_storage_map *smap, void *owner)
  48 {
  49         struct bpf_map *map = &smap->map;
  50
  51         return map->ops->map_owner_storage_ptr(owner);
  52 }
  53
  54 static bool selem_linked_to_storage_lockless(const struct bpf_local_storage_elem *selem)
  55 {
  56         return !hlist_unhashed_lockless(&selem->snode);
  57 }
  58
  59 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
  60 {
  61         return !hlist_unhashed(&selem->snode);
  62 }
  63
  64 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
  65 {
  66         return !hlist_unhashed_lockless(&selem->map_node);
  67 }
  68
  69 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
  70 {
  71         return !hlist_unhashed(&selem->map_node);
  72 }
  73
  74 struct bpf_local_storage_elem *
  75 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
  76                 void *value, bool charge_mem, gfp_t gfp_flags)
  77 {
  78         struct bpf_local_storage_elem *selem;
  79
  80         if (charge_mem && mem_charge(smap, owner, smap->elem_size))
  81                 return NULL;
  82
  83         selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
  84                                 gfp_flags | __GFP_NOWARN);
  85         if (selem) {
  86                 if (value)
  87                         copy_map_value(&smap->map, SDATA(selem)->data, value);
  88                 /* No need to call check_and_init_map_value as memory is zero init */
  89                 return selem;
  90         }
  91
  92         if (charge_mem)
  93                 mem_uncharge(smap, owner, smap->elem_size);
  94
  95         return NULL;
  96 }
  97
  98 void bpf_local_storage_free_rcu(struct rcu_head *rcu)
  99 {
 100         struct bpf_local_storage *local_storage;
 101
 102         /* If RCU Tasks Trace grace period implies RCU grace period, do
 103          * kfree(), else do kfree_rcu().
 104          */
 105         local_storage = container_of(rcu, struct bpf_local_storage, rcu);
 106         if (rcu_trace_implies_rcu_gp())
 107                 kfree(local_storage);
 108         else
 109                 kfree_rcu(local_storage, rcu);
 110 }
 111
 112 static void bpf_selem_free_fields_rcu(struct rcu_head *rcu)
 113 {
 114         struct bpf_local_storage_elem *selem;
 115         struct bpf_local_storage_map *smap;
 116
 117         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 118         /* protected by the rcu_barrier*() */
 119         smap = rcu_dereference_protected(SDATA(selem)->smap, true);
 120         bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
 121         kfree(selem);
 122 }
 123
 124 static void bpf_selem_free_fields_trace_rcu(struct rcu_head *rcu)
 125 {
 126         /* Free directly if Tasks Trace RCU GP also implies RCU GP */
 127         if (rcu_trace_implies_rcu_gp())
 128                 bpf_selem_free_fields_rcu(rcu);
 129         else
 130                 call_rcu(rcu, bpf_selem_free_fields_rcu);
 131 }
 132
 133 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 134 {
 135         struct bpf_local_storage_elem *selem;
 136
 137         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 138         if (rcu_trace_implies_rcu_gp())
 139                 kfree(selem);
 140         else
 141                 kfree_rcu(selem, rcu);
 142 }
 143
 144 /* local_storage->lock must be held and selem->local_storage == local_storage.
 145  * The caller must ensure selem->smap is still valid to be
 146  * dereferenced for its smap->elem_size and smap->cache_idx.
 147  */
 148 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
 149                                             struct bpf_local_storage_elem *selem,
 150                                             bool uncharge_mem, bool use_trace_rcu)
 151 {
 152         struct bpf_local_storage_map *smap;
 153         bool free_local_storage;
 154         struct btf_record *rec;
 155         void *owner;
 156
 157         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 158         owner = local_storage->owner;
 159
 160         /* All uncharging on the owner must be done first.
 161          * The owner may be freed once the last selem is unlinked
 162          * from local_storage.
 163          */
 164         if (uncharge_mem)
 165                 mem_uncharge(smap, owner, smap->elem_size);
 166
 167         free_local_storage = hlist_is_singular_node(&selem->snode,
 168                                                     &local_storage->list);
 169         if (free_local_storage) {
 170                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 171                 local_storage->owner = NULL;
 172
 173                 /* After this RCU_INIT, owner may be freed and cannot be used */
 174                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 175
 176                 /* local_storage is not freed now.  local_storage->lock is
 177                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 178                  * will be done by the caller.
 179                  *
 180                  * Although the unlock will be done under
 181                  * rcu_read_lock(),  it is more intuitive to
 182                  * read if the freeing of the storage is done
 183                  * after the raw_spin_unlock_bh(&local_storage->lock).
 184                  *
 185                  * Hence, a "bool free_local_storage" is returned
 186                  * to the caller which then calls then frees the storage after
 187                  * all the RCU grace periods have expired.
 188                  */
 189         }
 190         hlist_del_init_rcu(&selem->snode);
 191         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 192             SDATA(selem))
 193                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 194
 195         /* A different RCU callback is chosen whenever we need to free
 196          * additional fields in selem data before freeing selem.
 197          * bpf_local_storage_map_free only executes rcu_barrier to wait for RCU
 198          * callbacks when it has special fields, hence we can only conditionally
 199          * dereference smap, as by this time the map might have already been
 200          * freed without waiting for our call_rcu callback if it did not have
 201          * any special fields.
 202          */
 203         rec = smap->map.record;
 204         if (use_trace_rcu) {
 205                 if (!IS_ERR_OR_NULL(rec))
 206                         call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_fields_trace_rcu);
 207                 else
 208                         call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
 209         } else {
 210                 if (!IS_ERR_OR_NULL(rec))
 211                         call_rcu(&selem->rcu, bpf_selem_free_fields_rcu);
 212                 else
 213                         kfree_rcu(selem, rcu);
 214         }
 215
 216         return free_local_storage;
 217 }
 218
 219 static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
 220                                        bool use_trace_rcu)
 221 {
 222         struct bpf_local_storage *local_storage;
 223         bool free_local_storage = false;
 224         unsigned long flags;
 225
 226         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 227                 /* selem has already been unlinked from sk */
 228                 return;
 229
 230         local_storage = rcu_dereference_check(selem->local_storage,
 231                                               bpf_rcu_lock_held());
 232         raw_spin_lock_irqsave(&local_storage->lock, flags);
 233         if (likely(selem_linked_to_storage(selem)))
 234                 free_local_storage = bpf_selem_unlink_storage_nolock(
 235                         local_storage, selem, true, use_trace_rcu);
 236         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 237
 238         if (free_local_storage) {
 239                 if (use_trace_rcu)
 240                         call_rcu_tasks_trace(&local_storage->rcu,
 241                                      bpf_local_storage_free_rcu);
 242                 else
 243                         kfree_rcu(local_storage, rcu);
 244         }
 245 }
 246
 247 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 248                                    struct bpf_local_storage_elem *selem)
 249 {
 250         RCU_INIT_POINTER(selem->local_storage, local_storage);
 251         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 252 }
 253
 254 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 255 {
 256         struct bpf_local_storage_map *smap;
 257         struct bpf_local_storage_map_bucket *b;
 258         unsigned long flags;
 259
 260         if (unlikely(!selem_linked_to_map_lockless(selem)))
 261                 /* selem has already be unlinked from smap */
 262                 return;
 263
 264         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 265         b = select_bucket(smap, selem);
 266         raw_spin_lock_irqsave(&b->lock, flags);
 267         if (likely(selem_linked_to_map(selem)))
 268                 hlist_del_init_rcu(&selem->map_node);
 269         raw_spin_unlock_irqrestore(&b->lock, flags);
 270 }
 271
 272 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 273                         struct bpf_local_storage_elem *selem)
 274 {
 275         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 276         unsigned long flags;
 277
 278         raw_spin_lock_irqsave(&b->lock, flags);
 279         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 280         hlist_add_head_rcu(&selem->map_node, &b->list);
 281         raw_spin_unlock_irqrestore(&b->lock, flags);
 282 }
 283
 284 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool use_trace_rcu)
 285 {
 286         /* Always unlink from map before unlinking from local_storage
 287          * because selem will be freed after successfully unlinked from
 288          * the local_storage.
 289          */
 290         bpf_selem_unlink_map(selem);
 291         __bpf_selem_unlink_storage(selem, use_trace_rcu);
 292 }
 293
 294 /* If cacheit_lockit is false, this lookup function is lockless */
 295 struct bpf_local_storage_data *
 296 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 297                          struct bpf_local_storage_map *smap,
 298                          bool cacheit_lockit)
 299 {
 300         struct bpf_local_storage_data *sdata;
 301         struct bpf_local_storage_elem *selem;
 302
 303         /* Fast path (cache hit) */
 304         sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
 305                                       bpf_rcu_lock_held());
 306         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 307                 return sdata;
 308
 309         /* Slow path (cache miss) */
 310         hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
 311                                   rcu_read_lock_trace_held())
 312                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 313                         break;
 314
 315         if (!selem)
 316                 return NULL;
 317
 318         sdata = SDATA(selem);
 319         if (cacheit_lockit) {
 320                 unsigned long flags;
 321
 322                 /* spinlock is needed to avoid racing with the
 323                  * parallel delete.  Otherwise, publishing an already
 324                  * deleted sdata to the cache will become a use-after-free
 325                  * problem in the next bpf_local_storage_lookup().
 326                  */
 327                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 328                 if (selem_linked_to_storage(selem))
 329                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 330                                            sdata);
 331                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 332         }
 333
 334         return sdata;
 335 }
 336
 337 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 338                        u64 map_flags)
 339 {
 340         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 341                 /* elem already exists */
 342                 return -EEXIST;
 343
 344         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 345                 /* elem doesn't exist, cannot update it */
 346                 return -ENOENT;
 347
 348         return 0;
 349 }
 350
 351 int bpf_local_storage_alloc(void *owner,
 352                             struct bpf_local_storage_map *smap,
 353                             struct bpf_local_storage_elem *first_selem,
 354                             gfp_t gfp_flags)
 355 {
 356         struct bpf_local_storage *prev_storage, *storage;
 357         struct bpf_local_storage **owner_storage_ptr;
 358         int err;
 359
 360         err = mem_charge(smap, owner, sizeof(*storage));
 361         if (err)
 362                 return err;
 363
 364         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 365                                   gfp_flags | __GFP_NOWARN);
 366         if (!storage) {
 367                 err = -ENOMEM;
 368                 goto uncharge;
 369         }
 370
 371         INIT_HLIST_HEAD(&storage->list);
 372         raw_spin_lock_init(&storage->lock);
 373         storage->owner = owner;
 374
 375         bpf_selem_link_storage_nolock(storage, first_selem);
 376         bpf_selem_link_map(smap, first_selem);
 377
 378         owner_storage_ptr =
 379                 (struct bpf_local_storage **)owner_storage(smap, owner);
 380         /* Publish storage to the owner.
 381          * Instead of using any lock of the kernel object (i.e. owner),
 382          * cmpxchg will work with any kernel object regardless what
 383          * the running context is, bh, irq...etc.
 384          *
 385          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 386          * is protected by the storage->lock.  Hence, when freeing
 387          * the owner->storage, the storage->lock must be held before
 388          * setting owner->storage ptr to NULL.
 389          */
 390         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 391         if (unlikely(prev_storage)) {
 392                 bpf_selem_unlink_map(first_selem);
 393                 err = -EAGAIN;
 394                 goto uncharge;
 395
 396                 /* Note that even first_selem was linked to smap's
 397                  * bucket->list, first_selem can be freed immediately
 398                  * (instead of kfree_rcu) because
 399                  * bpf_local_storage_map_free() does a
 400                  * synchronize_rcu_mult (waiting for both sleepable and
 401                  * normal programs) before walking the bucket->list.
 402                  * Hence, no one is accessing selem from the
 403                  * bucket->list under rcu_read_lock().
 404                  */
 405         }
 406
 407         return 0;
 408
 409 uncharge:
 410         kfree(storage);
 411         mem_uncharge(smap, owner, sizeof(*storage));
 412         return err;
 413 }
 414
 415 /* sk cannot be going away because it is linking new elem
 416  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 417  * Otherwise, it will become a leak (and other memory issues
 418  * during map destruction).
 419  */
 420 struct bpf_local_storage_data *
 421 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 422                          void *value, u64 map_flags, gfp_t gfp_flags)
 423 {
 424         struct bpf_local_storage_data *old_sdata = NULL;
 425         struct bpf_local_storage_elem *selem = NULL;
 426         struct bpf_local_storage *local_storage;
 427         unsigned long flags;
 428         int err;
 429
 430         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 431         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 432             /* BPF_F_LOCK can only be used in a value with spin_lock */
 433             unlikely((map_flags & BPF_F_LOCK) &&
 434                      !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
 435                 return ERR_PTR(-EINVAL);
 436
 437         if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
 438                 return ERR_PTR(-EINVAL);
 439
 440         local_storage = rcu_dereference_check(*owner_storage(smap, owner),
 441                                               bpf_rcu_lock_held());
 442         if (!local_storage || hlist_empty(&local_storage->list)) {
 443                 /* Very first elem for the owner */
 444                 err = check_flags(NULL, map_flags);
 445                 if (err)
 446                         return ERR_PTR(err);
 447
 448                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 449                 if (!selem)
 450                         return ERR_PTR(-ENOMEM);
 451
 452                 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
 453                 if (err) {
 454                         kfree(selem);
 455                         mem_uncharge(smap, owner, smap->elem_size);
 456                         return ERR_PTR(err);
 457                 }
 458
 459                 return SDATA(selem);
 460         }
 461
 462         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 463                 /* Hoping to find an old_sdata to do inline update
 464                  * such that it can avoid taking the local_storage->lock
 465                  * and changing the lists.
 466                  */
 467                 old_sdata =
 468                         bpf_local_storage_lookup(local_storage, smap, false);
 469                 err = check_flags(old_sdata, map_flags);
 470                 if (err)
 471                         return ERR_PTR(err);
 472                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 473                         copy_map_value_locked(&smap->map, old_sdata->data,
 474                                               value, false);
 475                         return old_sdata;
 476                 }
 477         }
 478
 479         if (gfp_flags == GFP_KERNEL) {
 480                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 481                 if (!selem)
 482                         return ERR_PTR(-ENOMEM);
 483         }
 484
 485         raw_spin_lock_irqsave(&local_storage->lock, flags);
 486
 487         /* Recheck local_storage->list under local_storage->lock */
 488         if (unlikely(hlist_empty(&local_storage->list))) {
 489                 /* A parallel del is happening and local_storage is going
 490                  * away.  It has just been checked before, so very
 491                  * unlikely.  Return instead of retry to keep things
 492                  * simple.
 493                  */
 494                 err = -EAGAIN;
 495                 goto unlock_err;
 496         }
 497
 498         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 499         err = check_flags(old_sdata, map_flags);
 500         if (err)
 501                 goto unlock_err;
 502
 503         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 504                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 505                                       false);
 506                 selem = SELEM(old_sdata);
 507                 goto unlock;
 508         }
 509
 510         if (gfp_flags != GFP_KERNEL) {
 511                 /* local_storage->lock is held.  Hence, we are sure
 512                  * we can unlink and uncharge the old_sdata successfully
 513                  * later.  Hence, instead of charging the new selem now
 514                  * and then uncharge the old selem later (which may cause
 515                  * a potential but unnecessary charge failure),  avoid taking
 516                  * a charge at all here (the "!old_sdata" check) and the
 517                  * old_sdata will not be uncharged later during
 518                  * bpf_selem_unlink_storage_nolock().
 519                  */
 520                 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
 521                 if (!selem) {
 522                         err = -ENOMEM;
 523                         goto unlock_err;
 524                 }
 525         }
 526
 527         /* First, link the new selem to the map */
 528         bpf_selem_link_map(smap, selem);
 529
 530         /* Second, link (and publish) the new selem to local_storage */
 531         bpf_selem_link_storage_nolock(local_storage, selem);
 532
 533         /* Third, remove old selem, SELEM(old_sdata) */
 534         if (old_sdata) {
 535                 bpf_selem_unlink_map(SELEM(old_sdata));
 536                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 537                                                 false, true);
 538         }
 539
 540 unlock:
 541         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 542         return SDATA(selem);
 543
 544 unlock_err:
 545         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 546         if (selem) {
 547                 mem_uncharge(smap, owner, smap->elem_size);
 548                 kfree(selem);
 549         }
 550         return ERR_PTR(err);
 551 }
 552
 553 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 554 {
 555         u64 min_usage = U64_MAX;
 556         u16 i, res = 0;
 557
 558         spin_lock(&cache->idx_lock);
 559
 560         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 561                 if (cache->idx_usage_counts[i] < min_usage) {
 562                         min_usage = cache->idx_usage_counts[i];
 563                         res = i;
 564
 565                         /* Found a free cache_idx */
 566                         if (!min_usage)
 567                                 break;
 568                 }
 569         }
 570         cache->idx_usage_counts[res]++;
 571
 572         spin_unlock(&cache->idx_lock);
 573
 574         return res;
 575 }
 576
 577 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 578                                              u16 idx)
 579 {
 580         spin_lock(&cache->idx_lock);
 581         cache->idx_usage_counts[idx]--;
 582         spin_unlock(&cache->idx_lock);
 583 }
 584
 585 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 586 {
 587         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 588             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 589             attr->max_entries ||
 590             attr->key_size != sizeof(int) || !attr->value_size ||
 591             /* Enforce BTF for userspace sk dumping */
 592             !attr->btf_key_type_id || !attr->btf_value_type_id)
 593                 return -EINVAL;
 594
 595         if (!bpf_capable())
 596                 return -EPERM;
 597
 598         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 599                 return -E2BIG;
 600
 601         return 0;
 602 }
 603
 604 static struct bpf_local_storage_map *__bpf_local_storage_map_alloc(union bpf_attr *attr)
 605 {
 606         struct bpf_local_storage_map *smap;
 607         unsigned int i;
 608         u32 nbuckets;
 609
 610         smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
 611         if (!smap)
 612                 return ERR_PTR(-ENOMEM);
 613         bpf_map_init_from_attr(&smap->map, attr);
 614
 615         nbuckets = roundup_pow_of_two(num_possible_cpus());
 616         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 617         nbuckets = max_t(u32, 2, nbuckets);
 618         smap->bucket_log = ilog2(nbuckets);
 619
 620         smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
 621                                          nbuckets, GFP_USER | __GFP_NOWARN);
 622         if (!smap->buckets) {
 623                 bpf_map_area_free(smap);
 624                 return ERR_PTR(-ENOMEM);
 625         }
 626
 627         for (i = 0; i < nbuckets; i++) {
 628                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 629                 raw_spin_lock_init(&smap->buckets[i].lock);
 630         }
 631
 632         smap->elem_size = offsetof(struct bpf_local_storage_elem,
 633                                    sdata.data[attr->value_size]);
 634
 635         return smap;
 636 }
 637
 638 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 639                                     const struct btf *btf,
 640                                     const struct btf_type *key_type,
 641                                     const struct btf_type *value_type)
 642 {
 643         u32 int_data;
 644
 645         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 646                 return -EINVAL;
 647
 648         int_data = *(u32 *)(key_type + 1);
 649         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 650                 return -EINVAL;
 651
 652         return 0;
 653 }
 654
 655 bool bpf_local_storage_unlink_nolock(struct bpf_local_storage *local_storage)
 656 {
 657         struct bpf_local_storage_elem *selem;
 658         bool free_storage = false;
 659         struct hlist_node *n;
 660
 661         /* Neither the bpf_prog nor the bpf_map's syscall
 662          * could be modifying the local_storage->list now.
 663          * Thus, no elem can be added to or deleted from the
 664          * local_storage->list by the bpf_prog or by the bpf_map's syscall.
 665          *
 666          * It is racing with bpf_local_storage_map_free() alone
 667          * when unlinking elem from the local_storage->list and
 668          * the map's bucket->list.
 669          */
 670         hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
 671                 /* Always unlink from map before unlinking from
 672                  * local_storage.
 673                  */
 674                 bpf_selem_unlink_map(selem);
 675                 /* If local_storage list has only one element, the
 676                  * bpf_selem_unlink_storage_nolock() will return true.
 677                  * Otherwise, it will return false. The current loop iteration
 678                  * intends to remove all local storage. So the last iteration
 679                  * of the loop will set the free_cgroup_storage to true.
 680                  */
 681                 free_storage = bpf_selem_unlink_storage_nolock(
 682                         local_storage, selem, false, false);
 683         }
 684
 685         return free_storage;
 686 }
 687
 688 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
 689 {
 690         struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
 691         u64 usage = sizeof(*smap);
 692
 693         /* The dynamically callocated selems are not counted currently. */
 694         usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
 695         return usage;
 696 }
 697
 698 struct bpf_map *
 699 bpf_local_storage_map_alloc(union bpf_attr *attr,
 700                             struct bpf_local_storage_cache *cache)
 701 {
 702         struct bpf_local_storage_map *smap;
 703
 704         smap = __bpf_local_storage_map_alloc(attr);
 705         if (IS_ERR(smap))
 706                 return ERR_CAST(smap);
 707
 708         smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
 709         return &smap->map;
 710 }
 711
 712 void bpf_local_storage_map_free(struct bpf_map *map,
 713                                 struct bpf_local_storage_cache *cache,
 714                                 int __percpu *busy_counter)
 715 {
 716         struct bpf_local_storage_map_bucket *b;
 717         struct bpf_local_storage_elem *selem;
 718         struct bpf_local_storage_map *smap;
 719         unsigned int i;
 720
 721         smap = (struct bpf_local_storage_map *)map;
 722         bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
 723
 724         /* Note that this map might be concurrently cloned from
 725          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 726          * RCU read section to finish before proceeding. New RCU
 727          * read sections should be prevented via bpf_map_inc_not_zero.
 728          */
 729         synchronize_rcu();
 730
 731         /* bpf prog and the userspace can no longer access this map
 732          * now.  No new selem (of this map) can be added
 733          * to the owner->storage or to the map bucket's list.
 734          *
 735          * The elem of this map can be cleaned up here
 736          * or when the storage is freed e.g.
 737          * by bpf_sk_storage_free() during __sk_destruct().
 738          */
 739         for (i = 0; i < (1U << smap->bucket_log); i++) {
 740                 b = &smap->buckets[i];
 741
 742                 rcu_read_lock();
 743                 /* No one is adding to b->list now */
 744                 while ((selem = hlist_entry_safe(
 745                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 746                                 struct bpf_local_storage_elem, map_node))) {
 747                         if (busy_counter) {
 748                                 migrate_disable();
 749                                 this_cpu_inc(*busy_counter);
 750                         }
 751                         bpf_selem_unlink(selem, false);
 752                         if (busy_counter) {
 753                                 this_cpu_dec(*busy_counter);
 754                                 migrate_enable();
 755                         }
 756                         cond_resched_rcu();
 757                 }
 758                 rcu_read_unlock();
 759         }
 760
 761         /* While freeing the storage we may still need to access the map.
 762          *
 763          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 764          * which then made the above while((selem = ...)) loop
 765          * exit immediately.
 766          *
 767          * However, while freeing the storage one still needs to access the
 768          * smap->elem_size to do the uncharging in
 769          * bpf_selem_unlink_storage_nolock().
 770          *
 771          * Hence, wait another rcu grace period for the storage to be freed.
 772          */
 773         synchronize_rcu();
 774
 775         /* Only delay freeing of smap, buckets are not needed anymore */
 776         kvfree(smap->buckets);
 777
 778         /* When local storage has special fields, callbacks for
 779          * bpf_selem_free_fields_rcu and bpf_selem_free_fields_trace_rcu will
 780          * keep using the map BTF record, we need to execute an RCU barrier to
 781          * wait for them as the record will be freed right after our map_free
 782          * callback.
 783          */
 784         if (!IS_ERR_OR_NULL(smap->map.record)) {
 785                 rcu_barrier_tasks_trace();
 786                 /* We cannot skip rcu_barrier() when rcu_trace_implies_rcu_gp()
 787                  * is true, because while call_rcu invocation is skipped in that
 788                  * case in bpf_selem_free_fields_trace_rcu (and all local
 789                  * storage maps pass use_trace_rcu = true), there can be
 790                  * call_rcu callbacks based on use_trace_rcu = false in the
 791                  * while ((selem = ...)) loop above or when owner's free path
 792                  * calls bpf_local_storage_unlink_nolock.
 793                  */
 794                 rcu_barrier();
 795         }
 796         bpf_map_area_free(smap);
 797 }