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