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         /* 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         if (rcu_access_pointer(local_storage->smap) == smap)
 217                 RCU_INIT_POINTER(local_storage->smap, NULL);
 218
 219         return free_local_storage;
 220 }
 221
 222 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
 223                                      bool use_trace_rcu)
 224 {
 225         struct bpf_local_storage *local_storage;
 226         bool free_local_storage = false;
 227         unsigned long flags;
 228
 229         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 230                 /* selem has already been unlinked from sk */
 231                 return;
 232
 233         local_storage = rcu_dereference_check(selem->local_storage,
 234                                               bpf_rcu_lock_held());
 235         raw_spin_lock_irqsave(&local_storage->lock, flags);
 236         if (likely(selem_linked_to_storage(selem)))
 237                 free_local_storage = bpf_selem_unlink_storage_nolock(
 238                         local_storage, selem, true, use_trace_rcu);
 239         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 240
 241         if (free_local_storage) {
 242                 if (use_trace_rcu)
 243                         call_rcu_tasks_trace(&local_storage->rcu,
 244                                      bpf_local_storage_free_rcu);
 245                 else
 246                         kfree_rcu(local_storage, rcu);
 247         }
 248 }
 249
 250 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 251                                    struct bpf_local_storage_elem *selem)
 252 {
 253         RCU_INIT_POINTER(selem->local_storage, local_storage);
 254         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 255 }
 256
 257 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 258 {
 259         struct bpf_local_storage_map *smap;
 260         struct bpf_local_storage_map_bucket *b;
 261         unsigned long flags;
 262
 263         if (unlikely(!selem_linked_to_map_lockless(selem)))
 264                 /* selem has already be unlinked from smap */
 265                 return;
 266
 267         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 268         b = select_bucket(smap, selem);
 269         raw_spin_lock_irqsave(&b->lock, flags);
 270         if (likely(selem_linked_to_map(selem)))
 271                 hlist_del_init_rcu(&selem->map_node);
 272         raw_spin_unlock_irqrestore(&b->lock, flags);
 273 }
 274
 275 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 276                         struct bpf_local_storage_elem *selem)
 277 {
 278         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 279         unsigned long flags;
 280
 281         raw_spin_lock_irqsave(&b->lock, flags);
 282         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 283         hlist_add_head_rcu(&selem->map_node, &b->list);
 284         raw_spin_unlock_irqrestore(&b->lock, flags);
 285 }
 286
 287 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool use_trace_rcu)
 288 {
 289         /* Always unlink from map before unlinking from local_storage
 290          * because selem will be freed after successfully unlinked from
 291          * the local_storage.
 292          */
 293         bpf_selem_unlink_map(selem);
 294         bpf_selem_unlink_storage(selem, use_trace_rcu);
 295 }
 296
 297 /* If cacheit_lockit is false, this lookup function is lockless */
 298 struct bpf_local_storage_data *
 299 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 300                          struct bpf_local_storage_map *smap,
 301                          bool cacheit_lockit)
 302 {
 303         struct bpf_local_storage_data *sdata;
 304         struct bpf_local_storage_elem *selem;
 305
 306         /* Fast path (cache hit) */
 307         sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
 308                                       bpf_rcu_lock_held());
 309         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 310                 return sdata;
 311
 312         /* Slow path (cache miss) */
 313         hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
 314                                   rcu_read_lock_trace_held())
 315                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 316                         break;
 317
 318         if (!selem)
 319                 return NULL;
 320
 321         sdata = SDATA(selem);
 322         if (cacheit_lockit) {
 323                 unsigned long flags;
 324
 325                 /* spinlock is needed to avoid racing with the
 326                  * parallel delete.  Otherwise, publishing an already
 327                  * deleted sdata to the cache will become a use-after-free
 328                  * problem in the next bpf_local_storage_lookup().
 329                  */
 330                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 331                 if (selem_linked_to_storage(selem))
 332                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 333                                            sdata);
 334                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 335         }
 336
 337         return sdata;
 338 }
 339
 340 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 341                        u64 map_flags)
 342 {
 343         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 344                 /* elem already exists */
 345                 return -EEXIST;
 346
 347         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 348                 /* elem doesn't exist, cannot update it */
 349                 return -ENOENT;
 350
 351         return 0;
 352 }
 353
 354 int bpf_local_storage_alloc(void *owner,
 355                             struct bpf_local_storage_map *smap,
 356                             struct bpf_local_storage_elem *first_selem,
 357                             gfp_t gfp_flags)
 358 {
 359         struct bpf_local_storage *prev_storage, *storage;
 360         struct bpf_local_storage **owner_storage_ptr;
 361         int err;
 362
 363         err = mem_charge(smap, owner, sizeof(*storage));
 364         if (err)
 365                 return err;
 366
 367         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 368                                   gfp_flags | __GFP_NOWARN);
 369         if (!storage) {
 370                 err = -ENOMEM;
 371                 goto uncharge;
 372         }
 373
 374         RCU_INIT_POINTER(storage->smap, smap);
 375         INIT_HLIST_HEAD(&storage->list);
 376         raw_spin_lock_init(&storage->lock);
 377         storage->owner = owner;
 378
 379         bpf_selem_link_storage_nolock(storage, first_selem);
 380         bpf_selem_link_map(smap, first_selem);
 381
 382         owner_storage_ptr =
 383                 (struct bpf_local_storage **)owner_storage(smap, owner);
 384         /* Publish storage to the owner.
 385          * Instead of using any lock of the kernel object (i.e. owner),
 386          * cmpxchg will work with any kernel object regardless what
 387          * the running context is, bh, irq...etc.
 388          *
 389          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 390          * is protected by the storage->lock.  Hence, when freeing
 391          * the owner->storage, the storage->lock must be held before
 392          * setting owner->storage ptr to NULL.
 393          */
 394         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 395         if (unlikely(prev_storage)) {
 396                 bpf_selem_unlink_map(first_selem);
 397                 err = -EAGAIN;
 398                 goto uncharge;
 399
 400                 /* Note that even first_selem was linked to smap's
 401                  * bucket->list, first_selem can be freed immediately
 402                  * (instead of kfree_rcu) because
 403                  * bpf_local_storage_map_free() does a
 404                  * synchronize_rcu_mult (waiting for both sleepable and
 405                  * normal programs) before walking the bucket->list.
 406                  * Hence, no one is accessing selem from the
 407                  * bucket->list under rcu_read_lock().
 408                  */
 409         }
 410
 411         return 0;
 412
 413 uncharge:
 414         kfree(storage);
 415         mem_uncharge(smap, owner, sizeof(*storage));
 416         return err;
 417 }
 418
 419 /* sk cannot be going away because it is linking new elem
 420  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 421  * Otherwise, it will become a leak (and other memory issues
 422  * during map destruction).
 423  */
 424 struct bpf_local_storage_data *
 425 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 426                          void *value, u64 map_flags, gfp_t gfp_flags)
 427 {
 428         struct bpf_local_storage_data *old_sdata = NULL;
 429         struct bpf_local_storage_elem *selem = NULL;
 430         struct bpf_local_storage *local_storage;
 431         unsigned long flags;
 432         int err;
 433
 434         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 435         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 436             /* BPF_F_LOCK can only be used in a value with spin_lock */
 437             unlikely((map_flags & BPF_F_LOCK) &&
 438                      !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
 439                 return ERR_PTR(-EINVAL);
 440
 441         if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
 442                 return ERR_PTR(-EINVAL);
 443
 444         local_storage = rcu_dereference_check(*owner_storage(smap, owner),
 445                                               bpf_rcu_lock_held());
 446         if (!local_storage || hlist_empty(&local_storage->list)) {
 447                 /* Very first elem for the owner */
 448                 err = check_flags(NULL, map_flags);
 449                 if (err)
 450                         return ERR_PTR(err);
 451
 452                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 453                 if (!selem)
 454                         return ERR_PTR(-ENOMEM);
 455
 456                 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
 457                 if (err) {
 458                         kfree(selem);
 459                         mem_uncharge(smap, owner, smap->elem_size);
 460                         return ERR_PTR(err);
 461                 }
 462
 463                 return SDATA(selem);
 464         }
 465
 466         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 467                 /* Hoping to find an old_sdata to do inline update
 468                  * such that it can avoid taking the local_storage->lock
 469                  * and changing the lists.
 470                  */
 471                 old_sdata =
 472                         bpf_local_storage_lookup(local_storage, smap, false);
 473                 err = check_flags(old_sdata, map_flags);
 474                 if (err)
 475                         return ERR_PTR(err);
 476                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 477                         copy_map_value_locked(&smap->map, old_sdata->data,
 478                                               value, false);
 479                         return old_sdata;
 480                 }
 481         }
 482
 483         if (gfp_flags == GFP_KERNEL) {
 484                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 485                 if (!selem)
 486                         return ERR_PTR(-ENOMEM);
 487         }
 488
 489         raw_spin_lock_irqsave(&local_storage->lock, flags);
 490
 491         /* Recheck local_storage->list under local_storage->lock */
 492         if (unlikely(hlist_empty(&local_storage->list))) {
 493                 /* A parallel del is happening and local_storage is going
 494                  * away.  It has just been checked before, so very
 495                  * unlikely.  Return instead of retry to keep things
 496                  * simple.
 497                  */
 498                 err = -EAGAIN;
 499                 goto unlock_err;
 500         }
 501
 502         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 503         err = check_flags(old_sdata, map_flags);
 504         if (err)
 505                 goto unlock_err;
 506
 507         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 508                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 509                                       false);
 510                 selem = SELEM(old_sdata);
 511                 goto unlock;
 512         }
 513
 514         if (gfp_flags != GFP_KERNEL) {
 515                 /* local_storage->lock is held.  Hence, we are sure
 516                  * we can unlink and uncharge the old_sdata successfully
 517                  * later.  Hence, instead of charging the new selem now
 518                  * and then uncharge the old selem later (which may cause
 519                  * a potential but unnecessary charge failure),  avoid taking
 520                  * a charge at all here (the "!old_sdata" check) and the
 521                  * old_sdata will not be uncharged later during
 522                  * bpf_selem_unlink_storage_nolock().
 523                  */
 524                 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
 525                 if (!selem) {
 526                         err = -ENOMEM;
 527                         goto unlock_err;
 528                 }
 529         }
 530
 531         /* First, link the new selem to the map */
 532         bpf_selem_link_map(smap, selem);
 533
 534         /* Second, link (and publish) the new selem to local_storage */
 535         bpf_selem_link_storage_nolock(local_storage, selem);
 536
 537         /* Third, remove old selem, SELEM(old_sdata) */
 538         if (old_sdata) {
 539                 bpf_selem_unlink_map(SELEM(old_sdata));
 540                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 541                                                 false, true);
 542         }
 543
 544 unlock:
 545         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 546         return SDATA(selem);
 547
 548 unlock_err:
 549         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 550         if (selem) {
 551                 mem_uncharge(smap, owner, smap->elem_size);
 552                 kfree(selem);
 553         }
 554         return ERR_PTR(err);
 555 }
 556
 557 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 558 {
 559         u64 min_usage = U64_MAX;
 560         u16 i, res = 0;
 561
 562         spin_lock(&cache->idx_lock);
 563
 564         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 565                 if (cache->idx_usage_counts[i] < min_usage) {
 566                         min_usage = cache->idx_usage_counts[i];
 567                         res = i;
 568
 569                         /* Found a free cache_idx */
 570                         if (!min_usage)
 571                                 break;
 572                 }
 573         }
 574         cache->idx_usage_counts[res]++;
 575
 576         spin_unlock(&cache->idx_lock);
 577
 578         return res;
 579 }
 580
 581 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 582                                              u16 idx)
 583 {
 584         spin_lock(&cache->idx_lock);
 585         cache->idx_usage_counts[idx]--;
 586         spin_unlock(&cache->idx_lock);
 587 }
 588
 589 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 590 {
 591         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 592             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 593             attr->max_entries ||
 594             attr->key_size != sizeof(int) || !attr->value_size ||
 595             /* Enforce BTF for userspace sk dumping */
 596             !attr->btf_key_type_id || !attr->btf_value_type_id)
 597                 return -EINVAL;
 598
 599         if (!bpf_capable())
 600                 return -EPERM;
 601
 602         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 603                 return -E2BIG;
 604
 605         return 0;
 606 }
 607
 608 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 609                                     const struct btf *btf,
 610                                     const struct btf_type *key_type,
 611                                     const struct btf_type *value_type)
 612 {
 613         u32 int_data;
 614
 615         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 616                 return -EINVAL;
 617
 618         int_data = *(u32 *)(key_type + 1);
 619         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 620                 return -EINVAL;
 621
 622         return 0;
 623 }
 624
 625 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
 626 {
 627         struct bpf_local_storage_elem *selem;
 628         bool free_storage = false;
 629         struct hlist_node *n;
 630         unsigned long flags;
 631
 632         /* Neither the bpf_prog nor the bpf_map's syscall
 633          * could be modifying the local_storage->list now.
 634          * Thus, no elem can be added to or deleted from the
 635          * local_storage->list by the bpf_prog or by the bpf_map's syscall.
 636          *
 637          * It is racing with bpf_local_storage_map_free() alone
 638          * when unlinking elem from the local_storage->list and
 639          * the map's bucket->list.
 640          */
 641         raw_spin_lock_irqsave(&local_storage->lock, flags);
 642         hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
 643                 /* Always unlink from map before unlinking from
 644                  * local_storage.
 645                  */
 646                 bpf_selem_unlink_map(selem);
 647                 /* If local_storage list has only one element, the
 648                  * bpf_selem_unlink_storage_nolock() will return true.
 649                  * Otherwise, it will return false. The current loop iteration
 650                  * intends to remove all local storage. So the last iteration
 651                  * of the loop will set the free_cgroup_storage to true.
 652                  */
 653                 free_storage = bpf_selem_unlink_storage_nolock(
 654                         local_storage, selem, false, false);
 655         }
 656         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 657
 658         if (free_storage)
 659                 kfree_rcu(local_storage, rcu);
 660 }
 661
 662 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
 663 {
 664         struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
 665         u64 usage = sizeof(*smap);
 666
 667         /* The dynamically callocated selems are not counted currently. */
 668         usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
 669         return usage;
 670 }
 671
 672 struct bpf_map *
 673 bpf_local_storage_map_alloc(union bpf_attr *attr,
 674                             struct bpf_local_storage_cache *cache)
 675 {
 676         struct bpf_local_storage_map *smap;
 677         unsigned int i;
 678         u32 nbuckets;
 679
 680         smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
 681         if (!smap)
 682                 return ERR_PTR(-ENOMEM);
 683         bpf_map_init_from_attr(&smap->map, attr);
 684
 685         nbuckets = roundup_pow_of_two(num_possible_cpus());
 686         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 687         nbuckets = max_t(u32, 2, nbuckets);
 688         smap->bucket_log = ilog2(nbuckets);
 689
 690         smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
 691                                          nbuckets, GFP_USER | __GFP_NOWARN);
 692         if (!smap->buckets) {
 693                 bpf_map_area_free(smap);
 694                 return ERR_PTR(-ENOMEM);
 695         }
 696
 697         for (i = 0; i < nbuckets; i++) {
 698                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 699                 raw_spin_lock_init(&smap->buckets[i].lock);
 700         }
 701
 702         smap->elem_size = offsetof(struct bpf_local_storage_elem,
 703                                    sdata.data[attr->value_size]);
 704
 705         smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
 706         return &smap->map;
 707 }
 708
 709 void bpf_local_storage_map_free(struct bpf_map *map,
 710                                 struct bpf_local_storage_cache *cache,
 711                                 int __percpu *busy_counter)
 712 {
 713         struct bpf_local_storage_map_bucket *b;
 714         struct bpf_local_storage_elem *selem;
 715         struct bpf_local_storage_map *smap;
 716         unsigned int i;
 717
 718         smap = (struct bpf_local_storage_map *)map;
 719         bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
 720
 721         /* Note that this map might be concurrently cloned from
 722          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 723          * RCU read section to finish before proceeding. New RCU
 724          * read sections should be prevented via bpf_map_inc_not_zero.
 725          */
 726         synchronize_rcu();
 727
 728         /* bpf prog and the userspace can no longer access this map
 729          * now.  No new selem (of this map) can be added
 730          * to the owner->storage or to the map bucket's list.
 731          *
 732          * The elem of this map can be cleaned up here
 733          * or when the storage is freed e.g.
 734          * by bpf_sk_storage_free() during __sk_destruct().
 735          */
 736         for (i = 0; i < (1U << smap->bucket_log); i++) {
 737                 b = &smap->buckets[i];
 738
 739                 rcu_read_lock();
 740                 /* No one is adding to b->list now */
 741                 while ((selem = hlist_entry_safe(
 742                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 743                                 struct bpf_local_storage_elem, map_node))) {
 744                         if (busy_counter) {
 745                                 migrate_disable();
 746                                 this_cpu_inc(*busy_counter);
 747                         }
 748                         bpf_selem_unlink(selem, false);
 749                         if (busy_counter) {
 750                                 this_cpu_dec(*busy_counter);
 751                                 migrate_enable();
 752                         }
 753                         cond_resched_rcu();
 754                 }
 755                 rcu_read_unlock();
 756         }
 757
 758         /* While freeing the storage we may still need to access the map.
 759          *
 760          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 761          * which then made the above while((selem = ...)) loop
 762          * exit immediately.
 763          *
 764          * However, while freeing the storage one still needs to access the
 765          * smap->elem_size to do the uncharging in
 766          * bpf_selem_unlink_storage_nolock().
 767          *
 768          * Hence, wait another rcu grace period for the storage to be freed.
 769          */
 770         synchronize_rcu();
 771
 772         /* Only delay freeing of smap, buckets are not needed anymore */
 773         kvfree(smap->buckets);
 774
 775         /* When local storage has special fields, callbacks for
 776          * bpf_selem_free_fields_rcu and bpf_selem_free_fields_trace_rcu will
 777          * keep using the map BTF record, we need to execute an RCU barrier to
 778          * wait for them as the record will be freed right after our map_free
 779          * callback.
 780          */
 781         if (!IS_ERR_OR_NULL(smap->map.record)) {
 782                 rcu_barrier_tasks_trace();
 783                 /* We cannot skip rcu_barrier() when rcu_trace_implies_rcu_gp()
 784                  * is true, because while call_rcu invocation is skipped in that
 785                  * case in bpf_selem_free_fields_trace_rcu (and all local
 786                  * storage maps pass use_trace_rcu = true), there can be
 787                  * call_rcu callbacks based on use_trace_rcu = false in the
 788                  * while ((selem = ...)) loop above or when owner's free path
 789                  * calls bpf_local_storage_unlink_nolock.
 790                  */
 791                 rcu_barrier();
 792         }
 793         bpf_map_area_free(smap);
 794 }