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         if (smap->bpf_ma) {
  84                 migrate_disable();
  85                 selem = bpf_mem_cache_alloc_flags(&smap->selem_ma, gfp_flags);
  86                 migrate_enable();
  87                 if (selem)
  88                         /* Keep the original bpf_map_kzalloc behavior
  89                          * before started using the bpf_mem_cache_alloc.
  90                          *
  91                          * No need to use zero_map_value. The bpf_selem_free()
  92                          * only does bpf_mem_cache_free when there is
  93                          * no other bpf prog is using the selem.
  94                          */
  95                         memset(SDATA(selem)->data, 0, smap->map.value_size);
  96         } else {
  97                 selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
  98                                         gfp_flags | __GFP_NOWARN);
  99         }
 100
 101         if (selem) {
 102                 if (value)
 103                         copy_map_value(&smap->map, SDATA(selem)->data, value);
 104                 /* No need to call check_and_init_map_value as memory is zero init */
 105                 return selem;
 106         }
 107
 108         if (charge_mem)
 109                 mem_uncharge(smap, owner, smap->elem_size);
 110
 111         return NULL;
 112 }
 113
 114 static void bpf_local_storage_free_rcu(struct rcu_head *rcu)
 115 {
 116         struct bpf_local_storage *local_storage;
 117
 118         local_storage = container_of(rcu, struct bpf_local_storage, rcu);
 119         kfree(local_storage);
 120 }
 121
 122 static void bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
 123 {
 124         /* If RCU Tasks Trace grace period implies RCU grace period, do
 125          * kfree(), else do kfree_rcu().
 126          */
 127         if (rcu_trace_implies_rcu_gp())
 128                 bpf_local_storage_free_rcu(rcu);
 129         else
 130                 call_rcu(rcu, bpf_local_storage_free_rcu);
 131 }
 132
 133 static void bpf_local_storage_free(struct bpf_local_storage *local_storage,
 134                                    bool reuse_now)
 135 {
 136         if (!reuse_now)
 137                 call_rcu_tasks_trace(&local_storage->rcu,
 138                                      bpf_local_storage_free_trace_rcu);
 139         else
 140                 call_rcu(&local_storage->rcu, bpf_local_storage_free_rcu);
 141 }
 142
 143 /* rcu tasks trace callback for bpf_ma == false */
 144 static void __bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 145 {
 146         struct bpf_local_storage_elem *selem;
 147
 148         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 149         if (rcu_trace_implies_rcu_gp())
 150                 kfree(selem);
 151         else
 152                 kfree_rcu(selem, rcu);
 153 }
 154
 155 /* Handle bpf_ma == false */
 156 static void __bpf_selem_free(struct bpf_local_storage_elem *selem,
 157                              bool vanilla_rcu)
 158 {
 159         if (vanilla_rcu)
 160                 kfree_rcu(selem, rcu);
 161         else
 162                 call_rcu_tasks_trace(&selem->rcu, __bpf_selem_free_trace_rcu);
 163 }
 164
 165 static void bpf_selem_free_rcu(struct rcu_head *rcu)
 166 {
 167         struct bpf_local_storage_elem *selem;
 168
 169         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 170         bpf_mem_cache_raw_free(selem);
 171 }
 172
 173 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 174 {
 175         if (rcu_trace_implies_rcu_gp())
 176                 bpf_selem_free_rcu(rcu);
 177         else
 178                 call_rcu(rcu, bpf_selem_free_rcu);
 179 }
 180
 181 void bpf_selem_free(struct bpf_local_storage_elem *selem,
 182                     struct bpf_local_storage_map *smap,
 183                     bool reuse_now)
 184 {
 185         bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
 186
 187         if (!smap->bpf_ma) {
 188                 __bpf_selem_free(selem, reuse_now);
 189                 return;
 190         }
 191
 192         if (!reuse_now) {
 193                 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
 194         } else {
 195                 /* Instead of using the vanilla call_rcu(),
 196                  * bpf_mem_cache_free will be able to reuse selem
 197                  * immediately.
 198                  */
 199                 migrate_disable();
 200                 bpf_mem_cache_free(&smap->selem_ma, selem);
 201                 migrate_enable();
 202         }
 203 }
 204
 205 /* local_storage->lock must be held and selem->local_storage == local_storage.
 206  * The caller must ensure selem->smap is still valid to be
 207  * dereferenced for its smap->elem_size and smap->cache_idx.
 208  */
 209 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
 210                                             struct bpf_local_storage_elem *selem,
 211                                             bool uncharge_mem, bool reuse_now)
 212 {
 213         struct bpf_local_storage_map *smap;
 214         bool free_local_storage;
 215         void *owner;
 216
 217         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 218         owner = local_storage->owner;
 219
 220         /* All uncharging on the owner must be done first.
 221          * The owner may be freed once the last selem is unlinked
 222          * from local_storage.
 223          */
 224         if (uncharge_mem)
 225                 mem_uncharge(smap, owner, smap->elem_size);
 226
 227         free_local_storage = hlist_is_singular_node(&selem->snode,
 228                                                     &local_storage->list);
 229         if (free_local_storage) {
 230                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 231                 local_storage->owner = NULL;
 232
 233                 /* After this RCU_INIT, owner may be freed and cannot be used */
 234                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 235
 236                 /* local_storage is not freed now.  local_storage->lock is
 237                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 238                  * will be done by the caller.
 239                  *
 240                  * Although the unlock will be done under
 241                  * rcu_read_lock(),  it is more intuitive to
 242                  * read if the freeing of the storage is done
 243                  * after the raw_spin_unlock_bh(&local_storage->lock).
 244                  *
 245                  * Hence, a "bool free_local_storage" is returned
 246                  * to the caller which then calls then frees the storage after
 247                  * all the RCU grace periods have expired.
 248                  */
 249         }
 250         hlist_del_init_rcu(&selem->snode);
 251         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 252             SDATA(selem))
 253                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 254
 255         bpf_selem_free(selem, smap, reuse_now);
 256
 257         if (rcu_access_pointer(local_storage->smap) == smap)
 258                 RCU_INIT_POINTER(local_storage->smap, NULL);
 259
 260         return free_local_storage;
 261 }
 262
 263 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
 264                                      bool reuse_now)
 265 {
 266         struct bpf_local_storage *local_storage;
 267         bool free_local_storage = false;
 268         unsigned long flags;
 269
 270         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 271                 /* selem has already been unlinked from sk */
 272                 return;
 273
 274         local_storage = rcu_dereference_check(selem->local_storage,
 275                                               bpf_rcu_lock_held());
 276         raw_spin_lock_irqsave(&local_storage->lock, flags);
 277         if (likely(selem_linked_to_storage(selem)))
 278                 free_local_storage = bpf_selem_unlink_storage_nolock(
 279                         local_storage, selem, true, reuse_now);
 280         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 281
 282         if (free_local_storage)
 283                 bpf_local_storage_free(local_storage, reuse_now);
 284 }
 285
 286 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 287                                    struct bpf_local_storage_elem *selem)
 288 {
 289         RCU_INIT_POINTER(selem->local_storage, local_storage);
 290         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 291 }
 292
 293 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 294 {
 295         struct bpf_local_storage_map *smap;
 296         struct bpf_local_storage_map_bucket *b;
 297         unsigned long flags;
 298
 299         if (unlikely(!selem_linked_to_map_lockless(selem)))
 300                 /* selem has already be unlinked from smap */
 301                 return;
 302
 303         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 304         b = select_bucket(smap, selem);
 305         raw_spin_lock_irqsave(&b->lock, flags);
 306         if (likely(selem_linked_to_map(selem)))
 307                 hlist_del_init_rcu(&selem->map_node);
 308         raw_spin_unlock_irqrestore(&b->lock, flags);
 309 }
 310
 311 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 312                         struct bpf_local_storage_elem *selem)
 313 {
 314         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 315         unsigned long flags;
 316
 317         raw_spin_lock_irqsave(&b->lock, flags);
 318         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 319         hlist_add_head_rcu(&selem->map_node, &b->list);
 320         raw_spin_unlock_irqrestore(&b->lock, flags);
 321 }
 322
 323 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
 324 {
 325         /* Always unlink from map before unlinking from local_storage
 326          * because selem will be freed after successfully unlinked from
 327          * the local_storage.
 328          */
 329         bpf_selem_unlink_map(selem);
 330         bpf_selem_unlink_storage(selem, reuse_now);
 331 }
 332
 333 /* If cacheit_lockit is false, this lookup function is lockless */
 334 struct bpf_local_storage_data *
 335 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 336                          struct bpf_local_storage_map *smap,
 337                          bool cacheit_lockit)
 338 {
 339         struct bpf_local_storage_data *sdata;
 340         struct bpf_local_storage_elem *selem;
 341
 342         /* Fast path (cache hit) */
 343         sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
 344                                       bpf_rcu_lock_held());
 345         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 346                 return sdata;
 347
 348         /* Slow path (cache miss) */
 349         hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
 350                                   rcu_read_lock_trace_held())
 351                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 352                         break;
 353
 354         if (!selem)
 355                 return NULL;
 356
 357         sdata = SDATA(selem);
 358         if (cacheit_lockit) {
 359                 unsigned long flags;
 360
 361                 /* spinlock is needed to avoid racing with the
 362                  * parallel delete.  Otherwise, publishing an already
 363                  * deleted sdata to the cache will become a use-after-free
 364                  * problem in the next bpf_local_storage_lookup().
 365                  */
 366                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 367                 if (selem_linked_to_storage(selem))
 368                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 369                                            sdata);
 370                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 371         }
 372
 373         return sdata;
 374 }
 375
 376 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 377                        u64 map_flags)
 378 {
 379         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 380                 /* elem already exists */
 381                 return -EEXIST;
 382
 383         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 384                 /* elem doesn't exist, cannot update it */
 385                 return -ENOENT;
 386
 387         return 0;
 388 }
 389
 390 int bpf_local_storage_alloc(void *owner,
 391                             struct bpf_local_storage_map *smap,
 392                             struct bpf_local_storage_elem *first_selem,
 393                             gfp_t gfp_flags)
 394 {
 395         struct bpf_local_storage *prev_storage, *storage;
 396         struct bpf_local_storage **owner_storage_ptr;
 397         int err;
 398
 399         err = mem_charge(smap, owner, sizeof(*storage));
 400         if (err)
 401                 return err;
 402
 403         storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 404                                   gfp_flags | __GFP_NOWARN);
 405         if (!storage) {
 406                 err = -ENOMEM;
 407                 goto uncharge;
 408         }
 409
 410         RCU_INIT_POINTER(storage->smap, smap);
 411         INIT_HLIST_HEAD(&storage->list);
 412         raw_spin_lock_init(&storage->lock);
 413         storage->owner = owner;
 414
 415         bpf_selem_link_storage_nolock(storage, first_selem);
 416         bpf_selem_link_map(smap, first_selem);
 417
 418         owner_storage_ptr =
 419                 (struct bpf_local_storage **)owner_storage(smap, owner);
 420         /* Publish storage to the owner.
 421          * Instead of using any lock of the kernel object (i.e. owner),
 422          * cmpxchg will work with any kernel object regardless what
 423          * the running context is, bh, irq...etc.
 424          *
 425          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 426          * is protected by the storage->lock.  Hence, when freeing
 427          * the owner->storage, the storage->lock must be held before
 428          * setting owner->storage ptr to NULL.
 429          */
 430         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 431         if (unlikely(prev_storage)) {
 432                 bpf_selem_unlink_map(first_selem);
 433                 err = -EAGAIN;
 434                 goto uncharge;
 435
 436                 /* Note that even first_selem was linked to smap's
 437                  * bucket->list, first_selem can be freed immediately
 438                  * (instead of kfree_rcu) because
 439                  * bpf_local_storage_map_free() does a
 440                  * synchronize_rcu_mult (waiting for both sleepable and
 441                  * normal programs) before walking the bucket->list.
 442                  * Hence, no one is accessing selem from the
 443                  * bucket->list under rcu_read_lock().
 444                  */
 445         }
 446
 447         return 0;
 448
 449 uncharge:
 450         bpf_local_storage_free(storage, true);
 451         mem_uncharge(smap, owner, sizeof(*storage));
 452         return err;
 453 }
 454
 455 /* sk cannot be going away because it is linking new elem
 456  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 457  * Otherwise, it will become a leak (and other memory issues
 458  * during map destruction).
 459  */
 460 struct bpf_local_storage_data *
 461 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 462                          void *value, u64 map_flags, gfp_t gfp_flags)
 463 {
 464         struct bpf_local_storage_data *old_sdata = NULL;
 465         struct bpf_local_storage_elem *selem = NULL;
 466         struct bpf_local_storage *local_storage;
 467         unsigned long flags;
 468         int err;
 469
 470         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 471         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 472             /* BPF_F_LOCK can only be used in a value with spin_lock */
 473             unlikely((map_flags & BPF_F_LOCK) &&
 474                      !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
 475                 return ERR_PTR(-EINVAL);
 476
 477         if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
 478                 return ERR_PTR(-EINVAL);
 479
 480         local_storage = rcu_dereference_check(*owner_storage(smap, owner),
 481                                               bpf_rcu_lock_held());
 482         if (!local_storage || hlist_empty(&local_storage->list)) {
 483                 /* Very first elem for the owner */
 484                 err = check_flags(NULL, map_flags);
 485                 if (err)
 486                         return ERR_PTR(err);
 487
 488                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 489                 if (!selem)
 490                         return ERR_PTR(-ENOMEM);
 491
 492                 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
 493                 if (err) {
 494                         bpf_selem_free(selem, smap, true);
 495                         mem_uncharge(smap, owner, smap->elem_size);
 496                         return ERR_PTR(err);
 497                 }
 498
 499                 return SDATA(selem);
 500         }
 501
 502         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 503                 /* Hoping to find an old_sdata to do inline update
 504                  * such that it can avoid taking the local_storage->lock
 505                  * and changing the lists.
 506                  */
 507                 old_sdata =
 508                         bpf_local_storage_lookup(local_storage, smap, false);
 509                 err = check_flags(old_sdata, map_flags);
 510                 if (err)
 511                         return ERR_PTR(err);
 512                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 513                         copy_map_value_locked(&smap->map, old_sdata->data,
 514                                               value, false);
 515                         return old_sdata;
 516                 }
 517         }
 518
 519         if (gfp_flags == GFP_KERNEL) {
 520                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 521                 if (!selem)
 522                         return ERR_PTR(-ENOMEM);
 523         }
 524
 525         raw_spin_lock_irqsave(&local_storage->lock, flags);
 526
 527         /* Recheck local_storage->list under local_storage->lock */
 528         if (unlikely(hlist_empty(&local_storage->list))) {
 529                 /* A parallel del is happening and local_storage is going
 530                  * away.  It has just been checked before, so very
 531                  * unlikely.  Return instead of retry to keep things
 532                  * simple.
 533                  */
 534                 err = -EAGAIN;
 535                 goto unlock_err;
 536         }
 537
 538         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 539         err = check_flags(old_sdata, map_flags);
 540         if (err)
 541                 goto unlock_err;
 542
 543         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 544                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 545                                       false);
 546                 selem = SELEM(old_sdata);
 547                 goto unlock;
 548         }
 549
 550         if (gfp_flags != GFP_KERNEL) {
 551                 /* local_storage->lock is held.  Hence, we are sure
 552                  * we can unlink and uncharge the old_sdata successfully
 553                  * later.  Hence, instead of charging the new selem now
 554                  * and then uncharge the old selem later (which may cause
 555                  * a potential but unnecessary charge failure),  avoid taking
 556                  * a charge at all here (the "!old_sdata" check) and the
 557                  * old_sdata will not be uncharged later during
 558                  * bpf_selem_unlink_storage_nolock().
 559                  */
 560                 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
 561                 if (!selem) {
 562                         err = -ENOMEM;
 563                         goto unlock_err;
 564                 }
 565         }
 566
 567         /* First, link the new selem to the map */
 568         bpf_selem_link_map(smap, selem);
 569
 570         /* Second, link (and publish) the new selem to local_storage */
 571         bpf_selem_link_storage_nolock(local_storage, selem);
 572
 573         /* Third, remove old selem, SELEM(old_sdata) */
 574         if (old_sdata) {
 575                 bpf_selem_unlink_map(SELEM(old_sdata));
 576                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 577                                                 false, false);
 578         }
 579
 580 unlock:
 581         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 582         return SDATA(selem);
 583
 584 unlock_err:
 585         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 586         if (selem) {
 587                 mem_uncharge(smap, owner, smap->elem_size);
 588                 bpf_selem_free(selem, smap, true);
 589         }
 590         return ERR_PTR(err);
 591 }
 592
 593 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 594 {
 595         u64 min_usage = U64_MAX;
 596         u16 i, res = 0;
 597
 598         spin_lock(&cache->idx_lock);
 599
 600         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 601                 if (cache->idx_usage_counts[i] < min_usage) {
 602                         min_usage = cache->idx_usage_counts[i];
 603                         res = i;
 604
 605                         /* Found a free cache_idx */
 606                         if (!min_usage)
 607                                 break;
 608                 }
 609         }
 610         cache->idx_usage_counts[res]++;
 611
 612         spin_unlock(&cache->idx_lock);
 613
 614         return res;
 615 }
 616
 617 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 618                                              u16 idx)
 619 {
 620         spin_lock(&cache->idx_lock);
 621         cache->idx_usage_counts[idx]--;
 622         spin_unlock(&cache->idx_lock);
 623 }
 624
 625 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 626 {
 627         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 628             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 629             attr->max_entries ||
 630             attr->key_size != sizeof(int) || !attr->value_size ||
 631             /* Enforce BTF for userspace sk dumping */
 632             !attr->btf_key_type_id || !attr->btf_value_type_id)
 633                 return -EINVAL;
 634
 635         if (!bpf_capable())
 636                 return -EPERM;
 637
 638         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 639                 return -E2BIG;
 640
 641         return 0;
 642 }
 643
 644 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 645                                     const struct btf *btf,
 646                                     const struct btf_type *key_type,
 647                                     const struct btf_type *value_type)
 648 {
 649         u32 int_data;
 650
 651         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 652                 return -EINVAL;
 653
 654         int_data = *(u32 *)(key_type + 1);
 655         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 656                 return -EINVAL;
 657
 658         return 0;
 659 }
 660
 661 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
 662 {
 663         struct bpf_local_storage_elem *selem;
 664         bool free_storage = false;
 665         struct hlist_node *n;
 666         unsigned long flags;
 667
 668         /* Neither the bpf_prog nor the bpf_map's syscall
 669          * could be modifying the local_storage->list now.
 670          * Thus, no elem can be added to or deleted from the
 671          * local_storage->list by the bpf_prog or by the bpf_map's syscall.
 672          *
 673          * It is racing with bpf_local_storage_map_free() alone
 674          * when unlinking elem from the local_storage->list and
 675          * the map's bucket->list.
 676          */
 677         raw_spin_lock_irqsave(&local_storage->lock, flags);
 678         hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
 679                 /* Always unlink from map before unlinking from
 680                  * local_storage.
 681                  */
 682                 bpf_selem_unlink_map(selem);
 683                 /* If local_storage list has only one element, the
 684                  * bpf_selem_unlink_storage_nolock() will return true.
 685                  * Otherwise, it will return false. The current loop iteration
 686                  * intends to remove all local storage. So the last iteration
 687                  * of the loop will set the free_cgroup_storage to true.
 688                  */
 689                 free_storage = bpf_selem_unlink_storage_nolock(
 690                         local_storage, selem, false, true);
 691         }
 692         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 693
 694         if (free_storage)
 695                 bpf_local_storage_free(local_storage, true);
 696 }
 697
 698 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
 699 {
 700         struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
 701         u64 usage = sizeof(*smap);
 702
 703         /* The dynamically callocated selems are not counted currently. */
 704         usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
 705         return usage;
 706 }
 707
 708 /* When bpf_ma == true, the bpf_mem_alloc is used to allocate and free memory.
 709  * A deadlock free allocator is useful for storage that the bpf prog can easily
 710  * get a hold of the owner PTR_TO_BTF_ID in any context. eg. bpf_get_current_task_btf.
 711  * The task and cgroup storage fall into this case. The bpf_mem_alloc reuses
 712  * memory immediately. To be reuse-immediate safe, the owner destruction
 713  * code path needs to go through a rcu grace period before calling
 714  * bpf_local_storage_destroy().
 715  *
 716  * When bpf_ma == false, the kmalloc and kfree are used.
 717  */
 718 struct bpf_map *
 719 bpf_local_storage_map_alloc(union bpf_attr *attr,
 720                             struct bpf_local_storage_cache *cache,
 721                             bool bpf_ma)
 722 {
 723         struct bpf_local_storage_map *smap;
 724         unsigned int i;
 725         u32 nbuckets;
 726         int err;
 727
 728         smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
 729         if (!smap)
 730                 return ERR_PTR(-ENOMEM);
 731         bpf_map_init_from_attr(&smap->map, attr);
 732
 733         nbuckets = roundup_pow_of_two(num_possible_cpus());
 734         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 735         nbuckets = max_t(u32, 2, nbuckets);
 736         smap->bucket_log = ilog2(nbuckets);
 737
 738         smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
 739                                          nbuckets, GFP_USER | __GFP_NOWARN);
 740         if (!smap->buckets) {
 741                 err = -ENOMEM;
 742                 goto free_smap;
 743         }
 744
 745         for (i = 0; i < nbuckets; i++) {
 746                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 747                 raw_spin_lock_init(&smap->buckets[i].lock);
 748         }
 749
 750         smap->elem_size = offsetof(struct bpf_local_storage_elem,
 751                                    sdata.data[attr->value_size]);
 752
 753         smap->bpf_ma = bpf_ma;
 754         if (bpf_ma) {
 755                 err = bpf_mem_alloc_init(&smap->selem_ma, smap->elem_size, false);
 756                 if (err)
 757                         goto free_smap;
 758         }
 759
 760         smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
 761         return &smap->map;
 762
 763 free_smap:
 764         kvfree(smap->buckets);
 765         bpf_map_area_free(smap);
 766         return ERR_PTR(err);
 767 }
 768
 769 void bpf_local_storage_map_free(struct bpf_map *map,
 770                                 struct bpf_local_storage_cache *cache,
 771                                 int __percpu *busy_counter)
 772 {
 773         struct bpf_local_storage_map_bucket *b;
 774         struct bpf_local_storage_elem *selem;
 775         struct bpf_local_storage_map *smap;
 776         unsigned int i;
 777
 778         smap = (struct bpf_local_storage_map *)map;
 779         bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
 780
 781         /* Note that this map might be concurrently cloned from
 782          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 783          * RCU read section to finish before proceeding. New RCU
 784          * read sections should be prevented via bpf_map_inc_not_zero.
 785          */
 786         synchronize_rcu();
 787
 788         /* bpf prog and the userspace can no longer access this map
 789          * now.  No new selem (of this map) can be added
 790          * to the owner->storage or to the map bucket's list.
 791          *
 792          * The elem of this map can be cleaned up here
 793          * or when the storage is freed e.g.
 794          * by bpf_sk_storage_free() during __sk_destruct().
 795          */
 796         for (i = 0; i < (1U << smap->bucket_log); i++) {
 797                 b = &smap->buckets[i];
 798
 799                 rcu_read_lock();
 800                 /* No one is adding to b->list now */
 801                 while ((selem = hlist_entry_safe(
 802                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 803                                 struct bpf_local_storage_elem, map_node))) {
 804                         if (busy_counter) {
 805                                 migrate_disable();
 806                                 this_cpu_inc(*busy_counter);
 807                         }
 808                         bpf_selem_unlink(selem, true);
 809                         if (busy_counter) {
 810                                 this_cpu_dec(*busy_counter);
 811                                 migrate_enable();
 812                         }
 813                         cond_resched_rcu();
 814                 }
 815                 rcu_read_unlock();
 816         }
 817
 818         /* While freeing the storage we may still need to access the map.
 819          *
 820          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 821          * which then made the above while((selem = ...)) loop
 822          * exit immediately.
 823          *
 824          * However, while freeing the storage one still needs to access the
 825          * smap->elem_size to do the uncharging in
 826          * bpf_selem_unlink_storage_nolock().
 827          *
 828          * Hence, wait another rcu grace period for the storage to be freed.
 829          */
 830         synchronize_rcu();
 831
 832         if (smap->bpf_ma)
 833                 bpf_mem_alloc_destroy(&smap->selem_ma);
 834         kvfree(smap->buckets);
 835         bpf_map_area_free(smap);
 836 }