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 /* rcu tasks trace callback for bpf_ma == false */
 115 static void __bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
 116 {
 117         struct bpf_local_storage *local_storage;
 118
 119         /* If RCU Tasks Trace grace period implies RCU grace period, do
 120          * kfree(), else do kfree_rcu().
 121          */
 122         local_storage = container_of(rcu, struct bpf_local_storage, rcu);
 123         if (rcu_trace_implies_rcu_gp())
 124                 kfree(local_storage);
 125         else
 126                 kfree_rcu(local_storage, rcu);
 127 }
 128
 129 static void bpf_local_storage_free_rcu(struct rcu_head *rcu)
 130 {
 131         struct bpf_local_storage *local_storage;
 132
 133         local_storage = container_of(rcu, struct bpf_local_storage, rcu);
 134         bpf_mem_cache_raw_free(local_storage);
 135 }
 136
 137 static void bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
 138 {
 139         if (rcu_trace_implies_rcu_gp())
 140                 bpf_local_storage_free_rcu(rcu);
 141         else
 142                 call_rcu(rcu, bpf_local_storage_free_rcu);
 143 }
 144
 145 /* Handle bpf_ma == false */
 146 static void __bpf_local_storage_free(struct bpf_local_storage *local_storage,
 147                                      bool vanilla_rcu)
 148 {
 149         if (vanilla_rcu)
 150                 kfree_rcu(local_storage, rcu);
 151         else
 152                 call_rcu_tasks_trace(&local_storage->rcu,
 153                                      __bpf_local_storage_free_trace_rcu);
 154 }
 155
 156 static void bpf_local_storage_free(struct bpf_local_storage *local_storage,
 157                                    struct bpf_local_storage_map *smap,
 158                                    bool bpf_ma, bool reuse_now)
 159 {
 160         if (!local_storage)
 161                 return;
 162
 163         if (!bpf_ma) {
 164                 __bpf_local_storage_free(local_storage, reuse_now);
 165                 return;
 166         }
 167
 168         if (!reuse_now) {
 169                 call_rcu_tasks_trace(&local_storage->rcu,
 170                                      bpf_local_storage_free_trace_rcu);
 171                 return;
 172         }
 173
 174         if (smap) {
 175                 migrate_disable();
 176                 bpf_mem_cache_free(&smap->storage_ma, local_storage);
 177                 migrate_enable();
 178         } else {
 179                 /* smap could be NULL if the selem that triggered
 180                  * this 'local_storage' creation had been long gone.
 181                  * In this case, directly do call_rcu().
 182                  */
 183                 call_rcu(&local_storage->rcu, bpf_local_storage_free_rcu);
 184         }
 185 }
 186
 187 /* rcu tasks trace callback for bpf_ma == false */
 188 static void __bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 189 {
 190         struct bpf_local_storage_elem *selem;
 191
 192         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 193         if (rcu_trace_implies_rcu_gp())
 194                 kfree(selem);
 195         else
 196                 kfree_rcu(selem, rcu);
 197 }
 198
 199 /* Handle bpf_ma == false */
 200 static void __bpf_selem_free(struct bpf_local_storage_elem *selem,
 201                              bool vanilla_rcu)
 202 {
 203         if (vanilla_rcu)
 204                 kfree_rcu(selem, rcu);
 205         else
 206                 call_rcu_tasks_trace(&selem->rcu, __bpf_selem_free_trace_rcu);
 207 }
 208
 209 static void bpf_selem_free_rcu(struct rcu_head *rcu)
 210 {
 211         struct bpf_local_storage_elem *selem;
 212
 213         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 214         bpf_mem_cache_raw_free(selem);
 215 }
 216
 217 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 218 {
 219         if (rcu_trace_implies_rcu_gp())
 220                 bpf_selem_free_rcu(rcu);
 221         else
 222                 call_rcu(rcu, bpf_selem_free_rcu);
 223 }
 224
 225 void bpf_selem_free(struct bpf_local_storage_elem *selem,
 226                     struct bpf_local_storage_map *smap,
 227                     bool reuse_now)
 228 {
 229         bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
 230
 231         if (!smap->bpf_ma) {
 232                 __bpf_selem_free(selem, reuse_now);
 233                 return;
 234         }
 235
 236         if (!reuse_now) {
 237                 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
 238         } else {
 239                 /* Instead of using the vanilla call_rcu(),
 240                  * bpf_mem_cache_free will be able to reuse selem
 241                  * immediately.
 242                  */
 243                 migrate_disable();
 244                 bpf_mem_cache_free(&smap->selem_ma, selem);
 245                 migrate_enable();
 246         }
 247 }
 248
 249 /* local_storage->lock must be held and selem->local_storage == local_storage.
 250  * The caller must ensure selem->smap is still valid to be
 251  * dereferenced for its smap->elem_size and smap->cache_idx.
 252  */
 253 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
 254                                             struct bpf_local_storage_elem *selem,
 255                                             bool uncharge_mem, bool reuse_now)
 256 {
 257         struct bpf_local_storage_map *smap;
 258         bool free_local_storage;
 259         void *owner;
 260
 261         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 262         owner = local_storage->owner;
 263
 264         /* All uncharging on the owner must be done first.
 265          * The owner may be freed once the last selem is unlinked
 266          * from local_storage.
 267          */
 268         if (uncharge_mem)
 269                 mem_uncharge(smap, owner, smap->elem_size);
 270
 271         free_local_storage = hlist_is_singular_node(&selem->snode,
 272                                                     &local_storage->list);
 273         if (free_local_storage) {
 274                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 275                 local_storage->owner = NULL;
 276
 277                 /* After this RCU_INIT, owner may be freed and cannot be used */
 278                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 279
 280                 /* local_storage is not freed now.  local_storage->lock is
 281                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 282                  * will be done by the caller.
 283                  *
 284                  * Although the unlock will be done under
 285                  * rcu_read_lock(),  it is more intuitive to
 286                  * read if the freeing of the storage is done
 287                  * after the raw_spin_unlock_bh(&local_storage->lock).
 288                  *
 289                  * Hence, a "bool free_local_storage" is returned
 290                  * to the caller which then calls then frees the storage after
 291                  * all the RCU grace periods have expired.
 292                  */
 293         }
 294         hlist_del_init_rcu(&selem->snode);
 295         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 296             SDATA(selem))
 297                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 298
 299         bpf_selem_free(selem, smap, reuse_now);
 300
 301         if (rcu_access_pointer(local_storage->smap) == smap)
 302                 RCU_INIT_POINTER(local_storage->smap, NULL);
 303
 304         return free_local_storage;
 305 }
 306
 307 static bool check_storage_bpf_ma(struct bpf_local_storage *local_storage,
 308                                  struct bpf_local_storage_map *storage_smap,
 309                                  struct bpf_local_storage_elem *selem)
 310 {
 311
 312         struct bpf_local_storage_map *selem_smap;
 313
 314         /* local_storage->smap may be NULL. If it is, get the bpf_ma
 315          * from any selem in the local_storage->list. The bpf_ma of all
 316          * local_storage and selem should have the same value
 317          * for the same map type.
 318          *
 319          * If the local_storage->list is already empty, the caller will not
 320          * care about the bpf_ma value also because the caller is not
 321          * responsibile to free the local_storage.
 322          */
 323
 324         if (storage_smap)
 325                 return storage_smap->bpf_ma;
 326
 327         if (!selem) {
 328                 struct hlist_node *n;
 329
 330                 n = rcu_dereference_check(hlist_first_rcu(&local_storage->list),
 331                                           bpf_rcu_lock_held());
 332                 if (!n)
 333                         return false;
 334
 335                 selem = hlist_entry(n, struct bpf_local_storage_elem, snode);
 336         }
 337         selem_smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 338
 339         return selem_smap->bpf_ma;
 340 }
 341
 342 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
 343                                      bool reuse_now)
 344 {
 345         struct bpf_local_storage_map *storage_smap;
 346         struct bpf_local_storage *local_storage;
 347         bool bpf_ma, free_local_storage = false;
 348         unsigned long flags;
 349
 350         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 351                 /* selem has already been unlinked from sk */
 352                 return;
 353
 354         local_storage = rcu_dereference_check(selem->local_storage,
 355                                               bpf_rcu_lock_held());
 356         storage_smap = rcu_dereference_check(local_storage->smap,
 357                                              bpf_rcu_lock_held());
 358         bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, selem);
 359
 360         raw_spin_lock_irqsave(&local_storage->lock, flags);
 361         if (likely(selem_linked_to_storage(selem)))
 362                 free_local_storage = bpf_selem_unlink_storage_nolock(
 363                         local_storage, selem, true, reuse_now);
 364         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 365
 366         if (free_local_storage)
 367                 bpf_local_storage_free(local_storage, storage_smap, bpf_ma, reuse_now);
 368 }
 369
 370 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 371                                    struct bpf_local_storage_elem *selem)
 372 {
 373         RCU_INIT_POINTER(selem->local_storage, local_storage);
 374         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 375 }
 376
 377 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 378 {
 379         struct bpf_local_storage_map *smap;
 380         struct bpf_local_storage_map_bucket *b;
 381         unsigned long flags;
 382
 383         if (unlikely(!selem_linked_to_map_lockless(selem)))
 384                 /* selem has already be unlinked from smap */
 385                 return;
 386
 387         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 388         b = select_bucket(smap, selem);
 389         raw_spin_lock_irqsave(&b->lock, flags);
 390         if (likely(selem_linked_to_map(selem)))
 391                 hlist_del_init_rcu(&selem->map_node);
 392         raw_spin_unlock_irqrestore(&b->lock, flags);
 393 }
 394
 395 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 396                         struct bpf_local_storage_elem *selem)
 397 {
 398         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 399         unsigned long flags;
 400
 401         raw_spin_lock_irqsave(&b->lock, flags);
 402         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 403         hlist_add_head_rcu(&selem->map_node, &b->list);
 404         raw_spin_unlock_irqrestore(&b->lock, flags);
 405 }
 406
 407 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
 408 {
 409         /* Always unlink from map before unlinking from local_storage
 410          * because selem will be freed after successfully unlinked from
 411          * the local_storage.
 412          */
 413         bpf_selem_unlink_map(selem);
 414         bpf_selem_unlink_storage(selem, reuse_now);
 415 }
 416
 417 /* If cacheit_lockit is false, this lookup function is lockless */
 418 struct bpf_local_storage_data *
 419 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 420                          struct bpf_local_storage_map *smap,
 421                          bool cacheit_lockit)
 422 {
 423         struct bpf_local_storage_data *sdata;
 424         struct bpf_local_storage_elem *selem;
 425
 426         /* Fast path (cache hit) */
 427         sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
 428                                       bpf_rcu_lock_held());
 429         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 430                 return sdata;
 431
 432         /* Slow path (cache miss) */
 433         hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
 434                                   rcu_read_lock_trace_held())
 435                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 436                         break;
 437
 438         if (!selem)
 439                 return NULL;
 440
 441         sdata = SDATA(selem);
 442         if (cacheit_lockit) {
 443                 unsigned long flags;
 444
 445                 /* spinlock is needed to avoid racing with the
 446                  * parallel delete.  Otherwise, publishing an already
 447                  * deleted sdata to the cache will become a use-after-free
 448                  * problem in the next bpf_local_storage_lookup().
 449                  */
 450                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 451                 if (selem_linked_to_storage(selem))
 452                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 453                                            sdata);
 454                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 455         }
 456
 457         return sdata;
 458 }
 459
 460 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 461                        u64 map_flags)
 462 {
 463         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 464                 /* elem already exists */
 465                 return -EEXIST;
 466
 467         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 468                 /* elem doesn't exist, cannot update it */
 469                 return -ENOENT;
 470
 471         return 0;
 472 }
 473
 474 int bpf_local_storage_alloc(void *owner,
 475                             struct bpf_local_storage_map *smap,
 476                             struct bpf_local_storage_elem *first_selem,
 477                             gfp_t gfp_flags)
 478 {
 479         struct bpf_local_storage *prev_storage, *storage;
 480         struct bpf_local_storage **owner_storage_ptr;
 481         int err;
 482
 483         err = mem_charge(smap, owner, sizeof(*storage));
 484         if (err)
 485                 return err;
 486
 487         if (smap->bpf_ma) {
 488                 migrate_disable();
 489                 storage = bpf_mem_cache_alloc_flags(&smap->storage_ma, gfp_flags);
 490                 migrate_enable();
 491         } else {
 492                 storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 493                                           gfp_flags | __GFP_NOWARN);
 494         }
 495
 496         if (!storage) {
 497                 err = -ENOMEM;
 498                 goto uncharge;
 499         }
 500
 501         RCU_INIT_POINTER(storage->smap, smap);
 502         INIT_HLIST_HEAD(&storage->list);
 503         raw_spin_lock_init(&storage->lock);
 504         storage->owner = owner;
 505
 506         bpf_selem_link_storage_nolock(storage, first_selem);
 507         bpf_selem_link_map(smap, first_selem);
 508
 509         owner_storage_ptr =
 510                 (struct bpf_local_storage **)owner_storage(smap, owner);
 511         /* Publish storage to the owner.
 512          * Instead of using any lock of the kernel object (i.e. owner),
 513          * cmpxchg will work with any kernel object regardless what
 514          * the running context is, bh, irq...etc.
 515          *
 516          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 517          * is protected by the storage->lock.  Hence, when freeing
 518          * the owner->storage, the storage->lock must be held before
 519          * setting owner->storage ptr to NULL.
 520          */
 521         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 522         if (unlikely(prev_storage)) {
 523                 bpf_selem_unlink_map(first_selem);
 524                 err = -EAGAIN;
 525                 goto uncharge;
 526
 527                 /* Note that even first_selem was linked to smap's
 528                  * bucket->list, first_selem can be freed immediately
 529                  * (instead of kfree_rcu) because
 530                  * bpf_local_storage_map_free() does a
 531                  * synchronize_rcu_mult (waiting for both sleepable and
 532                  * normal programs) before walking the bucket->list.
 533                  * Hence, no one is accessing selem from the
 534                  * bucket->list under rcu_read_lock().
 535                  */
 536         }
 537
 538         return 0;
 539
 540 uncharge:
 541         bpf_local_storage_free(storage, smap, smap->bpf_ma, true);
 542         mem_uncharge(smap, owner, sizeof(*storage));
 543         return err;
 544 }
 545
 546 /* sk cannot be going away because it is linking new elem
 547  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 548  * Otherwise, it will become a leak (and other memory issues
 549  * during map destruction).
 550  */
 551 struct bpf_local_storage_data *
 552 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 553                          void *value, u64 map_flags, gfp_t gfp_flags)
 554 {
 555         struct bpf_local_storage_data *old_sdata = NULL;
 556         struct bpf_local_storage_elem *selem = NULL;
 557         struct bpf_local_storage *local_storage;
 558         unsigned long flags;
 559         int err;
 560
 561         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 562         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 563             /* BPF_F_LOCK can only be used in a value with spin_lock */
 564             unlikely((map_flags & BPF_F_LOCK) &&
 565                      !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
 566                 return ERR_PTR(-EINVAL);
 567
 568         if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
 569                 return ERR_PTR(-EINVAL);
 570
 571         local_storage = rcu_dereference_check(*owner_storage(smap, owner),
 572                                               bpf_rcu_lock_held());
 573         if (!local_storage || hlist_empty(&local_storage->list)) {
 574                 /* Very first elem for the owner */
 575                 err = check_flags(NULL, map_flags);
 576                 if (err)
 577                         return ERR_PTR(err);
 578
 579                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 580                 if (!selem)
 581                         return ERR_PTR(-ENOMEM);
 582
 583                 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
 584                 if (err) {
 585                         bpf_selem_free(selem, smap, true);
 586                         mem_uncharge(smap, owner, smap->elem_size);
 587                         return ERR_PTR(err);
 588                 }
 589
 590                 return SDATA(selem);
 591         }
 592
 593         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 594                 /* Hoping to find an old_sdata to do inline update
 595                  * such that it can avoid taking the local_storage->lock
 596                  * and changing the lists.
 597                  */
 598                 old_sdata =
 599                         bpf_local_storage_lookup(local_storage, smap, false);
 600                 err = check_flags(old_sdata, map_flags);
 601                 if (err)
 602                         return ERR_PTR(err);
 603                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 604                         copy_map_value_locked(&smap->map, old_sdata->data,
 605                                               value, false);
 606                         return old_sdata;
 607                 }
 608         }
 609
 610         if (gfp_flags == GFP_KERNEL) {
 611                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 612                 if (!selem)
 613                         return ERR_PTR(-ENOMEM);
 614         }
 615
 616         raw_spin_lock_irqsave(&local_storage->lock, flags);
 617
 618         /* Recheck local_storage->list under local_storage->lock */
 619         if (unlikely(hlist_empty(&local_storage->list))) {
 620                 /* A parallel del is happening and local_storage is going
 621                  * away.  It has just been checked before, so very
 622                  * unlikely.  Return instead of retry to keep things
 623                  * simple.
 624                  */
 625                 err = -EAGAIN;
 626                 goto unlock_err;
 627         }
 628
 629         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 630         err = check_flags(old_sdata, map_flags);
 631         if (err)
 632                 goto unlock_err;
 633
 634         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 635                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 636                                       false);
 637                 selem = SELEM(old_sdata);
 638                 goto unlock;
 639         }
 640
 641         if (gfp_flags != GFP_KERNEL) {
 642                 /* local_storage->lock is held.  Hence, we are sure
 643                  * we can unlink and uncharge the old_sdata successfully
 644                  * later.  Hence, instead of charging the new selem now
 645                  * and then uncharge the old selem later (which may cause
 646                  * a potential but unnecessary charge failure),  avoid taking
 647                  * a charge at all here (the "!old_sdata" check) and the
 648                  * old_sdata will not be uncharged later during
 649                  * bpf_selem_unlink_storage_nolock().
 650                  */
 651                 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
 652                 if (!selem) {
 653                         err = -ENOMEM;
 654                         goto unlock_err;
 655                 }
 656         }
 657
 658         /* First, link the new selem to the map */
 659         bpf_selem_link_map(smap, selem);
 660
 661         /* Second, link (and publish) the new selem to local_storage */
 662         bpf_selem_link_storage_nolock(local_storage, selem);
 663
 664         /* Third, remove old selem, SELEM(old_sdata) */
 665         if (old_sdata) {
 666                 bpf_selem_unlink_map(SELEM(old_sdata));
 667                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 668                                                 false, false);
 669         }
 670
 671 unlock:
 672         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 673         return SDATA(selem);
 674
 675 unlock_err:
 676         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 677         if (selem) {
 678                 mem_uncharge(smap, owner, smap->elem_size);
 679                 bpf_selem_free(selem, smap, true);
 680         }
 681         return ERR_PTR(err);
 682 }
 683
 684 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 685 {
 686         u64 min_usage = U64_MAX;
 687         u16 i, res = 0;
 688
 689         spin_lock(&cache->idx_lock);
 690
 691         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 692                 if (cache->idx_usage_counts[i] < min_usage) {
 693                         min_usage = cache->idx_usage_counts[i];
 694                         res = i;
 695
 696                         /* Found a free cache_idx */
 697                         if (!min_usage)
 698                                 break;
 699                 }
 700         }
 701         cache->idx_usage_counts[res]++;
 702
 703         spin_unlock(&cache->idx_lock);
 704
 705         return res;
 706 }
 707
 708 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 709                                              u16 idx)
 710 {
 711         spin_lock(&cache->idx_lock);
 712         cache->idx_usage_counts[idx]--;
 713         spin_unlock(&cache->idx_lock);
 714 }
 715
 716 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 717 {
 718         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 719             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 720             attr->max_entries ||
 721             attr->key_size != sizeof(int) || !attr->value_size ||
 722             /* Enforce BTF for userspace sk dumping */
 723             !attr->btf_key_type_id || !attr->btf_value_type_id)
 724                 return -EINVAL;
 725
 726         if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
 727                 return -E2BIG;
 728
 729         return 0;
 730 }
 731
 732 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
 733                                     const struct btf *btf,
 734                                     const struct btf_type *key_type,
 735                                     const struct btf_type *value_type)
 736 {
 737         u32 int_data;
 738
 739         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 740                 return -EINVAL;
 741
 742         int_data = *(u32 *)(key_type + 1);
 743         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 744                 return -EINVAL;
 745
 746         return 0;
 747 }
 748
 749 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
 750 {
 751         struct bpf_local_storage_map *storage_smap;
 752         struct bpf_local_storage_elem *selem;
 753         bool bpf_ma, free_storage = false;
 754         struct hlist_node *n;
 755         unsigned long flags;
 756
 757         storage_smap = rcu_dereference_check(local_storage->smap, bpf_rcu_lock_held());
 758         bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, NULL);
 759
 760         /* Neither the bpf_prog nor the bpf_map's syscall
 761          * could be modifying the local_storage->list now.
 762          * Thus, no elem can be added to or deleted from the
 763          * local_storage->list by the bpf_prog or by the bpf_map's syscall.
 764          *
 765          * It is racing with bpf_local_storage_map_free() alone
 766          * when unlinking elem from the local_storage->list and
 767          * the map's bucket->list.
 768          */
 769         raw_spin_lock_irqsave(&local_storage->lock, flags);
 770         hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
 771                 /* Always unlink from map before unlinking from
 772                  * local_storage.
 773                  */
 774                 bpf_selem_unlink_map(selem);
 775                 /* If local_storage list has only one element, the
 776                  * bpf_selem_unlink_storage_nolock() will return true.
 777                  * Otherwise, it will return false. The current loop iteration
 778                  * intends to remove all local storage. So the last iteration
 779                  * of the loop will set the free_cgroup_storage to true.
 780                  */
 781                 free_storage = bpf_selem_unlink_storage_nolock(
 782                         local_storage, selem, false, true);
 783         }
 784         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 785
 786         if (free_storage)
 787                 bpf_local_storage_free(local_storage, storage_smap, bpf_ma, true);
 788 }
 789
 790 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
 791 {
 792         struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
 793         u64 usage = sizeof(*smap);
 794
 795         /* The dynamically callocated selems are not counted currently. */
 796         usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
 797         return usage;
 798 }
 799
 800 /* When bpf_ma == true, the bpf_mem_alloc is used to allocate and free memory.
 801  * A deadlock free allocator is useful for storage that the bpf prog can easily
 802  * get a hold of the owner PTR_TO_BTF_ID in any context. eg. bpf_get_current_task_btf.
 803  * The task and cgroup storage fall into this case. The bpf_mem_alloc reuses
 804  * memory immediately. To be reuse-immediate safe, the owner destruction
 805  * code path needs to go through a rcu grace period before calling
 806  * bpf_local_storage_destroy().
 807  *
 808  * When bpf_ma == false, the kmalloc and kfree are used.
 809  */
 810 struct bpf_map *
 811 bpf_local_storage_map_alloc(union bpf_attr *attr,
 812                             struct bpf_local_storage_cache *cache,
 813                             bool bpf_ma)
 814 {
 815         struct bpf_local_storage_map *smap;
 816         unsigned int i;
 817         u32 nbuckets;
 818         int err;
 819
 820         smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
 821         if (!smap)
 822                 return ERR_PTR(-ENOMEM);
 823         bpf_map_init_from_attr(&smap->map, attr);
 824
 825         nbuckets = roundup_pow_of_two(num_possible_cpus());
 826         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 827         nbuckets = max_t(u32, 2, nbuckets);
 828         smap->bucket_log = ilog2(nbuckets);
 829
 830         smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
 831                                          nbuckets, GFP_USER | __GFP_NOWARN);
 832         if (!smap->buckets) {
 833                 err = -ENOMEM;
 834                 goto free_smap;
 835         }
 836
 837         for (i = 0; i < nbuckets; i++) {
 838                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 839                 raw_spin_lock_init(&smap->buckets[i].lock);
 840         }
 841
 842         smap->elem_size = offsetof(struct bpf_local_storage_elem,
 843                                    sdata.data[attr->value_size]);
 844
 845         smap->bpf_ma = bpf_ma;
 846         if (bpf_ma) {
 847                 err = bpf_mem_alloc_init(&smap->selem_ma, smap->elem_size, false);
 848                 if (err)
 849                         goto free_smap;
 850
 851                 err = bpf_mem_alloc_init(&smap->storage_ma, sizeof(struct bpf_local_storage), false);
 852                 if (err) {
 853                         bpf_mem_alloc_destroy(&smap->selem_ma);
 854                         goto free_smap;
 855                 }
 856         }
 857
 858         smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
 859         return &smap->map;
 860
 861 free_smap:
 862         kvfree(smap->buckets);
 863         bpf_map_area_free(smap);
 864         return ERR_PTR(err);
 865 }
 866
 867 void bpf_local_storage_map_free(struct bpf_map *map,
 868                                 struct bpf_local_storage_cache *cache,
 869                                 int __percpu *busy_counter)
 870 {
 871         struct bpf_local_storage_map_bucket *b;
 872         struct bpf_local_storage_elem *selem;
 873         struct bpf_local_storage_map *smap;
 874         unsigned int i;
 875
 876         smap = (struct bpf_local_storage_map *)map;
 877         bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
 878
 879         /* Note that this map might be concurrently cloned from
 880          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 881          * RCU read section to finish before proceeding. New RCU
 882          * read sections should be prevented via bpf_map_inc_not_zero.
 883          */
 884         synchronize_rcu();
 885
 886         /* bpf prog and the userspace can no longer access this map
 887          * now.  No new selem (of this map) can be added
 888          * to the owner->storage or to the map bucket's list.
 889          *
 890          * The elem of this map can be cleaned up here
 891          * or when the storage is freed e.g.
 892          * by bpf_sk_storage_free() during __sk_destruct().
 893          */
 894         for (i = 0; i < (1U << smap->bucket_log); i++) {
 895                 b = &smap->buckets[i];
 896
 897                 rcu_read_lock();
 898                 /* No one is adding to b->list now */
 899                 while ((selem = hlist_entry_safe(
 900                                 rcu_dereference_raw(hlist_first_rcu(&b->list)),
 901                                 struct bpf_local_storage_elem, map_node))) {
 902                         if (busy_counter) {
 903                                 migrate_disable();
 904                                 this_cpu_inc(*busy_counter);
 905                         }
 906                         bpf_selem_unlink(selem, true);
 907                         if (busy_counter) {
 908                                 this_cpu_dec(*busy_counter);
 909                                 migrate_enable();
 910                         }
 911                         cond_resched_rcu();
 912                 }
 913                 rcu_read_unlock();
 914         }
 915
 916         /* While freeing the storage we may still need to access the map.
 917          *
 918          * e.g. when bpf_sk_storage_free() has unlinked selem from the map
 919          * which then made the above while((selem = ...)) loop
 920          * exit immediately.
 921          *
 922          * However, while freeing the storage one still needs to access the
 923          * smap->elem_size to do the uncharging in
 924          * bpf_selem_unlink_storage_nolock().
 925          *
 926          * Hence, wait another rcu grace period for the storage to be freed.
 927          */
 928         synchronize_rcu();
 929
 930         if (smap->bpf_ma) {
 931                 bpf_mem_alloc_destroy(&smap->selem_ma);
 932                 bpf_mem_alloc_destroy(&smap->storage_ma);
 933         }
 934         kvfree(smap->buckets);
 935         bpf_map_area_free(smap);
 936 }