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 (!bpf_ma) {
 161                 __bpf_local_storage_free(local_storage, reuse_now);
 162                 return;
 163         }
 164
 165         if (!reuse_now) {
 166                 call_rcu_tasks_trace(&local_storage->rcu,
 167                                      bpf_local_storage_free_trace_rcu);
 168                 return;
 169         }
 170
 171         if (smap) {
 172                 migrate_disable();
 173                 bpf_mem_cache_free(&smap->storage_ma, local_storage);
 174                 migrate_enable();
 175         } else {
 176                 /* smap could be NULL if the selem that triggered
 177                  * this 'local_storage' creation had been long gone.
 178                  * In this case, directly do call_rcu().
 179                  */
 180                 call_rcu(&local_storage->rcu, bpf_local_storage_free_rcu);
 181         }
 182 }
 183
 184 /* rcu tasks trace callback for bpf_ma == false */
 185 static void __bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 186 {
 187         struct bpf_local_storage_elem *selem;
 188
 189         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 190         if (rcu_trace_implies_rcu_gp())
 191                 kfree(selem);
 192         else
 193                 kfree_rcu(selem, rcu);
 194 }
 195
 196 /* Handle bpf_ma == false */
 197 static void __bpf_selem_free(struct bpf_local_storage_elem *selem,
 198                              bool vanilla_rcu)
 199 {
 200         if (vanilla_rcu)
 201                 kfree_rcu(selem, rcu);
 202         else
 203                 call_rcu_tasks_trace(&selem->rcu, __bpf_selem_free_trace_rcu);
 204 }
 205
 206 static void bpf_selem_free_rcu(struct rcu_head *rcu)
 207 {
 208         struct bpf_local_storage_elem *selem;
 209
 210         selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
 211         bpf_mem_cache_raw_free(selem);
 212 }
 213
 214 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
 215 {
 216         if (rcu_trace_implies_rcu_gp())
 217                 bpf_selem_free_rcu(rcu);
 218         else
 219                 call_rcu(rcu, bpf_selem_free_rcu);
 220 }
 221
 222 void bpf_selem_free(struct bpf_local_storage_elem *selem,
 223                     struct bpf_local_storage_map *smap,
 224                     bool reuse_now)
 225 {
 226         bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
 227
 228         if (!smap->bpf_ma) {
 229                 __bpf_selem_free(selem, reuse_now);
 230                 return;
 231         }
 232
 233         if (!reuse_now) {
 234                 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
 235         } else {
 236                 /* Instead of using the vanilla call_rcu(),
 237                  * bpf_mem_cache_free will be able to reuse selem
 238                  * immediately.
 239                  */
 240                 migrate_disable();
 241                 bpf_mem_cache_free(&smap->selem_ma, selem);
 242                 migrate_enable();
 243         }
 244 }
 245
 246 /* local_storage->lock must be held and selem->local_storage == local_storage.
 247  * The caller must ensure selem->smap is still valid to be
 248  * dereferenced for its smap->elem_size and smap->cache_idx.
 249  */
 250 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
 251                                             struct bpf_local_storage_elem *selem,
 252                                             bool uncharge_mem, bool reuse_now)
 253 {
 254         struct bpf_local_storage_map *smap;
 255         bool free_local_storage;
 256         void *owner;
 257
 258         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 259         owner = local_storage->owner;
 260
 261         /* All uncharging on the owner must be done first.
 262          * The owner may be freed once the last selem is unlinked
 263          * from local_storage.
 264          */
 265         if (uncharge_mem)
 266                 mem_uncharge(smap, owner, smap->elem_size);
 267
 268         free_local_storage = hlist_is_singular_node(&selem->snode,
 269                                                     &local_storage->list);
 270         if (free_local_storage) {
 271                 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
 272                 local_storage->owner = NULL;
 273
 274                 /* After this RCU_INIT, owner may be freed and cannot be used */
 275                 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
 276
 277                 /* local_storage is not freed now.  local_storage->lock is
 278                  * still held and raw_spin_unlock_bh(&local_storage->lock)
 279                  * will be done by the caller.
 280                  *
 281                  * Although the unlock will be done under
 282                  * rcu_read_lock(),  it is more intuitive to
 283                  * read if the freeing of the storage is done
 284                  * after the raw_spin_unlock_bh(&local_storage->lock).
 285                  *
 286                  * Hence, a "bool free_local_storage" is returned
 287                  * to the caller which then calls then frees the storage after
 288                  * all the RCU grace periods have expired.
 289                  */
 290         }
 291         hlist_del_init_rcu(&selem->snode);
 292         if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
 293             SDATA(selem))
 294                 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
 295
 296         bpf_selem_free(selem, smap, reuse_now);
 297
 298         if (rcu_access_pointer(local_storage->smap) == smap)
 299                 RCU_INIT_POINTER(local_storage->smap, NULL);
 300
 301         return free_local_storage;
 302 }
 303
 304 static bool check_storage_bpf_ma(struct bpf_local_storage *local_storage,
 305                                  struct bpf_local_storage_map *storage_smap,
 306                                  struct bpf_local_storage_elem *selem)
 307 {
 308
 309         struct bpf_local_storage_map *selem_smap;
 310
 311         /* local_storage->smap may be NULL. If it is, get the bpf_ma
 312          * from any selem in the local_storage->list. The bpf_ma of all
 313          * local_storage and selem should have the same value
 314          * for the same map type.
 315          *
 316          * If the local_storage->list is already empty, the caller will not
 317          * care about the bpf_ma value also because the caller is not
 318          * responsibile to free the local_storage.
 319          */
 320
 321         if (storage_smap)
 322                 return storage_smap->bpf_ma;
 323
 324         if (!selem) {
 325                 struct hlist_node *n;
 326
 327                 n = rcu_dereference_check(hlist_first_rcu(&local_storage->list),
 328                                           bpf_rcu_lock_held());
 329                 if (!n)
 330                         return false;
 331
 332                 selem = hlist_entry(n, struct bpf_local_storage_elem, snode);
 333         }
 334         selem_smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 335
 336         return selem_smap->bpf_ma;
 337 }
 338
 339 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
 340                                      bool reuse_now)
 341 {
 342         struct bpf_local_storage_map *storage_smap;
 343         struct bpf_local_storage *local_storage;
 344         bool bpf_ma, free_local_storage = false;
 345         unsigned long flags;
 346
 347         if (unlikely(!selem_linked_to_storage_lockless(selem)))
 348                 /* selem has already been unlinked from sk */
 349                 return;
 350
 351         local_storage = rcu_dereference_check(selem->local_storage,
 352                                               bpf_rcu_lock_held());
 353         storage_smap = rcu_dereference_check(local_storage->smap,
 354                                              bpf_rcu_lock_held());
 355         bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, selem);
 356
 357         raw_spin_lock_irqsave(&local_storage->lock, flags);
 358         if (likely(selem_linked_to_storage(selem)))
 359                 free_local_storage = bpf_selem_unlink_storage_nolock(
 360                         local_storage, selem, true, reuse_now);
 361         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 362
 363         if (free_local_storage)
 364                 bpf_local_storage_free(local_storage, storage_smap, bpf_ma, reuse_now);
 365 }
 366
 367 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
 368                                    struct bpf_local_storage_elem *selem)
 369 {
 370         RCU_INIT_POINTER(selem->local_storage, local_storage);
 371         hlist_add_head_rcu(&selem->snode, &local_storage->list);
 372 }
 373
 374 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
 375 {
 376         struct bpf_local_storage_map *smap;
 377         struct bpf_local_storage_map_bucket *b;
 378         unsigned long flags;
 379
 380         if (unlikely(!selem_linked_to_map_lockless(selem)))
 381                 /* selem has already be unlinked from smap */
 382                 return;
 383
 384         smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
 385         b = select_bucket(smap, selem);
 386         raw_spin_lock_irqsave(&b->lock, flags);
 387         if (likely(selem_linked_to_map(selem)))
 388                 hlist_del_init_rcu(&selem->map_node);
 389         raw_spin_unlock_irqrestore(&b->lock, flags);
 390 }
 391
 392 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
 393                         struct bpf_local_storage_elem *selem)
 394 {
 395         struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
 396         unsigned long flags;
 397
 398         raw_spin_lock_irqsave(&b->lock, flags);
 399         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 400         hlist_add_head_rcu(&selem->map_node, &b->list);
 401         raw_spin_unlock_irqrestore(&b->lock, flags);
 402 }
 403
 404 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
 405 {
 406         /* Always unlink from map before unlinking from local_storage
 407          * because selem will be freed after successfully unlinked from
 408          * the local_storage.
 409          */
 410         bpf_selem_unlink_map(selem);
 411         bpf_selem_unlink_storage(selem, reuse_now);
 412 }
 413
 414 /* If cacheit_lockit is false, this lookup function is lockless */
 415 struct bpf_local_storage_data *
 416 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
 417                          struct bpf_local_storage_map *smap,
 418                          bool cacheit_lockit)
 419 {
 420         struct bpf_local_storage_data *sdata;
 421         struct bpf_local_storage_elem *selem;
 422
 423         /* Fast path (cache hit) */
 424         sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
 425                                       bpf_rcu_lock_held());
 426         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 427                 return sdata;
 428
 429         /* Slow path (cache miss) */
 430         hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
 431                                   rcu_read_lock_trace_held())
 432                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 433                         break;
 434
 435         if (!selem)
 436                 return NULL;
 437
 438         sdata = SDATA(selem);
 439         if (cacheit_lockit) {
 440                 unsigned long flags;
 441
 442                 /* spinlock is needed to avoid racing with the
 443                  * parallel delete.  Otherwise, publishing an already
 444                  * deleted sdata to the cache will become a use-after-free
 445                  * problem in the next bpf_local_storage_lookup().
 446                  */
 447                 raw_spin_lock_irqsave(&local_storage->lock, flags);
 448                 if (selem_linked_to_storage(selem))
 449                         rcu_assign_pointer(local_storage->cache[smap->cache_idx],
 450                                            sdata);
 451                 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 452         }
 453
 454         return sdata;
 455 }
 456
 457 static int check_flags(const struct bpf_local_storage_data *old_sdata,
 458                        u64 map_flags)
 459 {
 460         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 461                 /* elem already exists */
 462                 return -EEXIST;
 463
 464         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 465                 /* elem doesn't exist, cannot update it */
 466                 return -ENOENT;
 467
 468         return 0;
 469 }
 470
 471 int bpf_local_storage_alloc(void *owner,
 472                             struct bpf_local_storage_map *smap,
 473                             struct bpf_local_storage_elem *first_selem,
 474                             gfp_t gfp_flags)
 475 {
 476         struct bpf_local_storage *prev_storage, *storage;
 477         struct bpf_local_storage **owner_storage_ptr;
 478         int err;
 479
 480         err = mem_charge(smap, owner, sizeof(*storage));
 481         if (err)
 482                 return err;
 483
 484         if (smap->bpf_ma) {
 485                 migrate_disable();
 486                 storage = bpf_mem_cache_alloc_flags(&smap->storage_ma, gfp_flags);
 487                 migrate_enable();
 488         } else {
 489                 storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
 490                                           gfp_flags | __GFP_NOWARN);
 491         }
 492
 493         if (!storage) {
 494                 err = -ENOMEM;
 495                 goto uncharge;
 496         }
 497
 498         RCU_INIT_POINTER(storage->smap, smap);
 499         INIT_HLIST_HEAD(&storage->list);
 500         raw_spin_lock_init(&storage->lock);
 501         storage->owner = owner;
 502
 503         bpf_selem_link_storage_nolock(storage, first_selem);
 504         bpf_selem_link_map(smap, first_selem);
 505
 506         owner_storage_ptr =
 507                 (struct bpf_local_storage **)owner_storage(smap, owner);
 508         /* Publish storage to the owner.
 509          * Instead of using any lock of the kernel object (i.e. owner),
 510          * cmpxchg will work with any kernel object regardless what
 511          * the running context is, bh, irq...etc.
 512          *
 513          * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
 514          * is protected by the storage->lock.  Hence, when freeing
 515          * the owner->storage, the storage->lock must be held before
 516          * setting owner->storage ptr to NULL.
 517          */
 518         prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
 519         if (unlikely(prev_storage)) {
 520                 bpf_selem_unlink_map(first_selem);
 521                 err = -EAGAIN;
 522                 goto uncharge;
 523
 524                 /* Note that even first_selem was linked to smap's
 525                  * bucket->list, first_selem can be freed immediately
 526                  * (instead of kfree_rcu) because
 527                  * bpf_local_storage_map_free() does a
 528                  * synchronize_rcu_mult (waiting for both sleepable and
 529                  * normal programs) before walking the bucket->list.
 530                  * Hence, no one is accessing selem from the
 531                  * bucket->list under rcu_read_lock().
 532                  */
 533         }
 534
 535         return 0;
 536
 537 uncharge:
 538         bpf_local_storage_free(storage, smap, smap->bpf_ma, true);
 539         mem_uncharge(smap, owner, sizeof(*storage));
 540         return err;
 541 }
 542
 543 /* sk cannot be going away because it is linking new elem
 544  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 545  * Otherwise, it will become a leak (and other memory issues
 546  * during map destruction).
 547  */
 548 struct bpf_local_storage_data *
 549 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
 550                          void *value, u64 map_flags, gfp_t gfp_flags)
 551 {
 552         struct bpf_local_storage_data *old_sdata = NULL;
 553         struct bpf_local_storage_elem *selem = NULL;
 554         struct bpf_local_storage *local_storage;
 555         unsigned long flags;
 556         int err;
 557
 558         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 559         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 560             /* BPF_F_LOCK can only be used in a value with spin_lock */
 561             unlikely((map_flags & BPF_F_LOCK) &&
 562                      !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
 563                 return ERR_PTR(-EINVAL);
 564
 565         if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
 566                 return ERR_PTR(-EINVAL);
 567
 568         local_storage = rcu_dereference_check(*owner_storage(smap, owner),
 569                                               bpf_rcu_lock_held());
 570         if (!local_storage || hlist_empty(&local_storage->list)) {
 571                 /* Very first elem for the owner */
 572                 err = check_flags(NULL, map_flags);
 573                 if (err)
 574                         return ERR_PTR(err);
 575
 576                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 577                 if (!selem)
 578                         return ERR_PTR(-ENOMEM);
 579
 580                 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
 581                 if (err) {
 582                         bpf_selem_free(selem, smap, true);
 583                         mem_uncharge(smap, owner, smap->elem_size);
 584                         return ERR_PTR(err);
 585                 }
 586
 587                 return SDATA(selem);
 588         }
 589
 590         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 591                 /* Hoping to find an old_sdata to do inline update
 592                  * such that it can avoid taking the local_storage->lock
 593                  * and changing the lists.
 594                  */
 595                 old_sdata =
 596                         bpf_local_storage_lookup(local_storage, smap, false);
 597                 err = check_flags(old_sdata, map_flags);
 598                 if (err)
 599                         return ERR_PTR(err);
 600                 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
 601                         copy_map_value_locked(&smap->map, old_sdata->data,
 602                                               value, false);
 603                         return old_sdata;
 604                 }
 605         }
 606
 607         if (gfp_flags == GFP_KERNEL) {
 608                 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
 609                 if (!selem)
 610                         return ERR_PTR(-ENOMEM);
 611         }
 612
 613         raw_spin_lock_irqsave(&local_storage->lock, flags);
 614
 615         /* Recheck local_storage->list under local_storage->lock */
 616         if (unlikely(hlist_empty(&local_storage->list))) {
 617                 /* A parallel del is happening and local_storage is going
 618                  * away.  It has just been checked before, so very
 619                  * unlikely.  Return instead of retry to keep things
 620                  * simple.
 621                  */
 622                 err = -EAGAIN;
 623                 goto unlock_err;
 624         }
 625
 626         old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
 627         err = check_flags(old_sdata, map_flags);
 628         if (err)
 629                 goto unlock_err;
 630
 631         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 632                 copy_map_value_locked(&smap->map, old_sdata->data, value,
 633                                       false);
 634                 selem = SELEM(old_sdata);
 635                 goto unlock;
 636         }
 637
 638         if (gfp_flags != GFP_KERNEL) {
 639                 /* local_storage->lock is held.  Hence, we are sure
 640                  * we can unlink and uncharge the old_sdata successfully
 641                  * later.  Hence, instead of charging the new selem now
 642                  * and then uncharge the old selem later (which may cause
 643                  * a potential but unnecessary charge failure),  avoid taking
 644                  * a charge at all here (the "!old_sdata" check) and the
 645                  * old_sdata will not be uncharged later during
 646                  * bpf_selem_unlink_storage_nolock().
 647                  */
 648                 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
 649                 if (!selem) {
 650                         err = -ENOMEM;
 651                         goto unlock_err;
 652                 }
 653         }
 654
 655         /* First, link the new selem to the map */
 656         bpf_selem_link_map(smap, selem);
 657
 658         /* Second, link (and publish) the new selem to local_storage */
 659         bpf_selem_link_storage_nolock(local_storage, selem);
 660
 661         /* Third, remove old selem, SELEM(old_sdata) */
 662         if (old_sdata) {
 663                 bpf_selem_unlink_map(SELEM(old_sdata));
 664                 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
 665                                                 false, false);
 666         }
 667
 668 unlock:
 669         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 670         return SDATA(selem);
 671
 672 unlock_err:
 673         raw_spin_unlock_irqrestore(&local_storage->lock, flags);
 674         if (selem) {
 675                 mem_uncharge(smap, owner, smap->elem_size);
 676                 bpf_selem_free(selem, smap, true);
 677         }
 678         return ERR_PTR(err);
 679 }
 680
 681 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
 682 {
 683         u64 min_usage = U64_MAX;
 684         u16 i, res = 0;
 685
 686         spin_lock(&cache->idx_lock);
 687
 688         for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
 689                 if (cache->idx_usage_counts[i] < min_usage) {
 690                         min_usage = cache->idx_usage_counts[i];
 691                         res = i;
 692
 693                         /* Found a free cache_idx */
 694                         if (!min_usage)
 695                                 break;
 696                 }
 697         }
 698         cache->idx_usage_counts[res]++;
 699
 700         spin_unlock(&cache->idx_lock);
 701
 702         return res;
 703 }
 704
 705 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
 706                                              u16 idx)
 707 {
 708         spin_lock(&cache->idx_lock);
 709         cache->idx_usage_counts[idx]--;
 710         spin_unlock(&cache->idx_lock);
 711 }
 712
 713 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
 714 {
 715         if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
 716             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 717             attr->max_entries ||
 718             attr->key_size != sizeof(int) || !attr->value_size ||
 719             /* Enforce BTF for userspace sk dumping */
 720             !attr->btf_key_type_id || !attr->btf_value_type_id)
 721                 return -EINVAL;
 722
 723         if (!bpf_capable())
 724                 return -EPERM;
 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 }