net/core/bpf_sk_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 <net/bpf_sk_storage.h>
  10 #include <net/sock.h>
  11 #include <uapi/linux/sock_diag.h>
  12 #include <uapi/linux/btf.h>
  13
  14 #define SK_STORAGE_CREATE_FLAG_MASK                                     \
  15         (BPF_F_NO_PREALLOC | BPF_F_CLONE)
  16
  17 struct bucket {
  18         struct hlist_head list;
  19         raw_spinlock_t lock;
  20 };
  21
  22 /* Thp map is not the primary owner of a bpf_sk_storage_elem.
  23  * Instead, the sk->sk_bpf_storage is.
  24  *
  25  * The map (bpf_sk_storage_map) is for two purposes
  26  * 1. Define the size of the "sk local storage".  It is
  27  *    the map's value_size.
  28  *
  29  * 2. Maintain a list to keep track of all elems such
  30  *    that they can be cleaned up during the map destruction.
  31  *
  32  * When a bpf local storage is being looked up for a
  33  * particular sk,  the "bpf_map" pointer is actually used
  34  * as the "key" to search in the list of elem in
  35  * sk->sk_bpf_storage.
  36  *
  37  * Hence, consider sk->sk_bpf_storage is the mini-map
  38  * with the "bpf_map" pointer as the searching key.
  39  */
  40 struct bpf_sk_storage_map {
  41         struct bpf_map map;
  42         /* Lookup elem does not require accessing the map.
  43          *
  44          * Updating/Deleting requires a bucket lock to
  45          * link/unlink the elem from the map.  Having
  46          * multiple buckets to improve contention.
  47          */
  48         struct bucket *buckets;
  49         u32 bucket_log;
  50         u16 elem_size;
  51         u16 cache_idx;
  52 };
  53
  54 struct bpf_sk_storage_data {
  55         /* smap is used as the searching key when looking up
  56          * from sk->sk_bpf_storage.
  57          *
  58          * Put it in the same cacheline as the data to minimize
  59          * the number of cachelines access during the cache hit case.
  60          */
  61         struct bpf_sk_storage_map __rcu *smap;
  62         u8 data[] __aligned(8);
  63 };
  64
  65 /* Linked to bpf_sk_storage and bpf_sk_storage_map */
  66 struct bpf_sk_storage_elem {
  67         struct hlist_node map_node;     /* Linked to bpf_sk_storage_map */
  68         struct hlist_node snode;        /* Linked to bpf_sk_storage */
  69         struct bpf_sk_storage __rcu *sk_storage;
  70         struct rcu_head rcu;
  71         /* 8 bytes hole */
  72         /* The data is stored in aother cacheline to minimize
  73          * the number of cachelines access during a cache hit.
  74          */
  75         struct bpf_sk_storage_data sdata ____cacheline_aligned;
  76 };
  77
  78 #define SELEM(_SDATA) container_of((_SDATA), struct bpf_sk_storage_elem, sdata)
  79 #define SDATA(_SELEM) (&(_SELEM)->sdata)
  80 #define BPF_SK_STORAGE_CACHE_SIZE       16
  81
  82 static DEFINE_SPINLOCK(cache_idx_lock);
  83 static u64 cache_idx_usage_counts[BPF_SK_STORAGE_CACHE_SIZE];
  84
  85 struct bpf_sk_storage {
  86         struct bpf_sk_storage_data __rcu *cache[BPF_SK_STORAGE_CACHE_SIZE];
  87         struct hlist_head list; /* List of bpf_sk_storage_elem */
  88         struct sock *sk;        /* The sk that owns the the above "list" of
  89                                  * bpf_sk_storage_elem.
  90                                  */
  91         struct rcu_head rcu;
  92         raw_spinlock_t lock;    /* Protect adding/removing from the "list" */
  93 };
  94
  95 static struct bucket *select_bucket(struct bpf_sk_storage_map *smap,
  96                                     struct bpf_sk_storage_elem *selem)
  97 {
  98         return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
  99 }
 100
 101 static int omem_charge(struct sock *sk, unsigned int size)
 102 {
 103         /* same check as in sock_kmalloc() */
 104         if (size <= sysctl_optmem_max &&
 105             atomic_read(&sk->sk_omem_alloc) + size < sysctl_optmem_max) {
 106                 atomic_add(size, &sk->sk_omem_alloc);
 107                 return 0;
 108         }
 109
 110         return -ENOMEM;
 111 }
 112
 113 static bool selem_linked_to_sk(const struct bpf_sk_storage_elem *selem)
 114 {
 115         return !hlist_unhashed(&selem->snode);
 116 }
 117
 118 static bool selem_linked_to_map(const struct bpf_sk_storage_elem *selem)
 119 {
 120         return !hlist_unhashed(&selem->map_node);
 121 }
 122
 123 static struct bpf_sk_storage_elem *selem_alloc(struct bpf_sk_storage_map *smap,
 124                                                struct sock *sk, void *value,
 125                                                bool charge_omem)
 126 {
 127         struct bpf_sk_storage_elem *selem;
 128
 129         if (charge_omem && omem_charge(sk, smap->elem_size))
 130                 return NULL;
 131
 132         selem = kzalloc(smap->elem_size, GFP_ATOMIC | __GFP_NOWARN);
 133         if (selem) {
 134                 if (value)
 135                         memcpy(SDATA(selem)->data, value, smap->map.value_size);
 136                 return selem;
 137         }
 138
 139         if (charge_omem)
 140                 atomic_sub(smap->elem_size, &sk->sk_omem_alloc);
 141
 142         return NULL;
 143 }
 144
 145 /* sk_storage->lock must be held and selem->sk_storage == sk_storage.
 146  * The caller must ensure selem->smap is still valid to be
 147  * dereferenced for its smap->elem_size and smap->cache_idx.
 148  */
 149 static bool __selem_unlink_sk(struct bpf_sk_storage *sk_storage,
 150                               struct bpf_sk_storage_elem *selem,
 151                               bool uncharge_omem)
 152 {
 153         struct bpf_sk_storage_map *smap;
 154         bool free_sk_storage;
 155         struct sock *sk;
 156
 157         smap = rcu_dereference(SDATA(selem)->smap);
 158         sk = sk_storage->sk;
 159
 160         /* All uncharging on sk->sk_omem_alloc must be done first.
 161          * sk may be freed once the last selem is unlinked from sk_storage.
 162          */
 163         if (uncharge_omem)
 164                 atomic_sub(smap->elem_size, &sk->sk_omem_alloc);
 165
 166         free_sk_storage = hlist_is_singular_node(&selem->snode,
 167                                                  &sk_storage->list);
 168         if (free_sk_storage) {
 169                 atomic_sub(sizeof(struct bpf_sk_storage), &sk->sk_omem_alloc);
 170                 sk_storage->sk = NULL;
 171                 /* After this RCU_INIT, sk may be freed and cannot be used */
 172                 RCU_INIT_POINTER(sk->sk_bpf_storage, NULL);
 173
 174                 /* sk_storage is not freed now.  sk_storage->lock is
 175                  * still held and raw_spin_unlock_bh(&sk_storage->lock)
 176                  * will be done by the caller.
 177                  *
 178                  * Although the unlock will be done under
 179                  * rcu_read_lock(),  it is more intutivie to
 180                  * read if kfree_rcu(sk_storage, rcu) is done
 181                  * after the raw_spin_unlock_bh(&sk_storage->lock).
 182                  *
 183                  * Hence, a "bool free_sk_storage" is returned
 184                  * to the caller which then calls the kfree_rcu()
 185                  * after unlock.
 186                  */
 187         }
 188         hlist_del_init_rcu(&selem->snode);
 189         if (rcu_access_pointer(sk_storage->cache[smap->cache_idx]) ==
 190             SDATA(selem))
 191                 RCU_INIT_POINTER(sk_storage->cache[smap->cache_idx], NULL);
 192
 193         kfree_rcu(selem, rcu);
 194
 195         return free_sk_storage;
 196 }
 197
 198 static void selem_unlink_sk(struct bpf_sk_storage_elem *selem)
 199 {
 200         struct bpf_sk_storage *sk_storage;
 201         bool free_sk_storage = false;
 202
 203         if (unlikely(!selem_linked_to_sk(selem)))
 204                 /* selem has already been unlinked from sk */
 205                 return;
 206
 207         sk_storage = rcu_dereference(selem->sk_storage);
 208         raw_spin_lock_bh(&sk_storage->lock);
 209         if (likely(selem_linked_to_sk(selem)))
 210                 free_sk_storage = __selem_unlink_sk(sk_storage, selem, true);
 211         raw_spin_unlock_bh(&sk_storage->lock);
 212
 213         if (free_sk_storage)
 214                 kfree_rcu(sk_storage, rcu);
 215 }
 216
 217 static void __selem_link_sk(struct bpf_sk_storage *sk_storage,
 218                             struct bpf_sk_storage_elem *selem)
 219 {
 220         RCU_INIT_POINTER(selem->sk_storage, sk_storage);
 221         hlist_add_head(&selem->snode, &sk_storage->list);
 222 }
 223
 224 static void selem_unlink_map(struct bpf_sk_storage_elem *selem)
 225 {
 226         struct bpf_sk_storage_map *smap;
 227         struct bucket *b;
 228
 229         if (unlikely(!selem_linked_to_map(selem)))
 230                 /* selem has already be unlinked from smap */
 231                 return;
 232
 233         smap = rcu_dereference(SDATA(selem)->smap);
 234         b = select_bucket(smap, selem);
 235         raw_spin_lock_bh(&b->lock);
 236         if (likely(selem_linked_to_map(selem)))
 237                 hlist_del_init_rcu(&selem->map_node);
 238         raw_spin_unlock_bh(&b->lock);
 239 }
 240
 241 static void selem_link_map(struct bpf_sk_storage_map *smap,
 242                            struct bpf_sk_storage_elem *selem)
 243 {
 244         struct bucket *b = select_bucket(smap, selem);
 245
 246         raw_spin_lock_bh(&b->lock);
 247         RCU_INIT_POINTER(SDATA(selem)->smap, smap);
 248         hlist_add_head_rcu(&selem->map_node, &b->list);
 249         raw_spin_unlock_bh(&b->lock);
 250 }
 251
 252 static void selem_unlink(struct bpf_sk_storage_elem *selem)
 253 {
 254         /* Always unlink from map before unlinking from sk_storage
 255          * because selem will be freed after successfully unlinked from
 256          * the sk_storage.
 257          */
 258         selem_unlink_map(selem);
 259         selem_unlink_sk(selem);
 260 }
 261
 262 static struct bpf_sk_storage_data *
 263 __sk_storage_lookup(struct bpf_sk_storage *sk_storage,
 264                     struct bpf_sk_storage_map *smap,
 265                     bool cacheit_lockit)
 266 {
 267         struct bpf_sk_storage_data *sdata;
 268         struct bpf_sk_storage_elem *selem;
 269
 270         /* Fast path (cache hit) */
 271         sdata = rcu_dereference(sk_storage->cache[smap->cache_idx]);
 272         if (sdata && rcu_access_pointer(sdata->smap) == smap)
 273                 return sdata;
 274
 275         /* Slow path (cache miss) */
 276         hlist_for_each_entry_rcu(selem, &sk_storage->list, snode)
 277                 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
 278                         break;
 279
 280         if (!selem)
 281                 return NULL;
 282
 283         sdata = SDATA(selem);
 284         if (cacheit_lockit) {
 285                 /* spinlock is needed to avoid racing with the
 286                  * parallel delete.  Otherwise, publishing an already
 287                  * deleted sdata to the cache will become a use-after-free
 288                  * problem in the next __sk_storage_lookup().
 289                  */
 290                 raw_spin_lock_bh(&sk_storage->lock);
 291                 if (selem_linked_to_sk(selem))
 292                         rcu_assign_pointer(sk_storage->cache[smap->cache_idx],
 293                                            sdata);
 294                 raw_spin_unlock_bh(&sk_storage->lock);
 295         }
 296
 297         return sdata;
 298 }
 299
 300 static struct bpf_sk_storage_data *
 301 sk_storage_lookup(struct sock *sk, struct bpf_map *map, bool cacheit_lockit)
 302 {
 303         struct bpf_sk_storage *sk_storage;
 304         struct bpf_sk_storage_map *smap;
 305
 306         sk_storage = rcu_dereference(sk->sk_bpf_storage);
 307         if (!sk_storage)
 308                 return NULL;
 309
 310         smap = (struct bpf_sk_storage_map *)map;
 311         return __sk_storage_lookup(sk_storage, smap, cacheit_lockit);
 312 }
 313
 314 static int check_flags(const struct bpf_sk_storage_data *old_sdata,
 315                        u64 map_flags)
 316 {
 317         if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 318                 /* elem already exists */
 319                 return -EEXIST;
 320
 321         if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 322                 /* elem doesn't exist, cannot update it */
 323                 return -ENOENT;
 324
 325         return 0;
 326 }
 327
 328 static int sk_storage_alloc(struct sock *sk,
 329                             struct bpf_sk_storage_map *smap,
 330                             struct bpf_sk_storage_elem *first_selem)
 331 {
 332         struct bpf_sk_storage *prev_sk_storage, *sk_storage;
 333         int err;
 334
 335         err = omem_charge(sk, sizeof(*sk_storage));
 336         if (err)
 337                 return err;
 338
 339         sk_storage = kzalloc(sizeof(*sk_storage), GFP_ATOMIC | __GFP_NOWARN);
 340         if (!sk_storage) {
 341                 err = -ENOMEM;
 342                 goto uncharge;
 343         }
 344         INIT_HLIST_HEAD(&sk_storage->list);
 345         raw_spin_lock_init(&sk_storage->lock);
 346         sk_storage->sk = sk;
 347
 348         __selem_link_sk(sk_storage, first_selem);
 349         selem_link_map(smap, first_selem);
 350         /* Publish sk_storage to sk.  sk->sk_lock cannot be acquired.
 351          * Hence, atomic ops is used to set sk->sk_bpf_storage
 352          * from NULL to the newly allocated sk_storage ptr.
 353          *
 354          * From now on, the sk->sk_bpf_storage pointer is protected
 355          * by the sk_storage->lock.  Hence,  when freeing
 356          * the sk->sk_bpf_storage, the sk_storage->lock must
 357          * be held before setting sk->sk_bpf_storage to NULL.
 358          */
 359         prev_sk_storage = cmpxchg((struct bpf_sk_storage **)&sk->sk_bpf_storage,
 360                                   NULL, sk_storage);
 361         if (unlikely(prev_sk_storage)) {
 362                 selem_unlink_map(first_selem);
 363                 err = -EAGAIN;
 364                 goto uncharge;
 365
 366                 /* Note that even first_selem was linked to smap's
 367                  * bucket->list, first_selem can be freed immediately
 368                  * (instead of kfree_rcu) because
 369                  * bpf_sk_storage_map_free() does a
 370                  * synchronize_rcu() before walking the bucket->list.
 371                  * Hence, no one is accessing selem from the
 372                  * bucket->list under rcu_read_lock().
 373                  */
 374         }
 375
 376         return 0;
 377
 378 uncharge:
 379         kfree(sk_storage);
 380         atomic_sub(sizeof(*sk_storage), &sk->sk_omem_alloc);
 381         return err;
 382 }
 383
 384 /* sk cannot be going away because it is linking new elem
 385  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
 386  * Otherwise, it will become a leak (and other memory issues
 387  * during map destruction).
 388  */
 389 static struct bpf_sk_storage_data *sk_storage_update(struct sock *sk,
 390                                                      struct bpf_map *map,
 391                                                      void *value,
 392                                                      u64 map_flags)
 393 {
 394         struct bpf_sk_storage_data *old_sdata = NULL;
 395         struct bpf_sk_storage_elem *selem;
 396         struct bpf_sk_storage *sk_storage;
 397         struct bpf_sk_storage_map *smap;
 398         int err;
 399
 400         /* BPF_EXIST and BPF_NOEXIST cannot be both set */
 401         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
 402             /* BPF_F_LOCK can only be used in a value with spin_lock */
 403             unlikely((map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
 404                 return ERR_PTR(-EINVAL);
 405
 406         smap = (struct bpf_sk_storage_map *)map;
 407         sk_storage = rcu_dereference(sk->sk_bpf_storage);
 408         if (!sk_storage || hlist_empty(&sk_storage->list)) {
 409                 /* Very first elem for this sk */
 410                 err = check_flags(NULL, map_flags);
 411                 if (err)
 412                         return ERR_PTR(err);
 413
 414                 selem = selem_alloc(smap, sk, value, true);
 415                 if (!selem)
 416                         return ERR_PTR(-ENOMEM);
 417
 418                 err = sk_storage_alloc(sk, smap, selem);
 419                 if (err) {
 420                         kfree(selem);
 421                         atomic_sub(smap->elem_size, &sk->sk_omem_alloc);
 422                         return ERR_PTR(err);
 423                 }
 424
 425                 return SDATA(selem);
 426         }
 427
 428         if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
 429                 /* Hoping to find an old_sdata to do inline update
 430                  * such that it can avoid taking the sk_storage->lock
 431                  * and changing the lists.
 432                  */
 433                 old_sdata = __sk_storage_lookup(sk_storage, smap, false);
 434                 err = check_flags(old_sdata, map_flags);
 435                 if (err)
 436                         return ERR_PTR(err);
 437                 if (old_sdata && selem_linked_to_sk(SELEM(old_sdata))) {
 438                         copy_map_value_locked(map, old_sdata->data,
 439                                               value, false);
 440                         return old_sdata;
 441                 }
 442         }
 443
 444         raw_spin_lock_bh(&sk_storage->lock);
 445
 446         /* Recheck sk_storage->list under sk_storage->lock */
 447         if (unlikely(hlist_empty(&sk_storage->list))) {
 448                 /* A parallel del is happening and sk_storage is going
 449                  * away.  It has just been checked before, so very
 450                  * unlikely.  Return instead of retry to keep things
 451                  * simple.
 452                  */
 453                 err = -EAGAIN;
 454                 goto unlock_err;
 455         }
 456
 457         old_sdata = __sk_storage_lookup(sk_storage, smap, false);
 458         err = check_flags(old_sdata, map_flags);
 459         if (err)
 460                 goto unlock_err;
 461
 462         if (old_sdata && (map_flags & BPF_F_LOCK)) {
 463                 copy_map_value_locked(map, old_sdata->data, value, false);
 464                 selem = SELEM(old_sdata);
 465                 goto unlock;
 466         }
 467
 468         /* sk_storage->lock is held.  Hence, we are sure
 469          * we can unlink and uncharge the old_sdata successfully
 470          * later.  Hence, instead of charging the new selem now
 471          * and then uncharge the old selem later (which may cause
 472          * a potential but unnecessary charge failure),  avoid taking
 473          * a charge at all here (the "!old_sdata" check) and the
 474          * old_sdata will not be uncharged later during __selem_unlink_sk().
 475          */
 476         selem = selem_alloc(smap, sk, value, !old_sdata);
 477         if (!selem) {
 478                 err = -ENOMEM;
 479                 goto unlock_err;
 480         }
 481
 482         /* First, link the new selem to the map */
 483         selem_link_map(smap, selem);
 484
 485         /* Second, link (and publish) the new selem to sk_storage */
 486         __selem_link_sk(sk_storage, selem);
 487
 488         /* Third, remove old selem, SELEM(old_sdata) */
 489         if (old_sdata) {
 490                 selem_unlink_map(SELEM(old_sdata));
 491                 __selem_unlink_sk(sk_storage, SELEM(old_sdata), false);
 492         }
 493
 494 unlock:
 495         raw_spin_unlock_bh(&sk_storage->lock);
 496         return SDATA(selem);
 497
 498 unlock_err:
 499         raw_spin_unlock_bh(&sk_storage->lock);
 500         return ERR_PTR(err);
 501 }
 502
 503 static int sk_storage_delete(struct sock *sk, struct bpf_map *map)
 504 {
 505         struct bpf_sk_storage_data *sdata;
 506
 507         sdata = sk_storage_lookup(sk, map, false);
 508         if (!sdata)
 509                 return -ENOENT;
 510
 511         selem_unlink(SELEM(sdata));
 512
 513         return 0;
 514 }
 515
 516 static u16 cache_idx_get(void)
 517 {
 518         u64 min_usage = U64_MAX;
 519         u16 i, res = 0;
 520
 521         spin_lock(&cache_idx_lock);
 522
 523         for (i = 0; i < BPF_SK_STORAGE_CACHE_SIZE; i++) {
 524                 if (cache_idx_usage_counts[i] < min_usage) {
 525                         min_usage = cache_idx_usage_counts[i];
 526                         res = i;
 527
 528                         /* Found a free cache_idx */
 529                         if (!min_usage)
 530                                 break;
 531                 }
 532         }
 533         cache_idx_usage_counts[res]++;
 534
 535         spin_unlock(&cache_idx_lock);
 536
 537         return res;
 538 }
 539
 540 static void cache_idx_free(u16 idx)
 541 {
 542         spin_lock(&cache_idx_lock);
 543         cache_idx_usage_counts[idx]--;
 544         spin_unlock(&cache_idx_lock);
 545 }
 546
 547 /* Called by __sk_destruct() & bpf_sk_storage_clone() */
 548 void bpf_sk_storage_free(struct sock *sk)
 549 {
 550         struct bpf_sk_storage_elem *selem;
 551         struct bpf_sk_storage *sk_storage;
 552         bool free_sk_storage = false;
 553         struct hlist_node *n;
 554
 555         rcu_read_lock();
 556         sk_storage = rcu_dereference(sk->sk_bpf_storage);
 557         if (!sk_storage) {
 558                 rcu_read_unlock();
 559                 return;
 560         }
 561
 562         /* Netiher the bpf_prog nor the bpf-map's syscall
 563          * could be modifying the sk_storage->list now.
 564          * Thus, no elem can be added-to or deleted-from the
 565          * sk_storage->list by the bpf_prog or by the bpf-map's syscall.
 566          *
 567          * It is racing with bpf_sk_storage_map_free() alone
 568          * when unlinking elem from the sk_storage->list and
 569          * the map's bucket->list.
 570          */
 571         raw_spin_lock_bh(&sk_storage->lock);
 572         hlist_for_each_entry_safe(selem, n, &sk_storage->list, snode) {
 573                 /* Always unlink from map before unlinking from
 574                  * sk_storage.
 575                  */
 576                 selem_unlink_map(selem);
 577                 free_sk_storage = __selem_unlink_sk(sk_storage, selem, true);
 578         }
 579         raw_spin_unlock_bh(&sk_storage->lock);
 580         rcu_read_unlock();
 581
 582         if (free_sk_storage)
 583                 kfree_rcu(sk_storage, rcu);
 584 }
 585
 586 static void bpf_sk_storage_map_free(struct bpf_map *map)
 587 {
 588         struct bpf_sk_storage_elem *selem;
 589         struct bpf_sk_storage_map *smap;
 590         struct bucket *b;
 591         unsigned int i;
 592
 593         smap = (struct bpf_sk_storage_map *)map;
 594
 595         cache_idx_free(smap->cache_idx);
 596
 597         /* Note that this map might be concurrently cloned from
 598          * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
 599          * RCU read section to finish before proceeding. New RCU
 600          * read sections should be prevented via bpf_map_inc_not_zero.
 601          */
 602         synchronize_rcu();
 603
 604         /* bpf prog and the userspace can no longer access this map
 605          * now.  No new selem (of this map) can be added
 606          * to the sk->sk_bpf_storage or to the map bucket's list.
 607          *
 608          * The elem of this map can be cleaned up here
 609          * or
 610          * by bpf_sk_storage_free() during __sk_destruct().
 611          */
 612         for (i = 0; i < (1U << smap->bucket_log); i++) {
 613                 b = &smap->buckets[i];
 614
 615                 rcu_read_lock();
 616                 /* No one is adding to b->list now */
 617                 while ((selem = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(&b->list)),
 618                                                  struct bpf_sk_storage_elem,
 619                                                  map_node))) {
 620                         selem_unlink(selem);
 621                         cond_resched_rcu();
 622                 }
 623                 rcu_read_unlock();
 624         }
 625
 626         /* bpf_sk_storage_free() may still need to access the map.
 627          * e.g. bpf_sk_storage_free() has unlinked selem from the map
 628          * which then made the above while((selem = ...)) loop
 629          * exited immediately.
 630          *
 631          * However, the bpf_sk_storage_free() still needs to access
 632          * the smap->elem_size to do the uncharging in
 633          * __selem_unlink_sk().
 634          *
 635          * Hence, wait another rcu grace period for the
 636          * bpf_sk_storage_free() to finish.
 637          */
 638         synchronize_rcu();
 639
 640         kvfree(smap->buckets);
 641         kfree(map);
 642 }
 643
 644 /* U16_MAX is much more than enough for sk local storage
 645  * considering a tcp_sock is ~2k.
 646  */
 647 #define MAX_VALUE_SIZE                                                  \
 648         min_t(u32,                                                      \
 649               (KMALLOC_MAX_SIZE - MAX_BPF_STACK - sizeof(struct bpf_sk_storage_elem)), \
 650               (U16_MAX - sizeof(struct bpf_sk_storage_elem)))
 651
 652 static int bpf_sk_storage_map_alloc_check(union bpf_attr *attr)
 653 {
 654         if (attr->map_flags & ~SK_STORAGE_CREATE_FLAG_MASK ||
 655             !(attr->map_flags & BPF_F_NO_PREALLOC) ||
 656             attr->max_entries ||
 657             attr->key_size != sizeof(int) || !attr->value_size ||
 658             /* Enforce BTF for userspace sk dumping */
 659             !attr->btf_key_type_id || !attr->btf_value_type_id)
 660                 return -EINVAL;
 661
 662         if (!bpf_capable())
 663                 return -EPERM;
 664
 665         if (attr->value_size > MAX_VALUE_SIZE)
 666                 return -E2BIG;
 667
 668         return 0;
 669 }
 670
 671 static struct bpf_map *bpf_sk_storage_map_alloc(union bpf_attr *attr)
 672 {
 673         struct bpf_sk_storage_map *smap;
 674         unsigned int i;
 675         u32 nbuckets;
 676         u64 cost;
 677         int ret;
 678
 679         smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN);
 680         if (!smap)
 681                 return ERR_PTR(-ENOMEM);
 682         bpf_map_init_from_attr(&smap->map, attr);
 683
 684         nbuckets = roundup_pow_of_two(num_possible_cpus());
 685         /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
 686         nbuckets = max_t(u32, 2, nbuckets);
 687         smap->bucket_log = ilog2(nbuckets);
 688         cost = sizeof(*smap->buckets) * nbuckets + sizeof(*smap);
 689
 690         ret = bpf_map_charge_init(&smap->map.memory, cost);
 691         if (ret < 0) {
 692                 kfree(smap);
 693                 return ERR_PTR(ret);
 694         }
 695
 696         smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets,
 697                                  GFP_USER | __GFP_NOWARN);
 698         if (!smap->buckets) {
 699                 bpf_map_charge_finish(&smap->map.memory);
 700                 kfree(smap);
 701                 return ERR_PTR(-ENOMEM);
 702         }
 703
 704         for (i = 0; i < nbuckets; i++) {
 705                 INIT_HLIST_HEAD(&smap->buckets[i].list);
 706                 raw_spin_lock_init(&smap->buckets[i].lock);
 707         }
 708
 709         smap->elem_size = sizeof(struct bpf_sk_storage_elem) + attr->value_size;
 710         smap->cache_idx = cache_idx_get();
 711
 712         return &smap->map;
 713 }
 714
 715 static int notsupp_get_next_key(struct bpf_map *map, void *key,
 716                                 void *next_key)
 717 {
 718         return -ENOTSUPP;
 719 }
 720
 721 static int bpf_sk_storage_map_check_btf(const struct bpf_map *map,
 722                                         const struct btf *btf,
 723                                         const struct btf_type *key_type,
 724                                         const struct btf_type *value_type)
 725 {
 726         u32 int_data;
 727
 728         if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
 729                 return -EINVAL;
 730
 731         int_data = *(u32 *)(key_type + 1);
 732         if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
 733                 return -EINVAL;
 734
 735         return 0;
 736 }
 737
 738 static void *bpf_fd_sk_storage_lookup_elem(struct bpf_map *map, void *key)
 739 {
 740         struct bpf_sk_storage_data *sdata;
 741         struct socket *sock;
 742         int fd, err;
 743
 744         fd = *(int *)key;
 745         sock = sockfd_lookup(fd, &err);
 746         if (sock) {
 747                 sdata = sk_storage_lookup(sock->sk, map, true);
 748                 sockfd_put(sock);
 749                 return sdata ? sdata->data : NULL;
 750         }
 751
 752         return ERR_PTR(err);
 753 }
 754
 755 static int bpf_fd_sk_storage_update_elem(struct bpf_map *map, void *key,
 756                                          void *value, u64 map_flags)
 757 {
 758         struct bpf_sk_storage_data *sdata;
 759         struct socket *sock;
 760         int fd, err;
 761
 762         fd = *(int *)key;
 763         sock = sockfd_lookup(fd, &err);
 764         if (sock) {
 765                 sdata = sk_storage_update(sock->sk, map, value, map_flags);
 766                 sockfd_put(sock);
 767                 return PTR_ERR_OR_ZERO(sdata);
 768         }
 769
 770         return err;
 771 }
 772
 773 static int bpf_fd_sk_storage_delete_elem(struct bpf_map *map, void *key)
 774 {
 775         struct socket *sock;
 776         int fd, err;
 777
 778         fd = *(int *)key;
 779         sock = sockfd_lookup(fd, &err);
 780         if (sock) {
 781                 err = sk_storage_delete(sock->sk, map);
 782                 sockfd_put(sock);
 783                 return err;
 784         }
 785
 786         return err;
 787 }
 788
 789 static struct bpf_sk_storage_elem *
 790 bpf_sk_storage_clone_elem(struct sock *newsk,
 791                           struct bpf_sk_storage_map *smap,
 792                           struct bpf_sk_storage_elem *selem)
 793 {
 794         struct bpf_sk_storage_elem *copy_selem;
 795
 796         copy_selem = selem_alloc(smap, newsk, NULL, true);
 797         if (!copy_selem)
 798                 return NULL;
 799
 800         if (map_value_has_spin_lock(&smap->map))
 801                 copy_map_value_locked(&smap->map, SDATA(copy_selem)->data,
 802                                       SDATA(selem)->data, true);
 803         else
 804                 copy_map_value(&smap->map, SDATA(copy_selem)->data,
 805                                SDATA(selem)->data);
 806
 807         return copy_selem;
 808 }
 809
 810 int bpf_sk_storage_clone(const struct sock *sk, struct sock *newsk)
 811 {
 812         struct bpf_sk_storage *new_sk_storage = NULL;
 813         struct bpf_sk_storage *sk_storage;
 814         struct bpf_sk_storage_elem *selem;
 815         int ret = 0;
 816
 817         RCU_INIT_POINTER(newsk->sk_bpf_storage, NULL);
 818
 819         rcu_read_lock();
 820         sk_storage = rcu_dereference(sk->sk_bpf_storage);
 821
 822         if (!sk_storage || hlist_empty(&sk_storage->list))
 823                 goto out;
 824
 825         hlist_for_each_entry_rcu(selem, &sk_storage->list, snode) {
 826                 struct bpf_sk_storage_elem *copy_selem;
 827                 struct bpf_sk_storage_map *smap;
 828                 struct bpf_map *map;
 829
 830                 smap = rcu_dereference(SDATA(selem)->smap);
 831                 if (!(smap->map.map_flags & BPF_F_CLONE))
 832                         continue;
 833
 834                 /* Note that for lockless listeners adding new element
 835                  * here can race with cleanup in bpf_sk_storage_map_free.
 836                  * Try to grab map refcnt to make sure that it's still
 837                  * alive and prevent concurrent removal.
 838                  */
 839                 map = bpf_map_inc_not_zero(&smap->map);
 840                 if (IS_ERR(map))
 841                         continue;
 842
 843                 copy_selem = bpf_sk_storage_clone_elem(newsk, smap, selem);
 844                 if (!copy_selem) {
 845                         ret = -ENOMEM;
 846                         bpf_map_put(map);
 847                         goto out;
 848                 }
 849
 850                 if (new_sk_storage) {
 851                         selem_link_map(smap, copy_selem);
 852                         __selem_link_sk(new_sk_storage, copy_selem);
 853                 } else {
 854                         ret = sk_storage_alloc(newsk, smap, copy_selem);
 855                         if (ret) {
 856                                 kfree(copy_selem);
 857                                 atomic_sub(smap->elem_size,
 858                                            &newsk->sk_omem_alloc);
 859                                 bpf_map_put(map);
 860                                 goto out;
 861                         }
 862
 863                         new_sk_storage = rcu_dereference(copy_selem->sk_storage);
 864                 }
 865                 bpf_map_put(map);
 866         }
 867
 868 out:
 869         rcu_read_unlock();
 870
 871         /* In case of an error, don't free anything explicitly here, the
 872          * caller is responsible to call bpf_sk_storage_free.
 873          */
 874
 875         return ret;
 876 }
 877
 878 BPF_CALL_4(bpf_sk_storage_get, struct bpf_map *, map, struct sock *, sk,
 879            void *, value, u64, flags)
 880 {
 881         struct bpf_sk_storage_data *sdata;
 882
 883         if (flags > BPF_SK_STORAGE_GET_F_CREATE)
 884                 return (unsigned long)NULL;
 885
 886         sdata = sk_storage_lookup(sk, map, true);
 887         if (sdata)
 888                 return (unsigned long)sdata->data;
 889
 890         if (flags == BPF_SK_STORAGE_GET_F_CREATE &&
 891             /* Cannot add new elem to a going away sk.
 892              * Otherwise, the new elem may become a leak
 893              * (and also other memory issues during map
 894              *  destruction).
 895              */
 896             refcount_inc_not_zero(&sk->sk_refcnt)) {
 897                 sdata = sk_storage_update(sk, map, value, BPF_NOEXIST);
 898                 /* sk must be a fullsock (guaranteed by verifier),
 899                  * so sock_gen_put() is unnecessary.
 900                  */
 901                 sock_put(sk);
 902                 return IS_ERR(sdata) ?
 903                         (unsigned long)NULL : (unsigned long)sdata->data;
 904         }
 905
 906         return (unsigned long)NULL;
 907 }
 908
 909 BPF_CALL_2(bpf_sk_storage_delete, struct bpf_map *, map, struct sock *, sk)
 910 {
 911         if (refcount_inc_not_zero(&sk->sk_refcnt)) {
 912                 int err;
 913
 914                 err = sk_storage_delete(sk, map);
 915                 sock_put(sk);
 916                 return err;
 917         }
 918
 919         return -ENOENT;
 920 }
 921
 922 static int sk_storage_map_btf_id;
 923 const struct bpf_map_ops sk_storage_map_ops = {
 924         .map_alloc_check = bpf_sk_storage_map_alloc_check,
 925         .map_alloc = bpf_sk_storage_map_alloc,
 926         .map_free = bpf_sk_storage_map_free,
 927         .map_get_next_key = notsupp_get_next_key,
 928         .map_lookup_elem = bpf_fd_sk_storage_lookup_elem,
 929         .map_update_elem = bpf_fd_sk_storage_update_elem,
 930         .map_delete_elem = bpf_fd_sk_storage_delete_elem,
 931         .map_check_btf = bpf_sk_storage_map_check_btf,
 932         .map_btf_name = "bpf_sk_storage_map",
 933         .map_btf_id = &sk_storage_map_btf_id,
 934 };
 935
 936 const struct bpf_func_proto bpf_sk_storage_get_proto = {
 937         .func           = bpf_sk_storage_get,
 938         .gpl_only       = false,
 939         .ret_type       = RET_PTR_TO_MAP_VALUE_OR_NULL,
 940         .arg1_type      = ARG_CONST_MAP_PTR,
 941         .arg2_type      = ARG_PTR_TO_SOCKET,
 942         .arg3_type      = ARG_PTR_TO_MAP_VALUE_OR_NULL,
 943         .arg4_type      = ARG_ANYTHING,
 944 };
 945
 946 const struct bpf_func_proto bpf_sk_storage_delete_proto = {
 947         .func           = bpf_sk_storage_delete,
 948         .gpl_only       = false,
 949         .ret_type       = RET_INTEGER,
 950         .arg1_type      = ARG_CONST_MAP_PTR,
 951         .arg2_type      = ARG_PTR_TO_SOCKET,
 952 };
 953
 954 struct bpf_sk_storage_diag {
 955         u32 nr_maps;
 956         struct bpf_map *maps[];
 957 };
 958
 959 /* The reply will be like:
 960  * INET_DIAG_BPF_SK_STORAGES (nla_nest)
 961  *      SK_DIAG_BPF_STORAGE (nla_nest)
 962  *              SK_DIAG_BPF_STORAGE_MAP_ID (nla_put_u32)
 963  *              SK_DIAG_BPF_STORAGE_MAP_VALUE (nla_reserve_64bit)
 964  *      SK_DIAG_BPF_STORAGE (nla_nest)
 965  *              SK_DIAG_BPF_STORAGE_MAP_ID (nla_put_u32)
 966  *              SK_DIAG_BPF_STORAGE_MAP_VALUE (nla_reserve_64bit)
 967  *      ....
 968  */
 969 static int nla_value_size(u32 value_size)
 970 {
 971         /* SK_DIAG_BPF_STORAGE (nla_nest)
 972          *      SK_DIAG_BPF_STORAGE_MAP_ID (nla_put_u32)
 973          *      SK_DIAG_BPF_STORAGE_MAP_VALUE (nla_reserve_64bit)
 974          */
 975         return nla_total_size(0) + nla_total_size(sizeof(u32)) +
 976                 nla_total_size_64bit(value_size);
 977 }
 978
 979 void bpf_sk_storage_diag_free(struct bpf_sk_storage_diag *diag)
 980 {
 981         u32 i;
 982
 983         if (!diag)
 984                 return;
 985
 986         for (i = 0; i < diag->nr_maps; i++)
 987                 bpf_map_put(diag->maps[i]);
 988
 989         kfree(diag);
 990 }
 991 EXPORT_SYMBOL_GPL(bpf_sk_storage_diag_free);
 992
 993 static bool diag_check_dup(const struct bpf_sk_storage_diag *diag,
 994                            const struct bpf_map *map)
 995 {
 996         u32 i;
 997
 998         for (i = 0; i < diag->nr_maps; i++) {
 999                 if (diag->maps[i] == map)
1000                         return true;
1001         }
1002
1003         return false;
1004 }
1005
1006 struct bpf_sk_storage_diag *
1007 bpf_sk_storage_diag_alloc(const struct nlattr *nla_stgs)
1008 {
1009         struct bpf_sk_storage_diag *diag;
1010         struct nlattr *nla;
1011         u32 nr_maps = 0;
1012         int rem, err;
1013
1014         /* bpf_sk_storage_map is currently limited to CAP_SYS_ADMIN as
1015          * the map_alloc_check() side also does.
1016          */
1017         if (!bpf_capable())
1018                 return ERR_PTR(-EPERM);
1019
1020         nla_for_each_nested(nla, nla_stgs, rem) {
1021                 if (nla_type(nla) == SK_DIAG_BPF_STORAGE_REQ_MAP_FD)
1022                         nr_maps++;
1023         }
1024
1025         diag = kzalloc(sizeof(*diag) + sizeof(diag->maps[0]) * nr_maps,
1026                        GFP_KERNEL);
1027         if (!diag)
1028                 return ERR_PTR(-ENOMEM);
1029
1030         nla_for_each_nested(nla, nla_stgs, rem) {
1031                 struct bpf_map *map;
1032                 int map_fd;
1033
1034                 if (nla_type(nla) != SK_DIAG_BPF_STORAGE_REQ_MAP_FD)
1035                         continue;
1036
1037                 map_fd = nla_get_u32(nla);
1038                 map = bpf_map_get(map_fd);
1039                 if (IS_ERR(map)) {
1040                         err = PTR_ERR(map);
1041                         goto err_free;
1042                 }
1043                 if (map->map_type != BPF_MAP_TYPE_SK_STORAGE) {
1044                         bpf_map_put(map);
1045                         err = -EINVAL;
1046                         goto err_free;
1047                 }
1048                 if (diag_check_dup(diag, map)) {
1049                         bpf_map_put(map);
1050                         err = -EEXIST;
1051                         goto err_free;
1052                 }
1053                 diag->maps[diag->nr_maps++] = map;
1054         }
1055
1056         return diag;
1057
1058 err_free:
1059         bpf_sk_storage_diag_free(diag);
1060         return ERR_PTR(err);
1061 }
1062 EXPORT_SYMBOL_GPL(bpf_sk_storage_diag_alloc);
1063
1064 static int diag_get(struct bpf_sk_storage_data *sdata, struct sk_buff *skb)
1065 {
1066         struct nlattr *nla_stg, *nla_value;
1067         struct bpf_sk_storage_map *smap;
1068
1069         /* It cannot exceed max nlattr's payload */
1070         BUILD_BUG_ON(U16_MAX - NLA_HDRLEN < MAX_VALUE_SIZE);
1071
1072         nla_stg = nla_nest_start(skb, SK_DIAG_BPF_STORAGE);
1073         if (!nla_stg)
1074                 return -EMSGSIZE;
1075
1076         smap = rcu_dereference(sdata->smap);
1077         if (nla_put_u32(skb, SK_DIAG_BPF_STORAGE_MAP_ID, smap->map.id))
1078                 goto errout;
1079
1080         nla_value = nla_reserve_64bit(skb, SK_DIAG_BPF_STORAGE_MAP_VALUE,
1081                                       smap->map.value_size,
1082                                       SK_DIAG_BPF_STORAGE_PAD);
1083         if (!nla_value)
1084                 goto errout;
1085
1086         if (map_value_has_spin_lock(&smap->map))
1087                 copy_map_value_locked(&smap->map, nla_data(nla_value),
1088                                       sdata->data, true);
1089         else
1090                 copy_map_value(&smap->map, nla_data(nla_value), sdata->data);
1091
1092         nla_nest_end(skb, nla_stg);
1093         return 0;
1094
1095 errout:
1096         nla_nest_cancel(skb, nla_stg);
1097         return -EMSGSIZE;
1098 }
1099
1100 static int bpf_sk_storage_diag_put_all(struct sock *sk, struct sk_buff *skb,
1101                                        int stg_array_type,
1102                                        unsigned int *res_diag_size)
1103 {
1104         /* stg_array_type (e.g. INET_DIAG_BPF_SK_STORAGES) */
1105         unsigned int diag_size = nla_total_size(0);
1106         struct bpf_sk_storage *sk_storage;
1107         struct bpf_sk_storage_elem *selem;
1108         struct bpf_sk_storage_map *smap;
1109         struct nlattr *nla_stgs;
1110         unsigned int saved_len;
1111         int err = 0;
1112
1113         rcu_read_lock();
1114
1115         sk_storage = rcu_dereference(sk->sk_bpf_storage);
1116         if (!sk_storage || hlist_empty(&sk_storage->list)) {
1117                 rcu_read_unlock();
1118                 return 0;
1119         }
1120
1121         nla_stgs = nla_nest_start(skb, stg_array_type);
1122         if (!nla_stgs)
1123                 /* Continue to learn diag_size */
1124                 err = -EMSGSIZE;
1125
1126         saved_len = skb->len;
1127         hlist_for_each_entry_rcu(selem, &sk_storage->list, snode) {
1128                 smap = rcu_dereference(SDATA(selem)->smap);
1129                 diag_size += nla_value_size(smap->map.value_size);
1130
1131                 if (nla_stgs && diag_get(SDATA(selem), skb))
1132                         /* Continue to learn diag_size */
1133                         err = -EMSGSIZE;
1134         }
1135
1136         rcu_read_unlock();
1137
1138         if (nla_stgs) {
1139                 if (saved_len == skb->len)
1140                         nla_nest_cancel(skb, nla_stgs);
1141                 else
1142                         nla_nest_end(skb, nla_stgs);
1143         }
1144
1145         if (diag_size == nla_total_size(0)) {
1146                 *res_diag_size = 0;
1147                 return 0;
1148         }
1149
1150         *res_diag_size = diag_size;
1151         return err;
1152 }
1153
1154 int bpf_sk_storage_diag_put(struct bpf_sk_storage_diag *diag,
1155                             struct sock *sk, struct sk_buff *skb,
1156                             int stg_array_type,
1157                             unsigned int *res_diag_size)
1158 {
1159         /* stg_array_type (e.g. INET_DIAG_BPF_SK_STORAGES) */
1160         unsigned int diag_size = nla_total_size(0);
1161         struct bpf_sk_storage *sk_storage;
1162         struct bpf_sk_storage_data *sdata;
1163         struct nlattr *nla_stgs;
1164         unsigned int saved_len;
1165         int err = 0;
1166         u32 i;
1167
1168         *res_diag_size = 0;
1169
1170         /* No map has been specified.  Dump all. */
1171         if (!diag->nr_maps)
1172                 return bpf_sk_storage_diag_put_all(sk, skb, stg_array_type,
1173                                                    res_diag_size);
1174
1175         rcu_read_lock();
1176         sk_storage = rcu_dereference(sk->sk_bpf_storage);
1177         if (!sk_storage || hlist_empty(&sk_storage->list)) {
1178                 rcu_read_unlock();
1179                 return 0;
1180         }
1181
1182         nla_stgs = nla_nest_start(skb, stg_array_type);
1183         if (!nla_stgs)
1184                 /* Continue to learn diag_size */
1185                 err = -EMSGSIZE;
1186
1187         saved_len = skb->len;
1188         for (i = 0; i < diag->nr_maps; i++) {
1189                 sdata = __sk_storage_lookup(sk_storage,
1190                                 (struct bpf_sk_storage_map *)diag->maps[i],
1191                                 false);
1192
1193                 if (!sdata)
1194                         continue;
1195
1196                 diag_size += nla_value_size(diag->maps[i]->value_size);
1197
1198                 if (nla_stgs && diag_get(sdata, skb))
1199                         /* Continue to learn diag_size */
1200                         err = -EMSGSIZE;
1201         }
1202         rcu_read_unlock();
1203
1204         if (nla_stgs) {
1205                 if (saved_len == skb->len)
1206                         nla_nest_cancel(skb, nla_stgs);
1207                 else
1208                         nla_nest_end(skb, nla_stgs);
1209         }
1210
1211         if (diag_size == nla_total_size(0)) {
1212                 *res_diag_size = 0;
1213                 return 0;
1214         }
1215
1216         *res_diag_size = diag_size;
1217         return err;
1218 }
1219 EXPORT_SYMBOL_GPL(bpf_sk_storage_diag_put);