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