kernel/bpf/hashtab.c

   1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   2  * Copyright (c) 2016 Facebook
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of version 2 of the GNU General Public
   6  * License as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but
   9  * WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  11  * General Public License for more details.
  12  */
  13 #include <linux/bpf.h>
  14 #include <linux/jhash.h>
  15 #include <linux/filter.h>
  16 #include <linux/rculist_nulls.h>
  17 #include "percpu_freelist.h"
  18 #include "bpf_lru_list.h"
  19 #include "map_in_map.h"
  20
  21 #define HTAB_CREATE_FLAG_MASK \
  22         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE)
  23
  24 struct bucket {
  25         struct hlist_nulls_head head;
  26         raw_spinlock_t lock;
  27 };
  28
  29 struct bpf_htab {
  30         struct bpf_map map;
  31         struct bucket *buckets;
  32         void *elems;
  33         union {
  34                 struct pcpu_freelist freelist;
  35                 struct bpf_lru lru;
  36         };
  37         struct htab_elem *__percpu *extra_elems;
  38         atomic_t count; /* number of elements in this hashtable */
  39         u32 n_buckets;  /* number of hash buckets */
  40         u32 elem_size;  /* size of each element in bytes */
  41 };
  42
  43 /* each htab element is struct htab_elem + key + value */
  44 struct htab_elem {
  45         union {
  46                 struct hlist_nulls_node hash_node;
  47                 struct {
  48                         void *padding;
  49                         union {
  50                                 struct bpf_htab *htab;
  51                                 struct pcpu_freelist_node fnode;
  52                         };
  53                 };
  54         };
  55         union {
  56                 struct rcu_head rcu;
  57                 struct bpf_lru_node lru_node;
  58         };
  59         u32 hash;
  60         char key[0] __aligned(8);
  61 };
  62
  63 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  64
  65 static bool htab_is_lru(const struct bpf_htab *htab)
  66 {
  67         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  68                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  69 }
  70
  71 static bool htab_is_percpu(const struct bpf_htab *htab)
  72 {
  73         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  74                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  75 }
  76
  77 static bool htab_is_prealloc(const struct bpf_htab *htab)
  78 {
  79         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  80 }
  81
  82 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  83                                      void __percpu *pptr)
  84 {
  85         *(void __percpu **)(l->key + key_size) = pptr;
  86 }
  87
  88 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  89 {
  90         return *(void __percpu **)(l->key + key_size);
  91 }
  92
  93 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
  94 {
  95         return *(void **)(l->key + roundup(map->key_size, 8));
  96 }
  97
  98 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
  99 {
 100         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 101 }
 102
 103 static void htab_free_elems(struct bpf_htab *htab)
 104 {
 105         int i;
 106
 107         if (!htab_is_percpu(htab))
 108                 goto free_elems;
 109
 110         for (i = 0; i < htab->map.max_entries; i++) {
 111                 void __percpu *pptr;
 112
 113                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 114                                          htab->map.key_size);
 115                 free_percpu(pptr);
 116         }
 117 free_elems:
 118         bpf_map_area_free(htab->elems);
 119 }
 120
 121 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 122                                           u32 hash)
 123 {
 124         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 125         struct htab_elem *l;
 126
 127         if (node) {
 128                 l = container_of(node, struct htab_elem, lru_node);
 129                 memcpy(l->key, key, htab->map.key_size);
 130                 return l;
 131         }
 132
 133         return NULL;
 134 }
 135
 136 static int prealloc_init(struct bpf_htab *htab)
 137 {
 138         u32 num_entries = htab->map.max_entries;
 139         int err = -ENOMEM, i;
 140
 141         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 142                 num_entries += num_possible_cpus();
 143
 144         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 145                                          htab->map.numa_node);
 146         if (!htab->elems)
 147                 return -ENOMEM;
 148
 149         if (!htab_is_percpu(htab))
 150                 goto skip_percpu_elems;
 151
 152         for (i = 0; i < num_entries; i++) {
 153                 u32 size = round_up(htab->map.value_size, 8);
 154                 void __percpu *pptr;
 155
 156                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 157                 if (!pptr)
 158                         goto free_elems;
 159                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 160                                   pptr);
 161         }
 162
 163 skip_percpu_elems:
 164         if (htab_is_lru(htab))
 165                 err = bpf_lru_init(&htab->lru,
 166                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 167                                    offsetof(struct htab_elem, hash) -
 168                                    offsetof(struct htab_elem, lru_node),
 169                                    htab_lru_map_delete_node,
 170                                    htab);
 171         else
 172                 err = pcpu_freelist_init(&htab->freelist);
 173
 174         if (err)
 175                 goto free_elems;
 176
 177         if (htab_is_lru(htab))
 178                 bpf_lru_populate(&htab->lru, htab->elems,
 179                                  offsetof(struct htab_elem, lru_node),
 180                                  htab->elem_size, num_entries);
 181         else
 182                 pcpu_freelist_populate(&htab->freelist,
 183                                        htab->elems + offsetof(struct htab_elem, fnode),
 184                                        htab->elem_size, num_entries);
 185
 186         return 0;
 187
 188 free_elems:
 189         htab_free_elems(htab);
 190         return err;
 191 }
 192
 193 static void prealloc_destroy(struct bpf_htab *htab)
 194 {
 195         htab_free_elems(htab);
 196
 197         if (htab_is_lru(htab))
 198                 bpf_lru_destroy(&htab->lru);
 199         else
 200                 pcpu_freelist_destroy(&htab->freelist);
 201 }
 202
 203 static int alloc_extra_elems(struct bpf_htab *htab)
 204 {
 205         struct htab_elem *__percpu *pptr, *l_new;
 206         struct pcpu_freelist_node *l;
 207         int cpu;
 208
 209         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 210                                   GFP_USER | __GFP_NOWARN);
 211         if (!pptr)
 212                 return -ENOMEM;
 213
 214         for_each_possible_cpu(cpu) {
 215                 l = pcpu_freelist_pop(&htab->freelist);
 216                 /* pop will succeed, since prealloc_init()
 217                  * preallocated extra num_possible_cpus elements
 218                  */
 219                 l_new = container_of(l, struct htab_elem, fnode);
 220                 *per_cpu_ptr(pptr, cpu) = l_new;
 221         }
 222         htab->extra_elems = pptr;
 223         return 0;
 224 }
 225
 226 /* Called from syscall */
 227 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 228 {
 229         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 230                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 231         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 232                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 233         /* percpu_lru means each cpu has its own LRU list.
 234          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 235          * the map's value itself is percpu.  percpu_lru has
 236          * nothing to do with the map's value.
 237          */
 238         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 239         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 240         int numa_node = bpf_map_attr_numa_node(attr);
 241         struct bpf_htab *htab;
 242         int err, i;
 243         u64 cost;
 244
 245         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 246                      offsetof(struct htab_elem, hash_node.pprev));
 247         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 248                      offsetof(struct htab_elem, hash_node.pprev));
 249
 250         if (lru && !capable(CAP_SYS_ADMIN))
 251                 /* LRU implementation is much complicated than other
 252                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 253                  */
 254                 return ERR_PTR(-EPERM);
 255
 256         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
 257                 /* reserved bits should not be used */
 258                 return ERR_PTR(-EINVAL);
 259
 260         if (!lru && percpu_lru)
 261                 return ERR_PTR(-EINVAL);
 262
 263         if (lru && !prealloc)
 264                 return ERR_PTR(-ENOTSUPP);
 265
 266         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 267                 return ERR_PTR(-EINVAL);
 268
 269         htab = kzalloc(sizeof(*htab), GFP_USER);
 270         if (!htab)
 271                 return ERR_PTR(-ENOMEM);
 272
 273         /* mandatory map attributes */
 274         htab->map.map_type = attr->map_type;
 275         htab->map.key_size = attr->key_size;
 276         htab->map.value_size = attr->value_size;
 277         htab->map.max_entries = attr->max_entries;
 278         htab->map.map_flags = attr->map_flags;
 279         htab->map.numa_node = numa_node;
 280
 281         /* check sanity of attributes.
 282          * value_size == 0 may be allowed in the future to use map as a set
 283          */
 284         err = -EINVAL;
 285         if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
 286             htab->map.value_size == 0)
 287                 goto free_htab;
 288
 289         if (percpu_lru) {
 290                 /* ensure each CPU's lru list has >=1 elements.
 291                  * since we are at it, make each lru list has the same
 292                  * number of elements.
 293                  */
 294                 htab->map.max_entries = roundup(attr->max_entries,
 295                                                 num_possible_cpus());
 296                 if (htab->map.max_entries < attr->max_entries)
 297                         htab->map.max_entries = rounddown(attr->max_entries,
 298                                                           num_possible_cpus());
 299         }
 300
 301         /* hash table size must be power of 2 */
 302         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 303
 304         err = -E2BIG;
 305         if (htab->map.key_size > MAX_BPF_STACK)
 306                 /* eBPF programs initialize keys on stack, so they cannot be
 307                  * larger than max stack size
 308                  */
 309                 goto free_htab;
 310
 311         if (htab->map.value_size >= KMALLOC_MAX_SIZE -
 312             MAX_BPF_STACK - sizeof(struct htab_elem))
 313                 /* if value_size is bigger, the user space won't be able to
 314                  * access the elements via bpf syscall. This check also makes
 315                  * sure that the elem_size doesn't overflow and it's
 316                  * kmalloc-able later in htab_map_update_elem()
 317                  */
 318                 goto free_htab;
 319
 320         if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
 321                 /* make sure the size for pcpu_alloc() is reasonable */
 322                 goto free_htab;
 323
 324         htab->elem_size = sizeof(struct htab_elem) +
 325                           round_up(htab->map.key_size, 8);
 326         if (percpu)
 327                 htab->elem_size += sizeof(void *);
 328         else
 329                 htab->elem_size += round_up(htab->map.value_size, 8);
 330
 331         /* prevent zero size kmalloc and check for u32 overflow */
 332         if (htab->n_buckets == 0 ||
 333             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 334                 goto free_htab;
 335
 336         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 337                (u64) htab->elem_size * htab->map.max_entries;
 338
 339         if (percpu)
 340                 cost += (u64) round_up(htab->map.value_size, 8) *
 341                         num_possible_cpus() * htab->map.max_entries;
 342         else
 343                cost += (u64) htab->elem_size * num_possible_cpus();
 344
 345         if (cost >= U32_MAX - PAGE_SIZE)
 346                 /* make sure page count doesn't overflow */
 347                 goto free_htab;
 348
 349         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 350
 351         /* if map size is larger than memlock limit, reject it early */
 352         err = bpf_map_precharge_memlock(htab->map.pages);
 353         if (err)
 354                 goto free_htab;
 355
 356         err = -ENOMEM;
 357         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 358                                            sizeof(struct bucket),
 359                                            htab->map.numa_node);
 360         if (!htab->buckets)
 361                 goto free_htab;
 362
 363         for (i = 0; i < htab->n_buckets; i++) {
 364                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 365                 raw_spin_lock_init(&htab->buckets[i].lock);
 366         }
 367
 368         if (prealloc) {
 369                 err = prealloc_init(htab);
 370                 if (err)
 371                         goto free_buckets;
 372
 373                 if (!percpu && !lru) {
 374                         /* lru itself can remove the least used element, so
 375                          * there is no need for an extra elem during map_update.
 376                          */
 377                         err = alloc_extra_elems(htab);
 378                         if (err)
 379                                 goto free_prealloc;
 380                 }
 381         }
 382
 383         return &htab->map;
 384
 385 free_prealloc:
 386         prealloc_destroy(htab);
 387 free_buckets:
 388         bpf_map_area_free(htab->buckets);
 389 free_htab:
 390         kfree(htab);
 391         return ERR_PTR(err);
 392 }
 393
 394 static inline u32 htab_map_hash(const void *key, u32 key_len)
 395 {
 396         return jhash(key, key_len, 0);
 397 }
 398
 399 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 400 {
 401         return &htab->buckets[hash & (htab->n_buckets - 1)];
 402 }
 403
 404 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 405 {
 406         return &__select_bucket(htab, hash)->head;
 407 }
 408
 409 /* this lookup function can only be called with bucket lock taken */
 410 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 411                                          void *key, u32 key_size)
 412 {
 413         struct hlist_nulls_node *n;
 414         struct htab_elem *l;
 415
 416         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 417                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 418                         return l;
 419
 420         return NULL;
 421 }
 422
 423 /* can be called without bucket lock. it will repeat the loop in
 424  * the unlikely event when elements moved from one bucket into another
 425  * while link list is being walked
 426  */
 427 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 428                                                u32 hash, void *key,
 429                                                u32 key_size, u32 n_buckets)
 430 {
 431         struct hlist_nulls_node *n;
 432         struct htab_elem *l;
 433
 434 again:
 435         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 436                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 437                         return l;
 438
 439         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 440                 goto again;
 441
 442         return NULL;
 443 }
 444
 445 /* Called from syscall or from eBPF program directly, so
 446  * arguments have to match bpf_map_lookup_elem() exactly.
 447  * The return value is adjusted by BPF instructions
 448  * in htab_map_gen_lookup().
 449  */
 450 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 451 {
 452         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 453         struct hlist_nulls_head *head;
 454         struct htab_elem *l;
 455         u32 hash, key_size;
 456
 457         /* Must be called with rcu_read_lock. */
 458         WARN_ON_ONCE(!rcu_read_lock_held());
 459
 460         key_size = map->key_size;
 461
 462         hash = htab_map_hash(key, key_size);
 463
 464         head = select_bucket(htab, hash);
 465
 466         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 467
 468         return l;
 469 }
 470
 471 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 472 {
 473         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 474
 475         if (l)
 476                 return l->key + round_up(map->key_size, 8);
 477
 478         return NULL;
 479 }
 480
 481 /* inline bpf_map_lookup_elem() call.
 482  * Instead of:
 483  * bpf_prog
 484  *   bpf_map_lookup_elem
 485  *     map->ops->map_lookup_elem
 486  *       htab_map_lookup_elem
 487  *         __htab_map_lookup_elem
 488  * do:
 489  * bpf_prog
 490  *   __htab_map_lookup_elem
 491  */
 492 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 493 {
 494         struct bpf_insn *insn = insn_buf;
 495         const int ret = BPF_REG_0;
 496
 497         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 498         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 499         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 500                                 offsetof(struct htab_elem, key) +
 501                                 round_up(map->key_size, 8));
 502         return insn - insn_buf;
 503 }
 504
 505 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 506 {
 507         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 508
 509         if (l) {
 510                 bpf_lru_node_set_ref(&l->lru_node);
 511                 return l->key + round_up(map->key_size, 8);
 512         }
 513
 514         return NULL;
 515 }
 516
 517 /* It is called from the bpf_lru_list when the LRU needs to delete
 518  * older elements from the htab.
 519  */
 520 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 521 {
 522         struct bpf_htab *htab = (struct bpf_htab *)arg;
 523         struct htab_elem *l = NULL, *tgt_l;
 524         struct hlist_nulls_head *head;
 525         struct hlist_nulls_node *n;
 526         unsigned long flags;
 527         struct bucket *b;
 528
 529         tgt_l = container_of(node, struct htab_elem, lru_node);
 530         b = __select_bucket(htab, tgt_l->hash);
 531         head = &b->head;
 532
 533         raw_spin_lock_irqsave(&b->lock, flags);
 534
 535         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 536                 if (l == tgt_l) {
 537                         hlist_nulls_del_rcu(&l->hash_node);
 538                         break;
 539                 }
 540
 541         raw_spin_unlock_irqrestore(&b->lock, flags);
 542
 543         return l == tgt_l;
 544 }
 545
 546 /* Called from syscall */
 547 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 548 {
 549         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 550         struct hlist_nulls_head *head;
 551         struct htab_elem *l, *next_l;
 552         u32 hash, key_size;
 553         int i = 0;
 554
 555         WARN_ON_ONCE(!rcu_read_lock_held());
 556
 557         key_size = map->key_size;
 558
 559         if (!key)
 560                 goto find_first_elem;
 561
 562         hash = htab_map_hash(key, key_size);
 563
 564         head = select_bucket(htab, hash);
 565
 566         /* lookup the key */
 567         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 568
 569         if (!l)
 570                 goto find_first_elem;
 571
 572         /* key was found, get next key in the same bucket */
 573         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 574                                   struct htab_elem, hash_node);
 575
 576         if (next_l) {
 577                 /* if next elem in this hash list is non-zero, just return it */
 578                 memcpy(next_key, next_l->key, key_size);
 579                 return 0;
 580         }
 581
 582         /* no more elements in this hash list, go to the next bucket */
 583         i = hash & (htab->n_buckets - 1);
 584         i++;
 585
 586 find_first_elem:
 587         /* iterate over buckets */
 588         for (; i < htab->n_buckets; i++) {
 589                 head = select_bucket(htab, i);
 590
 591                 /* pick first element in the bucket */
 592                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 593                                           struct htab_elem, hash_node);
 594                 if (next_l) {
 595                         /* if it's not empty, just return it */
 596                         memcpy(next_key, next_l->key, key_size);
 597                         return 0;
 598                 }
 599         }
 600
 601         /* iterated over all buckets and all elements */
 602         return -ENOENT;
 603 }
 604
 605 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 606 {
 607         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 608                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 609         kfree(l);
 610 }
 611
 612 static void htab_elem_free_rcu(struct rcu_head *head)
 613 {
 614         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 615         struct bpf_htab *htab = l->htab;
 616
 617         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 618          * we're calling kfree, otherwise deadlock is possible if kprobes
 619          * are placed somewhere inside of slub
 620          */
 621         preempt_disable();
 622         __this_cpu_inc(bpf_prog_active);
 623         htab_elem_free(htab, l);
 624         __this_cpu_dec(bpf_prog_active);
 625         preempt_enable();
 626 }
 627
 628 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 629 {
 630         struct bpf_map *map = &htab->map;
 631
 632         if (map->ops->map_fd_put_ptr) {
 633                 void *ptr = fd_htab_map_get_ptr(map, l);
 634
 635                 map->ops->map_fd_put_ptr(ptr);
 636         }
 637
 638         if (htab_is_prealloc(htab)) {
 639                 pcpu_freelist_push(&htab->freelist, &l->fnode);
 640         } else {
 641                 atomic_dec(&htab->count);
 642                 l->htab = htab;
 643                 call_rcu(&l->rcu, htab_elem_free_rcu);
 644         }
 645 }
 646
 647 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 648                             void *value, bool onallcpus)
 649 {
 650         if (!onallcpus) {
 651                 /* copy true value_size bytes */
 652                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 653         } else {
 654                 u32 size = round_up(htab->map.value_size, 8);
 655                 int off = 0, cpu;
 656
 657                 for_each_possible_cpu(cpu) {
 658                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 659                                         value + off, size);
 660                         off += size;
 661                 }
 662         }
 663 }
 664
 665 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 666                                          void *value, u32 key_size, u32 hash,
 667                                          bool percpu, bool onallcpus,
 668                                          struct htab_elem *old_elem)
 669 {
 670         u32 size = htab->map.value_size;
 671         bool prealloc = htab_is_prealloc(htab);
 672         struct htab_elem *l_new, **pl_new;
 673         void __percpu *pptr;
 674
 675         if (prealloc) {
 676                 if (old_elem) {
 677                         /* if we're updating the existing element,
 678                          * use per-cpu extra elems to avoid freelist_pop/push
 679                          */
 680                         pl_new = this_cpu_ptr(htab->extra_elems);
 681                         l_new = *pl_new;
 682                         *pl_new = old_elem;
 683                 } else {
 684                         struct pcpu_freelist_node *l;
 685
 686                         l = pcpu_freelist_pop(&htab->freelist);
 687                         if (!l)
 688                                 return ERR_PTR(-E2BIG);
 689                         l_new = container_of(l, struct htab_elem, fnode);
 690                 }
 691         } else {
 692                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 693                         if (!old_elem) {
 694                                 /* when map is full and update() is replacing
 695                                  * old element, it's ok to allocate, since
 696                                  * old element will be freed immediately.
 697                                  * Otherwise return an error
 698                                  */
 699                                 atomic_dec(&htab->count);
 700                                 return ERR_PTR(-E2BIG);
 701                         }
 702                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 703                                      htab->map.numa_node);
 704                 if (!l_new)
 705                         return ERR_PTR(-ENOMEM);
 706         }
 707
 708         memcpy(l_new->key, key, key_size);
 709         if (percpu) {
 710                 /* round up value_size to 8 bytes */
 711                 size = round_up(size, 8);
 712
 713                 if (prealloc) {
 714                         pptr = htab_elem_get_ptr(l_new, key_size);
 715                 } else {
 716                         /* alloc_percpu zero-fills */
 717                         pptr = __alloc_percpu_gfp(size, 8,
 718                                                   GFP_ATOMIC | __GFP_NOWARN);
 719                         if (!pptr) {
 720                                 kfree(l_new);
 721                                 return ERR_PTR(-ENOMEM);
 722                         }
 723                 }
 724
 725                 pcpu_copy_value(htab, pptr, value, onallcpus);
 726
 727                 if (!prealloc)
 728                         htab_elem_set_ptr(l_new, key_size, pptr);
 729         } else {
 730                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 731         }
 732
 733         l_new->hash = hash;
 734         return l_new;
 735 }
 736
 737 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 738                        u64 map_flags)
 739 {
 740         if (l_old && map_flags == BPF_NOEXIST)
 741                 /* elem already exists */
 742                 return -EEXIST;
 743
 744         if (!l_old && map_flags == BPF_EXIST)
 745                 /* elem doesn't exist, cannot update it */
 746                 return -ENOENT;
 747
 748         return 0;
 749 }
 750
 751 /* Called from syscall or from eBPF program */
 752 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 753                                 u64 map_flags)
 754 {
 755         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 756         struct htab_elem *l_new = NULL, *l_old;
 757         struct hlist_nulls_head *head;
 758         unsigned long flags;
 759         struct bucket *b;
 760         u32 key_size, hash;
 761         int ret;
 762
 763         if (unlikely(map_flags > BPF_EXIST))
 764                 /* unknown flags */
 765                 return -EINVAL;
 766
 767         WARN_ON_ONCE(!rcu_read_lock_held());
 768
 769         key_size = map->key_size;
 770
 771         hash = htab_map_hash(key, key_size);
 772
 773         b = __select_bucket(htab, hash);
 774         head = &b->head;
 775
 776         /* bpf_map_update_elem() can be called in_irq() */
 777         raw_spin_lock_irqsave(&b->lock, flags);
 778
 779         l_old = lookup_elem_raw(head, hash, key, key_size);
 780
 781         ret = check_flags(htab, l_old, map_flags);
 782         if (ret)
 783                 goto err;
 784
 785         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 786                                 l_old);
 787         if (IS_ERR(l_new)) {
 788                 /* all pre-allocated elements are in use or memory exhausted */
 789                 ret = PTR_ERR(l_new);
 790                 goto err;
 791         }
 792
 793         /* add new element to the head of the list, so that
 794          * concurrent search will find it before old elem
 795          */
 796         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 797         if (l_old) {
 798                 hlist_nulls_del_rcu(&l_old->hash_node);
 799                 if (!htab_is_prealloc(htab))
 800                         free_htab_elem(htab, l_old);
 801         }
 802         ret = 0;
 803 err:
 804         raw_spin_unlock_irqrestore(&b->lock, flags);
 805         return ret;
 806 }
 807
 808 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 809                                     u64 map_flags)
 810 {
 811         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 812         struct htab_elem *l_new, *l_old = NULL;
 813         struct hlist_nulls_head *head;
 814         unsigned long flags;
 815         struct bucket *b;
 816         u32 key_size, hash;
 817         int ret;
 818
 819         if (unlikely(map_flags > BPF_EXIST))
 820                 /* unknown flags */
 821                 return -EINVAL;
 822
 823         WARN_ON_ONCE(!rcu_read_lock_held());
 824
 825         key_size = map->key_size;
 826
 827         hash = htab_map_hash(key, key_size);
 828
 829         b = __select_bucket(htab, hash);
 830         head = &b->head;
 831
 832         /* For LRU, we need to alloc before taking bucket's
 833          * spinlock because getting free nodes from LRU may need
 834          * to remove older elements from htab and this removal
 835          * operation will need a bucket lock.
 836          */
 837         l_new = prealloc_lru_pop(htab, key, hash);
 838         if (!l_new)
 839                 return -ENOMEM;
 840         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 841
 842         /* bpf_map_update_elem() can be called in_irq() */
 843         raw_spin_lock_irqsave(&b->lock, flags);
 844
 845         l_old = lookup_elem_raw(head, hash, key, key_size);
 846
 847         ret = check_flags(htab, l_old, map_flags);
 848         if (ret)
 849                 goto err;
 850
 851         /* add new element to the head of the list, so that
 852          * concurrent search will find it before old elem
 853          */
 854         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 855         if (l_old) {
 856                 bpf_lru_node_set_ref(&l_new->lru_node);
 857                 hlist_nulls_del_rcu(&l_old->hash_node);
 858         }
 859         ret = 0;
 860
 861 err:
 862         raw_spin_unlock_irqrestore(&b->lock, flags);
 863
 864         if (ret)
 865                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 866         else if (l_old)
 867                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 868
 869         return ret;
 870 }
 871
 872 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 873                                          void *value, u64 map_flags,
 874                                          bool onallcpus)
 875 {
 876         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 877         struct htab_elem *l_new = NULL, *l_old;
 878         struct hlist_nulls_head *head;
 879         unsigned long flags;
 880         struct bucket *b;
 881         u32 key_size, hash;
 882         int ret;
 883
 884         if (unlikely(map_flags > BPF_EXIST))
 885                 /* unknown flags */
 886                 return -EINVAL;
 887
 888         WARN_ON_ONCE(!rcu_read_lock_held());
 889
 890         key_size = map->key_size;
 891
 892         hash = htab_map_hash(key, key_size);
 893
 894         b = __select_bucket(htab, hash);
 895         head = &b->head;
 896
 897         /* bpf_map_update_elem() can be called in_irq() */
 898         raw_spin_lock_irqsave(&b->lock, flags);
 899
 900         l_old = lookup_elem_raw(head, hash, key, key_size);
 901
 902         ret = check_flags(htab, l_old, map_flags);
 903         if (ret)
 904                 goto err;
 905
 906         if (l_old) {
 907                 /* per-cpu hash map can update value in-place */
 908                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 909                                 value, onallcpus);
 910         } else {
 911                 l_new = alloc_htab_elem(htab, key, value, key_size,
 912                                         hash, true, onallcpus, NULL);
 913                 if (IS_ERR(l_new)) {
 914                         ret = PTR_ERR(l_new);
 915                         goto err;
 916                 }
 917                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 918         }
 919         ret = 0;
 920 err:
 921         raw_spin_unlock_irqrestore(&b->lock, flags);
 922         return ret;
 923 }
 924
 925 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 926                                              void *value, u64 map_flags,
 927                                              bool onallcpus)
 928 {
 929         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 930         struct htab_elem *l_new = NULL, *l_old;
 931         struct hlist_nulls_head *head;
 932         unsigned long flags;
 933         struct bucket *b;
 934         u32 key_size, hash;
 935         int ret;
 936
 937         if (unlikely(map_flags > BPF_EXIST))
 938                 /* unknown flags */
 939                 return -EINVAL;
 940
 941         WARN_ON_ONCE(!rcu_read_lock_held());
 942
 943         key_size = map->key_size;
 944
 945         hash = htab_map_hash(key, key_size);
 946
 947         b = __select_bucket(htab, hash);
 948         head = &b->head;
 949
 950         /* For LRU, we need to alloc before taking bucket's
 951          * spinlock because LRU's elem alloc may need
 952          * to remove older elem from htab and this removal
 953          * operation will need a bucket lock.
 954          */
 955         if (map_flags != BPF_EXIST) {
 956                 l_new = prealloc_lru_pop(htab, key, hash);
 957                 if (!l_new)
 958                         return -ENOMEM;
 959         }
 960
 961         /* bpf_map_update_elem() can be called in_irq() */
 962         raw_spin_lock_irqsave(&b->lock, flags);
 963
 964         l_old = lookup_elem_raw(head, hash, key, key_size);
 965
 966         ret = check_flags(htab, l_old, map_flags);
 967         if (ret)
 968                 goto err;
 969
 970         if (l_old) {
 971                 bpf_lru_node_set_ref(&l_old->lru_node);
 972
 973                 /* per-cpu hash map can update value in-place */
 974                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 975                                 value, onallcpus);
 976         } else {
 977                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
 978                                 value, onallcpus);
 979                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 980                 l_new = NULL;
 981         }
 982         ret = 0;
 983 err:
 984         raw_spin_unlock_irqrestore(&b->lock, flags);
 985         if (l_new)
 986                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 987         return ret;
 988 }
 989
 990 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 991                                        void *value, u64 map_flags)
 992 {
 993         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
 994 }
 995
 996 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 997                                            void *value, u64 map_flags)
 998 {
 999         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1000                                                  false);
1001 }
1002
1003 /* Called from syscall or from eBPF program */
1004 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1005 {
1006         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1007         struct hlist_nulls_head *head;
1008         struct bucket *b;
1009         struct htab_elem *l;
1010         unsigned long flags;
1011         u32 hash, key_size;
1012         int ret = -ENOENT;
1013
1014         WARN_ON_ONCE(!rcu_read_lock_held());
1015
1016         key_size = map->key_size;
1017
1018         hash = htab_map_hash(key, key_size);
1019         b = __select_bucket(htab, hash);
1020         head = &b->head;
1021
1022         raw_spin_lock_irqsave(&b->lock, flags);
1023
1024         l = lookup_elem_raw(head, hash, key, key_size);
1025
1026         if (l) {
1027                 hlist_nulls_del_rcu(&l->hash_node);
1028                 free_htab_elem(htab, l);
1029                 ret = 0;
1030         }
1031
1032         raw_spin_unlock_irqrestore(&b->lock, flags);
1033         return ret;
1034 }
1035
1036 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1037 {
1038         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1039         struct hlist_nulls_head *head;
1040         struct bucket *b;
1041         struct htab_elem *l;
1042         unsigned long flags;
1043         u32 hash, key_size;
1044         int ret = -ENOENT;
1045
1046         WARN_ON_ONCE(!rcu_read_lock_held());
1047
1048         key_size = map->key_size;
1049
1050         hash = htab_map_hash(key, key_size);
1051         b = __select_bucket(htab, hash);
1052         head = &b->head;
1053
1054         raw_spin_lock_irqsave(&b->lock, flags);
1055
1056         l = lookup_elem_raw(head, hash, key, key_size);
1057
1058         if (l) {
1059                 hlist_nulls_del_rcu(&l->hash_node);
1060                 ret = 0;
1061         }
1062
1063         raw_spin_unlock_irqrestore(&b->lock, flags);
1064         if (l)
1065                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1066         return ret;
1067 }
1068
1069 static void delete_all_elements(struct bpf_htab *htab)
1070 {
1071         int i;
1072
1073         for (i = 0; i < htab->n_buckets; i++) {
1074                 struct hlist_nulls_head *head = select_bucket(htab, i);
1075                 struct hlist_nulls_node *n;
1076                 struct htab_elem *l;
1077
1078                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1079                         hlist_nulls_del_rcu(&l->hash_node);
1080                         htab_elem_free(htab, l);
1081                 }
1082         }
1083 }
1084
1085 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1086 static void htab_map_free(struct bpf_map *map)
1087 {
1088         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1089
1090         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1091          * so the programs (can be more than one that used this map) were
1092          * disconnected from events. Wait for outstanding critical sections in
1093          * these programs to complete
1094          */
1095         synchronize_rcu();
1096
1097         /* some of free_htab_elem() callbacks for elements of this map may
1098          * not have executed. Wait for them.
1099          */
1100         rcu_barrier();
1101         if (!htab_is_prealloc(htab))
1102                 delete_all_elements(htab);
1103         else
1104                 prealloc_destroy(htab);
1105
1106         free_percpu(htab->extra_elems);
1107         bpf_map_area_free(htab->buckets);
1108         kfree(htab);
1109 }
1110
1111 const struct bpf_map_ops htab_map_ops = {
1112         .map_alloc = htab_map_alloc,
1113         .map_free = htab_map_free,
1114         .map_get_next_key = htab_map_get_next_key,
1115         .map_lookup_elem = htab_map_lookup_elem,
1116         .map_update_elem = htab_map_update_elem,
1117         .map_delete_elem = htab_map_delete_elem,
1118         .map_gen_lookup = htab_map_gen_lookup,
1119 };
1120
1121 const struct bpf_map_ops htab_lru_map_ops = {
1122         .map_alloc = htab_map_alloc,
1123         .map_free = htab_map_free,
1124         .map_get_next_key = htab_map_get_next_key,
1125         .map_lookup_elem = htab_lru_map_lookup_elem,
1126         .map_update_elem = htab_lru_map_update_elem,
1127         .map_delete_elem = htab_lru_map_delete_elem,
1128 };
1129
1130 /* Called from eBPF program */
1131 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1132 {
1133         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1134
1135         if (l)
1136                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1137         else
1138                 return NULL;
1139 }
1140
1141 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1142 {
1143         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1144
1145         if (l) {
1146                 bpf_lru_node_set_ref(&l->lru_node);
1147                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1148         }
1149
1150         return NULL;
1151 }
1152
1153 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1154 {
1155         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1156         struct htab_elem *l;
1157         void __percpu *pptr;
1158         int ret = -ENOENT;
1159         int cpu, off = 0;
1160         u32 size;
1161
1162         /* per_cpu areas are zero-filled and bpf programs can only
1163          * access 'value_size' of them, so copying rounded areas
1164          * will not leak any kernel data
1165          */
1166         size = round_up(map->value_size, 8);
1167         rcu_read_lock();
1168         l = __htab_map_lookup_elem(map, key);
1169         if (!l)
1170                 goto out;
1171         if (htab_is_lru(htab))
1172                 bpf_lru_node_set_ref(&l->lru_node);
1173         pptr = htab_elem_get_ptr(l, map->key_size);
1174         for_each_possible_cpu(cpu) {
1175                 bpf_long_memcpy(value + off,
1176                                 per_cpu_ptr(pptr, cpu), size);
1177                 off += size;
1178         }
1179         ret = 0;
1180 out:
1181         rcu_read_unlock();
1182         return ret;
1183 }
1184
1185 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1186                            u64 map_flags)
1187 {
1188         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1189         int ret;
1190
1191         rcu_read_lock();
1192         if (htab_is_lru(htab))
1193                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1194                                                         map_flags, true);
1195         else
1196                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1197                                                     true);
1198         rcu_read_unlock();
1199
1200         return ret;
1201 }
1202
1203 const struct bpf_map_ops htab_percpu_map_ops = {
1204         .map_alloc = htab_map_alloc,
1205         .map_free = htab_map_free,
1206         .map_get_next_key = htab_map_get_next_key,
1207         .map_lookup_elem = htab_percpu_map_lookup_elem,
1208         .map_update_elem = htab_percpu_map_update_elem,
1209         .map_delete_elem = htab_map_delete_elem,
1210 };
1211
1212 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1213         .map_alloc = htab_map_alloc,
1214         .map_free = htab_map_free,
1215         .map_get_next_key = htab_map_get_next_key,
1216         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1217         .map_update_elem = htab_lru_percpu_map_update_elem,
1218         .map_delete_elem = htab_lru_map_delete_elem,
1219 };
1220
1221 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1222 {
1223         struct bpf_map *map;
1224
1225         if (attr->value_size != sizeof(u32))
1226                 return ERR_PTR(-EINVAL);
1227
1228         /* pointer is stored internally */
1229         attr->value_size = sizeof(void *);
1230         map = htab_map_alloc(attr);
1231         attr->value_size = sizeof(u32);
1232
1233         return map;
1234 }
1235
1236 static void fd_htab_map_free(struct bpf_map *map)
1237 {
1238         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1239         struct hlist_nulls_node *n;
1240         struct hlist_nulls_head *head;
1241         struct htab_elem *l;
1242         int i;
1243
1244         for (i = 0; i < htab->n_buckets; i++) {
1245                 head = select_bucket(htab, i);
1246
1247                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1248                         void *ptr = fd_htab_map_get_ptr(map, l);
1249
1250                         map->ops->map_fd_put_ptr(ptr);
1251                 }
1252         }
1253
1254         htab_map_free(map);
1255 }
1256
1257 /* only called from syscall */
1258 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1259 {
1260         void **ptr;
1261         int ret = 0;
1262
1263         if (!map->ops->map_fd_sys_lookup_elem)
1264                 return -ENOTSUPP;
1265
1266         rcu_read_lock();
1267         ptr = htab_map_lookup_elem(map, key);
1268         if (ptr)
1269                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1270         else
1271                 ret = -ENOENT;
1272         rcu_read_unlock();
1273
1274         return ret;
1275 }
1276
1277 /* only called from syscall */
1278 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1279                                 void *key, void *value, u64 map_flags)
1280 {
1281         void *ptr;
1282         int ret;
1283         u32 ufd = *(u32 *)value;
1284
1285         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1286         if (IS_ERR(ptr))
1287                 return PTR_ERR(ptr);
1288
1289         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1290         if (ret)
1291                 map->ops->map_fd_put_ptr(ptr);
1292
1293         return ret;
1294 }
1295
1296 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1297 {
1298         struct bpf_map *map, *inner_map_meta;
1299
1300         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1301         if (IS_ERR(inner_map_meta))
1302                 return inner_map_meta;
1303
1304         map = fd_htab_map_alloc(attr);
1305         if (IS_ERR(map)) {
1306                 bpf_map_meta_free(inner_map_meta);
1307                 return map;
1308         }
1309
1310         map->inner_map_meta = inner_map_meta;
1311
1312         return map;
1313 }
1314
1315 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1316 {
1317         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1318
1319         if (!inner_map)
1320                 return NULL;
1321
1322         return READ_ONCE(*inner_map);
1323 }
1324
1325 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1326                                   struct bpf_insn *insn_buf)
1327 {
1328         struct bpf_insn *insn = insn_buf;
1329         const int ret = BPF_REG_0;
1330
1331         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1332         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1333         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1334                                 offsetof(struct htab_elem, key) +
1335                                 round_up(map->key_size, 8));
1336         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1337
1338         return insn - insn_buf;
1339 }
1340
1341 static void htab_of_map_free(struct bpf_map *map)
1342 {
1343         bpf_map_meta_free(map->inner_map_meta);
1344         fd_htab_map_free(map);
1345 }
1346
1347 const struct bpf_map_ops htab_of_maps_map_ops = {
1348         .map_alloc = htab_of_map_alloc,
1349         .map_free = htab_of_map_free,
1350         .map_get_next_key = htab_map_get_next_key,
1351         .map_lookup_elem = htab_of_map_lookup_elem,
1352         .map_delete_elem = htab_map_delete_elem,
1353         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1354         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1355         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1356         .map_gen_lookup = htab_of_map_gen_lookup,
1357 };