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         htab->elem_size = sizeof(struct htab_elem) +
 321                           round_up(htab->map.key_size, 8);
 322         if (percpu)
 323                 htab->elem_size += sizeof(void *);
 324         else
 325                 htab->elem_size += round_up(htab->map.value_size, 8);
 326
 327         /* prevent zero size kmalloc and check for u32 overflow */
 328         if (htab->n_buckets == 0 ||
 329             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 330                 goto free_htab;
 331
 332         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 333                (u64) htab->elem_size * htab->map.max_entries;
 334
 335         if (percpu)
 336                 cost += (u64) round_up(htab->map.value_size, 8) *
 337                         num_possible_cpus() * htab->map.max_entries;
 338         else
 339                cost += (u64) htab->elem_size * num_possible_cpus();
 340
 341         if (cost >= U32_MAX - PAGE_SIZE)
 342                 /* make sure page count doesn't overflow */
 343                 goto free_htab;
 344
 345         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 346
 347         /* if map size is larger than memlock limit, reject it early */
 348         err = bpf_map_precharge_memlock(htab->map.pages);
 349         if (err)
 350                 goto free_htab;
 351
 352         err = -ENOMEM;
 353         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 354                                            sizeof(struct bucket),
 355                                            htab->map.numa_node);
 356         if (!htab->buckets)
 357                 goto free_htab;
 358
 359         for (i = 0; i < htab->n_buckets; i++) {
 360                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 361                 raw_spin_lock_init(&htab->buckets[i].lock);
 362         }
 363
 364         if (prealloc) {
 365                 err = prealloc_init(htab);
 366                 if (err)
 367                         goto free_buckets;
 368
 369                 if (!percpu && !lru) {
 370                         /* lru itself can remove the least used element, so
 371                          * there is no need for an extra elem during map_update.
 372                          */
 373                         err = alloc_extra_elems(htab);
 374                         if (err)
 375                                 goto free_prealloc;
 376                 }
 377         }
 378
 379         return &htab->map;
 380
 381 free_prealloc:
 382         prealloc_destroy(htab);
 383 free_buckets:
 384         bpf_map_area_free(htab->buckets);
 385 free_htab:
 386         kfree(htab);
 387         return ERR_PTR(err);
 388 }
 389
 390 static inline u32 htab_map_hash(const void *key, u32 key_len)
 391 {
 392         return jhash(key, key_len, 0);
 393 }
 394
 395 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 396 {
 397         return &htab->buckets[hash & (htab->n_buckets - 1)];
 398 }
 399
 400 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 401 {
 402         return &__select_bucket(htab, hash)->head;
 403 }
 404
 405 /* this lookup function can only be called with bucket lock taken */
 406 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 407                                          void *key, u32 key_size)
 408 {
 409         struct hlist_nulls_node *n;
 410         struct htab_elem *l;
 411
 412         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 413                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 414                         return l;
 415
 416         return NULL;
 417 }
 418
 419 /* can be called without bucket lock. it will repeat the loop in
 420  * the unlikely event when elements moved from one bucket into another
 421  * while link list is being walked
 422  */
 423 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 424                                                u32 hash, void *key,
 425                                                u32 key_size, u32 n_buckets)
 426 {
 427         struct hlist_nulls_node *n;
 428         struct htab_elem *l;
 429
 430 again:
 431         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 432                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 433                         return l;
 434
 435         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 436                 goto again;
 437
 438         return NULL;
 439 }
 440
 441 /* Called from syscall or from eBPF program directly, so
 442  * arguments have to match bpf_map_lookup_elem() exactly.
 443  * The return value is adjusted by BPF instructions
 444  * in htab_map_gen_lookup().
 445  */
 446 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 447 {
 448         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 449         struct hlist_nulls_head *head;
 450         struct htab_elem *l;
 451         u32 hash, key_size;
 452
 453         /* Must be called with rcu_read_lock. */
 454         WARN_ON_ONCE(!rcu_read_lock_held());
 455
 456         key_size = map->key_size;
 457
 458         hash = htab_map_hash(key, key_size);
 459
 460         head = select_bucket(htab, hash);
 461
 462         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 463
 464         return l;
 465 }
 466
 467 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 468 {
 469         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 470
 471         if (l)
 472                 return l->key + round_up(map->key_size, 8);
 473
 474         return NULL;
 475 }
 476
 477 /* inline bpf_map_lookup_elem() call.
 478  * Instead of:
 479  * bpf_prog
 480  *   bpf_map_lookup_elem
 481  *     map->ops->map_lookup_elem
 482  *       htab_map_lookup_elem
 483  *         __htab_map_lookup_elem
 484  * do:
 485  * bpf_prog
 486  *   __htab_map_lookup_elem
 487  */
 488 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 489 {
 490         struct bpf_insn *insn = insn_buf;
 491         const int ret = BPF_REG_0;
 492
 493         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 494         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 495         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 496                                 offsetof(struct htab_elem, key) +
 497                                 round_up(map->key_size, 8));
 498         return insn - insn_buf;
 499 }
 500
 501 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 502 {
 503         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 504
 505         if (l) {
 506                 bpf_lru_node_set_ref(&l->lru_node);
 507                 return l->key + round_up(map->key_size, 8);
 508         }
 509
 510         return NULL;
 511 }
 512
 513 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 514                                    struct bpf_insn *insn_buf)
 515 {
 516         struct bpf_insn *insn = insn_buf;
 517         const int ret = BPF_REG_0;
 518         const int ref_reg = BPF_REG_1;
 519
 520         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 521         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 522         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 523                               offsetof(struct htab_elem, lru_node) +
 524                               offsetof(struct bpf_lru_node, ref));
 525         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 526         *insn++ = BPF_ST_MEM(BPF_B, ret,
 527                              offsetof(struct htab_elem, lru_node) +
 528                              offsetof(struct bpf_lru_node, ref),
 529                              1);
 530         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 531                                 offsetof(struct htab_elem, key) +
 532                                 round_up(map->key_size, 8));
 533         return insn - insn_buf;
 534 }
 535
 536 /* It is called from the bpf_lru_list when the LRU needs to delete
 537  * older elements from the htab.
 538  */
 539 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 540 {
 541         struct bpf_htab *htab = (struct bpf_htab *)arg;
 542         struct htab_elem *l = NULL, *tgt_l;
 543         struct hlist_nulls_head *head;
 544         struct hlist_nulls_node *n;
 545         unsigned long flags;
 546         struct bucket *b;
 547
 548         tgt_l = container_of(node, struct htab_elem, lru_node);
 549         b = __select_bucket(htab, tgt_l->hash);
 550         head = &b->head;
 551
 552         raw_spin_lock_irqsave(&b->lock, flags);
 553
 554         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 555                 if (l == tgt_l) {
 556                         hlist_nulls_del_rcu(&l->hash_node);
 557                         break;
 558                 }
 559
 560         raw_spin_unlock_irqrestore(&b->lock, flags);
 561
 562         return l == tgt_l;
 563 }
 564
 565 /* Called from syscall */
 566 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 567 {
 568         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 569         struct hlist_nulls_head *head;
 570         struct htab_elem *l, *next_l;
 571         u32 hash, key_size;
 572         int i = 0;
 573
 574         WARN_ON_ONCE(!rcu_read_lock_held());
 575
 576         key_size = map->key_size;
 577
 578         if (!key)
 579                 goto find_first_elem;
 580
 581         hash = htab_map_hash(key, key_size);
 582
 583         head = select_bucket(htab, hash);
 584
 585         /* lookup the key */
 586         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 587
 588         if (!l)
 589                 goto find_first_elem;
 590
 591         /* key was found, get next key in the same bucket */
 592         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 593                                   struct htab_elem, hash_node);
 594
 595         if (next_l) {
 596                 /* if next elem in this hash list is non-zero, just return it */
 597                 memcpy(next_key, next_l->key, key_size);
 598                 return 0;
 599         }
 600
 601         /* no more elements in this hash list, go to the next bucket */
 602         i = hash & (htab->n_buckets - 1);
 603         i++;
 604
 605 find_first_elem:
 606         /* iterate over buckets */
 607         for (; i < htab->n_buckets; i++) {
 608                 head = select_bucket(htab, i);
 609
 610                 /* pick first element in the bucket */
 611                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 612                                           struct htab_elem, hash_node);
 613                 if (next_l) {
 614                         /* if it's not empty, just return it */
 615                         memcpy(next_key, next_l->key, key_size);
 616                         return 0;
 617                 }
 618         }
 619
 620         /* iterated over all buckets and all elements */
 621         return -ENOENT;
 622 }
 623
 624 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 625 {
 626         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 627                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 628         kfree(l);
 629 }
 630
 631 static void htab_elem_free_rcu(struct rcu_head *head)
 632 {
 633         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 634         struct bpf_htab *htab = l->htab;
 635
 636         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 637          * we're calling kfree, otherwise deadlock is possible if kprobes
 638          * are placed somewhere inside of slub
 639          */
 640         preempt_disable();
 641         __this_cpu_inc(bpf_prog_active);
 642         htab_elem_free(htab, l);
 643         __this_cpu_dec(bpf_prog_active);
 644         preempt_enable();
 645 }
 646
 647 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 648 {
 649         struct bpf_map *map = &htab->map;
 650
 651         if (map->ops->map_fd_put_ptr) {
 652                 void *ptr = fd_htab_map_get_ptr(map, l);
 653
 654                 map->ops->map_fd_put_ptr(ptr);
 655         }
 656
 657         if (htab_is_prealloc(htab)) {
 658                 pcpu_freelist_push(&htab->freelist, &l->fnode);
 659         } else {
 660                 atomic_dec(&htab->count);
 661                 l->htab = htab;
 662                 call_rcu(&l->rcu, htab_elem_free_rcu);
 663         }
 664 }
 665
 666 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 667                             void *value, bool onallcpus)
 668 {
 669         if (!onallcpus) {
 670                 /* copy true value_size bytes */
 671                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 672         } else {
 673                 u32 size = round_up(htab->map.value_size, 8);
 674                 int off = 0, cpu;
 675
 676                 for_each_possible_cpu(cpu) {
 677                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 678                                         value + off, size);
 679                         off += size;
 680                 }
 681         }
 682 }
 683
 684 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 685 {
 686         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 687                BITS_PER_LONG == 64;
 688 }
 689
 690 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
 691 {
 692         u32 size = htab->map.value_size;
 693
 694         if (percpu || fd_htab_map_needs_adjust(htab))
 695                 size = round_up(size, 8);
 696         return size;
 697 }
 698
 699 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 700                                          void *value, u32 key_size, u32 hash,
 701                                          bool percpu, bool onallcpus,
 702                                          struct htab_elem *old_elem)
 703 {
 704         u32 size = htab_size_value(htab, percpu);
 705         bool prealloc = htab_is_prealloc(htab);
 706         struct htab_elem *l_new, **pl_new;
 707         void __percpu *pptr;
 708
 709         if (prealloc) {
 710                 if (old_elem) {
 711                         /* if we're updating the existing element,
 712                          * use per-cpu extra elems to avoid freelist_pop/push
 713                          */
 714                         pl_new = this_cpu_ptr(htab->extra_elems);
 715                         l_new = *pl_new;
 716                         *pl_new = old_elem;
 717                 } else {
 718                         struct pcpu_freelist_node *l;
 719
 720                         l = pcpu_freelist_pop(&htab->freelist);
 721                         if (!l)
 722                                 return ERR_PTR(-E2BIG);
 723                         l_new = container_of(l, struct htab_elem, fnode);
 724                 }
 725         } else {
 726                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 727                         if (!old_elem) {
 728                                 /* when map is full and update() is replacing
 729                                  * old element, it's ok to allocate, since
 730                                  * old element will be freed immediately.
 731                                  * Otherwise return an error
 732                                  */
 733                                 atomic_dec(&htab->count);
 734                                 return ERR_PTR(-E2BIG);
 735                         }
 736                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 737                                      htab->map.numa_node);
 738                 if (!l_new)
 739                         return ERR_PTR(-ENOMEM);
 740         }
 741
 742         memcpy(l_new->key, key, key_size);
 743         if (percpu) {
 744                 if (prealloc) {
 745                         pptr = htab_elem_get_ptr(l_new, key_size);
 746                 } else {
 747                         /* alloc_percpu zero-fills */
 748                         pptr = __alloc_percpu_gfp(size, 8,
 749                                                   GFP_ATOMIC | __GFP_NOWARN);
 750                         if (!pptr) {
 751                                 kfree(l_new);
 752                                 return ERR_PTR(-ENOMEM);
 753                         }
 754                 }
 755
 756                 pcpu_copy_value(htab, pptr, value, onallcpus);
 757
 758                 if (!prealloc)
 759                         htab_elem_set_ptr(l_new, key_size, pptr);
 760         } else {
 761                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 762         }
 763
 764         l_new->hash = hash;
 765         return l_new;
 766 }
 767
 768 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 769                        u64 map_flags)
 770 {
 771         if (l_old && map_flags == BPF_NOEXIST)
 772                 /* elem already exists */
 773                 return -EEXIST;
 774
 775         if (!l_old && map_flags == BPF_EXIST)
 776                 /* elem doesn't exist, cannot update it */
 777                 return -ENOENT;
 778
 779         return 0;
 780 }
 781
 782 /* Called from syscall or from eBPF program */
 783 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 784                                 u64 map_flags)
 785 {
 786         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 787         struct htab_elem *l_new = NULL, *l_old;
 788         struct hlist_nulls_head *head;
 789         unsigned long flags;
 790         struct bucket *b;
 791         u32 key_size, hash;
 792         int ret;
 793
 794         if (unlikely(map_flags > BPF_EXIST))
 795                 /* unknown flags */
 796                 return -EINVAL;
 797
 798         WARN_ON_ONCE(!rcu_read_lock_held());
 799
 800         key_size = map->key_size;
 801
 802         hash = htab_map_hash(key, key_size);
 803
 804         b = __select_bucket(htab, hash);
 805         head = &b->head;
 806
 807         /* bpf_map_update_elem() can be called in_irq() */
 808         raw_spin_lock_irqsave(&b->lock, flags);
 809
 810         l_old = lookup_elem_raw(head, hash, key, key_size);
 811
 812         ret = check_flags(htab, l_old, map_flags);
 813         if (ret)
 814                 goto err;
 815
 816         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 817                                 l_old);
 818         if (IS_ERR(l_new)) {
 819                 /* all pre-allocated elements are in use or memory exhausted */
 820                 ret = PTR_ERR(l_new);
 821                 goto err;
 822         }
 823
 824         /* add new element to the head of the list, so that
 825          * concurrent search will find it before old elem
 826          */
 827         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 828         if (l_old) {
 829                 hlist_nulls_del_rcu(&l_old->hash_node);
 830                 if (!htab_is_prealloc(htab))
 831                         free_htab_elem(htab, l_old);
 832         }
 833         ret = 0;
 834 err:
 835         raw_spin_unlock_irqrestore(&b->lock, flags);
 836         return ret;
 837 }
 838
 839 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 840                                     u64 map_flags)
 841 {
 842         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 843         struct htab_elem *l_new, *l_old = NULL;
 844         struct hlist_nulls_head *head;
 845         unsigned long flags;
 846         struct bucket *b;
 847         u32 key_size, hash;
 848         int ret;
 849
 850         if (unlikely(map_flags > BPF_EXIST))
 851                 /* unknown flags */
 852                 return -EINVAL;
 853
 854         WARN_ON_ONCE(!rcu_read_lock_held());
 855
 856         key_size = map->key_size;
 857
 858         hash = htab_map_hash(key, key_size);
 859
 860         b = __select_bucket(htab, hash);
 861         head = &b->head;
 862
 863         /* For LRU, we need to alloc before taking bucket's
 864          * spinlock because getting free nodes from LRU may need
 865          * to remove older elements from htab and this removal
 866          * operation will need a bucket lock.
 867          */
 868         l_new = prealloc_lru_pop(htab, key, hash);
 869         if (!l_new)
 870                 return -ENOMEM;
 871         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 872
 873         /* bpf_map_update_elem() can be called in_irq() */
 874         raw_spin_lock_irqsave(&b->lock, flags);
 875
 876         l_old = lookup_elem_raw(head, hash, key, key_size);
 877
 878         ret = check_flags(htab, l_old, map_flags);
 879         if (ret)
 880                 goto err;
 881
 882         /* add new element to the head of the list, so that
 883          * concurrent search will find it before old elem
 884          */
 885         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 886         if (l_old) {
 887                 bpf_lru_node_set_ref(&l_new->lru_node);
 888                 hlist_nulls_del_rcu(&l_old->hash_node);
 889         }
 890         ret = 0;
 891
 892 err:
 893         raw_spin_unlock_irqrestore(&b->lock, flags);
 894
 895         if (ret)
 896                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 897         else if (l_old)
 898                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 899
 900         return ret;
 901 }
 902
 903 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 904                                          void *value, u64 map_flags,
 905                                          bool onallcpus)
 906 {
 907         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 908         struct htab_elem *l_new = NULL, *l_old;
 909         struct hlist_nulls_head *head;
 910         unsigned long flags;
 911         struct bucket *b;
 912         u32 key_size, hash;
 913         int ret;
 914
 915         if (unlikely(map_flags > BPF_EXIST))
 916                 /* unknown flags */
 917                 return -EINVAL;
 918
 919         WARN_ON_ONCE(!rcu_read_lock_held());
 920
 921         key_size = map->key_size;
 922
 923         hash = htab_map_hash(key, key_size);
 924
 925         b = __select_bucket(htab, hash);
 926         head = &b->head;
 927
 928         /* bpf_map_update_elem() can be called in_irq() */
 929         raw_spin_lock_irqsave(&b->lock, flags);
 930
 931         l_old = lookup_elem_raw(head, hash, key, key_size);
 932
 933         ret = check_flags(htab, l_old, map_flags);
 934         if (ret)
 935                 goto err;
 936
 937         if (l_old) {
 938                 /* per-cpu hash map can update value in-place */
 939                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 940                                 value, onallcpus);
 941         } else {
 942                 l_new = alloc_htab_elem(htab, key, value, key_size,
 943                                         hash, true, onallcpus, NULL);
 944                 if (IS_ERR(l_new)) {
 945                         ret = PTR_ERR(l_new);
 946                         goto err;
 947                 }
 948                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 949         }
 950         ret = 0;
 951 err:
 952         raw_spin_unlock_irqrestore(&b->lock, flags);
 953         return ret;
 954 }
 955
 956 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 957                                              void *value, u64 map_flags,
 958                                              bool onallcpus)
 959 {
 960         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 961         struct htab_elem *l_new = NULL, *l_old;
 962         struct hlist_nulls_head *head;
 963         unsigned long flags;
 964         struct bucket *b;
 965         u32 key_size, hash;
 966         int ret;
 967
 968         if (unlikely(map_flags > BPF_EXIST))
 969                 /* unknown flags */
 970                 return -EINVAL;
 971
 972         WARN_ON_ONCE(!rcu_read_lock_held());
 973
 974         key_size = map->key_size;
 975
 976         hash = htab_map_hash(key, key_size);
 977
 978         b = __select_bucket(htab, hash);
 979         head = &b->head;
 980
 981         /* For LRU, we need to alloc before taking bucket's
 982          * spinlock because LRU's elem alloc may need
 983          * to remove older elem from htab and this removal
 984          * operation will need a bucket lock.
 985          */
 986         if (map_flags != BPF_EXIST) {
 987                 l_new = prealloc_lru_pop(htab, key, hash);
 988                 if (!l_new)
 989                         return -ENOMEM;
 990         }
 991
 992         /* bpf_map_update_elem() can be called in_irq() */
 993         raw_spin_lock_irqsave(&b->lock, flags);
 994
 995         l_old = lookup_elem_raw(head, hash, key, key_size);
 996
 997         ret = check_flags(htab, l_old, map_flags);
 998         if (ret)
 999                 goto err;
1000
1001         if (l_old) {
1002                 bpf_lru_node_set_ref(&l_old->lru_node);
1003
1004                 /* per-cpu hash map can update value in-place */
1005                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1006                                 value, onallcpus);
1007         } else {
1008                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1009                                 value, onallcpus);
1010                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1011                 l_new = NULL;
1012         }
1013         ret = 0;
1014 err:
1015         raw_spin_unlock_irqrestore(&b->lock, flags);
1016         if (l_new)
1017                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1018         return ret;
1019 }
1020
1021 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1022                                        void *value, u64 map_flags)
1023 {
1024         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1025 }
1026
1027 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1028                                            void *value, u64 map_flags)
1029 {
1030         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1031                                                  false);
1032 }
1033
1034 /* Called from syscall or from eBPF program */
1035 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1036 {
1037         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1038         struct hlist_nulls_head *head;
1039         struct bucket *b;
1040         struct htab_elem *l;
1041         unsigned long flags;
1042         u32 hash, key_size;
1043         int ret = -ENOENT;
1044
1045         WARN_ON_ONCE(!rcu_read_lock_held());
1046
1047         key_size = map->key_size;
1048
1049         hash = htab_map_hash(key, key_size);
1050         b = __select_bucket(htab, hash);
1051         head = &b->head;
1052
1053         raw_spin_lock_irqsave(&b->lock, flags);
1054
1055         l = lookup_elem_raw(head, hash, key, key_size);
1056
1057         if (l) {
1058                 hlist_nulls_del_rcu(&l->hash_node);
1059                 free_htab_elem(htab, l);
1060                 ret = 0;
1061         }
1062
1063         raw_spin_unlock_irqrestore(&b->lock, flags);
1064         return ret;
1065 }
1066
1067 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1068 {
1069         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1070         struct hlist_nulls_head *head;
1071         struct bucket *b;
1072         struct htab_elem *l;
1073         unsigned long flags;
1074         u32 hash, key_size;
1075         int ret = -ENOENT;
1076
1077         WARN_ON_ONCE(!rcu_read_lock_held());
1078
1079         key_size = map->key_size;
1080
1081         hash = htab_map_hash(key, key_size);
1082         b = __select_bucket(htab, hash);
1083         head = &b->head;
1084
1085         raw_spin_lock_irqsave(&b->lock, flags);
1086
1087         l = lookup_elem_raw(head, hash, key, key_size);
1088
1089         if (l) {
1090                 hlist_nulls_del_rcu(&l->hash_node);
1091                 ret = 0;
1092         }
1093
1094         raw_spin_unlock_irqrestore(&b->lock, flags);
1095         if (l)
1096                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1097         return ret;
1098 }
1099
1100 static void delete_all_elements(struct bpf_htab *htab)
1101 {
1102         int i;
1103
1104         for (i = 0; i < htab->n_buckets; i++) {
1105                 struct hlist_nulls_head *head = select_bucket(htab, i);
1106                 struct hlist_nulls_node *n;
1107                 struct htab_elem *l;
1108
1109                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1110                         hlist_nulls_del_rcu(&l->hash_node);
1111                         htab_elem_free(htab, l);
1112                 }
1113         }
1114 }
1115
1116 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1117 static void htab_map_free(struct bpf_map *map)
1118 {
1119         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1120
1121         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1122          * so the programs (can be more than one that used this map) were
1123          * disconnected from events. Wait for outstanding critical sections in
1124          * these programs to complete
1125          */
1126         synchronize_rcu();
1127
1128         /* some of free_htab_elem() callbacks for elements of this map may
1129          * not have executed. Wait for them.
1130          */
1131         rcu_barrier();
1132         if (!htab_is_prealloc(htab))
1133                 delete_all_elements(htab);
1134         else
1135                 prealloc_destroy(htab);
1136
1137         free_percpu(htab->extra_elems);
1138         bpf_map_area_free(htab->buckets);
1139         kfree(htab);
1140 }
1141
1142 const struct bpf_map_ops htab_map_ops = {
1143         .map_alloc = htab_map_alloc,
1144         .map_free = htab_map_free,
1145         .map_get_next_key = htab_map_get_next_key,
1146         .map_lookup_elem = htab_map_lookup_elem,
1147         .map_update_elem = htab_map_update_elem,
1148         .map_delete_elem = htab_map_delete_elem,
1149         .map_gen_lookup = htab_map_gen_lookup,
1150 };
1151
1152 const struct bpf_map_ops htab_lru_map_ops = {
1153         .map_alloc = htab_map_alloc,
1154         .map_free = htab_map_free,
1155         .map_get_next_key = htab_map_get_next_key,
1156         .map_lookup_elem = htab_lru_map_lookup_elem,
1157         .map_update_elem = htab_lru_map_update_elem,
1158         .map_delete_elem = htab_lru_map_delete_elem,
1159         .map_gen_lookup = htab_lru_map_gen_lookup,
1160 };
1161
1162 /* Called from eBPF program */
1163 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1164 {
1165         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1166
1167         if (l)
1168                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1169         else
1170                 return NULL;
1171 }
1172
1173 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1174 {
1175         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1176
1177         if (l) {
1178                 bpf_lru_node_set_ref(&l->lru_node);
1179                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1180         }
1181
1182         return NULL;
1183 }
1184
1185 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1186 {
1187         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1188         struct htab_elem *l;
1189         void __percpu *pptr;
1190         int ret = -ENOENT;
1191         int cpu, off = 0;
1192         u32 size;
1193
1194         /* per_cpu areas are zero-filled and bpf programs can only
1195          * access 'value_size' of them, so copying rounded areas
1196          * will not leak any kernel data
1197          */
1198         size = round_up(map->value_size, 8);
1199         rcu_read_lock();
1200         l = __htab_map_lookup_elem(map, key);
1201         if (!l)
1202                 goto out;
1203         if (htab_is_lru(htab))
1204                 bpf_lru_node_set_ref(&l->lru_node);
1205         pptr = htab_elem_get_ptr(l, map->key_size);
1206         for_each_possible_cpu(cpu) {
1207                 bpf_long_memcpy(value + off,
1208                                 per_cpu_ptr(pptr, cpu), size);
1209                 off += size;
1210         }
1211         ret = 0;
1212 out:
1213         rcu_read_unlock();
1214         return ret;
1215 }
1216
1217 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1218                            u64 map_flags)
1219 {
1220         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1221         int ret;
1222
1223         rcu_read_lock();
1224         if (htab_is_lru(htab))
1225                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1226                                                         map_flags, true);
1227         else
1228                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1229                                                     true);
1230         rcu_read_unlock();
1231
1232         return ret;
1233 }
1234
1235 const struct bpf_map_ops htab_percpu_map_ops = {
1236         .map_alloc = htab_map_alloc,
1237         .map_free = htab_map_free,
1238         .map_get_next_key = htab_map_get_next_key,
1239         .map_lookup_elem = htab_percpu_map_lookup_elem,
1240         .map_update_elem = htab_percpu_map_update_elem,
1241         .map_delete_elem = htab_map_delete_elem,
1242 };
1243
1244 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1245         .map_alloc = htab_map_alloc,
1246         .map_free = htab_map_free,
1247         .map_get_next_key = htab_map_get_next_key,
1248         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1249         .map_update_elem = htab_lru_percpu_map_update_elem,
1250         .map_delete_elem = htab_lru_map_delete_elem,
1251 };
1252
1253 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1254 {
1255         if (attr->value_size != sizeof(u32))
1256                 return ERR_PTR(-EINVAL);
1257         return htab_map_alloc(attr);
1258 }
1259
1260 static void fd_htab_map_free(struct bpf_map *map)
1261 {
1262         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1263         struct hlist_nulls_node *n;
1264         struct hlist_nulls_head *head;
1265         struct htab_elem *l;
1266         int i;
1267
1268         for (i = 0; i < htab->n_buckets; i++) {
1269                 head = select_bucket(htab, i);
1270
1271                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1272                         void *ptr = fd_htab_map_get_ptr(map, l);
1273
1274                         map->ops->map_fd_put_ptr(ptr);
1275                 }
1276         }
1277
1278         htab_map_free(map);
1279 }
1280
1281 /* only called from syscall */
1282 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1283 {
1284         void **ptr;
1285         int ret = 0;
1286
1287         if (!map->ops->map_fd_sys_lookup_elem)
1288                 return -ENOTSUPP;
1289
1290         rcu_read_lock();
1291         ptr = htab_map_lookup_elem(map, key);
1292         if (ptr)
1293                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1294         else
1295                 ret = -ENOENT;
1296         rcu_read_unlock();
1297
1298         return ret;
1299 }
1300
1301 /* only called from syscall */
1302 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1303                                 void *key, void *value, u64 map_flags)
1304 {
1305         void *ptr;
1306         int ret;
1307         u32 ufd = *(u32 *)value;
1308
1309         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1310         if (IS_ERR(ptr))
1311                 return PTR_ERR(ptr);
1312
1313         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1314         if (ret)
1315                 map->ops->map_fd_put_ptr(ptr);
1316
1317         return ret;
1318 }
1319
1320 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1321 {
1322         struct bpf_map *map, *inner_map_meta;
1323
1324         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1325         if (IS_ERR(inner_map_meta))
1326                 return inner_map_meta;
1327
1328         map = fd_htab_map_alloc(attr);
1329         if (IS_ERR(map)) {
1330                 bpf_map_meta_free(inner_map_meta);
1331                 return map;
1332         }
1333
1334         map->inner_map_meta = inner_map_meta;
1335
1336         return map;
1337 }
1338
1339 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1340 {
1341         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1342
1343         if (!inner_map)
1344                 return NULL;
1345
1346         return READ_ONCE(*inner_map);
1347 }
1348
1349 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1350                                   struct bpf_insn *insn_buf)
1351 {
1352         struct bpf_insn *insn = insn_buf;
1353         const int ret = BPF_REG_0;
1354
1355         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1356         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1357         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1358                                 offsetof(struct htab_elem, key) +
1359                                 round_up(map->key_size, 8));
1360         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1361
1362         return insn - insn_buf;
1363 }
1364
1365 static void htab_of_map_free(struct bpf_map *map)
1366 {
1367         bpf_map_meta_free(map->inner_map_meta);
1368         fd_htab_map_free(map);
1369 }
1370
1371 const struct bpf_map_ops htab_of_maps_map_ops = {
1372         .map_alloc = htab_of_map_alloc,
1373         .map_free = htab_of_map_free,
1374         .map_get_next_key = htab_map_get_next_key,
1375         .map_lookup_elem = htab_of_map_lookup_elem,
1376         .map_delete_elem = htab_map_delete_elem,
1377         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1378         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1379         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1380         .map_gen_lookup = htab_of_map_gen_lookup,
1381 };