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          BPF_F_RDONLY | BPF_F_WRONLY)
  24
  25 struct bucket {
  26         struct hlist_nulls_head head;
  27         raw_spinlock_t lock;
  28 };
  29
  30 struct bpf_htab {
  31         struct bpf_map map;
  32         struct bucket *buckets;
  33         void *elems;
  34         union {
  35                 struct pcpu_freelist freelist;
  36                 struct bpf_lru lru;
  37         };
  38         struct htab_elem *__percpu *extra_elems;
  39         atomic_t count; /* number of elements in this hashtable */
  40         u32 n_buckets;  /* number of hash buckets */
  41         u32 elem_size;  /* size of each element in bytes */
  42 };
  43
  44 /* each htab element is struct htab_elem + key + value */
  45 struct htab_elem {
  46         union {
  47                 struct hlist_nulls_node hash_node;
  48                 struct {
  49                         void *padding;
  50                         union {
  51                                 struct bpf_htab *htab;
  52                                 struct pcpu_freelist_node fnode;
  53                         };
  54                 };
  55         };
  56         union {
  57                 struct rcu_head rcu;
  58                 struct bpf_lru_node lru_node;
  59         };
  60         u32 hash;
  61         char key[0] __aligned(8);
  62 };
  63
  64 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  65
  66 static bool htab_is_lru(const struct bpf_htab *htab)
  67 {
  68         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  69                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  70 }
  71
  72 static bool htab_is_percpu(const struct bpf_htab *htab)
  73 {
  74         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  75                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  76 }
  77
  78 static bool htab_is_prealloc(const struct bpf_htab *htab)
  79 {
  80         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  81 }
  82
  83 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  84                                      void __percpu *pptr)
  85 {
  86         *(void __percpu **)(l->key + key_size) = pptr;
  87 }
  88
  89 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  90 {
  91         return *(void __percpu **)(l->key + key_size);
  92 }
  93
  94 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
  95 {
  96         return *(void **)(l->key + roundup(map->key_size, 8));
  97 }
  98
  99 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 100 {
 101         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 102 }
 103
 104 static void htab_free_elems(struct bpf_htab *htab)
 105 {
 106         int i;
 107
 108         if (!htab_is_percpu(htab))
 109                 goto free_elems;
 110
 111         for (i = 0; i < htab->map.max_entries; i++) {
 112                 void __percpu *pptr;
 113
 114                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 115                                          htab->map.key_size);
 116                 free_percpu(pptr);
 117         }
 118 free_elems:
 119         bpf_map_area_free(htab->elems);
 120 }
 121
 122 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 123                                           u32 hash)
 124 {
 125         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 126         struct htab_elem *l;
 127
 128         if (node) {
 129                 l = container_of(node, struct htab_elem, lru_node);
 130                 memcpy(l->key, key, htab->map.key_size);
 131                 return l;
 132         }
 133
 134         return NULL;
 135 }
 136
 137 static int prealloc_init(struct bpf_htab *htab)
 138 {
 139         u32 num_entries = htab->map.max_entries;
 140         int err = -ENOMEM, i;
 141
 142         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 143                 num_entries += num_possible_cpus();
 144
 145         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 146                                          htab->map.numa_node);
 147         if (!htab->elems)
 148                 return -ENOMEM;
 149
 150         if (!htab_is_percpu(htab))
 151                 goto skip_percpu_elems;
 152
 153         for (i = 0; i < num_entries; i++) {
 154                 u32 size = round_up(htab->map.value_size, 8);
 155                 void __percpu *pptr;
 156
 157                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 158                 if (!pptr)
 159                         goto free_elems;
 160                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 161                                   pptr);
 162         }
 163
 164 skip_percpu_elems:
 165         if (htab_is_lru(htab))
 166                 err = bpf_lru_init(&htab->lru,
 167                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 168                                    offsetof(struct htab_elem, hash) -
 169                                    offsetof(struct htab_elem, lru_node),
 170                                    htab_lru_map_delete_node,
 171                                    htab);
 172         else
 173                 err = pcpu_freelist_init(&htab->freelist);
 174
 175         if (err)
 176                 goto free_elems;
 177
 178         if (htab_is_lru(htab))
 179                 bpf_lru_populate(&htab->lru, htab->elems,
 180                                  offsetof(struct htab_elem, lru_node),
 181                                  htab->elem_size, num_entries);
 182         else
 183                 pcpu_freelist_populate(&htab->freelist,
 184                                        htab->elems + offsetof(struct htab_elem, fnode),
 185                                        htab->elem_size, num_entries);
 186
 187         return 0;
 188
 189 free_elems:
 190         htab_free_elems(htab);
 191         return err;
 192 }
 193
 194 static void prealloc_destroy(struct bpf_htab *htab)
 195 {
 196         htab_free_elems(htab);
 197
 198         if (htab_is_lru(htab))
 199                 bpf_lru_destroy(&htab->lru);
 200         else
 201                 pcpu_freelist_destroy(&htab->freelist);
 202 }
 203
 204 static int alloc_extra_elems(struct bpf_htab *htab)
 205 {
 206         struct htab_elem *__percpu *pptr, *l_new;
 207         struct pcpu_freelist_node *l;
 208         int cpu;
 209
 210         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 211                                   GFP_USER | __GFP_NOWARN);
 212         if (!pptr)
 213                 return -ENOMEM;
 214
 215         for_each_possible_cpu(cpu) {
 216                 l = pcpu_freelist_pop(&htab->freelist);
 217                 /* pop will succeed, since prealloc_init()
 218                  * preallocated extra num_possible_cpus elements
 219                  */
 220                 l_new = container_of(l, struct htab_elem, fnode);
 221                 *per_cpu_ptr(pptr, cpu) = l_new;
 222         }
 223         htab->extra_elems = pptr;
 224         return 0;
 225 }
 226
 227 /* Called from syscall */
 228 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 229 {
 230         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 231                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 232         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 233                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 234         /* percpu_lru means each cpu has its own LRU list.
 235          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 236          * the map's value itself is percpu.  percpu_lru has
 237          * nothing to do with the map's value.
 238          */
 239         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 240         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 241         int numa_node = bpf_map_attr_numa_node(attr);
 242         struct bpf_htab *htab;
 243         int err, i;
 244         u64 cost;
 245
 246         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 247                      offsetof(struct htab_elem, hash_node.pprev));
 248         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 249                      offsetof(struct htab_elem, hash_node.pprev));
 250
 251         if (lru && !capable(CAP_SYS_ADMIN))
 252                 /* LRU implementation is much complicated than other
 253                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 254                  */
 255                 return ERR_PTR(-EPERM);
 256
 257         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
 258                 /* reserved bits should not be used */
 259                 return ERR_PTR(-EINVAL);
 260
 261         if (!lru && percpu_lru)
 262                 return ERR_PTR(-EINVAL);
 263
 264         if (lru && !prealloc)
 265                 return ERR_PTR(-ENOTSUPP);
 266
 267         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 268                 return ERR_PTR(-EINVAL);
 269
 270         htab = kzalloc(sizeof(*htab), GFP_USER);
 271         if (!htab)
 272                 return ERR_PTR(-ENOMEM);
 273
 274         /* mandatory map attributes */
 275         htab->map.map_type = attr->map_type;
 276         htab->map.key_size = attr->key_size;
 277         htab->map.value_size = attr->value_size;
 278         htab->map.max_entries = attr->max_entries;
 279         htab->map.map_flags = attr->map_flags;
 280         htab->map.numa_node = numa_node;
 281
 282         /* check sanity of attributes.
 283          * value_size == 0 may be allowed in the future to use map as a set
 284          */
 285         err = -EINVAL;
 286         if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
 287             htab->map.value_size == 0)
 288                 goto free_htab;
 289
 290         if (percpu_lru) {
 291                 /* ensure each CPU's lru list has >=1 elements.
 292                  * since we are at it, make each lru list has the same
 293                  * number of elements.
 294                  */
 295                 htab->map.max_entries = roundup(attr->max_entries,
 296                                                 num_possible_cpus());
 297                 if (htab->map.max_entries < attr->max_entries)
 298                         htab->map.max_entries = rounddown(attr->max_entries,
 299                                                           num_possible_cpus());
 300         }
 301
 302         /* hash table size must be power of 2 */
 303         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 304
 305         err = -E2BIG;
 306         if (htab->map.key_size > MAX_BPF_STACK)
 307                 /* eBPF programs initialize keys on stack, so they cannot be
 308                  * larger than max stack size
 309                  */
 310                 goto free_htab;
 311
 312         if (htab->map.value_size >= KMALLOC_MAX_SIZE -
 313             MAX_BPF_STACK - sizeof(struct htab_elem))
 314                 /* if value_size is bigger, the user space won't be able to
 315                  * access the elements via bpf syscall. This check also makes
 316                  * sure that the elem_size doesn't overflow and it's
 317                  * kmalloc-able later in htab_map_update_elem()
 318                  */
 319                 goto free_htab;
 320
 321         htab->elem_size = sizeof(struct htab_elem) +
 322                           round_up(htab->map.key_size, 8);
 323         if (percpu)
 324                 htab->elem_size += sizeof(void *);
 325         else
 326                 htab->elem_size += round_up(htab->map.value_size, 8);
 327
 328         /* prevent zero size kmalloc and check for u32 overflow */
 329         if (htab->n_buckets == 0 ||
 330             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 331                 goto free_htab;
 332
 333         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 334                (u64) htab->elem_size * htab->map.max_entries;
 335
 336         if (percpu)
 337                 cost += (u64) round_up(htab->map.value_size, 8) *
 338                         num_possible_cpus() * htab->map.max_entries;
 339         else
 340                cost += (u64) htab->elem_size * num_possible_cpus();
 341
 342         if (cost >= U32_MAX - PAGE_SIZE)
 343                 /* make sure page count doesn't overflow */
 344                 goto free_htab;
 345
 346         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 347
 348         /* if map size is larger than memlock limit, reject it early */
 349         err = bpf_map_precharge_memlock(htab->map.pages);
 350         if (err)
 351                 goto free_htab;
 352
 353         err = -ENOMEM;
 354         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 355                                            sizeof(struct bucket),
 356                                            htab->map.numa_node);
 357         if (!htab->buckets)
 358                 goto free_htab;
 359
 360         for (i = 0; i < htab->n_buckets; i++) {
 361                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 362                 raw_spin_lock_init(&htab->buckets[i].lock);
 363         }
 364
 365         if (prealloc) {
 366                 err = prealloc_init(htab);
 367                 if (err)
 368                         goto free_buckets;
 369
 370                 if (!percpu && !lru) {
 371                         /* lru itself can remove the least used element, so
 372                          * there is no need for an extra elem during map_update.
 373                          */
 374                         err = alloc_extra_elems(htab);
 375                         if (err)
 376                                 goto free_prealloc;
 377                 }
 378         }
 379
 380         return &htab->map;
 381
 382 free_prealloc:
 383         prealloc_destroy(htab);
 384 free_buckets:
 385         bpf_map_area_free(htab->buckets);
 386 free_htab:
 387         kfree(htab);
 388         return ERR_PTR(err);
 389 }
 390
 391 static inline u32 htab_map_hash(const void *key, u32 key_len)
 392 {
 393         return jhash(key, key_len, 0);
 394 }
 395
 396 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 397 {
 398         return &htab->buckets[hash & (htab->n_buckets - 1)];
 399 }
 400
 401 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 402 {
 403         return &__select_bucket(htab, hash)->head;
 404 }
 405
 406 /* this lookup function can only be called with bucket lock taken */
 407 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 408                                          void *key, u32 key_size)
 409 {
 410         struct hlist_nulls_node *n;
 411         struct htab_elem *l;
 412
 413         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 414                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 415                         return l;
 416
 417         return NULL;
 418 }
 419
 420 /* can be called without bucket lock. it will repeat the loop in
 421  * the unlikely event when elements moved from one bucket into another
 422  * while link list is being walked
 423  */
 424 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 425                                                u32 hash, void *key,
 426                                                u32 key_size, u32 n_buckets)
 427 {
 428         struct hlist_nulls_node *n;
 429         struct htab_elem *l;
 430
 431 again:
 432         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 433                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 434                         return l;
 435
 436         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 437                 goto again;
 438
 439         return NULL;
 440 }
 441
 442 /* Called from syscall or from eBPF program directly, so
 443  * arguments have to match bpf_map_lookup_elem() exactly.
 444  * The return value is adjusted by BPF instructions
 445  * in htab_map_gen_lookup().
 446  */
 447 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 448 {
 449         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 450         struct hlist_nulls_head *head;
 451         struct htab_elem *l;
 452         u32 hash, key_size;
 453
 454         /* Must be called with rcu_read_lock. */
 455         WARN_ON_ONCE(!rcu_read_lock_held());
 456
 457         key_size = map->key_size;
 458
 459         hash = htab_map_hash(key, key_size);
 460
 461         head = select_bucket(htab, hash);
 462
 463         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 464
 465         return l;
 466 }
 467
 468 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 469 {
 470         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 471
 472         if (l)
 473                 return l->key + round_up(map->key_size, 8);
 474
 475         return NULL;
 476 }
 477
 478 /* inline bpf_map_lookup_elem() call.
 479  * Instead of:
 480  * bpf_prog
 481  *   bpf_map_lookup_elem
 482  *     map->ops->map_lookup_elem
 483  *       htab_map_lookup_elem
 484  *         __htab_map_lookup_elem
 485  * do:
 486  * bpf_prog
 487  *   __htab_map_lookup_elem
 488  */
 489 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 490 {
 491         struct bpf_insn *insn = insn_buf;
 492         const int ret = BPF_REG_0;
 493
 494         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 495         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 496         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 497                                 offsetof(struct htab_elem, key) +
 498                                 round_up(map->key_size, 8));
 499         return insn - insn_buf;
 500 }
 501
 502 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 503 {
 504         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 505
 506         if (l) {
 507                 bpf_lru_node_set_ref(&l->lru_node);
 508                 return l->key + round_up(map->key_size, 8);
 509         }
 510
 511         return NULL;
 512 }
 513
 514 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 515                                    struct bpf_insn *insn_buf)
 516 {
 517         struct bpf_insn *insn = insn_buf;
 518         const int ret = BPF_REG_0;
 519         const int ref_reg = BPF_REG_1;
 520
 521         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 522         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 523         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 524                               offsetof(struct htab_elem, lru_node) +
 525                               offsetof(struct bpf_lru_node, ref));
 526         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 527         *insn++ = BPF_ST_MEM(BPF_B, ret,
 528                              offsetof(struct htab_elem, lru_node) +
 529                              offsetof(struct bpf_lru_node, ref),
 530                              1);
 531         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 532                                 offsetof(struct htab_elem, key) +
 533                                 round_up(map->key_size, 8));
 534         return insn - insn_buf;
 535 }
 536
 537 /* It is called from the bpf_lru_list when the LRU needs to delete
 538  * older elements from the htab.
 539  */
 540 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 541 {
 542         struct bpf_htab *htab = (struct bpf_htab *)arg;
 543         struct htab_elem *l = NULL, *tgt_l;
 544         struct hlist_nulls_head *head;
 545         struct hlist_nulls_node *n;
 546         unsigned long flags;
 547         struct bucket *b;
 548
 549         tgt_l = container_of(node, struct htab_elem, lru_node);
 550         b = __select_bucket(htab, tgt_l->hash);
 551         head = &b->head;
 552
 553         raw_spin_lock_irqsave(&b->lock, flags);
 554
 555         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 556                 if (l == tgt_l) {
 557                         hlist_nulls_del_rcu(&l->hash_node);
 558                         break;
 559                 }
 560
 561         raw_spin_unlock_irqrestore(&b->lock, flags);
 562
 563         return l == tgt_l;
 564 }
 565
 566 /* Called from syscall */
 567 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 568 {
 569         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 570         struct hlist_nulls_head *head;
 571         struct htab_elem *l, *next_l;
 572         u32 hash, key_size;
 573         int i = 0;
 574
 575         WARN_ON_ONCE(!rcu_read_lock_held());
 576
 577         key_size = map->key_size;
 578
 579         if (!key)
 580                 goto find_first_elem;
 581
 582         hash = htab_map_hash(key, key_size);
 583
 584         head = select_bucket(htab, hash);
 585
 586         /* lookup the key */
 587         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 588
 589         if (!l)
 590                 goto find_first_elem;
 591
 592         /* key was found, get next key in the same bucket */
 593         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 594                                   struct htab_elem, hash_node);
 595
 596         if (next_l) {
 597                 /* if next elem in this hash list is non-zero, just return it */
 598                 memcpy(next_key, next_l->key, key_size);
 599                 return 0;
 600         }
 601
 602         /* no more elements in this hash list, go to the next bucket */
 603         i = hash & (htab->n_buckets - 1);
 604         i++;
 605
 606 find_first_elem:
 607         /* iterate over buckets */
 608         for (; i < htab->n_buckets; i++) {
 609                 head = select_bucket(htab, i);
 610
 611                 /* pick first element in the bucket */
 612                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 613                                           struct htab_elem, hash_node);
 614                 if (next_l) {
 615                         /* if it's not empty, just return it */
 616                         memcpy(next_key, next_l->key, key_size);
 617                         return 0;
 618                 }
 619         }
 620
 621         /* iterated over all buckets and all elements */
 622         return -ENOENT;
 623 }
 624
 625 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 626 {
 627         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 628                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 629         kfree(l);
 630 }
 631
 632 static void htab_elem_free_rcu(struct rcu_head *head)
 633 {
 634         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 635         struct bpf_htab *htab = l->htab;
 636
 637         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 638          * we're calling kfree, otherwise deadlock is possible if kprobes
 639          * are placed somewhere inside of slub
 640          */
 641         preempt_disable();
 642         __this_cpu_inc(bpf_prog_active);
 643         htab_elem_free(htab, l);
 644         __this_cpu_dec(bpf_prog_active);
 645         preempt_enable();
 646 }
 647
 648 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 649 {
 650         struct bpf_map *map = &htab->map;
 651
 652         if (map->ops->map_fd_put_ptr) {
 653                 void *ptr = fd_htab_map_get_ptr(map, l);
 654
 655                 map->ops->map_fd_put_ptr(ptr);
 656         }
 657
 658         if (htab_is_prealloc(htab)) {
 659                 pcpu_freelist_push(&htab->freelist, &l->fnode);
 660         } else {
 661                 atomic_dec(&htab->count);
 662                 l->htab = htab;
 663                 call_rcu(&l->rcu, htab_elem_free_rcu);
 664         }
 665 }
 666
 667 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 668                             void *value, bool onallcpus)
 669 {
 670         if (!onallcpus) {
 671                 /* copy true value_size bytes */
 672                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 673         } else {
 674                 u32 size = round_up(htab->map.value_size, 8);
 675                 int off = 0, cpu;
 676
 677                 for_each_possible_cpu(cpu) {
 678                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 679                                         value + off, size);
 680                         off += size;
 681                 }
 682         }
 683 }
 684
 685 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 686 {
 687         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 688                BITS_PER_LONG == 64;
 689 }
 690
 691 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
 692 {
 693         u32 size = htab->map.value_size;
 694
 695         if (percpu || fd_htab_map_needs_adjust(htab))
 696                 size = round_up(size, 8);
 697         return size;
 698 }
 699
 700 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 701                                          void *value, u32 key_size, u32 hash,
 702                                          bool percpu, bool onallcpus,
 703                                          struct htab_elem *old_elem)
 704 {
 705         u32 size = htab_size_value(htab, percpu);
 706         bool prealloc = htab_is_prealloc(htab);
 707         struct htab_elem *l_new, **pl_new;
 708         void __percpu *pptr;
 709
 710         if (prealloc) {
 711                 if (old_elem) {
 712                         /* if we're updating the existing element,
 713                          * use per-cpu extra elems to avoid freelist_pop/push
 714                          */
 715                         pl_new = this_cpu_ptr(htab->extra_elems);
 716                         l_new = *pl_new;
 717                         *pl_new = old_elem;
 718                 } else {
 719                         struct pcpu_freelist_node *l;
 720
 721                         l = pcpu_freelist_pop(&htab->freelist);
 722                         if (!l)
 723                                 return ERR_PTR(-E2BIG);
 724                         l_new = container_of(l, struct htab_elem, fnode);
 725                 }
 726         } else {
 727                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 728                         if (!old_elem) {
 729                                 /* when map is full and update() is replacing
 730                                  * old element, it's ok to allocate, since
 731                                  * old element will be freed immediately.
 732                                  * Otherwise return an error
 733                                  */
 734                                 atomic_dec(&htab->count);
 735                                 return ERR_PTR(-E2BIG);
 736                         }
 737                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 738                                      htab->map.numa_node);
 739                 if (!l_new)
 740                         return ERR_PTR(-ENOMEM);
 741         }
 742
 743         memcpy(l_new->key, key, key_size);
 744         if (percpu) {
 745                 if (prealloc) {
 746                         pptr = htab_elem_get_ptr(l_new, key_size);
 747                 } else {
 748                         /* alloc_percpu zero-fills */
 749                         pptr = __alloc_percpu_gfp(size, 8,
 750                                                   GFP_ATOMIC | __GFP_NOWARN);
 751                         if (!pptr) {
 752                                 kfree(l_new);
 753                                 return ERR_PTR(-ENOMEM);
 754                         }
 755                 }
 756
 757                 pcpu_copy_value(htab, pptr, value, onallcpus);
 758
 759                 if (!prealloc)
 760                         htab_elem_set_ptr(l_new, key_size, pptr);
 761         } else {
 762                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 763         }
 764
 765         l_new->hash = hash;
 766         return l_new;
 767 }
 768
 769 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 770                        u64 map_flags)
 771 {
 772         if (l_old && map_flags == BPF_NOEXIST)
 773                 /* elem already exists */
 774                 return -EEXIST;
 775
 776         if (!l_old && map_flags == BPF_EXIST)
 777                 /* elem doesn't exist, cannot update it */
 778                 return -ENOENT;
 779
 780         return 0;
 781 }
 782
 783 /* Called from syscall or from eBPF program */
 784 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 785                                 u64 map_flags)
 786 {
 787         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 788         struct htab_elem *l_new = NULL, *l_old;
 789         struct hlist_nulls_head *head;
 790         unsigned long flags;
 791         struct bucket *b;
 792         u32 key_size, hash;
 793         int ret;
 794
 795         if (unlikely(map_flags > BPF_EXIST))
 796                 /* unknown flags */
 797                 return -EINVAL;
 798
 799         WARN_ON_ONCE(!rcu_read_lock_held());
 800
 801         key_size = map->key_size;
 802
 803         hash = htab_map_hash(key, key_size);
 804
 805         b = __select_bucket(htab, hash);
 806         head = &b->head;
 807
 808         /* bpf_map_update_elem() can be called in_irq() */
 809         raw_spin_lock_irqsave(&b->lock, flags);
 810
 811         l_old = lookup_elem_raw(head, hash, key, key_size);
 812
 813         ret = check_flags(htab, l_old, map_flags);
 814         if (ret)
 815                 goto err;
 816
 817         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 818                                 l_old);
 819         if (IS_ERR(l_new)) {
 820                 /* all pre-allocated elements are in use or memory exhausted */
 821                 ret = PTR_ERR(l_new);
 822                 goto err;
 823         }
 824
 825         /* add new element to the head of the list, so that
 826          * concurrent search will find it before old elem
 827          */
 828         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 829         if (l_old) {
 830                 hlist_nulls_del_rcu(&l_old->hash_node);
 831                 if (!htab_is_prealloc(htab))
 832                         free_htab_elem(htab, l_old);
 833         }
 834         ret = 0;
 835 err:
 836         raw_spin_unlock_irqrestore(&b->lock, flags);
 837         return ret;
 838 }
 839
 840 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 841                                     u64 map_flags)
 842 {
 843         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 844         struct htab_elem *l_new, *l_old = NULL;
 845         struct hlist_nulls_head *head;
 846         unsigned long flags;
 847         struct bucket *b;
 848         u32 key_size, hash;
 849         int ret;
 850
 851         if (unlikely(map_flags > BPF_EXIST))
 852                 /* unknown flags */
 853                 return -EINVAL;
 854
 855         WARN_ON_ONCE(!rcu_read_lock_held());
 856
 857         key_size = map->key_size;
 858
 859         hash = htab_map_hash(key, key_size);
 860
 861         b = __select_bucket(htab, hash);
 862         head = &b->head;
 863
 864         /* For LRU, we need to alloc before taking bucket's
 865          * spinlock because getting free nodes from LRU may need
 866          * to remove older elements from htab and this removal
 867          * operation will need a bucket lock.
 868          */
 869         l_new = prealloc_lru_pop(htab, key, hash);
 870         if (!l_new)
 871                 return -ENOMEM;
 872         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 873
 874         /* bpf_map_update_elem() can be called in_irq() */
 875         raw_spin_lock_irqsave(&b->lock, flags);
 876
 877         l_old = lookup_elem_raw(head, hash, key, key_size);
 878
 879         ret = check_flags(htab, l_old, map_flags);
 880         if (ret)
 881                 goto err;
 882
 883         /* add new element to the head of the list, so that
 884          * concurrent search will find it before old elem
 885          */
 886         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 887         if (l_old) {
 888                 bpf_lru_node_set_ref(&l_new->lru_node);
 889                 hlist_nulls_del_rcu(&l_old->hash_node);
 890         }
 891         ret = 0;
 892
 893 err:
 894         raw_spin_unlock_irqrestore(&b->lock, flags);
 895
 896         if (ret)
 897                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 898         else if (l_old)
 899                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 900
 901         return ret;
 902 }
 903
 904 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 905                                          void *value, u64 map_flags,
 906                                          bool onallcpus)
 907 {
 908         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 909         struct htab_elem *l_new = NULL, *l_old;
 910         struct hlist_nulls_head *head;
 911         unsigned long flags;
 912         struct bucket *b;
 913         u32 key_size, hash;
 914         int ret;
 915
 916         if (unlikely(map_flags > BPF_EXIST))
 917                 /* unknown flags */
 918                 return -EINVAL;
 919
 920         WARN_ON_ONCE(!rcu_read_lock_held());
 921
 922         key_size = map->key_size;
 923
 924         hash = htab_map_hash(key, key_size);
 925
 926         b = __select_bucket(htab, hash);
 927         head = &b->head;
 928
 929         /* bpf_map_update_elem() can be called in_irq() */
 930         raw_spin_lock_irqsave(&b->lock, flags);
 931
 932         l_old = lookup_elem_raw(head, hash, key, key_size);
 933
 934         ret = check_flags(htab, l_old, map_flags);
 935         if (ret)
 936                 goto err;
 937
 938         if (l_old) {
 939                 /* per-cpu hash map can update value in-place */
 940                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 941                                 value, onallcpus);
 942         } else {
 943                 l_new = alloc_htab_elem(htab, key, value, key_size,
 944                                         hash, true, onallcpus, NULL);
 945                 if (IS_ERR(l_new)) {
 946                         ret = PTR_ERR(l_new);
 947                         goto err;
 948                 }
 949                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 950         }
 951         ret = 0;
 952 err:
 953         raw_spin_unlock_irqrestore(&b->lock, flags);
 954         return ret;
 955 }
 956
 957 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 958                                              void *value, u64 map_flags,
 959                                              bool onallcpus)
 960 {
 961         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 962         struct htab_elem *l_new = NULL, *l_old;
 963         struct hlist_nulls_head *head;
 964         unsigned long flags;
 965         struct bucket *b;
 966         u32 key_size, hash;
 967         int ret;
 968
 969         if (unlikely(map_flags > BPF_EXIST))
 970                 /* unknown flags */
 971                 return -EINVAL;
 972
 973         WARN_ON_ONCE(!rcu_read_lock_held());
 974
 975         key_size = map->key_size;
 976
 977         hash = htab_map_hash(key, key_size);
 978
 979         b = __select_bucket(htab, hash);
 980         head = &b->head;
 981
 982         /* For LRU, we need to alloc before taking bucket's
 983          * spinlock because LRU's elem alloc may need
 984          * to remove older elem from htab and this removal
 985          * operation will need a bucket lock.
 986          */
 987         if (map_flags != BPF_EXIST) {
 988                 l_new = prealloc_lru_pop(htab, key, hash);
 989                 if (!l_new)
 990                         return -ENOMEM;
 991         }
 992
 993         /* bpf_map_update_elem() can be called in_irq() */
 994         raw_spin_lock_irqsave(&b->lock, flags);
 995
 996         l_old = lookup_elem_raw(head, hash, key, key_size);
 997
 998         ret = check_flags(htab, l_old, map_flags);
 999         if (ret)
1000                 goto err;
1001
1002         if (l_old) {
1003                 bpf_lru_node_set_ref(&l_old->lru_node);
1004
1005                 /* per-cpu hash map can update value in-place */
1006                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1007                                 value, onallcpus);
1008         } else {
1009                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1010                                 value, onallcpus);
1011                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1012                 l_new = NULL;
1013         }
1014         ret = 0;
1015 err:
1016         raw_spin_unlock_irqrestore(&b->lock, flags);
1017         if (l_new)
1018                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1019         return ret;
1020 }
1021
1022 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1023                                        void *value, u64 map_flags)
1024 {
1025         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1026 }
1027
1028 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1029                                            void *value, u64 map_flags)
1030 {
1031         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1032                                                  false);
1033 }
1034
1035 /* Called from syscall or from eBPF program */
1036 static int htab_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                 free_htab_elem(htab, l);
1061                 ret = 0;
1062         }
1063
1064         raw_spin_unlock_irqrestore(&b->lock, flags);
1065         return ret;
1066 }
1067
1068 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1069 {
1070         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1071         struct hlist_nulls_head *head;
1072         struct bucket *b;
1073         struct htab_elem *l;
1074         unsigned long flags;
1075         u32 hash, key_size;
1076         int ret = -ENOENT;
1077
1078         WARN_ON_ONCE(!rcu_read_lock_held());
1079
1080         key_size = map->key_size;
1081
1082         hash = htab_map_hash(key, key_size);
1083         b = __select_bucket(htab, hash);
1084         head = &b->head;
1085
1086         raw_spin_lock_irqsave(&b->lock, flags);
1087
1088         l = lookup_elem_raw(head, hash, key, key_size);
1089
1090         if (l) {
1091                 hlist_nulls_del_rcu(&l->hash_node);
1092                 ret = 0;
1093         }
1094
1095         raw_spin_unlock_irqrestore(&b->lock, flags);
1096         if (l)
1097                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1098         return ret;
1099 }
1100
1101 static void delete_all_elements(struct bpf_htab *htab)
1102 {
1103         int i;
1104
1105         for (i = 0; i < htab->n_buckets; i++) {
1106                 struct hlist_nulls_head *head = select_bucket(htab, i);
1107                 struct hlist_nulls_node *n;
1108                 struct htab_elem *l;
1109
1110                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1111                         hlist_nulls_del_rcu(&l->hash_node);
1112                         htab_elem_free(htab, l);
1113                 }
1114         }
1115 }
1116
1117 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1118 static void htab_map_free(struct bpf_map *map)
1119 {
1120         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1121
1122         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1123          * so the programs (can be more than one that used this map) were
1124          * disconnected from events. Wait for outstanding critical sections in
1125          * these programs to complete
1126          */
1127         synchronize_rcu();
1128
1129         /* some of free_htab_elem() callbacks for elements of this map may
1130          * not have executed. Wait for them.
1131          */
1132         rcu_barrier();
1133         if (!htab_is_prealloc(htab))
1134                 delete_all_elements(htab);
1135         else
1136                 prealloc_destroy(htab);
1137
1138         free_percpu(htab->extra_elems);
1139         bpf_map_area_free(htab->buckets);
1140         kfree(htab);
1141 }
1142
1143 const struct bpf_map_ops htab_map_ops = {
1144         .map_alloc = htab_map_alloc,
1145         .map_free = htab_map_free,
1146         .map_get_next_key = htab_map_get_next_key,
1147         .map_lookup_elem = htab_map_lookup_elem,
1148         .map_update_elem = htab_map_update_elem,
1149         .map_delete_elem = htab_map_delete_elem,
1150         .map_gen_lookup = htab_map_gen_lookup,
1151 };
1152
1153 const struct bpf_map_ops htab_lru_map_ops = {
1154         .map_alloc = htab_map_alloc,
1155         .map_free = htab_map_free,
1156         .map_get_next_key = htab_map_get_next_key,
1157         .map_lookup_elem = htab_lru_map_lookup_elem,
1158         .map_update_elem = htab_lru_map_update_elem,
1159         .map_delete_elem = htab_lru_map_delete_elem,
1160         .map_gen_lookup = htab_lru_map_gen_lookup,
1161 };
1162
1163 /* Called from eBPF program */
1164 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1165 {
1166         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1167
1168         if (l)
1169                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1170         else
1171                 return NULL;
1172 }
1173
1174 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1175 {
1176         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1177
1178         if (l) {
1179                 bpf_lru_node_set_ref(&l->lru_node);
1180                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1181         }
1182
1183         return NULL;
1184 }
1185
1186 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1187 {
1188         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1189         struct htab_elem *l;
1190         void __percpu *pptr;
1191         int ret = -ENOENT;
1192         int cpu, off = 0;
1193         u32 size;
1194
1195         /* per_cpu areas are zero-filled and bpf programs can only
1196          * access 'value_size' of them, so copying rounded areas
1197          * will not leak any kernel data
1198          */
1199         size = round_up(map->value_size, 8);
1200         rcu_read_lock();
1201         l = __htab_map_lookup_elem(map, key);
1202         if (!l)
1203                 goto out;
1204         if (htab_is_lru(htab))
1205                 bpf_lru_node_set_ref(&l->lru_node);
1206         pptr = htab_elem_get_ptr(l, map->key_size);
1207         for_each_possible_cpu(cpu) {
1208                 bpf_long_memcpy(value + off,
1209                                 per_cpu_ptr(pptr, cpu), size);
1210                 off += size;
1211         }
1212         ret = 0;
1213 out:
1214         rcu_read_unlock();
1215         return ret;
1216 }
1217
1218 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1219                            u64 map_flags)
1220 {
1221         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1222         int ret;
1223
1224         rcu_read_lock();
1225         if (htab_is_lru(htab))
1226                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1227                                                         map_flags, true);
1228         else
1229                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1230                                                     true);
1231         rcu_read_unlock();
1232
1233         return ret;
1234 }
1235
1236 const struct bpf_map_ops htab_percpu_map_ops = {
1237         .map_alloc = htab_map_alloc,
1238         .map_free = htab_map_free,
1239         .map_get_next_key = htab_map_get_next_key,
1240         .map_lookup_elem = htab_percpu_map_lookup_elem,
1241         .map_update_elem = htab_percpu_map_update_elem,
1242         .map_delete_elem = htab_map_delete_elem,
1243 };
1244
1245 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1246         .map_alloc = htab_map_alloc,
1247         .map_free = htab_map_free,
1248         .map_get_next_key = htab_map_get_next_key,
1249         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1250         .map_update_elem = htab_lru_percpu_map_update_elem,
1251         .map_delete_elem = htab_lru_map_delete_elem,
1252 };
1253
1254 static struct bpf_map *fd_htab_map_alloc(union bpf_attr *attr)
1255 {
1256         if (attr->value_size != sizeof(u32))
1257                 return ERR_PTR(-EINVAL);
1258         return htab_map_alloc(attr);
1259 }
1260
1261 static void fd_htab_map_free(struct bpf_map *map)
1262 {
1263         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1264         struct hlist_nulls_node *n;
1265         struct hlist_nulls_head *head;
1266         struct htab_elem *l;
1267         int i;
1268
1269         for (i = 0; i < htab->n_buckets; i++) {
1270                 head = select_bucket(htab, i);
1271
1272                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1273                         void *ptr = fd_htab_map_get_ptr(map, l);
1274
1275                         map->ops->map_fd_put_ptr(ptr);
1276                 }
1277         }
1278
1279         htab_map_free(map);
1280 }
1281
1282 /* only called from syscall */
1283 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1284 {
1285         void **ptr;
1286         int ret = 0;
1287
1288         if (!map->ops->map_fd_sys_lookup_elem)
1289                 return -ENOTSUPP;
1290
1291         rcu_read_lock();
1292         ptr = htab_map_lookup_elem(map, key);
1293         if (ptr)
1294                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1295         else
1296                 ret = -ENOENT;
1297         rcu_read_unlock();
1298
1299         return ret;
1300 }
1301
1302 /* only called from syscall */
1303 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1304                                 void *key, void *value, u64 map_flags)
1305 {
1306         void *ptr;
1307         int ret;
1308         u32 ufd = *(u32 *)value;
1309
1310         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1311         if (IS_ERR(ptr))
1312                 return PTR_ERR(ptr);
1313
1314         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1315         if (ret)
1316                 map->ops->map_fd_put_ptr(ptr);
1317
1318         return ret;
1319 }
1320
1321 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1322 {
1323         struct bpf_map *map, *inner_map_meta;
1324
1325         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1326         if (IS_ERR(inner_map_meta))
1327                 return inner_map_meta;
1328
1329         map = fd_htab_map_alloc(attr);
1330         if (IS_ERR(map)) {
1331                 bpf_map_meta_free(inner_map_meta);
1332                 return map;
1333         }
1334
1335         map->inner_map_meta = inner_map_meta;
1336
1337         return map;
1338 }
1339
1340 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1341 {
1342         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1343
1344         if (!inner_map)
1345                 return NULL;
1346
1347         return READ_ONCE(*inner_map);
1348 }
1349
1350 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1351                                   struct bpf_insn *insn_buf)
1352 {
1353         struct bpf_insn *insn = insn_buf;
1354         const int ret = BPF_REG_0;
1355
1356         *insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
1357         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1358         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1359                                 offsetof(struct htab_elem, key) +
1360                                 round_up(map->key_size, 8));
1361         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1362
1363         return insn - insn_buf;
1364 }
1365
1366 static void htab_of_map_free(struct bpf_map *map)
1367 {
1368         bpf_map_meta_free(map->inner_map_meta);
1369         fd_htab_map_free(map);
1370 }
1371
1372 const struct bpf_map_ops htab_of_maps_map_ops = {
1373         .map_alloc = htab_of_map_alloc,
1374         .map_free = htab_of_map_free,
1375         .map_get_next_key = htab_map_get_next_key,
1376         .map_lookup_elem = htab_of_map_lookup_elem,
1377         .map_delete_elem = htab_map_delete_elem,
1378         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1379         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1380         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1381         .map_gen_lookup = htab_of_map_gen_lookup,
1382 };