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/btf.h>
  15 #include <linux/jhash.h>
  16 #include <linux/filter.h>
  17 #include <linux/rculist_nulls.h>
  18 #include <linux/random.h>
  19 #include <uapi/linux/btf.h>
  20 #include "percpu_freelist.h"
  21 #include "bpf_lru_list.h"
  22 #include "map_in_map.h"
  23
  24 #define HTAB_CREATE_FLAG_MASK                                           \
  25         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  26          BPF_F_RDONLY | BPF_F_WRONLY)
  27
  28 struct bucket {
  29         struct hlist_nulls_head head;
  30         raw_spinlock_t lock;
  31 };
  32
  33 struct bpf_htab {
  34         struct bpf_map map;
  35         struct bucket *buckets;
  36         void *elems;
  37         union {
  38                 struct pcpu_freelist freelist;
  39                 struct bpf_lru lru;
  40         };
  41         struct htab_elem *__percpu *extra_elems;
  42         atomic_t count; /* number of elements in this hashtable */
  43         u32 n_buckets;  /* number of hash buckets */
  44         u32 elem_size;  /* size of each element in bytes */
  45         u32 hashrnd;
  46 };
  47
  48 /* each htab element is struct htab_elem + key + value */
  49 struct htab_elem {
  50         union {
  51                 struct hlist_nulls_node hash_node;
  52                 struct {
  53                         void *padding;
  54                         union {
  55                                 struct bpf_htab *htab;
  56                                 struct pcpu_freelist_node fnode;
  57                         };
  58                 };
  59         };
  60         union {
  61                 struct rcu_head rcu;
  62                 struct bpf_lru_node lru_node;
  63         };
  64         u32 hash;
  65         char key[0] __aligned(8);
  66 };
  67
  68 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  69
  70 static bool htab_is_lru(const struct bpf_htab *htab)
  71 {
  72         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  73                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  74 }
  75
  76 static bool htab_is_percpu(const struct bpf_htab *htab)
  77 {
  78         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  79                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  80 }
  81
  82 static bool htab_is_prealloc(const struct bpf_htab *htab)
  83 {
  84         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  85 }
  86
  87 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  88                                      void __percpu *pptr)
  89 {
  90         *(void __percpu **)(l->key + key_size) = pptr;
  91 }
  92
  93 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  94 {
  95         return *(void __percpu **)(l->key + key_size);
  96 }
  97
  98 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
  99 {
 100         return *(void **)(l->key + roundup(map->key_size, 8));
 101 }
 102
 103 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 104 {
 105         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 106 }
 107
 108 static void htab_free_elems(struct bpf_htab *htab)
 109 {
 110         int i;
 111
 112         if (!htab_is_percpu(htab))
 113                 goto free_elems;
 114
 115         for (i = 0; i < htab->map.max_entries; i++) {
 116                 void __percpu *pptr;
 117
 118                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 119                                          htab->map.key_size);
 120                 free_percpu(pptr);
 121                 cond_resched();
 122         }
 123 free_elems:
 124         bpf_map_area_free(htab->elems);
 125 }
 126
 127 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 128                                           u32 hash)
 129 {
 130         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 131         struct htab_elem *l;
 132
 133         if (node) {
 134                 l = container_of(node, struct htab_elem, lru_node);
 135                 memcpy(l->key, key, htab->map.key_size);
 136                 return l;
 137         }
 138
 139         return NULL;
 140 }
 141
 142 static int prealloc_init(struct bpf_htab *htab)
 143 {
 144         u32 num_entries = htab->map.max_entries;
 145         int err = -ENOMEM, i;
 146
 147         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 148                 num_entries += num_possible_cpus();
 149
 150         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 151                                          htab->map.numa_node);
 152         if (!htab->elems)
 153                 return -ENOMEM;
 154
 155         if (!htab_is_percpu(htab))
 156                 goto skip_percpu_elems;
 157
 158         for (i = 0; i < num_entries; i++) {
 159                 u32 size = round_up(htab->map.value_size, 8);
 160                 void __percpu *pptr;
 161
 162                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 163                 if (!pptr)
 164                         goto free_elems;
 165                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 166                                   pptr);
 167                 cond_resched();
 168         }
 169
 170 skip_percpu_elems:
 171         if (htab_is_lru(htab))
 172                 err = bpf_lru_init(&htab->lru,
 173                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 174                                    offsetof(struct htab_elem, hash) -
 175                                    offsetof(struct htab_elem, lru_node),
 176                                    htab_lru_map_delete_node,
 177                                    htab);
 178         else
 179                 err = pcpu_freelist_init(&htab->freelist);
 180
 181         if (err)
 182                 goto free_elems;
 183
 184         if (htab_is_lru(htab))
 185                 bpf_lru_populate(&htab->lru, htab->elems,
 186                                  offsetof(struct htab_elem, lru_node),
 187                                  htab->elem_size, num_entries);
 188         else
 189                 pcpu_freelist_populate(&htab->freelist,
 190                                        htab->elems + offsetof(struct htab_elem, fnode),
 191                                        htab->elem_size, num_entries);
 192
 193         return 0;
 194
 195 free_elems:
 196         htab_free_elems(htab);
 197         return err;
 198 }
 199
 200 static void prealloc_destroy(struct bpf_htab *htab)
 201 {
 202         htab_free_elems(htab);
 203
 204         if (htab_is_lru(htab))
 205                 bpf_lru_destroy(&htab->lru);
 206         else
 207                 pcpu_freelist_destroy(&htab->freelist);
 208 }
 209
 210 static int alloc_extra_elems(struct bpf_htab *htab)
 211 {
 212         struct htab_elem *__percpu *pptr, *l_new;
 213         struct pcpu_freelist_node *l;
 214         int cpu;
 215
 216         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 217                                   GFP_USER | __GFP_NOWARN);
 218         if (!pptr)
 219                 return -ENOMEM;
 220
 221         for_each_possible_cpu(cpu) {
 222                 l = pcpu_freelist_pop(&htab->freelist);
 223                 /* pop will succeed, since prealloc_init()
 224                  * preallocated extra num_possible_cpus elements
 225                  */
 226                 l_new = container_of(l, struct htab_elem, fnode);
 227                 *per_cpu_ptr(pptr, cpu) = l_new;
 228         }
 229         htab->extra_elems = pptr;
 230         return 0;
 231 }
 232
 233 /* Called from syscall */
 234 static int htab_map_alloc_check(union bpf_attr *attr)
 235 {
 236         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 237                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 238         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 239                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 240         /* percpu_lru means each cpu has its own LRU list.
 241          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 242          * the map's value itself is percpu.  percpu_lru has
 243          * nothing to do with the map's value.
 244          */
 245         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 246         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 247         int numa_node = bpf_map_attr_numa_node(attr);
 248
 249         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 250                      offsetof(struct htab_elem, hash_node.pprev));
 251         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 252                      offsetof(struct htab_elem, hash_node.pprev));
 253
 254         if (lru && !capable(CAP_SYS_ADMIN))
 255                 /* LRU implementation is much complicated than other
 256                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 257                  */
 258                 return -EPERM;
 259
 260         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
 261                 /* reserved bits should not be used */
 262                 return -EINVAL;
 263
 264         if (!lru && percpu_lru)
 265                 return -EINVAL;
 266
 267         if (lru && !prealloc)
 268                 return -ENOTSUPP;
 269
 270         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 271                 return -EINVAL;
 272
 273         /* check sanity of attributes.
 274          * value_size == 0 may be allowed in the future to use map as a set
 275          */
 276         if (attr->max_entries == 0 || attr->key_size == 0 ||
 277             attr->value_size == 0)
 278                 return -EINVAL;
 279
 280         if (attr->key_size > MAX_BPF_STACK)
 281                 /* eBPF programs initialize keys on stack, so they cannot be
 282                  * larger than max stack size
 283                  */
 284                 return -E2BIG;
 285
 286         if (attr->value_size >= KMALLOC_MAX_SIZE -
 287             MAX_BPF_STACK - sizeof(struct htab_elem))
 288                 /* if value_size is bigger, the user space won't be able to
 289                  * access the elements via bpf syscall. This check also makes
 290                  * sure that the elem_size doesn't overflow and it's
 291                  * kmalloc-able later in htab_map_update_elem()
 292                  */
 293                 return -E2BIG;
 294
 295         return 0;
 296 }
 297
 298 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 299 {
 300         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 301                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 302         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 303                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 304         /* percpu_lru means each cpu has its own LRU list.
 305          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 306          * the map's value itself is percpu.  percpu_lru has
 307          * nothing to do with the map's value.
 308          */
 309         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 310         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 311         struct bpf_htab *htab;
 312         int err, i;
 313         u64 cost;
 314
 315         htab = kzalloc(sizeof(*htab), GFP_USER);
 316         if (!htab)
 317                 return ERR_PTR(-ENOMEM);
 318
 319         bpf_map_init_from_attr(&htab->map, attr);
 320
 321         if (percpu_lru) {
 322                 /* ensure each CPU's lru list has >=1 elements.
 323                  * since we are at it, make each lru list has the same
 324                  * number of elements.
 325                  */
 326                 htab->map.max_entries = roundup(attr->max_entries,
 327                                                 num_possible_cpus());
 328                 if (htab->map.max_entries < attr->max_entries)
 329                         htab->map.max_entries = rounddown(attr->max_entries,
 330                                                           num_possible_cpus());
 331         }
 332
 333         /* hash table size must be power of 2 */
 334         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 335
 336         htab->elem_size = sizeof(struct htab_elem) +
 337                           round_up(htab->map.key_size, 8);
 338         if (percpu)
 339                 htab->elem_size += sizeof(void *);
 340         else
 341                 htab->elem_size += round_up(htab->map.value_size, 8);
 342
 343         err = -E2BIG;
 344         /* prevent zero size kmalloc and check for u32 overflow */
 345         if (htab->n_buckets == 0 ||
 346             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 347                 goto free_htab;
 348
 349         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 350                (u64) htab->elem_size * htab->map.max_entries;
 351
 352         if (percpu)
 353                 cost += (u64) round_up(htab->map.value_size, 8) *
 354                         num_possible_cpus() * htab->map.max_entries;
 355         else
 356                cost += (u64) htab->elem_size * num_possible_cpus();
 357
 358         if (cost >= U32_MAX - PAGE_SIZE)
 359                 /* make sure page count doesn't overflow */
 360                 goto free_htab;
 361
 362         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 363
 364         /* if map size is larger than memlock limit, reject it early */
 365         err = bpf_map_precharge_memlock(htab->map.pages);
 366         if (err)
 367                 goto free_htab;
 368
 369         err = -ENOMEM;
 370         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 371                                            sizeof(struct bucket),
 372                                            htab->map.numa_node);
 373         if (!htab->buckets)
 374                 goto free_htab;
 375
 376         htab->hashrnd = get_random_int();
 377         for (i = 0; i < htab->n_buckets; i++) {
 378                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 379                 raw_spin_lock_init(&htab->buckets[i].lock);
 380         }
 381
 382         if (prealloc) {
 383                 err = prealloc_init(htab);
 384                 if (err)
 385                         goto free_buckets;
 386
 387                 if (!percpu && !lru) {
 388                         /* lru itself can remove the least used element, so
 389                          * there is no need for an extra elem during map_update.
 390                          */
 391                         err = alloc_extra_elems(htab);
 392                         if (err)
 393                                 goto free_prealloc;
 394                 }
 395         }
 396
 397         return &htab->map;
 398
 399 free_prealloc:
 400         prealloc_destroy(htab);
 401 free_buckets:
 402         bpf_map_area_free(htab->buckets);
 403 free_htab:
 404         kfree(htab);
 405         return ERR_PTR(err);
 406 }
 407
 408 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 409 {
 410         return jhash(key, key_len, hashrnd);
 411 }
 412
 413 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 414 {
 415         return &htab->buckets[hash & (htab->n_buckets - 1)];
 416 }
 417
 418 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 419 {
 420         return &__select_bucket(htab, hash)->head;
 421 }
 422
 423 /* this lookup function can only be called with bucket lock taken */
 424 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 425                                          void *key, u32 key_size)
 426 {
 427         struct hlist_nulls_node *n;
 428         struct htab_elem *l;
 429
 430         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 431                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 432                         return l;
 433
 434         return NULL;
 435 }
 436
 437 /* can be called without bucket lock. it will repeat the loop in
 438  * the unlikely event when elements moved from one bucket into another
 439  * while link list is being walked
 440  */
 441 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 442                                                u32 hash, void *key,
 443                                                u32 key_size, u32 n_buckets)
 444 {
 445         struct hlist_nulls_node *n;
 446         struct htab_elem *l;
 447
 448 again:
 449         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 450                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 451                         return l;
 452
 453         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 454                 goto again;
 455
 456         return NULL;
 457 }
 458
 459 /* Called from syscall or from eBPF program directly, so
 460  * arguments have to match bpf_map_lookup_elem() exactly.
 461  * The return value is adjusted by BPF instructions
 462  * in htab_map_gen_lookup().
 463  */
 464 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 465 {
 466         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 467         struct hlist_nulls_head *head;
 468         struct htab_elem *l;
 469         u32 hash, key_size;
 470
 471         /* Must be called with rcu_read_lock. */
 472         WARN_ON_ONCE(!rcu_read_lock_held());
 473
 474         key_size = map->key_size;
 475
 476         hash = htab_map_hash(key, key_size, htab->hashrnd);
 477
 478         head = select_bucket(htab, hash);
 479
 480         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 481
 482         return l;
 483 }
 484
 485 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 486 {
 487         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 488
 489         if (l)
 490                 return l->key + round_up(map->key_size, 8);
 491
 492         return NULL;
 493 }
 494
 495 /* inline bpf_map_lookup_elem() call.
 496  * Instead of:
 497  * bpf_prog
 498  *   bpf_map_lookup_elem
 499  *     map->ops->map_lookup_elem
 500  *       htab_map_lookup_elem
 501  *         __htab_map_lookup_elem
 502  * do:
 503  * bpf_prog
 504  *   __htab_map_lookup_elem
 505  */
 506 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 507 {
 508         struct bpf_insn *insn = insn_buf;
 509         const int ret = BPF_REG_0;
 510
 511         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 512                      (void *(*)(struct bpf_map *map, void *key))NULL));
 513         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 514         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 515         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 516                                 offsetof(struct htab_elem, key) +
 517                                 round_up(map->key_size, 8));
 518         return insn - insn_buf;
 519 }
 520
 521 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 522 {
 523         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 524
 525         if (l) {
 526                 bpf_lru_node_set_ref(&l->lru_node);
 527                 return l->key + round_up(map->key_size, 8);
 528         }
 529
 530         return NULL;
 531 }
 532
 533 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 534                                    struct bpf_insn *insn_buf)
 535 {
 536         struct bpf_insn *insn = insn_buf;
 537         const int ret = BPF_REG_0;
 538         const int ref_reg = BPF_REG_1;
 539
 540         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 541                      (void *(*)(struct bpf_map *map, void *key))NULL));
 542         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 543         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 544         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 545                               offsetof(struct htab_elem, lru_node) +
 546                               offsetof(struct bpf_lru_node, ref));
 547         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 548         *insn++ = BPF_ST_MEM(BPF_B, ret,
 549                              offsetof(struct htab_elem, lru_node) +
 550                              offsetof(struct bpf_lru_node, ref),
 551                              1);
 552         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 553                                 offsetof(struct htab_elem, key) +
 554                                 round_up(map->key_size, 8));
 555         return insn - insn_buf;
 556 }
 557
 558 /* It is called from the bpf_lru_list when the LRU needs to delete
 559  * older elements from the htab.
 560  */
 561 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 562 {
 563         struct bpf_htab *htab = (struct bpf_htab *)arg;
 564         struct htab_elem *l = NULL, *tgt_l;
 565         struct hlist_nulls_head *head;
 566         struct hlist_nulls_node *n;
 567         unsigned long flags;
 568         struct bucket *b;
 569
 570         tgt_l = container_of(node, struct htab_elem, lru_node);
 571         b = __select_bucket(htab, tgt_l->hash);
 572         head = &b->head;
 573
 574         raw_spin_lock_irqsave(&b->lock, flags);
 575
 576         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 577                 if (l == tgt_l) {
 578                         hlist_nulls_del_rcu(&l->hash_node);
 579                         break;
 580                 }
 581
 582         raw_spin_unlock_irqrestore(&b->lock, flags);
 583
 584         return l == tgt_l;
 585 }
 586
 587 /* Called from syscall */
 588 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 589 {
 590         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 591         struct hlist_nulls_head *head;
 592         struct htab_elem *l, *next_l;
 593         u32 hash, key_size;
 594         int i = 0;
 595
 596         WARN_ON_ONCE(!rcu_read_lock_held());
 597
 598         key_size = map->key_size;
 599
 600         if (!key)
 601                 goto find_first_elem;
 602
 603         hash = htab_map_hash(key, key_size, htab->hashrnd);
 604
 605         head = select_bucket(htab, hash);
 606
 607         /* lookup the key */
 608         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 609
 610         if (!l)
 611                 goto find_first_elem;
 612
 613         /* key was found, get next key in the same bucket */
 614         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 615                                   struct htab_elem, hash_node);
 616
 617         if (next_l) {
 618                 /* if next elem in this hash list is non-zero, just return it */
 619                 memcpy(next_key, next_l->key, key_size);
 620                 return 0;
 621         }
 622
 623         /* no more elements in this hash list, go to the next bucket */
 624         i = hash & (htab->n_buckets - 1);
 625         i++;
 626
 627 find_first_elem:
 628         /* iterate over buckets */
 629         for (; i < htab->n_buckets; i++) {
 630                 head = select_bucket(htab, i);
 631
 632                 /* pick first element in the bucket */
 633                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 634                                           struct htab_elem, hash_node);
 635                 if (next_l) {
 636                         /* if it's not empty, just return it */
 637                         memcpy(next_key, next_l->key, key_size);
 638                         return 0;
 639                 }
 640         }
 641
 642         /* iterated over all buckets and all elements */
 643         return -ENOENT;
 644 }
 645
 646 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 647 {
 648         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 649                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 650         kfree(l);
 651 }
 652
 653 static void htab_elem_free_rcu(struct rcu_head *head)
 654 {
 655         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 656         struct bpf_htab *htab = l->htab;
 657
 658         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 659          * we're calling kfree, otherwise deadlock is possible if kprobes
 660          * are placed somewhere inside of slub
 661          */
 662         preempt_disable();
 663         __this_cpu_inc(bpf_prog_active);
 664         htab_elem_free(htab, l);
 665         __this_cpu_dec(bpf_prog_active);
 666         preempt_enable();
 667 }
 668
 669 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 670 {
 671         struct bpf_map *map = &htab->map;
 672
 673         if (map->ops->map_fd_put_ptr) {
 674                 void *ptr = fd_htab_map_get_ptr(map, l);
 675
 676                 map->ops->map_fd_put_ptr(ptr);
 677         }
 678
 679         if (htab_is_prealloc(htab)) {
 680                 pcpu_freelist_push(&htab->freelist, &l->fnode);
 681         } else {
 682                 atomic_dec(&htab->count);
 683                 l->htab = htab;
 684                 call_rcu(&l->rcu, htab_elem_free_rcu);
 685         }
 686 }
 687
 688 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 689                             void *value, bool onallcpus)
 690 {
 691         if (!onallcpus) {
 692                 /* copy true value_size bytes */
 693                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 694         } else {
 695                 u32 size = round_up(htab->map.value_size, 8);
 696                 int off = 0, cpu;
 697
 698                 for_each_possible_cpu(cpu) {
 699                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 700                                         value + off, size);
 701                         off += size;
 702                 }
 703         }
 704 }
 705
 706 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 707 {
 708         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 709                BITS_PER_LONG == 64;
 710 }
 711
 712 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
 713 {
 714         u32 size = htab->map.value_size;
 715
 716         if (percpu || fd_htab_map_needs_adjust(htab))
 717                 size = round_up(size, 8);
 718         return size;
 719 }
 720
 721 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 722                                          void *value, u32 key_size, u32 hash,
 723                                          bool percpu, bool onallcpus,
 724                                          struct htab_elem *old_elem)
 725 {
 726         u32 size = htab_size_value(htab, percpu);
 727         bool prealloc = htab_is_prealloc(htab);
 728         struct htab_elem *l_new, **pl_new;
 729         void __percpu *pptr;
 730
 731         if (prealloc) {
 732                 if (old_elem) {
 733                         /* if we're updating the existing element,
 734                          * use per-cpu extra elems to avoid freelist_pop/push
 735                          */
 736                         pl_new = this_cpu_ptr(htab->extra_elems);
 737                         l_new = *pl_new;
 738                         *pl_new = old_elem;
 739                 } else {
 740                         struct pcpu_freelist_node *l;
 741
 742                         l = pcpu_freelist_pop(&htab->freelist);
 743                         if (!l)
 744                                 return ERR_PTR(-E2BIG);
 745                         l_new = container_of(l, struct htab_elem, fnode);
 746                 }
 747         } else {
 748                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 749                         if (!old_elem) {
 750                                 /* when map is full and update() is replacing
 751                                  * old element, it's ok to allocate, since
 752                                  * old element will be freed immediately.
 753                                  * Otherwise return an error
 754                                  */
 755                                 l_new = ERR_PTR(-E2BIG);
 756                                 goto dec_count;
 757                         }
 758                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 759                                      htab->map.numa_node);
 760                 if (!l_new) {
 761                         l_new = ERR_PTR(-ENOMEM);
 762                         goto dec_count;
 763                 }
 764         }
 765
 766         memcpy(l_new->key, key, key_size);
 767         if (percpu) {
 768                 if (prealloc) {
 769                         pptr = htab_elem_get_ptr(l_new, key_size);
 770                 } else {
 771                         /* alloc_percpu zero-fills */
 772                         pptr = __alloc_percpu_gfp(size, 8,
 773                                                   GFP_ATOMIC | __GFP_NOWARN);
 774                         if (!pptr) {
 775                                 kfree(l_new);
 776                                 l_new = ERR_PTR(-ENOMEM);
 777                                 goto dec_count;
 778                         }
 779                 }
 780
 781                 pcpu_copy_value(htab, pptr, value, onallcpus);
 782
 783                 if (!prealloc)
 784                         htab_elem_set_ptr(l_new, key_size, pptr);
 785         } else {
 786                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 787         }
 788
 789         l_new->hash = hash;
 790         return l_new;
 791 dec_count:
 792         atomic_dec(&htab->count);
 793         return l_new;
 794 }
 795
 796 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 797                        u64 map_flags)
 798 {
 799         if (l_old && map_flags == BPF_NOEXIST)
 800                 /* elem already exists */
 801                 return -EEXIST;
 802
 803         if (!l_old && map_flags == BPF_EXIST)
 804                 /* elem doesn't exist, cannot update it */
 805                 return -ENOENT;
 806
 807         return 0;
 808 }
 809
 810 /* Called from syscall or from eBPF program */
 811 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 812                                 u64 map_flags)
 813 {
 814         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 815         struct htab_elem *l_new = NULL, *l_old;
 816         struct hlist_nulls_head *head;
 817         unsigned long flags;
 818         struct bucket *b;
 819         u32 key_size, hash;
 820         int ret;
 821
 822         if (unlikely(map_flags > BPF_EXIST))
 823                 /* unknown flags */
 824                 return -EINVAL;
 825
 826         WARN_ON_ONCE(!rcu_read_lock_held());
 827
 828         key_size = map->key_size;
 829
 830         hash = htab_map_hash(key, key_size, htab->hashrnd);
 831
 832         b = __select_bucket(htab, hash);
 833         head = &b->head;
 834
 835         /* bpf_map_update_elem() can be called in_irq() */
 836         raw_spin_lock_irqsave(&b->lock, flags);
 837
 838         l_old = lookup_elem_raw(head, hash, key, key_size);
 839
 840         ret = check_flags(htab, l_old, map_flags);
 841         if (ret)
 842                 goto err;
 843
 844         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 845                                 l_old);
 846         if (IS_ERR(l_new)) {
 847                 /* all pre-allocated elements are in use or memory exhausted */
 848                 ret = PTR_ERR(l_new);
 849                 goto err;
 850         }
 851
 852         /* add new element to the head of the list, so that
 853          * concurrent search will find it before old elem
 854          */
 855         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 856         if (l_old) {
 857                 hlist_nulls_del_rcu(&l_old->hash_node);
 858                 if (!htab_is_prealloc(htab))
 859                         free_htab_elem(htab, l_old);
 860         }
 861         ret = 0;
 862 err:
 863         raw_spin_unlock_irqrestore(&b->lock, flags);
 864         return ret;
 865 }
 866
 867 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 868                                     u64 map_flags)
 869 {
 870         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 871         struct htab_elem *l_new, *l_old = NULL;
 872         struct hlist_nulls_head *head;
 873         unsigned long flags;
 874         struct bucket *b;
 875         u32 key_size, hash;
 876         int ret;
 877
 878         if (unlikely(map_flags > BPF_EXIST))
 879                 /* unknown flags */
 880                 return -EINVAL;
 881
 882         WARN_ON_ONCE(!rcu_read_lock_held());
 883
 884         key_size = map->key_size;
 885
 886         hash = htab_map_hash(key, key_size, htab->hashrnd);
 887
 888         b = __select_bucket(htab, hash);
 889         head = &b->head;
 890
 891         /* For LRU, we need to alloc before taking bucket's
 892          * spinlock because getting free nodes from LRU may need
 893          * to remove older elements from htab and this removal
 894          * operation will need a bucket lock.
 895          */
 896         l_new = prealloc_lru_pop(htab, key, hash);
 897         if (!l_new)
 898                 return -ENOMEM;
 899         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 900
 901         /* bpf_map_update_elem() can be called in_irq() */
 902         raw_spin_lock_irqsave(&b->lock, flags);
 903
 904         l_old = lookup_elem_raw(head, hash, key, key_size);
 905
 906         ret = check_flags(htab, l_old, map_flags);
 907         if (ret)
 908                 goto err;
 909
 910         /* add new element to the head of the list, so that
 911          * concurrent search will find it before old elem
 912          */
 913         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 914         if (l_old) {
 915                 bpf_lru_node_set_ref(&l_new->lru_node);
 916                 hlist_nulls_del_rcu(&l_old->hash_node);
 917         }
 918         ret = 0;
 919
 920 err:
 921         raw_spin_unlock_irqrestore(&b->lock, flags);
 922
 923         if (ret)
 924                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 925         else if (l_old)
 926                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 927
 928         return ret;
 929 }
 930
 931 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 932                                          void *value, u64 map_flags,
 933                                          bool onallcpus)
 934 {
 935         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 936         struct htab_elem *l_new = NULL, *l_old;
 937         struct hlist_nulls_head *head;
 938         unsigned long flags;
 939         struct bucket *b;
 940         u32 key_size, hash;
 941         int ret;
 942
 943         if (unlikely(map_flags > BPF_EXIST))
 944                 /* unknown flags */
 945                 return -EINVAL;
 946
 947         WARN_ON_ONCE(!rcu_read_lock_held());
 948
 949         key_size = map->key_size;
 950
 951         hash = htab_map_hash(key, key_size, htab->hashrnd);
 952
 953         b = __select_bucket(htab, hash);
 954         head = &b->head;
 955
 956         /* bpf_map_update_elem() can be called in_irq() */
 957         raw_spin_lock_irqsave(&b->lock, flags);
 958
 959         l_old = lookup_elem_raw(head, hash, key, key_size);
 960
 961         ret = check_flags(htab, l_old, map_flags);
 962         if (ret)
 963                 goto err;
 964
 965         if (l_old) {
 966                 /* per-cpu hash map can update value in-place */
 967                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 968                                 value, onallcpus);
 969         } else {
 970                 l_new = alloc_htab_elem(htab, key, value, key_size,
 971                                         hash, true, onallcpus, NULL);
 972                 if (IS_ERR(l_new)) {
 973                         ret = PTR_ERR(l_new);
 974                         goto err;
 975                 }
 976                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 977         }
 978         ret = 0;
 979 err:
 980         raw_spin_unlock_irqrestore(&b->lock, flags);
 981         return ret;
 982 }
 983
 984 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 985                                              void *value, u64 map_flags,
 986                                              bool onallcpus)
 987 {
 988         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 989         struct htab_elem *l_new = NULL, *l_old;
 990         struct hlist_nulls_head *head;
 991         unsigned long flags;
 992         struct bucket *b;
 993         u32 key_size, hash;
 994         int ret;
 995
 996         if (unlikely(map_flags > BPF_EXIST))
 997                 /* unknown flags */
 998                 return -EINVAL;
 999
1000         WARN_ON_ONCE(!rcu_read_lock_held());
1001
1002         key_size = map->key_size;
1003
1004         hash = htab_map_hash(key, key_size, htab->hashrnd);
1005
1006         b = __select_bucket(htab, hash);
1007         head = &b->head;
1008
1009         /* For LRU, we need to alloc before taking bucket's
1010          * spinlock because LRU's elem alloc may need
1011          * to remove older elem from htab and this removal
1012          * operation will need a bucket lock.
1013          */
1014         if (map_flags != BPF_EXIST) {
1015                 l_new = prealloc_lru_pop(htab, key, hash);
1016                 if (!l_new)
1017                         return -ENOMEM;
1018         }
1019
1020         /* bpf_map_update_elem() can be called in_irq() */
1021         raw_spin_lock_irqsave(&b->lock, flags);
1022
1023         l_old = lookup_elem_raw(head, hash, key, key_size);
1024
1025         ret = check_flags(htab, l_old, map_flags);
1026         if (ret)
1027                 goto err;
1028
1029         if (l_old) {
1030                 bpf_lru_node_set_ref(&l_old->lru_node);
1031
1032                 /* per-cpu hash map can update value in-place */
1033                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1034                                 value, onallcpus);
1035         } else {
1036                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1037                                 value, onallcpus);
1038                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1039                 l_new = NULL;
1040         }
1041         ret = 0;
1042 err:
1043         raw_spin_unlock_irqrestore(&b->lock, flags);
1044         if (l_new)
1045                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1046         return ret;
1047 }
1048
1049 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1050                                        void *value, u64 map_flags)
1051 {
1052         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1053 }
1054
1055 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1056                                            void *value, u64 map_flags)
1057 {
1058         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1059                                                  false);
1060 }
1061
1062 /* Called from syscall or from eBPF program */
1063 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1064 {
1065         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1066         struct hlist_nulls_head *head;
1067         struct bucket *b;
1068         struct htab_elem *l;
1069         unsigned long flags;
1070         u32 hash, key_size;
1071         int ret = -ENOENT;
1072
1073         WARN_ON_ONCE(!rcu_read_lock_held());
1074
1075         key_size = map->key_size;
1076
1077         hash = htab_map_hash(key, key_size, htab->hashrnd);
1078         b = __select_bucket(htab, hash);
1079         head = &b->head;
1080
1081         raw_spin_lock_irqsave(&b->lock, flags);
1082
1083         l = lookup_elem_raw(head, hash, key, key_size);
1084
1085         if (l) {
1086                 hlist_nulls_del_rcu(&l->hash_node);
1087                 free_htab_elem(htab, l);
1088                 ret = 0;
1089         }
1090
1091         raw_spin_unlock_irqrestore(&b->lock, flags);
1092         return ret;
1093 }
1094
1095 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1096 {
1097         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1098         struct hlist_nulls_head *head;
1099         struct bucket *b;
1100         struct htab_elem *l;
1101         unsigned long flags;
1102         u32 hash, key_size;
1103         int ret = -ENOENT;
1104
1105         WARN_ON_ONCE(!rcu_read_lock_held());
1106
1107         key_size = map->key_size;
1108
1109         hash = htab_map_hash(key, key_size, htab->hashrnd);
1110         b = __select_bucket(htab, hash);
1111         head = &b->head;
1112
1113         raw_spin_lock_irqsave(&b->lock, flags);
1114
1115         l = lookup_elem_raw(head, hash, key, key_size);
1116
1117         if (l) {
1118                 hlist_nulls_del_rcu(&l->hash_node);
1119                 ret = 0;
1120         }
1121
1122         raw_spin_unlock_irqrestore(&b->lock, flags);
1123         if (l)
1124                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1125         return ret;
1126 }
1127
1128 static void delete_all_elements(struct bpf_htab *htab)
1129 {
1130         int i;
1131
1132         for (i = 0; i < htab->n_buckets; i++) {
1133                 struct hlist_nulls_head *head = select_bucket(htab, i);
1134                 struct hlist_nulls_node *n;
1135                 struct htab_elem *l;
1136
1137                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1138                         hlist_nulls_del_rcu(&l->hash_node);
1139                         htab_elem_free(htab, l);
1140                 }
1141         }
1142 }
1143
1144 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1145 static void htab_map_free(struct bpf_map *map)
1146 {
1147         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1148
1149         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1150          * so the programs (can be more than one that used this map) were
1151          * disconnected from events. Wait for outstanding critical sections in
1152          * these programs to complete
1153          */
1154         synchronize_rcu();
1155
1156         /* some of free_htab_elem() callbacks for elements of this map may
1157          * not have executed. Wait for them.
1158          */
1159         rcu_barrier();
1160         if (!htab_is_prealloc(htab))
1161                 delete_all_elements(htab);
1162         else
1163                 prealloc_destroy(htab);
1164
1165         free_percpu(htab->extra_elems);
1166         bpf_map_area_free(htab->buckets);
1167         kfree(htab);
1168 }
1169
1170 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1171                                    struct seq_file *m)
1172 {
1173         void *value;
1174
1175         rcu_read_lock();
1176
1177         value = htab_map_lookup_elem(map, key);
1178         if (!value) {
1179                 rcu_read_unlock();
1180                 return;
1181         }
1182
1183         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1184         seq_puts(m, ": ");
1185         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1186         seq_puts(m, "\n");
1187
1188         rcu_read_unlock();
1189 }
1190
1191 const struct bpf_map_ops htab_map_ops = {
1192         .map_alloc_check = htab_map_alloc_check,
1193         .map_alloc = htab_map_alloc,
1194         .map_free = htab_map_free,
1195         .map_get_next_key = htab_map_get_next_key,
1196         .map_lookup_elem = htab_map_lookup_elem,
1197         .map_update_elem = htab_map_update_elem,
1198         .map_delete_elem = htab_map_delete_elem,
1199         .map_gen_lookup = htab_map_gen_lookup,
1200         .map_seq_show_elem = htab_map_seq_show_elem,
1201 };
1202
1203 const struct bpf_map_ops htab_lru_map_ops = {
1204         .map_alloc_check = htab_map_alloc_check,
1205         .map_alloc = htab_map_alloc,
1206         .map_free = htab_map_free,
1207         .map_get_next_key = htab_map_get_next_key,
1208         .map_lookup_elem = htab_lru_map_lookup_elem,
1209         .map_update_elem = htab_lru_map_update_elem,
1210         .map_delete_elem = htab_lru_map_delete_elem,
1211         .map_gen_lookup = htab_lru_map_gen_lookup,
1212         .map_seq_show_elem = htab_map_seq_show_elem,
1213 };
1214
1215 /* Called from eBPF program */
1216 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1217 {
1218         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1219
1220         if (l)
1221                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1222         else
1223                 return NULL;
1224 }
1225
1226 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1227 {
1228         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1229
1230         if (l) {
1231                 bpf_lru_node_set_ref(&l->lru_node);
1232                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1233         }
1234
1235         return NULL;
1236 }
1237
1238 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1239 {
1240         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1241         struct htab_elem *l;
1242         void __percpu *pptr;
1243         int ret = -ENOENT;
1244         int cpu, off = 0;
1245         u32 size;
1246
1247         /* per_cpu areas are zero-filled and bpf programs can only
1248          * access 'value_size' of them, so copying rounded areas
1249          * will not leak any kernel data
1250          */
1251         size = round_up(map->value_size, 8);
1252         rcu_read_lock();
1253         l = __htab_map_lookup_elem(map, key);
1254         if (!l)
1255                 goto out;
1256         if (htab_is_lru(htab))
1257                 bpf_lru_node_set_ref(&l->lru_node);
1258         pptr = htab_elem_get_ptr(l, map->key_size);
1259         for_each_possible_cpu(cpu) {
1260                 bpf_long_memcpy(value + off,
1261                                 per_cpu_ptr(pptr, cpu), size);
1262                 off += size;
1263         }
1264         ret = 0;
1265 out:
1266         rcu_read_unlock();
1267         return ret;
1268 }
1269
1270 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1271                            u64 map_flags)
1272 {
1273         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1274         int ret;
1275
1276         rcu_read_lock();
1277         if (htab_is_lru(htab))
1278                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1279                                                         map_flags, true);
1280         else
1281                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1282                                                     true);
1283         rcu_read_unlock();
1284
1285         return ret;
1286 }
1287
1288 const struct bpf_map_ops htab_percpu_map_ops = {
1289         .map_alloc_check = htab_map_alloc_check,
1290         .map_alloc = htab_map_alloc,
1291         .map_free = htab_map_free,
1292         .map_get_next_key = htab_map_get_next_key,
1293         .map_lookup_elem = htab_percpu_map_lookup_elem,
1294         .map_update_elem = htab_percpu_map_update_elem,
1295         .map_delete_elem = htab_map_delete_elem,
1296 };
1297
1298 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1299         .map_alloc_check = htab_map_alloc_check,
1300         .map_alloc = htab_map_alloc,
1301         .map_free = htab_map_free,
1302         .map_get_next_key = htab_map_get_next_key,
1303         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1304         .map_update_elem = htab_lru_percpu_map_update_elem,
1305         .map_delete_elem = htab_lru_map_delete_elem,
1306 };
1307
1308 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1309 {
1310         if (attr->value_size != sizeof(u32))
1311                 return -EINVAL;
1312         return htab_map_alloc_check(attr);
1313 }
1314
1315 static void fd_htab_map_free(struct bpf_map *map)
1316 {
1317         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1318         struct hlist_nulls_node *n;
1319         struct hlist_nulls_head *head;
1320         struct htab_elem *l;
1321         int i;
1322
1323         for (i = 0; i < htab->n_buckets; i++) {
1324                 head = select_bucket(htab, i);
1325
1326                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1327                         void *ptr = fd_htab_map_get_ptr(map, l);
1328
1329                         map->ops->map_fd_put_ptr(ptr);
1330                 }
1331         }
1332
1333         htab_map_free(map);
1334 }
1335
1336 /* only called from syscall */
1337 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1338 {
1339         void **ptr;
1340         int ret = 0;
1341
1342         if (!map->ops->map_fd_sys_lookup_elem)
1343                 return -ENOTSUPP;
1344
1345         rcu_read_lock();
1346         ptr = htab_map_lookup_elem(map, key);
1347         if (ptr)
1348                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1349         else
1350                 ret = -ENOENT;
1351         rcu_read_unlock();
1352
1353         return ret;
1354 }
1355
1356 /* only called from syscall */
1357 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1358                                 void *key, void *value, u64 map_flags)
1359 {
1360         void *ptr;
1361         int ret;
1362         u32 ufd = *(u32 *)value;
1363
1364         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1365         if (IS_ERR(ptr))
1366                 return PTR_ERR(ptr);
1367
1368         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1369         if (ret)
1370                 map->ops->map_fd_put_ptr(ptr);
1371
1372         return ret;
1373 }
1374
1375 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1376 {
1377         struct bpf_map *map, *inner_map_meta;
1378
1379         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1380         if (IS_ERR(inner_map_meta))
1381                 return inner_map_meta;
1382
1383         map = htab_map_alloc(attr);
1384         if (IS_ERR(map)) {
1385                 bpf_map_meta_free(inner_map_meta);
1386                 return map;
1387         }
1388
1389         map->inner_map_meta = inner_map_meta;
1390
1391         return map;
1392 }
1393
1394 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1395 {
1396         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1397
1398         if (!inner_map)
1399                 return NULL;
1400
1401         return READ_ONCE(*inner_map);
1402 }
1403
1404 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1405                                   struct bpf_insn *insn_buf)
1406 {
1407         struct bpf_insn *insn = insn_buf;
1408         const int ret = BPF_REG_0;
1409
1410         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1411                      (void *(*)(struct bpf_map *map, void *key))NULL));
1412         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1413         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1414         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1415                                 offsetof(struct htab_elem, key) +
1416                                 round_up(map->key_size, 8));
1417         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1418
1419         return insn - insn_buf;
1420 }
1421
1422 static void htab_of_map_free(struct bpf_map *map)
1423 {
1424         bpf_map_meta_free(map->inner_map_meta);
1425         fd_htab_map_free(map);
1426 }
1427
1428 const struct bpf_map_ops htab_of_maps_map_ops = {
1429         .map_alloc_check = fd_htab_map_alloc_check,
1430         .map_alloc = htab_of_map_alloc,
1431         .map_free = htab_of_map_free,
1432         .map_get_next_key = htab_map_get_next_key,
1433         .map_lookup_elem = htab_of_map_lookup_elem,
1434         .map_delete_elem = htab_map_delete_elem,
1435         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1436         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1437         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1438         .map_gen_lookup = htab_of_map_gen_lookup,
1439         .map_check_btf = map_check_no_btf,
1440 };