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 | BPF_F_ZERO_SEED)
  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         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 248         int numa_node = bpf_map_attr_numa_node(attr);
 249
 250         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 251                      offsetof(struct htab_elem, hash_node.pprev));
 252         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 253                      offsetof(struct htab_elem, hash_node.pprev));
 254
 255         if (lru && !capable(CAP_SYS_ADMIN))
 256                 /* LRU implementation is much complicated than other
 257                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 258                  */
 259                 return -EPERM;
 260
 261         if (zero_seed && !capable(CAP_SYS_ADMIN))
 262                 /* Guard against local DoS, and discourage production use. */
 263                 return -EPERM;
 264
 265         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
 266                 /* reserved bits should not be used */
 267                 return -EINVAL;
 268
 269         if (!lru && percpu_lru)
 270                 return -EINVAL;
 271
 272         if (lru && !prealloc)
 273                 return -ENOTSUPP;
 274
 275         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 276                 return -EINVAL;
 277
 278         /* check sanity of attributes.
 279          * value_size == 0 may be allowed in the future to use map as a set
 280          */
 281         if (attr->max_entries == 0 || attr->key_size == 0 ||
 282             attr->value_size == 0)
 283                 return -EINVAL;
 284
 285         if (attr->key_size > MAX_BPF_STACK)
 286                 /* eBPF programs initialize keys on stack, so they cannot be
 287                  * larger than max stack size
 288                  */
 289                 return -E2BIG;
 290
 291         if (attr->value_size >= KMALLOC_MAX_SIZE -
 292             MAX_BPF_STACK - sizeof(struct htab_elem))
 293                 /* if value_size is bigger, the user space won't be able to
 294                  * access the elements via bpf syscall. This check also makes
 295                  * sure that the elem_size doesn't overflow and it's
 296                  * kmalloc-able later in htab_map_update_elem()
 297                  */
 298                 return -E2BIG;
 299
 300         return 0;
 301 }
 302
 303 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 304 {
 305         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 306                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 307         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 308                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 309         /* percpu_lru means each cpu has its own LRU list.
 310          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 311          * the map's value itself is percpu.  percpu_lru has
 312          * nothing to do with the map's value.
 313          */
 314         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 315         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 316         struct bpf_htab *htab;
 317         int err, i;
 318         u64 cost;
 319
 320         htab = kzalloc(sizeof(*htab), GFP_USER);
 321         if (!htab)
 322                 return ERR_PTR(-ENOMEM);
 323
 324         bpf_map_init_from_attr(&htab->map, attr);
 325
 326         if (percpu_lru) {
 327                 /* ensure each CPU's lru list has >=1 elements.
 328                  * since we are at it, make each lru list has the same
 329                  * number of elements.
 330                  */
 331                 htab->map.max_entries = roundup(attr->max_entries,
 332                                                 num_possible_cpus());
 333                 if (htab->map.max_entries < attr->max_entries)
 334                         htab->map.max_entries = rounddown(attr->max_entries,
 335                                                           num_possible_cpus());
 336         }
 337
 338         /* hash table size must be power of 2 */
 339         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 340
 341         htab->elem_size = sizeof(struct htab_elem) +
 342                           round_up(htab->map.key_size, 8);
 343         if (percpu)
 344                 htab->elem_size += sizeof(void *);
 345         else
 346                 htab->elem_size += round_up(htab->map.value_size, 8);
 347
 348         err = -E2BIG;
 349         /* prevent zero size kmalloc and check for u32 overflow */
 350         if (htab->n_buckets == 0 ||
 351             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 352                 goto free_htab;
 353
 354         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 355                (u64) htab->elem_size * htab->map.max_entries;
 356
 357         if (percpu)
 358                 cost += (u64) round_up(htab->map.value_size, 8) *
 359                         num_possible_cpus() * htab->map.max_entries;
 360         else
 361                cost += (u64) htab->elem_size * num_possible_cpus();
 362
 363         if (cost >= U32_MAX - PAGE_SIZE)
 364                 /* make sure page count doesn't overflow */
 365                 goto free_htab;
 366
 367         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 368
 369         /* if map size is larger than memlock limit, reject it early */
 370         err = bpf_map_precharge_memlock(htab->map.pages);
 371         if (err)
 372                 goto free_htab;
 373
 374         err = -ENOMEM;
 375         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 376                                            sizeof(struct bucket),
 377                                            htab->map.numa_node);
 378         if (!htab->buckets)
 379                 goto free_htab;
 380
 381         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 382                 htab->hashrnd = 0;
 383         else
 384                 htab->hashrnd = get_random_int();
 385
 386         for (i = 0; i < htab->n_buckets; i++) {
 387                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 388                 raw_spin_lock_init(&htab->buckets[i].lock);
 389         }
 390
 391         if (prealloc) {
 392                 err = prealloc_init(htab);
 393                 if (err)
 394                         goto free_buckets;
 395
 396                 if (!percpu && !lru) {
 397                         /* lru itself can remove the least used element, so
 398                          * there is no need for an extra elem during map_update.
 399                          */
 400                         err = alloc_extra_elems(htab);
 401                         if (err)
 402                                 goto free_prealloc;
 403                 }
 404         }
 405
 406         return &htab->map;
 407
 408 free_prealloc:
 409         prealloc_destroy(htab);
 410 free_buckets:
 411         bpf_map_area_free(htab->buckets);
 412 free_htab:
 413         kfree(htab);
 414         return ERR_PTR(err);
 415 }
 416
 417 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 418 {
 419         return jhash(key, key_len, hashrnd);
 420 }
 421
 422 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 423 {
 424         return &htab->buckets[hash & (htab->n_buckets - 1)];
 425 }
 426
 427 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 428 {
 429         return &__select_bucket(htab, hash)->head;
 430 }
 431
 432 /* this lookup function can only be called with bucket lock taken */
 433 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 434                                          void *key, u32 key_size)
 435 {
 436         struct hlist_nulls_node *n;
 437         struct htab_elem *l;
 438
 439         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 440                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 441                         return l;
 442
 443         return NULL;
 444 }
 445
 446 /* can be called without bucket lock. it will repeat the loop in
 447  * the unlikely event when elements moved from one bucket into another
 448  * while link list is being walked
 449  */
 450 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 451                                                u32 hash, void *key,
 452                                                u32 key_size, u32 n_buckets)
 453 {
 454         struct hlist_nulls_node *n;
 455         struct htab_elem *l;
 456
 457 again:
 458         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 459                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 460                         return l;
 461
 462         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 463                 goto again;
 464
 465         return NULL;
 466 }
 467
 468 /* Called from syscall or from eBPF program directly, so
 469  * arguments have to match bpf_map_lookup_elem() exactly.
 470  * The return value is adjusted by BPF instructions
 471  * in htab_map_gen_lookup().
 472  */
 473 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 474 {
 475         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 476         struct hlist_nulls_head *head;
 477         struct htab_elem *l;
 478         u32 hash, key_size;
 479
 480         /* Must be called with rcu_read_lock. */
 481         WARN_ON_ONCE(!rcu_read_lock_held());
 482
 483         key_size = map->key_size;
 484
 485         hash = htab_map_hash(key, key_size, htab->hashrnd);
 486
 487         head = select_bucket(htab, hash);
 488
 489         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 490
 491         return l;
 492 }
 493
 494 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 495 {
 496         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 497
 498         if (l)
 499                 return l->key + round_up(map->key_size, 8);
 500
 501         return NULL;
 502 }
 503
 504 /* inline bpf_map_lookup_elem() call.
 505  * Instead of:
 506  * bpf_prog
 507  *   bpf_map_lookup_elem
 508  *     map->ops->map_lookup_elem
 509  *       htab_map_lookup_elem
 510  *         __htab_map_lookup_elem
 511  * do:
 512  * bpf_prog
 513  *   __htab_map_lookup_elem
 514  */
 515 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 516 {
 517         struct bpf_insn *insn = insn_buf;
 518         const int ret = BPF_REG_0;
 519
 520         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 521                      (void *(*)(struct bpf_map *map, void *key))NULL));
 522         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 523         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 524         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 525                                 offsetof(struct htab_elem, key) +
 526                                 round_up(map->key_size, 8));
 527         return insn - insn_buf;
 528 }
 529
 530 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 531 {
 532         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 533
 534         if (l) {
 535                 bpf_lru_node_set_ref(&l->lru_node);
 536                 return l->key + round_up(map->key_size, 8);
 537         }
 538
 539         return NULL;
 540 }
 541
 542 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 543                                    struct bpf_insn *insn_buf)
 544 {
 545         struct bpf_insn *insn = insn_buf;
 546         const int ret = BPF_REG_0;
 547         const int ref_reg = BPF_REG_1;
 548
 549         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 550                      (void *(*)(struct bpf_map *map, void *key))NULL));
 551         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 552         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 553         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 554                               offsetof(struct htab_elem, lru_node) +
 555                               offsetof(struct bpf_lru_node, ref));
 556         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 557         *insn++ = BPF_ST_MEM(BPF_B, ret,
 558                              offsetof(struct htab_elem, lru_node) +
 559                              offsetof(struct bpf_lru_node, ref),
 560                              1);
 561         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 562                                 offsetof(struct htab_elem, key) +
 563                                 round_up(map->key_size, 8));
 564         return insn - insn_buf;
 565 }
 566
 567 /* It is called from the bpf_lru_list when the LRU needs to delete
 568  * older elements from the htab.
 569  */
 570 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 571 {
 572         struct bpf_htab *htab = (struct bpf_htab *)arg;
 573         struct htab_elem *l = NULL, *tgt_l;
 574         struct hlist_nulls_head *head;
 575         struct hlist_nulls_node *n;
 576         unsigned long flags;
 577         struct bucket *b;
 578
 579         tgt_l = container_of(node, struct htab_elem, lru_node);
 580         b = __select_bucket(htab, tgt_l->hash);
 581         head = &b->head;
 582
 583         raw_spin_lock_irqsave(&b->lock, flags);
 584
 585         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 586                 if (l == tgt_l) {
 587                         hlist_nulls_del_rcu(&l->hash_node);
 588                         break;
 589                 }
 590
 591         raw_spin_unlock_irqrestore(&b->lock, flags);
 592
 593         return l == tgt_l;
 594 }
 595
 596 /* Called from syscall */
 597 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 598 {
 599         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 600         struct hlist_nulls_head *head;
 601         struct htab_elem *l, *next_l;
 602         u32 hash, key_size;
 603         int i = 0;
 604
 605         WARN_ON_ONCE(!rcu_read_lock_held());
 606
 607         key_size = map->key_size;
 608
 609         if (!key)
 610                 goto find_first_elem;
 611
 612         hash = htab_map_hash(key, key_size, htab->hashrnd);
 613
 614         head = select_bucket(htab, hash);
 615
 616         /* lookup the key */
 617         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 618
 619         if (!l)
 620                 goto find_first_elem;
 621
 622         /* key was found, get next key in the same bucket */
 623         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 624                                   struct htab_elem, hash_node);
 625
 626         if (next_l) {
 627                 /* if next elem in this hash list is non-zero, just return it */
 628                 memcpy(next_key, next_l->key, key_size);
 629                 return 0;
 630         }
 631
 632         /* no more elements in this hash list, go to the next bucket */
 633         i = hash & (htab->n_buckets - 1);
 634         i++;
 635
 636 find_first_elem:
 637         /* iterate over buckets */
 638         for (; i < htab->n_buckets; i++) {
 639                 head = select_bucket(htab, i);
 640
 641                 /* pick first element in the bucket */
 642                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 643                                           struct htab_elem, hash_node);
 644                 if (next_l) {
 645                         /* if it's not empty, just return it */
 646                         memcpy(next_key, next_l->key, key_size);
 647                         return 0;
 648                 }
 649         }
 650
 651         /* iterated over all buckets and all elements */
 652         return -ENOENT;
 653 }
 654
 655 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 656 {
 657         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 658                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 659         kfree(l);
 660 }
 661
 662 static void htab_elem_free_rcu(struct rcu_head *head)
 663 {
 664         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 665         struct bpf_htab *htab = l->htab;
 666
 667         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 668          * we're calling kfree, otherwise deadlock is possible if kprobes
 669          * are placed somewhere inside of slub
 670          */
 671         preempt_disable();
 672         __this_cpu_inc(bpf_prog_active);
 673         htab_elem_free(htab, l);
 674         __this_cpu_dec(bpf_prog_active);
 675         preempt_enable();
 676 }
 677
 678 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 679 {
 680         struct bpf_map *map = &htab->map;
 681
 682         if (map->ops->map_fd_put_ptr) {
 683                 void *ptr = fd_htab_map_get_ptr(map, l);
 684
 685                 map->ops->map_fd_put_ptr(ptr);
 686         }
 687
 688         if (htab_is_prealloc(htab)) {
 689                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 690         } else {
 691                 atomic_dec(&htab->count);
 692                 l->htab = htab;
 693                 call_rcu(&l->rcu, htab_elem_free_rcu);
 694         }
 695 }
 696
 697 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 698                             void *value, bool onallcpus)
 699 {
 700         if (!onallcpus) {
 701                 /* copy true value_size bytes */
 702                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 703         } else {
 704                 u32 size = round_up(htab->map.value_size, 8);
 705                 int off = 0, cpu;
 706
 707                 for_each_possible_cpu(cpu) {
 708                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 709                                         value + off, size);
 710                         off += size;
 711                 }
 712         }
 713 }
 714
 715 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 716 {
 717         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 718                BITS_PER_LONG == 64;
 719 }
 720
 721 static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
 722 {
 723         u32 size = htab->map.value_size;
 724
 725         if (percpu || fd_htab_map_needs_adjust(htab))
 726                 size = round_up(size, 8);
 727         return size;
 728 }
 729
 730 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 731                                          void *value, u32 key_size, u32 hash,
 732                                          bool percpu, bool onallcpus,
 733                                          struct htab_elem *old_elem)
 734 {
 735         u32 size = htab_size_value(htab, percpu);
 736         bool prealloc = htab_is_prealloc(htab);
 737         struct htab_elem *l_new, **pl_new;
 738         void __percpu *pptr;
 739
 740         if (prealloc) {
 741                 if (old_elem) {
 742                         /* if we're updating the existing element,
 743                          * use per-cpu extra elems to avoid freelist_pop/push
 744                          */
 745                         pl_new = this_cpu_ptr(htab->extra_elems);
 746                         l_new = *pl_new;
 747                         *pl_new = old_elem;
 748                 } else {
 749                         struct pcpu_freelist_node *l;
 750
 751                         l = __pcpu_freelist_pop(&htab->freelist);
 752                         if (!l)
 753                                 return ERR_PTR(-E2BIG);
 754                         l_new = container_of(l, struct htab_elem, fnode);
 755                 }
 756         } else {
 757                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 758                         if (!old_elem) {
 759                                 /* when map is full and update() is replacing
 760                                  * old element, it's ok to allocate, since
 761                                  * old element will be freed immediately.
 762                                  * Otherwise return an error
 763                                  */
 764                                 l_new = ERR_PTR(-E2BIG);
 765                                 goto dec_count;
 766                         }
 767                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 768                                      htab->map.numa_node);
 769                 if (!l_new) {
 770                         l_new = ERR_PTR(-ENOMEM);
 771                         goto dec_count;
 772                 }
 773         }
 774
 775         memcpy(l_new->key, key, key_size);
 776         if (percpu) {
 777                 if (prealloc) {
 778                         pptr = htab_elem_get_ptr(l_new, key_size);
 779                 } else {
 780                         /* alloc_percpu zero-fills */
 781                         pptr = __alloc_percpu_gfp(size, 8,
 782                                                   GFP_ATOMIC | __GFP_NOWARN);
 783                         if (!pptr) {
 784                                 kfree(l_new);
 785                                 l_new = ERR_PTR(-ENOMEM);
 786                                 goto dec_count;
 787                         }
 788                 }
 789
 790                 pcpu_copy_value(htab, pptr, value, onallcpus);
 791
 792                 if (!prealloc)
 793                         htab_elem_set_ptr(l_new, key_size, pptr);
 794         } else {
 795                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 796         }
 797
 798         l_new->hash = hash;
 799         return l_new;
 800 dec_count:
 801         atomic_dec(&htab->count);
 802         return l_new;
 803 }
 804
 805 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 806                        u64 map_flags)
 807 {
 808         if (l_old && map_flags == BPF_NOEXIST)
 809                 /* elem already exists */
 810                 return -EEXIST;
 811
 812         if (!l_old && map_flags == BPF_EXIST)
 813                 /* elem doesn't exist, cannot update it */
 814                 return -ENOENT;
 815
 816         return 0;
 817 }
 818
 819 /* Called from syscall or from eBPF program */
 820 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 821                                 u64 map_flags)
 822 {
 823         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 824         struct htab_elem *l_new = NULL, *l_old;
 825         struct hlist_nulls_head *head;
 826         unsigned long flags;
 827         struct bucket *b;
 828         u32 key_size, hash;
 829         int ret;
 830
 831         if (unlikely(map_flags > BPF_EXIST))
 832                 /* unknown flags */
 833                 return -EINVAL;
 834
 835         WARN_ON_ONCE(!rcu_read_lock_held());
 836
 837         key_size = map->key_size;
 838
 839         hash = htab_map_hash(key, key_size, htab->hashrnd);
 840
 841         b = __select_bucket(htab, hash);
 842         head = &b->head;
 843
 844         /* bpf_map_update_elem() can be called in_irq() */
 845         raw_spin_lock_irqsave(&b->lock, flags);
 846
 847         l_old = lookup_elem_raw(head, hash, key, key_size);
 848
 849         ret = check_flags(htab, l_old, map_flags);
 850         if (ret)
 851                 goto err;
 852
 853         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 854                                 l_old);
 855         if (IS_ERR(l_new)) {
 856                 /* all pre-allocated elements are in use or memory exhausted */
 857                 ret = PTR_ERR(l_new);
 858                 goto err;
 859         }
 860
 861         /* add new element to the head of the list, so that
 862          * concurrent search will find it before old elem
 863          */
 864         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 865         if (l_old) {
 866                 hlist_nulls_del_rcu(&l_old->hash_node);
 867                 if (!htab_is_prealloc(htab))
 868                         free_htab_elem(htab, l_old);
 869         }
 870         ret = 0;
 871 err:
 872         raw_spin_unlock_irqrestore(&b->lock, flags);
 873         return ret;
 874 }
 875
 876 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 877                                     u64 map_flags)
 878 {
 879         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 880         struct htab_elem *l_new, *l_old = NULL;
 881         struct hlist_nulls_head *head;
 882         unsigned long flags;
 883         struct bucket *b;
 884         u32 key_size, hash;
 885         int ret;
 886
 887         if (unlikely(map_flags > BPF_EXIST))
 888                 /* unknown flags */
 889                 return -EINVAL;
 890
 891         WARN_ON_ONCE(!rcu_read_lock_held());
 892
 893         key_size = map->key_size;
 894
 895         hash = htab_map_hash(key, key_size, htab->hashrnd);
 896
 897         b = __select_bucket(htab, hash);
 898         head = &b->head;
 899
 900         /* For LRU, we need to alloc before taking bucket's
 901          * spinlock because getting free nodes from LRU may need
 902          * to remove older elements from htab and this removal
 903          * operation will need a bucket lock.
 904          */
 905         l_new = prealloc_lru_pop(htab, key, hash);
 906         if (!l_new)
 907                 return -ENOMEM;
 908         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 909
 910         /* bpf_map_update_elem() can be called in_irq() */
 911         raw_spin_lock_irqsave(&b->lock, flags);
 912
 913         l_old = lookup_elem_raw(head, hash, key, key_size);
 914
 915         ret = check_flags(htab, l_old, map_flags);
 916         if (ret)
 917                 goto err;
 918
 919         /* add new element to the head of the list, so that
 920          * concurrent search will find it before old elem
 921          */
 922         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 923         if (l_old) {
 924                 bpf_lru_node_set_ref(&l_new->lru_node);
 925                 hlist_nulls_del_rcu(&l_old->hash_node);
 926         }
 927         ret = 0;
 928
 929 err:
 930         raw_spin_unlock_irqrestore(&b->lock, flags);
 931
 932         if (ret)
 933                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 934         else if (l_old)
 935                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 936
 937         return ret;
 938 }
 939
 940 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 941                                          void *value, u64 map_flags,
 942                                          bool onallcpus)
 943 {
 944         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 945         struct htab_elem *l_new = NULL, *l_old;
 946         struct hlist_nulls_head *head;
 947         unsigned long flags;
 948         struct bucket *b;
 949         u32 key_size, hash;
 950         int ret;
 951
 952         if (unlikely(map_flags > BPF_EXIST))
 953                 /* unknown flags */
 954                 return -EINVAL;
 955
 956         WARN_ON_ONCE(!rcu_read_lock_held());
 957
 958         key_size = map->key_size;
 959
 960         hash = htab_map_hash(key, key_size, htab->hashrnd);
 961
 962         b = __select_bucket(htab, hash);
 963         head = &b->head;
 964
 965         /* bpf_map_update_elem() can be called in_irq() */
 966         raw_spin_lock_irqsave(&b->lock, flags);
 967
 968         l_old = lookup_elem_raw(head, hash, key, key_size);
 969
 970         ret = check_flags(htab, l_old, map_flags);
 971         if (ret)
 972                 goto err;
 973
 974         if (l_old) {
 975                 /* per-cpu hash map can update value in-place */
 976                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 977                                 value, onallcpus);
 978         } else {
 979                 l_new = alloc_htab_elem(htab, key, value, key_size,
 980                                         hash, true, onallcpus, NULL);
 981                 if (IS_ERR(l_new)) {
 982                         ret = PTR_ERR(l_new);
 983                         goto err;
 984                 }
 985                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 986         }
 987         ret = 0;
 988 err:
 989         raw_spin_unlock_irqrestore(&b->lock, flags);
 990         return ret;
 991 }
 992
 993 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 994                                              void *value, u64 map_flags,
 995                                              bool onallcpus)
 996 {
 997         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 998         struct htab_elem *l_new = NULL, *l_old;
 999         struct hlist_nulls_head *head;
1000         unsigned long flags;
1001         struct bucket *b;
1002         u32 key_size, hash;
1003         int ret;
1004
1005         if (unlikely(map_flags > BPF_EXIST))
1006                 /* unknown flags */
1007                 return -EINVAL;
1008
1009         WARN_ON_ONCE(!rcu_read_lock_held());
1010
1011         key_size = map->key_size;
1012
1013         hash = htab_map_hash(key, key_size, htab->hashrnd);
1014
1015         b = __select_bucket(htab, hash);
1016         head = &b->head;
1017
1018         /* For LRU, we need to alloc before taking bucket's
1019          * spinlock because LRU's elem alloc may need
1020          * to remove older elem from htab and this removal
1021          * operation will need a bucket lock.
1022          */
1023         if (map_flags != BPF_EXIST) {
1024                 l_new = prealloc_lru_pop(htab, key, hash);
1025                 if (!l_new)
1026                         return -ENOMEM;
1027         }
1028
1029         /* bpf_map_update_elem() can be called in_irq() */
1030         raw_spin_lock_irqsave(&b->lock, flags);
1031
1032         l_old = lookup_elem_raw(head, hash, key, key_size);
1033
1034         ret = check_flags(htab, l_old, map_flags);
1035         if (ret)
1036                 goto err;
1037
1038         if (l_old) {
1039                 bpf_lru_node_set_ref(&l_old->lru_node);
1040
1041                 /* per-cpu hash map can update value in-place */
1042                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1043                                 value, onallcpus);
1044         } else {
1045                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1046                                 value, onallcpus);
1047                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1048                 l_new = NULL;
1049         }
1050         ret = 0;
1051 err:
1052         raw_spin_unlock_irqrestore(&b->lock, flags);
1053         if (l_new)
1054                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1055         return ret;
1056 }
1057
1058 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1059                                        void *value, u64 map_flags)
1060 {
1061         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1062 }
1063
1064 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1065                                            void *value, u64 map_flags)
1066 {
1067         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1068                                                  false);
1069 }
1070
1071 /* Called from syscall or from eBPF program */
1072 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1073 {
1074         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1075         struct hlist_nulls_head *head;
1076         struct bucket *b;
1077         struct htab_elem *l;
1078         unsigned long flags;
1079         u32 hash, key_size;
1080         int ret = -ENOENT;
1081
1082         WARN_ON_ONCE(!rcu_read_lock_held());
1083
1084         key_size = map->key_size;
1085
1086         hash = htab_map_hash(key, key_size, htab->hashrnd);
1087         b = __select_bucket(htab, hash);
1088         head = &b->head;
1089
1090         raw_spin_lock_irqsave(&b->lock, flags);
1091
1092         l = lookup_elem_raw(head, hash, key, key_size);
1093
1094         if (l) {
1095                 hlist_nulls_del_rcu(&l->hash_node);
1096                 free_htab_elem(htab, l);
1097                 ret = 0;
1098         }
1099
1100         raw_spin_unlock_irqrestore(&b->lock, flags);
1101         return ret;
1102 }
1103
1104 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1105 {
1106         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1107         struct hlist_nulls_head *head;
1108         struct bucket *b;
1109         struct htab_elem *l;
1110         unsigned long flags;
1111         u32 hash, key_size;
1112         int ret = -ENOENT;
1113
1114         WARN_ON_ONCE(!rcu_read_lock_held());
1115
1116         key_size = map->key_size;
1117
1118         hash = htab_map_hash(key, key_size, htab->hashrnd);
1119         b = __select_bucket(htab, hash);
1120         head = &b->head;
1121
1122         raw_spin_lock_irqsave(&b->lock, flags);
1123
1124         l = lookup_elem_raw(head, hash, key, key_size);
1125
1126         if (l) {
1127                 hlist_nulls_del_rcu(&l->hash_node);
1128                 ret = 0;
1129         }
1130
1131         raw_spin_unlock_irqrestore(&b->lock, flags);
1132         if (l)
1133                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1134         return ret;
1135 }
1136
1137 static void delete_all_elements(struct bpf_htab *htab)
1138 {
1139         int i;
1140
1141         for (i = 0; i < htab->n_buckets; i++) {
1142                 struct hlist_nulls_head *head = select_bucket(htab, i);
1143                 struct hlist_nulls_node *n;
1144                 struct htab_elem *l;
1145
1146                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1147                         hlist_nulls_del_rcu(&l->hash_node);
1148                         htab_elem_free(htab, l);
1149                 }
1150         }
1151 }
1152
1153 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1154 static void htab_map_free(struct bpf_map *map)
1155 {
1156         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1157
1158         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1159          * so the programs (can be more than one that used this map) were
1160          * disconnected from events. Wait for outstanding critical sections in
1161          * these programs to complete
1162          */
1163         synchronize_rcu();
1164
1165         /* some of free_htab_elem() callbacks for elements of this map may
1166          * not have executed. Wait for them.
1167          */
1168         rcu_barrier();
1169         if (!htab_is_prealloc(htab))
1170                 delete_all_elements(htab);
1171         else
1172                 prealloc_destroy(htab);
1173
1174         free_percpu(htab->extra_elems);
1175         bpf_map_area_free(htab->buckets);
1176         kfree(htab);
1177 }
1178
1179 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1180                                    struct seq_file *m)
1181 {
1182         void *value;
1183
1184         rcu_read_lock();
1185
1186         value = htab_map_lookup_elem(map, key);
1187         if (!value) {
1188                 rcu_read_unlock();
1189                 return;
1190         }
1191
1192         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1193         seq_puts(m, ": ");
1194         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1195         seq_puts(m, "\n");
1196
1197         rcu_read_unlock();
1198 }
1199
1200 const struct bpf_map_ops htab_map_ops = {
1201         .map_alloc_check = htab_map_alloc_check,
1202         .map_alloc = htab_map_alloc,
1203         .map_free = htab_map_free,
1204         .map_get_next_key = htab_map_get_next_key,
1205         .map_lookup_elem = htab_map_lookup_elem,
1206         .map_update_elem = htab_map_update_elem,
1207         .map_delete_elem = htab_map_delete_elem,
1208         .map_gen_lookup = htab_map_gen_lookup,
1209         .map_seq_show_elem = htab_map_seq_show_elem,
1210 };
1211
1212 const struct bpf_map_ops htab_lru_map_ops = {
1213         .map_alloc_check = htab_map_alloc_check,
1214         .map_alloc = htab_map_alloc,
1215         .map_free = htab_map_free,
1216         .map_get_next_key = htab_map_get_next_key,
1217         .map_lookup_elem = htab_lru_map_lookup_elem,
1218         .map_update_elem = htab_lru_map_update_elem,
1219         .map_delete_elem = htab_lru_map_delete_elem,
1220         .map_gen_lookup = htab_lru_map_gen_lookup,
1221         .map_seq_show_elem = htab_map_seq_show_elem,
1222 };
1223
1224 /* Called from eBPF program */
1225 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1226 {
1227         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1228
1229         if (l)
1230                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1231         else
1232                 return NULL;
1233 }
1234
1235 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1236 {
1237         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1238
1239         if (l) {
1240                 bpf_lru_node_set_ref(&l->lru_node);
1241                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1242         }
1243
1244         return NULL;
1245 }
1246
1247 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1248 {
1249         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1250         struct htab_elem *l;
1251         void __percpu *pptr;
1252         int ret = -ENOENT;
1253         int cpu, off = 0;
1254         u32 size;
1255
1256         /* per_cpu areas are zero-filled and bpf programs can only
1257          * access 'value_size' of them, so copying rounded areas
1258          * will not leak any kernel data
1259          */
1260         size = round_up(map->value_size, 8);
1261         rcu_read_lock();
1262         l = __htab_map_lookup_elem(map, key);
1263         if (!l)
1264                 goto out;
1265         if (htab_is_lru(htab))
1266                 bpf_lru_node_set_ref(&l->lru_node);
1267         pptr = htab_elem_get_ptr(l, map->key_size);
1268         for_each_possible_cpu(cpu) {
1269                 bpf_long_memcpy(value + off,
1270                                 per_cpu_ptr(pptr, cpu), size);
1271                 off += size;
1272         }
1273         ret = 0;
1274 out:
1275         rcu_read_unlock();
1276         return ret;
1277 }
1278
1279 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1280                            u64 map_flags)
1281 {
1282         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1283         int ret;
1284
1285         rcu_read_lock();
1286         if (htab_is_lru(htab))
1287                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1288                                                         map_flags, true);
1289         else
1290                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1291                                                     true);
1292         rcu_read_unlock();
1293
1294         return ret;
1295 }
1296
1297 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1298                                           struct seq_file *m)
1299 {
1300         struct htab_elem *l;
1301         void __percpu *pptr;
1302         int cpu;
1303
1304         rcu_read_lock();
1305
1306         l = __htab_map_lookup_elem(map, key);
1307         if (!l) {
1308                 rcu_read_unlock();
1309                 return;
1310         }
1311
1312         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1313         seq_puts(m, ": {\n");
1314         pptr = htab_elem_get_ptr(l, map->key_size);
1315         for_each_possible_cpu(cpu) {
1316                 seq_printf(m, "\tcpu%d: ", cpu);
1317                 btf_type_seq_show(map->btf, map->btf_value_type_id,
1318                                   per_cpu_ptr(pptr, cpu), m);
1319                 seq_puts(m, "\n");
1320         }
1321         seq_puts(m, "}\n");
1322
1323         rcu_read_unlock();
1324 }
1325
1326 const struct bpf_map_ops htab_percpu_map_ops = {
1327         .map_alloc_check = htab_map_alloc_check,
1328         .map_alloc = htab_map_alloc,
1329         .map_free = htab_map_free,
1330         .map_get_next_key = htab_map_get_next_key,
1331         .map_lookup_elem = htab_percpu_map_lookup_elem,
1332         .map_update_elem = htab_percpu_map_update_elem,
1333         .map_delete_elem = htab_map_delete_elem,
1334         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1335 };
1336
1337 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1338         .map_alloc_check = htab_map_alloc_check,
1339         .map_alloc = htab_map_alloc,
1340         .map_free = htab_map_free,
1341         .map_get_next_key = htab_map_get_next_key,
1342         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1343         .map_update_elem = htab_lru_percpu_map_update_elem,
1344         .map_delete_elem = htab_lru_map_delete_elem,
1345         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1346 };
1347
1348 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1349 {
1350         if (attr->value_size != sizeof(u32))
1351                 return -EINVAL;
1352         return htab_map_alloc_check(attr);
1353 }
1354
1355 static void fd_htab_map_free(struct bpf_map *map)
1356 {
1357         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1358         struct hlist_nulls_node *n;
1359         struct hlist_nulls_head *head;
1360         struct htab_elem *l;
1361         int i;
1362
1363         for (i = 0; i < htab->n_buckets; i++) {
1364                 head = select_bucket(htab, i);
1365
1366                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1367                         void *ptr = fd_htab_map_get_ptr(map, l);
1368
1369                         map->ops->map_fd_put_ptr(ptr);
1370                 }
1371         }
1372
1373         htab_map_free(map);
1374 }
1375
1376 /* only called from syscall */
1377 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1378 {
1379         void **ptr;
1380         int ret = 0;
1381
1382         if (!map->ops->map_fd_sys_lookup_elem)
1383                 return -ENOTSUPP;
1384
1385         rcu_read_lock();
1386         ptr = htab_map_lookup_elem(map, key);
1387         if (ptr)
1388                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1389         else
1390                 ret = -ENOENT;
1391         rcu_read_unlock();
1392
1393         return ret;
1394 }
1395
1396 /* only called from syscall */
1397 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1398                                 void *key, void *value, u64 map_flags)
1399 {
1400         void *ptr;
1401         int ret;
1402         u32 ufd = *(u32 *)value;
1403
1404         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1405         if (IS_ERR(ptr))
1406                 return PTR_ERR(ptr);
1407
1408         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1409         if (ret)
1410                 map->ops->map_fd_put_ptr(ptr);
1411
1412         return ret;
1413 }
1414
1415 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1416 {
1417         struct bpf_map *map, *inner_map_meta;
1418
1419         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1420         if (IS_ERR(inner_map_meta))
1421                 return inner_map_meta;
1422
1423         map = htab_map_alloc(attr);
1424         if (IS_ERR(map)) {
1425                 bpf_map_meta_free(inner_map_meta);
1426                 return map;
1427         }
1428
1429         map->inner_map_meta = inner_map_meta;
1430
1431         return map;
1432 }
1433
1434 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1435 {
1436         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1437
1438         if (!inner_map)
1439                 return NULL;
1440
1441         return READ_ONCE(*inner_map);
1442 }
1443
1444 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1445                                   struct bpf_insn *insn_buf)
1446 {
1447         struct bpf_insn *insn = insn_buf;
1448         const int ret = BPF_REG_0;
1449
1450         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1451                      (void *(*)(struct bpf_map *map, void *key))NULL));
1452         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1453         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1454         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1455                                 offsetof(struct htab_elem, key) +
1456                                 round_up(map->key_size, 8));
1457         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1458
1459         return insn - insn_buf;
1460 }
1461
1462 static void htab_of_map_free(struct bpf_map *map)
1463 {
1464         bpf_map_meta_free(map->inner_map_meta);
1465         fd_htab_map_free(map);
1466 }
1467
1468 const struct bpf_map_ops htab_of_maps_map_ops = {
1469         .map_alloc_check = fd_htab_map_alloc_check,
1470         .map_alloc = htab_of_map_alloc,
1471         .map_free = htab_of_map_free,
1472         .map_get_next_key = htab_map_get_next_key,
1473         .map_lookup_elem = htab_of_map_lookup_elem,
1474         .map_delete_elem = htab_map_delete_elem,
1475         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1476         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1477         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1478         .map_gen_lookup = htab_of_map_gen_lookup,
1479         .map_check_btf = map_check_no_btf,
1480 };