kernel/bpf/hashtab.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   3  * Copyright (c) 2016 Facebook
   4  */
   5 #include <linux/bpf.h>
   6 #include <linux/btf.h>
   7 #include <linux/jhash.h>
   8 #include <linux/filter.h>
   9 #include <linux/rculist_nulls.h>
  10 #include <linux/random.h>
  11 #include <uapi/linux/btf.h>
  12 #include "percpu_freelist.h"
  13 #include "bpf_lru_list.h"
  14 #include "map_in_map.h"
  15
  16 #define HTAB_CREATE_FLAG_MASK                                           \
  17         (BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |    \
  18          BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  19
  20 struct bucket {
  21         struct hlist_nulls_head head;
  22         raw_spinlock_t lock;
  23 };
  24
  25 struct bpf_htab {
  26         struct bpf_map map;
  27         struct bucket *buckets;
  28         void *elems;
  29         union {
  30                 struct pcpu_freelist freelist;
  31                 struct bpf_lru lru;
  32         };
  33         struct htab_elem *__percpu *extra_elems;
  34         atomic_t count; /* number of elements in this hashtable */
  35         u32 n_buckets;  /* number of hash buckets */
  36         u32 elem_size;  /* size of each element in bytes */
  37         u32 hashrnd;
  38 };
  39
  40 /* each htab element is struct htab_elem + key + value */
  41 struct htab_elem {
  42         union {
  43                 struct hlist_nulls_node hash_node;
  44                 struct {
  45                         void *padding;
  46                         union {
  47                                 struct bpf_htab *htab;
  48                                 struct pcpu_freelist_node fnode;
  49                         };
  50                 };
  51         };
  52         union {
  53                 struct rcu_head rcu;
  54                 struct bpf_lru_node lru_node;
  55         };
  56         u32 hash;
  57         char key[0] __aligned(8);
  58 };
  59
  60 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  61
  62 static bool htab_is_lru(const struct bpf_htab *htab)
  63 {
  64         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  65                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  66 }
  67
  68 static bool htab_is_percpu(const struct bpf_htab *htab)
  69 {
  70         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  71                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  72 }
  73
  74 static bool htab_is_prealloc(const struct bpf_htab *htab)
  75 {
  76         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  77 }
  78
  79 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  80                                      void __percpu *pptr)
  81 {
  82         *(void __percpu **)(l->key + key_size) = pptr;
  83 }
  84
  85 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  86 {
  87         return *(void __percpu **)(l->key + key_size);
  88 }
  89
  90 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
  91 {
  92         return *(void **)(l->key + roundup(map->key_size, 8));
  93 }
  94
  95 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
  96 {
  97         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
  98 }
  99
 100 static void htab_free_elems(struct bpf_htab *htab)
 101 {
 102         int i;
 103
 104         if (!htab_is_percpu(htab))
 105                 goto free_elems;
 106
 107         for (i = 0; i < htab->map.max_entries; i++) {
 108                 void __percpu *pptr;
 109
 110                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 111                                          htab->map.key_size);
 112                 free_percpu(pptr);
 113                 cond_resched();
 114         }
 115 free_elems:
 116         bpf_map_area_free(htab->elems);
 117 }
 118
 119 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 120                                           u32 hash)
 121 {
 122         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 123         struct htab_elem *l;
 124
 125         if (node) {
 126                 l = container_of(node, struct htab_elem, lru_node);
 127                 memcpy(l->key, key, htab->map.key_size);
 128                 return l;
 129         }
 130
 131         return NULL;
 132 }
 133
 134 static int prealloc_init(struct bpf_htab *htab)
 135 {
 136         u32 num_entries = htab->map.max_entries;
 137         int err = -ENOMEM, i;
 138
 139         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 140                 num_entries += num_possible_cpus();
 141
 142         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 143                                          htab->map.numa_node);
 144         if (!htab->elems)
 145                 return -ENOMEM;
 146
 147         if (!htab_is_percpu(htab))
 148                 goto skip_percpu_elems;
 149
 150         for (i = 0; i < num_entries; i++) {
 151                 u32 size = round_up(htab->map.value_size, 8);
 152                 void __percpu *pptr;
 153
 154                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 155                 if (!pptr)
 156                         goto free_elems;
 157                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 158                                   pptr);
 159                 cond_resched();
 160         }
 161
 162 skip_percpu_elems:
 163         if (htab_is_lru(htab))
 164                 err = bpf_lru_init(&htab->lru,
 165                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 166                                    offsetof(struct htab_elem, hash) -
 167                                    offsetof(struct htab_elem, lru_node),
 168                                    htab_lru_map_delete_node,
 169                                    htab);
 170         else
 171                 err = pcpu_freelist_init(&htab->freelist);
 172
 173         if (err)
 174                 goto free_elems;
 175
 176         if (htab_is_lru(htab))
 177                 bpf_lru_populate(&htab->lru, htab->elems,
 178                                  offsetof(struct htab_elem, lru_node),
 179                                  htab->elem_size, num_entries);
 180         else
 181                 pcpu_freelist_populate(&htab->freelist,
 182                                        htab->elems + offsetof(struct htab_elem, fnode),
 183                                        htab->elem_size, num_entries);
 184
 185         return 0;
 186
 187 free_elems:
 188         htab_free_elems(htab);
 189         return err;
 190 }
 191
 192 static void prealloc_destroy(struct bpf_htab *htab)
 193 {
 194         htab_free_elems(htab);
 195
 196         if (htab_is_lru(htab))
 197                 bpf_lru_destroy(&htab->lru);
 198         else
 199                 pcpu_freelist_destroy(&htab->freelist);
 200 }
 201
 202 static int alloc_extra_elems(struct bpf_htab *htab)
 203 {
 204         struct htab_elem *__percpu *pptr, *l_new;
 205         struct pcpu_freelist_node *l;
 206         int cpu;
 207
 208         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 209                                   GFP_USER | __GFP_NOWARN);
 210         if (!pptr)
 211                 return -ENOMEM;
 212
 213         for_each_possible_cpu(cpu) {
 214                 l = pcpu_freelist_pop(&htab->freelist);
 215                 /* pop will succeed, since prealloc_init()
 216                  * preallocated extra num_possible_cpus elements
 217                  */
 218                 l_new = container_of(l, struct htab_elem, fnode);
 219                 *per_cpu_ptr(pptr, cpu) = l_new;
 220         }
 221         htab->extra_elems = pptr;
 222         return 0;
 223 }
 224
 225 /* Called from syscall */
 226 static int htab_map_alloc_check(union bpf_attr *attr)
 227 {
 228         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 229                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 230         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 231                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 232         /* percpu_lru means each cpu has its own LRU list.
 233          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 234          * the map's value itself is percpu.  percpu_lru has
 235          * nothing to do with the map's value.
 236          */
 237         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 238         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 239         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 240         int numa_node = bpf_map_attr_numa_node(attr);
 241
 242         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 243                      offsetof(struct htab_elem, hash_node.pprev));
 244         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 245                      offsetof(struct htab_elem, hash_node.pprev));
 246
 247         if (lru && !capable(CAP_SYS_ADMIN))
 248                 /* LRU implementation is much complicated than other
 249                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 250                  */
 251                 return -EPERM;
 252
 253         if (zero_seed && !capable(CAP_SYS_ADMIN))
 254                 /* Guard against local DoS, and discourage production use. */
 255                 return -EPERM;
 256
 257         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 258             !bpf_map_flags_access_ok(attr->map_flags))
 259                 return -EINVAL;
 260
 261         if (!lru && percpu_lru)
 262                 return -EINVAL;
 263
 264         if (lru && !prealloc)
 265                 return -ENOTSUPP;
 266
 267         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 268                 return -EINVAL;
 269
 270         /* check sanity of attributes.
 271          * value_size == 0 may be allowed in the future to use map as a set
 272          */
 273         if (attr->max_entries == 0 || attr->key_size == 0 ||
 274             attr->value_size == 0)
 275                 return -EINVAL;
 276
 277         if (attr->key_size > MAX_BPF_STACK)
 278                 /* eBPF programs initialize keys on stack, so they cannot be
 279                  * larger than max stack size
 280                  */
 281                 return -E2BIG;
 282
 283         if (attr->value_size >= KMALLOC_MAX_SIZE -
 284             MAX_BPF_STACK - sizeof(struct htab_elem))
 285                 /* if value_size is bigger, the user space won't be able to
 286                  * access the elements via bpf syscall. This check also makes
 287                  * sure that the elem_size doesn't overflow and it's
 288                  * kmalloc-able later in htab_map_update_elem()
 289                  */
 290                 return -E2BIG;
 291
 292         return 0;
 293 }
 294
 295 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 296 {
 297         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 298                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 299         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 300                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 301         /* percpu_lru means each cpu has its own LRU list.
 302          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 303          * the map's value itself is percpu.  percpu_lru has
 304          * nothing to do with the map's value.
 305          */
 306         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 307         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 308         struct bpf_htab *htab;
 309         int err, i;
 310         u64 cost;
 311
 312         htab = kzalloc(sizeof(*htab), GFP_USER);
 313         if (!htab)
 314                 return ERR_PTR(-ENOMEM);
 315
 316         bpf_map_init_from_attr(&htab->map, attr);
 317
 318         if (percpu_lru) {
 319                 /* ensure each CPU's lru list has >=1 elements.
 320                  * since we are at it, make each lru list has the same
 321                  * number of elements.
 322                  */
 323                 htab->map.max_entries = roundup(attr->max_entries,
 324                                                 num_possible_cpus());
 325                 if (htab->map.max_entries < attr->max_entries)
 326                         htab->map.max_entries = rounddown(attr->max_entries,
 327                                                           num_possible_cpus());
 328         }
 329
 330         /* hash table size must be power of 2 */
 331         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 332
 333         htab->elem_size = sizeof(struct htab_elem) +
 334                           round_up(htab->map.key_size, 8);
 335         if (percpu)
 336                 htab->elem_size += sizeof(void *);
 337         else
 338                 htab->elem_size += round_up(htab->map.value_size, 8);
 339
 340         err = -E2BIG;
 341         /* prevent zero size kmalloc and check for u32 overflow */
 342         if (htab->n_buckets == 0 ||
 343             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 344                 goto free_htab;
 345
 346         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 347                (u64) htab->elem_size * htab->map.max_entries;
 348
 349         if (percpu)
 350                 cost += (u64) round_up(htab->map.value_size, 8) *
 351                         num_possible_cpus() * htab->map.max_entries;
 352         else
 353                cost += (u64) htab->elem_size * num_possible_cpus();
 354
 355         if (cost >= U32_MAX - PAGE_SIZE)
 356                 /* make sure page count doesn't overflow */
 357                 goto free_htab;
 358
 359         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 360
 361         /* if map size is larger than memlock limit, reject it early */
 362         err = bpf_map_precharge_memlock(htab->map.pages);
 363         if (err)
 364                 goto free_htab;
 365
 366         err = -ENOMEM;
 367         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 368                                            sizeof(struct bucket),
 369                                            htab->map.numa_node);
 370         if (!htab->buckets)
 371                 goto free_htab;
 372
 373         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 374                 htab->hashrnd = 0;
 375         else
 376                 htab->hashrnd = get_random_int();
 377
 378         for (i = 0; i < htab->n_buckets; i++) {
 379                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 380                 raw_spin_lock_init(&htab->buckets[i].lock);
 381         }
 382
 383         if (prealloc) {
 384                 err = prealloc_init(htab);
 385                 if (err)
 386                         goto free_buckets;
 387
 388                 if (!percpu && !lru) {
 389                         /* lru itself can remove the least used element, so
 390                          * there is no need for an extra elem during map_update.
 391                          */
 392                         err = alloc_extra_elems(htab);
 393                         if (err)
 394                                 goto free_prealloc;
 395                 }
 396         }
 397
 398         return &htab->map;
 399
 400 free_prealloc:
 401         prealloc_destroy(htab);
 402 free_buckets:
 403         bpf_map_area_free(htab->buckets);
 404 free_htab:
 405         kfree(htab);
 406         return ERR_PTR(err);
 407 }
 408
 409 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 410 {
 411         return jhash(key, key_len, hashrnd);
 412 }
 413
 414 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 415 {
 416         return &htab->buckets[hash & (htab->n_buckets - 1)];
 417 }
 418
 419 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 420 {
 421         return &__select_bucket(htab, hash)->head;
 422 }
 423
 424 /* this lookup function can only be called with bucket lock taken */
 425 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 426                                          void *key, u32 key_size)
 427 {
 428         struct hlist_nulls_node *n;
 429         struct htab_elem *l;
 430
 431         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 432                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 433                         return l;
 434
 435         return NULL;
 436 }
 437
 438 /* can be called without bucket lock. it will repeat the loop in
 439  * the unlikely event when elements moved from one bucket into another
 440  * while link list is being walked
 441  */
 442 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 443                                                u32 hash, void *key,
 444                                                u32 key_size, u32 n_buckets)
 445 {
 446         struct hlist_nulls_node *n;
 447         struct htab_elem *l;
 448
 449 again:
 450         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 451                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 452                         return l;
 453
 454         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 455                 goto again;
 456
 457         return NULL;
 458 }
 459
 460 /* Called from syscall or from eBPF program directly, so
 461  * arguments have to match bpf_map_lookup_elem() exactly.
 462  * The return value is adjusted by BPF instructions
 463  * in htab_map_gen_lookup().
 464  */
 465 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 466 {
 467         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 468         struct hlist_nulls_head *head;
 469         struct htab_elem *l;
 470         u32 hash, key_size;
 471
 472         /* Must be called with rcu_read_lock. */
 473         WARN_ON_ONCE(!rcu_read_lock_held());
 474
 475         key_size = map->key_size;
 476
 477         hash = htab_map_hash(key, key_size, htab->hashrnd);
 478
 479         head = select_bucket(htab, hash);
 480
 481         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 482
 483         return l;
 484 }
 485
 486 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 487 {
 488         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 489
 490         if (l)
 491                 return l->key + round_up(map->key_size, 8);
 492
 493         return NULL;
 494 }
 495
 496 /* inline bpf_map_lookup_elem() call.
 497  * Instead of:
 498  * bpf_prog
 499  *   bpf_map_lookup_elem
 500  *     map->ops->map_lookup_elem
 501  *       htab_map_lookup_elem
 502  *         __htab_map_lookup_elem
 503  * do:
 504  * bpf_prog
 505  *   __htab_map_lookup_elem
 506  */
 507 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 508 {
 509         struct bpf_insn *insn = insn_buf;
 510         const int ret = BPF_REG_0;
 511
 512         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 513                      (void *(*)(struct bpf_map *map, void *key))NULL));
 514         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 515         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 516         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 517                                 offsetof(struct htab_elem, key) +
 518                                 round_up(map->key_size, 8));
 519         return insn - insn_buf;
 520 }
 521
 522 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 523                                                         void *key, const bool mark)
 524 {
 525         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 526
 527         if (l) {
 528                 if (mark)
 529                         bpf_lru_node_set_ref(&l->lru_node);
 530                 return l->key + round_up(map->key_size, 8);
 531         }
 532
 533         return NULL;
 534 }
 535
 536 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 537 {
 538         return __htab_lru_map_lookup_elem(map, key, true);
 539 }
 540
 541 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 542 {
 543         return __htab_lru_map_lookup_elem(map, key, false);
 544 }
 545
 546 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 547                                    struct bpf_insn *insn_buf)
 548 {
 549         struct bpf_insn *insn = insn_buf;
 550         const int ret = BPF_REG_0;
 551         const int ref_reg = BPF_REG_1;
 552
 553         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 554                      (void *(*)(struct bpf_map *map, void *key))NULL));
 555         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 556         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 557         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 558                               offsetof(struct htab_elem, lru_node) +
 559                               offsetof(struct bpf_lru_node, ref));
 560         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 561         *insn++ = BPF_ST_MEM(BPF_B, ret,
 562                              offsetof(struct htab_elem, lru_node) +
 563                              offsetof(struct bpf_lru_node, ref),
 564                              1);
 565         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 566                                 offsetof(struct htab_elem, key) +
 567                                 round_up(map->key_size, 8));
 568         return insn - insn_buf;
 569 }
 570
 571 /* It is called from the bpf_lru_list when the LRU needs to delete
 572  * older elements from the htab.
 573  */
 574 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 575 {
 576         struct bpf_htab *htab = (struct bpf_htab *)arg;
 577         struct htab_elem *l = NULL, *tgt_l;
 578         struct hlist_nulls_head *head;
 579         struct hlist_nulls_node *n;
 580         unsigned long flags;
 581         struct bucket *b;
 582
 583         tgt_l = container_of(node, struct htab_elem, lru_node);
 584         b = __select_bucket(htab, tgt_l->hash);
 585         head = &b->head;
 586
 587         raw_spin_lock_irqsave(&b->lock, flags);
 588
 589         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 590                 if (l == tgt_l) {
 591                         hlist_nulls_del_rcu(&l->hash_node);
 592                         break;
 593                 }
 594
 595         raw_spin_unlock_irqrestore(&b->lock, flags);
 596
 597         return l == tgt_l;
 598 }
 599
 600 /* Called from syscall */
 601 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 602 {
 603         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 604         struct hlist_nulls_head *head;
 605         struct htab_elem *l, *next_l;
 606         u32 hash, key_size;
 607         int i = 0;
 608
 609         WARN_ON_ONCE(!rcu_read_lock_held());
 610
 611         key_size = map->key_size;
 612
 613         if (!key)
 614                 goto find_first_elem;
 615
 616         hash = htab_map_hash(key, key_size, htab->hashrnd);
 617
 618         head = select_bucket(htab, hash);
 619
 620         /* lookup the key */
 621         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 622
 623         if (!l)
 624                 goto find_first_elem;
 625
 626         /* key was found, get next key in the same bucket */
 627         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 628                                   struct htab_elem, hash_node);
 629
 630         if (next_l) {
 631                 /* if next elem in this hash list is non-zero, just return it */
 632                 memcpy(next_key, next_l->key, key_size);
 633                 return 0;
 634         }
 635
 636         /* no more elements in this hash list, go to the next bucket */
 637         i = hash & (htab->n_buckets - 1);
 638         i++;
 639
 640 find_first_elem:
 641         /* iterate over buckets */
 642         for (; i < htab->n_buckets; i++) {
 643                 head = select_bucket(htab, i);
 644
 645                 /* pick first element in the bucket */
 646                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 647                                           struct htab_elem, hash_node);
 648                 if (next_l) {
 649                         /* if it's not empty, just return it */
 650                         memcpy(next_key, next_l->key, key_size);
 651                         return 0;
 652                 }
 653         }
 654
 655         /* iterated over all buckets and all elements */
 656         return -ENOENT;
 657 }
 658
 659 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 660 {
 661         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 662                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 663         kfree(l);
 664 }
 665
 666 static void htab_elem_free_rcu(struct rcu_head *head)
 667 {
 668         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 669         struct bpf_htab *htab = l->htab;
 670
 671         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 672          * we're calling kfree, otherwise deadlock is possible if kprobes
 673          * are placed somewhere inside of slub
 674          */
 675         preempt_disable();
 676         __this_cpu_inc(bpf_prog_active);
 677         htab_elem_free(htab, l);
 678         __this_cpu_dec(bpf_prog_active);
 679         preempt_enable();
 680 }
 681
 682 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 683 {
 684         struct bpf_map *map = &htab->map;
 685
 686         if (map->ops->map_fd_put_ptr) {
 687                 void *ptr = fd_htab_map_get_ptr(map, l);
 688
 689                 map->ops->map_fd_put_ptr(ptr);
 690         }
 691
 692         if (htab_is_prealloc(htab)) {
 693                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 694         } else {
 695                 atomic_dec(&htab->count);
 696                 l->htab = htab;
 697                 call_rcu(&l->rcu, htab_elem_free_rcu);
 698         }
 699 }
 700
 701 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 702                             void *value, bool onallcpus)
 703 {
 704         if (!onallcpus) {
 705                 /* copy true value_size bytes */
 706                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 707         } else {
 708                 u32 size = round_up(htab->map.value_size, 8);
 709                 int off = 0, cpu;
 710
 711                 for_each_possible_cpu(cpu) {
 712                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 713                                         value + off, size);
 714                         off += size;
 715                 }
 716         }
 717 }
 718
 719 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 720 {
 721         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 722                BITS_PER_LONG == 64;
 723 }
 724
 725 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 726                                          void *value, u32 key_size, u32 hash,
 727                                          bool percpu, bool onallcpus,
 728                                          struct htab_elem *old_elem)
 729 {
 730         u32 size = htab->map.value_size;
 731         bool prealloc = htab_is_prealloc(htab);
 732         struct htab_elem *l_new, **pl_new;
 733         void __percpu *pptr;
 734
 735         if (prealloc) {
 736                 if (old_elem) {
 737                         /* if we're updating the existing element,
 738                          * use per-cpu extra elems to avoid freelist_pop/push
 739                          */
 740                         pl_new = this_cpu_ptr(htab->extra_elems);
 741                         l_new = *pl_new;
 742                         *pl_new = old_elem;
 743                 } else {
 744                         struct pcpu_freelist_node *l;
 745
 746                         l = __pcpu_freelist_pop(&htab->freelist);
 747                         if (!l)
 748                                 return ERR_PTR(-E2BIG);
 749                         l_new = container_of(l, struct htab_elem, fnode);
 750                 }
 751         } else {
 752                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 753                         if (!old_elem) {
 754                                 /* when map is full and update() is replacing
 755                                  * old element, it's ok to allocate, since
 756                                  * old element will be freed immediately.
 757                                  * Otherwise return an error
 758                                  */
 759                                 l_new = ERR_PTR(-E2BIG);
 760                                 goto dec_count;
 761                         }
 762                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 763                                      htab->map.numa_node);
 764                 if (!l_new) {
 765                         l_new = ERR_PTR(-ENOMEM);
 766                         goto dec_count;
 767                 }
 768                 check_and_init_map_lock(&htab->map,
 769                                         l_new->key + round_up(key_size, 8));
 770         }
 771
 772         memcpy(l_new->key, key, key_size);
 773         if (percpu) {
 774                 size = round_up(size, 8);
 775                 if (prealloc) {
 776                         pptr = htab_elem_get_ptr(l_new, key_size);
 777                 } else {
 778                         /* alloc_percpu zero-fills */
 779                         pptr = __alloc_percpu_gfp(size, 8,
 780                                                   GFP_ATOMIC | __GFP_NOWARN);
 781                         if (!pptr) {
 782                                 kfree(l_new);
 783                                 l_new = ERR_PTR(-ENOMEM);
 784                                 goto dec_count;
 785                         }
 786                 }
 787
 788                 pcpu_copy_value(htab, pptr, value, onallcpus);
 789
 790                 if (!prealloc)
 791                         htab_elem_set_ptr(l_new, key_size, pptr);
 792         } else if (fd_htab_map_needs_adjust(htab)) {
 793                 size = round_up(size, 8);
 794                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 795         } else {
 796                 copy_map_value(&htab->map,
 797                                l_new->key + round_up(key_size, 8),
 798                                value);
 799         }
 800
 801         l_new->hash = hash;
 802         return l_new;
 803 dec_count:
 804         atomic_dec(&htab->count);
 805         return l_new;
 806 }
 807
 808 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 809                        u64 map_flags)
 810 {
 811         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 812                 /* elem already exists */
 813                 return -EEXIST;
 814
 815         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 816                 /* elem doesn't exist, cannot update it */
 817                 return -ENOENT;
 818
 819         return 0;
 820 }
 821
 822 /* Called from syscall or from eBPF program */
 823 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 824                                 u64 map_flags)
 825 {
 826         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 827         struct htab_elem *l_new = NULL, *l_old;
 828         struct hlist_nulls_head *head;
 829         unsigned long flags;
 830         struct bucket *b;
 831         u32 key_size, hash;
 832         int ret;
 833
 834         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
 835                 /* unknown flags */
 836                 return -EINVAL;
 837
 838         WARN_ON_ONCE(!rcu_read_lock_held());
 839
 840         key_size = map->key_size;
 841
 842         hash = htab_map_hash(key, key_size, htab->hashrnd);
 843
 844         b = __select_bucket(htab, hash);
 845         head = &b->head;
 846
 847         if (unlikely(map_flags & BPF_F_LOCK)) {
 848                 if (unlikely(!map_value_has_spin_lock(map)))
 849                         return -EINVAL;
 850                 /* find an element without taking the bucket lock */
 851                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
 852                                               htab->n_buckets);
 853                 ret = check_flags(htab, l_old, map_flags);
 854                 if (ret)
 855                         return ret;
 856                 if (l_old) {
 857                         /* grab the element lock and update value in place */
 858                         copy_map_value_locked(map,
 859                                               l_old->key + round_up(key_size, 8),
 860                                               value, false);
 861                         return 0;
 862                 }
 863                 /* fall through, grab the bucket lock and lookup again.
 864                  * 99.9% chance that the element won't be found,
 865                  * but second lookup under lock has to be done.
 866                  */
 867         }
 868
 869         /* bpf_map_update_elem() can be called in_irq() */
 870         raw_spin_lock_irqsave(&b->lock, flags);
 871
 872         l_old = lookup_elem_raw(head, hash, key, key_size);
 873
 874         ret = check_flags(htab, l_old, map_flags);
 875         if (ret)
 876                 goto err;
 877
 878         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
 879                 /* first lookup without the bucket lock didn't find the element,
 880                  * but second lookup with the bucket lock found it.
 881                  * This case is highly unlikely, but has to be dealt with:
 882                  * grab the element lock in addition to the bucket lock
 883                  * and update element in place
 884                  */
 885                 copy_map_value_locked(map,
 886                                       l_old->key + round_up(key_size, 8),
 887                                       value, false);
 888                 ret = 0;
 889                 goto err;
 890         }
 891
 892         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 893                                 l_old);
 894         if (IS_ERR(l_new)) {
 895                 /* all pre-allocated elements are in use or memory exhausted */
 896                 ret = PTR_ERR(l_new);
 897                 goto err;
 898         }
 899
 900         /* add new element to the head of the list, so that
 901          * concurrent search will find it before old elem
 902          */
 903         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 904         if (l_old) {
 905                 hlist_nulls_del_rcu(&l_old->hash_node);
 906                 if (!htab_is_prealloc(htab))
 907                         free_htab_elem(htab, l_old);
 908         }
 909         ret = 0;
 910 err:
 911         raw_spin_unlock_irqrestore(&b->lock, flags);
 912         return ret;
 913 }
 914
 915 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 916                                     u64 map_flags)
 917 {
 918         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 919         struct htab_elem *l_new, *l_old = NULL;
 920         struct hlist_nulls_head *head;
 921         unsigned long flags;
 922         struct bucket *b;
 923         u32 key_size, hash;
 924         int ret;
 925
 926         if (unlikely(map_flags > BPF_EXIST))
 927                 /* unknown flags */
 928                 return -EINVAL;
 929
 930         WARN_ON_ONCE(!rcu_read_lock_held());
 931
 932         key_size = map->key_size;
 933
 934         hash = htab_map_hash(key, key_size, htab->hashrnd);
 935
 936         b = __select_bucket(htab, hash);
 937         head = &b->head;
 938
 939         /* For LRU, we need to alloc before taking bucket's
 940          * spinlock because getting free nodes from LRU may need
 941          * to remove older elements from htab and this removal
 942          * operation will need a bucket lock.
 943          */
 944         l_new = prealloc_lru_pop(htab, key, hash);
 945         if (!l_new)
 946                 return -ENOMEM;
 947         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 948
 949         /* bpf_map_update_elem() can be called in_irq() */
 950         raw_spin_lock_irqsave(&b->lock, flags);
 951
 952         l_old = lookup_elem_raw(head, hash, key, key_size);
 953
 954         ret = check_flags(htab, l_old, map_flags);
 955         if (ret)
 956                 goto err;
 957
 958         /* add new element to the head of the list, so that
 959          * concurrent search will find it before old elem
 960          */
 961         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 962         if (l_old) {
 963                 bpf_lru_node_set_ref(&l_new->lru_node);
 964                 hlist_nulls_del_rcu(&l_old->hash_node);
 965         }
 966         ret = 0;
 967
 968 err:
 969         raw_spin_unlock_irqrestore(&b->lock, flags);
 970
 971         if (ret)
 972                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 973         else if (l_old)
 974                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 975
 976         return ret;
 977 }
 978
 979 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 980                                          void *value, u64 map_flags,
 981                                          bool onallcpus)
 982 {
 983         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 984         struct htab_elem *l_new = NULL, *l_old;
 985         struct hlist_nulls_head *head;
 986         unsigned long flags;
 987         struct bucket *b;
 988         u32 key_size, hash;
 989         int ret;
 990
 991         if (unlikely(map_flags > BPF_EXIST))
 992                 /* unknown flags */
 993                 return -EINVAL;
 994
 995         WARN_ON_ONCE(!rcu_read_lock_held());
 996
 997         key_size = map->key_size;
 998
 999         hash = htab_map_hash(key, key_size, htab->hashrnd);
1000
1001         b = __select_bucket(htab, hash);
1002         head = &b->head;
1003
1004         /* bpf_map_update_elem() can be called in_irq() */
1005         raw_spin_lock_irqsave(&b->lock, flags);
1006
1007         l_old = lookup_elem_raw(head, hash, key, key_size);
1008
1009         ret = check_flags(htab, l_old, map_flags);
1010         if (ret)
1011                 goto err;
1012
1013         if (l_old) {
1014                 /* per-cpu hash map can update value in-place */
1015                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1016                                 value, onallcpus);
1017         } else {
1018                 l_new = alloc_htab_elem(htab, key, value, key_size,
1019                                         hash, true, onallcpus, NULL);
1020                 if (IS_ERR(l_new)) {
1021                         ret = PTR_ERR(l_new);
1022                         goto err;
1023                 }
1024                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1025         }
1026         ret = 0;
1027 err:
1028         raw_spin_unlock_irqrestore(&b->lock, flags);
1029         return ret;
1030 }
1031
1032 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1033                                              void *value, u64 map_flags,
1034                                              bool onallcpus)
1035 {
1036         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1037         struct htab_elem *l_new = NULL, *l_old;
1038         struct hlist_nulls_head *head;
1039         unsigned long flags;
1040         struct bucket *b;
1041         u32 key_size, hash;
1042         int ret;
1043
1044         if (unlikely(map_flags > BPF_EXIST))
1045                 /* unknown flags */
1046                 return -EINVAL;
1047
1048         WARN_ON_ONCE(!rcu_read_lock_held());
1049
1050         key_size = map->key_size;
1051
1052         hash = htab_map_hash(key, key_size, htab->hashrnd);
1053
1054         b = __select_bucket(htab, hash);
1055         head = &b->head;
1056
1057         /* For LRU, we need to alloc before taking bucket's
1058          * spinlock because LRU's elem alloc may need
1059          * to remove older elem from htab and this removal
1060          * operation will need a bucket lock.
1061          */
1062         if (map_flags != BPF_EXIST) {
1063                 l_new = prealloc_lru_pop(htab, key, hash);
1064                 if (!l_new)
1065                         return -ENOMEM;
1066         }
1067
1068         /* bpf_map_update_elem() can be called in_irq() */
1069         raw_spin_lock_irqsave(&b->lock, flags);
1070
1071         l_old = lookup_elem_raw(head, hash, key, key_size);
1072
1073         ret = check_flags(htab, l_old, map_flags);
1074         if (ret)
1075                 goto err;
1076
1077         if (l_old) {
1078                 bpf_lru_node_set_ref(&l_old->lru_node);
1079
1080                 /* per-cpu hash map can update value in-place */
1081                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1082                                 value, onallcpus);
1083         } else {
1084                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1085                                 value, onallcpus);
1086                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1087                 l_new = NULL;
1088         }
1089         ret = 0;
1090 err:
1091         raw_spin_unlock_irqrestore(&b->lock, flags);
1092         if (l_new)
1093                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1094         return ret;
1095 }
1096
1097 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1098                                        void *value, u64 map_flags)
1099 {
1100         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1101 }
1102
1103 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1104                                            void *value, u64 map_flags)
1105 {
1106         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1107                                                  false);
1108 }
1109
1110 /* Called from syscall or from eBPF program */
1111 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1112 {
1113         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1114         struct hlist_nulls_head *head;
1115         struct bucket *b;
1116         struct htab_elem *l;
1117         unsigned long flags;
1118         u32 hash, key_size;
1119         int ret = -ENOENT;
1120
1121         WARN_ON_ONCE(!rcu_read_lock_held());
1122
1123         key_size = map->key_size;
1124
1125         hash = htab_map_hash(key, key_size, htab->hashrnd);
1126         b = __select_bucket(htab, hash);
1127         head = &b->head;
1128
1129         raw_spin_lock_irqsave(&b->lock, flags);
1130
1131         l = lookup_elem_raw(head, hash, key, key_size);
1132
1133         if (l) {
1134                 hlist_nulls_del_rcu(&l->hash_node);
1135                 free_htab_elem(htab, l);
1136                 ret = 0;
1137         }
1138
1139         raw_spin_unlock_irqrestore(&b->lock, flags);
1140         return ret;
1141 }
1142
1143 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1144 {
1145         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1146         struct hlist_nulls_head *head;
1147         struct bucket *b;
1148         struct htab_elem *l;
1149         unsigned long flags;
1150         u32 hash, key_size;
1151         int ret = -ENOENT;
1152
1153         WARN_ON_ONCE(!rcu_read_lock_held());
1154
1155         key_size = map->key_size;
1156
1157         hash = htab_map_hash(key, key_size, htab->hashrnd);
1158         b = __select_bucket(htab, hash);
1159         head = &b->head;
1160
1161         raw_spin_lock_irqsave(&b->lock, flags);
1162
1163         l = lookup_elem_raw(head, hash, key, key_size);
1164
1165         if (l) {
1166                 hlist_nulls_del_rcu(&l->hash_node);
1167                 ret = 0;
1168         }
1169
1170         raw_spin_unlock_irqrestore(&b->lock, flags);
1171         if (l)
1172                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1173         return ret;
1174 }
1175
1176 static void delete_all_elements(struct bpf_htab *htab)
1177 {
1178         int i;
1179
1180         for (i = 0; i < htab->n_buckets; i++) {
1181                 struct hlist_nulls_head *head = select_bucket(htab, i);
1182                 struct hlist_nulls_node *n;
1183                 struct htab_elem *l;
1184
1185                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1186                         hlist_nulls_del_rcu(&l->hash_node);
1187                         htab_elem_free(htab, l);
1188                 }
1189         }
1190 }
1191
1192 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1193 static void htab_map_free(struct bpf_map *map)
1194 {
1195         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1196
1197         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1198          * so the programs (can be more than one that used this map) were
1199          * disconnected from events. Wait for outstanding critical sections in
1200          * these programs to complete
1201          */
1202         synchronize_rcu();
1203
1204         /* some of free_htab_elem() callbacks for elements of this map may
1205          * not have executed. Wait for them.
1206          */
1207         rcu_barrier();
1208         if (!htab_is_prealloc(htab))
1209                 delete_all_elements(htab);
1210         else
1211                 prealloc_destroy(htab);
1212
1213         free_percpu(htab->extra_elems);
1214         bpf_map_area_free(htab->buckets);
1215         kfree(htab);
1216 }
1217
1218 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1219                                    struct seq_file *m)
1220 {
1221         void *value;
1222
1223         rcu_read_lock();
1224
1225         value = htab_map_lookup_elem(map, key);
1226         if (!value) {
1227                 rcu_read_unlock();
1228                 return;
1229         }
1230
1231         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1232         seq_puts(m, ": ");
1233         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1234         seq_puts(m, "\n");
1235
1236         rcu_read_unlock();
1237 }
1238
1239 const struct bpf_map_ops htab_map_ops = {
1240         .map_alloc_check = htab_map_alloc_check,
1241         .map_alloc = htab_map_alloc,
1242         .map_free = htab_map_free,
1243         .map_get_next_key = htab_map_get_next_key,
1244         .map_lookup_elem = htab_map_lookup_elem,
1245         .map_update_elem = htab_map_update_elem,
1246         .map_delete_elem = htab_map_delete_elem,
1247         .map_gen_lookup = htab_map_gen_lookup,
1248         .map_seq_show_elem = htab_map_seq_show_elem,
1249 };
1250
1251 const struct bpf_map_ops htab_lru_map_ops = {
1252         .map_alloc_check = htab_map_alloc_check,
1253         .map_alloc = htab_map_alloc,
1254         .map_free = htab_map_free,
1255         .map_get_next_key = htab_map_get_next_key,
1256         .map_lookup_elem = htab_lru_map_lookup_elem,
1257         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1258         .map_update_elem = htab_lru_map_update_elem,
1259         .map_delete_elem = htab_lru_map_delete_elem,
1260         .map_gen_lookup = htab_lru_map_gen_lookup,
1261         .map_seq_show_elem = htab_map_seq_show_elem,
1262 };
1263
1264 /* Called from eBPF program */
1265 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1266 {
1267         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1268
1269         if (l)
1270                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1271         else
1272                 return NULL;
1273 }
1274
1275 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1276 {
1277         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1278
1279         if (l) {
1280                 bpf_lru_node_set_ref(&l->lru_node);
1281                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1282         }
1283
1284         return NULL;
1285 }
1286
1287 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1288 {
1289         struct htab_elem *l;
1290         void __percpu *pptr;
1291         int ret = -ENOENT;
1292         int cpu, off = 0;
1293         u32 size;
1294
1295         /* per_cpu areas are zero-filled and bpf programs can only
1296          * access 'value_size' of them, so copying rounded areas
1297          * will not leak any kernel data
1298          */
1299         size = round_up(map->value_size, 8);
1300         rcu_read_lock();
1301         l = __htab_map_lookup_elem(map, key);
1302         if (!l)
1303                 goto out;
1304         /* We do not mark LRU map element here in order to not mess up
1305          * eviction heuristics when user space does a map walk.
1306          */
1307         pptr = htab_elem_get_ptr(l, map->key_size);
1308         for_each_possible_cpu(cpu) {
1309                 bpf_long_memcpy(value + off,
1310                                 per_cpu_ptr(pptr, cpu), size);
1311                 off += size;
1312         }
1313         ret = 0;
1314 out:
1315         rcu_read_unlock();
1316         return ret;
1317 }
1318
1319 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1320                            u64 map_flags)
1321 {
1322         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1323         int ret;
1324
1325         rcu_read_lock();
1326         if (htab_is_lru(htab))
1327                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1328                                                         map_flags, true);
1329         else
1330                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1331                                                     true);
1332         rcu_read_unlock();
1333
1334         return ret;
1335 }
1336
1337 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1338                                           struct seq_file *m)
1339 {
1340         struct htab_elem *l;
1341         void __percpu *pptr;
1342         int cpu;
1343
1344         rcu_read_lock();
1345
1346         l = __htab_map_lookup_elem(map, key);
1347         if (!l) {
1348                 rcu_read_unlock();
1349                 return;
1350         }
1351
1352         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1353         seq_puts(m, ": {\n");
1354         pptr = htab_elem_get_ptr(l, map->key_size);
1355         for_each_possible_cpu(cpu) {
1356                 seq_printf(m, "\tcpu%d: ", cpu);
1357                 btf_type_seq_show(map->btf, map->btf_value_type_id,
1358                                   per_cpu_ptr(pptr, cpu), m);
1359                 seq_puts(m, "\n");
1360         }
1361         seq_puts(m, "}\n");
1362
1363         rcu_read_unlock();
1364 }
1365
1366 const struct bpf_map_ops htab_percpu_map_ops = {
1367         .map_alloc_check = htab_map_alloc_check,
1368         .map_alloc = htab_map_alloc,
1369         .map_free = htab_map_free,
1370         .map_get_next_key = htab_map_get_next_key,
1371         .map_lookup_elem = htab_percpu_map_lookup_elem,
1372         .map_update_elem = htab_percpu_map_update_elem,
1373         .map_delete_elem = htab_map_delete_elem,
1374         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1375 };
1376
1377 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1378         .map_alloc_check = htab_map_alloc_check,
1379         .map_alloc = htab_map_alloc,
1380         .map_free = htab_map_free,
1381         .map_get_next_key = htab_map_get_next_key,
1382         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1383         .map_update_elem = htab_lru_percpu_map_update_elem,
1384         .map_delete_elem = htab_lru_map_delete_elem,
1385         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1386 };
1387
1388 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1389 {
1390         if (attr->value_size != sizeof(u32))
1391                 return -EINVAL;
1392         return htab_map_alloc_check(attr);
1393 }
1394
1395 static void fd_htab_map_free(struct bpf_map *map)
1396 {
1397         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1398         struct hlist_nulls_node *n;
1399         struct hlist_nulls_head *head;
1400         struct htab_elem *l;
1401         int i;
1402
1403         for (i = 0; i < htab->n_buckets; i++) {
1404                 head = select_bucket(htab, i);
1405
1406                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1407                         void *ptr = fd_htab_map_get_ptr(map, l);
1408
1409                         map->ops->map_fd_put_ptr(ptr);
1410                 }
1411         }
1412
1413         htab_map_free(map);
1414 }
1415
1416 /* only called from syscall */
1417 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1418 {
1419         void **ptr;
1420         int ret = 0;
1421
1422         if (!map->ops->map_fd_sys_lookup_elem)
1423                 return -ENOTSUPP;
1424
1425         rcu_read_lock();
1426         ptr = htab_map_lookup_elem(map, key);
1427         if (ptr)
1428                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1429         else
1430                 ret = -ENOENT;
1431         rcu_read_unlock();
1432
1433         return ret;
1434 }
1435
1436 /* only called from syscall */
1437 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1438                                 void *key, void *value, u64 map_flags)
1439 {
1440         void *ptr;
1441         int ret;
1442         u32 ufd = *(u32 *)value;
1443
1444         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1445         if (IS_ERR(ptr))
1446                 return PTR_ERR(ptr);
1447
1448         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1449         if (ret)
1450                 map->ops->map_fd_put_ptr(ptr);
1451
1452         return ret;
1453 }
1454
1455 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1456 {
1457         struct bpf_map *map, *inner_map_meta;
1458
1459         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1460         if (IS_ERR(inner_map_meta))
1461                 return inner_map_meta;
1462
1463         map = htab_map_alloc(attr);
1464         if (IS_ERR(map)) {
1465                 bpf_map_meta_free(inner_map_meta);
1466                 return map;
1467         }
1468
1469         map->inner_map_meta = inner_map_meta;
1470
1471         return map;
1472 }
1473
1474 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1475 {
1476         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1477
1478         if (!inner_map)
1479                 return NULL;
1480
1481         return READ_ONCE(*inner_map);
1482 }
1483
1484 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1485                                   struct bpf_insn *insn_buf)
1486 {
1487         struct bpf_insn *insn = insn_buf;
1488         const int ret = BPF_REG_0;
1489
1490         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1491                      (void *(*)(struct bpf_map *map, void *key))NULL));
1492         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1493         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1494         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1495                                 offsetof(struct htab_elem, key) +
1496                                 round_up(map->key_size, 8));
1497         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1498
1499         return insn - insn_buf;
1500 }
1501
1502 static void htab_of_map_free(struct bpf_map *map)
1503 {
1504         bpf_map_meta_free(map->inner_map_meta);
1505         fd_htab_map_free(map);
1506 }
1507
1508 const struct bpf_map_ops htab_of_maps_map_ops = {
1509         .map_alloc_check = fd_htab_map_alloc_check,
1510         .map_alloc = htab_of_map_alloc,
1511         .map_free = htab_of_map_free,
1512         .map_get_next_key = htab_map_get_next_key,
1513         .map_lookup_elem = htab_of_map_lookup_elem,
1514         .map_delete_elem = htab_map_delete_elem,
1515         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1516         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1517         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1518         .map_gen_lookup = htab_of_map_gen_lookup,
1519         .map_check_btf = map_check_no_btf,
1520 };