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 #define BATCH_OPS(_name)                        \
  21         .map_lookup_batch =                     \
  22         _name##_map_lookup_batch,               \
  23         .map_lookup_and_delete_batch =          \
  24         _name##_map_lookup_and_delete_batch,    \
  25         .map_update_batch =                     \
  26         generic_map_update_batch,               \
  27         .map_delete_batch =                     \
  28         generic_map_delete_batch
  29
  30 struct bucket {
  31         struct hlist_nulls_head head;
  32         raw_spinlock_t lock;
  33 };
  34
  35 struct bpf_htab {
  36         struct bpf_map map;
  37         struct bucket *buckets;
  38         void *elems;
  39         union {
  40                 struct pcpu_freelist freelist;
  41                 struct bpf_lru lru;
  42         };
  43         struct htab_elem *__percpu *extra_elems;
  44         atomic_t count; /* number of elements in this hashtable */
  45         u32 n_buckets;  /* number of hash buckets */
  46         u32 elem_size;  /* size of each element in bytes */
  47         u32 hashrnd;
  48 };
  49
  50 /* each htab element is struct htab_elem + key + value */
  51 struct htab_elem {
  52         union {
  53                 struct hlist_nulls_node hash_node;
  54                 struct {
  55                         void *padding;
  56                         union {
  57                                 struct bpf_htab *htab;
  58                                 struct pcpu_freelist_node fnode;
  59                         };
  60                 };
  61         };
  62         union {
  63                 struct rcu_head rcu;
  64                 struct bpf_lru_node lru_node;
  65         };
  66         u32 hash;
  67         char key[0] __aligned(8);
  68 };
  69
  70 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  71
  72 static bool htab_is_lru(const struct bpf_htab *htab)
  73 {
  74         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  75                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  76 }
  77
  78 static bool htab_is_percpu(const struct bpf_htab *htab)
  79 {
  80         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  81                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  82 }
  83
  84 static bool htab_is_prealloc(const struct bpf_htab *htab)
  85 {
  86         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  87 }
  88
  89 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  90                                      void __percpu *pptr)
  91 {
  92         *(void __percpu **)(l->key + key_size) = pptr;
  93 }
  94
  95 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  96 {
  97         return *(void __percpu **)(l->key + key_size);
  98 }
  99
 100 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 101 {
 102         return *(void **)(l->key + roundup(map->key_size, 8));
 103 }
 104
 105 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 106 {
 107         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 108 }
 109
 110 static void htab_free_elems(struct bpf_htab *htab)
 111 {
 112         int i;
 113
 114         if (!htab_is_percpu(htab))
 115                 goto free_elems;
 116
 117         for (i = 0; i < htab->map.max_entries; i++) {
 118                 void __percpu *pptr;
 119
 120                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 121                                          htab->map.key_size);
 122                 free_percpu(pptr);
 123                 cond_resched();
 124         }
 125 free_elems:
 126         bpf_map_area_free(htab->elems);
 127 }
 128
 129 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 130                                           u32 hash)
 131 {
 132         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 133         struct htab_elem *l;
 134
 135         if (node) {
 136                 l = container_of(node, struct htab_elem, lru_node);
 137                 memcpy(l->key, key, htab->map.key_size);
 138                 return l;
 139         }
 140
 141         return NULL;
 142 }
 143
 144 static int prealloc_init(struct bpf_htab *htab)
 145 {
 146         u32 num_entries = htab->map.max_entries;
 147         int err = -ENOMEM, i;
 148
 149         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 150                 num_entries += num_possible_cpus();
 151
 152         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 153                                          htab->map.numa_node);
 154         if (!htab->elems)
 155                 return -ENOMEM;
 156
 157         if (!htab_is_percpu(htab))
 158                 goto skip_percpu_elems;
 159
 160         for (i = 0; i < num_entries; i++) {
 161                 u32 size = round_up(htab->map.value_size, 8);
 162                 void __percpu *pptr;
 163
 164                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 165                 if (!pptr)
 166                         goto free_elems;
 167                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 168                                   pptr);
 169                 cond_resched();
 170         }
 171
 172 skip_percpu_elems:
 173         if (htab_is_lru(htab))
 174                 err = bpf_lru_init(&htab->lru,
 175                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 176                                    offsetof(struct htab_elem, hash) -
 177                                    offsetof(struct htab_elem, lru_node),
 178                                    htab_lru_map_delete_node,
 179                                    htab);
 180         else
 181                 err = pcpu_freelist_init(&htab->freelist);
 182
 183         if (err)
 184                 goto free_elems;
 185
 186         if (htab_is_lru(htab))
 187                 bpf_lru_populate(&htab->lru, htab->elems,
 188                                  offsetof(struct htab_elem, lru_node),
 189                                  htab->elem_size, num_entries);
 190         else
 191                 pcpu_freelist_populate(&htab->freelist,
 192                                        htab->elems + offsetof(struct htab_elem, fnode),
 193                                        htab->elem_size, num_entries);
 194
 195         return 0;
 196
 197 free_elems:
 198         htab_free_elems(htab);
 199         return err;
 200 }
 201
 202 static void prealloc_destroy(struct bpf_htab *htab)
 203 {
 204         htab_free_elems(htab);
 205
 206         if (htab_is_lru(htab))
 207                 bpf_lru_destroy(&htab->lru);
 208         else
 209                 pcpu_freelist_destroy(&htab->freelist);
 210 }
 211
 212 static int alloc_extra_elems(struct bpf_htab *htab)
 213 {
 214         struct htab_elem *__percpu *pptr, *l_new;
 215         struct pcpu_freelist_node *l;
 216         int cpu;
 217
 218         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 219                                   GFP_USER | __GFP_NOWARN);
 220         if (!pptr)
 221                 return -ENOMEM;
 222
 223         for_each_possible_cpu(cpu) {
 224                 l = pcpu_freelist_pop(&htab->freelist);
 225                 /* pop will succeed, since prealloc_init()
 226                  * preallocated extra num_possible_cpus elements
 227                  */
 228                 l_new = container_of(l, struct htab_elem, fnode);
 229                 *per_cpu_ptr(pptr, cpu) = l_new;
 230         }
 231         htab->extra_elems = pptr;
 232         return 0;
 233 }
 234
 235 /* Called from syscall */
 236 static int htab_map_alloc_check(union bpf_attr *attr)
 237 {
 238         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 239                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 240         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 241                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 242         /* percpu_lru means each cpu has its own LRU list.
 243          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 244          * the map's value itself is percpu.  percpu_lru has
 245          * nothing to do with the map's value.
 246          */
 247         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 248         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 249         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 250         int numa_node = bpf_map_attr_numa_node(attr);
 251
 252         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 253                      offsetof(struct htab_elem, hash_node.pprev));
 254         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 255                      offsetof(struct htab_elem, hash_node.pprev));
 256
 257         if (lru && !capable(CAP_SYS_ADMIN))
 258                 /* LRU implementation is much complicated than other
 259                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 260                  */
 261                 return -EPERM;
 262
 263         if (zero_seed && !capable(CAP_SYS_ADMIN))
 264                 /* Guard against local DoS, and discourage production use. */
 265                 return -EPERM;
 266
 267         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 268             !bpf_map_flags_access_ok(attr->map_flags))
 269                 return -EINVAL;
 270
 271         if (!lru && percpu_lru)
 272                 return -EINVAL;
 273
 274         if (lru && !prealloc)
 275                 return -ENOTSUPP;
 276
 277         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 278                 return -EINVAL;
 279
 280         /* check sanity of attributes.
 281          * value_size == 0 may be allowed in the future to use map as a set
 282          */
 283         if (attr->max_entries == 0 || attr->key_size == 0 ||
 284             attr->value_size == 0)
 285                 return -EINVAL;
 286
 287         if (attr->key_size > MAX_BPF_STACK)
 288                 /* eBPF programs initialize keys on stack, so they cannot be
 289                  * larger than max stack size
 290                  */
 291                 return -E2BIG;
 292
 293         if (attr->value_size >= KMALLOC_MAX_SIZE -
 294             MAX_BPF_STACK - sizeof(struct htab_elem))
 295                 /* if value_size is bigger, the user space won't be able to
 296                  * access the elements via bpf syscall. This check also makes
 297                  * sure that the elem_size doesn't overflow and it's
 298                  * kmalloc-able later in htab_map_update_elem()
 299                  */
 300                 return -E2BIG;
 301
 302         return 0;
 303 }
 304
 305 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 306 {
 307         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 308                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 309         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 310                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 311         /* percpu_lru means each cpu has its own LRU list.
 312          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 313          * the map's value itself is percpu.  percpu_lru has
 314          * nothing to do with the map's value.
 315          */
 316         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 317         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 318         struct bpf_htab *htab;
 319         int err, i;
 320         u64 cost;
 321
 322         htab = kzalloc(sizeof(*htab), GFP_USER);
 323         if (!htab)
 324                 return ERR_PTR(-ENOMEM);
 325
 326         bpf_map_init_from_attr(&htab->map, attr);
 327
 328         if (percpu_lru) {
 329                 /* ensure each CPU's lru list has >=1 elements.
 330                  * since we are at it, make each lru list has the same
 331                  * number of elements.
 332                  */
 333                 htab->map.max_entries = roundup(attr->max_entries,
 334                                                 num_possible_cpus());
 335                 if (htab->map.max_entries < attr->max_entries)
 336                         htab->map.max_entries = rounddown(attr->max_entries,
 337                                                           num_possible_cpus());
 338         }
 339
 340         /* hash table size must be power of 2 */
 341         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 342
 343         htab->elem_size = sizeof(struct htab_elem) +
 344                           round_up(htab->map.key_size, 8);
 345         if (percpu)
 346                 htab->elem_size += sizeof(void *);
 347         else
 348                 htab->elem_size += round_up(htab->map.value_size, 8);
 349
 350         err = -E2BIG;
 351         /* prevent zero size kmalloc and check for u32 overflow */
 352         if (htab->n_buckets == 0 ||
 353             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 354                 goto free_htab;
 355
 356         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 357                (u64) htab->elem_size * htab->map.max_entries;
 358
 359         if (percpu)
 360                 cost += (u64) round_up(htab->map.value_size, 8) *
 361                         num_possible_cpus() * htab->map.max_entries;
 362         else
 363                cost += (u64) htab->elem_size * num_possible_cpus();
 364
 365         /* if map size is larger than memlock limit, reject it */
 366         err = bpf_map_charge_init(&htab->map.memory, cost);
 367         if (err)
 368                 goto free_htab;
 369
 370         err = -ENOMEM;
 371         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 372                                            sizeof(struct bucket),
 373                                            htab->map.numa_node);
 374         if (!htab->buckets)
 375                 goto free_charge;
 376
 377         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 378                 htab->hashrnd = 0;
 379         else
 380                 htab->hashrnd = get_random_int();
 381
 382         for (i = 0; i < htab->n_buckets; i++) {
 383                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 384                 raw_spin_lock_init(&htab->buckets[i].lock);
 385         }
 386
 387         if (prealloc) {
 388                 err = prealloc_init(htab);
 389                 if (err)
 390                         goto free_buckets;
 391
 392                 if (!percpu && !lru) {
 393                         /* lru itself can remove the least used element, so
 394                          * there is no need for an extra elem during map_update.
 395                          */
 396                         err = alloc_extra_elems(htab);
 397                         if (err)
 398                                 goto free_prealloc;
 399                 }
 400         }
 401
 402         return &htab->map;
 403
 404 free_prealloc:
 405         prealloc_destroy(htab);
 406 free_buckets:
 407         bpf_map_area_free(htab->buckets);
 408 free_charge:
 409         bpf_map_charge_finish(&htab->map.memory);
 410 free_htab:
 411         kfree(htab);
 412         return ERR_PTR(err);
 413 }
 414
 415 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 416 {
 417         return jhash(key, key_len, hashrnd);
 418 }
 419
 420 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 421 {
 422         return &htab->buckets[hash & (htab->n_buckets - 1)];
 423 }
 424
 425 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 426 {
 427         return &__select_bucket(htab, hash)->head;
 428 }
 429
 430 /* this lookup function can only be called with bucket lock taken */
 431 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 432                                          void *key, u32 key_size)
 433 {
 434         struct hlist_nulls_node *n;
 435         struct htab_elem *l;
 436
 437         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 438                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 439                         return l;
 440
 441         return NULL;
 442 }
 443
 444 /* can be called without bucket lock. it will repeat the loop in
 445  * the unlikely event when elements moved from one bucket into another
 446  * while link list is being walked
 447  */
 448 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 449                                                u32 hash, void *key,
 450                                                u32 key_size, u32 n_buckets)
 451 {
 452         struct hlist_nulls_node *n;
 453         struct htab_elem *l;
 454
 455 again:
 456         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 457                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 458                         return l;
 459
 460         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 461                 goto again;
 462
 463         return NULL;
 464 }
 465
 466 /* Called from syscall or from eBPF program directly, so
 467  * arguments have to match bpf_map_lookup_elem() exactly.
 468  * The return value is adjusted by BPF instructions
 469  * in htab_map_gen_lookup().
 470  */
 471 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 472 {
 473         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 474         struct hlist_nulls_head *head;
 475         struct htab_elem *l;
 476         u32 hash, key_size;
 477
 478         /* Must be called with rcu_read_lock. */
 479         WARN_ON_ONCE(!rcu_read_lock_held());
 480
 481         key_size = map->key_size;
 482
 483         hash = htab_map_hash(key, key_size, htab->hashrnd);
 484
 485         head = select_bucket(htab, hash);
 486
 487         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 488
 489         return l;
 490 }
 491
 492 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 493 {
 494         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 495
 496         if (l)
 497                 return l->key + round_up(map->key_size, 8);
 498
 499         return NULL;
 500 }
 501
 502 /* inline bpf_map_lookup_elem() call.
 503  * Instead of:
 504  * bpf_prog
 505  *   bpf_map_lookup_elem
 506  *     map->ops->map_lookup_elem
 507  *       htab_map_lookup_elem
 508  *         __htab_map_lookup_elem
 509  * do:
 510  * bpf_prog
 511  *   __htab_map_lookup_elem
 512  */
 513 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 514 {
 515         struct bpf_insn *insn = insn_buf;
 516         const int ret = BPF_REG_0;
 517
 518         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 519                      (void *(*)(struct bpf_map *map, void *key))NULL));
 520         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 521         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 522         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 523                                 offsetof(struct htab_elem, key) +
 524                                 round_up(map->key_size, 8));
 525         return insn - insn_buf;
 526 }
 527
 528 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 529                                                         void *key, const bool mark)
 530 {
 531         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 532
 533         if (l) {
 534                 if (mark)
 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 void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 543 {
 544         return __htab_lru_map_lookup_elem(map, key, true);
 545 }
 546
 547 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 548 {
 549         return __htab_lru_map_lookup_elem(map, key, false);
 550 }
 551
 552 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 553                                    struct bpf_insn *insn_buf)
 554 {
 555         struct bpf_insn *insn = insn_buf;
 556         const int ret = BPF_REG_0;
 557         const int ref_reg = BPF_REG_1;
 558
 559         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 560                      (void *(*)(struct bpf_map *map, void *key))NULL));
 561         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 562         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 563         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 564                               offsetof(struct htab_elem, lru_node) +
 565                               offsetof(struct bpf_lru_node, ref));
 566         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 567         *insn++ = BPF_ST_MEM(BPF_B, ret,
 568                              offsetof(struct htab_elem, lru_node) +
 569                              offsetof(struct bpf_lru_node, ref),
 570                              1);
 571         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 572                                 offsetof(struct htab_elem, key) +
 573                                 round_up(map->key_size, 8));
 574         return insn - insn_buf;
 575 }
 576
 577 /* It is called from the bpf_lru_list when the LRU needs to delete
 578  * older elements from the htab.
 579  */
 580 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 581 {
 582         struct bpf_htab *htab = (struct bpf_htab *)arg;
 583         struct htab_elem *l = NULL, *tgt_l;
 584         struct hlist_nulls_head *head;
 585         struct hlist_nulls_node *n;
 586         unsigned long flags;
 587         struct bucket *b;
 588
 589         tgt_l = container_of(node, struct htab_elem, lru_node);
 590         b = __select_bucket(htab, tgt_l->hash);
 591         head = &b->head;
 592
 593         raw_spin_lock_irqsave(&b->lock, flags);
 594
 595         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 596                 if (l == tgt_l) {
 597                         hlist_nulls_del_rcu(&l->hash_node);
 598                         break;
 599                 }
 600
 601         raw_spin_unlock_irqrestore(&b->lock, flags);
 602
 603         return l == tgt_l;
 604 }
 605
 606 /* Called from syscall */
 607 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 608 {
 609         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 610         struct hlist_nulls_head *head;
 611         struct htab_elem *l, *next_l;
 612         u32 hash, key_size;
 613         int i = 0;
 614
 615         WARN_ON_ONCE(!rcu_read_lock_held());
 616
 617         key_size = map->key_size;
 618
 619         if (!key)
 620                 goto find_first_elem;
 621
 622         hash = htab_map_hash(key, key_size, htab->hashrnd);
 623
 624         head = select_bucket(htab, hash);
 625
 626         /* lookup the key */
 627         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 628
 629         if (!l)
 630                 goto find_first_elem;
 631
 632         /* key was found, get next key in the same bucket */
 633         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 634                                   struct htab_elem, hash_node);
 635
 636         if (next_l) {
 637                 /* if next elem in this hash list is non-zero, just return it */
 638                 memcpy(next_key, next_l->key, key_size);
 639                 return 0;
 640         }
 641
 642         /* no more elements in this hash list, go to the next bucket */
 643         i = hash & (htab->n_buckets - 1);
 644         i++;
 645
 646 find_first_elem:
 647         /* iterate over buckets */
 648         for (; i < htab->n_buckets; i++) {
 649                 head = select_bucket(htab, i);
 650
 651                 /* pick first element in the bucket */
 652                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 653                                           struct htab_elem, hash_node);
 654                 if (next_l) {
 655                         /* if it's not empty, just return it */
 656                         memcpy(next_key, next_l->key, key_size);
 657                         return 0;
 658                 }
 659         }
 660
 661         /* iterated over all buckets and all elements */
 662         return -ENOENT;
 663 }
 664
 665 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 666 {
 667         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 668                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 669         kfree(l);
 670 }
 671
 672 static void htab_elem_free_rcu(struct rcu_head *head)
 673 {
 674         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 675         struct bpf_htab *htab = l->htab;
 676
 677         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 678          * we're calling kfree, otherwise deadlock is possible if kprobes
 679          * are placed somewhere inside of slub
 680          */
 681         preempt_disable();
 682         __this_cpu_inc(bpf_prog_active);
 683         htab_elem_free(htab, l);
 684         __this_cpu_dec(bpf_prog_active);
 685         preempt_enable();
 686 }
 687
 688 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 689 {
 690         struct bpf_map *map = &htab->map;
 691
 692         if (map->ops->map_fd_put_ptr) {
 693                 void *ptr = fd_htab_map_get_ptr(map, l);
 694
 695                 map->ops->map_fd_put_ptr(ptr);
 696         }
 697
 698         if (htab_is_prealloc(htab)) {
 699                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 700         } else {
 701                 atomic_dec(&htab->count);
 702                 l->htab = htab;
 703                 call_rcu(&l->rcu, htab_elem_free_rcu);
 704         }
 705 }
 706
 707 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 708                             void *value, bool onallcpus)
 709 {
 710         if (!onallcpus) {
 711                 /* copy true value_size bytes */
 712                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 713         } else {
 714                 u32 size = round_up(htab->map.value_size, 8);
 715                 int off = 0, cpu;
 716
 717                 for_each_possible_cpu(cpu) {
 718                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 719                                         value + off, size);
 720                         off += size;
 721                 }
 722         }
 723 }
 724
 725 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 726 {
 727         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 728                BITS_PER_LONG == 64;
 729 }
 730
 731 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 732                                          void *value, u32 key_size, u32 hash,
 733                                          bool percpu, bool onallcpus,
 734                                          struct htab_elem *old_elem)
 735 {
 736         u32 size = htab->map.value_size;
 737         bool prealloc = htab_is_prealloc(htab);
 738         struct htab_elem *l_new, **pl_new;
 739         void __percpu *pptr;
 740
 741         if (prealloc) {
 742                 if (old_elem) {
 743                         /* if we're updating the existing element,
 744                          * use per-cpu extra elems to avoid freelist_pop/push
 745                          */
 746                         pl_new = this_cpu_ptr(htab->extra_elems);
 747                         l_new = *pl_new;
 748                         *pl_new = old_elem;
 749                 } else {
 750                         struct pcpu_freelist_node *l;
 751
 752                         l = __pcpu_freelist_pop(&htab->freelist);
 753                         if (!l)
 754                                 return ERR_PTR(-E2BIG);
 755                         l_new = container_of(l, struct htab_elem, fnode);
 756                 }
 757         } else {
 758                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 759                         if (!old_elem) {
 760                                 /* when map is full and update() is replacing
 761                                  * old element, it's ok to allocate, since
 762                                  * old element will be freed immediately.
 763                                  * Otherwise return an error
 764                                  */
 765                                 l_new = ERR_PTR(-E2BIG);
 766                                 goto dec_count;
 767                         }
 768                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 769                                      htab->map.numa_node);
 770                 if (!l_new) {
 771                         l_new = ERR_PTR(-ENOMEM);
 772                         goto dec_count;
 773                 }
 774                 check_and_init_map_lock(&htab->map,
 775                                         l_new->key + round_up(key_size, 8));
 776         }
 777
 778         memcpy(l_new->key, key, key_size);
 779         if (percpu) {
 780                 size = round_up(size, 8);
 781                 if (prealloc) {
 782                         pptr = htab_elem_get_ptr(l_new, key_size);
 783                 } else {
 784                         /* alloc_percpu zero-fills */
 785                         pptr = __alloc_percpu_gfp(size, 8,
 786                                                   GFP_ATOMIC | __GFP_NOWARN);
 787                         if (!pptr) {
 788                                 kfree(l_new);
 789                                 l_new = ERR_PTR(-ENOMEM);
 790                                 goto dec_count;
 791                         }
 792                 }
 793
 794                 pcpu_copy_value(htab, pptr, value, onallcpus);
 795
 796                 if (!prealloc)
 797                         htab_elem_set_ptr(l_new, key_size, pptr);
 798         } else if (fd_htab_map_needs_adjust(htab)) {
 799                 size = round_up(size, 8);
 800                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 801         } else {
 802                 copy_map_value(&htab->map,
 803                                l_new->key + round_up(key_size, 8),
 804                                value);
 805         }
 806
 807         l_new->hash = hash;
 808         return l_new;
 809 dec_count:
 810         atomic_dec(&htab->count);
 811         return l_new;
 812 }
 813
 814 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 815                        u64 map_flags)
 816 {
 817         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 818                 /* elem already exists */
 819                 return -EEXIST;
 820
 821         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 822                 /* elem doesn't exist, cannot update it */
 823                 return -ENOENT;
 824
 825         return 0;
 826 }
 827
 828 /* Called from syscall or from eBPF program */
 829 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 830                                 u64 map_flags)
 831 {
 832         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 833         struct htab_elem *l_new = NULL, *l_old;
 834         struct hlist_nulls_head *head;
 835         unsigned long flags;
 836         struct bucket *b;
 837         u32 key_size, hash;
 838         int ret;
 839
 840         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
 841                 /* unknown flags */
 842                 return -EINVAL;
 843
 844         WARN_ON_ONCE(!rcu_read_lock_held());
 845
 846         key_size = map->key_size;
 847
 848         hash = htab_map_hash(key, key_size, htab->hashrnd);
 849
 850         b = __select_bucket(htab, hash);
 851         head = &b->head;
 852
 853         if (unlikely(map_flags & BPF_F_LOCK)) {
 854                 if (unlikely(!map_value_has_spin_lock(map)))
 855                         return -EINVAL;
 856                 /* find an element without taking the bucket lock */
 857                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
 858                                               htab->n_buckets);
 859                 ret = check_flags(htab, l_old, map_flags);
 860                 if (ret)
 861                         return ret;
 862                 if (l_old) {
 863                         /* grab the element lock and update value in place */
 864                         copy_map_value_locked(map,
 865                                               l_old->key + round_up(key_size, 8),
 866                                               value, false);
 867                         return 0;
 868                 }
 869                 /* fall through, grab the bucket lock and lookup again.
 870                  * 99.9% chance that the element won't be found,
 871                  * but second lookup under lock has to be done.
 872                  */
 873         }
 874
 875         /* bpf_map_update_elem() can be called in_irq() */
 876         raw_spin_lock_irqsave(&b->lock, flags);
 877
 878         l_old = lookup_elem_raw(head, hash, key, key_size);
 879
 880         ret = check_flags(htab, l_old, map_flags);
 881         if (ret)
 882                 goto err;
 883
 884         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
 885                 /* first lookup without the bucket lock didn't find the element,
 886                  * but second lookup with the bucket lock found it.
 887                  * This case is highly unlikely, but has to be dealt with:
 888                  * grab the element lock in addition to the bucket lock
 889                  * and update element in place
 890                  */
 891                 copy_map_value_locked(map,
 892                                       l_old->key + round_up(key_size, 8),
 893                                       value, false);
 894                 ret = 0;
 895                 goto err;
 896         }
 897
 898         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 899                                 l_old);
 900         if (IS_ERR(l_new)) {
 901                 /* all pre-allocated elements are in use or memory exhausted */
 902                 ret = PTR_ERR(l_new);
 903                 goto err;
 904         }
 905
 906         /* add new element to the head of the list, so that
 907          * concurrent search will find it before old elem
 908          */
 909         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 910         if (l_old) {
 911                 hlist_nulls_del_rcu(&l_old->hash_node);
 912                 if (!htab_is_prealloc(htab))
 913                         free_htab_elem(htab, l_old);
 914         }
 915         ret = 0;
 916 err:
 917         raw_spin_unlock_irqrestore(&b->lock, flags);
 918         return ret;
 919 }
 920
 921 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 922                                     u64 map_flags)
 923 {
 924         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 925         struct htab_elem *l_new, *l_old = NULL;
 926         struct hlist_nulls_head *head;
 927         unsigned long flags;
 928         struct bucket *b;
 929         u32 key_size, hash;
 930         int ret;
 931
 932         if (unlikely(map_flags > BPF_EXIST))
 933                 /* unknown flags */
 934                 return -EINVAL;
 935
 936         WARN_ON_ONCE(!rcu_read_lock_held());
 937
 938         key_size = map->key_size;
 939
 940         hash = htab_map_hash(key, key_size, htab->hashrnd);
 941
 942         b = __select_bucket(htab, hash);
 943         head = &b->head;
 944
 945         /* For LRU, we need to alloc before taking bucket's
 946          * spinlock because getting free nodes from LRU may need
 947          * to remove older elements from htab and this removal
 948          * operation will need a bucket lock.
 949          */
 950         l_new = prealloc_lru_pop(htab, key, hash);
 951         if (!l_new)
 952                 return -ENOMEM;
 953         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 954
 955         /* bpf_map_update_elem() can be called in_irq() */
 956         raw_spin_lock_irqsave(&b->lock, flags);
 957
 958         l_old = lookup_elem_raw(head, hash, key, key_size);
 959
 960         ret = check_flags(htab, l_old, map_flags);
 961         if (ret)
 962                 goto err;
 963
 964         /* add new element to the head of the list, so that
 965          * concurrent search will find it before old elem
 966          */
 967         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 968         if (l_old) {
 969                 bpf_lru_node_set_ref(&l_new->lru_node);
 970                 hlist_nulls_del_rcu(&l_old->hash_node);
 971         }
 972         ret = 0;
 973
 974 err:
 975         raw_spin_unlock_irqrestore(&b->lock, flags);
 976
 977         if (ret)
 978                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 979         else if (l_old)
 980                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 981
 982         return ret;
 983 }
 984
 985 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 986                                          void *value, u64 map_flags,
 987                                          bool onallcpus)
 988 {
 989         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 990         struct htab_elem *l_new = NULL, *l_old;
 991         struct hlist_nulls_head *head;
 992         unsigned long flags;
 993         struct bucket *b;
 994         u32 key_size, hash;
 995         int ret;
 996
 997         if (unlikely(map_flags > BPF_EXIST))
 998                 /* unknown flags */
 999                 return -EINVAL;
1000
1001         WARN_ON_ONCE(!rcu_read_lock_held());
1002
1003         key_size = map->key_size;
1004
1005         hash = htab_map_hash(key, key_size, htab->hashrnd);
1006
1007         b = __select_bucket(htab, hash);
1008         head = &b->head;
1009
1010         /* bpf_map_update_elem() can be called in_irq() */
1011         raw_spin_lock_irqsave(&b->lock, flags);
1012
1013         l_old = lookup_elem_raw(head, hash, key, key_size);
1014
1015         ret = check_flags(htab, l_old, map_flags);
1016         if (ret)
1017                 goto err;
1018
1019         if (l_old) {
1020                 /* per-cpu hash map can update value in-place */
1021                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1022                                 value, onallcpus);
1023         } else {
1024                 l_new = alloc_htab_elem(htab, key, value, key_size,
1025                                         hash, true, onallcpus, NULL);
1026                 if (IS_ERR(l_new)) {
1027                         ret = PTR_ERR(l_new);
1028                         goto err;
1029                 }
1030                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1031         }
1032         ret = 0;
1033 err:
1034         raw_spin_unlock_irqrestore(&b->lock, flags);
1035         return ret;
1036 }
1037
1038 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1039                                              void *value, u64 map_flags,
1040                                              bool onallcpus)
1041 {
1042         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1043         struct htab_elem *l_new = NULL, *l_old;
1044         struct hlist_nulls_head *head;
1045         unsigned long flags;
1046         struct bucket *b;
1047         u32 key_size, hash;
1048         int ret;
1049
1050         if (unlikely(map_flags > BPF_EXIST))
1051                 /* unknown flags */
1052                 return -EINVAL;
1053
1054         WARN_ON_ONCE(!rcu_read_lock_held());
1055
1056         key_size = map->key_size;
1057
1058         hash = htab_map_hash(key, key_size, htab->hashrnd);
1059
1060         b = __select_bucket(htab, hash);
1061         head = &b->head;
1062
1063         /* For LRU, we need to alloc before taking bucket's
1064          * spinlock because LRU's elem alloc may need
1065          * to remove older elem from htab and this removal
1066          * operation will need a bucket lock.
1067          */
1068         if (map_flags != BPF_EXIST) {
1069                 l_new = prealloc_lru_pop(htab, key, hash);
1070                 if (!l_new)
1071                         return -ENOMEM;
1072         }
1073
1074         /* bpf_map_update_elem() can be called in_irq() */
1075         raw_spin_lock_irqsave(&b->lock, flags);
1076
1077         l_old = lookup_elem_raw(head, hash, key, key_size);
1078
1079         ret = check_flags(htab, l_old, map_flags);
1080         if (ret)
1081                 goto err;
1082
1083         if (l_old) {
1084                 bpf_lru_node_set_ref(&l_old->lru_node);
1085
1086                 /* per-cpu hash map can update value in-place */
1087                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1088                                 value, onallcpus);
1089         } else {
1090                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1091                                 value, onallcpus);
1092                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1093                 l_new = NULL;
1094         }
1095         ret = 0;
1096 err:
1097         raw_spin_unlock_irqrestore(&b->lock, flags);
1098         if (l_new)
1099                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1100         return ret;
1101 }
1102
1103 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1104                                        void *value, u64 map_flags)
1105 {
1106         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1107 }
1108
1109 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1110                                            void *value, u64 map_flags)
1111 {
1112         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1113                                                  false);
1114 }
1115
1116 /* Called from syscall or from eBPF program */
1117 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1118 {
1119         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1120         struct hlist_nulls_head *head;
1121         struct bucket *b;
1122         struct htab_elem *l;
1123         unsigned long flags;
1124         u32 hash, key_size;
1125         int ret = -ENOENT;
1126
1127         WARN_ON_ONCE(!rcu_read_lock_held());
1128
1129         key_size = map->key_size;
1130
1131         hash = htab_map_hash(key, key_size, htab->hashrnd);
1132         b = __select_bucket(htab, hash);
1133         head = &b->head;
1134
1135         raw_spin_lock_irqsave(&b->lock, flags);
1136
1137         l = lookup_elem_raw(head, hash, key, key_size);
1138
1139         if (l) {
1140                 hlist_nulls_del_rcu(&l->hash_node);
1141                 free_htab_elem(htab, l);
1142                 ret = 0;
1143         }
1144
1145         raw_spin_unlock_irqrestore(&b->lock, flags);
1146         return ret;
1147 }
1148
1149 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1150 {
1151         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1152         struct hlist_nulls_head *head;
1153         struct bucket *b;
1154         struct htab_elem *l;
1155         unsigned long flags;
1156         u32 hash, key_size;
1157         int ret = -ENOENT;
1158
1159         WARN_ON_ONCE(!rcu_read_lock_held());
1160
1161         key_size = map->key_size;
1162
1163         hash = htab_map_hash(key, key_size, htab->hashrnd);
1164         b = __select_bucket(htab, hash);
1165         head = &b->head;
1166
1167         raw_spin_lock_irqsave(&b->lock, flags);
1168
1169         l = lookup_elem_raw(head, hash, key, key_size);
1170
1171         if (l) {
1172                 hlist_nulls_del_rcu(&l->hash_node);
1173                 ret = 0;
1174         }
1175
1176         raw_spin_unlock_irqrestore(&b->lock, flags);
1177         if (l)
1178                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1179         return ret;
1180 }
1181
1182 static void delete_all_elements(struct bpf_htab *htab)
1183 {
1184         int i;
1185
1186         for (i = 0; i < htab->n_buckets; i++) {
1187                 struct hlist_nulls_head *head = select_bucket(htab, i);
1188                 struct hlist_nulls_node *n;
1189                 struct htab_elem *l;
1190
1191                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1192                         hlist_nulls_del_rcu(&l->hash_node);
1193                         htab_elem_free(htab, l);
1194                 }
1195         }
1196 }
1197
1198 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1199 static void htab_map_free(struct bpf_map *map)
1200 {
1201         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1202
1203         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1204          * so the programs (can be more than one that used this map) were
1205          * disconnected from events. Wait for outstanding critical sections in
1206          * these programs to complete
1207          */
1208         synchronize_rcu();
1209
1210         /* some of free_htab_elem() callbacks for elements of this map may
1211          * not have executed. Wait for them.
1212          */
1213         rcu_barrier();
1214         if (!htab_is_prealloc(htab))
1215                 delete_all_elements(htab);
1216         else
1217                 prealloc_destroy(htab);
1218
1219         free_percpu(htab->extra_elems);
1220         bpf_map_area_free(htab->buckets);
1221         kfree(htab);
1222 }
1223
1224 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1225                                    struct seq_file *m)
1226 {
1227         void *value;
1228
1229         rcu_read_lock();
1230
1231         value = htab_map_lookup_elem(map, key);
1232         if (!value) {
1233                 rcu_read_unlock();
1234                 return;
1235         }
1236
1237         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1238         seq_puts(m, ": ");
1239         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1240         seq_puts(m, "\n");
1241
1242         rcu_read_unlock();
1243 }
1244
1245 static int
1246 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1247                                    const union bpf_attr *attr,
1248                                    union bpf_attr __user *uattr,
1249                                    bool do_delete, bool is_lru_map,
1250                                    bool is_percpu)
1251 {
1252         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1253         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1254         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1255         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1256         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1257         void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1258         u32 batch, max_count, size, bucket_size;
1259         u64 elem_map_flags, map_flags;
1260         struct hlist_nulls_head *head;
1261         struct hlist_nulls_node *n;
1262         unsigned long flags;
1263         struct htab_elem *l;
1264         struct bucket *b;
1265         int ret = 0;
1266
1267         elem_map_flags = attr->batch.elem_flags;
1268         if ((elem_map_flags & ~BPF_F_LOCK) ||
1269             ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1270                 return -EINVAL;
1271
1272         map_flags = attr->batch.flags;
1273         if (map_flags)
1274                 return -EINVAL;
1275
1276         max_count = attr->batch.count;
1277         if (!max_count)
1278                 return 0;
1279
1280         if (put_user(0, &uattr->batch.count))
1281                 return -EFAULT;
1282
1283         batch = 0;
1284         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1285                 return -EFAULT;
1286
1287         if (batch >= htab->n_buckets)
1288                 return -ENOENT;
1289
1290         key_size = htab->map.key_size;
1291         roundup_key_size = round_up(htab->map.key_size, 8);
1292         value_size = htab->map.value_size;
1293         size = round_up(value_size, 8);
1294         if (is_percpu)
1295                 value_size = size * num_possible_cpus();
1296         total = 0;
1297         /* while experimenting with hash tables with sizes ranging from 10 to
1298          * 1000, it was observed that a bucket can have upto 5 entries.
1299          */
1300         bucket_size = 5;
1301
1302 alloc:
1303         /* We cannot do copy_from_user or copy_to_user inside
1304          * the rcu_read_lock. Allocate enough space here.
1305          */
1306         keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1307         values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1308         if (!keys || !values) {
1309                 ret = -ENOMEM;
1310                 goto after_loop;
1311         }
1312
1313 again:
1314         preempt_disable();
1315         this_cpu_inc(bpf_prog_active);
1316         rcu_read_lock();
1317 again_nocopy:
1318         dst_key = keys;
1319         dst_val = values;
1320         b = &htab->buckets[batch];
1321         head = &b->head;
1322         raw_spin_lock_irqsave(&b->lock, flags);
1323
1324         bucket_cnt = 0;
1325         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1326                 bucket_cnt++;
1327
1328         if (bucket_cnt > (max_count - total)) {
1329                 if (total == 0)
1330                         ret = -ENOSPC;
1331                 raw_spin_unlock_irqrestore(&b->lock, flags);
1332                 rcu_read_unlock();
1333                 this_cpu_dec(bpf_prog_active);
1334                 preempt_enable();
1335                 goto after_loop;
1336         }
1337
1338         if (bucket_cnt > bucket_size) {
1339                 bucket_size = bucket_cnt;
1340                 raw_spin_unlock_irqrestore(&b->lock, flags);
1341                 rcu_read_unlock();
1342                 this_cpu_dec(bpf_prog_active);
1343                 preempt_enable();
1344                 kvfree(keys);
1345                 kvfree(values);
1346                 goto alloc;
1347         }
1348
1349         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1350                 memcpy(dst_key, l->key, key_size);
1351
1352                 if (is_percpu) {
1353                         int off = 0, cpu;
1354                         void __percpu *pptr;
1355
1356                         pptr = htab_elem_get_ptr(l, map->key_size);
1357                         for_each_possible_cpu(cpu) {
1358                                 bpf_long_memcpy(dst_val + off,
1359                                                 per_cpu_ptr(pptr, cpu), size);
1360                                 off += size;
1361                         }
1362                 } else {
1363                         value = l->key + roundup_key_size;
1364                         if (elem_map_flags & BPF_F_LOCK)
1365                                 copy_map_value_locked(map, dst_val, value,
1366                                                       true);
1367                         else
1368                                 copy_map_value(map, dst_val, value);
1369                         check_and_init_map_lock(map, dst_val);
1370                 }
1371                 if (do_delete) {
1372                         hlist_nulls_del_rcu(&l->hash_node);
1373                         if (is_lru_map)
1374                                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1375                         else
1376                                 free_htab_elem(htab, l);
1377                 }
1378                 dst_key += key_size;
1379                 dst_val += value_size;
1380         }
1381
1382         raw_spin_unlock_irqrestore(&b->lock, flags);
1383         /* If we are not copying data, we can go to next bucket and avoid
1384          * unlocking the rcu.
1385          */
1386         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1387                 batch++;
1388                 goto again_nocopy;
1389         }
1390
1391         rcu_read_unlock();
1392         this_cpu_dec(bpf_prog_active);
1393         preempt_enable();
1394         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1395             key_size * bucket_cnt) ||
1396             copy_to_user(uvalues + total * value_size, values,
1397             value_size * bucket_cnt))) {
1398                 ret = -EFAULT;
1399                 goto after_loop;
1400         }
1401
1402         total += bucket_cnt;
1403         batch++;
1404         if (batch >= htab->n_buckets) {
1405                 ret = -ENOENT;
1406                 goto after_loop;
1407         }
1408         goto again;
1409
1410 after_loop:
1411         if (ret == -EFAULT)
1412                 goto out;
1413
1414         /* copy # of entries and next batch */
1415         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1416         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1417             put_user(total, &uattr->batch.count))
1418                 ret = -EFAULT;
1419
1420 out:
1421         kvfree(keys);
1422         kvfree(values);
1423         return ret;
1424 }
1425
1426 static int
1427 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1428                              union bpf_attr __user *uattr)
1429 {
1430         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1431                                                   false, true);
1432 }
1433
1434 static int
1435 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1436                                         const union bpf_attr *attr,
1437                                         union bpf_attr __user *uattr)
1438 {
1439         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1440                                                   false, true);
1441 }
1442
1443 static int
1444 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1445                       union bpf_attr __user *uattr)
1446 {
1447         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1448                                                   false, false);
1449 }
1450
1451 static int
1452 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1453                                  const union bpf_attr *attr,
1454                                  union bpf_attr __user *uattr)
1455 {
1456         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1457                                                   false, false);
1458 }
1459
1460 static int
1461 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1462                                  const union bpf_attr *attr,
1463                                  union bpf_attr __user *uattr)
1464 {
1465         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1466                                                   true, true);
1467 }
1468
1469 static int
1470 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1471                                             const union bpf_attr *attr,
1472                                             union bpf_attr __user *uattr)
1473 {
1474         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1475                                                   true, true);
1476 }
1477
1478 static int
1479 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1480                           union bpf_attr __user *uattr)
1481 {
1482         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1483                                                   true, false);
1484 }
1485
1486 static int
1487 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1488                                      const union bpf_attr *attr,
1489                                      union bpf_attr __user *uattr)
1490 {
1491         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1492                                                   true, false);
1493 }
1494
1495 const struct bpf_map_ops htab_map_ops = {
1496         .map_alloc_check = htab_map_alloc_check,
1497         .map_alloc = htab_map_alloc,
1498         .map_free = htab_map_free,
1499         .map_get_next_key = htab_map_get_next_key,
1500         .map_lookup_elem = htab_map_lookup_elem,
1501         .map_update_elem = htab_map_update_elem,
1502         .map_delete_elem = htab_map_delete_elem,
1503         .map_gen_lookup = htab_map_gen_lookup,
1504         .map_seq_show_elem = htab_map_seq_show_elem,
1505         BATCH_OPS(htab),
1506 };
1507
1508 const struct bpf_map_ops htab_lru_map_ops = {
1509         .map_alloc_check = htab_map_alloc_check,
1510         .map_alloc = htab_map_alloc,
1511         .map_free = htab_map_free,
1512         .map_get_next_key = htab_map_get_next_key,
1513         .map_lookup_elem = htab_lru_map_lookup_elem,
1514         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1515         .map_update_elem = htab_lru_map_update_elem,
1516         .map_delete_elem = htab_lru_map_delete_elem,
1517         .map_gen_lookup = htab_lru_map_gen_lookup,
1518         .map_seq_show_elem = htab_map_seq_show_elem,
1519         BATCH_OPS(htab_lru),
1520 };
1521
1522 /* Called from eBPF program */
1523 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1524 {
1525         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1526
1527         if (l)
1528                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1529         else
1530                 return NULL;
1531 }
1532
1533 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1534 {
1535         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1536
1537         if (l) {
1538                 bpf_lru_node_set_ref(&l->lru_node);
1539                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1540         }
1541
1542         return NULL;
1543 }
1544
1545 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1546 {
1547         struct htab_elem *l;
1548         void __percpu *pptr;
1549         int ret = -ENOENT;
1550         int cpu, off = 0;
1551         u32 size;
1552
1553         /* per_cpu areas are zero-filled and bpf programs can only
1554          * access 'value_size' of them, so copying rounded areas
1555          * will not leak any kernel data
1556          */
1557         size = round_up(map->value_size, 8);
1558         rcu_read_lock();
1559         l = __htab_map_lookup_elem(map, key);
1560         if (!l)
1561                 goto out;
1562         /* We do not mark LRU map element here in order to not mess up
1563          * eviction heuristics when user space does a map walk.
1564          */
1565         pptr = htab_elem_get_ptr(l, map->key_size);
1566         for_each_possible_cpu(cpu) {
1567                 bpf_long_memcpy(value + off,
1568                                 per_cpu_ptr(pptr, cpu), size);
1569                 off += size;
1570         }
1571         ret = 0;
1572 out:
1573         rcu_read_unlock();
1574         return ret;
1575 }
1576
1577 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1578                            u64 map_flags)
1579 {
1580         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1581         int ret;
1582
1583         rcu_read_lock();
1584         if (htab_is_lru(htab))
1585                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1586                                                         map_flags, true);
1587         else
1588                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1589                                                     true);
1590         rcu_read_unlock();
1591
1592         return ret;
1593 }
1594
1595 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1596                                           struct seq_file *m)
1597 {
1598         struct htab_elem *l;
1599         void __percpu *pptr;
1600         int cpu;
1601
1602         rcu_read_lock();
1603
1604         l = __htab_map_lookup_elem(map, key);
1605         if (!l) {
1606                 rcu_read_unlock();
1607                 return;
1608         }
1609
1610         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1611         seq_puts(m, ": {\n");
1612         pptr = htab_elem_get_ptr(l, map->key_size);
1613         for_each_possible_cpu(cpu) {
1614                 seq_printf(m, "\tcpu%d: ", cpu);
1615                 btf_type_seq_show(map->btf, map->btf_value_type_id,
1616                                   per_cpu_ptr(pptr, cpu), m);
1617                 seq_puts(m, "\n");
1618         }
1619         seq_puts(m, "}\n");
1620
1621         rcu_read_unlock();
1622 }
1623
1624 const struct bpf_map_ops htab_percpu_map_ops = {
1625         .map_alloc_check = htab_map_alloc_check,
1626         .map_alloc = htab_map_alloc,
1627         .map_free = htab_map_free,
1628         .map_get_next_key = htab_map_get_next_key,
1629         .map_lookup_elem = htab_percpu_map_lookup_elem,
1630         .map_update_elem = htab_percpu_map_update_elem,
1631         .map_delete_elem = htab_map_delete_elem,
1632         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1633         BATCH_OPS(htab_percpu),
1634 };
1635
1636 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1637         .map_alloc_check = htab_map_alloc_check,
1638         .map_alloc = htab_map_alloc,
1639         .map_free = htab_map_free,
1640         .map_get_next_key = htab_map_get_next_key,
1641         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1642         .map_update_elem = htab_lru_percpu_map_update_elem,
1643         .map_delete_elem = htab_lru_map_delete_elem,
1644         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1645         BATCH_OPS(htab_lru_percpu),
1646 };
1647
1648 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1649 {
1650         if (attr->value_size != sizeof(u32))
1651                 return -EINVAL;
1652         return htab_map_alloc_check(attr);
1653 }
1654
1655 static void fd_htab_map_free(struct bpf_map *map)
1656 {
1657         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1658         struct hlist_nulls_node *n;
1659         struct hlist_nulls_head *head;
1660         struct htab_elem *l;
1661         int i;
1662
1663         for (i = 0; i < htab->n_buckets; i++) {
1664                 head = select_bucket(htab, i);
1665
1666                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1667                         void *ptr = fd_htab_map_get_ptr(map, l);
1668
1669                         map->ops->map_fd_put_ptr(ptr);
1670                 }
1671         }
1672
1673         htab_map_free(map);
1674 }
1675
1676 /* only called from syscall */
1677 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1678 {
1679         void **ptr;
1680         int ret = 0;
1681
1682         if (!map->ops->map_fd_sys_lookup_elem)
1683                 return -ENOTSUPP;
1684
1685         rcu_read_lock();
1686         ptr = htab_map_lookup_elem(map, key);
1687         if (ptr)
1688                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1689         else
1690                 ret = -ENOENT;
1691         rcu_read_unlock();
1692
1693         return ret;
1694 }
1695
1696 /* only called from syscall */
1697 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1698                                 void *key, void *value, u64 map_flags)
1699 {
1700         void *ptr;
1701         int ret;
1702         u32 ufd = *(u32 *)value;
1703
1704         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1705         if (IS_ERR(ptr))
1706                 return PTR_ERR(ptr);
1707
1708         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1709         if (ret)
1710                 map->ops->map_fd_put_ptr(ptr);
1711
1712         return ret;
1713 }
1714
1715 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1716 {
1717         struct bpf_map *map, *inner_map_meta;
1718
1719         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1720         if (IS_ERR(inner_map_meta))
1721                 return inner_map_meta;
1722
1723         map = htab_map_alloc(attr);
1724         if (IS_ERR(map)) {
1725                 bpf_map_meta_free(inner_map_meta);
1726                 return map;
1727         }
1728
1729         map->inner_map_meta = inner_map_meta;
1730
1731         return map;
1732 }
1733
1734 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1735 {
1736         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1737
1738         if (!inner_map)
1739                 return NULL;
1740
1741         return READ_ONCE(*inner_map);
1742 }
1743
1744 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1745                                   struct bpf_insn *insn_buf)
1746 {
1747         struct bpf_insn *insn = insn_buf;
1748         const int ret = BPF_REG_0;
1749
1750         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1751                      (void *(*)(struct bpf_map *map, void *key))NULL));
1752         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1753         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1754         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1755                                 offsetof(struct htab_elem, key) +
1756                                 round_up(map->key_size, 8));
1757         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1758
1759         return insn - insn_buf;
1760 }
1761
1762 static void htab_of_map_free(struct bpf_map *map)
1763 {
1764         bpf_map_meta_free(map->inner_map_meta);
1765         fd_htab_map_free(map);
1766 }
1767
1768 const struct bpf_map_ops htab_of_maps_map_ops = {
1769         .map_alloc_check = fd_htab_map_alloc_check,
1770         .map_alloc = htab_of_map_alloc,
1771         .map_free = htab_of_map_free,
1772         .map_get_next_key = htab_map_get_next_key,
1773         .map_lookup_elem = htab_of_map_lookup_elem,
1774         .map_delete_elem = htab_map_delete_elem,
1775         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1776         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1777         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1778         .map_gen_lookup = htab_of_map_gen_lookup,
1779         .map_check_btf = map_check_no_btf,
1780 };