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                                 struct htab_elem *batch_flink;
  60                         };
  61                 };
  62         };
  63         union {
  64                 struct rcu_head rcu;
  65                 struct bpf_lru_node lru_node;
  66         };
  67         u32 hash;
  68         char key[0] __aligned(8);
  69 };
  70
  71 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  72
  73 static bool htab_is_lru(const struct bpf_htab *htab)
  74 {
  75         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  76                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  77 }
  78
  79 static bool htab_is_percpu(const struct bpf_htab *htab)
  80 {
  81         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  82                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  83 }
  84
  85 static bool htab_is_prealloc(const struct bpf_htab *htab)
  86 {
  87         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  88 }
  89
  90 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  91                                      void __percpu *pptr)
  92 {
  93         *(void __percpu **)(l->key + key_size) = pptr;
  94 }
  95
  96 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  97 {
  98         return *(void __percpu **)(l->key + key_size);
  99 }
 100
 101 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 102 {
 103         return *(void **)(l->key + roundup(map->key_size, 8));
 104 }
 105
 106 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 107 {
 108         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 109 }
 110
 111 static void htab_free_elems(struct bpf_htab *htab)
 112 {
 113         int i;
 114
 115         if (!htab_is_percpu(htab))
 116                 goto free_elems;
 117
 118         for (i = 0; i < htab->map.max_entries; i++) {
 119                 void __percpu *pptr;
 120
 121                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 122                                          htab->map.key_size);
 123                 free_percpu(pptr);
 124                 cond_resched();
 125         }
 126 free_elems:
 127         bpf_map_area_free(htab->elems);
 128 }
 129
 130 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 131  * (bucket_lock). If both locks need to be acquired together, the lock
 132  * order is always lru_lock -> bucket_lock and this only happens in
 133  * bpf_lru_list.c logic. For example, certain code path of
 134  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 135  * will acquire lru_lock first followed by acquiring bucket_lock.
 136  *
 137  * In hashtab.c, to avoid deadlock, lock acquisition of
 138  * bucket_lock followed by lru_lock is not allowed. In such cases,
 139  * bucket_lock needs to be released first before acquiring lru_lock.
 140  */
 141 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 142                                           u32 hash)
 143 {
 144         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 145         struct htab_elem *l;
 146
 147         if (node) {
 148                 l = container_of(node, struct htab_elem, lru_node);
 149                 memcpy(l->key, key, htab->map.key_size);
 150                 return l;
 151         }
 152
 153         return NULL;
 154 }
 155
 156 static int prealloc_init(struct bpf_htab *htab)
 157 {
 158         u32 num_entries = htab->map.max_entries;
 159         int err = -ENOMEM, i;
 160
 161         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 162                 num_entries += num_possible_cpus();
 163
 164         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 165                                          htab->map.numa_node);
 166         if (!htab->elems)
 167                 return -ENOMEM;
 168
 169         if (!htab_is_percpu(htab))
 170                 goto skip_percpu_elems;
 171
 172         for (i = 0; i < num_entries; i++) {
 173                 u32 size = round_up(htab->map.value_size, 8);
 174                 void __percpu *pptr;
 175
 176                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 177                 if (!pptr)
 178                         goto free_elems;
 179                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 180                                   pptr);
 181                 cond_resched();
 182         }
 183
 184 skip_percpu_elems:
 185         if (htab_is_lru(htab))
 186                 err = bpf_lru_init(&htab->lru,
 187                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 188                                    offsetof(struct htab_elem, hash) -
 189                                    offsetof(struct htab_elem, lru_node),
 190                                    htab_lru_map_delete_node,
 191                                    htab);
 192         else
 193                 err = pcpu_freelist_init(&htab->freelist);
 194
 195         if (err)
 196                 goto free_elems;
 197
 198         if (htab_is_lru(htab))
 199                 bpf_lru_populate(&htab->lru, htab->elems,
 200                                  offsetof(struct htab_elem, lru_node),
 201                                  htab->elem_size, num_entries);
 202         else
 203                 pcpu_freelist_populate(&htab->freelist,
 204                                        htab->elems + offsetof(struct htab_elem, fnode),
 205                                        htab->elem_size, num_entries);
 206
 207         return 0;
 208
 209 free_elems:
 210         htab_free_elems(htab);
 211         return err;
 212 }
 213
 214 static void prealloc_destroy(struct bpf_htab *htab)
 215 {
 216         htab_free_elems(htab);
 217
 218         if (htab_is_lru(htab))
 219                 bpf_lru_destroy(&htab->lru);
 220         else
 221                 pcpu_freelist_destroy(&htab->freelist);
 222 }
 223
 224 static int alloc_extra_elems(struct bpf_htab *htab)
 225 {
 226         struct htab_elem *__percpu *pptr, *l_new;
 227         struct pcpu_freelist_node *l;
 228         int cpu;
 229
 230         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 231                                   GFP_USER | __GFP_NOWARN);
 232         if (!pptr)
 233                 return -ENOMEM;
 234
 235         for_each_possible_cpu(cpu) {
 236                 l = pcpu_freelist_pop(&htab->freelist);
 237                 /* pop will succeed, since prealloc_init()
 238                  * preallocated extra num_possible_cpus elements
 239                  */
 240                 l_new = container_of(l, struct htab_elem, fnode);
 241                 *per_cpu_ptr(pptr, cpu) = l_new;
 242         }
 243         htab->extra_elems = pptr;
 244         return 0;
 245 }
 246
 247 /* Called from syscall */
 248 static int htab_map_alloc_check(union bpf_attr *attr)
 249 {
 250         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 251                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 252         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 253                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 254         /* percpu_lru means each cpu has its own LRU list.
 255          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 256          * the map's value itself is percpu.  percpu_lru has
 257          * nothing to do with the map's value.
 258          */
 259         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 260         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 261         bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 262         int numa_node = bpf_map_attr_numa_node(attr);
 263
 264         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 265                      offsetof(struct htab_elem, hash_node.pprev));
 266         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 267                      offsetof(struct htab_elem, hash_node.pprev));
 268
 269         if (lru && !capable(CAP_SYS_ADMIN))
 270                 /* LRU implementation is much complicated than other
 271                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 272                  */
 273                 return -EPERM;
 274
 275         if (zero_seed && !capable(CAP_SYS_ADMIN))
 276                 /* Guard against local DoS, and discourage production use. */
 277                 return -EPERM;
 278
 279         if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 280             !bpf_map_flags_access_ok(attr->map_flags))
 281                 return -EINVAL;
 282
 283         if (!lru && percpu_lru)
 284                 return -EINVAL;
 285
 286         if (lru && !prealloc)
 287                 return -ENOTSUPP;
 288
 289         if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 290                 return -EINVAL;
 291
 292         /* check sanity of attributes.
 293          * value_size == 0 may be allowed in the future to use map as a set
 294          */
 295         if (attr->max_entries == 0 || attr->key_size == 0 ||
 296             attr->value_size == 0)
 297                 return -EINVAL;
 298
 299         if (attr->key_size > MAX_BPF_STACK)
 300                 /* eBPF programs initialize keys on stack, so they cannot be
 301                  * larger than max stack size
 302                  */
 303                 return -E2BIG;
 304
 305         if (attr->value_size >= KMALLOC_MAX_SIZE -
 306             MAX_BPF_STACK - sizeof(struct htab_elem))
 307                 /* if value_size is bigger, the user space won't be able to
 308                  * access the elements via bpf syscall. This check also makes
 309                  * sure that the elem_size doesn't overflow and it's
 310                  * kmalloc-able later in htab_map_update_elem()
 311                  */
 312                 return -E2BIG;
 313
 314         return 0;
 315 }
 316
 317 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 318 {
 319         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 320                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 321         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 322                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 323         /* percpu_lru means each cpu has its own LRU list.
 324          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 325          * the map's value itself is percpu.  percpu_lru has
 326          * nothing to do with the map's value.
 327          */
 328         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 329         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 330         struct bpf_htab *htab;
 331         int err, i;
 332         u64 cost;
 333
 334         htab = kzalloc(sizeof(*htab), GFP_USER);
 335         if (!htab)
 336                 return ERR_PTR(-ENOMEM);
 337
 338         bpf_map_init_from_attr(&htab->map, attr);
 339
 340         if (percpu_lru) {
 341                 /* ensure each CPU's lru list has >=1 elements.
 342                  * since we are at it, make each lru list has the same
 343                  * number of elements.
 344                  */
 345                 htab->map.max_entries = roundup(attr->max_entries,
 346                                                 num_possible_cpus());
 347                 if (htab->map.max_entries < attr->max_entries)
 348                         htab->map.max_entries = rounddown(attr->max_entries,
 349                                                           num_possible_cpus());
 350         }
 351
 352         /* hash table size must be power of 2 */
 353         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 354
 355         htab->elem_size = sizeof(struct htab_elem) +
 356                           round_up(htab->map.key_size, 8);
 357         if (percpu)
 358                 htab->elem_size += sizeof(void *);
 359         else
 360                 htab->elem_size += round_up(htab->map.value_size, 8);
 361
 362         err = -E2BIG;
 363         /* prevent zero size kmalloc and check for u32 overflow */
 364         if (htab->n_buckets == 0 ||
 365             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 366                 goto free_htab;
 367
 368         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 369                (u64) htab->elem_size * htab->map.max_entries;
 370
 371         if (percpu)
 372                 cost += (u64) round_up(htab->map.value_size, 8) *
 373                         num_possible_cpus() * htab->map.max_entries;
 374         else
 375                cost += (u64) htab->elem_size * num_possible_cpus();
 376
 377         /* if map size is larger than memlock limit, reject it */
 378         err = bpf_map_charge_init(&htab->map.memory, cost);
 379         if (err)
 380                 goto free_htab;
 381
 382         err = -ENOMEM;
 383         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 384                                            sizeof(struct bucket),
 385                                            htab->map.numa_node);
 386         if (!htab->buckets)
 387                 goto free_charge;
 388
 389         if (htab->map.map_flags & BPF_F_ZERO_SEED)
 390                 htab->hashrnd = 0;
 391         else
 392                 htab->hashrnd = get_random_int();
 393
 394         for (i = 0; i < htab->n_buckets; i++) {
 395                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 396                 raw_spin_lock_init(&htab->buckets[i].lock);
 397         }
 398
 399         if (prealloc) {
 400                 err = prealloc_init(htab);
 401                 if (err)
 402                         goto free_buckets;
 403
 404                 if (!percpu && !lru) {
 405                         /* lru itself can remove the least used element, so
 406                          * there is no need for an extra elem during map_update.
 407                          */
 408                         err = alloc_extra_elems(htab);
 409                         if (err)
 410                                 goto free_prealloc;
 411                 }
 412         }
 413
 414         return &htab->map;
 415
 416 free_prealloc:
 417         prealloc_destroy(htab);
 418 free_buckets:
 419         bpf_map_area_free(htab->buckets);
 420 free_charge:
 421         bpf_map_charge_finish(&htab->map.memory);
 422 free_htab:
 423         kfree(htab);
 424         return ERR_PTR(err);
 425 }
 426
 427 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 428 {
 429         return jhash(key, key_len, hashrnd);
 430 }
 431
 432 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 433 {
 434         return &htab->buckets[hash & (htab->n_buckets - 1)];
 435 }
 436
 437 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 438 {
 439         return &__select_bucket(htab, hash)->head;
 440 }
 441
 442 /* this lookup function can only be called with bucket lock taken */
 443 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 444                                          void *key, u32 key_size)
 445 {
 446         struct hlist_nulls_node *n;
 447         struct htab_elem *l;
 448
 449         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 450                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 451                         return l;
 452
 453         return NULL;
 454 }
 455
 456 /* can be called without bucket lock. it will repeat the loop in
 457  * the unlikely event when elements moved from one bucket into another
 458  * while link list is being walked
 459  */
 460 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 461                                                u32 hash, void *key,
 462                                                u32 key_size, u32 n_buckets)
 463 {
 464         struct hlist_nulls_node *n;
 465         struct htab_elem *l;
 466
 467 again:
 468         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 469                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 470                         return l;
 471
 472         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 473                 goto again;
 474
 475         return NULL;
 476 }
 477
 478 /* Called from syscall or from eBPF program directly, so
 479  * arguments have to match bpf_map_lookup_elem() exactly.
 480  * The return value is adjusted by BPF instructions
 481  * in htab_map_gen_lookup().
 482  */
 483 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 484 {
 485         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 486         struct hlist_nulls_head *head;
 487         struct htab_elem *l;
 488         u32 hash, key_size;
 489
 490         /* Must be called with rcu_read_lock. */
 491         WARN_ON_ONCE(!rcu_read_lock_held());
 492
 493         key_size = map->key_size;
 494
 495         hash = htab_map_hash(key, key_size, htab->hashrnd);
 496
 497         head = select_bucket(htab, hash);
 498
 499         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 500
 501         return l;
 502 }
 503
 504 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 505 {
 506         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 507
 508         if (l)
 509                 return l->key + round_up(map->key_size, 8);
 510
 511         return NULL;
 512 }
 513
 514 /* inline bpf_map_lookup_elem() call.
 515  * Instead of:
 516  * bpf_prog
 517  *   bpf_map_lookup_elem
 518  *     map->ops->map_lookup_elem
 519  *       htab_map_lookup_elem
 520  *         __htab_map_lookup_elem
 521  * do:
 522  * bpf_prog
 523  *   __htab_map_lookup_elem
 524  */
 525 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 526 {
 527         struct bpf_insn *insn = insn_buf;
 528         const int ret = BPF_REG_0;
 529
 530         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 531                      (void *(*)(struct bpf_map *map, void *key))NULL));
 532         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 533         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 534         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 535                                 offsetof(struct htab_elem, key) +
 536                                 round_up(map->key_size, 8));
 537         return insn - insn_buf;
 538 }
 539
 540 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 541                                                         void *key, const bool mark)
 542 {
 543         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 544
 545         if (l) {
 546                 if (mark)
 547                         bpf_lru_node_set_ref(&l->lru_node);
 548                 return l->key + round_up(map->key_size, 8);
 549         }
 550
 551         return NULL;
 552 }
 553
 554 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 555 {
 556         return __htab_lru_map_lookup_elem(map, key, true);
 557 }
 558
 559 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 560 {
 561         return __htab_lru_map_lookup_elem(map, key, false);
 562 }
 563
 564 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 565                                    struct bpf_insn *insn_buf)
 566 {
 567         struct bpf_insn *insn = insn_buf;
 568         const int ret = BPF_REG_0;
 569         const int ref_reg = BPF_REG_1;
 570
 571         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 572                      (void *(*)(struct bpf_map *map, void *key))NULL));
 573         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 574         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 575         *insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 576                               offsetof(struct htab_elem, lru_node) +
 577                               offsetof(struct bpf_lru_node, ref));
 578         *insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 579         *insn++ = BPF_ST_MEM(BPF_B, ret,
 580                              offsetof(struct htab_elem, lru_node) +
 581                              offsetof(struct bpf_lru_node, ref),
 582                              1);
 583         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 584                                 offsetof(struct htab_elem, key) +
 585                                 round_up(map->key_size, 8));
 586         return insn - insn_buf;
 587 }
 588
 589 /* It is called from the bpf_lru_list when the LRU needs to delete
 590  * older elements from the htab.
 591  */
 592 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 593 {
 594         struct bpf_htab *htab = (struct bpf_htab *)arg;
 595         struct htab_elem *l = NULL, *tgt_l;
 596         struct hlist_nulls_head *head;
 597         struct hlist_nulls_node *n;
 598         unsigned long flags;
 599         struct bucket *b;
 600
 601         tgt_l = container_of(node, struct htab_elem, lru_node);
 602         b = __select_bucket(htab, tgt_l->hash);
 603         head = &b->head;
 604
 605         raw_spin_lock_irqsave(&b->lock, flags);
 606
 607         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 608                 if (l == tgt_l) {
 609                         hlist_nulls_del_rcu(&l->hash_node);
 610                         break;
 611                 }
 612
 613         raw_spin_unlock_irqrestore(&b->lock, flags);
 614
 615         return l == tgt_l;
 616 }
 617
 618 /* Called from syscall */
 619 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 620 {
 621         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 622         struct hlist_nulls_head *head;
 623         struct htab_elem *l, *next_l;
 624         u32 hash, key_size;
 625         int i = 0;
 626
 627         WARN_ON_ONCE(!rcu_read_lock_held());
 628
 629         key_size = map->key_size;
 630
 631         if (!key)
 632                 goto find_first_elem;
 633
 634         hash = htab_map_hash(key, key_size, htab->hashrnd);
 635
 636         head = select_bucket(htab, hash);
 637
 638         /* lookup the key */
 639         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 640
 641         if (!l)
 642                 goto find_first_elem;
 643
 644         /* key was found, get next key in the same bucket */
 645         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 646                                   struct htab_elem, hash_node);
 647
 648         if (next_l) {
 649                 /* if next elem in this hash list is non-zero, just return it */
 650                 memcpy(next_key, next_l->key, key_size);
 651                 return 0;
 652         }
 653
 654         /* no more elements in this hash list, go to the next bucket */
 655         i = hash & (htab->n_buckets - 1);
 656         i++;
 657
 658 find_first_elem:
 659         /* iterate over buckets */
 660         for (; i < htab->n_buckets; i++) {
 661                 head = select_bucket(htab, i);
 662
 663                 /* pick first element in the bucket */
 664                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 665                                           struct htab_elem, hash_node);
 666                 if (next_l) {
 667                         /* if it's not empty, just return it */
 668                         memcpy(next_key, next_l->key, key_size);
 669                         return 0;
 670                 }
 671         }
 672
 673         /* iterated over all buckets and all elements */
 674         return -ENOENT;
 675 }
 676
 677 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 678 {
 679         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 680                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 681         kfree(l);
 682 }
 683
 684 static void htab_elem_free_rcu(struct rcu_head *head)
 685 {
 686         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 687         struct bpf_htab *htab = l->htab;
 688
 689         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 690          * we're calling kfree, otherwise deadlock is possible if kprobes
 691          * are placed somewhere inside of slub
 692          */
 693         preempt_disable();
 694         __this_cpu_inc(bpf_prog_active);
 695         htab_elem_free(htab, l);
 696         __this_cpu_dec(bpf_prog_active);
 697         preempt_enable();
 698 }
 699
 700 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 701 {
 702         struct bpf_map *map = &htab->map;
 703
 704         if (map->ops->map_fd_put_ptr) {
 705                 void *ptr = fd_htab_map_get_ptr(map, l);
 706
 707                 map->ops->map_fd_put_ptr(ptr);
 708         }
 709
 710         if (htab_is_prealloc(htab)) {
 711                 __pcpu_freelist_push(&htab->freelist, &l->fnode);
 712         } else {
 713                 atomic_dec(&htab->count);
 714                 l->htab = htab;
 715                 call_rcu(&l->rcu, htab_elem_free_rcu);
 716         }
 717 }
 718
 719 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 720                             void *value, bool onallcpus)
 721 {
 722         if (!onallcpus) {
 723                 /* copy true value_size bytes */
 724                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 725         } else {
 726                 u32 size = round_up(htab->map.value_size, 8);
 727                 int off = 0, cpu;
 728
 729                 for_each_possible_cpu(cpu) {
 730                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 731                                         value + off, size);
 732                         off += size;
 733                 }
 734         }
 735 }
 736
 737 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 738 {
 739         return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 740                BITS_PER_LONG == 64;
 741 }
 742
 743 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 744                                          void *value, u32 key_size, u32 hash,
 745                                          bool percpu, bool onallcpus,
 746                                          struct htab_elem *old_elem)
 747 {
 748         u32 size = htab->map.value_size;
 749         bool prealloc = htab_is_prealloc(htab);
 750         struct htab_elem *l_new, **pl_new;
 751         void __percpu *pptr;
 752
 753         if (prealloc) {
 754                 if (old_elem) {
 755                         /* if we're updating the existing element,
 756                          * use per-cpu extra elems to avoid freelist_pop/push
 757                          */
 758                         pl_new = this_cpu_ptr(htab->extra_elems);
 759                         l_new = *pl_new;
 760                         *pl_new = old_elem;
 761                 } else {
 762                         struct pcpu_freelist_node *l;
 763
 764                         l = __pcpu_freelist_pop(&htab->freelist);
 765                         if (!l)
 766                                 return ERR_PTR(-E2BIG);
 767                         l_new = container_of(l, struct htab_elem, fnode);
 768                 }
 769         } else {
 770                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 771                         if (!old_elem) {
 772                                 /* when map is full and update() is replacing
 773                                  * old element, it's ok to allocate, since
 774                                  * old element will be freed immediately.
 775                                  * Otherwise return an error
 776                                  */
 777                                 l_new = ERR_PTR(-E2BIG);
 778                                 goto dec_count;
 779                         }
 780                 l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 781                                      htab->map.numa_node);
 782                 if (!l_new) {
 783                         l_new = ERR_PTR(-ENOMEM);
 784                         goto dec_count;
 785                 }
 786                 check_and_init_map_lock(&htab->map,
 787                                         l_new->key + round_up(key_size, 8));
 788         }
 789
 790         memcpy(l_new->key, key, key_size);
 791         if (percpu) {
 792                 size = round_up(size, 8);
 793                 if (prealloc) {
 794                         pptr = htab_elem_get_ptr(l_new, key_size);
 795                 } else {
 796                         /* alloc_percpu zero-fills */
 797                         pptr = __alloc_percpu_gfp(size, 8,
 798                                                   GFP_ATOMIC | __GFP_NOWARN);
 799                         if (!pptr) {
 800                                 kfree(l_new);
 801                                 l_new = ERR_PTR(-ENOMEM);
 802                                 goto dec_count;
 803                         }
 804                 }
 805
 806                 pcpu_copy_value(htab, pptr, value, onallcpus);
 807
 808                 if (!prealloc)
 809                         htab_elem_set_ptr(l_new, key_size, pptr);
 810         } else if (fd_htab_map_needs_adjust(htab)) {
 811                 size = round_up(size, 8);
 812                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 813         } else {
 814                 copy_map_value(&htab->map,
 815                                l_new->key + round_up(key_size, 8),
 816                                value);
 817         }
 818
 819         l_new->hash = hash;
 820         return l_new;
 821 dec_count:
 822         atomic_dec(&htab->count);
 823         return l_new;
 824 }
 825
 826 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 827                        u64 map_flags)
 828 {
 829         if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 830                 /* elem already exists */
 831                 return -EEXIST;
 832
 833         if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 834                 /* elem doesn't exist, cannot update it */
 835                 return -ENOENT;
 836
 837         return 0;
 838 }
 839
 840 /* Called from syscall or from eBPF program */
 841 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 842                                 u64 map_flags)
 843 {
 844         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 845         struct htab_elem *l_new = NULL, *l_old;
 846         struct hlist_nulls_head *head;
 847         unsigned long flags;
 848         struct bucket *b;
 849         u32 key_size, hash;
 850         int ret;
 851
 852         if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
 853                 /* unknown flags */
 854                 return -EINVAL;
 855
 856         WARN_ON_ONCE(!rcu_read_lock_held());
 857
 858         key_size = map->key_size;
 859
 860         hash = htab_map_hash(key, key_size, htab->hashrnd);
 861
 862         b = __select_bucket(htab, hash);
 863         head = &b->head;
 864
 865         if (unlikely(map_flags & BPF_F_LOCK)) {
 866                 if (unlikely(!map_value_has_spin_lock(map)))
 867                         return -EINVAL;
 868                 /* find an element without taking the bucket lock */
 869                 l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
 870                                               htab->n_buckets);
 871                 ret = check_flags(htab, l_old, map_flags);
 872                 if (ret)
 873                         return ret;
 874                 if (l_old) {
 875                         /* grab the element lock and update value in place */
 876                         copy_map_value_locked(map,
 877                                               l_old->key + round_up(key_size, 8),
 878                                               value, false);
 879                         return 0;
 880                 }
 881                 /* fall through, grab the bucket lock and lookup again.
 882                  * 99.9% chance that the element won't be found,
 883                  * but second lookup under lock has to be done.
 884                  */
 885         }
 886
 887         /* bpf_map_update_elem() can be called in_irq() */
 888         raw_spin_lock_irqsave(&b->lock, flags);
 889
 890         l_old = lookup_elem_raw(head, hash, key, key_size);
 891
 892         ret = check_flags(htab, l_old, map_flags);
 893         if (ret)
 894                 goto err;
 895
 896         if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
 897                 /* first lookup without the bucket lock didn't find the element,
 898                  * but second lookup with the bucket lock found it.
 899                  * This case is highly unlikely, but has to be dealt with:
 900                  * grab the element lock in addition to the bucket lock
 901                  * and update element in place
 902                  */
 903                 copy_map_value_locked(map,
 904                                       l_old->key + round_up(key_size, 8),
 905                                       value, false);
 906                 ret = 0;
 907                 goto err;
 908         }
 909
 910         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 911                                 l_old);
 912         if (IS_ERR(l_new)) {
 913                 /* all pre-allocated elements are in use or memory exhausted */
 914                 ret = PTR_ERR(l_new);
 915                 goto err;
 916         }
 917
 918         /* add new element to the head of the list, so that
 919          * concurrent search will find it before old elem
 920          */
 921         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 922         if (l_old) {
 923                 hlist_nulls_del_rcu(&l_old->hash_node);
 924                 if (!htab_is_prealloc(htab))
 925                         free_htab_elem(htab, l_old);
 926         }
 927         ret = 0;
 928 err:
 929         raw_spin_unlock_irqrestore(&b->lock, flags);
 930         return ret;
 931 }
 932
 933 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 934                                     u64 map_flags)
 935 {
 936         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 937         struct htab_elem *l_new, *l_old = NULL;
 938         struct hlist_nulls_head *head;
 939         unsigned long flags;
 940         struct bucket *b;
 941         u32 key_size, hash;
 942         int ret;
 943
 944         if (unlikely(map_flags > BPF_EXIST))
 945                 /* unknown flags */
 946                 return -EINVAL;
 947
 948         WARN_ON_ONCE(!rcu_read_lock_held());
 949
 950         key_size = map->key_size;
 951
 952         hash = htab_map_hash(key, key_size, htab->hashrnd);
 953
 954         b = __select_bucket(htab, hash);
 955         head = &b->head;
 956
 957         /* For LRU, we need to alloc before taking bucket's
 958          * spinlock because getting free nodes from LRU may need
 959          * to remove older elements from htab and this removal
 960          * operation will need a bucket lock.
 961          */
 962         l_new = prealloc_lru_pop(htab, key, hash);
 963         if (!l_new)
 964                 return -ENOMEM;
 965         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 966
 967         /* bpf_map_update_elem() can be called in_irq() */
 968         raw_spin_lock_irqsave(&b->lock, flags);
 969
 970         l_old = lookup_elem_raw(head, hash, key, key_size);
 971
 972         ret = check_flags(htab, l_old, map_flags);
 973         if (ret)
 974                 goto err;
 975
 976         /* add new element to the head of the list, so that
 977          * concurrent search will find it before old elem
 978          */
 979         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 980         if (l_old) {
 981                 bpf_lru_node_set_ref(&l_new->lru_node);
 982                 hlist_nulls_del_rcu(&l_old->hash_node);
 983         }
 984         ret = 0;
 985
 986 err:
 987         raw_spin_unlock_irqrestore(&b->lock, flags);
 988
 989         if (ret)
 990                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 991         else if (l_old)
 992                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 993
 994         return ret;
 995 }
 996
 997 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 998                                          void *value, u64 map_flags,
 999                                          bool onallcpus)
1000 {
1001         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1002         struct htab_elem *l_new = NULL, *l_old;
1003         struct hlist_nulls_head *head;
1004         unsigned long flags;
1005         struct bucket *b;
1006         u32 key_size, hash;
1007         int ret;
1008
1009         if (unlikely(map_flags > BPF_EXIST))
1010                 /* unknown flags */
1011                 return -EINVAL;
1012
1013         WARN_ON_ONCE(!rcu_read_lock_held());
1014
1015         key_size = map->key_size;
1016
1017         hash = htab_map_hash(key, key_size, htab->hashrnd);
1018
1019         b = __select_bucket(htab, hash);
1020         head = &b->head;
1021
1022         /* bpf_map_update_elem() can be called in_irq() */
1023         raw_spin_lock_irqsave(&b->lock, flags);
1024
1025         l_old = lookup_elem_raw(head, hash, key, key_size);
1026
1027         ret = check_flags(htab, l_old, map_flags);
1028         if (ret)
1029                 goto err;
1030
1031         if (l_old) {
1032                 /* per-cpu hash map can update value in-place */
1033                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1034                                 value, onallcpus);
1035         } else {
1036                 l_new = alloc_htab_elem(htab, key, value, key_size,
1037                                         hash, true, onallcpus, NULL);
1038                 if (IS_ERR(l_new)) {
1039                         ret = PTR_ERR(l_new);
1040                         goto err;
1041                 }
1042                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1043         }
1044         ret = 0;
1045 err:
1046         raw_spin_unlock_irqrestore(&b->lock, flags);
1047         return ret;
1048 }
1049
1050 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1051                                              void *value, u64 map_flags,
1052                                              bool onallcpus)
1053 {
1054         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1055         struct htab_elem *l_new = NULL, *l_old;
1056         struct hlist_nulls_head *head;
1057         unsigned long flags;
1058         struct bucket *b;
1059         u32 key_size, hash;
1060         int ret;
1061
1062         if (unlikely(map_flags > BPF_EXIST))
1063                 /* unknown flags */
1064                 return -EINVAL;
1065
1066         WARN_ON_ONCE(!rcu_read_lock_held());
1067
1068         key_size = map->key_size;
1069
1070         hash = htab_map_hash(key, key_size, htab->hashrnd);
1071
1072         b = __select_bucket(htab, hash);
1073         head = &b->head;
1074
1075         /* For LRU, we need to alloc before taking bucket's
1076          * spinlock because LRU's elem alloc may need
1077          * to remove older elem from htab and this removal
1078          * operation will need a bucket lock.
1079          */
1080         if (map_flags != BPF_EXIST) {
1081                 l_new = prealloc_lru_pop(htab, key, hash);
1082                 if (!l_new)
1083                         return -ENOMEM;
1084         }
1085
1086         /* bpf_map_update_elem() can be called in_irq() */
1087         raw_spin_lock_irqsave(&b->lock, flags);
1088
1089         l_old = lookup_elem_raw(head, hash, key, key_size);
1090
1091         ret = check_flags(htab, l_old, map_flags);
1092         if (ret)
1093                 goto err;
1094
1095         if (l_old) {
1096                 bpf_lru_node_set_ref(&l_old->lru_node);
1097
1098                 /* per-cpu hash map can update value in-place */
1099                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1100                                 value, onallcpus);
1101         } else {
1102                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1103                                 value, onallcpus);
1104                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1105                 l_new = NULL;
1106         }
1107         ret = 0;
1108 err:
1109         raw_spin_unlock_irqrestore(&b->lock, flags);
1110         if (l_new)
1111                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1112         return ret;
1113 }
1114
1115 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1116                                        void *value, u64 map_flags)
1117 {
1118         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1119 }
1120
1121 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1122                                            void *value, u64 map_flags)
1123 {
1124         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1125                                                  false);
1126 }
1127
1128 /* Called from syscall or from eBPF program */
1129 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1130 {
1131         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1132         struct hlist_nulls_head *head;
1133         struct bucket *b;
1134         struct htab_elem *l;
1135         unsigned long flags;
1136         u32 hash, key_size;
1137         int ret = -ENOENT;
1138
1139         WARN_ON_ONCE(!rcu_read_lock_held());
1140
1141         key_size = map->key_size;
1142
1143         hash = htab_map_hash(key, key_size, htab->hashrnd);
1144         b = __select_bucket(htab, hash);
1145         head = &b->head;
1146
1147         raw_spin_lock_irqsave(&b->lock, flags);
1148
1149         l = lookup_elem_raw(head, hash, key, key_size);
1150
1151         if (l) {
1152                 hlist_nulls_del_rcu(&l->hash_node);
1153                 free_htab_elem(htab, l);
1154                 ret = 0;
1155         }
1156
1157         raw_spin_unlock_irqrestore(&b->lock, flags);
1158         return ret;
1159 }
1160
1161 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1162 {
1163         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1164         struct hlist_nulls_head *head;
1165         struct bucket *b;
1166         struct htab_elem *l;
1167         unsigned long flags;
1168         u32 hash, key_size;
1169         int ret = -ENOENT;
1170
1171         WARN_ON_ONCE(!rcu_read_lock_held());
1172
1173         key_size = map->key_size;
1174
1175         hash = htab_map_hash(key, key_size, htab->hashrnd);
1176         b = __select_bucket(htab, hash);
1177         head = &b->head;
1178
1179         raw_spin_lock_irqsave(&b->lock, flags);
1180
1181         l = lookup_elem_raw(head, hash, key, key_size);
1182
1183         if (l) {
1184                 hlist_nulls_del_rcu(&l->hash_node);
1185                 ret = 0;
1186         }
1187
1188         raw_spin_unlock_irqrestore(&b->lock, flags);
1189         if (l)
1190                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1191         return ret;
1192 }
1193
1194 static void delete_all_elements(struct bpf_htab *htab)
1195 {
1196         int i;
1197
1198         for (i = 0; i < htab->n_buckets; i++) {
1199                 struct hlist_nulls_head *head = select_bucket(htab, i);
1200                 struct hlist_nulls_node *n;
1201                 struct htab_elem *l;
1202
1203                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1204                         hlist_nulls_del_rcu(&l->hash_node);
1205                         htab_elem_free(htab, l);
1206                 }
1207         }
1208 }
1209
1210 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1211 static void htab_map_free(struct bpf_map *map)
1212 {
1213         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1214
1215         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1216          * so the programs (can be more than one that used this map) were
1217          * disconnected from events. Wait for outstanding critical sections in
1218          * these programs to complete
1219          */
1220         synchronize_rcu();
1221
1222         /* some of free_htab_elem() callbacks for elements of this map may
1223          * not have executed. Wait for them.
1224          */
1225         rcu_barrier();
1226         if (!htab_is_prealloc(htab))
1227                 delete_all_elements(htab);
1228         else
1229                 prealloc_destroy(htab);
1230
1231         free_percpu(htab->extra_elems);
1232         bpf_map_area_free(htab->buckets);
1233         kfree(htab);
1234 }
1235
1236 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1237                                    struct seq_file *m)
1238 {
1239         void *value;
1240
1241         rcu_read_lock();
1242
1243         value = htab_map_lookup_elem(map, key);
1244         if (!value) {
1245                 rcu_read_unlock();
1246                 return;
1247         }
1248
1249         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1250         seq_puts(m, ": ");
1251         btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1252         seq_puts(m, "\n");
1253
1254         rcu_read_unlock();
1255 }
1256
1257 static int
1258 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1259                                    const union bpf_attr *attr,
1260                                    union bpf_attr __user *uattr,
1261                                    bool do_delete, bool is_lru_map,
1262                                    bool is_percpu)
1263 {
1264         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1265         u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1266         void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1267         void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1268         void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1269         void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1270         u32 batch, max_count, size, bucket_size;
1271         struct htab_elem *node_to_free = NULL;
1272         u64 elem_map_flags, map_flags;
1273         struct hlist_nulls_head *head;
1274         struct hlist_nulls_node *n;
1275         unsigned long flags = 0;
1276         bool locked = false;
1277         struct htab_elem *l;
1278         struct bucket *b;
1279         int ret = 0;
1280
1281         elem_map_flags = attr->batch.elem_flags;
1282         if ((elem_map_flags & ~BPF_F_LOCK) ||
1283             ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1284                 return -EINVAL;
1285
1286         map_flags = attr->batch.flags;
1287         if (map_flags)
1288                 return -EINVAL;
1289
1290         max_count = attr->batch.count;
1291         if (!max_count)
1292                 return 0;
1293
1294         if (put_user(0, &uattr->batch.count))
1295                 return -EFAULT;
1296
1297         batch = 0;
1298         if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1299                 return -EFAULT;
1300
1301         if (batch >= htab->n_buckets)
1302                 return -ENOENT;
1303
1304         key_size = htab->map.key_size;
1305         roundup_key_size = round_up(htab->map.key_size, 8);
1306         value_size = htab->map.value_size;
1307         size = round_up(value_size, 8);
1308         if (is_percpu)
1309                 value_size = size * num_possible_cpus();
1310         total = 0;
1311         /* while experimenting with hash tables with sizes ranging from 10 to
1312          * 1000, it was observed that a bucket can have upto 5 entries.
1313          */
1314         bucket_size = 5;
1315
1316 alloc:
1317         /* We cannot do copy_from_user or copy_to_user inside
1318          * the rcu_read_lock. Allocate enough space here.
1319          */
1320         keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1321         values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1322         if (!keys || !values) {
1323                 ret = -ENOMEM;
1324                 goto after_loop;
1325         }
1326
1327 again:
1328         preempt_disable();
1329         this_cpu_inc(bpf_prog_active);
1330         rcu_read_lock();
1331 again_nocopy:
1332         dst_key = keys;
1333         dst_val = values;
1334         b = &htab->buckets[batch];
1335         head = &b->head;
1336         /* do not grab the lock unless need it (bucket_cnt > 0). */
1337         if (locked)
1338                 raw_spin_lock_irqsave(&b->lock, flags);
1339
1340         bucket_cnt = 0;
1341         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1342                 bucket_cnt++;
1343
1344         if (bucket_cnt && !locked) {
1345                 locked = true;
1346                 goto again_nocopy;
1347         }
1348
1349         if (bucket_cnt > (max_count - total)) {
1350                 if (total == 0)
1351                         ret = -ENOSPC;
1352                 /* Note that since bucket_cnt > 0 here, it is implicit
1353                  * that the locked was grabbed, so release it.
1354                  */
1355                 raw_spin_unlock_irqrestore(&b->lock, flags);
1356                 rcu_read_unlock();
1357                 this_cpu_dec(bpf_prog_active);
1358                 preempt_enable();
1359                 goto after_loop;
1360         }
1361
1362         if (bucket_cnt > bucket_size) {
1363                 bucket_size = bucket_cnt;
1364                 /* Note that since bucket_cnt > 0 here, it is implicit
1365                  * that the locked was grabbed, so release it.
1366                  */
1367                 raw_spin_unlock_irqrestore(&b->lock, flags);
1368                 rcu_read_unlock();
1369                 this_cpu_dec(bpf_prog_active);
1370                 preempt_enable();
1371                 kvfree(keys);
1372                 kvfree(values);
1373                 goto alloc;
1374         }
1375
1376         /* Next block is only safe to run if you have grabbed the lock */
1377         if (!locked)
1378                 goto next_batch;
1379
1380         hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1381                 memcpy(dst_key, l->key, key_size);
1382
1383                 if (is_percpu) {
1384                         int off = 0, cpu;
1385                         void __percpu *pptr;
1386
1387                         pptr = htab_elem_get_ptr(l, map->key_size);
1388                         for_each_possible_cpu(cpu) {
1389                                 bpf_long_memcpy(dst_val + off,
1390                                                 per_cpu_ptr(pptr, cpu), size);
1391                                 off += size;
1392                         }
1393                 } else {
1394                         value = l->key + roundup_key_size;
1395                         if (elem_map_flags & BPF_F_LOCK)
1396                                 copy_map_value_locked(map, dst_val, value,
1397                                                       true);
1398                         else
1399                                 copy_map_value(map, dst_val, value);
1400                         check_and_init_map_lock(map, dst_val);
1401                 }
1402                 if (do_delete) {
1403                         hlist_nulls_del_rcu(&l->hash_node);
1404
1405                         /* bpf_lru_push_free() will acquire lru_lock, which
1406                          * may cause deadlock. See comments in function
1407                          * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1408                          * after releasing the bucket lock.
1409                          */
1410                         if (is_lru_map) {
1411                                 l->batch_flink = node_to_free;
1412                                 node_to_free = l;
1413                         } else {
1414                                 free_htab_elem(htab, l);
1415                         }
1416                 }
1417                 dst_key += key_size;
1418                 dst_val += value_size;
1419         }
1420
1421         raw_spin_unlock_irqrestore(&b->lock, flags);
1422         locked = false;
1423
1424         while (node_to_free) {
1425                 l = node_to_free;
1426                 node_to_free = node_to_free->batch_flink;
1427                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1428         }
1429
1430 next_batch:
1431         /* If we are not copying data, we can go to next bucket and avoid
1432          * unlocking the rcu.
1433          */
1434         if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1435                 batch++;
1436                 goto again_nocopy;
1437         }
1438
1439         rcu_read_unlock();
1440         this_cpu_dec(bpf_prog_active);
1441         preempt_enable();
1442         if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1443             key_size * bucket_cnt) ||
1444             copy_to_user(uvalues + total * value_size, values,
1445             value_size * bucket_cnt))) {
1446                 ret = -EFAULT;
1447                 goto after_loop;
1448         }
1449
1450         total += bucket_cnt;
1451         batch++;
1452         if (batch >= htab->n_buckets) {
1453                 ret = -ENOENT;
1454                 goto after_loop;
1455         }
1456         goto again;
1457
1458 after_loop:
1459         if (ret == -EFAULT)
1460                 goto out;
1461
1462         /* copy # of entries and next batch */
1463         ubatch = u64_to_user_ptr(attr->batch.out_batch);
1464         if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1465             put_user(total, &uattr->batch.count))
1466                 ret = -EFAULT;
1467
1468 out:
1469         kvfree(keys);
1470         kvfree(values);
1471         return ret;
1472 }
1473
1474 static int
1475 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1476                              union bpf_attr __user *uattr)
1477 {
1478         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1479                                                   false, true);
1480 }
1481
1482 static int
1483 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1484                                         const union bpf_attr *attr,
1485                                         union bpf_attr __user *uattr)
1486 {
1487         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1488                                                   false, true);
1489 }
1490
1491 static int
1492 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1493                       union bpf_attr __user *uattr)
1494 {
1495         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1496                                                   false, false);
1497 }
1498
1499 static int
1500 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1501                                  const union bpf_attr *attr,
1502                                  union bpf_attr __user *uattr)
1503 {
1504         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1505                                                   false, false);
1506 }
1507
1508 static int
1509 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1510                                  const union bpf_attr *attr,
1511                                  union bpf_attr __user *uattr)
1512 {
1513         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1514                                                   true, true);
1515 }
1516
1517 static int
1518 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1519                                             const union bpf_attr *attr,
1520                                             union bpf_attr __user *uattr)
1521 {
1522         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1523                                                   true, true);
1524 }
1525
1526 static int
1527 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1528                           union bpf_attr __user *uattr)
1529 {
1530         return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1531                                                   true, false);
1532 }
1533
1534 static int
1535 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1536                                      const union bpf_attr *attr,
1537                                      union bpf_attr __user *uattr)
1538 {
1539         return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1540                                                   true, false);
1541 }
1542
1543 const struct bpf_map_ops htab_map_ops = {
1544         .map_alloc_check = htab_map_alloc_check,
1545         .map_alloc = htab_map_alloc,
1546         .map_free = htab_map_free,
1547         .map_get_next_key = htab_map_get_next_key,
1548         .map_lookup_elem = htab_map_lookup_elem,
1549         .map_update_elem = htab_map_update_elem,
1550         .map_delete_elem = htab_map_delete_elem,
1551         .map_gen_lookup = htab_map_gen_lookup,
1552         .map_seq_show_elem = htab_map_seq_show_elem,
1553         BATCH_OPS(htab),
1554 };
1555
1556 const struct bpf_map_ops htab_lru_map_ops = {
1557         .map_alloc_check = htab_map_alloc_check,
1558         .map_alloc = htab_map_alloc,
1559         .map_free = htab_map_free,
1560         .map_get_next_key = htab_map_get_next_key,
1561         .map_lookup_elem = htab_lru_map_lookup_elem,
1562         .map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1563         .map_update_elem = htab_lru_map_update_elem,
1564         .map_delete_elem = htab_lru_map_delete_elem,
1565         .map_gen_lookup = htab_lru_map_gen_lookup,
1566         .map_seq_show_elem = htab_map_seq_show_elem,
1567         BATCH_OPS(htab_lru),
1568 };
1569
1570 /* Called from eBPF program */
1571 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1572 {
1573         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1574
1575         if (l)
1576                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1577         else
1578                 return NULL;
1579 }
1580
1581 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1582 {
1583         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1584
1585         if (l) {
1586                 bpf_lru_node_set_ref(&l->lru_node);
1587                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1588         }
1589
1590         return NULL;
1591 }
1592
1593 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1594 {
1595         struct htab_elem *l;
1596         void __percpu *pptr;
1597         int ret = -ENOENT;
1598         int cpu, off = 0;
1599         u32 size;
1600
1601         /* per_cpu areas are zero-filled and bpf programs can only
1602          * access 'value_size' of them, so copying rounded areas
1603          * will not leak any kernel data
1604          */
1605         size = round_up(map->value_size, 8);
1606         rcu_read_lock();
1607         l = __htab_map_lookup_elem(map, key);
1608         if (!l)
1609                 goto out;
1610         /* We do not mark LRU map element here in order to not mess up
1611          * eviction heuristics when user space does a map walk.
1612          */
1613         pptr = htab_elem_get_ptr(l, map->key_size);
1614         for_each_possible_cpu(cpu) {
1615                 bpf_long_memcpy(value + off,
1616                                 per_cpu_ptr(pptr, cpu), size);
1617                 off += size;
1618         }
1619         ret = 0;
1620 out:
1621         rcu_read_unlock();
1622         return ret;
1623 }
1624
1625 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1626                            u64 map_flags)
1627 {
1628         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1629         int ret;
1630
1631         rcu_read_lock();
1632         if (htab_is_lru(htab))
1633                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1634                                                         map_flags, true);
1635         else
1636                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1637                                                     true);
1638         rcu_read_unlock();
1639
1640         return ret;
1641 }
1642
1643 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1644                                           struct seq_file *m)
1645 {
1646         struct htab_elem *l;
1647         void __percpu *pptr;
1648         int cpu;
1649
1650         rcu_read_lock();
1651
1652         l = __htab_map_lookup_elem(map, key);
1653         if (!l) {
1654                 rcu_read_unlock();
1655                 return;
1656         }
1657
1658         btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1659         seq_puts(m, ": {\n");
1660         pptr = htab_elem_get_ptr(l, map->key_size);
1661         for_each_possible_cpu(cpu) {
1662                 seq_printf(m, "\tcpu%d: ", cpu);
1663                 btf_type_seq_show(map->btf, map->btf_value_type_id,
1664                                   per_cpu_ptr(pptr, cpu), m);
1665                 seq_puts(m, "\n");
1666         }
1667         seq_puts(m, "}\n");
1668
1669         rcu_read_unlock();
1670 }
1671
1672 const struct bpf_map_ops htab_percpu_map_ops = {
1673         .map_alloc_check = htab_map_alloc_check,
1674         .map_alloc = htab_map_alloc,
1675         .map_free = htab_map_free,
1676         .map_get_next_key = htab_map_get_next_key,
1677         .map_lookup_elem = htab_percpu_map_lookup_elem,
1678         .map_update_elem = htab_percpu_map_update_elem,
1679         .map_delete_elem = htab_map_delete_elem,
1680         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1681         BATCH_OPS(htab_percpu),
1682 };
1683
1684 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1685         .map_alloc_check = htab_map_alloc_check,
1686         .map_alloc = htab_map_alloc,
1687         .map_free = htab_map_free,
1688         .map_get_next_key = htab_map_get_next_key,
1689         .map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1690         .map_update_elem = htab_lru_percpu_map_update_elem,
1691         .map_delete_elem = htab_lru_map_delete_elem,
1692         .map_seq_show_elem = htab_percpu_map_seq_show_elem,
1693         BATCH_OPS(htab_lru_percpu),
1694 };
1695
1696 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1697 {
1698         if (attr->value_size != sizeof(u32))
1699                 return -EINVAL;
1700         return htab_map_alloc_check(attr);
1701 }
1702
1703 static void fd_htab_map_free(struct bpf_map *map)
1704 {
1705         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1706         struct hlist_nulls_node *n;
1707         struct hlist_nulls_head *head;
1708         struct htab_elem *l;
1709         int i;
1710
1711         for (i = 0; i < htab->n_buckets; i++) {
1712                 head = select_bucket(htab, i);
1713
1714                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1715                         void *ptr = fd_htab_map_get_ptr(map, l);
1716
1717                         map->ops->map_fd_put_ptr(ptr);
1718                 }
1719         }
1720
1721         htab_map_free(map);
1722 }
1723
1724 /* only called from syscall */
1725 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1726 {
1727         void **ptr;
1728         int ret = 0;
1729
1730         if (!map->ops->map_fd_sys_lookup_elem)
1731                 return -ENOTSUPP;
1732
1733         rcu_read_lock();
1734         ptr = htab_map_lookup_elem(map, key);
1735         if (ptr)
1736                 *value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1737         else
1738                 ret = -ENOENT;
1739         rcu_read_unlock();
1740
1741         return ret;
1742 }
1743
1744 /* only called from syscall */
1745 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1746                                 void *key, void *value, u64 map_flags)
1747 {
1748         void *ptr;
1749         int ret;
1750         u32 ufd = *(u32 *)value;
1751
1752         ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1753         if (IS_ERR(ptr))
1754                 return PTR_ERR(ptr);
1755
1756         ret = htab_map_update_elem(map, key, &ptr, map_flags);
1757         if (ret)
1758                 map->ops->map_fd_put_ptr(ptr);
1759
1760         return ret;
1761 }
1762
1763 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1764 {
1765         struct bpf_map *map, *inner_map_meta;
1766
1767         inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1768         if (IS_ERR(inner_map_meta))
1769                 return inner_map_meta;
1770
1771         map = htab_map_alloc(attr);
1772         if (IS_ERR(map)) {
1773                 bpf_map_meta_free(inner_map_meta);
1774                 return map;
1775         }
1776
1777         map->inner_map_meta = inner_map_meta;
1778
1779         return map;
1780 }
1781
1782 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1783 {
1784         struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1785
1786         if (!inner_map)
1787                 return NULL;
1788
1789         return READ_ONCE(*inner_map);
1790 }
1791
1792 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1793                                   struct bpf_insn *insn_buf)
1794 {
1795         struct bpf_insn *insn = insn_buf;
1796         const int ret = BPF_REG_0;
1797
1798         BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1799                      (void *(*)(struct bpf_map *map, void *key))NULL));
1800         *insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1801         *insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1802         *insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1803                                 offsetof(struct htab_elem, key) +
1804                                 round_up(map->key_size, 8));
1805         *insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1806
1807         return insn - insn_buf;
1808 }
1809
1810 static void htab_of_map_free(struct bpf_map *map)
1811 {
1812         bpf_map_meta_free(map->inner_map_meta);
1813         fd_htab_map_free(map);
1814 }
1815
1816 const struct bpf_map_ops htab_of_maps_map_ops = {
1817         .map_alloc_check = fd_htab_map_alloc_check,
1818         .map_alloc = htab_of_map_alloc,
1819         .map_free = htab_of_map_free,
1820         .map_get_next_key = htab_map_get_next_key,
1821         .map_lookup_elem = htab_of_map_lookup_elem,
1822         .map_delete_elem = htab_map_delete_elem,
1823         .map_fd_get_ptr = bpf_map_fd_get_ptr,
1824         .map_fd_put_ptr = bpf_map_fd_put_ptr,
1825         .map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1826         .map_gen_lookup = htab_of_map_gen_lookup,
1827         .map_check_btf = map_check_no_btf,
1828 };