kernel/bpf/hashtab.c

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