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
  20 struct bucket {
  21         struct hlist_nulls_head head;
  22         raw_spinlock_t lock;
  23 };
  24
  25 struct bpf_htab {
  26         struct bpf_map map;
  27         struct bucket *buckets;
  28         void *elems;
  29         union {
  30                 struct pcpu_freelist freelist;
  31                 struct bpf_lru lru;
  32         };
  33         struct htab_elem *__percpu *extra_elems;
  34         atomic_t count; /* number of elements in this hashtable */
  35         u32 n_buckets;  /* number of hash buckets */
  36         u32 elem_size;  /* size of each element in bytes */
  37 };
  38
  39 /* each htab element is struct htab_elem + key + value */
  40 struct htab_elem {
  41         union {
  42                 struct hlist_nulls_node hash_node;
  43                 struct {
  44                         void *padding;
  45                         union {
  46                                 struct bpf_htab *htab;
  47                                 struct pcpu_freelist_node fnode;
  48                         };
  49                 };
  50         };
  51         union {
  52                 struct rcu_head rcu;
  53                 struct bpf_lru_node lru_node;
  54         };
  55         u32 hash;
  56         char key[0] __aligned(8);
  57 };
  58
  59 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  60
  61 static bool htab_is_lru(const struct bpf_htab *htab)
  62 {
  63         return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  64                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  65 }
  66
  67 static bool htab_is_percpu(const struct bpf_htab *htab)
  68 {
  69         return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  70                 htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  71 }
  72
  73 static bool htab_is_prealloc(const struct bpf_htab *htab)
  74 {
  75         return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  76 }
  77
  78 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  79                                      void __percpu *pptr)
  80 {
  81         *(void __percpu **)(l->key + key_size) = pptr;
  82 }
  83
  84 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  85 {
  86         return *(void __percpu **)(l->key + key_size);
  87 }
  88
  89 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
  90 {
  91         return (struct htab_elem *) (htab->elems + i * htab->elem_size);
  92 }
  93
  94 static void htab_free_elems(struct bpf_htab *htab)
  95 {
  96         int i;
  97
  98         if (!htab_is_percpu(htab))
  99                 goto free_elems;
 100
 101         for (i = 0; i < htab->map.max_entries; i++) {
 102                 void __percpu *pptr;
 103
 104                 pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 105                                          htab->map.key_size);
 106                 free_percpu(pptr);
 107         }
 108 free_elems:
 109         bpf_map_area_free(htab->elems);
 110 }
 111
 112 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 113                                           u32 hash)
 114 {
 115         struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 116         struct htab_elem *l;
 117
 118         if (node) {
 119                 l = container_of(node, struct htab_elem, lru_node);
 120                 memcpy(l->key, key, htab->map.key_size);
 121                 return l;
 122         }
 123
 124         return NULL;
 125 }
 126
 127 static int prealloc_init(struct bpf_htab *htab)
 128 {
 129         u32 num_entries = htab->map.max_entries;
 130         int err = -ENOMEM, i;
 131
 132         if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 133                 num_entries += num_possible_cpus();
 134
 135         htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries);
 136         if (!htab->elems)
 137                 return -ENOMEM;
 138
 139         if (!htab_is_percpu(htab))
 140                 goto skip_percpu_elems;
 141
 142         for (i = 0; i < num_entries; i++) {
 143                 u32 size = round_up(htab->map.value_size, 8);
 144                 void __percpu *pptr;
 145
 146                 pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 147                 if (!pptr)
 148                         goto free_elems;
 149                 htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 150                                   pptr);
 151         }
 152
 153 skip_percpu_elems:
 154         if (htab_is_lru(htab))
 155                 err = bpf_lru_init(&htab->lru,
 156                                    htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 157                                    offsetof(struct htab_elem, hash) -
 158                                    offsetof(struct htab_elem, lru_node),
 159                                    htab_lru_map_delete_node,
 160                                    htab);
 161         else
 162                 err = pcpu_freelist_init(&htab->freelist);
 163
 164         if (err)
 165                 goto free_elems;
 166
 167         if (htab_is_lru(htab))
 168                 bpf_lru_populate(&htab->lru, htab->elems,
 169                                  offsetof(struct htab_elem, lru_node),
 170                                  htab->elem_size, num_entries);
 171         else
 172                 pcpu_freelist_populate(&htab->freelist,
 173                                        htab->elems + offsetof(struct htab_elem, fnode),
 174                                        htab->elem_size, num_entries);
 175
 176         return 0;
 177
 178 free_elems:
 179         htab_free_elems(htab);
 180         return err;
 181 }
 182
 183 static void prealloc_destroy(struct bpf_htab *htab)
 184 {
 185         htab_free_elems(htab);
 186
 187         if (htab_is_lru(htab))
 188                 bpf_lru_destroy(&htab->lru);
 189         else
 190                 pcpu_freelist_destroy(&htab->freelist);
 191 }
 192
 193 static int alloc_extra_elems(struct bpf_htab *htab)
 194 {
 195         struct htab_elem *__percpu *pptr, *l_new;
 196         struct pcpu_freelist_node *l;
 197         int cpu;
 198
 199         pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 200                                   GFP_USER | __GFP_NOWARN);
 201         if (!pptr)
 202                 return -ENOMEM;
 203
 204         for_each_possible_cpu(cpu) {
 205                 l = pcpu_freelist_pop(&htab->freelist);
 206                 /* pop will succeed, since prealloc_init()
 207                  * preallocated extra num_possible_cpus elements
 208                  */
 209                 l_new = container_of(l, struct htab_elem, fnode);
 210                 *per_cpu_ptr(pptr, cpu) = l_new;
 211         }
 212         htab->extra_elems = pptr;
 213         return 0;
 214 }
 215
 216 /* Called from syscall */
 217 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 218 {
 219         bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 220                        attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 221         bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 222                     attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 223         /* percpu_lru means each cpu has its own LRU list.
 224          * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 225          * the map's value itself is percpu.  percpu_lru has
 226          * nothing to do with the map's value.
 227          */
 228         bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 229         bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 230         struct bpf_htab *htab;
 231         int err, i;
 232         u64 cost;
 233
 234         BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 235                      offsetof(struct htab_elem, hash_node.pprev));
 236         BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 237                      offsetof(struct htab_elem, hash_node.pprev));
 238
 239         if (lru && !capable(CAP_SYS_ADMIN))
 240                 /* LRU implementation is much complicated than other
 241                  * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 242                  */
 243                 return ERR_PTR(-EPERM);
 244
 245         if (attr->map_flags & ~(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU))
 246                 /* reserved bits should not be used */
 247                 return ERR_PTR(-EINVAL);
 248
 249         if (!lru && percpu_lru)
 250                 return ERR_PTR(-EINVAL);
 251
 252         if (lru && !prealloc)
 253                 return ERR_PTR(-ENOTSUPP);
 254
 255         htab = kzalloc(sizeof(*htab), GFP_USER);
 256         if (!htab)
 257                 return ERR_PTR(-ENOMEM);
 258
 259         /* mandatory map attributes */
 260         htab->map.map_type = attr->map_type;
 261         htab->map.key_size = attr->key_size;
 262         htab->map.value_size = attr->value_size;
 263         htab->map.max_entries = attr->max_entries;
 264         htab->map.map_flags = attr->map_flags;
 265
 266         /* check sanity of attributes.
 267          * value_size == 0 may be allowed in the future to use map as a set
 268          */
 269         err = -EINVAL;
 270         if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
 271             htab->map.value_size == 0)
 272                 goto free_htab;
 273
 274         if (percpu_lru) {
 275                 /* ensure each CPU's lru list has >=1 elements.
 276                  * since we are at it, make each lru list has the same
 277                  * number of elements.
 278                  */
 279                 htab->map.max_entries = roundup(attr->max_entries,
 280                                                 num_possible_cpus());
 281                 if (htab->map.max_entries < attr->max_entries)
 282                         htab->map.max_entries = rounddown(attr->max_entries,
 283                                                           num_possible_cpus());
 284         }
 285
 286         /* hash table size must be power of 2 */
 287         htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 288
 289         err = -E2BIG;
 290         if (htab->map.key_size > MAX_BPF_STACK)
 291                 /* eBPF programs initialize keys on stack, so they cannot be
 292                  * larger than max stack size
 293                  */
 294                 goto free_htab;
 295
 296         if (htab->map.value_size >= KMALLOC_MAX_SIZE -
 297             MAX_BPF_STACK - sizeof(struct htab_elem))
 298                 /* if value_size is bigger, the user space won't be able to
 299                  * access the elements via bpf syscall. This check also makes
 300                  * sure that the elem_size doesn't overflow and it's
 301                  * kmalloc-able later in htab_map_update_elem()
 302                  */
 303                 goto free_htab;
 304
 305         if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
 306                 /* make sure the size for pcpu_alloc() is reasonable */
 307                 goto free_htab;
 308
 309         htab->elem_size = sizeof(struct htab_elem) +
 310                           round_up(htab->map.key_size, 8);
 311         if (percpu)
 312                 htab->elem_size += sizeof(void *);
 313         else
 314                 htab->elem_size += round_up(htab->map.value_size, 8);
 315
 316         /* prevent zero size kmalloc and check for u32 overflow */
 317         if (htab->n_buckets == 0 ||
 318             htab->n_buckets > U32_MAX / sizeof(struct bucket))
 319                 goto free_htab;
 320
 321         cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 322                (u64) htab->elem_size * htab->map.max_entries;
 323
 324         if (percpu)
 325                 cost += (u64) round_up(htab->map.value_size, 8) *
 326                         num_possible_cpus() * htab->map.max_entries;
 327         else
 328                cost += (u64) htab->elem_size * num_possible_cpus();
 329
 330         if (cost >= U32_MAX - PAGE_SIZE)
 331                 /* make sure page count doesn't overflow */
 332                 goto free_htab;
 333
 334         htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 335
 336         /* if map size is larger than memlock limit, reject it early */
 337         err = bpf_map_precharge_memlock(htab->map.pages);
 338         if (err)
 339                 goto free_htab;
 340
 341         err = -ENOMEM;
 342         htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 343                                            sizeof(struct bucket));
 344         if (!htab->buckets)
 345                 goto free_htab;
 346
 347         for (i = 0; i < htab->n_buckets; i++) {
 348                 INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 349                 raw_spin_lock_init(&htab->buckets[i].lock);
 350         }
 351
 352         if (prealloc) {
 353                 err = prealloc_init(htab);
 354                 if (err)
 355                         goto free_buckets;
 356
 357                 if (!percpu && !lru) {
 358                         /* lru itself can remove the least used element, so
 359                          * there is no need for an extra elem during map_update.
 360                          */
 361                         err = alloc_extra_elems(htab);
 362                         if (err)
 363                                 goto free_prealloc;
 364                 }
 365         }
 366
 367         return &htab->map;
 368
 369 free_prealloc:
 370         prealloc_destroy(htab);
 371 free_buckets:
 372         bpf_map_area_free(htab->buckets);
 373 free_htab:
 374         kfree(htab);
 375         return ERR_PTR(err);
 376 }
 377
 378 static inline u32 htab_map_hash(const void *key, u32 key_len)
 379 {
 380         return jhash(key, key_len, 0);
 381 }
 382
 383 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 384 {
 385         return &htab->buckets[hash & (htab->n_buckets - 1)];
 386 }
 387
 388 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 389 {
 390         return &__select_bucket(htab, hash)->head;
 391 }
 392
 393 /* this lookup function can only be called with bucket lock taken */
 394 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 395                                          void *key, u32 key_size)
 396 {
 397         struct hlist_nulls_node *n;
 398         struct htab_elem *l;
 399
 400         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 401                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 402                         return l;
 403
 404         return NULL;
 405 }
 406
 407 /* can be called without bucket lock. it will repeat the loop in
 408  * the unlikely event when elements moved from one bucket into another
 409  * while link list is being walked
 410  */
 411 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 412                                                u32 hash, void *key,
 413                                                u32 key_size, u32 n_buckets)
 414 {
 415         struct hlist_nulls_node *n;
 416         struct htab_elem *l;
 417
 418 again:
 419         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 420                 if (l->hash == hash && !memcmp(&l->key, key, key_size))
 421                         return l;
 422
 423         if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 424                 goto again;
 425
 426         return NULL;
 427 }
 428
 429 /* Called from syscall or from eBPF program */
 430 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 431 {
 432         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 433         struct hlist_nulls_head *head;
 434         struct htab_elem *l;
 435         u32 hash, key_size;
 436
 437         /* Must be called with rcu_read_lock. */
 438         WARN_ON_ONCE(!rcu_read_lock_held());
 439
 440         key_size = map->key_size;
 441
 442         hash = htab_map_hash(key, key_size);
 443
 444         head = select_bucket(htab, hash);
 445
 446         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 447
 448         return l;
 449 }
 450
 451 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 452 {
 453         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 454
 455         if (l)
 456                 return l->key + round_up(map->key_size, 8);
 457
 458         return NULL;
 459 }
 460
 461 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 462 {
 463         struct htab_elem *l = __htab_map_lookup_elem(map, key);
 464
 465         if (l) {
 466                 bpf_lru_node_set_ref(&l->lru_node);
 467                 return l->key + round_up(map->key_size, 8);
 468         }
 469
 470         return NULL;
 471 }
 472
 473 /* It is called from the bpf_lru_list when the LRU needs to delete
 474  * older elements from the htab.
 475  */
 476 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 477 {
 478         struct bpf_htab *htab = (struct bpf_htab *)arg;
 479         struct htab_elem *l = NULL, *tgt_l;
 480         struct hlist_nulls_head *head;
 481         struct hlist_nulls_node *n;
 482         unsigned long flags;
 483         struct bucket *b;
 484
 485         tgt_l = container_of(node, struct htab_elem, lru_node);
 486         b = __select_bucket(htab, tgt_l->hash);
 487         head = &b->head;
 488
 489         raw_spin_lock_irqsave(&b->lock, flags);
 490
 491         hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 492                 if (l == tgt_l) {
 493                         hlist_nulls_del_rcu(&l->hash_node);
 494                         break;
 495                 }
 496
 497         raw_spin_unlock_irqrestore(&b->lock, flags);
 498
 499         return l == tgt_l;
 500 }
 501
 502 /* Called from syscall */
 503 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 504 {
 505         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 506         struct hlist_nulls_head *head;
 507         struct htab_elem *l, *next_l;
 508         u32 hash, key_size;
 509         int i;
 510
 511         WARN_ON_ONCE(!rcu_read_lock_held());
 512
 513         key_size = map->key_size;
 514
 515         hash = htab_map_hash(key, key_size);
 516
 517         head = select_bucket(htab, hash);
 518
 519         /* lookup the key */
 520         l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 521
 522         if (!l) {
 523                 i = 0;
 524                 goto find_first_elem;
 525         }
 526
 527         /* key was found, get next key in the same bucket */
 528         next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 529                                   struct htab_elem, hash_node);
 530
 531         if (next_l) {
 532                 /* if next elem in this hash list is non-zero, just return it */
 533                 memcpy(next_key, next_l->key, key_size);
 534                 return 0;
 535         }
 536
 537         /* no more elements in this hash list, go to the next bucket */
 538         i = hash & (htab->n_buckets - 1);
 539         i++;
 540
 541 find_first_elem:
 542         /* iterate over buckets */
 543         for (; i < htab->n_buckets; i++) {
 544                 head = select_bucket(htab, i);
 545
 546                 /* pick first element in the bucket */
 547                 next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 548                                           struct htab_elem, hash_node);
 549                 if (next_l) {
 550                         /* if it's not empty, just return it */
 551                         memcpy(next_key, next_l->key, key_size);
 552                         return 0;
 553                 }
 554         }
 555
 556         /* iterated over all buckets and all elements */
 557         return -ENOENT;
 558 }
 559
 560 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 561 {
 562         if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 563                 free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 564         kfree(l);
 565 }
 566
 567 static void htab_elem_free_rcu(struct rcu_head *head)
 568 {
 569         struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 570         struct bpf_htab *htab = l->htab;
 571
 572         /* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 573          * we're calling kfree, otherwise deadlock is possible if kprobes
 574          * are placed somewhere inside of slub
 575          */
 576         preempt_disable();
 577         __this_cpu_inc(bpf_prog_active);
 578         htab_elem_free(htab, l);
 579         __this_cpu_dec(bpf_prog_active);
 580         preempt_enable();
 581 }
 582
 583 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 584 {
 585         if (htab_is_prealloc(htab)) {
 586                 pcpu_freelist_push(&htab->freelist, &l->fnode);
 587         } else {
 588                 atomic_dec(&htab->count);
 589                 l->htab = htab;
 590                 call_rcu(&l->rcu, htab_elem_free_rcu);
 591         }
 592 }
 593
 594 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 595                             void *value, bool onallcpus)
 596 {
 597         if (!onallcpus) {
 598                 /* copy true value_size bytes */
 599                 memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 600         } else {
 601                 u32 size = round_up(htab->map.value_size, 8);
 602                 int off = 0, cpu;
 603
 604                 for_each_possible_cpu(cpu) {
 605                         bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 606                                         value + off, size);
 607                         off += size;
 608                 }
 609         }
 610 }
 611
 612 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 613                                          void *value, u32 key_size, u32 hash,
 614                                          bool percpu, bool onallcpus,
 615                                          struct htab_elem *old_elem)
 616 {
 617         u32 size = htab->map.value_size;
 618         bool prealloc = htab_is_prealloc(htab);
 619         struct htab_elem *l_new, **pl_new;
 620         void __percpu *pptr;
 621
 622         if (prealloc) {
 623                 if (old_elem) {
 624                         /* if we're updating the existing element,
 625                          * use per-cpu extra elems to avoid freelist_pop/push
 626                          */
 627                         pl_new = this_cpu_ptr(htab->extra_elems);
 628                         l_new = *pl_new;
 629                         *pl_new = old_elem;
 630                 } else {
 631                         struct pcpu_freelist_node *l;
 632
 633                         l = pcpu_freelist_pop(&htab->freelist);
 634                         if (!l)
 635                                 return ERR_PTR(-E2BIG);
 636                         l_new = container_of(l, struct htab_elem, fnode);
 637                 }
 638         } else {
 639                 if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 640                         if (!old_elem) {
 641                                 /* when map is full and update() is replacing
 642                                  * old element, it's ok to allocate, since
 643                                  * old element will be freed immediately.
 644                                  * Otherwise return an error
 645                                  */
 646                                 atomic_dec(&htab->count);
 647                                 return ERR_PTR(-E2BIG);
 648                         }
 649                 l_new = kmalloc(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN);
 650                 if (!l_new)
 651                         return ERR_PTR(-ENOMEM);
 652         }
 653
 654         memcpy(l_new->key, key, key_size);
 655         if (percpu) {
 656                 /* round up value_size to 8 bytes */
 657                 size = round_up(size, 8);
 658
 659                 if (prealloc) {
 660                         pptr = htab_elem_get_ptr(l_new, key_size);
 661                 } else {
 662                         /* alloc_percpu zero-fills */
 663                         pptr = __alloc_percpu_gfp(size, 8,
 664                                                   GFP_ATOMIC | __GFP_NOWARN);
 665                         if (!pptr) {
 666                                 kfree(l_new);
 667                                 return ERR_PTR(-ENOMEM);
 668                         }
 669                 }
 670
 671                 pcpu_copy_value(htab, pptr, value, onallcpus);
 672
 673                 if (!prealloc)
 674                         htab_elem_set_ptr(l_new, key_size, pptr);
 675         } else {
 676                 memcpy(l_new->key + round_up(key_size, 8), value, size);
 677         }
 678
 679         l_new->hash = hash;
 680         return l_new;
 681 }
 682
 683 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 684                        u64 map_flags)
 685 {
 686         if (l_old && map_flags == BPF_NOEXIST)
 687                 /* elem already exists */
 688                 return -EEXIST;
 689
 690         if (!l_old && map_flags == BPF_EXIST)
 691                 /* elem doesn't exist, cannot update it */
 692                 return -ENOENT;
 693
 694         return 0;
 695 }
 696
 697 /* Called from syscall or from eBPF program */
 698 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 699                                 u64 map_flags)
 700 {
 701         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 702         struct htab_elem *l_new = NULL, *l_old;
 703         struct hlist_nulls_head *head;
 704         unsigned long flags;
 705         struct bucket *b;
 706         u32 key_size, hash;
 707         int ret;
 708
 709         if (unlikely(map_flags > BPF_EXIST))
 710                 /* unknown flags */
 711                 return -EINVAL;
 712
 713         WARN_ON_ONCE(!rcu_read_lock_held());
 714
 715         key_size = map->key_size;
 716
 717         hash = htab_map_hash(key, key_size);
 718
 719         b = __select_bucket(htab, hash);
 720         head = &b->head;
 721
 722         /* bpf_map_update_elem() can be called in_irq() */
 723         raw_spin_lock_irqsave(&b->lock, flags);
 724
 725         l_old = lookup_elem_raw(head, hash, key, key_size);
 726
 727         ret = check_flags(htab, l_old, map_flags);
 728         if (ret)
 729                 goto err;
 730
 731         l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 732                                 l_old);
 733         if (IS_ERR(l_new)) {
 734                 /* all pre-allocated elements are in use or memory exhausted */
 735                 ret = PTR_ERR(l_new);
 736                 goto err;
 737         }
 738
 739         /* add new element to the head of the list, so that
 740          * concurrent search will find it before old elem
 741          */
 742         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 743         if (l_old) {
 744                 hlist_nulls_del_rcu(&l_old->hash_node);
 745                 if (!htab_is_prealloc(htab))
 746                         free_htab_elem(htab, l_old);
 747         }
 748         ret = 0;
 749 err:
 750         raw_spin_unlock_irqrestore(&b->lock, flags);
 751         return ret;
 752 }
 753
 754 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 755                                     u64 map_flags)
 756 {
 757         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 758         struct htab_elem *l_new, *l_old = NULL;
 759         struct hlist_nulls_head *head;
 760         unsigned long flags;
 761         struct bucket *b;
 762         u32 key_size, hash;
 763         int ret;
 764
 765         if (unlikely(map_flags > BPF_EXIST))
 766                 /* unknown flags */
 767                 return -EINVAL;
 768
 769         WARN_ON_ONCE(!rcu_read_lock_held());
 770
 771         key_size = map->key_size;
 772
 773         hash = htab_map_hash(key, key_size);
 774
 775         b = __select_bucket(htab, hash);
 776         head = &b->head;
 777
 778         /* For LRU, we need to alloc before taking bucket's
 779          * spinlock because getting free nodes from LRU may need
 780          * to remove older elements from htab and this removal
 781          * operation will need a bucket lock.
 782          */
 783         l_new = prealloc_lru_pop(htab, key, hash);
 784         if (!l_new)
 785                 return -ENOMEM;
 786         memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 787
 788         /* bpf_map_update_elem() can be called in_irq() */
 789         raw_spin_lock_irqsave(&b->lock, flags);
 790
 791         l_old = lookup_elem_raw(head, hash, key, key_size);
 792
 793         ret = check_flags(htab, l_old, map_flags);
 794         if (ret)
 795                 goto err;
 796
 797         /* add new element to the head of the list, so that
 798          * concurrent search will find it before old elem
 799          */
 800         hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 801         if (l_old) {
 802                 bpf_lru_node_set_ref(&l_new->lru_node);
 803                 hlist_nulls_del_rcu(&l_old->hash_node);
 804         }
 805         ret = 0;
 806
 807 err:
 808         raw_spin_unlock_irqrestore(&b->lock, flags);
 809
 810         if (ret)
 811                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 812         else if (l_old)
 813                 bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 814
 815         return ret;
 816 }
 817
 818 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 819                                          void *value, u64 map_flags,
 820                                          bool onallcpus)
 821 {
 822         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 823         struct htab_elem *l_new = NULL, *l_old;
 824         struct hlist_nulls_head *head;
 825         unsigned long flags;
 826         struct bucket *b;
 827         u32 key_size, hash;
 828         int ret;
 829
 830         if (unlikely(map_flags > BPF_EXIST))
 831                 /* unknown flags */
 832                 return -EINVAL;
 833
 834         WARN_ON_ONCE(!rcu_read_lock_held());
 835
 836         key_size = map->key_size;
 837
 838         hash = htab_map_hash(key, key_size);
 839
 840         b = __select_bucket(htab, hash);
 841         head = &b->head;
 842
 843         /* bpf_map_update_elem() can be called in_irq() */
 844         raw_spin_lock_irqsave(&b->lock, flags);
 845
 846         l_old = lookup_elem_raw(head, hash, key, key_size);
 847
 848         ret = check_flags(htab, l_old, map_flags);
 849         if (ret)
 850                 goto err;
 851
 852         if (l_old) {
 853                 /* per-cpu hash map can update value in-place */
 854                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 855                                 value, onallcpus);
 856         } else {
 857                 l_new = alloc_htab_elem(htab, key, value, key_size,
 858                                         hash, true, onallcpus, NULL);
 859                 if (IS_ERR(l_new)) {
 860                         ret = PTR_ERR(l_new);
 861                         goto err;
 862                 }
 863                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 864         }
 865         ret = 0;
 866 err:
 867         raw_spin_unlock_irqrestore(&b->lock, flags);
 868         return ret;
 869 }
 870
 871 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 872                                              void *value, u64 map_flags,
 873                                              bool onallcpus)
 874 {
 875         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 876         struct htab_elem *l_new = NULL, *l_old;
 877         struct hlist_nulls_head *head;
 878         unsigned long flags;
 879         struct bucket *b;
 880         u32 key_size, hash;
 881         int ret;
 882
 883         if (unlikely(map_flags > BPF_EXIST))
 884                 /* unknown flags */
 885                 return -EINVAL;
 886
 887         WARN_ON_ONCE(!rcu_read_lock_held());
 888
 889         key_size = map->key_size;
 890
 891         hash = htab_map_hash(key, key_size);
 892
 893         b = __select_bucket(htab, hash);
 894         head = &b->head;
 895
 896         /* For LRU, we need to alloc before taking bucket's
 897          * spinlock because LRU's elem alloc may need
 898          * to remove older elem from htab and this removal
 899          * operation will need a bucket lock.
 900          */
 901         if (map_flags != BPF_EXIST) {
 902                 l_new = prealloc_lru_pop(htab, key, hash);
 903                 if (!l_new)
 904                         return -ENOMEM;
 905         }
 906
 907         /* bpf_map_update_elem() can be called in_irq() */
 908         raw_spin_lock_irqsave(&b->lock, flags);
 909
 910         l_old = lookup_elem_raw(head, hash, key, key_size);
 911
 912         ret = check_flags(htab, l_old, map_flags);
 913         if (ret)
 914                 goto err;
 915
 916         if (l_old) {
 917                 bpf_lru_node_set_ref(&l_old->lru_node);
 918
 919                 /* per-cpu hash map can update value in-place */
 920                 pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 921                                 value, onallcpus);
 922         } else {
 923                 pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
 924                                 value, onallcpus);
 925                 hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 926                 l_new = NULL;
 927         }
 928         ret = 0;
 929 err:
 930         raw_spin_unlock_irqrestore(&b->lock, flags);
 931         if (l_new)
 932                 bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 933         return ret;
 934 }
 935
 936 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 937                                        void *value, u64 map_flags)
 938 {
 939         return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
 940 }
 941
 942 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 943                                            void *value, u64 map_flags)
 944 {
 945         return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
 946                                                  false);
 947 }
 948
 949 /* Called from syscall or from eBPF program */
 950 static int htab_map_delete_elem(struct bpf_map *map, void *key)
 951 {
 952         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 953         struct hlist_nulls_head *head;
 954         struct bucket *b;
 955         struct htab_elem *l;
 956         unsigned long flags;
 957         u32 hash, key_size;
 958         int ret = -ENOENT;
 959
 960         WARN_ON_ONCE(!rcu_read_lock_held());
 961
 962         key_size = map->key_size;
 963
 964         hash = htab_map_hash(key, key_size);
 965         b = __select_bucket(htab, hash);
 966         head = &b->head;
 967
 968         raw_spin_lock_irqsave(&b->lock, flags);
 969
 970         l = lookup_elem_raw(head, hash, key, key_size);
 971
 972         if (l) {
 973                 hlist_nulls_del_rcu(&l->hash_node);
 974                 free_htab_elem(htab, l);
 975                 ret = 0;
 976         }
 977
 978         raw_spin_unlock_irqrestore(&b->lock, flags);
 979         return ret;
 980 }
 981
 982 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
 983 {
 984         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 985         struct hlist_nulls_head *head;
 986         struct bucket *b;
 987         struct htab_elem *l;
 988         unsigned long flags;
 989         u32 hash, key_size;
 990         int ret = -ENOENT;
 991
 992         WARN_ON_ONCE(!rcu_read_lock_held());
 993
 994         key_size = map->key_size;
 995
 996         hash = htab_map_hash(key, key_size);
 997         b = __select_bucket(htab, hash);
 998         head = &b->head;
 999
1000         raw_spin_lock_irqsave(&b->lock, flags);
1001
1002         l = lookup_elem_raw(head, hash, key, key_size);
1003
1004         if (l) {
1005                 hlist_nulls_del_rcu(&l->hash_node);
1006                 ret = 0;
1007         }
1008
1009         raw_spin_unlock_irqrestore(&b->lock, flags);
1010         if (l)
1011                 bpf_lru_push_free(&htab->lru, &l->lru_node);
1012         return ret;
1013 }
1014
1015 static void delete_all_elements(struct bpf_htab *htab)
1016 {
1017         int i;
1018
1019         for (i = 0; i < htab->n_buckets; i++) {
1020                 struct hlist_nulls_head *head = select_bucket(htab, i);
1021                 struct hlist_nulls_node *n;
1022                 struct htab_elem *l;
1023
1024                 hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1025                         hlist_nulls_del_rcu(&l->hash_node);
1026                         htab_elem_free(htab, l);
1027                 }
1028         }
1029 }
1030 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1031 static void htab_map_free(struct bpf_map *map)
1032 {
1033         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1034
1035         /* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1036          * so the programs (can be more than one that used this map) were
1037          * disconnected from events. Wait for outstanding critical sections in
1038          * these programs to complete
1039          */
1040         synchronize_rcu();
1041
1042         /* some of free_htab_elem() callbacks for elements of this map may
1043          * not have executed. Wait for them.
1044          */
1045         rcu_barrier();
1046         if (!htab_is_prealloc(htab))
1047                 delete_all_elements(htab);
1048         else
1049                 prealloc_destroy(htab);
1050
1051         free_percpu(htab->extra_elems);
1052         bpf_map_area_free(htab->buckets);
1053         kfree(htab);
1054 }
1055
1056 static const struct bpf_map_ops htab_ops = {
1057         .map_alloc = htab_map_alloc,
1058         .map_free = htab_map_free,
1059         .map_get_next_key = htab_map_get_next_key,
1060         .map_lookup_elem = htab_map_lookup_elem,
1061         .map_update_elem = htab_map_update_elem,
1062         .map_delete_elem = htab_map_delete_elem,
1063 };
1064
1065 static struct bpf_map_type_list htab_type __ro_after_init = {
1066         .ops = &htab_ops,
1067         .type = BPF_MAP_TYPE_HASH,
1068 };
1069
1070 static const struct bpf_map_ops htab_lru_ops = {
1071         .map_alloc = htab_map_alloc,
1072         .map_free = htab_map_free,
1073         .map_get_next_key = htab_map_get_next_key,
1074         .map_lookup_elem = htab_lru_map_lookup_elem,
1075         .map_update_elem = htab_lru_map_update_elem,
1076         .map_delete_elem = htab_lru_map_delete_elem,
1077 };
1078
1079 static struct bpf_map_type_list htab_lru_type __ro_after_init = {
1080         .ops = &htab_lru_ops,
1081         .type = BPF_MAP_TYPE_LRU_HASH,
1082 };
1083
1084 /* Called from eBPF program */
1085 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1086 {
1087         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1088
1089         if (l)
1090                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1091         else
1092                 return NULL;
1093 }
1094
1095 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1096 {
1097         struct htab_elem *l = __htab_map_lookup_elem(map, key);
1098
1099         if (l) {
1100                 bpf_lru_node_set_ref(&l->lru_node);
1101                 return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1102         }
1103
1104         return NULL;
1105 }
1106
1107 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1108 {
1109         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1110         struct htab_elem *l;
1111         void __percpu *pptr;
1112         int ret = -ENOENT;
1113         int cpu, off = 0;
1114         u32 size;
1115
1116         /* per_cpu areas are zero-filled and bpf programs can only
1117          * access 'value_size' of them, so copying rounded areas
1118          * will not leak any kernel data
1119          */
1120         size = round_up(map->value_size, 8);
1121         rcu_read_lock();
1122         l = __htab_map_lookup_elem(map, key);
1123         if (!l)
1124                 goto out;
1125         if (htab_is_lru(htab))
1126                 bpf_lru_node_set_ref(&l->lru_node);
1127         pptr = htab_elem_get_ptr(l, map->key_size);
1128         for_each_possible_cpu(cpu) {
1129                 bpf_long_memcpy(value + off,
1130                                 per_cpu_ptr(pptr, cpu), size);
1131                 off += size;
1132         }
1133         ret = 0;
1134 out:
1135         rcu_read_unlock();
1136         return ret;
1137 }
1138
1139 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1140                            u64 map_flags)
1141 {
1142         struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1143         int ret;
1144
1145         rcu_read_lock();
1146         if (htab_is_lru(htab))
1147                 ret = __htab_lru_percpu_map_update_elem(map, key, value,
1148                                                         map_flags, true);
1149         else
1150                 ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1151                                                     true);
1152         rcu_read_unlock();
1153
1154         return ret;
1155 }
1156
1157 static const struct bpf_map_ops htab_percpu_ops = {
1158         .map_alloc = htab_map_alloc,
1159         .map_free = htab_map_free,
1160         .map_get_next_key = htab_map_get_next_key,
1161         .map_lookup_elem = htab_percpu_map_lookup_elem,
1162         .map_update_elem = htab_percpu_map_update_elem,
1163         .map_delete_elem = htab_map_delete_elem,
1164 };
1165
1166 static struct bpf_map_type_list htab_percpu_type __ro_after_init = {
1167         .ops = &htab_percpu_ops,
1168         .type = BPF_MAP_TYPE_PERCPU_HASH,
1169 };
1170
1171 static const struct bpf_map_ops htab_lru_percpu_ops = {
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_percpu_map_lookup_elem,
1176         .map_update_elem = htab_lru_percpu_map_update_elem,
1177         .map_delete_elem = htab_lru_map_delete_elem,
1178 };
1179
1180 static struct bpf_map_type_list htab_lru_percpu_type __ro_after_init = {
1181         .ops = &htab_lru_percpu_ops,
1182         .type = BPF_MAP_TYPE_LRU_PERCPU_HASH,
1183 };
1184
1185 static int __init register_htab_map(void)
1186 {
1187         bpf_register_map_type(&htab_type);
1188         bpf_register_map_type(&htab_percpu_type);
1189         bpf_register_map_type(&htab_lru_type);
1190         bpf_register_map_type(&htab_lru_percpu_type);
1191         return 0;
1192 }
1193 late_initcall(register_htab_map);