lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
   5  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   6  *
   7  * Based on the following paper:
   8  * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
   9  *
  10  * Code partially derived from nft_hash
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2 as
  14  * published by the Free Software Foundation.
  15  */
  16
  17 #include <linux/kernel.h>
  18 #include <linux/init.h>
  19 #include <linux/log2.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/mm.h>
  24 #include <linux/jhash.h>
  25 #include <linux/random.h>
  26 #include <linux/rhashtable.h>
  27 #include <linux/err.h>
  28
  29 #define HASH_DEFAULT_SIZE       64UL
  30 #define HASH_MIN_SIZE           4UL
  31 #define BUCKET_LOCKS_PER_CPU   128UL
  32
  33 /* Base bits plus 1 bit for nulls marker */
  34 #define HASH_RESERVED_SPACE     (RHT_BASE_BITS + 1)
  35
  36 enum {
  37         RHT_LOCK_NORMAL,
  38         RHT_LOCK_NESTED,
  39 };
  40
  41 /* The bucket lock is selected based on the hash and protects mutations
  42  * on a group of hash buckets.
  43  *
  44  * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
  45  * a single lock always covers both buckets which may both contains
  46  * entries which link to the same bucket of the old table during resizing.
  47  * This allows to simplify the locking as locking the bucket in both
  48  * tables during resize always guarantee protection.
  49  *
  50  * IMPORTANT: When holding the bucket lock of both the old and new table
  51  * during expansions and shrinking, the old bucket lock must always be
  52  * acquired first.
  53  */
  54 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  55 {
  56         return &tbl->locks[hash & tbl->locks_mask];
  57 }
  58
  59 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  60 {
  61         return (void *) he - ht->p.head_offset;
  62 }
  63
  64 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  65 {
  66         return hash & (tbl->size - 1);
  67 }
  68
  69 static u32 obj_raw_hashfn(struct rhashtable *ht,
  70                           const struct bucket_table *tbl, const void *ptr)
  71 {
  72         u32 hash;
  73
  74         if (unlikely(!ht->p.key_len))
  75                 hash = ht->p.obj_hashfn(ptr, tbl->hash_rnd);
  76         else
  77                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  78                                     tbl->hash_rnd);
  79
  80         return hash >> HASH_RESERVED_SPACE;
  81 }
  82
  83 static u32 key_hashfn(struct rhashtable *ht, const struct bucket_table *tbl,
  84                       const void *key, u32 len)
  85 {
  86         return rht_bucket_index(tbl, ht->p.hashfn(key, len, tbl->hash_rnd) >>
  87                                      HASH_RESERVED_SPACE);
  88 }
  89
  90 static u32 head_hashfn(struct rhashtable *ht,
  91                        const struct bucket_table *tbl,
  92                        const struct rhash_head *he)
  93 {
  94         return rht_bucket_index(tbl, obj_raw_hashfn(ht, tbl, rht_obj(ht, he)));
  95 }
  96
  97 #ifdef CONFIG_PROVE_LOCKING
  98 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  99
 100 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
 101 {
 102         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
 103 }
 104 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
 105
 106 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
 107 {
 108         spinlock_t *lock = bucket_lock(tbl, hash);
 109
 110         return (debug_locks) ? lockdep_is_held(lock) : 1;
 111 }
 112 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
 113 #else
 114 #define ASSERT_RHT_MUTEX(HT)
 115 #endif
 116
 117
 118 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 119 {
 120         unsigned int i, size;
 121 #if defined(CONFIG_PROVE_LOCKING)
 122         unsigned int nr_pcpus = 2;
 123 #else
 124         unsigned int nr_pcpus = num_possible_cpus();
 125 #endif
 126
 127         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 128         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 129
 130         /* Never allocate more than 0.5 locks per bucket */
 131         size = min_t(unsigned int, size, tbl->size >> 1);
 132
 133         if (sizeof(spinlock_t) != 0) {
 134 #ifdef CONFIG_NUMA
 135                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 136                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 137                 else
 138 #endif
 139                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 140                                            GFP_KERNEL);
 141                 if (!tbl->locks)
 142                         return -ENOMEM;
 143                 for (i = 0; i < size; i++)
 144                         spin_lock_init(&tbl->locks[i]);
 145         }
 146         tbl->locks_mask = size - 1;
 147
 148         return 0;
 149 }
 150
 151 static void bucket_table_free(const struct bucket_table *tbl)
 152 {
 153         if (tbl)
 154                 kvfree(tbl->locks);
 155
 156         kvfree(tbl);
 157 }
 158
 159 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 160                                                size_t nbuckets)
 161 {
 162         struct bucket_table *tbl = NULL;
 163         size_t size;
 164         int i;
 165
 166         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 167         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER))
 168                 tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
 169         if (tbl == NULL)
 170                 tbl = vzalloc(size);
 171         if (tbl == NULL)
 172                 return NULL;
 173
 174         tbl->size = nbuckets;
 175
 176         if (alloc_bucket_locks(ht, tbl) < 0) {
 177                 bucket_table_free(tbl);
 178                 return NULL;
 179         }
 180
 181         for (i = 0; i < nbuckets; i++)
 182                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 183
 184         return tbl;
 185 }
 186
 187 /**
 188  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 189  * @ht:         hash table
 190  * @new_size:   new table size
 191  */
 192 static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 193 {
 194         /* Expand table when exceeding 75% load */
 195         return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
 196                (!ht->p.max_shift || atomic_read(&ht->shift) < ht->p.max_shift);
 197 }
 198
 199 /**
 200  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 201  * @ht:         hash table
 202  * @new_size:   new table size
 203  */
 204 static bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 205 {
 206         /* Shrink table beneath 30% load */
 207         return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
 208                (atomic_read(&ht->shift) > ht->p.min_shift);
 209 }
 210
 211 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash)
 212 {
 213         struct bucket_table *new_tbl = rht_dereference(ht->future_tbl, ht);
 214         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 215         struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
 216         int err = -ENOENT;
 217         struct rhash_head *head, *next, *entry;
 218         spinlock_t *new_bucket_lock;
 219         unsigned new_hash;
 220
 221         rht_for_each(entry, old_tbl, old_hash) {
 222                 err = 0;
 223                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 224
 225                 if (rht_is_a_nulls(next))
 226                         break;
 227
 228                 pprev = &entry->next;
 229         }
 230
 231         if (err)
 232                 goto out;
 233
 234         new_hash = head_hashfn(ht, new_tbl, entry);
 235
 236         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 237
 238         spin_lock_nested(new_bucket_lock, RHT_LOCK_NESTED);
 239         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 240                                       new_tbl, new_hash);
 241
 242         if (rht_is_a_nulls(head))
 243                 INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash);
 244         else
 245                 RCU_INIT_POINTER(entry->next, head);
 246
 247         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 248         spin_unlock(new_bucket_lock);
 249
 250         rcu_assign_pointer(*pprev, next);
 251
 252 out:
 253         return err;
 254 }
 255
 256 static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash)
 257 {
 258         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 259         spinlock_t *old_bucket_lock;
 260
 261         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 262
 263         spin_lock_bh(old_bucket_lock);
 264         while (!rhashtable_rehash_one(ht, old_hash))
 265                 ;
 266         spin_unlock_bh(old_bucket_lock);
 267 }
 268
 269 static void rhashtable_rehash(struct rhashtable *ht,
 270                               struct bucket_table *new_tbl)
 271 {
 272         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 273         unsigned old_hash;
 274
 275         get_random_bytes(&new_tbl->hash_rnd, sizeof(new_tbl->hash_rnd));
 276
 277         /* Make insertions go into the new, empty table right away. Deletions
 278          * and lookups will be attempted in both tables until we synchronize.
 279          * The synchronize_rcu() guarantees for the new table to be picked up
 280          * so no new additions go into the old table while we relink.
 281          */
 282         rcu_assign_pointer(ht->future_tbl, new_tbl);
 283
 284         for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
 285                 rhashtable_rehash_chain(ht, old_hash);
 286
 287         /* Publish the new table pointer. */
 288         rcu_assign_pointer(ht->tbl, new_tbl);
 289
 290         /* Wait for readers. All new readers will see the new
 291          * table, and thus no references to the old table will
 292          * remain.
 293          */
 294         synchronize_rcu();
 295
 296         bucket_table_free(old_tbl);
 297 }
 298
 299 /**
 300  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 301  * @ht:         the hash table to expand
 302  *
 303  * A secondary bucket array is allocated and the hash entries are migrated.
 304  *
 305  * This function may only be called in a context where it is safe to call
 306  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 307  *
 308  * The caller must ensure that no concurrent resizing occurs by holding
 309  * ht->mutex.
 310  *
 311  * It is valid to have concurrent insertions and deletions protected by per
 312  * bucket locks or concurrent RCU protected lookups and traversals.
 313  */
 314 int rhashtable_expand(struct rhashtable *ht)
 315 {
 316         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 317
 318         ASSERT_RHT_MUTEX(ht);
 319
 320         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 321         if (new_tbl == NULL)
 322                 return -ENOMEM;
 323
 324         new_tbl->hash_rnd = old_tbl->hash_rnd;
 325
 326         atomic_inc(&ht->shift);
 327
 328         rhashtable_rehash(ht, new_tbl);
 329
 330         return 0;
 331 }
 332 EXPORT_SYMBOL_GPL(rhashtable_expand);
 333
 334 /**
 335  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 336  * @ht:         the hash table to shrink
 337  *
 338  * This function may only be called in a context where it is safe to call
 339  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 340  *
 341  * The caller must ensure that no concurrent resizing occurs by holding
 342  * ht->mutex.
 343  *
 344  * The caller must ensure that no concurrent table mutations take place.
 345  * It is however valid to have concurrent lookups if they are RCU protected.
 346  *
 347  * It is valid to have concurrent insertions and deletions protected by per
 348  * bucket locks or concurrent RCU protected lookups and traversals.
 349  */
 350 int rhashtable_shrink(struct rhashtable *ht)
 351 {
 352         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 353
 354         ASSERT_RHT_MUTEX(ht);
 355
 356         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 357         if (new_tbl == NULL)
 358                 return -ENOMEM;
 359
 360         new_tbl->hash_rnd = tbl->hash_rnd;
 361
 362         atomic_dec(&ht->shift);
 363
 364         rhashtable_rehash(ht, new_tbl);
 365
 366         return 0;
 367 }
 368 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 369
 370 static void rht_deferred_worker(struct work_struct *work)
 371 {
 372         struct rhashtable *ht;
 373         struct bucket_table *tbl;
 374         struct rhashtable_walker *walker;
 375
 376         ht = container_of(work, struct rhashtable, run_work);
 377         mutex_lock(&ht->mutex);
 378         if (ht->being_destroyed)
 379                 goto unlock;
 380
 381         tbl = rht_dereference(ht->tbl, ht);
 382
 383         list_for_each_entry(walker, &ht->walkers, list)
 384                 walker->resize = true;
 385
 386         if (rht_grow_above_75(ht, tbl->size))
 387                 rhashtable_expand(ht);
 388         else if (rht_shrink_below_30(ht, tbl->size))
 389                 rhashtable_shrink(ht);
 390 unlock:
 391         mutex_unlock(&ht->mutex);
 392 }
 393
 394 static bool __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 395                                 bool (*compare)(void *, void *), void *arg)
 396 {
 397         struct bucket_table *tbl, *old_tbl;
 398         struct rhash_head *head;
 399         bool no_resize_running;
 400         unsigned hash;
 401         bool success = true;
 402
 403         rcu_read_lock();
 404
 405         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 406         hash = obj_raw_hashfn(ht, old_tbl, rht_obj(ht, obj));
 407
 408         spin_lock_bh(bucket_lock(old_tbl, hash));
 409
 410         /* Because we have already taken the bucket lock in old_tbl,
 411          * if we find that future_tbl is not yet visible then that
 412          * guarantees all other insertions of the same entry will
 413          * also grab the bucket lock in old_tbl because until the
 414          * rehash completes ht->tbl won't be changed.
 415          */
 416         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 417         if (tbl != old_tbl) {
 418                 hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj));
 419                 spin_lock_nested(bucket_lock(tbl, hash), RHT_LOCK_NESTED);
 420         }
 421
 422         if (compare &&
 423             rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 424                                       compare, arg)) {
 425                 success = false;
 426                 goto exit;
 427         }
 428
 429         no_resize_running = tbl == old_tbl;
 430
 431         hash = rht_bucket_index(tbl, hash);
 432         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 433
 434         if (rht_is_a_nulls(head))
 435                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 436         else
 437                 RCU_INIT_POINTER(obj->next, head);
 438
 439         rcu_assign_pointer(tbl->buckets[hash], obj);
 440
 441         atomic_inc(&ht->nelems);
 442         if (no_resize_running && rht_grow_above_75(ht, tbl->size))
 443                 schedule_work(&ht->run_work);
 444
 445 exit:
 446         if (tbl != old_tbl) {
 447                 hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj));
 448                 spin_unlock(bucket_lock(tbl, hash));
 449         }
 450
 451         hash = obj_raw_hashfn(ht, old_tbl, rht_obj(ht, obj));
 452         spin_unlock_bh(bucket_lock(old_tbl, hash));
 453
 454         rcu_read_unlock();
 455
 456         return success;
 457 }
 458
 459 /**
 460  * rhashtable_insert - insert object into hash table
 461  * @ht:         hash table
 462  * @obj:        pointer to hash head inside object
 463  *
 464  * Will take a per bucket spinlock to protect against mutual mutations
 465  * on the same bucket. Multiple insertions may occur in parallel unless
 466  * they map to the same bucket lock.
 467  *
 468  * It is safe to call this function from atomic context.
 469  *
 470  * Will trigger an automatic deferred table resizing if the size grows
 471  * beyond the watermark indicated by grow_decision() which can be passed
 472  * to rhashtable_init().
 473  */
 474 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 475 {
 476         __rhashtable_insert(ht, obj, NULL, NULL);
 477 }
 478 EXPORT_SYMBOL_GPL(rhashtable_insert);
 479
 480 static bool __rhashtable_remove(struct rhashtable *ht,
 481                                 struct bucket_table *tbl,
 482                                 struct rhash_head *obj)
 483 {
 484         struct rhash_head __rcu **pprev;
 485         struct rhash_head *he;
 486         spinlock_t * lock;
 487         unsigned hash;
 488         bool ret = false;
 489
 490         hash = obj_raw_hashfn(ht, tbl, rht_obj(ht, obj));
 491         lock = bucket_lock(tbl, hash);
 492         hash = rht_bucket_index(tbl, hash);
 493
 494         spin_lock_bh(lock);
 495
 496         pprev = &tbl->buckets[hash];
 497         rht_for_each(he, tbl, hash) {
 498                 if (he != obj) {
 499                         pprev = &he->next;
 500                         continue;
 501                 }
 502
 503                 rcu_assign_pointer(*pprev, obj->next);
 504                 ret = true;
 505                 break;
 506         }
 507
 508         spin_unlock_bh(lock);
 509
 510         return ret;
 511 }
 512
 513 /**
 514  * rhashtable_remove - remove object from hash table
 515  * @ht:         hash table
 516  * @obj:        pointer to hash head inside object
 517  *
 518  * Since the hash chain is single linked, the removal operation needs to
 519  * walk the bucket chain upon removal. The removal operation is thus
 520  * considerable slow if the hash table is not correctly sized.
 521  *
 522  * Will automatically shrink the table via rhashtable_expand() if the
 523  * shrink_decision function specified at rhashtable_init() returns true.
 524  *
 525  * The caller must ensure that no concurrent table mutations occur. It is
 526  * however valid to have concurrent lookups if they are RCU protected.
 527  */
 528 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 529 {
 530         struct bucket_table *tbl, *old_tbl;
 531         bool ret;
 532
 533         rcu_read_lock();
 534
 535         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 536         ret = __rhashtable_remove(ht, old_tbl, obj);
 537
 538         /* Because we have already taken (and released) the bucket
 539          * lock in old_tbl, if we find that future_tbl is not yet
 540          * visible then that guarantees the entry to still be in
 541          * old_tbl if it exists.
 542          */
 543         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 544         if (!ret && old_tbl != tbl)
 545                 ret = __rhashtable_remove(ht, tbl, obj);
 546
 547         if (ret) {
 548                 bool no_resize_running = tbl == old_tbl;
 549
 550                 atomic_dec(&ht->nelems);
 551                 if (no_resize_running && rht_shrink_below_30(ht, tbl->size))
 552                         schedule_work(&ht->run_work);
 553         }
 554
 555         rcu_read_unlock();
 556
 557         return ret;
 558 }
 559 EXPORT_SYMBOL_GPL(rhashtable_remove);
 560
 561 struct rhashtable_compare_arg {
 562         struct rhashtable *ht;
 563         const void *key;
 564 };
 565
 566 static bool rhashtable_compare(void *ptr, void *arg)
 567 {
 568         struct rhashtable_compare_arg *x = arg;
 569         struct rhashtable *ht = x->ht;
 570
 571         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 572 }
 573
 574 /**
 575  * rhashtable_lookup - lookup key in hash table
 576  * @ht:         hash table
 577  * @key:        pointer to key
 578  *
 579  * Computes the hash value for the key and traverses the bucket chain looking
 580  * for a entry with an identical key. The first matching entry is returned.
 581  *
 582  * This lookup function may only be used for fixed key hash table (key_len
 583  * parameter set). It will BUG() if used inappropriately.
 584  *
 585  * Lookups may occur in parallel with hashtable mutations and resizing.
 586  */
 587 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 588 {
 589         struct rhashtable_compare_arg arg = {
 590                 .ht = ht,
 591                 .key = key,
 592         };
 593
 594         BUG_ON(!ht->p.key_len);
 595
 596         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 597 }
 598 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 599
 600 /**
 601  * rhashtable_lookup_compare - search hash table with compare function
 602  * @ht:         hash table
 603  * @key:        the pointer to the key
 604  * @compare:    compare function, must return true on match
 605  * @arg:        argument passed on to compare function
 606  *
 607  * Traverses the bucket chain behind the provided hash value and calls the
 608  * specified compare function for each entry.
 609  *
 610  * Lookups may occur in parallel with hashtable mutations and resizing.
 611  *
 612  * Returns the first entry on which the compare function returned true.
 613  */
 614 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 615                                 bool (*compare)(void *, void *), void *arg)
 616 {
 617         const struct bucket_table *tbl, *old_tbl;
 618         struct rhash_head *he;
 619         u32 hash;
 620
 621         rcu_read_lock();
 622
 623         tbl = rht_dereference_rcu(ht->tbl, ht);
 624         hash = key_hashfn(ht, tbl, key, ht->p.key_len);
 625 restart:
 626         rht_for_each_rcu(he, tbl, hash) {
 627                 if (!compare(rht_obj(ht, he), arg))
 628                         continue;
 629                 rcu_read_unlock();
 630                 return rht_obj(ht, he);
 631         }
 632
 633         old_tbl = tbl;
 634         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 635         if (unlikely(tbl != old_tbl))
 636                 goto restart;
 637         rcu_read_unlock();
 638
 639         return NULL;
 640 }
 641 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 642
 643 /**
 644  * rhashtable_lookup_insert - lookup and insert object into hash table
 645  * @ht:         hash table
 646  * @obj:        pointer to hash head inside object
 647  *
 648  * Locks down the bucket chain in both the old and new table if a resize
 649  * is in progress to ensure that writers can't remove from the old table
 650  * and can't insert to the new table during the atomic operation of search
 651  * and insertion. Searches for duplicates in both the old and new table if
 652  * a resize is in progress.
 653  *
 654  * This lookup function may only be used for fixed key hash table (key_len
 655  * parameter set). It will BUG() if used inappropriately.
 656  *
 657  * It is safe to call this function from atomic context.
 658  *
 659  * Will trigger an automatic deferred table resizing if the size grows
 660  * beyond the watermark indicated by grow_decision() which can be passed
 661  * to rhashtable_init().
 662  */
 663 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 664 {
 665         struct rhashtable_compare_arg arg = {
 666                 .ht = ht,
 667                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 668         };
 669
 670         BUG_ON(!ht->p.key_len);
 671
 672         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 673                                                 &arg);
 674 }
 675 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 676
 677 /**
 678  * rhashtable_lookup_compare_insert - search and insert object to hash table
 679  *                                    with compare function
 680  * @ht:         hash table
 681  * @obj:        pointer to hash head inside object
 682  * @compare:    compare function, must return true on match
 683  * @arg:        argument passed on to compare function
 684  *
 685  * Locks down the bucket chain in both the old and new table if a resize
 686  * is in progress to ensure that writers can't remove from the old table
 687  * and can't insert to the new table during the atomic operation of search
 688  * and insertion. Searches for duplicates in both the old and new table if
 689  * a resize is in progress.
 690  *
 691  * Lookups may occur in parallel with hashtable mutations and resizing.
 692  *
 693  * Will trigger an automatic deferred table resizing if the size grows
 694  * beyond the watermark indicated by grow_decision() which can be passed
 695  * to rhashtable_init().
 696  */
 697 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 698                                       struct rhash_head *obj,
 699                                       bool (*compare)(void *, void *),
 700                                       void *arg)
 701 {
 702         BUG_ON(!ht->p.key_len);
 703
 704         return __rhashtable_insert(ht, obj, compare, arg);
 705 }
 706 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 707
 708 /**
 709  * rhashtable_walk_init - Initialise an iterator
 710  * @ht:         Table to walk over
 711  * @iter:       Hash table Iterator
 712  *
 713  * This function prepares a hash table walk.
 714  *
 715  * Note that if you restart a walk after rhashtable_walk_stop you
 716  * may see the same object twice.  Also, you may miss objects if
 717  * there are removals in between rhashtable_walk_stop and the next
 718  * call to rhashtable_walk_start.
 719  *
 720  * For a completely stable walk you should construct your own data
 721  * structure outside the hash table.
 722  *
 723  * This function may sleep so you must not call it from interrupt
 724  * context or with spin locks held.
 725  *
 726  * You must call rhashtable_walk_exit if this function returns
 727  * successfully.
 728  */
 729 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 730 {
 731         iter->ht = ht;
 732         iter->p = NULL;
 733         iter->slot = 0;
 734         iter->skip = 0;
 735
 736         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 737         if (!iter->walker)
 738                 return -ENOMEM;
 739
 740         INIT_LIST_HEAD(&iter->walker->list);
 741         iter->walker->resize = false;
 742
 743         mutex_lock(&ht->mutex);
 744         list_add(&iter->walker->list, &ht->walkers);
 745         mutex_unlock(&ht->mutex);
 746
 747         return 0;
 748 }
 749 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 750
 751 /**
 752  * rhashtable_walk_exit - Free an iterator
 753  * @iter:       Hash table Iterator
 754  *
 755  * This function frees resources allocated by rhashtable_walk_init.
 756  */
 757 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 758 {
 759         mutex_lock(&iter->ht->mutex);
 760         list_del(&iter->walker->list);
 761         mutex_unlock(&iter->ht->mutex);
 762         kfree(iter->walker);
 763 }
 764 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 765
 766 /**
 767  * rhashtable_walk_start - Start a hash table walk
 768  * @iter:       Hash table iterator
 769  *
 770  * Start a hash table walk.  Note that we take the RCU lock in all
 771  * cases including when we return an error.  So you must always call
 772  * rhashtable_walk_stop to clean up.
 773  *
 774  * Returns zero if successful.
 775  *
 776  * Returns -EAGAIN if resize event occured.  Note that the iterator
 777  * will rewind back to the beginning and you may use it immediately
 778  * by calling rhashtable_walk_next.
 779  */
 780 int rhashtable_walk_start(struct rhashtable_iter *iter)
 781 {
 782         rcu_read_lock();
 783
 784         if (iter->walker->resize) {
 785                 iter->slot = 0;
 786                 iter->skip = 0;
 787                 iter->walker->resize = false;
 788                 return -EAGAIN;
 789         }
 790
 791         return 0;
 792 }
 793 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 794
 795 /**
 796  * rhashtable_walk_next - Return the next object and advance the iterator
 797  * @iter:       Hash table iterator
 798  *
 799  * Note that you must call rhashtable_walk_stop when you are finished
 800  * with the walk.
 801  *
 802  * Returns the next object or NULL when the end of the table is reached.
 803  *
 804  * Returns -EAGAIN if resize event occured.  Note that the iterator
 805  * will rewind back to the beginning and you may continue to use it.
 806  */
 807 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 808 {
 809         const struct bucket_table *tbl;
 810         struct rhashtable *ht = iter->ht;
 811         struct rhash_head *p = iter->p;
 812         void *obj = NULL;
 813
 814         tbl = rht_dereference_rcu(ht->tbl, ht);
 815
 816         if (p) {
 817                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 818                 goto next;
 819         }
 820
 821         for (; iter->slot < tbl->size; iter->slot++) {
 822                 int skip = iter->skip;
 823
 824                 rht_for_each_rcu(p, tbl, iter->slot) {
 825                         if (!skip)
 826                                 break;
 827                         skip--;
 828                 }
 829
 830 next:
 831                 if (!rht_is_a_nulls(p)) {
 832                         iter->skip++;
 833                         iter->p = p;
 834                         obj = rht_obj(ht, p);
 835                         goto out;
 836                 }
 837
 838                 iter->skip = 0;
 839         }
 840
 841         iter->p = NULL;
 842
 843 out:
 844         if (iter->walker->resize) {
 845                 iter->p = NULL;
 846                 iter->slot = 0;
 847                 iter->skip = 0;
 848                 iter->walker->resize = false;
 849                 return ERR_PTR(-EAGAIN);
 850         }
 851
 852         return obj;
 853 }
 854 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 855
 856 /**
 857  * rhashtable_walk_stop - Finish a hash table walk
 858  * @iter:       Hash table iterator
 859  *
 860  * Finish a hash table walk.
 861  */
 862 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 863 {
 864         rcu_read_unlock();
 865         iter->p = NULL;
 866 }
 867 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 868
 869 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 870 {
 871         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 872                    1UL << params->min_shift);
 873 }
 874
 875 /**
 876  * rhashtable_init - initialize a new hash table
 877  * @ht:         hash table to be initialized
 878  * @params:     configuration parameters
 879  *
 880  * Initializes a new hash table based on the provided configuration
 881  * parameters. A table can be configured either with a variable or
 882  * fixed length key:
 883  *
 884  * Configuration Example 1: Fixed length keys
 885  * struct test_obj {
 886  *      int                     key;
 887  *      void *                  my_member;
 888  *      struct rhash_head       node;
 889  * };
 890  *
 891  * struct rhashtable_params params = {
 892  *      .head_offset = offsetof(struct test_obj, node),
 893  *      .key_offset = offsetof(struct test_obj, key),
 894  *      .key_len = sizeof(int),
 895  *      .hashfn = jhash,
 896  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 897  * };
 898  *
 899  * Configuration Example 2: Variable length keys
 900  * struct test_obj {
 901  *      [...]
 902  *      struct rhash_head       node;
 903  * };
 904  *
 905  * u32 my_hash_fn(const void *data, u32 seed)
 906  * {
 907  *      struct test_obj *obj = data;
 908  *
 909  *      return [... hash ...];
 910  * }
 911  *
 912  * struct rhashtable_params params = {
 913  *      .head_offset = offsetof(struct test_obj, node),
 914  *      .hashfn = jhash,
 915  *      .obj_hashfn = my_hash_fn,
 916  * };
 917  */
 918 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
 919 {
 920         struct bucket_table *tbl;
 921         size_t size;
 922
 923         size = HASH_DEFAULT_SIZE;
 924
 925         if ((params->key_len && !params->hashfn) ||
 926             (!params->key_len && !params->obj_hashfn))
 927                 return -EINVAL;
 928
 929         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 930                 return -EINVAL;
 931
 932         params->min_shift = max_t(size_t, params->min_shift,
 933                                   ilog2(HASH_MIN_SIZE));
 934
 935         if (params->nelem_hint)
 936                 size = rounded_hashtable_size(params);
 937
 938         memset(ht, 0, sizeof(*ht));
 939         mutex_init(&ht->mutex);
 940         memcpy(&ht->p, params, sizeof(*params));
 941         INIT_LIST_HEAD(&ht->walkers);
 942
 943         if (params->locks_mul)
 944                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 945         else
 946                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 947
 948         tbl = bucket_table_alloc(ht, size);
 949         if (tbl == NULL)
 950                 return -ENOMEM;
 951
 952         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 953
 954         atomic_set(&ht->nelems, 0);
 955         atomic_set(&ht->shift, ilog2(tbl->size));
 956         RCU_INIT_POINTER(ht->tbl, tbl);
 957         RCU_INIT_POINTER(ht->future_tbl, tbl);
 958
 959         INIT_WORK(&ht->run_work, rht_deferred_worker);
 960
 961         return 0;
 962 }
 963 EXPORT_SYMBOL_GPL(rhashtable_init);
 964
 965 /**
 966  * rhashtable_destroy - destroy hash table
 967  * @ht:         the hash table to destroy
 968  *
 969  * Frees the bucket array. This function is not rcu safe, therefore the caller
 970  * has to make sure that no resizing may happen by unpublishing the hashtable
 971  * and waiting for the quiescent cycle before releasing the bucket array.
 972  */
 973 void rhashtable_destroy(struct rhashtable *ht)
 974 {
 975         ht->being_destroyed = true;
 976
 977         cancel_work_sync(&ht->run_work);
 978
 979         mutex_lock(&ht->mutex);
 980         bucket_table_free(rht_dereference(ht->tbl, ht));
 981         mutex_unlock(&ht->mutex);
 982 }
 983 EXPORT_SYMBOL_GPL(rhashtable_destroy);