lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
   5  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
   6  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   7  *
   8  * Code partially derived from nft_hash
   9  * Rewritten with rehash code from br_multicast plus single list
  10  * pointer as suggested by Josh Triplett
  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           4U
  31 #define BUCKET_LOCKS_PER_CPU   128UL
  32
  33 static u32 head_hashfn(struct rhashtable *ht,
  34                        const struct bucket_table *tbl,
  35                        const struct rhash_head *he)
  36 {
  37         return rht_head_hashfn(ht, tbl, he, ht->p);
  38 }
  39
  40 #ifdef CONFIG_PROVE_LOCKING
  41 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  42
  43 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  44 {
  45         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  46 }
  47 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  48
  49 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  50 {
  51         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  52
  53         return (debug_locks) ? lockdep_is_held(lock) : 1;
  54 }
  55 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  56 #else
  57 #define ASSERT_RHT_MUTEX(HT)
  58 #endif
  59
  60
  61 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  62                               gfp_t gfp)
  63 {
  64         unsigned int i, size;
  65 #if defined(CONFIG_PROVE_LOCKING)
  66         unsigned int nr_pcpus = 2;
  67 #else
  68         unsigned int nr_pcpus = num_possible_cpus();
  69 #endif
  70
  71         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
  72         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  73
  74         /* Never allocate more than 0.5 locks per bucket */
  75         size = min_t(unsigned int, size, tbl->size >> 1);
  76
  77         if (sizeof(spinlock_t) != 0) {
  78 #ifdef CONFIG_NUMA
  79                 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
  80                     gfp == GFP_KERNEL)
  81                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
  82                 else
  83 #endif
  84                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  85                                            gfp);
  86                 if (!tbl->locks)
  87                         return -ENOMEM;
  88                 for (i = 0; i < size; i++)
  89                         spin_lock_init(&tbl->locks[i]);
  90         }
  91         tbl->locks_mask = size - 1;
  92
  93         return 0;
  94 }
  95
  96 static void bucket_table_free(const struct bucket_table *tbl)
  97 {
  98         if (tbl)
  99                 kvfree(tbl->locks);
 100
 101         kvfree(tbl);
 102 }
 103
 104 static void bucket_table_free_rcu(struct rcu_head *head)
 105 {
 106         bucket_table_free(container_of(head, struct bucket_table, rcu));
 107 }
 108
 109 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 110                                                size_t nbuckets,
 111                                                gfp_t gfp)
 112 {
 113         struct bucket_table *tbl = NULL;
 114         size_t size;
 115         int i;
 116
 117         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 118         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 119             gfp != GFP_KERNEL)
 120                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 121         if (tbl == NULL && gfp == GFP_KERNEL)
 122                 tbl = vzalloc(size);
 123         if (tbl == NULL)
 124                 return NULL;
 125
 126         tbl->size = nbuckets;
 127
 128         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 129                 bucket_table_free(tbl);
 130                 return NULL;
 131         }
 132
 133         INIT_LIST_HEAD(&tbl->walkers);
 134
 135         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 136
 137         for (i = 0; i < nbuckets; i++)
 138                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 139
 140         return tbl;
 141 }
 142
 143 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 144                                                   struct bucket_table *tbl)
 145 {
 146         struct bucket_table *new_tbl;
 147
 148         do {
 149                 new_tbl = tbl;
 150                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 151         } while (tbl);
 152
 153         return new_tbl;
 154 }
 155
 156 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash)
 157 {
 158         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 159         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 160                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 161         struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
 162         int err = -ENOENT;
 163         struct rhash_head *head, *next, *entry;
 164         spinlock_t *new_bucket_lock;
 165         unsigned new_hash;
 166
 167         rht_for_each(entry, old_tbl, old_hash) {
 168                 err = 0;
 169                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 170
 171                 if (rht_is_a_nulls(next))
 172                         break;
 173
 174                 pprev = &entry->next;
 175         }
 176
 177         if (err)
 178                 goto out;
 179
 180         new_hash = head_hashfn(ht, new_tbl, entry);
 181
 182         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 183
 184         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 185         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 186                                       new_tbl, new_hash);
 187
 188         if (rht_is_a_nulls(head))
 189                 INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash);
 190         else
 191                 RCU_INIT_POINTER(entry->next, head);
 192
 193         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 194         spin_unlock(new_bucket_lock);
 195
 196         rcu_assign_pointer(*pprev, next);
 197
 198 out:
 199         return err;
 200 }
 201
 202 static void rhashtable_rehash_chain(struct rhashtable *ht, unsigned old_hash)
 203 {
 204         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 205         spinlock_t *old_bucket_lock;
 206
 207         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 208
 209         spin_lock_bh(old_bucket_lock);
 210         while (!rhashtable_rehash_one(ht, old_hash))
 211                 ;
 212         old_tbl->rehash++;
 213         spin_unlock_bh(old_bucket_lock);
 214 }
 215
 216 static int rhashtable_rehash_attach(struct rhashtable *ht,
 217                                     struct bucket_table *old_tbl,
 218                                     struct bucket_table *new_tbl)
 219 {
 220         /* Protect future_tbl using the first bucket lock. */
 221         spin_lock_bh(old_tbl->locks);
 222
 223         /* Did somebody beat us to it? */
 224         if (rcu_access_pointer(old_tbl->future_tbl)) {
 225                 spin_unlock_bh(old_tbl->locks);
 226                 return -EEXIST;
 227         }
 228
 229         /* Make insertions go into the new, empty table right away. Deletions
 230          * and lookups will be attempted in both tables until we synchronize.
 231          */
 232         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 233
 234         /* Ensure the new table is visible to readers. */
 235         smp_wmb();
 236
 237         spin_unlock_bh(old_tbl->locks);
 238
 239         return 0;
 240 }
 241
 242 static int rhashtable_rehash_table(struct rhashtable *ht)
 243 {
 244         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 245         struct bucket_table *new_tbl;
 246         struct rhashtable_walker *walker;
 247         unsigned old_hash;
 248
 249         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 250         if (!new_tbl)
 251                 return 0;
 252
 253         for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
 254                 rhashtable_rehash_chain(ht, old_hash);
 255
 256         /* Publish the new table pointer. */
 257         rcu_assign_pointer(ht->tbl, new_tbl);
 258
 259         list_for_each_entry(walker, &old_tbl->walkers, list)
 260                 walker->tbl = NULL;
 261
 262         /* Wait for readers. All new readers will see the new
 263          * table, and thus no references to the old table will
 264          * remain.
 265          */
 266         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 267
 268         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 269 }
 270
 271 /**
 272  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 273  * @ht:         the hash table to expand
 274  *
 275  * A secondary bucket array is allocated and the hash entries are migrated.
 276  *
 277  * This function may only be called in a context where it is safe to call
 278  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 279  *
 280  * The caller must ensure that no concurrent resizing occurs by holding
 281  * ht->mutex.
 282  *
 283  * It is valid to have concurrent insertions and deletions protected by per
 284  * bucket locks or concurrent RCU protected lookups and traversals.
 285  */
 286 static int rhashtable_expand(struct rhashtable *ht)
 287 {
 288         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 289         int err;
 290
 291         ASSERT_RHT_MUTEX(ht);
 292
 293         old_tbl = rhashtable_last_table(ht, old_tbl);
 294
 295         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 296         if (new_tbl == NULL)
 297                 return -ENOMEM;
 298
 299         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 300         if (err)
 301                 bucket_table_free(new_tbl);
 302
 303         return err;
 304 }
 305
 306 /**
 307  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 308  * @ht:         the hash table to shrink
 309  *
 310  * This function shrinks the hash table to fit, i.e., the smallest
 311  * size would not cause it to expand right away automatically.
 312  *
 313  * The caller must ensure that no concurrent resizing occurs by holding
 314  * ht->mutex.
 315  *
 316  * The caller must ensure that no concurrent table mutations take place.
 317  * It is however valid to have concurrent lookups if they are RCU protected.
 318  *
 319  * It is valid to have concurrent insertions and deletions protected by per
 320  * bucket locks or concurrent RCU protected lookups and traversals.
 321  */
 322 static int rhashtable_shrink(struct rhashtable *ht)
 323 {
 324         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 325         unsigned size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
 326         int err;
 327
 328         ASSERT_RHT_MUTEX(ht);
 329
 330         if (size < ht->p.min_size)
 331                 size = ht->p.min_size;
 332
 333         if (old_tbl->size <= size)
 334                 return 0;
 335
 336         if (rht_dereference(old_tbl->future_tbl, ht))
 337                 return -EEXIST;
 338
 339         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 340         if (new_tbl == NULL)
 341                 return -ENOMEM;
 342
 343         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 344         if (err)
 345                 bucket_table_free(new_tbl);
 346
 347         return err;
 348 }
 349
 350 static void rht_deferred_worker(struct work_struct *work)
 351 {
 352         struct rhashtable *ht;
 353         struct bucket_table *tbl;
 354         int err = 0;
 355
 356         ht = container_of(work, struct rhashtable, run_work);
 357         mutex_lock(&ht->mutex);
 358         if (ht->being_destroyed)
 359                 goto unlock;
 360
 361         tbl = rht_dereference(ht->tbl, ht);
 362         tbl = rhashtable_last_table(ht, tbl);
 363
 364         if (rht_grow_above_75(ht, tbl))
 365                 rhashtable_expand(ht);
 366         else if (rht_shrink_below_30(ht, tbl))
 367                 rhashtable_shrink(ht);
 368
 369         err = rhashtable_rehash_table(ht);
 370
 371 unlock:
 372         mutex_unlock(&ht->mutex);
 373
 374         if (err)
 375                 schedule_work(&ht->run_work);
 376 }
 377
 378 static bool rhashtable_check_elasticity(struct rhashtable *ht,
 379                                         struct bucket_table *tbl,
 380                                         unsigned hash)
 381 {
 382         unsigned elasticity = ht->elasticity;
 383         struct rhash_head *head;
 384
 385         rht_for_each(head, tbl, hash)
 386                 if (!--elasticity)
 387                         return true;
 388
 389         return false;
 390 }
 391
 392 int rhashtable_insert_rehash(struct rhashtable *ht)
 393 {
 394         struct bucket_table *old_tbl;
 395         struct bucket_table *new_tbl;
 396         struct bucket_table *tbl;
 397         unsigned int size;
 398         int err;
 399
 400         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 401         tbl = rhashtable_last_table(ht, old_tbl);
 402
 403         size = tbl->size;
 404
 405         if (rht_grow_above_75(ht, tbl))
 406                 size *= 2;
 407         /* More than two rehashes (not resizes) detected. */
 408         else if (WARN_ON(old_tbl != tbl && old_tbl->size == size))
 409                 return -EBUSY;
 410
 411         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 412         if (new_tbl == NULL)
 413                 return -ENOMEM;
 414
 415         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 416         if (err) {
 417                 bucket_table_free(new_tbl);
 418                 if (err == -EEXIST)
 419                         err = 0;
 420         } else
 421                 schedule_work(&ht->run_work);
 422
 423         return err;
 424 }
 425 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
 426
 427 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 428                            struct rhash_head *obj,
 429                            struct bucket_table *tbl)
 430 {
 431         struct rhash_head *head;
 432         unsigned hash;
 433         int err;
 434
 435         tbl = rhashtable_last_table(ht, tbl);
 436         hash = head_hashfn(ht, tbl, obj);
 437         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 438
 439         err = -EEXIST;
 440         if (key && rhashtable_lookup_fast(ht, key, ht->p))
 441                 goto exit;
 442
 443         err = -EAGAIN;
 444         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 445             rht_grow_above_100(ht, tbl))
 446                 goto exit;
 447
 448         err = 0;
 449
 450         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 451
 452         RCU_INIT_POINTER(obj->next, head);
 453
 454         rcu_assign_pointer(tbl->buckets[hash], obj);
 455
 456         atomic_inc(&ht->nelems);
 457
 458 exit:
 459         spin_unlock(rht_bucket_lock(tbl, hash));
 460
 461         return err;
 462 }
 463 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 464
 465 /**
 466  * rhashtable_walk_init - Initialise an iterator
 467  * @ht:         Table to walk over
 468  * @iter:       Hash table Iterator
 469  *
 470  * This function prepares a hash table walk.
 471  *
 472  * Note that if you restart a walk after rhashtable_walk_stop you
 473  * may see the same object twice.  Also, you may miss objects if
 474  * there are removals in between rhashtable_walk_stop and the next
 475  * call to rhashtable_walk_start.
 476  *
 477  * For a completely stable walk you should construct your own data
 478  * structure outside the hash table.
 479  *
 480  * This function may sleep so you must not call it from interrupt
 481  * context or with spin locks held.
 482  *
 483  * You must call rhashtable_walk_exit if this function returns
 484  * successfully.
 485  */
 486 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 487 {
 488         iter->ht = ht;
 489         iter->p = NULL;
 490         iter->slot = 0;
 491         iter->skip = 0;
 492
 493         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 494         if (!iter->walker)
 495                 return -ENOMEM;
 496
 497         mutex_lock(&ht->mutex);
 498         iter->walker->tbl = rht_dereference(ht->tbl, ht);
 499         list_add(&iter->walker->list, &iter->walker->tbl->walkers);
 500         mutex_unlock(&ht->mutex);
 501
 502         return 0;
 503 }
 504 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 505
 506 /**
 507  * rhashtable_walk_exit - Free an iterator
 508  * @iter:       Hash table Iterator
 509  *
 510  * This function frees resources allocated by rhashtable_walk_init.
 511  */
 512 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 513 {
 514         mutex_lock(&iter->ht->mutex);
 515         if (iter->walker->tbl)
 516                 list_del(&iter->walker->list);
 517         mutex_unlock(&iter->ht->mutex);
 518         kfree(iter->walker);
 519 }
 520 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 521
 522 /**
 523  * rhashtable_walk_start - Start a hash table walk
 524  * @iter:       Hash table iterator
 525  *
 526  * Start a hash table walk.  Note that we take the RCU lock in all
 527  * cases including when we return an error.  So you must always call
 528  * rhashtable_walk_stop to clean up.
 529  *
 530  * Returns zero if successful.
 531  *
 532  * Returns -EAGAIN if resize event occured.  Note that the iterator
 533  * will rewind back to the beginning and you may use it immediately
 534  * by calling rhashtable_walk_next.
 535  */
 536 int rhashtable_walk_start(struct rhashtable_iter *iter)
 537         __acquires(RCU)
 538 {
 539         struct rhashtable *ht = iter->ht;
 540
 541         mutex_lock(&ht->mutex);
 542
 543         if (iter->walker->tbl)
 544                 list_del(&iter->walker->list);
 545
 546         rcu_read_lock();
 547
 548         mutex_unlock(&ht->mutex);
 549
 550         if (!iter->walker->tbl) {
 551                 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
 552                 return -EAGAIN;
 553         }
 554
 555         return 0;
 556 }
 557 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 558
 559 /**
 560  * rhashtable_walk_next - Return the next object and advance the iterator
 561  * @iter:       Hash table iterator
 562  *
 563  * Note that you must call rhashtable_walk_stop when you are finished
 564  * with the walk.
 565  *
 566  * Returns the next object or NULL when the end of the table is reached.
 567  *
 568  * Returns -EAGAIN if resize event occured.  Note that the iterator
 569  * will rewind back to the beginning and you may continue to use it.
 570  */
 571 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 572 {
 573         struct bucket_table *tbl = iter->walker->tbl;
 574         struct rhashtable *ht = iter->ht;
 575         struct rhash_head *p = iter->p;
 576         void *obj = NULL;
 577
 578         if (p) {
 579                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 580                 goto next;
 581         }
 582
 583         for (; iter->slot < tbl->size; iter->slot++) {
 584                 int skip = iter->skip;
 585
 586                 rht_for_each_rcu(p, tbl, iter->slot) {
 587                         if (!skip)
 588                                 break;
 589                         skip--;
 590                 }
 591
 592 next:
 593                 if (!rht_is_a_nulls(p)) {
 594                         iter->skip++;
 595                         iter->p = p;
 596                         obj = rht_obj(ht, p);
 597                         goto out;
 598                 }
 599
 600                 iter->skip = 0;
 601         }
 602
 603         /* Ensure we see any new tables. */
 604         smp_rmb();
 605
 606         iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 607         if (iter->walker->tbl) {
 608                 iter->slot = 0;
 609                 iter->skip = 0;
 610                 return ERR_PTR(-EAGAIN);
 611         }
 612
 613         iter->p = NULL;
 614
 615 out:
 616
 617         return obj;
 618 }
 619 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 620
 621 /**
 622  * rhashtable_walk_stop - Finish a hash table walk
 623  * @iter:       Hash table iterator
 624  *
 625  * Finish a hash table walk.
 626  */
 627 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 628         __releases(RCU)
 629 {
 630         struct rhashtable *ht;
 631         struct bucket_table *tbl = iter->walker->tbl;
 632
 633         if (!tbl)
 634                 goto out;
 635
 636         ht = iter->ht;
 637
 638         mutex_lock(&ht->mutex);
 639         if (tbl->rehash < tbl->size)
 640                 list_add(&iter->walker->list, &tbl->walkers);
 641         else
 642                 iter->walker->tbl = NULL;
 643         mutex_unlock(&ht->mutex);
 644
 645         iter->p = NULL;
 646
 647 out:
 648         rcu_read_unlock();
 649 }
 650 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 651
 652 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 653 {
 654         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 655                    (unsigned long)params->min_size);
 656 }
 657
 658 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 659 {
 660         return jhash2(key, length, seed);
 661 }
 662
 663 /**
 664  * rhashtable_init - initialize a new hash table
 665  * @ht:         hash table to be initialized
 666  * @params:     configuration parameters
 667  *
 668  * Initializes a new hash table based on the provided configuration
 669  * parameters. A table can be configured either with a variable or
 670  * fixed length key:
 671  *
 672  * Configuration Example 1: Fixed length keys
 673  * struct test_obj {
 674  *      int                     key;
 675  *      void *                  my_member;
 676  *      struct rhash_head       node;
 677  * };
 678  *
 679  * struct rhashtable_params params = {
 680  *      .head_offset = offsetof(struct test_obj, node),
 681  *      .key_offset = offsetof(struct test_obj, key),
 682  *      .key_len = sizeof(int),
 683  *      .hashfn = jhash,
 684  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 685  * };
 686  *
 687  * Configuration Example 2: Variable length keys
 688  * struct test_obj {
 689  *      [...]
 690  *      struct rhash_head       node;
 691  * };
 692  *
 693  * u32 my_hash_fn(const void *data, u32 seed)
 694  * {
 695  *      struct test_obj *obj = data;
 696  *
 697  *      return [... hash ...];
 698  * }
 699  *
 700  * struct rhashtable_params params = {
 701  *      .head_offset = offsetof(struct test_obj, node),
 702  *      .hashfn = jhash,
 703  *      .obj_hashfn = my_hash_fn,
 704  * };
 705  */
 706 int rhashtable_init(struct rhashtable *ht,
 707                     const struct rhashtable_params *params)
 708 {
 709         struct bucket_table *tbl;
 710         size_t size;
 711
 712         size = HASH_DEFAULT_SIZE;
 713
 714         if ((!params->key_len && !params->obj_hashfn) ||
 715             (params->obj_hashfn && !params->obj_cmpfn))
 716                 return -EINVAL;
 717
 718         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 719                 return -EINVAL;
 720
 721         if (params->nelem_hint)
 722                 size = rounded_hashtable_size(params);
 723
 724         memset(ht, 0, sizeof(*ht));
 725         mutex_init(&ht->mutex);
 726         memcpy(&ht->p, params, sizeof(*params));
 727
 728         if (params->min_size)
 729                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 730
 731         if (params->max_size)
 732                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 733
 734         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 735
 736         if (!params->insecure_elasticity)
 737                 ht->elasticity = 16;
 738
 739         if (params->locks_mul)
 740                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 741         else
 742                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 743
 744         ht->key_len = ht->p.key_len;
 745         if (!params->hashfn) {
 746                 ht->p.hashfn = jhash;
 747
 748                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 749                         ht->key_len /= sizeof(u32);
 750                         ht->p.hashfn = rhashtable_jhash2;
 751                 }
 752         }
 753
 754         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 755         if (tbl == NULL)
 756                 return -ENOMEM;
 757
 758         atomic_set(&ht->nelems, 0);
 759
 760         RCU_INIT_POINTER(ht->tbl, tbl);
 761
 762         INIT_WORK(&ht->run_work, rht_deferred_worker);
 763
 764         return 0;
 765 }
 766 EXPORT_SYMBOL_GPL(rhashtable_init);
 767
 768 /**
 769  * rhashtable_destroy - destroy hash table
 770  * @ht:         the hash table to destroy
 771  *
 772  * Frees the bucket array. This function is not rcu safe, therefore the caller
 773  * has to make sure that no resizing may happen by unpublishing the hashtable
 774  * and waiting for the quiescent cycle before releasing the bucket array.
 775  */
 776 void rhashtable_destroy(struct rhashtable *ht)
 777 {
 778         ht->being_destroyed = true;
 779
 780         cancel_work_sync(&ht->run_work);
 781
 782         mutex_lock(&ht->mutex);
 783         bucket_table_free(rht_dereference(ht->tbl, ht));
 784         mutex_unlock(&ht->mutex);
 785 }
 786 EXPORT_SYMBOL_GPL(rhashtable_destroy);