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         spin_lock(&ht->lock);
 260         list_for_each_entry(walker, &old_tbl->walkers, list)
 261                 walker->tbl = NULL;
 262         spin_unlock(&ht->lock);
 263
 264         /* Wait for readers. All new readers will see the new
 265          * table, and thus no references to the old table will
 266          * remain.
 267          */
 268         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 269
 270         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 271 }
 272
 273 /**
 274  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 275  * @ht:         the hash table to expand
 276  *
 277  * A secondary bucket array is allocated and the hash entries are migrated.
 278  *
 279  * This function may only be called in a context where it is safe to call
 280  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 281  *
 282  * The caller must ensure that no concurrent resizing occurs by holding
 283  * ht->mutex.
 284  *
 285  * It is valid to have concurrent insertions and deletions protected by per
 286  * bucket locks or concurrent RCU protected lookups and traversals.
 287  */
 288 static int rhashtable_expand(struct rhashtable *ht)
 289 {
 290         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 291         int err;
 292
 293         ASSERT_RHT_MUTEX(ht);
 294
 295         old_tbl = rhashtable_last_table(ht, old_tbl);
 296
 297         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 298         if (new_tbl == NULL)
 299                 return -ENOMEM;
 300
 301         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 302         if (err)
 303                 bucket_table_free(new_tbl);
 304
 305         return err;
 306 }
 307
 308 /**
 309  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 310  * @ht:         the hash table to shrink
 311  *
 312  * This function shrinks the hash table to fit, i.e., the smallest
 313  * size would not cause it to expand right away automatically.
 314  *
 315  * The caller must ensure that no concurrent resizing occurs by holding
 316  * ht->mutex.
 317  *
 318  * The caller must ensure that no concurrent table mutations take place.
 319  * It is however valid to have concurrent lookups if they are RCU protected.
 320  *
 321  * It is valid to have concurrent insertions and deletions protected by per
 322  * bucket locks or concurrent RCU protected lookups and traversals.
 323  */
 324 static int rhashtable_shrink(struct rhashtable *ht)
 325 {
 326         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 327         unsigned size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
 328         int err;
 329
 330         ASSERT_RHT_MUTEX(ht);
 331
 332         if (size < ht->p.min_size)
 333                 size = ht->p.min_size;
 334
 335         if (old_tbl->size <= size)
 336                 return 0;
 337
 338         if (rht_dereference(old_tbl->future_tbl, ht))
 339                 return -EEXIST;
 340
 341         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 342         if (new_tbl == NULL)
 343                 return -ENOMEM;
 344
 345         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 346         if (err)
 347                 bucket_table_free(new_tbl);
 348
 349         return err;
 350 }
 351
 352 static void rht_deferred_worker(struct work_struct *work)
 353 {
 354         struct rhashtable *ht;
 355         struct bucket_table *tbl;
 356         int err = 0;
 357
 358         ht = container_of(work, struct rhashtable, run_work);
 359         mutex_lock(&ht->mutex);
 360         if (ht->being_destroyed)
 361                 goto unlock;
 362
 363         tbl = rht_dereference(ht->tbl, ht);
 364         tbl = rhashtable_last_table(ht, tbl);
 365
 366         if (rht_grow_above_75(ht, tbl))
 367                 rhashtable_expand(ht);
 368         else if (rht_shrink_below_30(ht, tbl))
 369                 rhashtable_shrink(ht);
 370
 371         err = rhashtable_rehash_table(ht);
 372
 373 unlock:
 374         mutex_unlock(&ht->mutex);
 375
 376         if (err)
 377                 schedule_work(&ht->run_work);
 378 }
 379
 380 static bool rhashtable_check_elasticity(struct rhashtable *ht,
 381                                         struct bucket_table *tbl,
 382                                         unsigned hash)
 383 {
 384         unsigned elasticity = ht->elasticity;
 385         struct rhash_head *head;
 386
 387         rht_for_each(head, tbl, hash)
 388                 if (!--elasticity)
 389                         return true;
 390
 391         return false;
 392 }
 393
 394 int rhashtable_insert_rehash(struct rhashtable *ht)
 395 {
 396         struct bucket_table *old_tbl;
 397         struct bucket_table *new_tbl;
 398         struct bucket_table *tbl;
 399         unsigned int size;
 400         int err;
 401
 402         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 403         tbl = rhashtable_last_table(ht, old_tbl);
 404
 405         size = tbl->size;
 406
 407         if (rht_grow_above_75(ht, tbl))
 408                 size *= 2;
 409         /* More than two rehashes (not resizes) detected. */
 410         else if (WARN_ON(old_tbl != tbl && old_tbl->size == size))
 411                 return -EBUSY;
 412
 413         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 414         if (new_tbl == NULL)
 415                 return -ENOMEM;
 416
 417         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 418         if (err) {
 419                 bucket_table_free(new_tbl);
 420                 if (err == -EEXIST)
 421                         err = 0;
 422         } else
 423                 schedule_work(&ht->run_work);
 424
 425         return err;
 426 }
 427 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
 428
 429 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 430                            struct rhash_head *obj,
 431                            struct bucket_table *tbl)
 432 {
 433         struct rhash_head *head;
 434         unsigned hash;
 435         int err;
 436
 437         tbl = rhashtable_last_table(ht, tbl);
 438         hash = head_hashfn(ht, tbl, obj);
 439         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 440
 441         err = -EEXIST;
 442         if (key && rhashtable_lookup_fast(ht, key, ht->p))
 443                 goto exit;
 444
 445         err = -EAGAIN;
 446         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 447             rht_grow_above_100(ht, tbl))
 448                 goto exit;
 449
 450         err = 0;
 451
 452         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 453
 454         RCU_INIT_POINTER(obj->next, head);
 455
 456         rcu_assign_pointer(tbl->buckets[hash], obj);
 457
 458         atomic_inc(&ht->nelems);
 459
 460 exit:
 461         spin_unlock(rht_bucket_lock(tbl, hash));
 462
 463         return err;
 464 }
 465 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 466
 467 /**
 468  * rhashtable_walk_init - Initialise an iterator
 469  * @ht:         Table to walk over
 470  * @iter:       Hash table Iterator
 471  *
 472  * This function prepares a hash table walk.
 473  *
 474  * Note that if you restart a walk after rhashtable_walk_stop you
 475  * may see the same object twice.  Also, you may miss objects if
 476  * there are removals in between rhashtable_walk_stop and the next
 477  * call to rhashtable_walk_start.
 478  *
 479  * For a completely stable walk you should construct your own data
 480  * structure outside the hash table.
 481  *
 482  * This function may sleep so you must not call it from interrupt
 483  * context or with spin locks held.
 484  *
 485  * You must call rhashtable_walk_exit if this function returns
 486  * successfully.
 487  */
 488 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 489 {
 490         iter->ht = ht;
 491         iter->p = NULL;
 492         iter->slot = 0;
 493         iter->skip = 0;
 494
 495         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 496         if (!iter->walker)
 497                 return -ENOMEM;
 498
 499         mutex_lock(&ht->mutex);
 500         iter->walker->tbl = rht_dereference(ht->tbl, ht);
 501         list_add(&iter->walker->list, &iter->walker->tbl->walkers);
 502         mutex_unlock(&ht->mutex);
 503
 504         return 0;
 505 }
 506 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 507
 508 /**
 509  * rhashtable_walk_exit - Free an iterator
 510  * @iter:       Hash table Iterator
 511  *
 512  * This function frees resources allocated by rhashtable_walk_init.
 513  */
 514 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 515 {
 516         mutex_lock(&iter->ht->mutex);
 517         if (iter->walker->tbl)
 518                 list_del(&iter->walker->list);
 519         mutex_unlock(&iter->ht->mutex);
 520         kfree(iter->walker);
 521 }
 522 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 523
 524 /**
 525  * rhashtable_walk_start - Start a hash table walk
 526  * @iter:       Hash table iterator
 527  *
 528  * Start a hash table walk.  Note that we take the RCU lock in all
 529  * cases including when we return an error.  So you must always call
 530  * rhashtable_walk_stop to clean up.
 531  *
 532  * Returns zero if successful.
 533  *
 534  * Returns -EAGAIN if resize event occured.  Note that the iterator
 535  * will rewind back to the beginning and you may use it immediately
 536  * by calling rhashtable_walk_next.
 537  */
 538 int rhashtable_walk_start(struct rhashtable_iter *iter)
 539         __acquires(RCU)
 540 {
 541         struct rhashtable *ht = iter->ht;
 542
 543         mutex_lock(&ht->mutex);
 544
 545         if (iter->walker->tbl)
 546                 list_del(&iter->walker->list);
 547
 548         rcu_read_lock();
 549
 550         mutex_unlock(&ht->mutex);
 551
 552         if (!iter->walker->tbl) {
 553                 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
 554                 return -EAGAIN;
 555         }
 556
 557         return 0;
 558 }
 559 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 560
 561 /**
 562  * rhashtable_walk_next - Return the next object and advance the iterator
 563  * @iter:       Hash table iterator
 564  *
 565  * Note that you must call rhashtable_walk_stop when you are finished
 566  * with the walk.
 567  *
 568  * Returns the next object or NULL when the end of the table is reached.
 569  *
 570  * Returns -EAGAIN if resize event occured.  Note that the iterator
 571  * will rewind back to the beginning and you may continue to use it.
 572  */
 573 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 574 {
 575         struct bucket_table *tbl = iter->walker->tbl;
 576         struct rhashtable *ht = iter->ht;
 577         struct rhash_head *p = iter->p;
 578         void *obj = NULL;
 579
 580         if (p) {
 581                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 582                 goto next;
 583         }
 584
 585         for (; iter->slot < tbl->size; iter->slot++) {
 586                 int skip = iter->skip;
 587
 588                 rht_for_each_rcu(p, tbl, iter->slot) {
 589                         if (!skip)
 590                                 break;
 591                         skip--;
 592                 }
 593
 594 next:
 595                 if (!rht_is_a_nulls(p)) {
 596                         iter->skip++;
 597                         iter->p = p;
 598                         obj = rht_obj(ht, p);
 599                         goto out;
 600                 }
 601
 602                 iter->skip = 0;
 603         }
 604
 605         /* Ensure we see any new tables. */
 606         smp_rmb();
 607
 608         iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 609         if (iter->walker->tbl) {
 610                 iter->slot = 0;
 611                 iter->skip = 0;
 612                 return ERR_PTR(-EAGAIN);
 613         }
 614
 615         iter->p = NULL;
 616
 617 out:
 618
 619         return obj;
 620 }
 621 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 622
 623 /**
 624  * rhashtable_walk_stop - Finish a hash table walk
 625  * @iter:       Hash table iterator
 626  *
 627  * Finish a hash table walk.
 628  */
 629 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 630         __releases(RCU)
 631 {
 632         struct rhashtable *ht;
 633         struct bucket_table *tbl = iter->walker->tbl;
 634
 635         if (!tbl)
 636                 goto out;
 637
 638         ht = iter->ht;
 639
 640         spin_lock(&ht->lock);
 641         if (tbl->rehash < tbl->size)
 642                 list_add(&iter->walker->list, &tbl->walkers);
 643         else
 644                 iter->walker->tbl = NULL;
 645         spin_unlock(&ht->lock);
 646
 647         iter->p = NULL;
 648
 649 out:
 650         rcu_read_unlock();
 651 }
 652 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 653
 654 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 655 {
 656         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 657                    (unsigned long)params->min_size);
 658 }
 659
 660 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 661 {
 662         return jhash2(key, length, seed);
 663 }
 664
 665 /**
 666  * rhashtable_init - initialize a new hash table
 667  * @ht:         hash table to be initialized
 668  * @params:     configuration parameters
 669  *
 670  * Initializes a new hash table based on the provided configuration
 671  * parameters. A table can be configured either with a variable or
 672  * fixed length key:
 673  *
 674  * Configuration Example 1: Fixed length keys
 675  * struct test_obj {
 676  *      int                     key;
 677  *      void *                  my_member;
 678  *      struct rhash_head       node;
 679  * };
 680  *
 681  * struct rhashtable_params params = {
 682  *      .head_offset = offsetof(struct test_obj, node),
 683  *      .key_offset = offsetof(struct test_obj, key),
 684  *      .key_len = sizeof(int),
 685  *      .hashfn = jhash,
 686  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 687  * };
 688  *
 689  * Configuration Example 2: Variable length keys
 690  * struct test_obj {
 691  *      [...]
 692  *      struct rhash_head       node;
 693  * };
 694  *
 695  * u32 my_hash_fn(const void *data, u32 seed)
 696  * {
 697  *      struct test_obj *obj = data;
 698  *
 699  *      return [... hash ...];
 700  * }
 701  *
 702  * struct rhashtable_params params = {
 703  *      .head_offset = offsetof(struct test_obj, node),
 704  *      .hashfn = jhash,
 705  *      .obj_hashfn = my_hash_fn,
 706  * };
 707  */
 708 int rhashtable_init(struct rhashtable *ht,
 709                     const struct rhashtable_params *params)
 710 {
 711         struct bucket_table *tbl;
 712         size_t size;
 713
 714         size = HASH_DEFAULT_SIZE;
 715
 716         if ((!params->key_len && !params->obj_hashfn) ||
 717             (params->obj_hashfn && !params->obj_cmpfn))
 718                 return -EINVAL;
 719
 720         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 721                 return -EINVAL;
 722
 723         if (params->nelem_hint)
 724                 size = rounded_hashtable_size(params);
 725
 726         memset(ht, 0, sizeof(*ht));
 727         mutex_init(&ht->mutex);
 728         spin_lock_init(&ht->lock);
 729         memcpy(&ht->p, params, sizeof(*params));
 730
 731         if (params->min_size)
 732                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 733
 734         if (params->max_size)
 735                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 736
 737         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 738
 739         if (!params->insecure_elasticity)
 740                 ht->elasticity = 16;
 741
 742         if (params->locks_mul)
 743                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 744         else
 745                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 746
 747         ht->key_len = ht->p.key_len;
 748         if (!params->hashfn) {
 749                 ht->p.hashfn = jhash;
 750
 751                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 752                         ht->key_len /= sizeof(u32);
 753                         ht->p.hashfn = rhashtable_jhash2;
 754                 }
 755         }
 756
 757         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 758         if (tbl == NULL)
 759                 return -ENOMEM;
 760
 761         atomic_set(&ht->nelems, 0);
 762
 763         RCU_INIT_POINTER(ht->tbl, tbl);
 764
 765         INIT_WORK(&ht->run_work, rht_deferred_worker);
 766
 767         return 0;
 768 }
 769 EXPORT_SYMBOL_GPL(rhashtable_init);
 770
 771 /**
 772  * rhashtable_destroy - destroy hash table
 773  * @ht:         the hash table to destroy
 774  *
 775  * Frees the bucket array. This function is not rcu safe, therefore the caller
 776  * has to make sure that no resizing may happen by unpublishing the hashtable
 777  * and waiting for the quiescent cycle before releasing the bucket array.
 778  */
 779 void rhashtable_destroy(struct rhashtable *ht)
 780 {
 781         ht->being_destroyed = true;
 782
 783         cancel_work_sync(&ht->run_work);
 784
 785         mutex_lock(&ht->mutex);
 786         bucket_table_free(rht_dereference(ht->tbl, ht));
 787         mutex_unlock(&ht->mutex);
 788 }
 789 EXPORT_SYMBOL_GPL(rhashtable_destroy);