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 int 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 int 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,
 203                                     unsigned int old_hash)
 204 {
 205         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 206         spinlock_t *old_bucket_lock;
 207
 208         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 209
 210         spin_lock_bh(old_bucket_lock);
 211         while (!rhashtable_rehash_one(ht, old_hash))
 212                 ;
 213         old_tbl->rehash++;
 214         spin_unlock_bh(old_bucket_lock);
 215 }
 216
 217 static int rhashtable_rehash_attach(struct rhashtable *ht,
 218                                     struct bucket_table *old_tbl,
 219                                     struct bucket_table *new_tbl)
 220 {
 221         /* Protect future_tbl using the first bucket lock. */
 222         spin_lock_bh(old_tbl->locks);
 223
 224         /* Did somebody beat us to it? */
 225         if (rcu_access_pointer(old_tbl->future_tbl)) {
 226                 spin_unlock_bh(old_tbl->locks);
 227                 return -EEXIST;
 228         }
 229
 230         /* Make insertions go into the new, empty table right away. Deletions
 231          * and lookups will be attempted in both tables until we synchronize.
 232          */
 233         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 234
 235         /* Ensure the new table is visible to readers. */
 236         smp_wmb();
 237
 238         spin_unlock_bh(old_tbl->locks);
 239
 240         return 0;
 241 }
 242
 243 static int rhashtable_rehash_table(struct rhashtable *ht)
 244 {
 245         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 246         struct bucket_table *new_tbl;
 247         struct rhashtable_walker *walker;
 248         unsigned int old_hash;
 249
 250         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 251         if (!new_tbl)
 252                 return 0;
 253
 254         for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
 255                 rhashtable_rehash_chain(ht, old_hash);
 256
 257         /* Publish the new table pointer. */
 258         rcu_assign_pointer(ht->tbl, new_tbl);
 259
 260         spin_lock(&ht->lock);
 261         list_for_each_entry(walker, &old_tbl->walkers, list)
 262                 walker->tbl = NULL;
 263         spin_unlock(&ht->lock);
 264
 265         /* Wait for readers. All new readers will see the new
 266          * table, and thus no references to the old table will
 267          * remain.
 268          */
 269         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 270
 271         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 272 }
 273
 274 /**
 275  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 276  * @ht:         the hash table to expand
 277  *
 278  * A secondary bucket array is allocated and the hash entries are migrated.
 279  *
 280  * This function may only be called in a context where it is safe to call
 281  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 282  *
 283  * The caller must ensure that no concurrent resizing occurs by holding
 284  * ht->mutex.
 285  *
 286  * It is valid to have concurrent insertions and deletions protected by per
 287  * bucket locks or concurrent RCU protected lookups and traversals.
 288  */
 289 static int rhashtable_expand(struct rhashtable *ht)
 290 {
 291         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 292         int err;
 293
 294         ASSERT_RHT_MUTEX(ht);
 295
 296         old_tbl = rhashtable_last_table(ht, old_tbl);
 297
 298         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 299         if (new_tbl == NULL)
 300                 return -ENOMEM;
 301
 302         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 303         if (err)
 304                 bucket_table_free(new_tbl);
 305
 306         return err;
 307 }
 308
 309 /**
 310  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 311  * @ht:         the hash table to shrink
 312  *
 313  * This function shrinks the hash table to fit, i.e., the smallest
 314  * size would not cause it to expand right away automatically.
 315  *
 316  * The caller must ensure that no concurrent resizing occurs by holding
 317  * ht->mutex.
 318  *
 319  * The caller must ensure that no concurrent table mutations take place.
 320  * It is however valid to have concurrent lookups if they are RCU protected.
 321  *
 322  * It is valid to have concurrent insertions and deletions protected by per
 323  * bucket locks or concurrent RCU protected lookups and traversals.
 324  */
 325 static int rhashtable_shrink(struct rhashtable *ht)
 326 {
 327         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 328         unsigned int size;
 329         int err;
 330
 331         ASSERT_RHT_MUTEX(ht);
 332
 333         size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
 334         if (size < ht->p.min_size)
 335                 size = ht->p.min_size;
 336
 337         if (old_tbl->size <= size)
 338                 return 0;
 339
 340         if (rht_dereference(old_tbl->future_tbl, ht))
 341                 return -EEXIST;
 342
 343         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 344         if (new_tbl == NULL)
 345                 return -ENOMEM;
 346
 347         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 348         if (err)
 349                 bucket_table_free(new_tbl);
 350
 351         return err;
 352 }
 353
 354 static void rht_deferred_worker(struct work_struct *work)
 355 {
 356         struct rhashtable *ht;
 357         struct bucket_table *tbl;
 358         int err = 0;
 359
 360         ht = container_of(work, struct rhashtable, run_work);
 361         mutex_lock(&ht->mutex);
 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 (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 369                 rhashtable_shrink(ht);
 370
 371         err = rhashtable_rehash_table(ht);
 372
 373         mutex_unlock(&ht->mutex);
 374
 375         if (err)
 376                 schedule_work(&ht->run_work);
 377 }
 378
 379 static bool rhashtable_check_elasticity(struct rhashtable *ht,
 380                                         struct bucket_table *tbl,
 381                                         unsigned int hash)
 382 {
 383         unsigned int elasticity = ht->elasticity;
 384         struct rhash_head *head;
 385
 386         rht_for_each(head, tbl, hash)
 387                 if (!--elasticity)
 388                         return true;
 389
 390         return false;
 391 }
 392
 393 int rhashtable_insert_rehash(struct rhashtable *ht)
 394 {
 395         struct bucket_table *old_tbl;
 396         struct bucket_table *new_tbl;
 397         struct bucket_table *tbl;
 398         unsigned int size;
 399         int err;
 400
 401         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 402         tbl = rhashtable_last_table(ht, old_tbl);
 403
 404         size = tbl->size;
 405
 406         if (rht_grow_above_75(ht, tbl))
 407                 size *= 2;
 408         /* Do not schedule more than one rehash */
 409         else if (old_tbl != tbl)
 410                 return -EBUSY;
 411
 412         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 413         if (new_tbl == NULL) {
 414                 /* Schedule async resize/rehash to try allocation
 415                  * non-atomic context.
 416                  */
 417                 schedule_work(&ht->run_work);
 418                 return -ENOMEM;
 419         }
 420
 421         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 422         if (err) {
 423                 bucket_table_free(new_tbl);
 424                 if (err == -EEXIST)
 425                         err = 0;
 426         } else
 427                 schedule_work(&ht->run_work);
 428
 429         return err;
 430 }
 431 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
 432
 433 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 434                            struct rhash_head *obj,
 435                            struct bucket_table *tbl)
 436 {
 437         struct rhash_head *head;
 438         unsigned int hash;
 439         int err;
 440
 441         tbl = rhashtable_last_table(ht, tbl);
 442         hash = head_hashfn(ht, tbl, obj);
 443         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 444
 445         err = -EEXIST;
 446         if (key && rhashtable_lookup_fast(ht, key, ht->p))
 447                 goto exit;
 448
 449         err = -EAGAIN;
 450         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 451             rht_grow_above_100(ht, tbl))
 452                 goto exit;
 453
 454         err = 0;
 455
 456         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 457
 458         RCU_INIT_POINTER(obj->next, head);
 459
 460         rcu_assign_pointer(tbl->buckets[hash], obj);
 461
 462         atomic_inc(&ht->nelems);
 463
 464 exit:
 465         spin_unlock(rht_bucket_lock(tbl, hash));
 466
 467         return err;
 468 }
 469 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 470
 471 /**
 472  * rhashtable_walk_init - Initialise an iterator
 473  * @ht:         Table to walk over
 474  * @iter:       Hash table Iterator
 475  *
 476  * This function prepares a hash table walk.
 477  *
 478  * Note that if you restart a walk after rhashtable_walk_stop you
 479  * may see the same object twice.  Also, you may miss objects if
 480  * there are removals in between rhashtable_walk_stop and the next
 481  * call to rhashtable_walk_start.
 482  *
 483  * For a completely stable walk you should construct your own data
 484  * structure outside the hash table.
 485  *
 486  * This function may sleep so you must not call it from interrupt
 487  * context or with spin locks held.
 488  *
 489  * You must call rhashtable_walk_exit if this function returns
 490  * successfully.
 491  */
 492 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 493 {
 494         iter->ht = ht;
 495         iter->p = NULL;
 496         iter->slot = 0;
 497         iter->skip = 0;
 498
 499         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 500         if (!iter->walker)
 501                 return -ENOMEM;
 502
 503         mutex_lock(&ht->mutex);
 504         iter->walker->tbl = rht_dereference(ht->tbl, ht);
 505         list_add(&iter->walker->list, &iter->walker->tbl->walkers);
 506         mutex_unlock(&ht->mutex);
 507
 508         return 0;
 509 }
 510 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 511
 512 /**
 513  * rhashtable_walk_exit - Free an iterator
 514  * @iter:       Hash table Iterator
 515  *
 516  * This function frees resources allocated by rhashtable_walk_init.
 517  */
 518 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 519 {
 520         mutex_lock(&iter->ht->mutex);
 521         if (iter->walker->tbl)
 522                 list_del(&iter->walker->list);
 523         mutex_unlock(&iter->ht->mutex);
 524         kfree(iter->walker);
 525 }
 526 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 527
 528 /**
 529  * rhashtable_walk_start - Start a hash table walk
 530  * @iter:       Hash table iterator
 531  *
 532  * Start a hash table walk.  Note that we take the RCU lock in all
 533  * cases including when we return an error.  So you must always call
 534  * rhashtable_walk_stop to clean up.
 535  *
 536  * Returns zero if successful.
 537  *
 538  * Returns -EAGAIN if resize event occured.  Note that the iterator
 539  * will rewind back to the beginning and you may use it immediately
 540  * by calling rhashtable_walk_next.
 541  */
 542 int rhashtable_walk_start(struct rhashtable_iter *iter)
 543         __acquires(RCU)
 544 {
 545         struct rhashtable *ht = iter->ht;
 546
 547         mutex_lock(&ht->mutex);
 548
 549         if (iter->walker->tbl)
 550                 list_del(&iter->walker->list);
 551
 552         rcu_read_lock();
 553
 554         mutex_unlock(&ht->mutex);
 555
 556         if (!iter->walker->tbl) {
 557                 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
 558                 return -EAGAIN;
 559         }
 560
 561         return 0;
 562 }
 563 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 564
 565 /**
 566  * rhashtable_walk_next - Return the next object and advance the iterator
 567  * @iter:       Hash table iterator
 568  *
 569  * Note that you must call rhashtable_walk_stop when you are finished
 570  * with the walk.
 571  *
 572  * Returns the next object or NULL when the end of the table is reached.
 573  *
 574  * Returns -EAGAIN if resize event occured.  Note that the iterator
 575  * will rewind back to the beginning and you may continue to use it.
 576  */
 577 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 578 {
 579         struct bucket_table *tbl = iter->walker->tbl;
 580         struct rhashtable *ht = iter->ht;
 581         struct rhash_head *p = iter->p;
 582         void *obj = NULL;
 583
 584         if (p) {
 585                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 586                 goto next;
 587         }
 588
 589         for (; iter->slot < tbl->size; iter->slot++) {
 590                 int skip = iter->skip;
 591
 592                 rht_for_each_rcu(p, tbl, iter->slot) {
 593                         if (!skip)
 594                                 break;
 595                         skip--;
 596                 }
 597
 598 next:
 599                 if (!rht_is_a_nulls(p)) {
 600                         iter->skip++;
 601                         iter->p = p;
 602                         obj = rht_obj(ht, p);
 603                         goto out;
 604                 }
 605
 606                 iter->skip = 0;
 607         }
 608
 609         /* Ensure we see any new tables. */
 610         smp_rmb();
 611
 612         iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 613         if (iter->walker->tbl) {
 614                 iter->slot = 0;
 615                 iter->skip = 0;
 616                 return ERR_PTR(-EAGAIN);
 617         }
 618
 619         iter->p = NULL;
 620
 621 out:
 622
 623         return obj;
 624 }
 625 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 626
 627 /**
 628  * rhashtable_walk_stop - Finish a hash table walk
 629  * @iter:       Hash table iterator
 630  *
 631  * Finish a hash table walk.
 632  */
 633 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 634         __releases(RCU)
 635 {
 636         struct rhashtable *ht;
 637         struct bucket_table *tbl = iter->walker->tbl;
 638
 639         if (!tbl)
 640                 goto out;
 641
 642         ht = iter->ht;
 643
 644         spin_lock(&ht->lock);
 645         if (tbl->rehash < tbl->size)
 646                 list_add(&iter->walker->list, &tbl->walkers);
 647         else
 648                 iter->walker->tbl = NULL;
 649         spin_unlock(&ht->lock);
 650
 651         iter->p = NULL;
 652
 653 out:
 654         rcu_read_unlock();
 655 }
 656 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 657
 658 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 659 {
 660         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 661                    (unsigned long)params->min_size);
 662 }
 663
 664 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 665 {
 666         return jhash2(key, length, seed);
 667 }
 668
 669 /**
 670  * rhashtable_init - initialize a new hash table
 671  * @ht:         hash table to be initialized
 672  * @params:     configuration parameters
 673  *
 674  * Initializes a new hash table based on the provided configuration
 675  * parameters. A table can be configured either with a variable or
 676  * fixed length key:
 677  *
 678  * Configuration Example 1: Fixed length keys
 679  * struct test_obj {
 680  *      int                     key;
 681  *      void *                  my_member;
 682  *      struct rhash_head       node;
 683  * };
 684  *
 685  * struct rhashtable_params params = {
 686  *      .head_offset = offsetof(struct test_obj, node),
 687  *      .key_offset = offsetof(struct test_obj, key),
 688  *      .key_len = sizeof(int),
 689  *      .hashfn = jhash,
 690  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 691  * };
 692  *
 693  * Configuration Example 2: Variable length keys
 694  * struct test_obj {
 695  *      [...]
 696  *      struct rhash_head       node;
 697  * };
 698  *
 699  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 700  * {
 701  *      struct test_obj *obj = data;
 702  *
 703  *      return [... hash ...];
 704  * }
 705  *
 706  * struct rhashtable_params params = {
 707  *      .head_offset = offsetof(struct test_obj, node),
 708  *      .hashfn = jhash,
 709  *      .obj_hashfn = my_hash_fn,
 710  * };
 711  */
 712 int rhashtable_init(struct rhashtable *ht,
 713                     const struct rhashtable_params *params)
 714 {
 715         struct bucket_table *tbl;
 716         size_t size;
 717
 718         size = HASH_DEFAULT_SIZE;
 719
 720         if ((!params->key_len && !params->obj_hashfn) ||
 721             (params->obj_hashfn && !params->obj_cmpfn))
 722                 return -EINVAL;
 723
 724         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 725                 return -EINVAL;
 726
 727         if (params->nelem_hint)
 728                 size = rounded_hashtable_size(params);
 729
 730         memset(ht, 0, sizeof(*ht));
 731         mutex_init(&ht->mutex);
 732         spin_lock_init(&ht->lock);
 733         memcpy(&ht->p, params, sizeof(*params));
 734
 735         if (params->min_size)
 736                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 737
 738         if (params->max_size)
 739                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 740
 741         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 742
 743         /* The maximum (not average) chain length grows with the
 744          * size of the hash table, at a rate of (log N)/(log log N).
 745          * The value of 16 is selected so that even if the hash
 746          * table grew to 2^32 you would not expect the maximum
 747          * chain length to exceed it unless we are under attack
 748          * (or extremely unlucky).
 749          *
 750          * As this limit is only to detect attacks, we don't need
 751          * to set it to a lower value as you'd need the chain
 752          * length to vastly exceed 16 to have any real effect
 753          * on the system.
 754          */
 755         if (!params->insecure_elasticity)
 756                 ht->elasticity = 16;
 757
 758         if (params->locks_mul)
 759                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 760         else
 761                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 762
 763         ht->key_len = ht->p.key_len;
 764         if (!params->hashfn) {
 765                 ht->p.hashfn = jhash;
 766
 767                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 768                         ht->key_len /= sizeof(u32);
 769                         ht->p.hashfn = rhashtable_jhash2;
 770                 }
 771         }
 772
 773         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 774         if (tbl == NULL)
 775                 return -ENOMEM;
 776
 777         atomic_set(&ht->nelems, 0);
 778
 779         RCU_INIT_POINTER(ht->tbl, tbl);
 780
 781         INIT_WORK(&ht->run_work, rht_deferred_worker);
 782
 783         return 0;
 784 }
 785 EXPORT_SYMBOL_GPL(rhashtable_init);
 786
 787 /**
 788  * rhashtable_free_and_destroy - free elements and destroy hash table
 789  * @ht:         the hash table to destroy
 790  * @free_fn:    callback to release resources of element
 791  * @arg:        pointer passed to free_fn
 792  *
 793  * Stops an eventual async resize. If defined, invokes free_fn for each
 794  * element to releasal resources. Please note that RCU protected
 795  * readers may still be accessing the elements. Releasing of resources
 796  * must occur in a compatible manner. Then frees the bucket array.
 797  *
 798  * This function will eventually sleep to wait for an async resize
 799  * to complete. The caller is responsible that no further write operations
 800  * occurs in parallel.
 801  */
 802 void rhashtable_free_and_destroy(struct rhashtable *ht,
 803                                  void (*free_fn)(void *ptr, void *arg),
 804                                  void *arg)
 805 {
 806         const struct bucket_table *tbl;
 807         unsigned int i;
 808
 809         cancel_work_sync(&ht->run_work);
 810
 811         mutex_lock(&ht->mutex);
 812         tbl = rht_dereference(ht->tbl, ht);
 813         if (free_fn) {
 814                 for (i = 0; i < tbl->size; i++) {
 815                         struct rhash_head *pos, *next;
 816
 817                         for (pos = rht_dereference(tbl->buckets[i], ht),
 818                              next = !rht_is_a_nulls(pos) ?
 819                                         rht_dereference(pos->next, ht) : NULL;
 820                              !rht_is_a_nulls(pos);
 821                              pos = next,
 822                              next = !rht_is_a_nulls(pos) ?
 823                                         rht_dereference(pos->next, ht) : NULL)
 824                                 free_fn(rht_obj(ht, pos), arg);
 825                 }
 826         }
 827
 828         bucket_table_free(tbl);
 829         mutex_unlock(&ht->mutex);
 830 }
 831 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
 832
 833 void rhashtable_destroy(struct rhashtable *ht)
 834 {
 835         return rhashtable_free_and_destroy(ht, NULL, NULL);
 836 }
 837 EXPORT_SYMBOL_GPL(rhashtable_destroy);