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/atomic.h>
  18 #include <linux/kernel.h>
  19 #include <linux/init.h>
  20 #include <linux/log2.h>
  21 #include <linux/sched.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/mm.h>
  25 #include <linux/jhash.h>
  26 #include <linux/random.h>
  27 #include <linux/rhashtable.h>
  28 #include <linux/err.h>
  29 #include <linux/export.h>
  30
  31 #define HASH_DEFAULT_SIZE       64UL
  32 #define HASH_MIN_SIZE           4U
  33 #define BUCKET_LOCKS_PER_CPU   128UL
  34
  35 static u32 head_hashfn(struct rhashtable *ht,
  36                        const struct bucket_table *tbl,
  37                        const struct rhash_head *he)
  38 {
  39         return rht_head_hashfn(ht, tbl, he, ht->p);
  40 }
  41
  42 #ifdef CONFIG_PROVE_LOCKING
  43 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  44
  45 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  46 {
  47         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  48 }
  49 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  50
  51 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  52 {
  53         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  54
  55         return (debug_locks) ? lockdep_is_held(lock) : 1;
  56 }
  57 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  58 #else
  59 #define ASSERT_RHT_MUTEX(HT)
  60 #endif
  61
  62
  63 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  64                               gfp_t gfp)
  65 {
  66         unsigned int i, size;
  67 #if defined(CONFIG_PROVE_LOCKING)
  68         unsigned int nr_pcpus = 2;
  69 #else
  70         unsigned int nr_pcpus = num_possible_cpus();
  71 #endif
  72
  73         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
  74         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  75
  76         /* Never allocate more than 0.5 locks per bucket */
  77         size = min_t(unsigned int, size, tbl->size >> 1);
  78
  79         if (sizeof(spinlock_t) != 0) {
  80 #ifdef CONFIG_NUMA
  81                 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
  82                     gfp == GFP_KERNEL)
  83                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
  84                 else
  85 #endif
  86                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  87                                            gfp);
  88                 if (!tbl->locks)
  89                         return -ENOMEM;
  90                 for (i = 0; i < size; i++)
  91                         spin_lock_init(&tbl->locks[i]);
  92         }
  93         tbl->locks_mask = size - 1;
  94
  95         return 0;
  96 }
  97
  98 static void bucket_table_free(const struct bucket_table *tbl)
  99 {
 100         if (tbl)
 101                 kvfree(tbl->locks);
 102
 103         kvfree(tbl);
 104 }
 105
 106 static void bucket_table_free_rcu(struct rcu_head *head)
 107 {
 108         bucket_table_free(container_of(head, struct bucket_table, rcu));
 109 }
 110
 111 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 112                                                size_t nbuckets,
 113                                                gfp_t gfp)
 114 {
 115         struct bucket_table *tbl = NULL;
 116         size_t size;
 117         int i;
 118
 119         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 120         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 121             gfp != GFP_KERNEL)
 122                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 123         if (tbl == NULL && gfp == GFP_KERNEL)
 124                 tbl = vzalloc(size);
 125         if (tbl == NULL)
 126                 return NULL;
 127
 128         tbl->size = nbuckets;
 129
 130         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 131                 bucket_table_free(tbl);
 132                 return NULL;
 133         }
 134
 135         INIT_LIST_HEAD(&tbl->walkers);
 136
 137         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 138
 139         for (i = 0; i < nbuckets; i++)
 140                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 141
 142         return tbl;
 143 }
 144
 145 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 146                                                   struct bucket_table *tbl)
 147 {
 148         struct bucket_table *new_tbl;
 149
 150         do {
 151                 new_tbl = tbl;
 152                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 153         } while (tbl);
 154
 155         return new_tbl;
 156 }
 157
 158 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 159 {
 160         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 161         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 162                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 163         struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
 164         int err = -ENOENT;
 165         struct rhash_head *head, *next, *entry;
 166         spinlock_t *new_bucket_lock;
 167         unsigned int new_hash;
 168
 169         rht_for_each(entry, old_tbl, old_hash) {
 170                 err = 0;
 171                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 172
 173                 if (rht_is_a_nulls(next))
 174                         break;
 175
 176                 pprev = &entry->next;
 177         }
 178
 179         if (err)
 180                 goto out;
 181
 182         new_hash = head_hashfn(ht, new_tbl, entry);
 183
 184         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 185
 186         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 187         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 188                                       new_tbl, new_hash);
 189
 190         RCU_INIT_POINTER(entry->next, head);
 191
 192         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 193         spin_unlock(new_bucket_lock);
 194
 195         rcu_assign_pointer(*pprev, next);
 196
 197 out:
 198         return err;
 199 }
 200
 201 static void rhashtable_rehash_chain(struct rhashtable *ht,
 202                                     unsigned int 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 int 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                 cond_resched();
 256         }
 257
 258         /* Publish the new table pointer. */
 259         rcu_assign_pointer(ht->tbl, new_tbl);
 260
 261         spin_lock(&ht->lock);
 262         list_for_each_entry(walker, &old_tbl->walkers, list)
 263                 walker->tbl = NULL;
 264         spin_unlock(&ht->lock);
 265
 266         /* Wait for readers. All new readers will see the new
 267          * table, and thus no references to the old table will
 268          * remain.
 269          */
 270         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 271
 272         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 273 }
 274
 275 /**
 276  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 277  * @ht:         the hash table to expand
 278  *
 279  * A secondary bucket array is allocated and the hash entries are migrated.
 280  *
 281  * This function may only be called in a context where it is safe to call
 282  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 283  *
 284  * The caller must ensure that no concurrent resizing occurs by holding
 285  * ht->mutex.
 286  *
 287  * It is valid to have concurrent insertions and deletions protected by per
 288  * bucket locks or concurrent RCU protected lookups and traversals.
 289  */
 290 static int rhashtable_expand(struct rhashtable *ht)
 291 {
 292         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 293         int err;
 294
 295         ASSERT_RHT_MUTEX(ht);
 296
 297         old_tbl = rhashtable_last_table(ht, old_tbl);
 298
 299         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 300         if (new_tbl == NULL)
 301                 return -ENOMEM;
 302
 303         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 304         if (err)
 305                 bucket_table_free(new_tbl);
 306
 307         return err;
 308 }
 309
 310 /**
 311  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 312  * @ht:         the hash table to shrink
 313  *
 314  * This function shrinks the hash table to fit, i.e., the smallest
 315  * size would not cause it to expand right away automatically.
 316  *
 317  * The caller must ensure that no concurrent resizing occurs by holding
 318  * ht->mutex.
 319  *
 320  * The caller must ensure that no concurrent table mutations take place.
 321  * It is however valid to have concurrent lookups if they are RCU protected.
 322  *
 323  * It is valid to have concurrent insertions and deletions protected by per
 324  * bucket locks or concurrent RCU protected lookups and traversals.
 325  */
 326 static int rhashtable_shrink(struct rhashtable *ht)
 327 {
 328         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 329         unsigned int size;
 330         int err;
 331
 332         ASSERT_RHT_MUTEX(ht);
 333
 334         size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
 335         if (size < ht->p.min_size)
 336                 size = ht->p.min_size;
 337
 338         if (old_tbl->size <= size)
 339                 return 0;
 340
 341         if (rht_dereference(old_tbl->future_tbl, ht))
 342                 return -EEXIST;
 343
 344         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 345         if (new_tbl == NULL)
 346                 return -ENOMEM;
 347
 348         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 349         if (err)
 350                 bucket_table_free(new_tbl);
 351
 352         return err;
 353 }
 354
 355 static void rht_deferred_worker(struct work_struct *work)
 356 {
 357         struct rhashtable *ht;
 358         struct bucket_table *tbl;
 359         int err = 0;
 360
 361         ht = container_of(work, struct rhashtable, run_work);
 362         mutex_lock(&ht->mutex);
 363
 364         tbl = rht_dereference(ht->tbl, ht);
 365         tbl = rhashtable_last_table(ht, tbl);
 366
 367         if (rht_grow_above_75(ht, tbl))
 368                 rhashtable_expand(ht);
 369         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 370                 rhashtable_shrink(ht);
 371
 372         err = rhashtable_rehash_table(ht);
 373
 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 int hash)
 383 {
 384         unsigned int 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                              struct bucket_table *tbl)
 396 {
 397         struct bucket_table *old_tbl;
 398         struct bucket_table *new_tbl;
 399         unsigned int size;
 400         int err;
 401
 402         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 403
 404         size = tbl->size;
 405
 406         err = -EBUSY;
 407
 408         if (rht_grow_above_75(ht, tbl))
 409                 size *= 2;
 410         /* Do not schedule more than one rehash */
 411         else if (old_tbl != tbl)
 412                 goto fail;
 413
 414         err = -ENOMEM;
 415
 416         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 417         if (new_tbl == NULL)
 418                 goto fail;
 419
 420         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 421         if (err) {
 422                 bucket_table_free(new_tbl);
 423                 if (err == -EEXIST)
 424                         err = 0;
 425         } else
 426                 schedule_work(&ht->run_work);
 427
 428         return err;
 429
 430 fail:
 431         /* Do not fail the insert if someone else did a rehash. */
 432         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 433                 return 0;
 434
 435         /* Schedule async rehash to retry allocation in process context. */
 436         if (err == -ENOMEM)
 437                 schedule_work(&ht->run_work);
 438
 439         return err;
 440 }
 441 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
 442
 443 struct bucket_table *rhashtable_insert_slow(struct rhashtable *ht,
 444                                             const void *key,
 445                                             struct rhash_head *obj,
 446                                             struct bucket_table *tbl,
 447                                             void **data)
 448 {
 449         struct rhash_head *head;
 450         unsigned int hash;
 451         int err;
 452
 453         tbl = rhashtable_last_table(ht, tbl);
 454         hash = head_hashfn(ht, tbl, obj);
 455         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 456
 457         err = -EEXIST;
 458         if (key) {
 459                 *data = rhashtable_lookup_fast(ht, key, ht->p);
 460                 if (*data)
 461                         goto exit;
 462         }
 463
 464         err = -E2BIG;
 465         if (unlikely(rht_grow_above_max(ht, tbl)))
 466                 goto exit;
 467
 468         err = -EAGAIN;
 469         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 470             rht_grow_above_100(ht, tbl))
 471                 goto exit;
 472
 473         err = 0;
 474
 475         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 476
 477         RCU_INIT_POINTER(obj->next, head);
 478
 479         rcu_assign_pointer(tbl->buckets[hash], obj);
 480
 481         atomic_inc(&ht->nelems);
 482
 483 exit:
 484         spin_unlock(rht_bucket_lock(tbl, hash));
 485
 486         if (err == 0)
 487                 return NULL;
 488         else if (err == -EAGAIN)
 489                 return tbl;
 490         else
 491                 return ERR_PTR(err);
 492 }
 493 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 494
 495 /**
 496  * rhashtable_walk_init - Initialise an iterator
 497  * @ht:         Table to walk over
 498  * @iter:       Hash table Iterator
 499  *
 500  * This function prepares a hash table walk.
 501  *
 502  * Note that if you restart a walk after rhashtable_walk_stop you
 503  * may see the same object twice.  Also, you may miss objects if
 504  * there are removals in between rhashtable_walk_stop and the next
 505  * call to rhashtable_walk_start.
 506  *
 507  * For a completely stable walk you should construct your own data
 508  * structure outside the hash table.
 509  *
 510  * This function may sleep so you must not call it from interrupt
 511  * context or with spin locks held.
 512  *
 513  * You must call rhashtable_walk_exit if this function returns
 514  * successfully.
 515  */
 516 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 517 {
 518         iter->ht = ht;
 519         iter->p = NULL;
 520         iter->slot = 0;
 521         iter->skip = 0;
 522
 523         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 524         if (!iter->walker)
 525                 return -ENOMEM;
 526
 527         spin_lock(&ht->lock);
 528         iter->walker->tbl =
 529                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 530         list_add(&iter->walker->list, &iter->walker->tbl->walkers);
 531         spin_unlock(&ht->lock);
 532
 533         return 0;
 534 }
 535 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 536
 537 /**
 538  * rhashtable_walk_exit - Free an iterator
 539  * @iter:       Hash table Iterator
 540  *
 541  * This function frees resources allocated by rhashtable_walk_init.
 542  */
 543 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 544 {
 545         spin_lock(&iter->ht->lock);
 546         if (iter->walker->tbl)
 547                 list_del(&iter->walker->list);
 548         spin_unlock(&iter->ht->lock);
 549         kfree(iter->walker);
 550 }
 551 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 552
 553 /**
 554  * rhashtable_walk_start - Start a hash table walk
 555  * @iter:       Hash table iterator
 556  *
 557  * Start a hash table walk.  Note that we take the RCU lock in all
 558  * cases including when we return an error.  So you must always call
 559  * rhashtable_walk_stop to clean up.
 560  *
 561  * Returns zero if successful.
 562  *
 563  * Returns -EAGAIN if resize event occured.  Note that the iterator
 564  * will rewind back to the beginning and you may use it immediately
 565  * by calling rhashtable_walk_next.
 566  */
 567 int rhashtable_walk_start(struct rhashtable_iter *iter)
 568         __acquires(RCU)
 569 {
 570         struct rhashtable *ht = iter->ht;
 571
 572         rcu_read_lock();
 573
 574         spin_lock(&ht->lock);
 575         if (iter->walker->tbl)
 576                 list_del(&iter->walker->list);
 577         spin_unlock(&ht->lock);
 578
 579         if (!iter->walker->tbl) {
 580                 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
 581                 return -EAGAIN;
 582         }
 583
 584         return 0;
 585 }
 586 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 587
 588 /**
 589  * rhashtable_walk_next - Return the next object and advance the iterator
 590  * @iter:       Hash table iterator
 591  *
 592  * Note that you must call rhashtable_walk_stop when you are finished
 593  * with the walk.
 594  *
 595  * Returns the next object or NULL when the end of the table is reached.
 596  *
 597  * Returns -EAGAIN if resize event occured.  Note that the iterator
 598  * will rewind back to the beginning and you may continue to use it.
 599  */
 600 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 601 {
 602         struct bucket_table *tbl = iter->walker->tbl;
 603         struct rhashtable *ht = iter->ht;
 604         struct rhash_head *p = iter->p;
 605
 606         if (p) {
 607                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 608                 goto next;
 609         }
 610
 611         for (; iter->slot < tbl->size; iter->slot++) {
 612                 int skip = iter->skip;
 613
 614                 rht_for_each_rcu(p, tbl, iter->slot) {
 615                         if (!skip)
 616                                 break;
 617                         skip--;
 618                 }
 619
 620 next:
 621                 if (!rht_is_a_nulls(p)) {
 622                         iter->skip++;
 623                         iter->p = p;
 624                         return rht_obj(ht, p);
 625                 }
 626
 627                 iter->skip = 0;
 628         }
 629
 630         iter->p = NULL;
 631
 632         /* Ensure we see any new tables. */
 633         smp_rmb();
 634
 635         iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 636         if (iter->walker->tbl) {
 637                 iter->slot = 0;
 638                 iter->skip = 0;
 639                 return ERR_PTR(-EAGAIN);
 640         }
 641
 642         return NULL;
 643 }
 644 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 645
 646 /**
 647  * rhashtable_walk_stop - Finish a hash table walk
 648  * @iter:       Hash table iterator
 649  *
 650  * Finish a hash table walk.
 651  */
 652 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 653         __releases(RCU)
 654 {
 655         struct rhashtable *ht;
 656         struct bucket_table *tbl = iter->walker->tbl;
 657
 658         if (!tbl)
 659                 goto out;
 660
 661         ht = iter->ht;
 662
 663         spin_lock(&ht->lock);
 664         if (tbl->rehash < tbl->size)
 665                 list_add(&iter->walker->list, &tbl->walkers);
 666         else
 667                 iter->walker->tbl = NULL;
 668         spin_unlock(&ht->lock);
 669
 670         iter->p = NULL;
 671
 672 out:
 673         rcu_read_unlock();
 674 }
 675 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 676
 677 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 678 {
 679         size_t retsize;
 680
 681         if (params->nelem_hint)
 682                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 683                               (unsigned long)params->min_size);
 684         else
 685                 retsize = max(HASH_DEFAULT_SIZE,
 686                               (unsigned long)params->min_size);
 687
 688         return retsize;
 689 }
 690
 691 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 692 {
 693         return jhash2(key, length, seed);
 694 }
 695
 696 /**
 697  * rhashtable_init - initialize a new hash table
 698  * @ht:         hash table to be initialized
 699  * @params:     configuration parameters
 700  *
 701  * Initializes a new hash table based on the provided configuration
 702  * parameters. A table can be configured either with a variable or
 703  * fixed length key:
 704  *
 705  * Configuration Example 1: Fixed length keys
 706  * struct test_obj {
 707  *      int                     key;
 708  *      void *                  my_member;
 709  *      struct rhash_head       node;
 710  * };
 711  *
 712  * struct rhashtable_params params = {
 713  *      .head_offset = offsetof(struct test_obj, node),
 714  *      .key_offset = offsetof(struct test_obj, key),
 715  *      .key_len = sizeof(int),
 716  *      .hashfn = jhash,
 717  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 718  * };
 719  *
 720  * Configuration Example 2: Variable length keys
 721  * struct test_obj {
 722  *      [...]
 723  *      struct rhash_head       node;
 724  * };
 725  *
 726  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 727  * {
 728  *      struct test_obj *obj = data;
 729  *
 730  *      return [... hash ...];
 731  * }
 732  *
 733  * struct rhashtable_params params = {
 734  *      .head_offset = offsetof(struct test_obj, node),
 735  *      .hashfn = jhash,
 736  *      .obj_hashfn = my_hash_fn,
 737  * };
 738  */
 739 int rhashtable_init(struct rhashtable *ht,
 740                     const struct rhashtable_params *params)
 741 {
 742         struct bucket_table *tbl;
 743         size_t size;
 744
 745         if ((!params->key_len && !params->obj_hashfn) ||
 746             (params->obj_hashfn && !params->obj_cmpfn))
 747                 return -EINVAL;
 748
 749         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 750                 return -EINVAL;
 751
 752         memset(ht, 0, sizeof(*ht));
 753         mutex_init(&ht->mutex);
 754         spin_lock_init(&ht->lock);
 755         memcpy(&ht->p, params, sizeof(*params));
 756
 757         if (params->min_size)
 758                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 759
 760         if (params->max_size)
 761                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 762
 763         if (params->insecure_max_entries)
 764                 ht->p.insecure_max_entries =
 765                         rounddown_pow_of_two(params->insecure_max_entries);
 766         else
 767                 ht->p.insecure_max_entries = ht->p.max_size * 2;
 768
 769         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 770
 771         size = rounded_hashtable_size(&ht->p);
 772
 773         /* The maximum (not average) chain length grows with the
 774          * size of the hash table, at a rate of (log N)/(log log N).
 775          * The value of 16 is selected so that even if the hash
 776          * table grew to 2^32 you would not expect the maximum
 777          * chain length to exceed it unless we are under attack
 778          * (or extremely unlucky).
 779          *
 780          * As this limit is only to detect attacks, we don't need
 781          * to set it to a lower value as you'd need the chain
 782          * length to vastly exceed 16 to have any real effect
 783          * on the system.
 784          */
 785         if (!params->insecure_elasticity)
 786                 ht->elasticity = 16;
 787
 788         if (params->locks_mul)
 789                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 790         else
 791                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 792
 793         ht->key_len = ht->p.key_len;
 794         if (!params->hashfn) {
 795                 ht->p.hashfn = jhash;
 796
 797                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 798                         ht->key_len /= sizeof(u32);
 799                         ht->p.hashfn = rhashtable_jhash2;
 800                 }
 801         }
 802
 803         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 804         if (tbl == NULL)
 805                 return -ENOMEM;
 806
 807         atomic_set(&ht->nelems, 0);
 808
 809         RCU_INIT_POINTER(ht->tbl, tbl);
 810
 811         INIT_WORK(&ht->run_work, rht_deferred_worker);
 812
 813         return 0;
 814 }
 815 EXPORT_SYMBOL_GPL(rhashtable_init);
 816
 817 /**
 818  * rhashtable_free_and_destroy - free elements and destroy hash table
 819  * @ht:         the hash table to destroy
 820  * @free_fn:    callback to release resources of element
 821  * @arg:        pointer passed to free_fn
 822  *
 823  * Stops an eventual async resize. If defined, invokes free_fn for each
 824  * element to releasal resources. Please note that RCU protected
 825  * readers may still be accessing the elements. Releasing of resources
 826  * must occur in a compatible manner. Then frees the bucket array.
 827  *
 828  * This function will eventually sleep to wait for an async resize
 829  * to complete. The caller is responsible that no further write operations
 830  * occurs in parallel.
 831  */
 832 void rhashtable_free_and_destroy(struct rhashtable *ht,
 833                                  void (*free_fn)(void *ptr, void *arg),
 834                                  void *arg)
 835 {
 836         const struct bucket_table *tbl;
 837         unsigned int i;
 838
 839         cancel_work_sync(&ht->run_work);
 840
 841         mutex_lock(&ht->mutex);
 842         tbl = rht_dereference(ht->tbl, ht);
 843         if (free_fn) {
 844                 for (i = 0; i < tbl->size; i++) {
 845                         struct rhash_head *pos, *next;
 846
 847                         cond_resched();
 848                         for (pos = rht_dereference(tbl->buckets[i], ht),
 849                              next = !rht_is_a_nulls(pos) ?
 850                                         rht_dereference(pos->next, ht) : NULL;
 851                              !rht_is_a_nulls(pos);
 852                              pos = next,
 853                              next = !rht_is_a_nulls(pos) ?
 854                                         rht_dereference(pos->next, ht) : NULL)
 855                                 free_fn(rht_obj(ht, pos), arg);
 856                 }
 857         }
 858
 859         bucket_table_free(tbl);
 860         mutex_unlock(&ht->mutex);
 861 }
 862 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
 863
 864 void rhashtable_destroy(struct rhashtable *ht)
 865 {
 866         return rhashtable_free_and_destroy(ht, NULL, NULL);
 867 }
 868 EXPORT_SYMBOL_GPL(rhashtable_destroy);