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         spin_unlock_bh(old_tbl->locks);
 235
 236         return 0;
 237 }
 238
 239 static int rhashtable_rehash_table(struct rhashtable *ht)
 240 {
 241         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 242         struct bucket_table *new_tbl;
 243         struct rhashtable_walker *walker;
 244         unsigned int old_hash;
 245
 246         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 247         if (!new_tbl)
 248                 return 0;
 249
 250         for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
 251                 rhashtable_rehash_chain(ht, old_hash);
 252
 253         /* Publish the new table pointer. */
 254         rcu_assign_pointer(ht->tbl, new_tbl);
 255
 256         spin_lock(&ht->lock);
 257         list_for_each_entry(walker, &old_tbl->walkers, list)
 258                 walker->tbl = NULL;
 259         spin_unlock(&ht->lock);
 260
 261         /* Wait for readers. All new readers will see the new
 262          * table, and thus no references to the old table will
 263          * remain.
 264          */
 265         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 266
 267         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 268 }
 269
 270 /**
 271  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 272  * @ht:         the hash table to expand
 273  *
 274  * A secondary bucket array is allocated and the hash entries are migrated.
 275  *
 276  * This function may only be called in a context where it is safe to call
 277  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 278  *
 279  * The caller must ensure that no concurrent resizing occurs by holding
 280  * ht->mutex.
 281  *
 282  * It is valid to have concurrent insertions and deletions protected by per
 283  * bucket locks or concurrent RCU protected lookups and traversals.
 284  */
 285 static int rhashtable_expand(struct rhashtable *ht)
 286 {
 287         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 288         int err;
 289
 290         ASSERT_RHT_MUTEX(ht);
 291
 292         old_tbl = rhashtable_last_table(ht, old_tbl);
 293
 294         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 295         if (new_tbl == NULL)
 296                 return -ENOMEM;
 297
 298         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 299         if (err)
 300                 bucket_table_free(new_tbl);
 301
 302         return err;
 303 }
 304
 305 /**
 306  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 307  * @ht:         the hash table to shrink
 308  *
 309  * This function shrinks the hash table to fit, i.e., the smallest
 310  * size would not cause it to expand right away automatically.
 311  *
 312  * The caller must ensure that no concurrent resizing occurs by holding
 313  * ht->mutex.
 314  *
 315  * The caller must ensure that no concurrent table mutations take place.
 316  * It is however valid to have concurrent lookups if they are RCU protected.
 317  *
 318  * It is valid to have concurrent insertions and deletions protected by per
 319  * bucket locks or concurrent RCU protected lookups and traversals.
 320  */
 321 static int rhashtable_shrink(struct rhashtable *ht)
 322 {
 323         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 324         unsigned int size;
 325         int err;
 326
 327         ASSERT_RHT_MUTEX(ht);
 328
 329         size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
 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
 359         tbl = rht_dereference(ht->tbl, ht);
 360         tbl = rhashtable_last_table(ht, tbl);
 361
 362         if (rht_grow_above_75(ht, tbl))
 363                 rhashtable_expand(ht);
 364         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 365                 rhashtable_shrink(ht);
 366
 367         err = rhashtable_rehash_table(ht);
 368
 369         mutex_unlock(&ht->mutex);
 370
 371         if (err)
 372                 schedule_work(&ht->run_work);
 373 }
 374
 375 static bool rhashtable_check_elasticity(struct rhashtable *ht,
 376                                         struct bucket_table *tbl,
 377                                         unsigned int hash)
 378 {
 379         unsigned int elasticity = ht->elasticity;
 380         struct rhash_head *head;
 381
 382         rht_for_each(head, tbl, hash)
 383                 if (!--elasticity)
 384                         return true;
 385
 386         return false;
 387 }
 388
 389 int rhashtable_insert_rehash(struct rhashtable *ht,
 390                              struct bucket_table *tbl)
 391 {
 392         struct bucket_table *old_tbl;
 393         struct bucket_table *new_tbl;
 394         unsigned int size;
 395         int err;
 396
 397         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 398
 399         size = tbl->size;
 400
 401         err = -EBUSY;
 402
 403         if (rht_grow_above_75(ht, tbl))
 404                 size *= 2;
 405         /* Do not schedule more than one rehash */
 406         else if (old_tbl != tbl)
 407                 goto fail;
 408
 409         err = -ENOMEM;
 410
 411         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 412         if (new_tbl == NULL)
 413                 goto fail;
 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 fail:
 426         /* Do not fail the insert if someone else did a rehash. */
 427         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 428                 return 0;
 429
 430         /* Schedule async rehash to retry allocation in process context. */
 431         if (err == -ENOMEM)
 432                 schedule_work(&ht->run_work);
 433
 434         return err;
 435 }
 436 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
 437
 438 struct bucket_table *rhashtable_insert_slow(struct rhashtable *ht,
 439                                             const void *key,
 440                                             struct rhash_head *obj,
 441                                             struct bucket_table *tbl,
 442                                             void **data)
 443 {
 444         struct rhash_head *head;
 445         unsigned int hash;
 446         int err;
 447
 448         tbl = rhashtable_last_table(ht, tbl);
 449         hash = head_hashfn(ht, tbl, obj);
 450         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 451
 452         err = -EEXIST;
 453         if (key) {
 454                 *data = rhashtable_lookup_fast(ht, key, ht->p);
 455                 if (*data)
 456                         goto exit;
 457         }
 458
 459         err = -E2BIG;
 460         if (unlikely(rht_grow_above_max(ht, tbl)))
 461                 goto exit;
 462
 463         err = -EAGAIN;
 464         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 465             rht_grow_above_100(ht, tbl))
 466                 goto exit;
 467
 468         err = 0;
 469
 470         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 471
 472         RCU_INIT_POINTER(obj->next, head);
 473
 474         rcu_assign_pointer(tbl->buckets[hash], obj);
 475
 476         atomic_inc(&ht->nelems);
 477
 478 exit:
 479         spin_unlock(rht_bucket_lock(tbl, hash));
 480
 481         if (err == 0)
 482                 return NULL;
 483         else if (err == -EAGAIN)
 484                 return tbl;
 485         else
 486                 return ERR_PTR(err);
 487 }
 488 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 489
 490 /**
 491  * rhashtable_walk_init - Initialise an iterator
 492  * @ht:         Table to walk over
 493  * @iter:       Hash table Iterator
 494  * @gfp:        GFP flags for allocations
 495  *
 496  * This function prepares a hash table walk.
 497  *
 498  * Note that if you restart a walk after rhashtable_walk_stop you
 499  * may see the same object twice.  Also, you may miss objects if
 500  * there are removals in between rhashtable_walk_stop and the next
 501  * call to rhashtable_walk_start.
 502  *
 503  * For a completely stable walk you should construct your own data
 504  * structure outside the hash table.
 505  *
 506  * This function may sleep so you must not call it from interrupt
 507  * context or with spin locks held.
 508  *
 509  * You must call rhashtable_walk_exit if this function returns
 510  * successfully.
 511  */
 512 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter,
 513                          gfp_t gfp)
 514 {
 515         iter->ht = ht;
 516         iter->p = NULL;
 517         iter->slot = 0;
 518         iter->skip = 0;
 519
 520         iter->walker = kmalloc(sizeof(*iter->walker), gfp);
 521         if (!iter->walker)
 522                 return -ENOMEM;
 523
 524         spin_lock(&ht->lock);
 525         iter->walker->tbl =
 526                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 527         list_add(&iter->walker->list, &iter->walker->tbl->walkers);
 528         spin_unlock(&ht->lock);
 529
 530         return 0;
 531 }
 532 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 533
 534 /**
 535  * rhashtable_walk_exit - Free an iterator
 536  * @iter:       Hash table Iterator
 537  *
 538  * This function frees resources allocated by rhashtable_walk_init.
 539  */
 540 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 541 {
 542         spin_lock(&iter->ht->lock);
 543         if (iter->walker->tbl)
 544                 list_del(&iter->walker->list);
 545         spin_unlock(&iter->ht->lock);
 546         kfree(iter->walker);
 547 }
 548 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 549
 550 /**
 551  * rhashtable_walk_start - Start a hash table walk
 552  * @iter:       Hash table iterator
 553  *
 554  * Start a hash table walk.  Note that we take the RCU lock in all
 555  * cases including when we return an error.  So you must always call
 556  * rhashtable_walk_stop to clean up.
 557  *
 558  * Returns zero if successful.
 559  *
 560  * Returns -EAGAIN if resize event occured.  Note that the iterator
 561  * will rewind back to the beginning and you may use it immediately
 562  * by calling rhashtable_walk_next.
 563  */
 564 int rhashtable_walk_start(struct rhashtable_iter *iter)
 565         __acquires(RCU)
 566 {
 567         struct rhashtable *ht = iter->ht;
 568
 569         rcu_read_lock();
 570
 571         spin_lock(&ht->lock);
 572         if (iter->walker->tbl)
 573                 list_del(&iter->walker->list);
 574         spin_unlock(&ht->lock);
 575
 576         if (!iter->walker->tbl) {
 577                 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
 578                 return -EAGAIN;
 579         }
 580
 581         return 0;
 582 }
 583 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 584
 585 /**
 586  * rhashtable_walk_next - Return the next object and advance the iterator
 587  * @iter:       Hash table iterator
 588  *
 589  * Note that you must call rhashtable_walk_stop when you are finished
 590  * with the walk.
 591  *
 592  * Returns the next object or NULL when the end of the table is reached.
 593  *
 594  * Returns -EAGAIN if resize event occured.  Note that the iterator
 595  * will rewind back to the beginning and you may continue to use it.
 596  */
 597 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 598 {
 599         struct bucket_table *tbl = iter->walker->tbl;
 600         struct rhashtable *ht = iter->ht;
 601         struct rhash_head *p = iter->p;
 602
 603         if (p) {
 604                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 605                 goto next;
 606         }
 607
 608         for (; iter->slot < tbl->size; iter->slot++) {
 609                 int skip = iter->skip;
 610
 611                 rht_for_each_rcu(p, tbl, iter->slot) {
 612                         if (!skip)
 613                                 break;
 614                         skip--;
 615                 }
 616
 617 next:
 618                 if (!rht_is_a_nulls(p)) {
 619                         iter->skip++;
 620                         iter->p = p;
 621                         return rht_obj(ht, p);
 622                 }
 623
 624                 iter->skip = 0;
 625         }
 626
 627         iter->p = NULL;
 628
 629         /* Ensure we see any new tables. */
 630         smp_rmb();
 631
 632         iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 633         if (iter->walker->tbl) {
 634                 iter->slot = 0;
 635                 iter->skip = 0;
 636                 return ERR_PTR(-EAGAIN);
 637         }
 638
 639         return NULL;
 640 }
 641 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 642
 643 /**
 644  * rhashtable_walk_stop - Finish a hash table walk
 645  * @iter:       Hash table iterator
 646  *
 647  * Finish a hash table walk.
 648  */
 649 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 650         __releases(RCU)
 651 {
 652         struct rhashtable *ht;
 653         struct bucket_table *tbl = iter->walker->tbl;
 654
 655         if (!tbl)
 656                 goto out;
 657
 658         ht = iter->ht;
 659
 660         spin_lock(&ht->lock);
 661         if (tbl->rehash < tbl->size)
 662                 list_add(&iter->walker->list, &tbl->walkers);
 663         else
 664                 iter->walker->tbl = NULL;
 665         spin_unlock(&ht->lock);
 666
 667         iter->p = NULL;
 668
 669 out:
 670         rcu_read_unlock();
 671 }
 672 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 673
 674 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 675 {
 676         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 677                    (unsigned long)params->min_size);
 678 }
 679
 680 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 681 {
 682         return jhash2(key, length, seed);
 683 }
 684
 685 /**
 686  * rhashtable_init - initialize a new hash table
 687  * @ht:         hash table to be initialized
 688  * @params:     configuration parameters
 689  *
 690  * Initializes a new hash table based on the provided configuration
 691  * parameters. A table can be configured either with a variable or
 692  * fixed length key:
 693  *
 694  * Configuration Example 1: Fixed length keys
 695  * struct test_obj {
 696  *      int                     key;
 697  *      void *                  my_member;
 698  *      struct rhash_head       node;
 699  * };
 700  *
 701  * struct rhashtable_params params = {
 702  *      .head_offset = offsetof(struct test_obj, node),
 703  *      .key_offset = offsetof(struct test_obj, key),
 704  *      .key_len = sizeof(int),
 705  *      .hashfn = jhash,
 706  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 707  * };
 708  *
 709  * Configuration Example 2: Variable length keys
 710  * struct test_obj {
 711  *      [...]
 712  *      struct rhash_head       node;
 713  * };
 714  *
 715  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 716  * {
 717  *      struct test_obj *obj = data;
 718  *
 719  *      return [... hash ...];
 720  * }
 721  *
 722  * struct rhashtable_params params = {
 723  *      .head_offset = offsetof(struct test_obj, node),
 724  *      .hashfn = jhash,
 725  *      .obj_hashfn = my_hash_fn,
 726  * };
 727  */
 728 int rhashtable_init(struct rhashtable *ht,
 729                     const struct rhashtable_params *params)
 730 {
 731         struct bucket_table *tbl;
 732         size_t size;
 733
 734         size = HASH_DEFAULT_SIZE;
 735
 736         if ((!params->key_len && !params->obj_hashfn) ||
 737             (params->obj_hashfn && !params->obj_cmpfn))
 738                 return -EINVAL;
 739
 740         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 741                 return -EINVAL;
 742
 743         memset(ht, 0, sizeof(*ht));
 744         mutex_init(&ht->mutex);
 745         spin_lock_init(&ht->lock);
 746         memcpy(&ht->p, params, sizeof(*params));
 747
 748         if (params->min_size)
 749                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 750
 751         if (params->max_size)
 752                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 753
 754         if (params->insecure_max_entries)
 755                 ht->p.insecure_max_entries =
 756                         rounddown_pow_of_two(params->insecure_max_entries);
 757         else
 758                 ht->p.insecure_max_entries = ht->p.max_size * 2;
 759
 760         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 761
 762         if (params->nelem_hint)
 763                 size = rounded_hashtable_size(&ht->p);
 764
 765         /* The maximum (not average) chain length grows with the
 766          * size of the hash table, at a rate of (log N)/(log log N).
 767          * The value of 16 is selected so that even if the hash
 768          * table grew to 2^32 you would not expect the maximum
 769          * chain length to exceed it unless we are under attack
 770          * (or extremely unlucky).
 771          *
 772          * As this limit is only to detect attacks, we don't need
 773          * to set it to a lower value as you'd need the chain
 774          * length to vastly exceed 16 to have any real effect
 775          * on the system.
 776          */
 777         if (!params->insecure_elasticity)
 778                 ht->elasticity = 16;
 779
 780         if (params->locks_mul)
 781                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 782         else
 783                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 784
 785         ht->key_len = ht->p.key_len;
 786         if (!params->hashfn) {
 787                 ht->p.hashfn = jhash;
 788
 789                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 790                         ht->key_len /= sizeof(u32);
 791                         ht->p.hashfn = rhashtable_jhash2;
 792                 }
 793         }
 794
 795         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 796         if (tbl == NULL)
 797                 return -ENOMEM;
 798
 799         atomic_set(&ht->nelems, 0);
 800
 801         RCU_INIT_POINTER(ht->tbl, tbl);
 802
 803         INIT_WORK(&ht->run_work, rht_deferred_worker);
 804
 805         return 0;
 806 }
 807 EXPORT_SYMBOL_GPL(rhashtable_init);
 808
 809 /**
 810  * rhashtable_free_and_destroy - free elements and destroy hash table
 811  * @ht:         the hash table to destroy
 812  * @free_fn:    callback to release resources of element
 813  * @arg:        pointer passed to free_fn
 814  *
 815  * Stops an eventual async resize. If defined, invokes free_fn for each
 816  * element to releasal resources. Please note that RCU protected
 817  * readers may still be accessing the elements. Releasing of resources
 818  * must occur in a compatible manner. Then frees the bucket array.
 819  *
 820  * This function will eventually sleep to wait for an async resize
 821  * to complete. The caller is responsible that no further write operations
 822  * occurs in parallel.
 823  */
 824 void rhashtable_free_and_destroy(struct rhashtable *ht,
 825                                  void (*free_fn)(void *ptr, void *arg),
 826                                  void *arg)
 827 {
 828         const struct bucket_table *tbl;
 829         unsigned int i;
 830
 831         cancel_work_sync(&ht->run_work);
 832
 833         mutex_lock(&ht->mutex);
 834         tbl = rht_dereference(ht->tbl, ht);
 835         if (free_fn) {
 836                 for (i = 0; i < tbl->size; i++) {
 837                         struct rhash_head *pos, *next;
 838
 839                         for (pos = rht_dereference(tbl->buckets[i], ht),
 840                              next = !rht_is_a_nulls(pos) ?
 841                                         rht_dereference(pos->next, ht) : NULL;
 842                              !rht_is_a_nulls(pos);
 843                              pos = next,
 844                              next = !rht_is_a_nulls(pos) ?
 845                                         rht_dereference(pos->next, ht) : NULL)
 846                                 free_fn(rht_obj(ht, pos), arg);
 847                 }
 848         }
 849
 850         bucket_table_free(tbl);
 851         mutex_unlock(&ht->mutex);
 852 }
 853 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
 854
 855 void rhashtable_destroy(struct rhashtable *ht)
 856 {
 857         return rhashtable_free_and_destroy(ht, NULL, NULL);
 858 }
 859 EXPORT_SYMBOL_GPL(rhashtable_destroy);