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    32UL
  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, 64UL);
  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                 if (gfp != GFP_KERNEL)
  87                         gfp |= __GFP_NOWARN | __GFP_NORETRY;
  88
  89                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  90                                            gfp);
  91                 if (!tbl->locks)
  92                         return -ENOMEM;
  93                 for (i = 0; i < size; i++)
  94                         spin_lock_init(&tbl->locks[i]);
  95         }
  96         tbl->locks_mask = size - 1;
  97
  98         return 0;
  99 }
 100
 101 static void bucket_table_free(const struct bucket_table *tbl)
 102 {
 103         if (tbl)
 104                 kvfree(tbl->locks);
 105
 106         kvfree(tbl);
 107 }
 108
 109 static void bucket_table_free_rcu(struct rcu_head *head)
 110 {
 111         bucket_table_free(container_of(head, struct bucket_table, rcu));
 112 }
 113
 114 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 115                                                size_t nbuckets,
 116                                                gfp_t gfp)
 117 {
 118         struct bucket_table *tbl = NULL;
 119         size_t size;
 120         int i;
 121
 122         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 123         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 124             gfp != GFP_KERNEL)
 125                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 126         if (tbl == NULL && gfp == GFP_KERNEL)
 127                 tbl = vzalloc(size);
 128         if (tbl == NULL)
 129                 return NULL;
 130
 131         tbl->size = nbuckets;
 132
 133         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 134                 bucket_table_free(tbl);
 135                 return NULL;
 136         }
 137
 138         INIT_LIST_HEAD(&tbl->walkers);
 139
 140         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 141
 142         for (i = 0; i < nbuckets; i++)
 143                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 144
 145         return tbl;
 146 }
 147
 148 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 149                                                   struct bucket_table *tbl)
 150 {
 151         struct bucket_table *new_tbl;
 152
 153         do {
 154                 new_tbl = tbl;
 155                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 156         } while (tbl);
 157
 158         return new_tbl;
 159 }
 160
 161 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 162 {
 163         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 164         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 165                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 166         struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
 167         int err = -ENOENT;
 168         struct rhash_head *head, *next, *entry;
 169         spinlock_t *new_bucket_lock;
 170         unsigned int new_hash;
 171
 172         rht_for_each(entry, old_tbl, old_hash) {
 173                 err = 0;
 174                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 175
 176                 if (rht_is_a_nulls(next))
 177                         break;
 178
 179                 pprev = &entry->next;
 180         }
 181
 182         if (err)
 183                 goto out;
 184
 185         new_hash = head_hashfn(ht, new_tbl, entry);
 186
 187         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 188
 189         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 190         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 191                                       new_tbl, new_hash);
 192
 193         RCU_INIT_POINTER(entry->next, head);
 194
 195         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 196         spin_unlock(new_bucket_lock);
 197
 198         rcu_assign_pointer(*pprev, next);
 199
 200 out:
 201         return err;
 202 }
 203
 204 static void rhashtable_rehash_chain(struct rhashtable *ht,
 205                                     unsigned int old_hash)
 206 {
 207         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 208         spinlock_t *old_bucket_lock;
 209
 210         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 211
 212         spin_lock_bh(old_bucket_lock);
 213         while (!rhashtable_rehash_one(ht, old_hash))
 214                 ;
 215         old_tbl->rehash++;
 216         spin_unlock_bh(old_bucket_lock);
 217 }
 218
 219 static int rhashtable_rehash_attach(struct rhashtable *ht,
 220                                     struct bucket_table *old_tbl,
 221                                     struct bucket_table *new_tbl)
 222 {
 223         /* Protect future_tbl using the first bucket lock. */
 224         spin_lock_bh(old_tbl->locks);
 225
 226         /* Did somebody beat us to it? */
 227         if (rcu_access_pointer(old_tbl->future_tbl)) {
 228                 spin_unlock_bh(old_tbl->locks);
 229                 return -EEXIST;
 230         }
 231
 232         /* Make insertions go into the new, empty table right away. Deletions
 233          * and lookups will be attempted in both tables until we synchronize.
 234          */
 235         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 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
 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 int nelems = atomic_read(&ht->nelems);
 328         unsigned int size = 0;
 329         int err;
 330
 331         ASSERT_RHT_MUTEX(ht);
 332
 333         if (nelems)
 334                 size = roundup_pow_of_two(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 {
 448         struct rhash_head *head;
 449         unsigned int hash;
 450         int err;
 451
 452         tbl = rhashtable_last_table(ht, tbl);
 453         hash = head_hashfn(ht, tbl, obj);
 454         spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
 455
 456         err = -EEXIST;
 457         if (key && rhashtable_lookup_fast(ht, key, ht->p))
 458                 goto exit;
 459
 460         err = -E2BIG;
 461         if (unlikely(rht_grow_above_max(ht, tbl)))
 462                 goto exit;
 463
 464         err = -EAGAIN;
 465         if (rhashtable_check_elasticity(ht, tbl, hash) ||
 466             rht_grow_above_100(ht, tbl))
 467                 goto exit;
 468
 469         err = 0;
 470
 471         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 472
 473         RCU_INIT_POINTER(obj->next, head);
 474
 475         rcu_assign_pointer(tbl->buckets[hash], obj);
 476
 477         atomic_inc(&ht->nelems);
 478
 479 exit:
 480         spin_unlock(rht_bucket_lock(tbl, hash));
 481
 482         if (err == 0)
 483                 return NULL;
 484         else if (err == -EAGAIN)
 485                 return tbl;
 486         else
 487                 return ERR_PTR(err);
 488 }
 489 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 490
 491 /**
 492  * rhashtable_walk_enter - Initialise an iterator
 493  * @ht:         Table to walk over
 494  * @iter:       Hash table Iterator
 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 after this function returns.
 510  */
 511 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 512 {
 513         iter->ht = ht;
 514         iter->p = NULL;
 515         iter->slot = 0;
 516         iter->skip = 0;
 517
 518         spin_lock(&ht->lock);
 519         iter->walker.tbl =
 520                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 521         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 522         spin_unlock(&ht->lock);
 523 }
 524 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 525
 526 /**
 527  * rhashtable_walk_exit - Free an iterator
 528  * @iter:       Hash table Iterator
 529  *
 530  * This function frees resources allocated by rhashtable_walk_init.
 531  */
 532 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 533 {
 534         spin_lock(&iter->ht->lock);
 535         if (iter->walker.tbl)
 536                 list_del(&iter->walker.list);
 537         spin_unlock(&iter->ht->lock);
 538 }
 539 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 540
 541 /**
 542  * rhashtable_walk_start - Start a hash table walk
 543  * @iter:       Hash table iterator
 544  *
 545  * Start a hash table walk.  Note that we take the RCU lock in all
 546  * cases including when we return an error.  So you must always call
 547  * rhashtable_walk_stop to clean up.
 548  *
 549  * Returns zero if successful.
 550  *
 551  * Returns -EAGAIN if resize event occured.  Note that the iterator
 552  * will rewind back to the beginning and you may use it immediately
 553  * by calling rhashtable_walk_next.
 554  */
 555 int rhashtable_walk_start(struct rhashtable_iter *iter)
 556         __acquires(RCU)
 557 {
 558         struct rhashtable *ht = iter->ht;
 559
 560         rcu_read_lock();
 561
 562         spin_lock(&ht->lock);
 563         if (iter->walker.tbl)
 564                 list_del(&iter->walker.list);
 565         spin_unlock(&ht->lock);
 566
 567         if (!iter->walker.tbl) {
 568                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 569                 return -EAGAIN;
 570         }
 571
 572         return 0;
 573 }
 574 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 575
 576 /**
 577  * rhashtable_walk_next - Return the next object and advance the iterator
 578  * @iter:       Hash table iterator
 579  *
 580  * Note that you must call rhashtable_walk_stop when you are finished
 581  * with the walk.
 582  *
 583  * Returns the next object or NULL when the end of the table is reached.
 584  *
 585  * Returns -EAGAIN if resize event occured.  Note that the iterator
 586  * will rewind back to the beginning and you may continue to use it.
 587  */
 588 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 589 {
 590         struct bucket_table *tbl = iter->walker.tbl;
 591         struct rhashtable *ht = iter->ht;
 592         struct rhash_head *p = iter->p;
 593
 594         if (p) {
 595                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 596                 goto next;
 597         }
 598
 599         for (; iter->slot < tbl->size; iter->slot++) {
 600                 int skip = iter->skip;
 601
 602                 rht_for_each_rcu(p, tbl, iter->slot) {
 603                         if (!skip)
 604                                 break;
 605                         skip--;
 606                 }
 607
 608 next:
 609                 if (!rht_is_a_nulls(p)) {
 610                         iter->skip++;
 611                         iter->p = p;
 612                         return rht_obj(ht, p);
 613                 }
 614
 615                 iter->skip = 0;
 616         }
 617
 618         iter->p = NULL;
 619
 620         /* Ensure we see any new tables. */
 621         smp_rmb();
 622
 623         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 624         if (iter->walker.tbl) {
 625                 iter->slot = 0;
 626                 iter->skip = 0;
 627                 return ERR_PTR(-EAGAIN);
 628         }
 629
 630         return NULL;
 631 }
 632 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 633
 634 /**
 635  * rhashtable_walk_stop - Finish a hash table walk
 636  * @iter:       Hash table iterator
 637  *
 638  * Finish a hash table walk.
 639  */
 640 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 641         __releases(RCU)
 642 {
 643         struct rhashtable *ht;
 644         struct bucket_table *tbl = iter->walker.tbl;
 645
 646         if (!tbl)
 647                 goto out;
 648
 649         ht = iter->ht;
 650
 651         spin_lock(&ht->lock);
 652         if (tbl->rehash < tbl->size)
 653                 list_add(&iter->walker.list, &tbl->walkers);
 654         else
 655                 iter->walker.tbl = NULL;
 656         spin_unlock(&ht->lock);
 657
 658         iter->p = NULL;
 659
 660 out:
 661         rcu_read_unlock();
 662 }
 663 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 664
 665 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 666 {
 667         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 668                    (unsigned long)params->min_size);
 669 }
 670
 671 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 672 {
 673         return jhash2(key, length, seed);
 674 }
 675
 676 /**
 677  * rhashtable_init - initialize a new hash table
 678  * @ht:         hash table to be initialized
 679  * @params:     configuration parameters
 680  *
 681  * Initializes a new hash table based on the provided configuration
 682  * parameters. A table can be configured either with a variable or
 683  * fixed length key:
 684  *
 685  * Configuration Example 1: Fixed length keys
 686  * struct test_obj {
 687  *      int                     key;
 688  *      void *                  my_member;
 689  *      struct rhash_head       node;
 690  * };
 691  *
 692  * struct rhashtable_params params = {
 693  *      .head_offset = offsetof(struct test_obj, node),
 694  *      .key_offset = offsetof(struct test_obj, key),
 695  *      .key_len = sizeof(int),
 696  *      .hashfn = jhash,
 697  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 698  * };
 699  *
 700  * Configuration Example 2: Variable length keys
 701  * struct test_obj {
 702  *      [...]
 703  *      struct rhash_head       node;
 704  * };
 705  *
 706  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 707  * {
 708  *      struct test_obj *obj = data;
 709  *
 710  *      return [... hash ...];
 711  * }
 712  *
 713  * struct rhashtable_params params = {
 714  *      .head_offset = offsetof(struct test_obj, node),
 715  *      .hashfn = jhash,
 716  *      .obj_hashfn = my_hash_fn,
 717  * };
 718  */
 719 int rhashtable_init(struct rhashtable *ht,
 720                     const struct rhashtable_params *params)
 721 {
 722         struct bucket_table *tbl;
 723         size_t size;
 724
 725         size = HASH_DEFAULT_SIZE;
 726
 727         if ((!params->key_len && !params->obj_hashfn) ||
 728             (params->obj_hashfn && !params->obj_cmpfn))
 729                 return -EINVAL;
 730
 731         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 732                 return -EINVAL;
 733
 734         memset(ht, 0, sizeof(*ht));
 735         mutex_init(&ht->mutex);
 736         spin_lock_init(&ht->lock);
 737         memcpy(&ht->p, params, sizeof(*params));
 738
 739         if (params->min_size)
 740                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 741
 742         if (params->max_size)
 743                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 744
 745         if (params->insecure_max_entries)
 746                 ht->p.insecure_max_entries =
 747                         rounddown_pow_of_two(params->insecure_max_entries);
 748         else
 749                 ht->p.insecure_max_entries = ht->p.max_size * 2;
 750
 751         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 752
 753         if (params->nelem_hint)
 754                 size = rounded_hashtable_size(&ht->p);
 755
 756         /* The maximum (not average) chain length grows with the
 757          * size of the hash table, at a rate of (log N)/(log log N).
 758          * The value of 16 is selected so that even if the hash
 759          * table grew to 2^32 you would not expect the maximum
 760          * chain length to exceed it unless we are under attack
 761          * (or extremely unlucky).
 762          *
 763          * As this limit is only to detect attacks, we don't need
 764          * to set it to a lower value as you'd need the chain
 765          * length to vastly exceed 16 to have any real effect
 766          * on the system.
 767          */
 768         if (!params->insecure_elasticity)
 769                 ht->elasticity = 16;
 770
 771         if (params->locks_mul)
 772                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 773         else
 774                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 775
 776         ht->key_len = ht->p.key_len;
 777         if (!params->hashfn) {
 778                 ht->p.hashfn = jhash;
 779
 780                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 781                         ht->key_len /= sizeof(u32);
 782                         ht->p.hashfn = rhashtable_jhash2;
 783                 }
 784         }
 785
 786         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 787         if (tbl == NULL)
 788                 return -ENOMEM;
 789
 790         atomic_set(&ht->nelems, 0);
 791
 792         RCU_INIT_POINTER(ht->tbl, tbl);
 793
 794         INIT_WORK(&ht->run_work, rht_deferred_worker);
 795
 796         return 0;
 797 }
 798 EXPORT_SYMBOL_GPL(rhashtable_init);
 799
 800 /**
 801  * rhashtable_free_and_destroy - free elements and destroy hash table
 802  * @ht:         the hash table to destroy
 803  * @free_fn:    callback to release resources of element
 804  * @arg:        pointer passed to free_fn
 805  *
 806  * Stops an eventual async resize. If defined, invokes free_fn for each
 807  * element to releasal resources. Please note that RCU protected
 808  * readers may still be accessing the elements. Releasing of resources
 809  * must occur in a compatible manner. Then frees the bucket array.
 810  *
 811  * This function will eventually sleep to wait for an async resize
 812  * to complete. The caller is responsible that no further write operations
 813  * occurs in parallel.
 814  */
 815 void rhashtable_free_and_destroy(struct rhashtable *ht,
 816                                  void (*free_fn)(void *ptr, void *arg),
 817                                  void *arg)
 818 {
 819         const struct bucket_table *tbl;
 820         unsigned int i;
 821
 822         cancel_work_sync(&ht->run_work);
 823
 824         mutex_lock(&ht->mutex);
 825         tbl = rht_dereference(ht->tbl, ht);
 826         if (free_fn) {
 827                 for (i = 0; i < tbl->size; i++) {
 828                         struct rhash_head *pos, *next;
 829
 830                         for (pos = rht_dereference(tbl->buckets[i], ht),
 831                              next = !rht_is_a_nulls(pos) ?
 832                                         rht_dereference(pos->next, ht) : NULL;
 833                              !rht_is_a_nulls(pos);
 834                              pos = next,
 835                              next = !rht_is_a_nulls(pos) ?
 836                                         rht_dereference(pos->next, ht) : NULL)
 837                                 free_fn(rht_obj(ht, pos), arg);
 838                 }
 839         }
 840
 841         bucket_table_free(tbl);
 842         mutex_unlock(&ht->mutex);
 843 }
 844 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
 845
 846 void rhashtable_destroy(struct rhashtable *ht)
 847 {
 848         return rhashtable_free_and_destroy(ht, NULL, NULL);
 849 }
 850 EXPORT_SYMBOL_GPL(rhashtable_destroy);