lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2014 Thomas Graf <tgraf@suug.ch>
   5  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   6  *
   7  * Based on the following paper:
   8  * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
   9  *
  10  * Code partially derived from nft_hash
  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/slab.h>
  21 #include <linux/vmalloc.h>
  22 #include <linux/mm.h>
  23 #include <linux/jhash.h>
  24 #include <linux/random.h>
  25 #include <linux/rhashtable.h>
  26
  27 #define HASH_DEFAULT_SIZE       64UL
  28 #define HASH_MIN_SIZE           4UL
  29 #define BUCKET_LOCKS_PER_CPU   128UL
  30
  31 enum {
  32         RHT_LOCK_NORMAL,
  33         RHT_LOCK_NESTED,
  34         RHT_LOCK_NESTED2,
  35 };
  36
  37 /* The bucket lock is selected based on the hash and protects mutations
  38  * on a group of hash buckets.
  39  *
  40  * IMPORTANT: When holding the bucket lock of both the old and new table
  41  * during expansions and shrinking, the old bucket lock must always be
  42  * acquired first.
  43  */
  44 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  45 {
  46         return &tbl->locks[hash & tbl->locks_mask];
  47 }
  48
  49 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  50 #define ASSERT_BUCKET_LOCK(TBL, HASH) \
  51         BUG_ON(!lockdep_rht_bucket_is_held(TBL, HASH))
  52
  53 #ifdef CONFIG_PROVE_LOCKING
  54 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  55 {
  56         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  57 }
  58 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  59
  60 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  61 {
  62         spinlock_t *lock = bucket_lock(tbl, hash);
  63
  64         return (debug_locks) ? lockdep_is_held(lock) : 1;
  65 }
  66 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  67 #endif
  68
  69 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  70 {
  71         return (void *) he - ht->p.head_offset;
  72 }
  73
  74 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  75 {
  76         return hash & (tbl->size - 1);
  77 }
  78
  79 static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr)
  80 {
  81         u32 hash;
  82
  83         if (unlikely(!ht->p.key_len))
  84                 hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
  85         else
  86                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  87                                     ht->p.hash_rnd);
  88
  89         return hash;
  90 }
  91
  92 static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len)
  93 {
  94         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
  95         u32 hash;
  96
  97         hash = ht->p.hashfn(key, len, ht->p.hash_rnd);
  98
  99         return rht_bucket_index(tbl, hash);
 100 }
 101
 102 static u32 head_hashfn(const struct rhashtable *ht,
 103                        const struct bucket_table *tbl,
 104                        const struct rhash_head *he)
 105 {
 106         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
 107 }
 108
 109 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 110 {
 111         struct rhash_head __rcu **pprev;
 112
 113         for (pprev = &tbl->buckets[n];
 114              rht_dereference_bucket(*pprev, tbl, n);
 115              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 116                 ;
 117
 118         return pprev;
 119 }
 120
 121 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 122 {
 123         unsigned int i, size;
 124 #if defined(CONFIG_PROVE_LOCKING)
 125         unsigned int nr_pcpus = 2;
 126 #else
 127         unsigned int nr_pcpus = num_possible_cpus();
 128 #endif
 129
 130         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 131         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 132
 133         /* Never allocate more than one lock per bucket */
 134         size = min_t(unsigned int, size, tbl->size);
 135
 136         if (sizeof(spinlock_t) != 0) {
 137 #ifdef CONFIG_NUMA
 138                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 139                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 140                 else
 141 #endif
 142                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 143                                            GFP_KERNEL);
 144                 if (!tbl->locks)
 145                         return -ENOMEM;
 146                 for (i = 0; i < size; i++)
 147                         spin_lock_init(&tbl->locks[i]);
 148         }
 149         tbl->locks_mask = size - 1;
 150
 151         return 0;
 152 }
 153
 154 static void bucket_table_free(const struct bucket_table *tbl)
 155 {
 156         if (tbl)
 157                 kvfree(tbl->locks);
 158
 159         kvfree(tbl);
 160 }
 161
 162 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 163                                                size_t nbuckets)
 164 {
 165         struct bucket_table *tbl;
 166         size_t size;
 167
 168         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 169         tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
 170         if (tbl == NULL)
 171                 tbl = vzalloc(size);
 172
 173         if (tbl == NULL)
 174                 return NULL;
 175
 176         tbl->size = nbuckets;
 177
 178         if (alloc_bucket_locks(ht, tbl) < 0) {
 179                 bucket_table_free(tbl);
 180                 return NULL;
 181         }
 182
 183         return tbl;
 184 }
 185
 186 /**
 187  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 188  * @ht:         hash table
 189  * @new_size:   new table size
 190  */
 191 bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 192 {
 193         /* Expand table when exceeding 75% load */
 194         return atomic_read(&ht->nelems) > (new_size / 4 * 3);
 195 }
 196 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 197
 198 /**
 199  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 200  * @ht:         hash table
 201  * @new_size:   new table size
 202  */
 203 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 204 {
 205         /* Shrink table beneath 30% load */
 206         return atomic_read(&ht->nelems) < (new_size * 3 / 10);
 207 }
 208 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 209
 210 static void hashtable_chain_unzip(const struct rhashtable *ht,
 211                                   const struct bucket_table *new_tbl,
 212                                   struct bucket_table *old_tbl,
 213                                   size_t old_hash)
 214 {
 215         struct rhash_head *he, *p, *next;
 216         spinlock_t *new_bucket_lock, *new_bucket_lock2 = NULL;
 217         unsigned int new_hash, new_hash2;
 218
 219         ASSERT_BUCKET_LOCK(old_tbl, old_hash);
 220
 221         /* Old bucket empty, no work needed. */
 222         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 223                                    old_hash);
 224         if (!p)
 225                 return;
 226
 227         new_hash = new_hash2 = head_hashfn(ht, new_tbl, p);
 228         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 229
 230         /* Advance the old bucket pointer one or more times until it
 231          * reaches a node that doesn't hash to the same bucket as the
 232          * previous node p. Call the previous node p;
 233          */
 234         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 235                 new_hash2 = head_hashfn(ht, new_tbl, he);
 236                 if (new_hash != new_hash2)
 237                         break;
 238                 p = he;
 239         }
 240         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 241
 242         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 243
 244         /* If we have encountered an entry that maps to a different bucket in
 245          * the new table, lock down that bucket as well as we might cut off
 246          * the end of the chain.
 247          */
 248         new_bucket_lock2 = bucket_lock(new_tbl, new_hash);
 249         if (new_bucket_lock != new_bucket_lock2)
 250                 spin_lock_bh_nested(new_bucket_lock2, RHT_LOCK_NESTED2);
 251
 252         /* Find the subsequent node which does hash to the same
 253          * bucket as node P, or NULL if no such node exists.
 254          */
 255         next = NULL;
 256         if (he) {
 257                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 258                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 259                                 next = he;
 260                                 break;
 261                         }
 262                 }
 263         }
 264
 265         /* Set p's next pointer to that subsequent node pointer,
 266          * bypassing the nodes which do not hash to p's bucket
 267          */
 268         rcu_assign_pointer(p->next, next);
 269
 270         if (new_bucket_lock != new_bucket_lock2)
 271                 spin_unlock_bh(new_bucket_lock2);
 272         spin_unlock_bh(new_bucket_lock);
 273 }
 274
 275 static void link_old_to_new(struct bucket_table *new_tbl,
 276                             unsigned int new_hash, struct rhash_head *entry)
 277 {
 278         spinlock_t *new_bucket_lock;
 279
 280         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 281
 282         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 283         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 284         spin_unlock_bh(new_bucket_lock);
 285 }
 286
 287 /**
 288  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 289  * @ht:         the hash table to expand
 290  *
 291  * A secondary bucket array is allocated and the hash entries are migrated
 292  * while keeping them on both lists until the end of the RCU grace period.
 293  *
 294  * This function may only be called in a context where it is safe to call
 295  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 296  *
 297  * The caller must ensure that no concurrent resizing occurs by holding
 298  * ht->mutex.
 299  *
 300  * It is valid to have concurrent insertions and deletions protected by per
 301  * bucket locks or concurrent RCU protected lookups and traversals.
 302  */
 303 int rhashtable_expand(struct rhashtable *ht)
 304 {
 305         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 306         struct rhash_head *he;
 307         spinlock_t *old_bucket_lock;
 308         unsigned int new_hash, old_hash;
 309         bool complete = false;
 310
 311         ASSERT_RHT_MUTEX(ht);
 312
 313         if (ht->p.max_shift && ht->shift >= ht->p.max_shift)
 314                 return 0;
 315
 316         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 317         if (new_tbl == NULL)
 318                 return -ENOMEM;
 319
 320         ht->shift++;
 321
 322         /* Make insertions go into the new, empty table right away. Deletions
 323          * and lookups will be attempted in both tables until we synchronize.
 324          * The synchronize_rcu() guarantees for the new table to be picked up
 325          * so no new additions go into the old table while we relink.
 326          */
 327         rcu_assign_pointer(ht->future_tbl, new_tbl);
 328         synchronize_rcu();
 329
 330         /* For each new bucket, search the corresponding old bucket for the
 331          * first entry that hashes to the new bucket, and link the end of
 332          * newly formed bucket chain (containing entries added to future
 333          * table) to that entry. Since all the entries which will end up in
 334          * the new bucket appear in the same old bucket, this constructs an
 335          * entirely valid new hash table, but with multiple buckets
 336          * "zipped" together into a single imprecise chain.
 337          */
 338         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 339                 old_hash = rht_bucket_index(old_tbl, new_hash);
 340                 old_bucket_lock = bucket_lock(old_tbl, old_hash);
 341
 342                 spin_lock_bh(old_bucket_lock);
 343                 rht_for_each(he, old_tbl, old_hash) {
 344                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 345                                 link_old_to_new(new_tbl, new_hash, he);
 346                                 break;
 347                         }
 348                 }
 349                 spin_unlock_bh(old_bucket_lock);
 350         }
 351
 352         /* Publish the new table pointer. Lookups may now traverse
 353          * the new table, but they will not benefit from any
 354          * additional efficiency until later steps unzip the buckets.
 355          */
 356         rcu_assign_pointer(ht->tbl, new_tbl);
 357
 358         /* Unzip interleaved hash chains */
 359         while (!complete && !ht->being_destroyed) {
 360                 /* Wait for readers. All new readers will see the new
 361                  * table, and thus no references to the old table will
 362                  * remain.
 363                  */
 364                 synchronize_rcu();
 365
 366                 /* For each bucket in the old table (each of which
 367                  * contains items from multiple buckets of the new
 368                  * table): ...
 369                  */
 370                 complete = true;
 371                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 372                         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 373                         spin_lock_bh(old_bucket_lock);
 374
 375                         hashtable_chain_unzip(ht, new_tbl, old_tbl, old_hash);
 376                         if (old_tbl->buckets[old_hash] != NULL)
 377                                 complete = false;
 378
 379                         spin_unlock_bh(old_bucket_lock);
 380                 }
 381         }
 382
 383         bucket_table_free(old_tbl);
 384         return 0;
 385 }
 386 EXPORT_SYMBOL_GPL(rhashtable_expand);
 387
 388 /**
 389  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 390  * @ht:         the hash table to shrink
 391  *
 392  * This function may only be called in a context where it is safe to call
 393  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 394  *
 395  * The caller must ensure that no concurrent resizing occurs by holding
 396  * ht->mutex.
 397  *
 398  * The caller must ensure that no concurrent table mutations take place.
 399  * It is however valid to have concurrent lookups if they are RCU protected.
 400  *
 401  * It is valid to have concurrent insertions and deletions protected by per
 402  * bucket locks or concurrent RCU protected lookups and traversals.
 403  */
 404 int rhashtable_shrink(struct rhashtable *ht)
 405 {
 406         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 407         spinlock_t *new_bucket_lock, *old_bucket_lock1, *old_bucket_lock2;
 408         unsigned int new_hash;
 409
 410         ASSERT_RHT_MUTEX(ht);
 411
 412         if (ht->shift <= ht->p.min_shift)
 413                 return 0;
 414
 415         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 416         if (new_tbl == NULL)
 417                 return -ENOMEM;
 418
 419         rcu_assign_pointer(ht->future_tbl, new_tbl);
 420         synchronize_rcu();
 421
 422         /* Link the first entry in the old bucket to the end of the
 423          * bucket in the new table. As entries are concurrently being
 424          * added to the new table, lock down the new bucket. As we
 425          * always divide the size in half when shrinking, each bucket
 426          * in the new table maps to exactly two buckets in the old
 427          * table.
 428          *
 429          * As removals can occur concurrently on the old table, we need
 430          * to lock down both matching buckets in the old table.
 431          */
 432         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 433                 old_bucket_lock1 = bucket_lock(tbl, new_hash);
 434                 old_bucket_lock2 = bucket_lock(tbl, new_hash + new_tbl->size);
 435                 new_bucket_lock = bucket_lock(new_tbl, new_hash);
 436
 437                 spin_lock_bh(old_bucket_lock1);
 438                 spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED);
 439                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2);
 440
 441                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 442                                    tbl->buckets[new_hash]);
 443                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 444                                    tbl->buckets[new_hash + new_tbl->size]);
 445
 446                 spin_unlock_bh(new_bucket_lock);
 447                 spin_unlock_bh(old_bucket_lock2);
 448                 spin_unlock_bh(old_bucket_lock1);
 449         }
 450
 451         /* Publish the new, valid hash table */
 452         rcu_assign_pointer(ht->tbl, new_tbl);
 453         ht->shift--;
 454
 455         /* Wait for readers. No new readers will have references to the
 456          * old hash table.
 457          */
 458         synchronize_rcu();
 459
 460         bucket_table_free(tbl);
 461
 462         return 0;
 463 }
 464 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 465
 466 static void rht_deferred_worker(struct work_struct *work)
 467 {
 468         struct rhashtable *ht;
 469         struct bucket_table *tbl;
 470
 471         ht = container_of(work, struct rhashtable, run_work.work);
 472         mutex_lock(&ht->mutex);
 473         tbl = rht_dereference(ht->tbl, ht);
 474
 475         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 476                 rhashtable_expand(ht);
 477         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 478                 rhashtable_shrink(ht);
 479
 480         mutex_unlock(&ht->mutex);
 481 }
 482
 483 /**
 484  * rhashtable_insert - insert object into hash hash table
 485  * @ht:         hash table
 486  * @obj:        pointer to hash head inside object
 487  *
 488  * Will take a per bucket spinlock to protect against mutual mutations
 489  * on the same bucket. Multiple insertions may occur in parallel unless
 490  * they map to the same bucket lock.
 491  *
 492  * It is safe to call this function from atomic context.
 493  *
 494  * Will trigger an automatic deferred table resizing if the size grows
 495  * beyond the watermark indicated by grow_decision() which can be passed
 496  * to rhashtable_init().
 497  */
 498 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 499 {
 500         struct bucket_table *tbl;
 501         spinlock_t *lock;
 502         unsigned hash;
 503
 504         rcu_read_lock();
 505
 506         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 507         hash = head_hashfn(ht, tbl, obj);
 508         lock = bucket_lock(tbl, hash);
 509
 510         spin_lock_bh(lock);
 511         RCU_INIT_POINTER(obj->next, tbl->buckets[hash]);
 512         rcu_assign_pointer(tbl->buckets[hash], obj);
 513         spin_unlock_bh(lock);
 514
 515         atomic_inc(&ht->nelems);
 516
 517         /* Only grow the table if no resizing is currently in progress. */
 518         if (ht->tbl != ht->future_tbl &&
 519             ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 520                 schedule_delayed_work(&ht->run_work, 0);
 521
 522         rcu_read_unlock();
 523 }
 524 EXPORT_SYMBOL_GPL(rhashtable_insert);
 525
 526 /**
 527  * rhashtable_remove - remove object from hash table
 528  * @ht:         hash table
 529  * @obj:        pointer to hash head inside object
 530  *
 531  * Since the hash chain is single linked, the removal operation needs to
 532  * walk the bucket chain upon removal. The removal operation is thus
 533  * considerable slow if the hash table is not correctly sized.
 534  *
 535  * Will automatically shrink the table via rhashtable_expand() if the the
 536  * shrink_decision function specified at rhashtable_init() returns true.
 537  *
 538  * The caller must ensure that no concurrent table mutations occur. It is
 539  * however valid to have concurrent lookups if they are RCU protected.
 540  */
 541 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 542 {
 543         struct bucket_table *tbl;
 544         struct rhash_head __rcu **pprev;
 545         struct rhash_head *he;
 546         spinlock_t *lock;
 547         unsigned int hash;
 548
 549         rcu_read_lock();
 550         tbl = rht_dereference_rcu(ht->tbl, ht);
 551         hash = head_hashfn(ht, tbl, obj);
 552
 553         lock = bucket_lock(tbl, hash);
 554         spin_lock_bh(lock);
 555
 556 restart:
 557         pprev = &tbl->buckets[hash];
 558         rht_for_each(he, tbl, hash) {
 559                 if (he != obj) {
 560                         pprev = &he->next;
 561                         continue;
 562                 }
 563
 564                 rcu_assign_pointer(*pprev, obj->next);
 565                 atomic_dec(&ht->nelems);
 566
 567                 spin_unlock_bh(lock);
 568
 569                 if (ht->tbl != ht->future_tbl &&
 570                     ht->p.shrink_decision &&
 571                     ht->p.shrink_decision(ht, tbl->size))
 572                         schedule_delayed_work(&ht->run_work, 0);
 573
 574                 rcu_read_unlock();
 575
 576                 return true;
 577         }
 578
 579         if (tbl != rht_dereference_rcu(ht->tbl, ht)) {
 580                 spin_unlock_bh(lock);
 581
 582                 tbl = rht_dereference_rcu(ht->tbl, ht);
 583                 hash = head_hashfn(ht, tbl, obj);
 584
 585                 lock = bucket_lock(tbl, hash);
 586                 spin_lock_bh(lock);
 587                 goto restart;
 588         }
 589
 590         spin_unlock_bh(lock);
 591         rcu_read_unlock();
 592
 593         return false;
 594 }
 595 EXPORT_SYMBOL_GPL(rhashtable_remove);
 596
 597 /**
 598  * rhashtable_lookup - lookup key in hash table
 599  * @ht:         hash table
 600  * @key:        pointer to key
 601  *
 602  * Computes the hash value for the key and traverses the bucket chain looking
 603  * for a entry with an identical key. The first matching entry is returned.
 604  *
 605  * This lookup function may only be used for fixed key hash table (key_len
 606  * paramter set). It will BUG() if used inappropriately.
 607  *
 608  * Lookups may occur in parallel with hashtable mutations and resizing.
 609  */
 610 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 611 {
 612         const struct bucket_table *tbl, *old_tbl;
 613         struct rhash_head *he;
 614         u32 hash;
 615
 616         BUG_ON(!ht->p.key_len);
 617
 618         rcu_read_lock();
 619         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 620         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 621         hash = key_hashfn(ht, key, ht->p.key_len);
 622 restart:
 623         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 624                 if (memcmp(rht_obj(ht, he) + ht->p.key_offset, key,
 625                            ht->p.key_len))
 626                         continue;
 627                 rcu_read_unlock();
 628                 return rht_obj(ht, he);
 629         }
 630
 631         if (unlikely(tbl != old_tbl)) {
 632                 tbl = old_tbl;
 633                 goto restart;
 634         }
 635
 636         rcu_read_unlock();
 637         return NULL;
 638 }
 639 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 640
 641 /**
 642  * rhashtable_lookup_compare - search hash table with compare function
 643  * @ht:         hash table
 644  * @key:        the pointer to the key
 645  * @compare:    compare function, must return true on match
 646  * @arg:        argument passed on to compare function
 647  *
 648  * Traverses the bucket chain behind the provided hash value and calls the
 649  * specified compare function for each entry.
 650  *
 651  * Lookups may occur in parallel with hashtable mutations and resizing.
 652  *
 653  * Returns the first entry on which the compare function returned true.
 654  */
 655 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 656                                 bool (*compare)(void *, void *), void *arg)
 657 {
 658         const struct bucket_table *tbl, *old_tbl;
 659         struct rhash_head *he;
 660         u32 hash;
 661
 662         rcu_read_lock();
 663
 664         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 665         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 666         hash = key_hashfn(ht, key, ht->p.key_len);
 667 restart:
 668         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 669                 if (!compare(rht_obj(ht, he), arg))
 670                         continue;
 671                 rcu_read_unlock();
 672                 return rht_obj(ht, he);
 673         }
 674
 675         if (unlikely(tbl != old_tbl)) {
 676                 tbl = old_tbl;
 677                 goto restart;
 678         }
 679         rcu_read_unlock();
 680
 681         return NULL;
 682 }
 683 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 684
 685 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 686 {
 687         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 688                    1UL << params->min_shift);
 689 }
 690
 691 /**
 692  * rhashtable_init - initialize a new hash table
 693  * @ht:         hash table to be initialized
 694  * @params:     configuration parameters
 695  *
 696  * Initializes a new hash table based on the provided configuration
 697  * parameters. A table can be configured either with a variable or
 698  * fixed length key:
 699  *
 700  * Configuration Example 1: Fixed length keys
 701  * struct test_obj {
 702  *      int                     key;
 703  *      void *                  my_member;
 704  *      struct rhash_head       node;
 705  * };
 706  *
 707  * struct rhashtable_params params = {
 708  *      .head_offset = offsetof(struct test_obj, node),
 709  *      .key_offset = offsetof(struct test_obj, key),
 710  *      .key_len = sizeof(int),
 711  *      .hashfn = jhash,
 712  * };
 713  *
 714  * Configuration Example 2: Variable length keys
 715  * struct test_obj {
 716  *      [...]
 717  *      struct rhash_head       node;
 718  * };
 719  *
 720  * u32 my_hash_fn(const void *data, u32 seed)
 721  * {
 722  *      struct test_obj *obj = data;
 723  *
 724  *      return [... hash ...];
 725  * }
 726  *
 727  * struct rhashtable_params params = {
 728  *      .head_offset = offsetof(struct test_obj, node),
 729  *      .hashfn = jhash,
 730  *      .obj_hashfn = my_hash_fn,
 731  * };
 732  */
 733 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
 734 {
 735         struct bucket_table *tbl;
 736         size_t size;
 737
 738         size = HASH_DEFAULT_SIZE;
 739
 740         if ((params->key_len && !params->hashfn) ||
 741             (!params->key_len && !params->obj_hashfn))
 742                 return -EINVAL;
 743
 744         params->min_shift = max_t(size_t, params->min_shift,
 745                                   ilog2(HASH_MIN_SIZE));
 746
 747         if (params->nelem_hint)
 748                 size = rounded_hashtable_size(params);
 749
 750         memset(ht, 0, sizeof(*ht));
 751         mutex_init(&ht->mutex);
 752         memcpy(&ht->p, params, sizeof(*params));
 753
 754         if (params->locks_mul)
 755                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 756         else
 757                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 758
 759         tbl = bucket_table_alloc(ht, size);
 760         if (tbl == NULL)
 761                 return -ENOMEM;
 762
 763         ht->shift = ilog2(tbl->size);
 764         RCU_INIT_POINTER(ht->tbl, tbl);
 765         RCU_INIT_POINTER(ht->future_tbl, tbl);
 766
 767         if (!ht->p.hash_rnd)
 768                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
 769
 770         if (ht->p.grow_decision || ht->p.shrink_decision)
 771                 INIT_DEFERRABLE_WORK(&ht->run_work, rht_deferred_worker);
 772
 773         return 0;
 774 }
 775 EXPORT_SYMBOL_GPL(rhashtable_init);
 776
 777 /**
 778  * rhashtable_destroy - destroy hash table
 779  * @ht:         the hash table to destroy
 780  *
 781  * Frees the bucket array. This function is not rcu safe, therefore the caller
 782  * has to make sure that no resizing may happen by unpublishing the hashtable
 783  * and waiting for the quiescent cycle before releasing the bucket array.
 784  */
 785 void rhashtable_destroy(struct rhashtable *ht)
 786 {
 787         ht->being_destroyed = true;
 788
 789         mutex_lock(&ht->mutex);
 790
 791         cancel_delayed_work(&ht->run_work);
 792         bucket_table_free(rht_dereference(ht->tbl, ht));
 793
 794         mutex_unlock(&ht->mutex);
 795 }
 796 EXPORT_SYMBOL_GPL(rhashtable_destroy);
 797
 798 /**************************************************************************
 799  * Self Test
 800  **************************************************************************/
 801
 802 #ifdef CONFIG_TEST_RHASHTABLE
 803
 804 #define TEST_HT_SIZE    8
 805 #define TEST_ENTRIES    2048
 806 #define TEST_PTR        ((void *) 0xdeadbeef)
 807 #define TEST_NEXPANDS   4
 808
 809 struct test_obj {
 810         void                    *ptr;
 811         int                     value;
 812         struct rhash_head       node;
 813 };
 814
 815 static int __init test_rht_lookup(struct rhashtable *ht)
 816 {
 817         unsigned int i;
 818
 819         for (i = 0; i < TEST_ENTRIES * 2; i++) {
 820                 struct test_obj *obj;
 821                 bool expected = !(i % 2);
 822                 u32 key = i;
 823
 824                 obj = rhashtable_lookup(ht, &key);
 825
 826                 if (expected && !obj) {
 827                         pr_warn("Test failed: Could not find key %u\n", key);
 828                         return -ENOENT;
 829                 } else if (!expected && obj) {
 830                         pr_warn("Test failed: Unexpected entry found for key %u\n",
 831                                 key);
 832                         return -EEXIST;
 833                 } else if (expected && obj) {
 834                         if (obj->ptr != TEST_PTR || obj->value != i) {
 835                                 pr_warn("Test failed: Lookup value mismatch %p!=%p, %u!=%u\n",
 836                                         obj->ptr, TEST_PTR, obj->value, i);
 837                                 return -EINVAL;
 838                         }
 839                 }
 840         }
 841
 842         return 0;
 843 }
 844
 845 static void test_bucket_stats(struct rhashtable *ht, bool quiet)
 846 {
 847         unsigned int cnt, rcu_cnt, i, total = 0;
 848         struct rhash_head *pos;
 849         struct test_obj *obj;
 850         struct bucket_table *tbl;
 851
 852         tbl = rht_dereference_rcu(ht->tbl, ht);
 853         for (i = 0; i < tbl->size; i++) {
 854                 rcu_cnt = cnt = 0;
 855
 856                 if (!quiet)
 857                         pr_info(" [%#4x/%zu]", i, tbl->size);
 858
 859                 rht_for_each_entry_rcu(obj, pos, tbl, i, node) {
 860                         cnt++;
 861                         total++;
 862                         if (!quiet)
 863                                 pr_cont(" [%p],", obj);
 864                 }
 865
 866                 rht_for_each_entry_rcu(obj, pos, tbl, i, node)
 867                         rcu_cnt++;
 868
 869                 if (rcu_cnt != cnt)
 870                         pr_warn("Test failed: Chain count mismach %d != %d",
 871                                 cnt, rcu_cnt);
 872
 873                 if (!quiet)
 874                         pr_cont("\n  [%#x] first element: %p, chain length: %u\n",
 875                                 i, tbl->buckets[i], cnt);
 876         }
 877
 878         pr_info("  Traversal complete: counted=%u, nelems=%u, entries=%d\n",
 879                 total, atomic_read(&ht->nelems), TEST_ENTRIES);
 880
 881         if (total != atomic_read(&ht->nelems) || total != TEST_ENTRIES)
 882                 pr_warn("Test failed: Total count mismatch ^^^");
 883 }
 884
 885 static int __init test_rhashtable(struct rhashtable *ht)
 886 {
 887         struct bucket_table *tbl;
 888         struct test_obj *obj;
 889         struct rhash_head *pos, *next;
 890         int err;
 891         unsigned int i;
 892
 893         /*
 894          * Insertion Test:
 895          * Insert TEST_ENTRIES into table with all keys even numbers
 896          */
 897         pr_info("  Adding %d keys\n", TEST_ENTRIES);
 898         for (i = 0; i < TEST_ENTRIES; i++) {
 899                 struct test_obj *obj;
 900
 901                 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
 902                 if (!obj) {
 903                         err = -ENOMEM;
 904                         goto error;
 905                 }
 906
 907                 obj->ptr = TEST_PTR;
 908                 obj->value = i * 2;
 909
 910                 rhashtable_insert(ht, &obj->node);
 911         }
 912
 913         rcu_read_lock();
 914         test_bucket_stats(ht, true);
 915         test_rht_lookup(ht);
 916         rcu_read_unlock();
 917
 918         for (i = 0; i < TEST_NEXPANDS; i++) {
 919                 pr_info("  Table expansion iteration %u...\n", i);
 920                 mutex_lock(&ht->mutex);
 921                 rhashtable_expand(ht);
 922                 mutex_unlock(&ht->mutex);
 923
 924                 rcu_read_lock();
 925                 pr_info("  Verifying lookups...\n");
 926                 test_rht_lookup(ht);
 927                 rcu_read_unlock();
 928         }
 929
 930         for (i = 0; i < TEST_NEXPANDS; i++) {
 931                 pr_info("  Table shrinkage iteration %u...\n", i);
 932                 mutex_lock(&ht->mutex);
 933                 rhashtable_shrink(ht);
 934                 mutex_unlock(&ht->mutex);
 935
 936                 rcu_read_lock();
 937                 pr_info("  Verifying lookups...\n");
 938                 test_rht_lookup(ht);
 939                 rcu_read_unlock();
 940         }
 941
 942         rcu_read_lock();
 943         test_bucket_stats(ht, true);
 944         rcu_read_unlock();
 945
 946         pr_info("  Deleting %d keys\n", TEST_ENTRIES);
 947         for (i = 0; i < TEST_ENTRIES; i++) {
 948                 u32 key = i * 2;
 949
 950                 obj = rhashtable_lookup(ht, &key);
 951                 BUG_ON(!obj);
 952
 953                 rhashtable_remove(ht, &obj->node);
 954                 kfree(obj);
 955         }
 956
 957         return 0;
 958
 959 error:
 960         tbl = rht_dereference_rcu(ht->tbl, ht);
 961         for (i = 0; i < tbl->size; i++)
 962                 rht_for_each_entry_safe(obj, pos, next, tbl, i, node)
 963                         kfree(obj);
 964
 965         return err;
 966 }
 967
 968 static int __init test_rht_init(void)
 969 {
 970         struct rhashtable ht;
 971         struct rhashtable_params params = {
 972                 .nelem_hint = TEST_HT_SIZE,
 973                 .head_offset = offsetof(struct test_obj, node),
 974                 .key_offset = offsetof(struct test_obj, value),
 975                 .key_len = sizeof(int),
 976                 .hashfn = jhash,
 977                 .grow_decision = rht_grow_above_75,
 978                 .shrink_decision = rht_shrink_below_30,
 979         };
 980         int err;
 981
 982         pr_info("Running resizable hashtable tests...\n");
 983
 984         err = rhashtable_init(&ht, &params);
 985         if (err < 0) {
 986                 pr_warn("Test failed: Unable to initialize hashtable: %d\n",
 987                         err);
 988                 return err;
 989         }
 990
 991         err = test_rhashtable(&ht);
 992
 993         rhashtable_destroy(&ht);
 994
 995         return err;
 996 }
 997
 998 subsys_initcall(test_rht_init);
 999
1000 #endif /* CONFIG_TEST_RHASHTABLE */