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 /* Base bits plus 1 bit for nulls marker */
  32 #define HASH_RESERVED_SPACE     (RHT_BASE_BITS + 1)
  33
  34 enum {
  35         RHT_LOCK_NORMAL,
  36         RHT_LOCK_NESTED,
  37         RHT_LOCK_NESTED2,
  38 };
  39
  40 /* The bucket lock is selected based on the hash and protects mutations
  41  * on a group of hash buckets.
  42  *
  43  * IMPORTANT: When holding the bucket lock of both the old and new table
  44  * during expansions and shrinking, the old bucket lock must always be
  45  * acquired first.
  46  */
  47 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  48 {
  49         return &tbl->locks[hash & tbl->locks_mask];
  50 }
  51
  52 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  53 #define ASSERT_BUCKET_LOCK(TBL, HASH) \
  54         BUG_ON(!lockdep_rht_bucket_is_held(TBL, HASH))
  55
  56 #ifdef CONFIG_PROVE_LOCKING
  57 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  58 {
  59         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  60 }
  61 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  62
  63 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  64 {
  65         spinlock_t *lock = bucket_lock(tbl, hash);
  66
  67         return (debug_locks) ? lockdep_is_held(lock) : 1;
  68 }
  69 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  70 #endif
  71
  72 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  73 {
  74         return (void *) he - ht->p.head_offset;
  75 }
  76
  77 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  78 {
  79         return hash & (tbl->size - 1);
  80 }
  81
  82 static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr)
  83 {
  84         u32 hash;
  85
  86         if (unlikely(!ht->p.key_len))
  87                 hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
  88         else
  89                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  90                                     ht->p.hash_rnd);
  91
  92         return hash >> HASH_RESERVED_SPACE;
  93 }
  94
  95 static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len)
  96 {
  97         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
  98         u32 hash;
  99
 100         hash = ht->p.hashfn(key, len, ht->p.hash_rnd);
 101         hash >>= HASH_RESERVED_SPACE;
 102
 103         return rht_bucket_index(tbl, hash);
 104 }
 105
 106 static u32 head_hashfn(const struct rhashtable *ht,
 107                        const struct bucket_table *tbl,
 108                        const struct rhash_head *he)
 109 {
 110         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
 111 }
 112
 113 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 114 {
 115         struct rhash_head __rcu **pprev;
 116
 117         for (pprev = &tbl->buckets[n];
 118              !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n));
 119              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 120                 ;
 121
 122         return pprev;
 123 }
 124
 125 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 126 {
 127         unsigned int i, size;
 128 #if defined(CONFIG_PROVE_LOCKING)
 129         unsigned int nr_pcpus = 2;
 130 #else
 131         unsigned int nr_pcpus = num_possible_cpus();
 132 #endif
 133
 134         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 135         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 136
 137         /* Never allocate more than one lock per bucket */
 138         size = min_t(unsigned int, size, tbl->size);
 139
 140         if (sizeof(spinlock_t) != 0) {
 141 #ifdef CONFIG_NUMA
 142                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 143                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 144                 else
 145 #endif
 146                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 147                                            GFP_KERNEL);
 148                 if (!tbl->locks)
 149                         return -ENOMEM;
 150                 for (i = 0; i < size; i++)
 151                         spin_lock_init(&tbl->locks[i]);
 152         }
 153         tbl->locks_mask = size - 1;
 154
 155         return 0;
 156 }
 157
 158 static void bucket_table_free(const struct bucket_table *tbl)
 159 {
 160         if (tbl)
 161                 kvfree(tbl->locks);
 162
 163         kvfree(tbl);
 164 }
 165
 166 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 167                                                size_t nbuckets)
 168 {
 169         struct bucket_table *tbl;
 170         size_t size;
 171         int i;
 172
 173         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 174         tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
 175         if (tbl == NULL)
 176                 tbl = vzalloc(size);
 177
 178         if (tbl == NULL)
 179                 return NULL;
 180
 181         tbl->size = nbuckets;
 182
 183         if (alloc_bucket_locks(ht, tbl) < 0) {
 184                 bucket_table_free(tbl);
 185                 return NULL;
 186         }
 187
 188         for (i = 0; i < nbuckets; i++)
 189                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 190
 191         return tbl;
 192 }
 193
 194 /**
 195  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 196  * @ht:         hash table
 197  * @new_size:   new table size
 198  */
 199 bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 200 {
 201         /* Expand table when exceeding 75% load */
 202         return atomic_read(&ht->nelems) > (new_size / 4 * 3);
 203 }
 204 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 205
 206 /**
 207  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 208  * @ht:         hash table
 209  * @new_size:   new table size
 210  */
 211 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 212 {
 213         /* Shrink table beneath 30% load */
 214         return atomic_read(&ht->nelems) < (new_size * 3 / 10);
 215 }
 216 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 217
 218 static void hashtable_chain_unzip(const struct rhashtable *ht,
 219                                   const struct bucket_table *new_tbl,
 220                                   struct bucket_table *old_tbl,
 221                                   size_t old_hash)
 222 {
 223         struct rhash_head *he, *p, *next;
 224         spinlock_t *new_bucket_lock, *new_bucket_lock2 = NULL;
 225         unsigned int new_hash, new_hash2;
 226
 227         ASSERT_BUCKET_LOCK(old_tbl, old_hash);
 228
 229         /* Old bucket empty, no work needed. */
 230         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 231                                    old_hash);
 232         if (rht_is_a_nulls(p))
 233                 return;
 234
 235         new_hash = new_hash2 = head_hashfn(ht, new_tbl, p);
 236         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 237
 238         /* Advance the old bucket pointer one or more times until it
 239          * reaches a node that doesn't hash to the same bucket as the
 240          * previous node p. Call the previous node p;
 241          */
 242         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 243                 new_hash2 = head_hashfn(ht, new_tbl, he);
 244                 if (new_hash != new_hash2)
 245                         break;
 246                 p = he;
 247         }
 248         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 249
 250         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 251
 252         /* If we have encountered an entry that maps to a different bucket in
 253          * the new table, lock down that bucket as well as we might cut off
 254          * the end of the chain.
 255          */
 256         new_bucket_lock2 = bucket_lock(new_tbl, new_hash);
 257         if (new_bucket_lock != new_bucket_lock2)
 258                 spin_lock_bh_nested(new_bucket_lock2, RHT_LOCK_NESTED2);
 259
 260         /* Find the subsequent node which does hash to the same
 261          * bucket as node P, or NULL if no such node exists.
 262          */
 263         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 264         if (!rht_is_a_nulls(he)) {
 265                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 266                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 267                                 next = he;
 268                                 break;
 269                         }
 270                 }
 271         }
 272
 273         /* Set p's next pointer to that subsequent node pointer,
 274          * bypassing the nodes which do not hash to p's bucket
 275          */
 276         rcu_assign_pointer(p->next, next);
 277
 278         if (new_bucket_lock != new_bucket_lock2)
 279                 spin_unlock_bh(new_bucket_lock2);
 280         spin_unlock_bh(new_bucket_lock);
 281 }
 282
 283 static void link_old_to_new(struct bucket_table *new_tbl,
 284                             unsigned int new_hash, struct rhash_head *entry)
 285 {
 286         spinlock_t *new_bucket_lock;
 287
 288         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 289
 290         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 291         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 292         spin_unlock_bh(new_bucket_lock);
 293 }
 294
 295 /**
 296  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 297  * @ht:         the hash table to expand
 298  *
 299  * A secondary bucket array is allocated and the hash entries are migrated
 300  * while keeping them on both lists until the end of the RCU grace period.
 301  *
 302  * This function may only be called in a context where it is safe to call
 303  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 304  *
 305  * The caller must ensure that no concurrent resizing occurs by holding
 306  * ht->mutex.
 307  *
 308  * It is valid to have concurrent insertions and deletions protected by per
 309  * bucket locks or concurrent RCU protected lookups and traversals.
 310  */
 311 int rhashtable_expand(struct rhashtable *ht)
 312 {
 313         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 314         struct rhash_head *he;
 315         spinlock_t *old_bucket_lock;
 316         unsigned int new_hash, old_hash;
 317         bool complete = false;
 318
 319         ASSERT_RHT_MUTEX(ht);
 320
 321         if (ht->p.max_shift && ht->shift >= ht->p.max_shift)
 322                 return 0;
 323
 324         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 325         if (new_tbl == NULL)
 326                 return -ENOMEM;
 327
 328         ht->shift++;
 329
 330         /* Make insertions go into the new, empty table right away. Deletions
 331          * and lookups will be attempted in both tables until we synchronize.
 332          * The synchronize_rcu() guarantees for the new table to be picked up
 333          * so no new additions go into the old table while we relink.
 334          */
 335         rcu_assign_pointer(ht->future_tbl, new_tbl);
 336         synchronize_rcu();
 337
 338         /* For each new bucket, search the corresponding old bucket for the
 339          * first entry that hashes to the new bucket, and link the end of
 340          * newly formed bucket chain (containing entries added to future
 341          * table) to that entry. Since all the entries which will end up in
 342          * the new bucket appear in the same old bucket, this constructs an
 343          * entirely valid new hash table, but with multiple buckets
 344          * "zipped" together into a single imprecise chain.
 345          */
 346         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 347                 old_hash = rht_bucket_index(old_tbl, new_hash);
 348                 old_bucket_lock = bucket_lock(old_tbl, old_hash);
 349
 350                 spin_lock_bh(old_bucket_lock);
 351                 rht_for_each(he, old_tbl, old_hash) {
 352                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 353                                 link_old_to_new(new_tbl, new_hash, he);
 354                                 break;
 355                         }
 356                 }
 357                 spin_unlock_bh(old_bucket_lock);
 358         }
 359
 360         /* Publish the new table pointer. Lookups may now traverse
 361          * the new table, but they will not benefit from any
 362          * additional efficiency until later steps unzip the buckets.
 363          */
 364         rcu_assign_pointer(ht->tbl, new_tbl);
 365
 366         /* Unzip interleaved hash chains */
 367         while (!complete && !ht->being_destroyed) {
 368                 /* Wait for readers. All new readers will see the new
 369                  * table, and thus no references to the old table will
 370                  * remain.
 371                  */
 372                 synchronize_rcu();
 373
 374                 /* For each bucket in the old table (each of which
 375                  * contains items from multiple buckets of the new
 376                  * table): ...
 377                  */
 378                 complete = true;
 379                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 380                         struct rhash_head *head;
 381
 382                         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 383                         spin_lock_bh(old_bucket_lock);
 384
 385                         hashtable_chain_unzip(ht, new_tbl, old_tbl, old_hash);
 386                         head = rht_dereference_bucket(old_tbl->buckets[old_hash],
 387                                                       old_tbl, old_hash);
 388                         if (!rht_is_a_nulls(head))
 389                                 complete = false;
 390
 391                         spin_unlock_bh(old_bucket_lock);
 392                 }
 393         }
 394
 395         bucket_table_free(old_tbl);
 396         return 0;
 397 }
 398 EXPORT_SYMBOL_GPL(rhashtable_expand);
 399
 400 /**
 401  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 402  * @ht:         the hash table to shrink
 403  *
 404  * This function may only be called in a context where it is safe to call
 405  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 406  *
 407  * The caller must ensure that no concurrent resizing occurs by holding
 408  * ht->mutex.
 409  *
 410  * The caller must ensure that no concurrent table mutations take place.
 411  * It is however valid to have concurrent lookups if they are RCU protected.
 412  *
 413  * It is valid to have concurrent insertions and deletions protected by per
 414  * bucket locks or concurrent RCU protected lookups and traversals.
 415  */
 416 int rhashtable_shrink(struct rhashtable *ht)
 417 {
 418         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 419         spinlock_t *new_bucket_lock, *old_bucket_lock1, *old_bucket_lock2;
 420         unsigned int new_hash;
 421
 422         ASSERT_RHT_MUTEX(ht);
 423
 424         if (ht->shift <= ht->p.min_shift)
 425                 return 0;
 426
 427         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 428         if (new_tbl == NULL)
 429                 return -ENOMEM;
 430
 431         rcu_assign_pointer(ht->future_tbl, new_tbl);
 432         synchronize_rcu();
 433
 434         /* Link the first entry in the old bucket to the end of the
 435          * bucket in the new table. As entries are concurrently being
 436          * added to the new table, lock down the new bucket. As we
 437          * always divide the size in half when shrinking, each bucket
 438          * in the new table maps to exactly two buckets in the old
 439          * table.
 440          *
 441          * As removals can occur concurrently on the old table, we need
 442          * to lock down both matching buckets in the old table.
 443          */
 444         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 445                 old_bucket_lock1 = bucket_lock(tbl, new_hash);
 446                 old_bucket_lock2 = bucket_lock(tbl, new_hash + new_tbl->size);
 447                 new_bucket_lock = bucket_lock(new_tbl, new_hash);
 448
 449                 spin_lock_bh(old_bucket_lock1);
 450                 spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED);
 451                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2);
 452
 453                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 454                                    tbl->buckets[new_hash]);
 455                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 456                                    tbl->buckets[new_hash + new_tbl->size]);
 457
 458                 spin_unlock_bh(new_bucket_lock);
 459                 spin_unlock_bh(old_bucket_lock2);
 460                 spin_unlock_bh(old_bucket_lock1);
 461         }
 462
 463         /* Publish the new, valid hash table */
 464         rcu_assign_pointer(ht->tbl, new_tbl);
 465         ht->shift--;
 466
 467         /* Wait for readers. No new readers will have references to the
 468          * old hash table.
 469          */
 470         synchronize_rcu();
 471
 472         bucket_table_free(tbl);
 473
 474         return 0;
 475 }
 476 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 477
 478 static void rht_deferred_worker(struct work_struct *work)
 479 {
 480         struct rhashtable *ht;
 481         struct bucket_table *tbl;
 482
 483         ht = container_of(work, struct rhashtable, run_work.work);
 484         mutex_lock(&ht->mutex);
 485         tbl = rht_dereference(ht->tbl, ht);
 486
 487         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 488                 rhashtable_expand(ht);
 489         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 490                 rhashtable_shrink(ht);
 491
 492         mutex_unlock(&ht->mutex);
 493 }
 494
 495 /**
 496  * rhashtable_insert - insert object into hash hash table
 497  * @ht:         hash table
 498  * @obj:        pointer to hash head inside object
 499  *
 500  * Will take a per bucket spinlock to protect against mutual mutations
 501  * on the same bucket. Multiple insertions may occur in parallel unless
 502  * they map to the same bucket lock.
 503  *
 504  * It is safe to call this function from atomic context.
 505  *
 506  * Will trigger an automatic deferred table resizing if the size grows
 507  * beyond the watermark indicated by grow_decision() which can be passed
 508  * to rhashtable_init().
 509  */
 510 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 511 {
 512         struct bucket_table *tbl;
 513         struct rhash_head *head;
 514         spinlock_t *lock;
 515         unsigned hash;
 516
 517         rcu_read_lock();
 518
 519         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 520         hash = head_hashfn(ht, tbl, obj);
 521         lock = bucket_lock(tbl, hash);
 522
 523         spin_lock_bh(lock);
 524         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 525         if (rht_is_a_nulls(head))
 526                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 527         else
 528                 RCU_INIT_POINTER(obj->next, head);
 529
 530         rcu_assign_pointer(tbl->buckets[hash], obj);
 531         spin_unlock_bh(lock);
 532
 533         atomic_inc(&ht->nelems);
 534
 535         /* Only grow the table if no resizing is currently in progress. */
 536         if (ht->tbl != ht->future_tbl &&
 537             ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 538                 schedule_delayed_work(&ht->run_work, 0);
 539
 540         rcu_read_unlock();
 541 }
 542 EXPORT_SYMBOL_GPL(rhashtable_insert);
 543
 544 /**
 545  * rhashtable_remove - remove object from hash table
 546  * @ht:         hash table
 547  * @obj:        pointer to hash head inside object
 548  *
 549  * Since the hash chain is single linked, the removal operation needs to
 550  * walk the bucket chain upon removal. The removal operation is thus
 551  * considerable slow if the hash table is not correctly sized.
 552  *
 553  * Will automatically shrink the table via rhashtable_expand() if the the
 554  * shrink_decision function specified at rhashtable_init() returns true.
 555  *
 556  * The caller must ensure that no concurrent table mutations occur. It is
 557  * however valid to have concurrent lookups if they are RCU protected.
 558  */
 559 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 560 {
 561         struct bucket_table *tbl;
 562         struct rhash_head __rcu **pprev;
 563         struct rhash_head *he;
 564         spinlock_t *lock;
 565         unsigned int hash;
 566
 567         rcu_read_lock();
 568         tbl = rht_dereference_rcu(ht->tbl, ht);
 569         hash = head_hashfn(ht, tbl, obj);
 570
 571         lock = bucket_lock(tbl, hash);
 572         spin_lock_bh(lock);
 573
 574 restart:
 575         pprev = &tbl->buckets[hash];
 576         rht_for_each(he, tbl, hash) {
 577                 if (he != obj) {
 578                         pprev = &he->next;
 579                         continue;
 580                 }
 581
 582                 rcu_assign_pointer(*pprev, obj->next);
 583                 atomic_dec(&ht->nelems);
 584
 585                 spin_unlock_bh(lock);
 586
 587                 if (ht->tbl != ht->future_tbl &&
 588                     ht->p.shrink_decision &&
 589                     ht->p.shrink_decision(ht, tbl->size))
 590                         schedule_delayed_work(&ht->run_work, 0);
 591
 592                 rcu_read_unlock();
 593
 594                 return true;
 595         }
 596
 597         if (tbl != rht_dereference_rcu(ht->tbl, ht)) {
 598                 spin_unlock_bh(lock);
 599
 600                 tbl = rht_dereference_rcu(ht->tbl, ht);
 601                 hash = head_hashfn(ht, tbl, obj);
 602
 603                 lock = bucket_lock(tbl, hash);
 604                 spin_lock_bh(lock);
 605                 goto restart;
 606         }
 607
 608         spin_unlock_bh(lock);
 609         rcu_read_unlock();
 610
 611         return false;
 612 }
 613 EXPORT_SYMBOL_GPL(rhashtable_remove);
 614
 615 /**
 616  * rhashtable_lookup - lookup key in hash table
 617  * @ht:         hash table
 618  * @key:        pointer to key
 619  *
 620  * Computes the hash value for the key and traverses the bucket chain looking
 621  * for a entry with an identical key. The first matching entry is returned.
 622  *
 623  * This lookup function may only be used for fixed key hash table (key_len
 624  * paramter set). It will BUG() if used inappropriately.
 625  *
 626  * Lookups may occur in parallel with hashtable mutations and resizing.
 627  */
 628 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 629 {
 630         const struct bucket_table *tbl, *old_tbl;
 631         struct rhash_head *he;
 632         u32 hash;
 633
 634         BUG_ON(!ht->p.key_len);
 635
 636         rcu_read_lock();
 637         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 638         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 639         hash = key_hashfn(ht, key, ht->p.key_len);
 640 restart:
 641         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 642                 if (memcmp(rht_obj(ht, he) + ht->p.key_offset, key,
 643                            ht->p.key_len))
 644                         continue;
 645                 rcu_read_unlock();
 646                 return rht_obj(ht, he);
 647         }
 648
 649         if (unlikely(tbl != old_tbl)) {
 650                 tbl = old_tbl;
 651                 goto restart;
 652         }
 653
 654         rcu_read_unlock();
 655         return NULL;
 656 }
 657 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 658
 659 /**
 660  * rhashtable_lookup_compare - search hash table with compare function
 661  * @ht:         hash table
 662  * @key:        the pointer to the key
 663  * @compare:    compare function, must return true on match
 664  * @arg:        argument passed on to compare function
 665  *
 666  * Traverses the bucket chain behind the provided hash value and calls the
 667  * specified compare function for each entry.
 668  *
 669  * Lookups may occur in parallel with hashtable mutations and resizing.
 670  *
 671  * Returns the first entry on which the compare function returned true.
 672  */
 673 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 674                                 bool (*compare)(void *, void *), void *arg)
 675 {
 676         const struct bucket_table *tbl, *old_tbl;
 677         struct rhash_head *he;
 678         u32 hash;
 679
 680         rcu_read_lock();
 681
 682         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 683         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 684         hash = key_hashfn(ht, key, ht->p.key_len);
 685 restart:
 686         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 687                 if (!compare(rht_obj(ht, he), arg))
 688                         continue;
 689                 rcu_read_unlock();
 690                 return rht_obj(ht, he);
 691         }
 692
 693         if (unlikely(tbl != old_tbl)) {
 694                 tbl = old_tbl;
 695                 goto restart;
 696         }
 697         rcu_read_unlock();
 698
 699         return NULL;
 700 }
 701 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 702
 703 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 704 {
 705         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 706                    1UL << params->min_shift);
 707 }
 708
 709 /**
 710  * rhashtable_init - initialize a new hash table
 711  * @ht:         hash table to be initialized
 712  * @params:     configuration parameters
 713  *
 714  * Initializes a new hash table based on the provided configuration
 715  * parameters. A table can be configured either with a variable or
 716  * fixed length key:
 717  *
 718  * Configuration Example 1: Fixed length keys
 719  * struct test_obj {
 720  *      int                     key;
 721  *      void *                  my_member;
 722  *      struct rhash_head       node;
 723  * };
 724  *
 725  * struct rhashtable_params params = {
 726  *      .head_offset = offsetof(struct test_obj, node),
 727  *      .key_offset = offsetof(struct test_obj, key),
 728  *      .key_len = sizeof(int),
 729  *      .hashfn = jhash,
 730  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 731  * };
 732  *
 733  * Configuration Example 2: Variable length keys
 734  * struct test_obj {
 735  *      [...]
 736  *      struct rhash_head       node;
 737  * };
 738  *
 739  * u32 my_hash_fn(const void *data, u32 seed)
 740  * {
 741  *      struct test_obj *obj = data;
 742  *
 743  *      return [... hash ...];
 744  * }
 745  *
 746  * struct rhashtable_params params = {
 747  *      .head_offset = offsetof(struct test_obj, node),
 748  *      .hashfn = jhash,
 749  *      .obj_hashfn = my_hash_fn,
 750  * };
 751  */
 752 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
 753 {
 754         struct bucket_table *tbl;
 755         size_t size;
 756
 757         size = HASH_DEFAULT_SIZE;
 758
 759         if ((params->key_len && !params->hashfn) ||
 760             (!params->key_len && !params->obj_hashfn))
 761                 return -EINVAL;
 762
 763         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 764                 return -EINVAL;
 765
 766         params->min_shift = max_t(size_t, params->min_shift,
 767                                   ilog2(HASH_MIN_SIZE));
 768
 769         if (params->nelem_hint)
 770                 size = rounded_hashtable_size(params);
 771
 772         memset(ht, 0, sizeof(*ht));
 773         mutex_init(&ht->mutex);
 774         memcpy(&ht->p, params, sizeof(*params));
 775
 776         if (params->locks_mul)
 777                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 778         else
 779                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 780
 781         tbl = bucket_table_alloc(ht, size);
 782         if (tbl == NULL)
 783                 return -ENOMEM;
 784
 785         ht->shift = ilog2(tbl->size);
 786         RCU_INIT_POINTER(ht->tbl, tbl);
 787         RCU_INIT_POINTER(ht->future_tbl, tbl);
 788
 789         if (!ht->p.hash_rnd)
 790                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
 791
 792         if (ht->p.grow_decision || ht->p.shrink_decision)
 793                 INIT_DEFERRABLE_WORK(&ht->run_work, rht_deferred_worker);
 794
 795         return 0;
 796 }
 797 EXPORT_SYMBOL_GPL(rhashtable_init);
 798
 799 /**
 800  * rhashtable_destroy - destroy hash table
 801  * @ht:         the hash table to destroy
 802  *
 803  * Frees the bucket array. This function is not rcu safe, therefore the caller
 804  * has to make sure that no resizing may happen by unpublishing the hashtable
 805  * and waiting for the quiescent cycle before releasing the bucket array.
 806  */
 807 void rhashtable_destroy(struct rhashtable *ht)
 808 {
 809         ht->being_destroyed = true;
 810
 811         mutex_lock(&ht->mutex);
 812
 813         cancel_delayed_work(&ht->run_work);
 814         bucket_table_free(rht_dereference(ht->tbl, ht));
 815
 816         mutex_unlock(&ht->mutex);
 817 }
 818 EXPORT_SYMBOL_GPL(rhashtable_destroy);
 819
 820 /**************************************************************************
 821  * Self Test
 822  **************************************************************************/
 823
 824 #ifdef CONFIG_TEST_RHASHTABLE
 825
 826 #define TEST_HT_SIZE    8
 827 #define TEST_ENTRIES    2048
 828 #define TEST_PTR        ((void *) 0xdeadbeef)
 829 #define TEST_NEXPANDS   4
 830
 831 struct test_obj {
 832         void                    *ptr;
 833         int                     value;
 834         struct rhash_head       node;
 835 };
 836
 837 static int __init test_rht_lookup(struct rhashtable *ht)
 838 {
 839         unsigned int i;
 840
 841         for (i = 0; i < TEST_ENTRIES * 2; i++) {
 842                 struct test_obj *obj;
 843                 bool expected = !(i % 2);
 844                 u32 key = i;
 845
 846                 obj = rhashtable_lookup(ht, &key);
 847
 848                 if (expected && !obj) {
 849                         pr_warn("Test failed: Could not find key %u\n", key);
 850                         return -ENOENT;
 851                 } else if (!expected && obj) {
 852                         pr_warn("Test failed: Unexpected entry found for key %u\n",
 853                                 key);
 854                         return -EEXIST;
 855                 } else if (expected && obj) {
 856                         if (obj->ptr != TEST_PTR || obj->value != i) {
 857                                 pr_warn("Test failed: Lookup value mismatch %p!=%p, %u!=%u\n",
 858                                         obj->ptr, TEST_PTR, obj->value, i);
 859                                 return -EINVAL;
 860                         }
 861                 }
 862         }
 863
 864         return 0;
 865 }
 866
 867 static void test_bucket_stats(struct rhashtable *ht, bool quiet)
 868 {
 869         unsigned int cnt, rcu_cnt, i, total = 0;
 870         struct rhash_head *pos;
 871         struct test_obj *obj;
 872         struct bucket_table *tbl;
 873
 874         tbl = rht_dereference_rcu(ht->tbl, ht);
 875         for (i = 0; i < tbl->size; i++) {
 876                 rcu_cnt = cnt = 0;
 877
 878                 if (!quiet)
 879                         pr_info(" [%#4x/%zu]", i, tbl->size);
 880
 881                 rht_for_each_entry_rcu(obj, pos, tbl, i, node) {
 882                         cnt++;
 883                         total++;
 884                         if (!quiet)
 885                                 pr_cont(" [%p],", obj);
 886                 }
 887
 888                 rht_for_each_entry_rcu(obj, pos, tbl, i, node)
 889                         rcu_cnt++;
 890
 891                 if (rcu_cnt != cnt)
 892                         pr_warn("Test failed: Chain count mismach %d != %d",
 893                                 cnt, rcu_cnt);
 894
 895                 if (!quiet)
 896                         pr_cont("\n  [%#x] first element: %p, chain length: %u\n",
 897                                 i, tbl->buckets[i], cnt);
 898         }
 899
 900         pr_info("  Traversal complete: counted=%u, nelems=%u, entries=%d\n",
 901                 total, atomic_read(&ht->nelems), TEST_ENTRIES);
 902
 903         if (total != atomic_read(&ht->nelems) || total != TEST_ENTRIES)
 904                 pr_warn("Test failed: Total count mismatch ^^^");
 905 }
 906
 907 static int __init test_rhashtable(struct rhashtable *ht)
 908 {
 909         struct bucket_table *tbl;
 910         struct test_obj *obj;
 911         struct rhash_head *pos, *next;
 912         int err;
 913         unsigned int i;
 914
 915         /*
 916          * Insertion Test:
 917          * Insert TEST_ENTRIES into table with all keys even numbers
 918          */
 919         pr_info("  Adding %d keys\n", TEST_ENTRIES);
 920         for (i = 0; i < TEST_ENTRIES; i++) {
 921                 struct test_obj *obj;
 922
 923                 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
 924                 if (!obj) {
 925                         err = -ENOMEM;
 926                         goto error;
 927                 }
 928
 929                 obj->ptr = TEST_PTR;
 930                 obj->value = i * 2;
 931
 932                 rhashtable_insert(ht, &obj->node);
 933         }
 934
 935         rcu_read_lock();
 936         test_bucket_stats(ht, true);
 937         test_rht_lookup(ht);
 938         rcu_read_unlock();
 939
 940         for (i = 0; i < TEST_NEXPANDS; i++) {
 941                 pr_info("  Table expansion iteration %u...\n", i);
 942                 mutex_lock(&ht->mutex);
 943                 rhashtable_expand(ht);
 944                 mutex_unlock(&ht->mutex);
 945
 946                 rcu_read_lock();
 947                 pr_info("  Verifying lookups...\n");
 948                 test_rht_lookup(ht);
 949                 rcu_read_unlock();
 950         }
 951
 952         for (i = 0; i < TEST_NEXPANDS; i++) {
 953                 pr_info("  Table shrinkage iteration %u...\n", i);
 954                 mutex_lock(&ht->mutex);
 955                 rhashtable_shrink(ht);
 956                 mutex_unlock(&ht->mutex);
 957
 958                 rcu_read_lock();
 959                 pr_info("  Verifying lookups...\n");
 960                 test_rht_lookup(ht);
 961                 rcu_read_unlock();
 962         }
 963
 964         rcu_read_lock();
 965         test_bucket_stats(ht, true);
 966         rcu_read_unlock();
 967
 968         pr_info("  Deleting %d keys\n", TEST_ENTRIES);
 969         for (i = 0; i < TEST_ENTRIES; i++) {
 970                 u32 key = i * 2;
 971
 972                 obj = rhashtable_lookup(ht, &key);
 973                 BUG_ON(!obj);
 974
 975                 rhashtable_remove(ht, &obj->node);
 976                 kfree(obj);
 977         }
 978
 979         return 0;
 980
 981 error:
 982         tbl = rht_dereference_rcu(ht->tbl, ht);
 983         for (i = 0; i < tbl->size; i++)
 984                 rht_for_each_entry_safe(obj, pos, next, tbl, i, node)
 985                         kfree(obj);
 986
 987         return err;
 988 }
 989
 990 static int __init test_rht_init(void)
 991 {
 992         struct rhashtable ht;
 993         struct rhashtable_params params = {
 994                 .nelem_hint = TEST_HT_SIZE,
 995                 .head_offset = offsetof(struct test_obj, node),
 996                 .key_offset = offsetof(struct test_obj, value),
 997                 .key_len = sizeof(int),
 998                 .hashfn = jhash,
 999                 .nulls_base = (3U << RHT_BASE_SHIFT),
1000                 .grow_decision = rht_grow_above_75,
1001                 .shrink_decision = rht_shrink_below_30,
1002         };
1003         int err;
1004
1005         pr_info("Running resizable hashtable tests...\n");
1006
1007         err = rhashtable_init(&ht, &params);
1008         if (err < 0) {
1009                 pr_warn("Test failed: Unable to initialize hashtable: %d\n",
1010                         err);
1011                 return err;
1012         }
1013
1014         err = test_rhashtable(&ht);
1015
1016         rhashtable_destroy(&ht);
1017
1018         return err;
1019 }
1020
1021 subsys_initcall(test_rht_init);
1022
1023 #endif /* CONFIG_TEST_RHASHTABLE */