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                (ht->p.max_shift && atomic_read(&ht->shift) < ht->p.max_shift);
 204 }
 205 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 206
 207 /**
 208  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 209  * @ht:         hash table
 210  * @new_size:   new table size
 211  */
 212 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 213 {
 214         /* Shrink table beneath 30% load */
 215         return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
 216                (atomic_read(&ht->shift) > ht->p.min_shift);
 217 }
 218 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 219
 220 static void hashtable_chain_unzip(const struct rhashtable *ht,
 221                                   const struct bucket_table *new_tbl,
 222                                   struct bucket_table *old_tbl,
 223                                   size_t old_hash)
 224 {
 225         struct rhash_head *he, *p, *next;
 226         spinlock_t *new_bucket_lock, *new_bucket_lock2 = NULL;
 227         unsigned int new_hash, new_hash2;
 228
 229         ASSERT_BUCKET_LOCK(old_tbl, old_hash);
 230
 231         /* Old bucket empty, no work needed. */
 232         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 233                                    old_hash);
 234         if (rht_is_a_nulls(p))
 235                 return;
 236
 237         new_hash = new_hash2 = head_hashfn(ht, new_tbl, p);
 238         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 239
 240         /* Advance the old bucket pointer one or more times until it
 241          * reaches a node that doesn't hash to the same bucket as the
 242          * previous node p. Call the previous node p;
 243          */
 244         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 245                 new_hash2 = head_hashfn(ht, new_tbl, he);
 246                 if (new_hash != new_hash2)
 247                         break;
 248                 p = he;
 249         }
 250         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 251
 252         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 253
 254         /* If we have encountered an entry that maps to a different bucket in
 255          * the new table, lock down that bucket as well as we might cut off
 256          * the end of the chain.
 257          */
 258         new_bucket_lock2 = bucket_lock(new_tbl, new_hash);
 259         if (new_bucket_lock != new_bucket_lock2)
 260                 spin_lock_bh_nested(new_bucket_lock2, RHT_LOCK_NESTED2);
 261
 262         /* Find the subsequent node which does hash to the same
 263          * bucket as node P, or NULL if no such node exists.
 264          */
 265         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 266         if (!rht_is_a_nulls(he)) {
 267                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 268                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 269                                 next = he;
 270                                 break;
 271                         }
 272                 }
 273         }
 274
 275         /* Set p's next pointer to that subsequent node pointer,
 276          * bypassing the nodes which do not hash to p's bucket
 277          */
 278         rcu_assign_pointer(p->next, next);
 279
 280         if (new_bucket_lock != new_bucket_lock2)
 281                 spin_unlock_bh(new_bucket_lock2);
 282         spin_unlock_bh(new_bucket_lock);
 283 }
 284
 285 static void link_old_to_new(struct bucket_table *new_tbl,
 286                             unsigned int new_hash, struct rhash_head *entry)
 287 {
 288         spinlock_t *new_bucket_lock;
 289
 290         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 291
 292         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 293         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 294         spin_unlock_bh(new_bucket_lock);
 295 }
 296
 297 /**
 298  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 299  * @ht:         the hash table to expand
 300  *
 301  * A secondary bucket array is allocated and the hash entries are migrated
 302  * while keeping them on both lists until the end of the RCU grace period.
 303  *
 304  * This function may only be called in a context where it is safe to call
 305  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 306  *
 307  * The caller must ensure that no concurrent resizing occurs by holding
 308  * ht->mutex.
 309  *
 310  * It is valid to have concurrent insertions and deletions protected by per
 311  * bucket locks or concurrent RCU protected lookups and traversals.
 312  */
 313 int rhashtable_expand(struct rhashtable *ht)
 314 {
 315         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 316         struct rhash_head *he;
 317         spinlock_t *old_bucket_lock;
 318         unsigned int new_hash, old_hash;
 319         bool complete = false;
 320
 321         ASSERT_RHT_MUTEX(ht);
 322
 323         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 324         if (new_tbl == NULL)
 325                 return -ENOMEM;
 326
 327         atomic_inc(&ht->shift);
 328
 329         /* Make insertions go into the new, empty table right away. Deletions
 330          * and lookups will be attempted in both tables until we synchronize.
 331          * The synchronize_rcu() guarantees for the new table to be picked up
 332          * so no new additions go into the old table while we relink.
 333          */
 334         rcu_assign_pointer(ht->future_tbl, new_tbl);
 335         synchronize_rcu();
 336
 337         /* For each new bucket, search the corresponding old bucket for the
 338          * first entry that hashes to the new bucket, and link the end of
 339          * newly formed bucket chain (containing entries added to future
 340          * table) to that entry. Since all the entries which will end up in
 341          * the new bucket appear in the same old bucket, this constructs an
 342          * entirely valid new hash table, but with multiple buckets
 343          * "zipped" together into a single imprecise chain.
 344          */
 345         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 346                 old_hash = rht_bucket_index(old_tbl, new_hash);
 347                 old_bucket_lock = bucket_lock(old_tbl, old_hash);
 348
 349                 spin_lock_bh(old_bucket_lock);
 350                 rht_for_each(he, old_tbl, old_hash) {
 351                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 352                                 link_old_to_new(new_tbl, new_hash, he);
 353                                 break;
 354                         }
 355                 }
 356                 spin_unlock_bh(old_bucket_lock);
 357         }
 358
 359         /* Publish the new table pointer. Lookups may now traverse
 360          * the new table, but they will not benefit from any
 361          * additional efficiency until later steps unzip the buckets.
 362          */
 363         rcu_assign_pointer(ht->tbl, new_tbl);
 364
 365         /* Unzip interleaved hash chains */
 366         while (!complete && !ht->being_destroyed) {
 367                 /* Wait for readers. All new readers will see the new
 368                  * table, and thus no references to the old table will
 369                  * remain.
 370                  */
 371                 synchronize_rcu();
 372
 373                 /* For each bucket in the old table (each of which
 374                  * contains items from multiple buckets of the new
 375                  * table): ...
 376                  */
 377                 complete = true;
 378                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 379                         struct rhash_head *head;
 380
 381                         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 382                         spin_lock_bh(old_bucket_lock);
 383
 384                         hashtable_chain_unzip(ht, new_tbl, old_tbl, old_hash);
 385                         head = rht_dereference_bucket(old_tbl->buckets[old_hash],
 386                                                       old_tbl, old_hash);
 387                         if (!rht_is_a_nulls(head))
 388                                 complete = false;
 389
 390                         spin_unlock_bh(old_bucket_lock);
 391                 }
 392         }
 393
 394         bucket_table_free(old_tbl);
 395         return 0;
 396 }
 397 EXPORT_SYMBOL_GPL(rhashtable_expand);
 398
 399 /**
 400  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 401  * @ht:         the hash table to shrink
 402  *
 403  * This function may only be called in a context where it is safe to call
 404  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 405  *
 406  * The caller must ensure that no concurrent resizing occurs by holding
 407  * ht->mutex.
 408  *
 409  * The caller must ensure that no concurrent table mutations take place.
 410  * It is however valid to have concurrent lookups if they are RCU protected.
 411  *
 412  * It is valid to have concurrent insertions and deletions protected by per
 413  * bucket locks or concurrent RCU protected lookups and traversals.
 414  */
 415 int rhashtable_shrink(struct rhashtable *ht)
 416 {
 417         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 418         spinlock_t *new_bucket_lock, *old_bucket_lock1, *old_bucket_lock2;
 419         unsigned int new_hash;
 420
 421         ASSERT_RHT_MUTEX(ht);
 422
 423         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 424         if (new_tbl == NULL)
 425                 return -ENOMEM;
 426
 427         rcu_assign_pointer(ht->future_tbl, new_tbl);
 428         synchronize_rcu();
 429
 430         /* Link the first entry in the old bucket to the end of the
 431          * bucket in the new table. As entries are concurrently being
 432          * added to the new table, lock down the new bucket. As we
 433          * always divide the size in half when shrinking, each bucket
 434          * in the new table maps to exactly two buckets in the old
 435          * table.
 436          *
 437          * As removals can occur concurrently on the old table, we need
 438          * to lock down both matching buckets in the old table.
 439          */
 440         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 441                 old_bucket_lock1 = bucket_lock(tbl, new_hash);
 442                 old_bucket_lock2 = bucket_lock(tbl, new_hash + new_tbl->size);
 443                 new_bucket_lock = bucket_lock(new_tbl, new_hash);
 444
 445                 spin_lock_bh(old_bucket_lock1);
 446                 spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED);
 447                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2);
 448
 449                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 450                                    tbl->buckets[new_hash]);
 451                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 452                                    tbl->buckets[new_hash + new_tbl->size]);
 453
 454                 spin_unlock_bh(new_bucket_lock);
 455                 spin_unlock_bh(old_bucket_lock2);
 456                 spin_unlock_bh(old_bucket_lock1);
 457         }
 458
 459         /* Publish the new, valid hash table */
 460         rcu_assign_pointer(ht->tbl, new_tbl);
 461         atomic_dec(&ht->shift);
 462
 463         /* Wait for readers. No new readers will have references to the
 464          * old hash table.
 465          */
 466         synchronize_rcu();
 467
 468         bucket_table_free(tbl);
 469
 470         return 0;
 471 }
 472 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 473
 474 static void rht_deferred_worker(struct work_struct *work)
 475 {
 476         struct rhashtable *ht;
 477         struct bucket_table *tbl;
 478
 479         ht = container_of(work, struct rhashtable, run_work.work);
 480         mutex_lock(&ht->mutex);
 481         tbl = rht_dereference(ht->tbl, ht);
 482
 483         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 484                 rhashtable_expand(ht);
 485         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 486                 rhashtable_shrink(ht);
 487
 488         mutex_unlock(&ht->mutex);
 489 }
 490
 491 static void rhashtable_wakeup_worker(struct rhashtable *ht)
 492 {
 493         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 494         struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 495         size_t size = tbl->size;
 496
 497         /* Only adjust the table if no resizing is currently in progress. */
 498         if (tbl == new_tbl &&
 499             ((ht->p.grow_decision && ht->p.grow_decision(ht, size)) ||
 500              (ht->p.shrink_decision && ht->p.shrink_decision(ht, size))))
 501                 schedule_delayed_work(&ht->run_work, 0);
 502 }
 503
 504 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 505                                 struct bucket_table *tbl, u32 hash)
 506 {
 507         struct rhash_head *head = rht_dereference_bucket(tbl->buckets[hash],
 508                                                          tbl, hash);
 509
 510         if (rht_is_a_nulls(head))
 511                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 512         else
 513                 RCU_INIT_POINTER(obj->next, head);
 514
 515         rcu_assign_pointer(tbl->buckets[hash], obj);
 516
 517         atomic_inc(&ht->nelems);
 518
 519         rhashtable_wakeup_worker(ht);
 520 }
 521
 522 /**
 523  * rhashtable_insert - insert object into hash table
 524  * @ht:         hash table
 525  * @obj:        pointer to hash head inside object
 526  *
 527  * Will take a per bucket spinlock to protect against mutual mutations
 528  * on the same bucket. Multiple insertions may occur in parallel unless
 529  * they map to the same bucket lock.
 530  *
 531  * It is safe to call this function from atomic context.
 532  *
 533  * Will trigger an automatic deferred table resizing if the size grows
 534  * beyond the watermark indicated by grow_decision() which can be passed
 535  * to rhashtable_init().
 536  */
 537 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 538 {
 539         struct bucket_table *tbl;
 540         spinlock_t *lock;
 541         unsigned hash;
 542
 543         rcu_read_lock();
 544
 545         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 546         hash = head_hashfn(ht, tbl, obj);
 547         lock = bucket_lock(tbl, hash);
 548
 549         spin_lock_bh(lock);
 550         __rhashtable_insert(ht, obj, tbl, hash);
 551         spin_unlock_bh(lock);
 552
 553         rcu_read_unlock();
 554 }
 555 EXPORT_SYMBOL_GPL(rhashtable_insert);
 556
 557 /**
 558  * rhashtable_remove - remove object from hash table
 559  * @ht:         hash table
 560  * @obj:        pointer to hash head inside object
 561  *
 562  * Since the hash chain is single linked, the removal operation needs to
 563  * walk the bucket chain upon removal. The removal operation is thus
 564  * considerable slow if the hash table is not correctly sized.
 565  *
 566  * Will automatically shrink the table via rhashtable_expand() if the
 567  * shrink_decision function specified at rhashtable_init() returns true.
 568  *
 569  * The caller must ensure that no concurrent table mutations occur. It is
 570  * however valid to have concurrent lookups if they are RCU protected.
 571  */
 572 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 573 {
 574         struct bucket_table *tbl;
 575         struct rhash_head __rcu **pprev;
 576         struct rhash_head *he;
 577         spinlock_t *lock;
 578         unsigned int hash;
 579
 580         rcu_read_lock();
 581         tbl = rht_dereference_rcu(ht->tbl, ht);
 582         hash = head_hashfn(ht, tbl, obj);
 583
 584         lock = bucket_lock(tbl, hash);
 585         spin_lock_bh(lock);
 586
 587 restart:
 588         pprev = &tbl->buckets[hash];
 589         rht_for_each(he, tbl, hash) {
 590                 if (he != obj) {
 591                         pprev = &he->next;
 592                         continue;
 593                 }
 594
 595                 rcu_assign_pointer(*pprev, obj->next);
 596                 atomic_dec(&ht->nelems);
 597
 598                 spin_unlock_bh(lock);
 599
 600                 rhashtable_wakeup_worker(ht);
 601
 602                 rcu_read_unlock();
 603
 604                 return true;
 605         }
 606
 607         if (tbl != rht_dereference_rcu(ht->future_tbl, ht)) {
 608                 spin_unlock_bh(lock);
 609
 610                 tbl = rht_dereference_rcu(ht->future_tbl, ht);
 611                 hash = head_hashfn(ht, tbl, obj);
 612
 613                 lock = bucket_lock(tbl, hash);
 614                 spin_lock_bh(lock);
 615                 goto restart;
 616         }
 617
 618         spin_unlock_bh(lock);
 619         rcu_read_unlock();
 620
 621         return false;
 622 }
 623 EXPORT_SYMBOL_GPL(rhashtable_remove);
 624
 625 struct rhashtable_compare_arg {
 626         struct rhashtable *ht;
 627         const void *key;
 628 };
 629
 630 static bool rhashtable_compare(void *ptr, void *arg)
 631 {
 632         struct rhashtable_compare_arg *x = arg;
 633         struct rhashtable *ht = x->ht;
 634
 635         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 636 }
 637
 638 /**
 639  * rhashtable_lookup - lookup key in hash table
 640  * @ht:         hash table
 641  * @key:        pointer to key
 642  *
 643  * Computes the hash value for the key and traverses the bucket chain looking
 644  * for a entry with an identical key. The first matching entry is returned.
 645  *
 646  * This lookup function may only be used for fixed key hash table (key_len
 647  * parameter set). It will BUG() if used inappropriately.
 648  *
 649  * Lookups may occur in parallel with hashtable mutations and resizing.
 650  */
 651 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 652 {
 653         struct rhashtable_compare_arg arg = {
 654                 .ht = ht,
 655                 .key = key,
 656         };
 657
 658         BUG_ON(!ht->p.key_len);
 659
 660         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 661 }
 662 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 663
 664 /**
 665  * rhashtable_lookup_compare - search hash table with compare function
 666  * @ht:         hash table
 667  * @key:        the pointer to the key
 668  * @compare:    compare function, must return true on match
 669  * @arg:        argument passed on to compare function
 670  *
 671  * Traverses the bucket chain behind the provided hash value and calls the
 672  * specified compare function for each entry.
 673  *
 674  * Lookups may occur in parallel with hashtable mutations and resizing.
 675  *
 676  * Returns the first entry on which the compare function returned true.
 677  */
 678 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 679                                 bool (*compare)(void *, void *), void *arg)
 680 {
 681         const struct bucket_table *tbl, *old_tbl;
 682         struct rhash_head *he;
 683         u32 hash;
 684
 685         rcu_read_lock();
 686
 687         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 688         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 689         hash = key_hashfn(ht, key, ht->p.key_len);
 690 restart:
 691         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 692                 if (!compare(rht_obj(ht, he), arg))
 693                         continue;
 694                 rcu_read_unlock();
 695                 return rht_obj(ht, he);
 696         }
 697
 698         if (unlikely(tbl != old_tbl)) {
 699                 tbl = old_tbl;
 700                 goto restart;
 701         }
 702         rcu_read_unlock();
 703
 704         return NULL;
 705 }
 706 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 707
 708 /**
 709  * rhashtable_lookup_insert - lookup and insert object into hash table
 710  * @ht:         hash table
 711  * @obj:        pointer to hash head inside object
 712  *
 713  * Locks down the bucket chain in both the old and new table if a resize
 714  * is in progress to ensure that writers can't remove from the old table
 715  * and can't insert to the new table during the atomic operation of search
 716  * and insertion. Searches for duplicates in both the old and new table if
 717  * a resize is in progress.
 718  *
 719  * This lookup function may only be used for fixed key hash table (key_len
 720  * parameter set). It will BUG() if used inappropriately.
 721  *
 722  * It is safe to call this function from atomic context.
 723  *
 724  * Will trigger an automatic deferred table resizing if the size grows
 725  * beyond the watermark indicated by grow_decision() which can be passed
 726  * to rhashtable_init().
 727  */
 728 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 729 {
 730         struct bucket_table *new_tbl, *old_tbl;
 731         spinlock_t *new_bucket_lock, *old_bucket_lock;
 732         u32 new_hash, old_hash;
 733         bool success = true;
 734
 735         BUG_ON(!ht->p.key_len);
 736
 737         rcu_read_lock();
 738
 739         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 740         old_hash = head_hashfn(ht, old_tbl, obj);
 741         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 742         spin_lock_bh(old_bucket_lock);
 743
 744         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 745         new_hash = head_hashfn(ht, new_tbl, obj);
 746         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 747         if (unlikely(old_tbl != new_tbl))
 748                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 749
 750         if (rhashtable_lookup(ht, rht_obj(ht, obj) + ht->p.key_offset)) {
 751                 success = false;
 752                 goto exit;
 753         }
 754
 755         __rhashtable_insert(ht, obj, new_tbl, new_hash);
 756
 757 exit:
 758         if (unlikely(old_tbl != new_tbl))
 759                 spin_unlock_bh(new_bucket_lock);
 760         spin_unlock_bh(old_bucket_lock);
 761
 762         rcu_read_unlock();
 763
 764         return success;
 765 }
 766 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 767
 768 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 769 {
 770         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 771                    1UL << params->min_shift);
 772 }
 773
 774 /**
 775  * rhashtable_init - initialize a new hash table
 776  * @ht:         hash table to be initialized
 777  * @params:     configuration parameters
 778  *
 779  * Initializes a new hash table based on the provided configuration
 780  * parameters. A table can be configured either with a variable or
 781  * fixed length key:
 782  *
 783  * Configuration Example 1: Fixed length keys
 784  * struct test_obj {
 785  *      int                     key;
 786  *      void *                  my_member;
 787  *      struct rhash_head       node;
 788  * };
 789  *
 790  * struct rhashtable_params params = {
 791  *      .head_offset = offsetof(struct test_obj, node),
 792  *      .key_offset = offsetof(struct test_obj, key),
 793  *      .key_len = sizeof(int),
 794  *      .hashfn = jhash,
 795  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 796  * };
 797  *
 798  * Configuration Example 2: Variable length keys
 799  * struct test_obj {
 800  *      [...]
 801  *      struct rhash_head       node;
 802  * };
 803  *
 804  * u32 my_hash_fn(const void *data, u32 seed)
 805  * {
 806  *      struct test_obj *obj = data;
 807  *
 808  *      return [... hash ...];
 809  * }
 810  *
 811  * struct rhashtable_params params = {
 812  *      .head_offset = offsetof(struct test_obj, node),
 813  *      .hashfn = jhash,
 814  *      .obj_hashfn = my_hash_fn,
 815  * };
 816  */
 817 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
 818 {
 819         struct bucket_table *tbl;
 820         size_t size;
 821
 822         size = HASH_DEFAULT_SIZE;
 823
 824         if ((params->key_len && !params->hashfn) ||
 825             (!params->key_len && !params->obj_hashfn))
 826                 return -EINVAL;
 827
 828         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 829                 return -EINVAL;
 830
 831         params->min_shift = max_t(size_t, params->min_shift,
 832                                   ilog2(HASH_MIN_SIZE));
 833
 834         if (params->nelem_hint)
 835                 size = rounded_hashtable_size(params);
 836
 837         memset(ht, 0, sizeof(*ht));
 838         mutex_init(&ht->mutex);
 839         memcpy(&ht->p, params, sizeof(*params));
 840
 841         if (params->locks_mul)
 842                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 843         else
 844                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 845
 846         tbl = bucket_table_alloc(ht, size);
 847         if (tbl == NULL)
 848                 return -ENOMEM;
 849
 850         atomic_set(&ht->nelems, 0);
 851         atomic_set(&ht->shift, ilog2(tbl->size));
 852         RCU_INIT_POINTER(ht->tbl, tbl);
 853         RCU_INIT_POINTER(ht->future_tbl, tbl);
 854
 855         if (!ht->p.hash_rnd)
 856                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
 857
 858         if (ht->p.grow_decision || ht->p.shrink_decision)
 859                 INIT_DEFERRABLE_WORK(&ht->run_work, rht_deferred_worker);
 860
 861         return 0;
 862 }
 863 EXPORT_SYMBOL_GPL(rhashtable_init);
 864
 865 /**
 866  * rhashtable_destroy - destroy hash table
 867  * @ht:         the hash table to destroy
 868  *
 869  * Frees the bucket array. This function is not rcu safe, therefore the caller
 870  * has to make sure that no resizing may happen by unpublishing the hashtable
 871  * and waiting for the quiescent cycle before releasing the bucket array.
 872  */
 873 void rhashtable_destroy(struct rhashtable *ht)
 874 {
 875         ht->being_destroyed = true;
 876
 877         mutex_lock(&ht->mutex);
 878
 879         cancel_delayed_work(&ht->run_work);
 880         bucket_table_free(rht_dereference(ht->tbl, ht));
 881
 882         mutex_unlock(&ht->mutex);
 883 }
 884 EXPORT_SYMBOL_GPL(rhashtable_destroy);
 885
 886 /**************************************************************************
 887  * Self Test
 888  **************************************************************************/
 889
 890 #ifdef CONFIG_TEST_RHASHTABLE
 891
 892 #define TEST_HT_SIZE    8
 893 #define TEST_ENTRIES    2048
 894 #define TEST_PTR        ((void *) 0xdeadbeef)
 895 #define TEST_NEXPANDS   4
 896
 897 struct test_obj {
 898         void                    *ptr;
 899         int                     value;
 900         struct rhash_head       node;
 901 };
 902
 903 static int __init test_rht_lookup(struct rhashtable *ht)
 904 {
 905         unsigned int i;
 906
 907         for (i = 0; i < TEST_ENTRIES * 2; i++) {
 908                 struct test_obj *obj;
 909                 bool expected = !(i % 2);
 910                 u32 key = i;
 911
 912                 obj = rhashtable_lookup(ht, &key);
 913
 914                 if (expected && !obj) {
 915                         pr_warn("Test failed: Could not find key %u\n", key);
 916                         return -ENOENT;
 917                 } else if (!expected && obj) {
 918                         pr_warn("Test failed: Unexpected entry found for key %u\n",
 919                                 key);
 920                         return -EEXIST;
 921                 } else if (expected && obj) {
 922                         if (obj->ptr != TEST_PTR || obj->value != i) {
 923                                 pr_warn("Test failed: Lookup value mismatch %p!=%p, %u!=%u\n",
 924                                         obj->ptr, TEST_PTR, obj->value, i);
 925                                 return -EINVAL;
 926                         }
 927                 }
 928         }
 929
 930         return 0;
 931 }
 932
 933 static void test_bucket_stats(struct rhashtable *ht, bool quiet)
 934 {
 935         unsigned int cnt, rcu_cnt, i, total = 0;
 936         struct rhash_head *pos;
 937         struct test_obj *obj;
 938         struct bucket_table *tbl;
 939
 940         tbl = rht_dereference_rcu(ht->tbl, ht);
 941         for (i = 0; i < tbl->size; i++) {
 942                 rcu_cnt = cnt = 0;
 943
 944                 if (!quiet)
 945                         pr_info(" [%#4x/%zu]", i, tbl->size);
 946
 947                 rht_for_each_entry_rcu(obj, pos, tbl, i, node) {
 948                         cnt++;
 949                         total++;
 950                         if (!quiet)
 951                                 pr_cont(" [%p],", obj);
 952                 }
 953
 954                 rht_for_each_entry_rcu(obj, pos, tbl, i, node)
 955                         rcu_cnt++;
 956
 957                 if (rcu_cnt != cnt)
 958                         pr_warn("Test failed: Chain count mismach %d != %d",
 959                                 cnt, rcu_cnt);
 960
 961                 if (!quiet)
 962                         pr_cont("\n  [%#x] first element: %p, chain length: %u\n",
 963                                 i, tbl->buckets[i], cnt);
 964         }
 965
 966         pr_info("  Traversal complete: counted=%u, nelems=%u, entries=%d\n",
 967                 total, atomic_read(&ht->nelems), TEST_ENTRIES);
 968
 969         if (total != atomic_read(&ht->nelems) || total != TEST_ENTRIES)
 970                 pr_warn("Test failed: Total count mismatch ^^^");
 971 }
 972
 973 static int __init test_rhashtable(struct rhashtable *ht)
 974 {
 975         struct bucket_table *tbl;
 976         struct test_obj *obj;
 977         struct rhash_head *pos, *next;
 978         int err;
 979         unsigned int i;
 980
 981         /*
 982          * Insertion Test:
 983          * Insert TEST_ENTRIES into table with all keys even numbers
 984          */
 985         pr_info("  Adding %d keys\n", TEST_ENTRIES);
 986         for (i = 0; i < TEST_ENTRIES; i++) {
 987                 struct test_obj *obj;
 988
 989                 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
 990                 if (!obj) {
 991                         err = -ENOMEM;
 992                         goto error;
 993                 }
 994
 995                 obj->ptr = TEST_PTR;
 996                 obj->value = i * 2;
 997
 998                 rhashtable_insert(ht, &obj->node);
 999         }
1000
1001         rcu_read_lock();
1002         test_bucket_stats(ht, true);
1003         test_rht_lookup(ht);
1004         rcu_read_unlock();
1005
1006         for (i = 0; i < TEST_NEXPANDS; i++) {
1007                 pr_info("  Table expansion iteration %u...\n", i);
1008                 mutex_lock(&ht->mutex);
1009                 rhashtable_expand(ht);
1010                 mutex_unlock(&ht->mutex);
1011
1012                 rcu_read_lock();
1013                 pr_info("  Verifying lookups...\n");
1014                 test_rht_lookup(ht);
1015                 rcu_read_unlock();
1016         }
1017
1018         for (i = 0; i < TEST_NEXPANDS; i++) {
1019                 pr_info("  Table shrinkage iteration %u...\n", i);
1020                 mutex_lock(&ht->mutex);
1021                 rhashtable_shrink(ht);
1022                 mutex_unlock(&ht->mutex);
1023
1024                 rcu_read_lock();
1025                 pr_info("  Verifying lookups...\n");
1026                 test_rht_lookup(ht);
1027                 rcu_read_unlock();
1028         }
1029
1030         rcu_read_lock();
1031         test_bucket_stats(ht, true);
1032         rcu_read_unlock();
1033
1034         pr_info("  Deleting %d keys\n", TEST_ENTRIES);
1035         for (i = 0; i < TEST_ENTRIES; i++) {
1036                 u32 key = i * 2;
1037
1038                 obj = rhashtable_lookup(ht, &key);
1039                 BUG_ON(!obj);
1040
1041                 rhashtable_remove(ht, &obj->node);
1042                 kfree(obj);
1043         }
1044
1045         return 0;
1046
1047 error:
1048         tbl = rht_dereference_rcu(ht->tbl, ht);
1049         for (i = 0; i < tbl->size; i++)
1050                 rht_for_each_entry_safe(obj, pos, next, tbl, i, node)
1051                         kfree(obj);
1052
1053         return err;
1054 }
1055
1056 static int __init test_rht_init(void)
1057 {
1058         struct rhashtable ht;
1059         struct rhashtable_params params = {
1060                 .nelem_hint = TEST_HT_SIZE,
1061                 .head_offset = offsetof(struct test_obj, node),
1062                 .key_offset = offsetof(struct test_obj, value),
1063                 .key_len = sizeof(int),
1064                 .hashfn = jhash,
1065                 .nulls_base = (3U << RHT_BASE_SHIFT),
1066                 .grow_decision = rht_grow_above_75,
1067                 .shrink_decision = rht_shrink_below_30,
1068         };
1069         int err;
1070
1071         pr_info("Running resizable hashtable tests...\n");
1072
1073         err = rhashtable_init(&ht, &params);
1074         if (err < 0) {
1075                 pr_warn("Test failed: Unable to initialize hashtable: %d\n",
1076                         err);
1077                 return err;
1078         }
1079
1080         err = test_rhashtable(&ht);
1081
1082         rhashtable_destroy(&ht);
1083
1084         return err;
1085 }
1086
1087 subsys_initcall(test_rht_init);
1088
1089 #endif /* CONFIG_TEST_RHASHTABLE */