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         return ht->p.hashfn(key, len, ht->p.hash_rnd) >> HASH_RESERVED_SPACE;
  98 }
  99
 100 static u32 head_hashfn(const struct rhashtable *ht,
 101                        const struct bucket_table *tbl,
 102                        const struct rhash_head *he)
 103 {
 104         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
 105 }
 106
 107 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 108 {
 109         struct rhash_head __rcu **pprev;
 110
 111         for (pprev = &tbl->buckets[n];
 112              !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n));
 113              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 114                 ;
 115
 116         return pprev;
 117 }
 118
 119 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 120 {
 121         unsigned int i, size;
 122 #if defined(CONFIG_PROVE_LOCKING)
 123         unsigned int nr_pcpus = 2;
 124 #else
 125         unsigned int nr_pcpus = num_possible_cpus();
 126 #endif
 127
 128         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 129         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 130
 131         /* Never allocate more than one lock per bucket */
 132         size = min_t(unsigned int, size, tbl->size);
 133
 134         if (sizeof(spinlock_t) != 0) {
 135 #ifdef CONFIG_NUMA
 136                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 137                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 138                 else
 139 #endif
 140                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 141                                            GFP_KERNEL);
 142                 if (!tbl->locks)
 143                         return -ENOMEM;
 144                 for (i = 0; i < size; i++)
 145                         spin_lock_init(&tbl->locks[i]);
 146         }
 147         tbl->locks_mask = size - 1;
 148
 149         return 0;
 150 }
 151
 152 static void bucket_table_free(const struct bucket_table *tbl)
 153 {
 154         if (tbl)
 155                 kvfree(tbl->locks);
 156
 157         kvfree(tbl);
 158 }
 159
 160 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 161                                                size_t nbuckets)
 162 {
 163         struct bucket_table *tbl;
 164         size_t size;
 165         int i;
 166
 167         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 168         tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
 169         if (tbl == NULL)
 170                 tbl = vzalloc(size);
 171
 172         if (tbl == NULL)
 173                 return NULL;
 174
 175         tbl->size = nbuckets;
 176
 177         if (alloc_bucket_locks(ht, tbl) < 0) {
 178                 bucket_table_free(tbl);
 179                 return NULL;
 180         }
 181
 182         for (i = 0; i < nbuckets; i++)
 183                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 184
 185         return tbl;
 186 }
 187
 188 /**
 189  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 190  * @ht:         hash table
 191  * @new_size:   new table size
 192  */
 193 bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 194 {
 195         /* Expand table when exceeding 75% load */
 196         return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
 197                (ht->p.max_shift && atomic_read(&ht->shift) < ht->p.max_shift);
 198 }
 199 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 200
 201 /**
 202  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 203  * @ht:         hash table
 204  * @new_size:   new table size
 205  */
 206 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 207 {
 208         /* Shrink table beneath 30% load */
 209         return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
 210                (atomic_read(&ht->shift) > ht->p.min_shift);
 211 }
 212 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 213
 214 static void hashtable_chain_unzip(const struct rhashtable *ht,
 215                                   const struct bucket_table *new_tbl,
 216                                   struct bucket_table *old_tbl,
 217                                   size_t old_hash)
 218 {
 219         struct rhash_head *he, *p, *next;
 220         spinlock_t *new_bucket_lock, *new_bucket_lock2 = NULL;
 221         unsigned int new_hash, new_hash2;
 222
 223         ASSERT_BUCKET_LOCK(old_tbl, old_hash);
 224
 225         /* Old bucket empty, no work needed. */
 226         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 227                                    old_hash);
 228         if (rht_is_a_nulls(p))
 229                 return;
 230
 231         new_hash = new_hash2 = head_hashfn(ht, new_tbl, p);
 232         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 233
 234         /* Advance the old bucket pointer one or more times until it
 235          * reaches a node that doesn't hash to the same bucket as the
 236          * previous node p. Call the previous node p;
 237          */
 238         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 239                 new_hash2 = head_hashfn(ht, new_tbl, he);
 240                 if (new_hash != new_hash2)
 241                         break;
 242                 p = he;
 243         }
 244         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 245
 246         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 247
 248         /* If we have encountered an entry that maps to a different bucket in
 249          * the new table, lock down that bucket as well as we might cut off
 250          * the end of the chain.
 251          */
 252         new_bucket_lock2 = bucket_lock(new_tbl, new_hash);
 253         if (new_bucket_lock != new_bucket_lock2)
 254                 spin_lock_bh_nested(new_bucket_lock2, RHT_LOCK_NESTED2);
 255
 256         /* Find the subsequent node which does hash to the same
 257          * bucket as node P, or NULL if no such node exists.
 258          */
 259         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 260         if (!rht_is_a_nulls(he)) {
 261                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 262                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 263                                 next = he;
 264                                 break;
 265                         }
 266                 }
 267         }
 268
 269         /* Set p's next pointer to that subsequent node pointer,
 270          * bypassing the nodes which do not hash to p's bucket
 271          */
 272         rcu_assign_pointer(p->next, next);
 273
 274         if (new_bucket_lock != new_bucket_lock2)
 275                 spin_unlock_bh(new_bucket_lock2);
 276         spin_unlock_bh(new_bucket_lock);
 277 }
 278
 279 static void link_old_to_new(struct bucket_table *new_tbl,
 280                             unsigned int new_hash, struct rhash_head *entry)
 281 {
 282         spinlock_t *new_bucket_lock;
 283
 284         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 285
 286         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 287         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 288         spin_unlock_bh(new_bucket_lock);
 289 }
 290
 291 /**
 292  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 293  * @ht:         the hash table to expand
 294  *
 295  * A secondary bucket array is allocated and the hash entries are migrated
 296  * while keeping them on both lists until the end of the RCU grace period.
 297  *
 298  * This function may only be called in a context where it is safe to call
 299  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 300  *
 301  * The caller must ensure that no concurrent resizing occurs by holding
 302  * ht->mutex.
 303  *
 304  * It is valid to have concurrent insertions and deletions protected by per
 305  * bucket locks or concurrent RCU protected lookups and traversals.
 306  */
 307 int rhashtable_expand(struct rhashtable *ht)
 308 {
 309         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 310         struct rhash_head *he;
 311         spinlock_t *old_bucket_lock;
 312         unsigned int new_hash, old_hash;
 313         bool complete = false;
 314
 315         ASSERT_RHT_MUTEX(ht);
 316
 317         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 318         if (new_tbl == NULL)
 319                 return -ENOMEM;
 320
 321         atomic_inc(&ht->shift);
 322
 323         /* Make insertions go into the new, empty table right away. Deletions
 324          * and lookups will be attempted in both tables until we synchronize.
 325          * The synchronize_rcu() guarantees for the new table to be picked up
 326          * so no new additions go into the old table while we relink.
 327          */
 328         rcu_assign_pointer(ht->future_tbl, new_tbl);
 329         synchronize_rcu();
 330
 331         /* For each new bucket, search the corresponding old bucket for the
 332          * first entry that hashes to the new bucket, and link the end of
 333          * newly formed bucket chain (containing entries added to future
 334          * table) to that entry. Since all the entries which will end up in
 335          * the new bucket appear in the same old bucket, this constructs an
 336          * entirely valid new hash table, but with multiple buckets
 337          * "zipped" together into a single imprecise chain.
 338          */
 339         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 340                 old_hash = rht_bucket_index(old_tbl, new_hash);
 341                 old_bucket_lock = bucket_lock(old_tbl, old_hash);
 342
 343                 spin_lock_bh(old_bucket_lock);
 344                 rht_for_each(he, old_tbl, old_hash) {
 345                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 346                                 link_old_to_new(new_tbl, new_hash, he);
 347                                 break;
 348                         }
 349                 }
 350                 spin_unlock_bh(old_bucket_lock);
 351         }
 352
 353         /* Publish the new table pointer. Lookups may now traverse
 354          * the new table, but they will not benefit from any
 355          * additional efficiency until later steps unzip the buckets.
 356          */
 357         rcu_assign_pointer(ht->tbl, new_tbl);
 358
 359         /* Unzip interleaved hash chains */
 360         while (!complete && !ht->being_destroyed) {
 361                 /* Wait for readers. All new readers will see the new
 362                  * table, and thus no references to the old table will
 363                  * remain.
 364                  */
 365                 synchronize_rcu();
 366
 367                 /* For each bucket in the old table (each of which
 368                  * contains items from multiple buckets of the new
 369                  * table): ...
 370                  */
 371                 complete = true;
 372                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 373                         struct rhash_head *head;
 374
 375                         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 376                         spin_lock_bh(old_bucket_lock);
 377
 378                         hashtable_chain_unzip(ht, new_tbl, old_tbl, old_hash);
 379                         head = rht_dereference_bucket(old_tbl->buckets[old_hash],
 380                                                       old_tbl, old_hash);
 381                         if (!rht_is_a_nulls(head))
 382                                 complete = false;
 383
 384                         spin_unlock_bh(old_bucket_lock);
 385                 }
 386         }
 387
 388         bucket_table_free(old_tbl);
 389         return 0;
 390 }
 391 EXPORT_SYMBOL_GPL(rhashtable_expand);
 392
 393 /**
 394  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 395  * @ht:         the hash table to shrink
 396  *
 397  * This function may only be called in a context where it is safe to call
 398  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 399  *
 400  * The caller must ensure that no concurrent resizing occurs by holding
 401  * ht->mutex.
 402  *
 403  * The caller must ensure that no concurrent table mutations take place.
 404  * It is however valid to have concurrent lookups if they are RCU protected.
 405  *
 406  * It is valid to have concurrent insertions and deletions protected by per
 407  * bucket locks or concurrent RCU protected lookups and traversals.
 408  */
 409 int rhashtable_shrink(struct rhashtable *ht)
 410 {
 411         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 412         spinlock_t *new_bucket_lock, *old_bucket_lock1, *old_bucket_lock2;
 413         unsigned int new_hash;
 414
 415         ASSERT_RHT_MUTEX(ht);
 416
 417         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 418         if (new_tbl == NULL)
 419                 return -ENOMEM;
 420
 421         rcu_assign_pointer(ht->future_tbl, new_tbl);
 422         synchronize_rcu();
 423
 424         /* Link the first entry in the old bucket to the end of the
 425          * bucket in the new table. As entries are concurrently being
 426          * added to the new table, lock down the new bucket. As we
 427          * always divide the size in half when shrinking, each bucket
 428          * in the new table maps to exactly two buckets in the old
 429          * table.
 430          *
 431          * As removals can occur concurrently on the old table, we need
 432          * to lock down both matching buckets in the old table.
 433          */
 434         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 435                 old_bucket_lock1 = bucket_lock(tbl, new_hash);
 436                 old_bucket_lock2 = bucket_lock(tbl, new_hash + new_tbl->size);
 437                 new_bucket_lock = bucket_lock(new_tbl, new_hash);
 438
 439                 spin_lock_bh(old_bucket_lock1);
 440
 441                 /* Depending on the lock per buckets mapping, the bucket in
 442                  * the lower and upper region may map to the same lock.
 443                  */
 444                 if (old_bucket_lock1 != old_bucket_lock2) {
 445                         spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED);
 446                         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2);
 447                 } else {
 448                         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 449                 }
 450
 451                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 452                                    tbl->buckets[new_hash]);
 453                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 454                                    tbl->buckets[new_hash + new_tbl->size]);
 455
 456                 spin_unlock_bh(new_bucket_lock);
 457                 if (old_bucket_lock1 != old_bucket_lock2)
 458                         spin_unlock_bh(old_bucket_lock2);
 459                 spin_unlock_bh(old_bucket_lock1);
 460         }
 461
 462         /* Publish the new, valid hash table */
 463         rcu_assign_pointer(ht->tbl, new_tbl);
 464         atomic_dec(&ht->shift);
 465
 466         /* Wait for readers. No new readers will have references to the
 467          * old hash table.
 468          */
 469         synchronize_rcu();
 470
 471         bucket_table_free(tbl);
 472
 473         return 0;
 474 }
 475 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 476
 477 static void rht_deferred_worker(struct work_struct *work)
 478 {
 479         struct rhashtable *ht;
 480         struct bucket_table *tbl;
 481         struct rhashtable_walker *walker;
 482
 483         ht = container_of(work, struct rhashtable, run_work);
 484         mutex_lock(&ht->mutex);
 485         if (ht->being_destroyed)
 486                 goto unlock;
 487
 488         tbl = rht_dereference(ht->tbl, ht);
 489
 490         list_for_each_entry(walker, &ht->walkers, list)
 491                 walker->resize = true;
 492
 493         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 494                 rhashtable_expand(ht);
 495         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 496                 rhashtable_shrink(ht);
 497
 498 unlock:
 499         mutex_unlock(&ht->mutex);
 500 }
 501
 502 static void rhashtable_wakeup_worker(struct rhashtable *ht)
 503 {
 504         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 505         struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 506         size_t size = tbl->size;
 507
 508         /* Only adjust the table if no resizing is currently in progress. */
 509         if (tbl == new_tbl &&
 510             ((ht->p.grow_decision && ht->p.grow_decision(ht, size)) ||
 511              (ht->p.shrink_decision && ht->p.shrink_decision(ht, size))))
 512                 schedule_work(&ht->run_work);
 513 }
 514
 515 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 516                                 struct bucket_table *tbl, u32 hash)
 517 {
 518         struct rhash_head *head = rht_dereference_bucket(tbl->buckets[hash],
 519                                                          tbl, hash);
 520
 521         if (rht_is_a_nulls(head))
 522                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 523         else
 524                 RCU_INIT_POINTER(obj->next, head);
 525
 526         rcu_assign_pointer(tbl->buckets[hash], obj);
 527
 528         atomic_inc(&ht->nelems);
 529
 530         rhashtable_wakeup_worker(ht);
 531 }
 532
 533 /**
 534  * rhashtable_insert - insert object into hash table
 535  * @ht:         hash table
 536  * @obj:        pointer to hash head inside object
 537  *
 538  * Will take a per bucket spinlock to protect against mutual mutations
 539  * on the same bucket. Multiple insertions may occur in parallel unless
 540  * they map to the same bucket lock.
 541  *
 542  * It is safe to call this function from atomic context.
 543  *
 544  * Will trigger an automatic deferred table resizing if the size grows
 545  * beyond the watermark indicated by grow_decision() which can be passed
 546  * to rhashtable_init().
 547  */
 548 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 549 {
 550         struct bucket_table *tbl;
 551         spinlock_t *lock;
 552         unsigned hash;
 553
 554         rcu_read_lock();
 555
 556         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 557         hash = head_hashfn(ht, tbl, obj);
 558         lock = bucket_lock(tbl, hash);
 559
 560         spin_lock_bh(lock);
 561         __rhashtable_insert(ht, obj, tbl, hash);
 562         spin_unlock_bh(lock);
 563
 564         rcu_read_unlock();
 565 }
 566 EXPORT_SYMBOL_GPL(rhashtable_insert);
 567
 568 /**
 569  * rhashtable_remove - remove object from hash table
 570  * @ht:         hash table
 571  * @obj:        pointer to hash head inside object
 572  *
 573  * Since the hash chain is single linked, the removal operation needs to
 574  * walk the bucket chain upon removal. The removal operation is thus
 575  * considerable slow if the hash table is not correctly sized.
 576  *
 577  * Will automatically shrink the table via rhashtable_expand() if the
 578  * shrink_decision function specified at rhashtable_init() returns true.
 579  *
 580  * The caller must ensure that no concurrent table mutations occur. It is
 581  * however valid to have concurrent lookups if they are RCU protected.
 582  */
 583 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 584 {
 585         struct bucket_table *tbl;
 586         struct rhash_head __rcu **pprev;
 587         struct rhash_head *he;
 588         spinlock_t *lock;
 589         unsigned int hash;
 590         bool ret = false;
 591
 592         rcu_read_lock();
 593         tbl = rht_dereference_rcu(ht->tbl, ht);
 594         hash = head_hashfn(ht, tbl, obj);
 595
 596         lock = bucket_lock(tbl, hash);
 597         spin_lock_bh(lock);
 598
 599 restart:
 600         pprev = &tbl->buckets[hash];
 601         rht_for_each(he, tbl, hash) {
 602                 if (he != obj) {
 603                         pprev = &he->next;
 604                         continue;
 605                 }
 606
 607                 rcu_assign_pointer(*pprev, obj->next);
 608
 609                 ret = true;
 610                 break;
 611         }
 612
 613         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 614          * 'future_tbl' only exists for a short period of time during
 615          * resizing. Thus traversing both is fine and the added cost is
 616          * very rare.
 617          */
 618         if (tbl != rht_dereference_rcu(ht->future_tbl, ht)) {
 619                 spin_unlock_bh(lock);
 620
 621                 tbl = rht_dereference_rcu(ht->future_tbl, ht);
 622                 hash = head_hashfn(ht, tbl, obj);
 623
 624                 lock = bucket_lock(tbl, hash);
 625                 spin_lock_bh(lock);
 626                 goto restart;
 627         }
 628
 629         spin_unlock_bh(lock);
 630
 631         if (ret) {
 632                 atomic_dec(&ht->nelems);
 633                 rhashtable_wakeup_worker(ht);
 634         }
 635
 636         rcu_read_unlock();
 637
 638         return ret;
 639 }
 640 EXPORT_SYMBOL_GPL(rhashtable_remove);
 641
 642 struct rhashtable_compare_arg {
 643         struct rhashtable *ht;
 644         const void *key;
 645 };
 646
 647 static bool rhashtable_compare(void *ptr, void *arg)
 648 {
 649         struct rhashtable_compare_arg *x = arg;
 650         struct rhashtable *ht = x->ht;
 651
 652         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 653 }
 654
 655 /**
 656  * rhashtable_lookup - lookup key in hash table
 657  * @ht:         hash table
 658  * @key:        pointer to key
 659  *
 660  * Computes the hash value for the key and traverses the bucket chain looking
 661  * for a entry with an identical key. The first matching entry is returned.
 662  *
 663  * This lookup function may only be used for fixed key hash table (key_len
 664  * parameter set). It will BUG() if used inappropriately.
 665  *
 666  * Lookups may occur in parallel with hashtable mutations and resizing.
 667  */
 668 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 669 {
 670         struct rhashtable_compare_arg arg = {
 671                 .ht = ht,
 672                 .key = key,
 673         };
 674
 675         BUG_ON(!ht->p.key_len);
 676
 677         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 678 }
 679 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 680
 681 /**
 682  * rhashtable_lookup_compare - search hash table with compare function
 683  * @ht:         hash table
 684  * @key:        the pointer to the key
 685  * @compare:    compare function, must return true on match
 686  * @arg:        argument passed on to compare function
 687  *
 688  * Traverses the bucket chain behind the provided hash value and calls the
 689  * specified compare function for each entry.
 690  *
 691  * Lookups may occur in parallel with hashtable mutations and resizing.
 692  *
 693  * Returns the first entry on which the compare function returned true.
 694  */
 695 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 696                                 bool (*compare)(void *, void *), void *arg)
 697 {
 698         const struct bucket_table *tbl, *old_tbl;
 699         struct rhash_head *he;
 700         u32 hash;
 701
 702         rcu_read_lock();
 703
 704         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 705         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 706         hash = key_hashfn(ht, key, ht->p.key_len);
 707 restart:
 708         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 709                 if (!compare(rht_obj(ht, he), arg))
 710                         continue;
 711                 rcu_read_unlock();
 712                 return rht_obj(ht, he);
 713         }
 714
 715         if (unlikely(tbl != old_tbl)) {
 716                 tbl = old_tbl;
 717                 goto restart;
 718         }
 719         rcu_read_unlock();
 720
 721         return NULL;
 722 }
 723 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 724
 725 /**
 726  * rhashtable_lookup_insert - lookup and insert object into hash table
 727  * @ht:         hash table
 728  * @obj:        pointer to hash head inside object
 729  *
 730  * Locks down the bucket chain in both the old and new table if a resize
 731  * is in progress to ensure that writers can't remove from the old table
 732  * and can't insert to the new table during the atomic operation of search
 733  * and insertion. Searches for duplicates in both the old and new table if
 734  * a resize is in progress.
 735  *
 736  * This lookup function may only be used for fixed key hash table (key_len
 737  * parameter set). It will BUG() if used inappropriately.
 738  *
 739  * It is safe to call this function from atomic context.
 740  *
 741  * Will trigger an automatic deferred table resizing if the size grows
 742  * beyond the watermark indicated by grow_decision() which can be passed
 743  * to rhashtable_init().
 744  */
 745 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 746 {
 747         struct rhashtable_compare_arg arg = {
 748                 .ht = ht,
 749                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 750         };
 751
 752         BUG_ON(!ht->p.key_len);
 753
 754         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 755                                                 &arg);
 756 }
 757 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 758
 759 /**
 760  * rhashtable_lookup_compare_insert - search and insert object to hash table
 761  *                                    with compare function
 762  * @ht:         hash table
 763  * @obj:        pointer to hash head inside object
 764  * @compare:    compare function, must return true on match
 765  * @arg:        argument passed on to compare function
 766  *
 767  * Locks down the bucket chain in both the old and new table if a resize
 768  * is in progress to ensure that writers can't remove from the old table
 769  * and can't insert to the new table during the atomic operation of search
 770  * and insertion. Searches for duplicates in both the old and new table if
 771  * a resize is in progress.
 772  *
 773  * Lookups may occur in parallel with hashtable mutations and resizing.
 774  *
 775  * Will trigger an automatic deferred table resizing if the size grows
 776  * beyond the watermark indicated by grow_decision() which can be passed
 777  * to rhashtable_init().
 778  */
 779 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 780                                       struct rhash_head *obj,
 781                                       bool (*compare)(void *, void *),
 782                                       void *arg)
 783 {
 784         struct bucket_table *new_tbl, *old_tbl;
 785         spinlock_t *new_bucket_lock, *old_bucket_lock;
 786         u32 new_hash, old_hash;
 787         bool success = true;
 788
 789         BUG_ON(!ht->p.key_len);
 790
 791         rcu_read_lock();
 792
 793         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 794         old_hash = head_hashfn(ht, old_tbl, obj);
 795         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 796         spin_lock_bh(old_bucket_lock);
 797
 798         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 799         new_hash = head_hashfn(ht, new_tbl, obj);
 800         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 801         if (unlikely(old_tbl != new_tbl))
 802                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 803
 804         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 805                                       compare, arg)) {
 806                 success = false;
 807                 goto exit;
 808         }
 809
 810         __rhashtable_insert(ht, obj, new_tbl, new_hash);
 811
 812 exit:
 813         if (unlikely(old_tbl != new_tbl))
 814                 spin_unlock_bh(new_bucket_lock);
 815         spin_unlock_bh(old_bucket_lock);
 816
 817         rcu_read_unlock();
 818
 819         return success;
 820 }
 821 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 822
 823 /**
 824  * rhashtable_walk_init - Initialise an iterator
 825  * @ht:         Table to walk over
 826  * @iter:       Hash table Iterator
 827  *
 828  * This function prepares a hash table walk.
 829  *
 830  * Note that if you restart a walk after rhashtable_walk_stop you
 831  * may see the same object twice.  Also, you may miss objects if
 832  * there are removals in between rhashtable_walk_stop and the next
 833  * call to rhashtable_walk_start.
 834  *
 835  * For a completely stable walk you should construct your own data
 836  * structure outside the hash table.
 837  *
 838  * This function may sleep so you must not call it from interrupt
 839  * context or with spin locks held.
 840  *
 841  * You must call rhashtable_walk_exit if this function returns
 842  * successfully.
 843  */
 844 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 845 {
 846         iter->ht = ht;
 847         iter->p = NULL;
 848         iter->slot = 0;
 849         iter->skip = 0;
 850
 851         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 852         if (!iter->walker)
 853                 return -ENOMEM;
 854
 855         mutex_lock(&ht->mutex);
 856         list_add(&iter->walker->list, &ht->walkers);
 857         mutex_unlock(&ht->mutex);
 858
 859         return 0;
 860 }
 861 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 862
 863 /**
 864  * rhashtable_walk_exit - Free an iterator
 865  * @iter:       Hash table Iterator
 866  *
 867  * This function frees resources allocated by rhashtable_walk_init.
 868  */
 869 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 870 {
 871         mutex_lock(&iter->ht->mutex);
 872         list_del(&iter->walker->list);
 873         mutex_unlock(&iter->ht->mutex);
 874         kfree(iter->walker);
 875 }
 876 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 877
 878 /**
 879  * rhashtable_walk_start - Start a hash table walk
 880  * @iter:       Hash table iterator
 881  *
 882  * Start a hash table walk.  Note that we take the RCU lock in all
 883  * cases including when we return an error.  So you must always call
 884  * rhashtable_walk_stop to clean up.
 885  *
 886  * Returns zero if successful.
 887  *
 888  * Returns -EAGAIN if resize event occured.  Note that the iterator
 889  * will rewind back to the beginning and you may use it immediately
 890  * by calling rhashtable_walk_next.
 891  */
 892 int rhashtable_walk_start(struct rhashtable_iter *iter)
 893 {
 894         rcu_read_lock();
 895
 896         if (iter->walker->resize) {
 897                 iter->slot = 0;
 898                 iter->skip = 0;
 899                 iter->walker->resize = false;
 900                 return -EAGAIN;
 901         }
 902
 903         return 0;
 904 }
 905 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 906
 907 /**
 908  * rhashtable_walk_next - Return the next object and advance the iterator
 909  * @iter:       Hash table iterator
 910  *
 911  * Note that you must call rhashtable_walk_stop when you are finished
 912  * with the walk.
 913  *
 914  * Returns the next object or NULL when the end of the table is reached.
 915  *
 916  * Returns -EAGAIN if resize event occured.  Note that the iterator
 917  * will rewind back to the beginning and you may continue to use it.
 918  */
 919 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 920 {
 921         const struct bucket_table *tbl;
 922         struct rhashtable *ht = iter->ht;
 923         struct rhash_head *p = iter->p;
 924         void *obj = NULL;
 925
 926         tbl = rht_dereference_rcu(ht->tbl, ht);
 927
 928         if (p) {
 929                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 930                 goto next;
 931         }
 932
 933         for (; iter->slot < tbl->size; iter->slot++) {
 934                 int skip = iter->skip;
 935
 936                 rht_for_each_rcu(p, tbl, iter->slot) {
 937                         if (!skip)
 938                                 break;
 939                         skip--;
 940                 }
 941
 942 next:
 943                 if (!rht_is_a_nulls(p)) {
 944                         iter->skip++;
 945                         iter->p = p;
 946                         obj = rht_obj(ht, p);
 947                         goto out;
 948                 }
 949
 950                 iter->skip = 0;
 951         }
 952
 953         iter->p = NULL;
 954
 955 out:
 956         if (iter->walker->resize) {
 957                 iter->p = NULL;
 958                 iter->slot = 0;
 959                 iter->skip = 0;
 960                 iter->walker->resize = false;
 961                 return ERR_PTR(-EAGAIN);
 962         }
 963
 964         return obj;
 965 }
 966 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 967
 968 /**
 969  * rhashtable_walk_stop - Finish a hash table walk
 970  * @iter:       Hash table iterator
 971  *
 972  * Finish a hash table walk.
 973  */
 974 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 975 {
 976         rcu_read_unlock();
 977         iter->p = NULL;
 978 }
 979 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 980
 981 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 982 {
 983         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 984                    1UL << params->min_shift);
 985 }
 986
 987 /**
 988  * rhashtable_init - initialize a new hash table
 989  * @ht:         hash table to be initialized
 990  * @params:     configuration parameters
 991  *
 992  * Initializes a new hash table based on the provided configuration
 993  * parameters. A table can be configured either with a variable or
 994  * fixed length key:
 995  *
 996  * Configuration Example 1: Fixed length keys
 997  * struct test_obj {
 998  *      int                     key;
 999  *      void *                  my_member;
1000  *      struct rhash_head       node;
1001  * };
1002  *
1003  * struct rhashtable_params params = {
1004  *      .head_offset = offsetof(struct test_obj, node),
1005  *      .key_offset = offsetof(struct test_obj, key),
1006  *      .key_len = sizeof(int),
1007  *      .hashfn = jhash,
1008  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1009  * };
1010  *
1011  * Configuration Example 2: Variable length keys
1012  * struct test_obj {
1013  *      [...]
1014  *      struct rhash_head       node;
1015  * };
1016  *
1017  * u32 my_hash_fn(const void *data, u32 seed)
1018  * {
1019  *      struct test_obj *obj = data;
1020  *
1021  *      return [... hash ...];
1022  * }
1023  *
1024  * struct rhashtable_params params = {
1025  *      .head_offset = offsetof(struct test_obj, node),
1026  *      .hashfn = jhash,
1027  *      .obj_hashfn = my_hash_fn,
1028  * };
1029  */
1030 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1031 {
1032         struct bucket_table *tbl;
1033         size_t size;
1034
1035         size = HASH_DEFAULT_SIZE;
1036
1037         if ((params->key_len && !params->hashfn) ||
1038             (!params->key_len && !params->obj_hashfn))
1039                 return -EINVAL;
1040
1041         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1042                 return -EINVAL;
1043
1044         params->min_shift = max_t(size_t, params->min_shift,
1045                                   ilog2(HASH_MIN_SIZE));
1046
1047         if (params->nelem_hint)
1048                 size = rounded_hashtable_size(params);
1049
1050         memset(ht, 0, sizeof(*ht));
1051         mutex_init(&ht->mutex);
1052         memcpy(&ht->p, params, sizeof(*params));
1053         INIT_LIST_HEAD(&ht->walkers);
1054
1055         if (params->locks_mul)
1056                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1057         else
1058                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1059
1060         tbl = bucket_table_alloc(ht, size);
1061         if (tbl == NULL)
1062                 return -ENOMEM;
1063
1064         atomic_set(&ht->nelems, 0);
1065         atomic_set(&ht->shift, ilog2(tbl->size));
1066         RCU_INIT_POINTER(ht->tbl, tbl);
1067         RCU_INIT_POINTER(ht->future_tbl, tbl);
1068
1069         if (!ht->p.hash_rnd)
1070                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1071
1072         if (ht->p.grow_decision || ht->p.shrink_decision)
1073                 INIT_WORK(&ht->run_work, rht_deferred_worker);
1074
1075         return 0;
1076 }
1077 EXPORT_SYMBOL_GPL(rhashtable_init);
1078
1079 /**
1080  * rhashtable_destroy - destroy hash table
1081  * @ht:         the hash table to destroy
1082  *
1083  * Frees the bucket array. This function is not rcu safe, therefore the caller
1084  * has to make sure that no resizing may happen by unpublishing the hashtable
1085  * and waiting for the quiescent cycle before releasing the bucket array.
1086  */
1087 void rhashtable_destroy(struct rhashtable *ht)
1088 {
1089         ht->being_destroyed = true;
1090
1091         if (ht->p.grow_decision || ht->p.shrink_decision)
1092                 cancel_work_sync(&ht->run_work);
1093
1094         mutex_lock(&ht->mutex);
1095         bucket_table_free(rht_dereference(ht->tbl, ht));
1096         mutex_unlock(&ht->mutex);
1097 }
1098 EXPORT_SYMBOL_GPL(rhashtable_destroy);