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
 447                 /* Depending on the lock per buckets mapping, the bucket in
 448                  * the lower and upper region may map to the same lock.
 449                  */
 450                 if (old_bucket_lock1 != old_bucket_lock2) {
 451                         spin_lock_bh_nested(old_bucket_lock2, RHT_LOCK_NESTED);
 452                         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED2);
 453                 } else {
 454                         spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 455                 }
 456
 457                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 458                                    tbl->buckets[new_hash]);
 459                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 460                                    tbl->buckets[new_hash + new_tbl->size]);
 461
 462                 spin_unlock_bh(new_bucket_lock);
 463                 if (old_bucket_lock1 != old_bucket_lock2)
 464                         spin_unlock_bh(old_bucket_lock2);
 465                 spin_unlock_bh(old_bucket_lock1);
 466         }
 467
 468         /* Publish the new, valid hash table */
 469         rcu_assign_pointer(ht->tbl, new_tbl);
 470         atomic_dec(&ht->shift);
 471
 472         /* Wait for readers. No new readers will have references to the
 473          * old hash table.
 474          */
 475         synchronize_rcu();
 476
 477         bucket_table_free(tbl);
 478
 479         return 0;
 480 }
 481 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 482
 483 static void rht_deferred_worker(struct work_struct *work)
 484 {
 485         struct rhashtable *ht;
 486         struct bucket_table *tbl;
 487         struct rhashtable_walker *walker;
 488
 489         ht = container_of(work, struct rhashtable, run_work);
 490         mutex_lock(&ht->mutex);
 491         if (ht->being_destroyed)
 492                 goto unlock;
 493
 494         tbl = rht_dereference(ht->tbl, ht);
 495
 496         list_for_each_entry(walker, &ht->walkers, list)
 497                 walker->resize = true;
 498
 499         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 500                 rhashtable_expand(ht);
 501         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 502                 rhashtable_shrink(ht);
 503
 504 unlock:
 505         mutex_unlock(&ht->mutex);
 506 }
 507
 508 static void rhashtable_wakeup_worker(struct rhashtable *ht)
 509 {
 510         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 511         struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 512         size_t size = tbl->size;
 513
 514         /* Only adjust the table if no resizing is currently in progress. */
 515         if (tbl == new_tbl &&
 516             ((ht->p.grow_decision && ht->p.grow_decision(ht, size)) ||
 517              (ht->p.shrink_decision && ht->p.shrink_decision(ht, size))))
 518                 schedule_work(&ht->run_work);
 519 }
 520
 521 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 522                                 struct bucket_table *tbl, u32 hash)
 523 {
 524         struct rhash_head *head = rht_dereference_bucket(tbl->buckets[hash],
 525                                                          tbl, hash);
 526
 527         if (rht_is_a_nulls(head))
 528                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 529         else
 530                 RCU_INIT_POINTER(obj->next, head);
 531
 532         rcu_assign_pointer(tbl->buckets[hash], obj);
 533
 534         atomic_inc(&ht->nelems);
 535
 536         rhashtable_wakeup_worker(ht);
 537 }
 538
 539 /**
 540  * rhashtable_insert - insert object into hash table
 541  * @ht:         hash table
 542  * @obj:        pointer to hash head inside object
 543  *
 544  * Will take a per bucket spinlock to protect against mutual mutations
 545  * on the same bucket. Multiple insertions may occur in parallel unless
 546  * they map to the same bucket lock.
 547  *
 548  * It is safe to call this function from atomic context.
 549  *
 550  * Will trigger an automatic deferred table resizing if the size grows
 551  * beyond the watermark indicated by grow_decision() which can be passed
 552  * to rhashtable_init().
 553  */
 554 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 555 {
 556         struct bucket_table *tbl;
 557         spinlock_t *lock;
 558         unsigned hash;
 559
 560         rcu_read_lock();
 561
 562         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 563         hash = head_hashfn(ht, tbl, obj);
 564         lock = bucket_lock(tbl, hash);
 565
 566         spin_lock_bh(lock);
 567         __rhashtable_insert(ht, obj, tbl, hash);
 568         spin_unlock_bh(lock);
 569
 570         rcu_read_unlock();
 571 }
 572 EXPORT_SYMBOL_GPL(rhashtable_insert);
 573
 574 /**
 575  * rhashtable_remove - remove object from hash table
 576  * @ht:         hash table
 577  * @obj:        pointer to hash head inside object
 578  *
 579  * Since the hash chain is single linked, the removal operation needs to
 580  * walk the bucket chain upon removal. The removal operation is thus
 581  * considerable slow if the hash table is not correctly sized.
 582  *
 583  * Will automatically shrink the table via rhashtable_expand() if the
 584  * shrink_decision function specified at rhashtable_init() returns true.
 585  *
 586  * The caller must ensure that no concurrent table mutations occur. It is
 587  * however valid to have concurrent lookups if they are RCU protected.
 588  */
 589 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 590 {
 591         struct bucket_table *tbl;
 592         struct rhash_head __rcu **pprev;
 593         struct rhash_head *he;
 594         spinlock_t *lock;
 595         unsigned int hash;
 596         bool ret = false;
 597
 598         rcu_read_lock();
 599         tbl = rht_dereference_rcu(ht->tbl, ht);
 600         hash = head_hashfn(ht, tbl, obj);
 601
 602         lock = bucket_lock(tbl, hash);
 603         spin_lock_bh(lock);
 604
 605 restart:
 606         pprev = &tbl->buckets[hash];
 607         rht_for_each(he, tbl, hash) {
 608                 if (he != obj) {
 609                         pprev = &he->next;
 610                         continue;
 611                 }
 612
 613                 rcu_assign_pointer(*pprev, obj->next);
 614
 615                 ret = true;
 616                 break;
 617         }
 618
 619         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 620          * 'future_tbl' only exists for a short period of time during
 621          * resizing. Thus traversing both is fine and the added cost is
 622          * very rare.
 623          */
 624         if (tbl != rht_dereference_rcu(ht->future_tbl, ht)) {
 625                 spin_unlock_bh(lock);
 626
 627                 tbl = rht_dereference_rcu(ht->future_tbl, ht);
 628                 hash = head_hashfn(ht, tbl, obj);
 629
 630                 lock = bucket_lock(tbl, hash);
 631                 spin_lock_bh(lock);
 632                 goto restart;
 633         }
 634
 635         spin_unlock_bh(lock);
 636
 637         if (ret) {
 638                 atomic_dec(&ht->nelems);
 639                 rhashtable_wakeup_worker(ht);
 640         }
 641
 642         rcu_read_unlock();
 643
 644         return ret;
 645 }
 646 EXPORT_SYMBOL_GPL(rhashtable_remove);
 647
 648 struct rhashtable_compare_arg {
 649         struct rhashtable *ht;
 650         const void *key;
 651 };
 652
 653 static bool rhashtable_compare(void *ptr, void *arg)
 654 {
 655         struct rhashtable_compare_arg *x = arg;
 656         struct rhashtable *ht = x->ht;
 657
 658         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 659 }
 660
 661 /**
 662  * rhashtable_lookup - lookup key in hash table
 663  * @ht:         hash table
 664  * @key:        pointer to key
 665  *
 666  * Computes the hash value for the key and traverses the bucket chain looking
 667  * for a entry with an identical key. The first matching entry is returned.
 668  *
 669  * This lookup function may only be used for fixed key hash table (key_len
 670  * parameter set). It will BUG() if used inappropriately.
 671  *
 672  * Lookups may occur in parallel with hashtable mutations and resizing.
 673  */
 674 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 675 {
 676         struct rhashtable_compare_arg arg = {
 677                 .ht = ht,
 678                 .key = key,
 679         };
 680
 681         BUG_ON(!ht->p.key_len);
 682
 683         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 684 }
 685 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 686
 687 /**
 688  * rhashtable_lookup_compare - search hash table with compare function
 689  * @ht:         hash table
 690  * @key:        the pointer to the key
 691  * @compare:    compare function, must return true on match
 692  * @arg:        argument passed on to compare function
 693  *
 694  * Traverses the bucket chain behind the provided hash value and calls the
 695  * specified compare function for each entry.
 696  *
 697  * Lookups may occur in parallel with hashtable mutations and resizing.
 698  *
 699  * Returns the first entry on which the compare function returned true.
 700  */
 701 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 702                                 bool (*compare)(void *, void *), void *arg)
 703 {
 704         const struct bucket_table *tbl, *old_tbl;
 705         struct rhash_head *he;
 706         u32 hash;
 707
 708         rcu_read_lock();
 709
 710         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 711         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 712         hash = key_hashfn(ht, key, ht->p.key_len);
 713 restart:
 714         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 715                 if (!compare(rht_obj(ht, he), arg))
 716                         continue;
 717                 rcu_read_unlock();
 718                 return rht_obj(ht, he);
 719         }
 720
 721         if (unlikely(tbl != old_tbl)) {
 722                 tbl = old_tbl;
 723                 goto restart;
 724         }
 725         rcu_read_unlock();
 726
 727         return NULL;
 728 }
 729 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 730
 731 /**
 732  * rhashtable_lookup_insert - lookup and insert object into hash table
 733  * @ht:         hash table
 734  * @obj:        pointer to hash head inside object
 735  *
 736  * Locks down the bucket chain in both the old and new table if a resize
 737  * is in progress to ensure that writers can't remove from the old table
 738  * and can't insert to the new table during the atomic operation of search
 739  * and insertion. Searches for duplicates in both the old and new table if
 740  * a resize is in progress.
 741  *
 742  * This lookup function may only be used for fixed key hash table (key_len
 743  * parameter set). It will BUG() if used inappropriately.
 744  *
 745  * It is safe to call this function from atomic context.
 746  *
 747  * Will trigger an automatic deferred table resizing if the size grows
 748  * beyond the watermark indicated by grow_decision() which can be passed
 749  * to rhashtable_init().
 750  */
 751 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 752 {
 753         struct rhashtable_compare_arg arg = {
 754                 .ht = ht,
 755                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 756         };
 757
 758         BUG_ON(!ht->p.key_len);
 759
 760         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 761                                                 &arg);
 762 }
 763 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 764
 765 /**
 766  * rhashtable_lookup_compare_insert - search and insert object to hash table
 767  *                                    with compare function
 768  * @ht:         hash table
 769  * @obj:        pointer to hash head inside object
 770  * @compare:    compare function, must return true on match
 771  * @arg:        argument passed on to compare function
 772  *
 773  * Locks down the bucket chain in both the old and new table if a resize
 774  * is in progress to ensure that writers can't remove from the old table
 775  * and can't insert to the new table during the atomic operation of search
 776  * and insertion. Searches for duplicates in both the old and new table if
 777  * a resize is in progress.
 778  *
 779  * Lookups may occur in parallel with hashtable mutations and resizing.
 780  *
 781  * Will trigger an automatic deferred table resizing if the size grows
 782  * beyond the watermark indicated by grow_decision() which can be passed
 783  * to rhashtable_init().
 784  */
 785 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 786                                       struct rhash_head *obj,
 787                                       bool (*compare)(void *, void *),
 788                                       void *arg)
 789 {
 790         struct bucket_table *new_tbl, *old_tbl;
 791         spinlock_t *new_bucket_lock, *old_bucket_lock;
 792         u32 new_hash, old_hash;
 793         bool success = true;
 794
 795         BUG_ON(!ht->p.key_len);
 796
 797         rcu_read_lock();
 798
 799         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 800         old_hash = head_hashfn(ht, old_tbl, obj);
 801         old_bucket_lock = bucket_lock(old_tbl, old_hash);
 802         spin_lock_bh(old_bucket_lock);
 803
 804         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 805         new_hash = head_hashfn(ht, new_tbl, obj);
 806         new_bucket_lock = bucket_lock(new_tbl, new_hash);
 807         if (unlikely(old_tbl != new_tbl))
 808                 spin_lock_bh_nested(new_bucket_lock, RHT_LOCK_NESTED);
 809
 810         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 811                                       compare, arg)) {
 812                 success = false;
 813                 goto exit;
 814         }
 815
 816         __rhashtable_insert(ht, obj, new_tbl, new_hash);
 817
 818 exit:
 819         if (unlikely(old_tbl != new_tbl))
 820                 spin_unlock_bh(new_bucket_lock);
 821         spin_unlock_bh(old_bucket_lock);
 822
 823         rcu_read_unlock();
 824
 825         return success;
 826 }
 827 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 828
 829 /**
 830  * rhashtable_walk_init - Initialise an iterator
 831  * @ht:         Table to walk over
 832  * @iter:       Hash table Iterator
 833  *
 834  * This function prepares a hash table walk.
 835  *
 836  * Note that if you restart a walk after rhashtable_walk_stop you
 837  * may see the same object twice.  Also, you may miss objects if
 838  * there are removals in between rhashtable_walk_stop and the next
 839  * call to rhashtable_walk_start.
 840  *
 841  * For a completely stable walk you should construct your own data
 842  * structure outside the hash table.
 843  *
 844  * This function may sleep so you must not call it from interrupt
 845  * context or with spin locks held.
 846  *
 847  * You must call rhashtable_walk_exit if this function returns
 848  * successfully.
 849  */
 850 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 851 {
 852         iter->ht = ht;
 853         iter->p = NULL;
 854         iter->slot = 0;
 855         iter->skip = 0;
 856
 857         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 858         if (!iter->walker)
 859                 return -ENOMEM;
 860
 861         mutex_lock(&ht->mutex);
 862         list_add(&iter->walker->list, &ht->walkers);
 863         mutex_unlock(&ht->mutex);
 864
 865         return 0;
 866 }
 867 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 868
 869 /**
 870  * rhashtable_walk_exit - Free an iterator
 871  * @iter:       Hash table Iterator
 872  *
 873  * This function frees resources allocated by rhashtable_walk_init.
 874  */
 875 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 876 {
 877         mutex_lock(&iter->ht->mutex);
 878         list_del(&iter->walker->list);
 879         mutex_unlock(&iter->ht->mutex);
 880         kfree(iter->walker);
 881 }
 882 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 883
 884 /**
 885  * rhashtable_walk_start - Start a hash table walk
 886  * @iter:       Hash table iterator
 887  *
 888  * Start a hash table walk.  Note that we take the RCU lock in all
 889  * cases including when we return an error.  So you must always call
 890  * rhashtable_walk_stop to clean up.
 891  *
 892  * Returns zero if successful.
 893  *
 894  * Returns -EAGAIN if resize event occured.  Note that the iterator
 895  * will rewind back to the beginning and you may use it immediately
 896  * by calling rhashtable_walk_next.
 897  */
 898 int rhashtable_walk_start(struct rhashtable_iter *iter)
 899 {
 900         rcu_read_lock();
 901
 902         if (iter->walker->resize) {
 903                 iter->slot = 0;
 904                 iter->skip = 0;
 905                 iter->walker->resize = false;
 906                 return -EAGAIN;
 907         }
 908
 909         return 0;
 910 }
 911 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 912
 913 /**
 914  * rhashtable_walk_next - Return the next object and advance the iterator
 915  * @iter:       Hash table iterator
 916  *
 917  * Note that you must call rhashtable_walk_stop when you are finished
 918  * with the walk.
 919  *
 920  * Returns the next object or NULL when the end of the table is reached.
 921  *
 922  * Returns -EAGAIN if resize event occured.  Note that the iterator
 923  * will rewind back to the beginning and you may continue to use it.
 924  */
 925 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 926 {
 927         const struct bucket_table *tbl;
 928         struct rhashtable *ht = iter->ht;
 929         struct rhash_head *p = iter->p;
 930         void *obj = NULL;
 931
 932         tbl = rht_dereference_rcu(ht->tbl, ht);
 933
 934         if (p) {
 935                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 936                 goto next;
 937         }
 938
 939         for (; iter->slot < tbl->size; iter->slot++) {
 940                 int skip = iter->skip;
 941
 942                 rht_for_each_rcu(p, tbl, iter->slot) {
 943                         if (!skip)
 944                                 break;
 945                         skip--;
 946                 }
 947
 948 next:
 949                 if (!rht_is_a_nulls(p)) {
 950                         iter->skip++;
 951                         iter->p = p;
 952                         obj = rht_obj(ht, p);
 953                         goto out;
 954                 }
 955
 956                 iter->skip = 0;
 957         }
 958
 959         iter->p = NULL;
 960
 961 out:
 962         if (iter->walker->resize) {
 963                 iter->p = NULL;
 964                 iter->slot = 0;
 965                 iter->skip = 0;
 966                 iter->walker->resize = false;
 967                 return ERR_PTR(-EAGAIN);
 968         }
 969
 970         return obj;
 971 }
 972 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 973
 974 /**
 975  * rhashtable_walk_stop - Finish a hash table walk
 976  * @iter:       Hash table iterator
 977  *
 978  * Finish a hash table walk.
 979  */
 980 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 981 {
 982         rcu_read_unlock();
 983         iter->p = NULL;
 984 }
 985 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 986
 987 static size_t rounded_hashtable_size(struct rhashtable_params *params)
 988 {
 989         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 990                    1UL << params->min_shift);
 991 }
 992
 993 /**
 994  * rhashtable_init - initialize a new hash table
 995  * @ht:         hash table to be initialized
 996  * @params:     configuration parameters
 997  *
 998  * Initializes a new hash table based on the provided configuration
 999  * parameters. A table can be configured either with a variable or
1000  * fixed length key:
1001  *
1002  * Configuration Example 1: Fixed length keys
1003  * struct test_obj {
1004  *      int                     key;
1005  *      void *                  my_member;
1006  *      struct rhash_head       node;
1007  * };
1008  *
1009  * struct rhashtable_params params = {
1010  *      .head_offset = offsetof(struct test_obj, node),
1011  *      .key_offset = offsetof(struct test_obj, key),
1012  *      .key_len = sizeof(int),
1013  *      .hashfn = jhash,
1014  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1015  * };
1016  *
1017  * Configuration Example 2: Variable length keys
1018  * struct test_obj {
1019  *      [...]
1020  *      struct rhash_head       node;
1021  * };
1022  *
1023  * u32 my_hash_fn(const void *data, u32 seed)
1024  * {
1025  *      struct test_obj *obj = data;
1026  *
1027  *      return [... hash ...];
1028  * }
1029  *
1030  * struct rhashtable_params params = {
1031  *      .head_offset = offsetof(struct test_obj, node),
1032  *      .hashfn = jhash,
1033  *      .obj_hashfn = my_hash_fn,
1034  * };
1035  */
1036 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1037 {
1038         struct bucket_table *tbl;
1039         size_t size;
1040
1041         size = HASH_DEFAULT_SIZE;
1042
1043         if ((params->key_len && !params->hashfn) ||
1044             (!params->key_len && !params->obj_hashfn))
1045                 return -EINVAL;
1046
1047         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1048                 return -EINVAL;
1049
1050         params->min_shift = max_t(size_t, params->min_shift,
1051                                   ilog2(HASH_MIN_SIZE));
1052
1053         if (params->nelem_hint)
1054                 size = rounded_hashtable_size(params);
1055
1056         memset(ht, 0, sizeof(*ht));
1057         mutex_init(&ht->mutex);
1058         memcpy(&ht->p, params, sizeof(*params));
1059         INIT_LIST_HEAD(&ht->walkers);
1060
1061         if (params->locks_mul)
1062                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1063         else
1064                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1065
1066         tbl = bucket_table_alloc(ht, size);
1067         if (tbl == NULL)
1068                 return -ENOMEM;
1069
1070         atomic_set(&ht->nelems, 0);
1071         atomic_set(&ht->shift, ilog2(tbl->size));
1072         RCU_INIT_POINTER(ht->tbl, tbl);
1073         RCU_INIT_POINTER(ht->future_tbl, tbl);
1074
1075         if (!ht->p.hash_rnd)
1076                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1077
1078         if (ht->p.grow_decision || ht->p.shrink_decision)
1079                 INIT_WORK(&ht->run_work, rht_deferred_worker);
1080
1081         return 0;
1082 }
1083 EXPORT_SYMBOL_GPL(rhashtable_init);
1084
1085 /**
1086  * rhashtable_destroy - destroy hash table
1087  * @ht:         the hash table to destroy
1088  *
1089  * Frees the bucket array. This function is not rcu safe, therefore the caller
1090  * has to make sure that no resizing may happen by unpublishing the hashtable
1091  * and waiting for the quiescent cycle before releasing the bucket array.
1092  */
1093 void rhashtable_destroy(struct rhashtable *ht)
1094 {
1095         ht->being_destroyed = true;
1096
1097         if (ht->p.grow_decision || ht->p.shrink_decision)
1098                 cancel_work_sync(&ht->run_work);
1099
1100         mutex_lock(&ht->mutex);
1101         bucket_table_free(rht_dereference(ht->tbl, ht));
1102         mutex_unlock(&ht->mutex);
1103 }
1104 EXPORT_SYMBOL_GPL(rhashtable_destroy);