lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2014-2015 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 };
  38
  39 /* The bucket lock is selected based on the hash and protects mutations
  40  * on a group of hash buckets.
  41  *
  42  * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
  43  * a single lock always covers both buckets which may both contains
  44  * entries which link to the same bucket of the old table during resizing.
  45  * This allows to simplify the locking as locking the bucket in both
  46  * tables during resize always guarantee protection.
  47  *
  48  * IMPORTANT: When holding the bucket lock of both the old and new table
  49  * during expansions and shrinking, the old bucket lock must always be
  50  * acquired first.
  51  */
  52 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  53 {
  54         return &tbl->locks[hash & tbl->locks_mask];
  55 }
  56
  57 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  58 {
  59         return (void *) he - ht->p.head_offset;
  60 }
  61
  62 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  63 {
  64         return hash & (tbl->size - 1);
  65 }
  66
  67 static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr)
  68 {
  69         u32 hash;
  70
  71         if (unlikely(!ht->p.key_len))
  72                 hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
  73         else
  74                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  75                                     ht->p.hash_rnd);
  76
  77         return hash >> HASH_RESERVED_SPACE;
  78 }
  79
  80 static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len)
  81 {
  82         return ht->p.hashfn(key, len, ht->p.hash_rnd) >> HASH_RESERVED_SPACE;
  83 }
  84
  85 static u32 head_hashfn(const struct rhashtable *ht,
  86                        const struct bucket_table *tbl,
  87                        const struct rhash_head *he)
  88 {
  89         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
  90 }
  91
  92 #ifdef CONFIG_PROVE_LOCKING
  93 static void debug_dump_buckets(const struct rhashtable *ht,
  94                                const struct bucket_table *tbl)
  95 {
  96         struct rhash_head *he;
  97         unsigned int i, hash;
  98
  99         for (i = 0; i < tbl->size; i++) {
 100                 pr_warn(" [Bucket %d] ", i);
 101                 rht_for_each_rcu(he, tbl, i) {
 102                         hash = head_hashfn(ht, tbl, he);
 103                         pr_cont("[hash = %#x, lock = %p] ",
 104                                 hash, bucket_lock(tbl, hash));
 105                 }
 106                 pr_cont("\n");
 107         }
 108
 109 }
 110
 111 static void debug_dump_table(struct rhashtable *ht,
 112                              const struct bucket_table *tbl,
 113                              unsigned int hash)
 114 {
 115         struct bucket_table *old_tbl, *future_tbl;
 116
 117         pr_emerg("BUG: lock for hash %#x in table %p not held\n",
 118                  hash, tbl);
 119
 120         rcu_read_lock();
 121         future_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 122         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 123         if (future_tbl != old_tbl) {
 124                 pr_warn("Future table %p (size: %zd)\n",
 125                         future_tbl, future_tbl->size);
 126                 debug_dump_buckets(ht, future_tbl);
 127         }
 128
 129         pr_warn("Table %p (size: %zd)\n", old_tbl, old_tbl->size);
 130         debug_dump_buckets(ht, old_tbl);
 131
 132         rcu_read_unlock();
 133 }
 134
 135 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
 136 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)                               \
 137         do {                                                            \
 138                 if (unlikely(!lockdep_rht_bucket_is_held(TBL, HASH))) { \
 139                         debug_dump_table(HT, TBL, HASH);                \
 140                         BUG();                                          \
 141                 }                                                       \
 142         } while (0)
 143
 144 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
 145 {
 146         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
 147 }
 148 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
 149
 150 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
 151 {
 152         spinlock_t *lock = bucket_lock(tbl, hash);
 153
 154         return (debug_locks) ? lockdep_is_held(lock) : 1;
 155 }
 156 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
 157 #else
 158 #define ASSERT_RHT_MUTEX(HT)
 159 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)
 160 #endif
 161
 162
 163 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 164 {
 165         struct rhash_head __rcu **pprev;
 166
 167         for (pprev = &tbl->buckets[n];
 168              !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n));
 169              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 170                 ;
 171
 172         return pprev;
 173 }
 174
 175 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 176 {
 177         unsigned int i, size;
 178 #if defined(CONFIG_PROVE_LOCKING)
 179         unsigned int nr_pcpus = 2;
 180 #else
 181         unsigned int nr_pcpus = num_possible_cpus();
 182 #endif
 183
 184         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 185         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 186
 187         /* Never allocate more than 0.5 locks per bucket */
 188         size = min_t(unsigned int, size, tbl->size >> 1);
 189
 190         if (sizeof(spinlock_t) != 0) {
 191 #ifdef CONFIG_NUMA
 192                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 193                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 194                 else
 195 #endif
 196                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 197                                            GFP_KERNEL);
 198                 if (!tbl->locks)
 199                         return -ENOMEM;
 200                 for (i = 0; i < size; i++)
 201                         spin_lock_init(&tbl->locks[i]);
 202         }
 203         tbl->locks_mask = size - 1;
 204
 205         return 0;
 206 }
 207
 208 static void bucket_table_free(const struct bucket_table *tbl)
 209 {
 210         if (tbl)
 211                 kvfree(tbl->locks);
 212
 213         kvfree(tbl);
 214 }
 215
 216 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 217                                                size_t nbuckets)
 218 {
 219         struct bucket_table *tbl;
 220         size_t size;
 221         int i;
 222
 223         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 224         tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
 225         if (tbl == NULL)
 226                 tbl = vzalloc(size);
 227
 228         if (tbl == NULL)
 229                 return NULL;
 230
 231         tbl->size = nbuckets;
 232
 233         if (alloc_bucket_locks(ht, tbl) < 0) {
 234                 bucket_table_free(tbl);
 235                 return NULL;
 236         }
 237
 238         for (i = 0; i < nbuckets; i++)
 239                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 240
 241         return tbl;
 242 }
 243
 244 /**
 245  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 246  * @ht:         hash table
 247  * @new_size:   new table size
 248  */
 249 bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 250 {
 251         /* Expand table when exceeding 75% load */
 252         return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
 253                (ht->p.max_shift && atomic_read(&ht->shift) < ht->p.max_shift);
 254 }
 255 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 256
 257 /**
 258  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 259  * @ht:         hash table
 260  * @new_size:   new table size
 261  */
 262 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 263 {
 264         /* Shrink table beneath 30% load */
 265         return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
 266                (atomic_read(&ht->shift) > ht->p.min_shift);
 267 }
 268 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 269
 270 static void lock_buckets(struct bucket_table *new_tbl,
 271                          struct bucket_table *old_tbl, unsigned int hash)
 272         __acquires(old_bucket_lock)
 273 {
 274         spin_lock_bh(bucket_lock(old_tbl, hash));
 275         if (new_tbl != old_tbl)
 276                 spin_lock_bh_nested(bucket_lock(new_tbl, hash),
 277                                     RHT_LOCK_NESTED);
 278 }
 279
 280 static void unlock_buckets(struct bucket_table *new_tbl,
 281                            struct bucket_table *old_tbl, unsigned int hash)
 282         __releases(old_bucket_lock)
 283 {
 284         if (new_tbl != old_tbl)
 285                 spin_unlock_bh(bucket_lock(new_tbl, hash));
 286         spin_unlock_bh(bucket_lock(old_tbl, hash));
 287 }
 288
 289 /**
 290  * Unlink entries on bucket which hash to different bucket.
 291  *
 292  * Returns true if no more work needs to be performed on the bucket.
 293  */
 294 static bool hashtable_chain_unzip(struct rhashtable *ht,
 295                                   const struct bucket_table *new_tbl,
 296                                   struct bucket_table *old_tbl,
 297                                   size_t old_hash)
 298 {
 299         struct rhash_head *he, *p, *next;
 300         unsigned int new_hash, new_hash2;
 301
 302         ASSERT_BUCKET_LOCK(ht, old_tbl, old_hash);
 303
 304         /* Old bucket empty, no work needed. */
 305         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 306                                    old_hash);
 307         if (rht_is_a_nulls(p))
 308                 return false;
 309
 310         new_hash = head_hashfn(ht, new_tbl, p);
 311         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 312
 313         /* Advance the old bucket pointer one or more times until it
 314          * reaches a node that doesn't hash to the same bucket as the
 315          * previous node p. Call the previous node p;
 316          */
 317         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 318                 new_hash2 = head_hashfn(ht, new_tbl, he);
 319                 ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash2);
 320
 321                 if (new_hash != new_hash2)
 322                         break;
 323                 p = he;
 324         }
 325         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 326
 327         /* Find the subsequent node which does hash to the same
 328          * bucket as node P, or NULL if no such node exists.
 329          */
 330         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 331         if (!rht_is_a_nulls(he)) {
 332                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 333                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 334                                 next = he;
 335                                 break;
 336                         }
 337                 }
 338         }
 339
 340         /* Set p's next pointer to that subsequent node pointer,
 341          * bypassing the nodes which do not hash to p's bucket
 342          */
 343         rcu_assign_pointer(p->next, next);
 344
 345         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 346                                    old_hash);
 347
 348         return !rht_is_a_nulls(p);
 349 }
 350
 351 static void link_old_to_new(struct rhashtable *ht, struct bucket_table *new_tbl,
 352                             unsigned int new_hash, struct rhash_head *entry)
 353 {
 354         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 355
 356         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 357 }
 358
 359 /**
 360  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 361  * @ht:         the hash table to expand
 362  *
 363  * A secondary bucket array is allocated and the hash entries are migrated
 364  * while keeping them on both lists until the end of the RCU grace period.
 365  *
 366  * This function may only be called in a context where it is safe to call
 367  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 368  *
 369  * The caller must ensure that no concurrent resizing occurs by holding
 370  * ht->mutex.
 371  *
 372  * It is valid to have concurrent insertions and deletions protected by per
 373  * bucket locks or concurrent RCU protected lookups and traversals.
 374  */
 375 int rhashtable_expand(struct rhashtable *ht)
 376 {
 377         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 378         struct rhash_head *he;
 379         unsigned int new_hash, old_hash;
 380         bool complete = false;
 381
 382         ASSERT_RHT_MUTEX(ht);
 383
 384         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 385         if (new_tbl == NULL)
 386                 return -ENOMEM;
 387
 388         atomic_inc(&ht->shift);
 389
 390         /* Make insertions go into the new, empty table right away. Deletions
 391          * and lookups will be attempted in both tables until we synchronize.
 392          * The synchronize_rcu() guarantees for the new table to be picked up
 393          * so no new additions go into the old table while we relink.
 394          */
 395         rcu_assign_pointer(ht->future_tbl, new_tbl);
 396         synchronize_rcu();
 397
 398         /* For each new bucket, search the corresponding old bucket for the
 399          * first entry that hashes to the new bucket, and link the end of
 400          * newly formed bucket chain (containing entries added to future
 401          * table) to that entry. Since all the entries which will end up in
 402          * the new bucket appear in the same old bucket, this constructs an
 403          * entirely valid new hash table, but with multiple buckets
 404          * "zipped" together into a single imprecise chain.
 405          */
 406         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 407                 old_hash = rht_bucket_index(old_tbl, new_hash);
 408                 lock_buckets(new_tbl, old_tbl, new_hash);
 409                 rht_for_each(he, old_tbl, old_hash) {
 410                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 411                                 link_old_to_new(ht, new_tbl, new_hash, he);
 412                                 break;
 413                         }
 414                 }
 415                 unlock_buckets(new_tbl, old_tbl, new_hash);
 416         }
 417
 418         /* Unzip interleaved hash chains */
 419         while (!complete && !ht->being_destroyed) {
 420                 /* Wait for readers. All new readers will see the new
 421                  * table, and thus no references to the old table will
 422                  * remain.
 423                  */
 424                 synchronize_rcu();
 425
 426                 /* For each bucket in the old table (each of which
 427                  * contains items from multiple buckets of the new
 428                  * table): ...
 429                  */
 430                 complete = true;
 431                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 432                         lock_buckets(new_tbl, old_tbl, old_hash);
 433
 434                         if (hashtable_chain_unzip(ht, new_tbl, old_tbl,
 435                                                   old_hash))
 436                                 complete = false;
 437
 438                         unlock_buckets(new_tbl, old_tbl, old_hash);
 439                 }
 440         }
 441
 442         rcu_assign_pointer(ht->tbl, new_tbl);
 443         synchronize_rcu();
 444
 445         bucket_table_free(old_tbl);
 446         return 0;
 447 }
 448 EXPORT_SYMBOL_GPL(rhashtable_expand);
 449
 450 /**
 451  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 452  * @ht:         the hash table to shrink
 453  *
 454  * This function may only be called in a context where it is safe to call
 455  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 456  *
 457  * The caller must ensure that no concurrent resizing occurs by holding
 458  * ht->mutex.
 459  *
 460  * The caller must ensure that no concurrent table mutations take place.
 461  * It is however valid to have concurrent lookups if they are RCU protected.
 462  *
 463  * It is valid to have concurrent insertions and deletions protected by per
 464  * bucket locks or concurrent RCU protected lookups and traversals.
 465  */
 466 int rhashtable_shrink(struct rhashtable *ht)
 467 {
 468         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 469         unsigned int new_hash;
 470
 471         ASSERT_RHT_MUTEX(ht);
 472
 473         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 474         if (new_tbl == NULL)
 475                 return -ENOMEM;
 476
 477         rcu_assign_pointer(ht->future_tbl, new_tbl);
 478         synchronize_rcu();
 479
 480         /* Link the first entry in the old bucket to the end of the
 481          * bucket in the new table. As entries are concurrently being
 482          * added to the new table, lock down the new bucket. As we
 483          * always divide the size in half when shrinking, each bucket
 484          * in the new table maps to exactly two buckets in the old
 485          * table.
 486          */
 487         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 488                 lock_buckets(new_tbl, tbl, new_hash);
 489
 490                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 491                                    tbl->buckets[new_hash]);
 492                 ASSERT_BUCKET_LOCK(ht, tbl, new_hash + new_tbl->size);
 493                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 494                                    tbl->buckets[new_hash + new_tbl->size]);
 495
 496                 unlock_buckets(new_tbl, tbl, new_hash);
 497         }
 498
 499         /* Publish the new, valid hash table */
 500         rcu_assign_pointer(ht->tbl, new_tbl);
 501         atomic_dec(&ht->shift);
 502
 503         /* Wait for readers. No new readers will have references to the
 504          * old hash table.
 505          */
 506         synchronize_rcu();
 507
 508         bucket_table_free(tbl);
 509
 510         return 0;
 511 }
 512 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 513
 514 static void rht_deferred_worker(struct work_struct *work)
 515 {
 516         struct rhashtable *ht;
 517         struct bucket_table *tbl;
 518         struct rhashtable_walker *walker;
 519
 520         ht = container_of(work, struct rhashtable, run_work);
 521         mutex_lock(&ht->mutex);
 522         if (ht->being_destroyed)
 523                 goto unlock;
 524
 525         tbl = rht_dereference(ht->tbl, ht);
 526
 527         list_for_each_entry(walker, &ht->walkers, list)
 528                 walker->resize = true;
 529
 530         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 531                 rhashtable_expand(ht);
 532         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 533                 rhashtable_shrink(ht);
 534
 535 unlock:
 536         mutex_unlock(&ht->mutex);
 537 }
 538
 539 static void rhashtable_wakeup_worker(struct rhashtable *ht)
 540 {
 541         struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 542         struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 543         size_t size = tbl->size;
 544
 545         /* Only adjust the table if no resizing is currently in progress. */
 546         if (tbl == new_tbl &&
 547             ((ht->p.grow_decision && ht->p.grow_decision(ht, size)) ||
 548              (ht->p.shrink_decision && ht->p.shrink_decision(ht, size))))
 549                 schedule_work(&ht->run_work);
 550 }
 551
 552 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 553                                 struct bucket_table *tbl, u32 hash)
 554 {
 555         struct rhash_head *head;
 556
 557         hash = rht_bucket_index(tbl, hash);
 558         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 559
 560         ASSERT_BUCKET_LOCK(ht, tbl, hash);
 561
 562         if (rht_is_a_nulls(head))
 563                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 564         else
 565                 RCU_INIT_POINTER(obj->next, head);
 566
 567         rcu_assign_pointer(tbl->buckets[hash], obj);
 568
 569         atomic_inc(&ht->nelems);
 570
 571         rhashtable_wakeup_worker(ht);
 572 }
 573
 574 /**
 575  * rhashtable_insert - insert object into hash table
 576  * @ht:         hash table
 577  * @obj:        pointer to hash head inside object
 578  *
 579  * Will take a per bucket spinlock to protect against mutual mutations
 580  * on the same bucket. Multiple insertions may occur in parallel unless
 581  * they map to the same bucket lock.
 582  *
 583  * It is safe to call this function from atomic context.
 584  *
 585  * Will trigger an automatic deferred table resizing if the size grows
 586  * beyond the watermark indicated by grow_decision() which can be passed
 587  * to rhashtable_init().
 588  */
 589 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 590 {
 591         struct bucket_table *tbl, *old_tbl;
 592         unsigned hash;
 593
 594         rcu_read_lock();
 595
 596         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 597         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 598         hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 599
 600         lock_buckets(tbl, old_tbl, hash);
 601         __rhashtable_insert(ht, obj, tbl, hash);
 602         unlock_buckets(tbl, old_tbl, hash);
 603
 604         rcu_read_unlock();
 605 }
 606 EXPORT_SYMBOL_GPL(rhashtable_insert);
 607
 608 /**
 609  * rhashtable_remove - remove object from hash table
 610  * @ht:         hash table
 611  * @obj:        pointer to hash head inside object
 612  *
 613  * Since the hash chain is single linked, the removal operation needs to
 614  * walk the bucket chain upon removal. The removal operation is thus
 615  * considerable slow if the hash table is not correctly sized.
 616  *
 617  * Will automatically shrink the table via rhashtable_expand() if the
 618  * shrink_decision function specified at rhashtable_init() returns true.
 619  *
 620  * The caller must ensure that no concurrent table mutations occur. It is
 621  * however valid to have concurrent lookups if they are RCU protected.
 622  */
 623 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 624 {
 625         struct bucket_table *tbl, *new_tbl, *old_tbl;
 626         struct rhash_head __rcu **pprev;
 627         struct rhash_head *he, *he2;
 628         unsigned int hash, new_hash;
 629         bool ret = false;
 630
 631         rcu_read_lock();
 632         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 633         tbl = new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 634         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 635
 636         lock_buckets(new_tbl, old_tbl, new_hash);
 637 restart:
 638         hash = rht_bucket_index(tbl, new_hash);
 639         pprev = &tbl->buckets[hash];
 640         rht_for_each(he, tbl, hash) {
 641                 if (he != obj) {
 642                         pprev = &he->next;
 643                         continue;
 644                 }
 645
 646                 ASSERT_BUCKET_LOCK(ht, tbl, hash);
 647
 648                 if (old_tbl->size > new_tbl->size && tbl == old_tbl &&
 649                     !rht_is_a_nulls(obj->next) &&
 650                     head_hashfn(ht, tbl, obj->next) != hash) {
 651                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 652                 } else if (unlikely(old_tbl->size < new_tbl->size && tbl == new_tbl)) {
 653                         rht_for_each_continue(he2, obj->next, tbl, hash) {
 654                                 if (head_hashfn(ht, tbl, he2) == hash) {
 655                                         rcu_assign_pointer(*pprev, he2);
 656                                         goto found;
 657                                 }
 658                         }
 659
 660                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 661                 } else {
 662                         rcu_assign_pointer(*pprev, obj->next);
 663                 }
 664
 665 found:
 666                 ret = true;
 667                 break;
 668         }
 669
 670         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 671          * 'future_tbl' only exists for a short period of time during
 672          * resizing. Thus traversing both is fine and the added cost is
 673          * very rare.
 674          */
 675         if (tbl != old_tbl) {
 676                 tbl = old_tbl;
 677                 goto restart;
 678         }
 679
 680         unlock_buckets(new_tbl, old_tbl, new_hash);
 681
 682         if (ret) {
 683                 atomic_dec(&ht->nelems);
 684                 rhashtable_wakeup_worker(ht);
 685         }
 686
 687         rcu_read_unlock();
 688
 689         return ret;
 690 }
 691 EXPORT_SYMBOL_GPL(rhashtable_remove);
 692
 693 struct rhashtable_compare_arg {
 694         struct rhashtable *ht;
 695         const void *key;
 696 };
 697
 698 static bool rhashtable_compare(void *ptr, void *arg)
 699 {
 700         struct rhashtable_compare_arg *x = arg;
 701         struct rhashtable *ht = x->ht;
 702
 703         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 704 }
 705
 706 /**
 707  * rhashtable_lookup - lookup key in hash table
 708  * @ht:         hash table
 709  * @key:        pointer to key
 710  *
 711  * Computes the hash value for the key and traverses the bucket chain looking
 712  * for a entry with an identical key. The first matching entry is returned.
 713  *
 714  * This lookup function may only be used for fixed key hash table (key_len
 715  * parameter set). It will BUG() if used inappropriately.
 716  *
 717  * Lookups may occur in parallel with hashtable mutations and resizing.
 718  */
 719 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 720 {
 721         struct rhashtable_compare_arg arg = {
 722                 .ht = ht,
 723                 .key = key,
 724         };
 725
 726         BUG_ON(!ht->p.key_len);
 727
 728         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 729 }
 730 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 731
 732 /**
 733  * rhashtable_lookup_compare - search hash table with compare function
 734  * @ht:         hash table
 735  * @key:        the pointer to the key
 736  * @compare:    compare function, must return true on match
 737  * @arg:        argument passed on to compare function
 738  *
 739  * Traverses the bucket chain behind the provided hash value and calls the
 740  * specified compare function for each entry.
 741  *
 742  * Lookups may occur in parallel with hashtable mutations and resizing.
 743  *
 744  * Returns the first entry on which the compare function returned true.
 745  */
 746 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 747                                 bool (*compare)(void *, void *), void *arg)
 748 {
 749         const struct bucket_table *tbl, *old_tbl;
 750         struct rhash_head *he;
 751         u32 hash;
 752
 753         rcu_read_lock();
 754
 755         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 756         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 757         hash = key_hashfn(ht, key, ht->p.key_len);
 758 restart:
 759         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 760                 if (!compare(rht_obj(ht, he), arg))
 761                         continue;
 762                 rcu_read_unlock();
 763                 return rht_obj(ht, he);
 764         }
 765
 766         if (unlikely(tbl != old_tbl)) {
 767                 tbl = old_tbl;
 768                 goto restart;
 769         }
 770         rcu_read_unlock();
 771
 772         return NULL;
 773 }
 774 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 775
 776 /**
 777  * rhashtable_lookup_insert - lookup and insert object into hash table
 778  * @ht:         hash table
 779  * @obj:        pointer to hash head inside object
 780  *
 781  * Locks down the bucket chain in both the old and new table if a resize
 782  * is in progress to ensure that writers can't remove from the old table
 783  * and can't insert to the new table during the atomic operation of search
 784  * and insertion. Searches for duplicates in both the old and new table if
 785  * a resize is in progress.
 786  *
 787  * This lookup function may only be used for fixed key hash table (key_len
 788  * parameter set). It will BUG() if used inappropriately.
 789  *
 790  * It is safe to call this function from atomic context.
 791  *
 792  * Will trigger an automatic deferred table resizing if the size grows
 793  * beyond the watermark indicated by grow_decision() which can be passed
 794  * to rhashtable_init().
 795  */
 796 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 797 {
 798         struct rhashtable_compare_arg arg = {
 799                 .ht = ht,
 800                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 801         };
 802
 803         BUG_ON(!ht->p.key_len);
 804
 805         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 806                                                 &arg);
 807 }
 808 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 809
 810 /**
 811  * rhashtable_lookup_compare_insert - search and insert object to hash table
 812  *                                    with compare function
 813  * @ht:         hash table
 814  * @obj:        pointer to hash head inside object
 815  * @compare:    compare function, must return true on match
 816  * @arg:        argument passed on to compare function
 817  *
 818  * Locks down the bucket chain in both the old and new table if a resize
 819  * is in progress to ensure that writers can't remove from the old table
 820  * and can't insert to the new table during the atomic operation of search
 821  * and insertion. Searches for duplicates in both the old and new table if
 822  * a resize is in progress.
 823  *
 824  * Lookups may occur in parallel with hashtable mutations and resizing.
 825  *
 826  * Will trigger an automatic deferred table resizing if the size grows
 827  * beyond the watermark indicated by grow_decision() which can be passed
 828  * to rhashtable_init().
 829  */
 830 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 831                                       struct rhash_head *obj,
 832                                       bool (*compare)(void *, void *),
 833                                       void *arg)
 834 {
 835         struct bucket_table *new_tbl, *old_tbl;
 836         u32 new_hash;
 837         bool success = true;
 838
 839         BUG_ON(!ht->p.key_len);
 840
 841         rcu_read_lock();
 842         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 843         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 844         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 845
 846         lock_buckets(new_tbl, old_tbl, new_hash);
 847
 848         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 849                                       compare, arg)) {
 850                 success = false;
 851                 goto exit;
 852         }
 853
 854         __rhashtable_insert(ht, obj, new_tbl, new_hash);
 855
 856 exit:
 857         unlock_buckets(new_tbl, old_tbl, new_hash);
 858         rcu_read_unlock();
 859
 860         return success;
 861 }
 862 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 863
 864 /**
 865  * rhashtable_walk_init - Initialise an iterator
 866  * @ht:         Table to walk over
 867  * @iter:       Hash table Iterator
 868  *
 869  * This function prepares a hash table walk.
 870  *
 871  * Note that if you restart a walk after rhashtable_walk_stop you
 872  * may see the same object twice.  Also, you may miss objects if
 873  * there are removals in between rhashtable_walk_stop and the next
 874  * call to rhashtable_walk_start.
 875  *
 876  * For a completely stable walk you should construct your own data
 877  * structure outside the hash table.
 878  *
 879  * This function may sleep so you must not call it from interrupt
 880  * context or with spin locks held.
 881  *
 882  * You must call rhashtable_walk_exit if this function returns
 883  * successfully.
 884  */
 885 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 886 {
 887         iter->ht = ht;
 888         iter->p = NULL;
 889         iter->slot = 0;
 890         iter->skip = 0;
 891
 892         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 893         if (!iter->walker)
 894                 return -ENOMEM;
 895
 896         mutex_lock(&ht->mutex);
 897         list_add(&iter->walker->list, &ht->walkers);
 898         mutex_unlock(&ht->mutex);
 899
 900         return 0;
 901 }
 902 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 903
 904 /**
 905  * rhashtable_walk_exit - Free an iterator
 906  * @iter:       Hash table Iterator
 907  *
 908  * This function frees resources allocated by rhashtable_walk_init.
 909  */
 910 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 911 {
 912         mutex_lock(&iter->ht->mutex);
 913         list_del(&iter->walker->list);
 914         mutex_unlock(&iter->ht->mutex);
 915         kfree(iter->walker);
 916 }
 917 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 918
 919 /**
 920  * rhashtable_walk_start - Start a hash table walk
 921  * @iter:       Hash table iterator
 922  *
 923  * Start a hash table walk.  Note that we take the RCU lock in all
 924  * cases including when we return an error.  So you must always call
 925  * rhashtable_walk_stop to clean up.
 926  *
 927  * Returns zero if successful.
 928  *
 929  * Returns -EAGAIN if resize event occured.  Note that the iterator
 930  * will rewind back to the beginning and you may use it immediately
 931  * by calling rhashtable_walk_next.
 932  */
 933 int rhashtable_walk_start(struct rhashtable_iter *iter)
 934 {
 935         rcu_read_lock();
 936
 937         if (iter->walker->resize) {
 938                 iter->slot = 0;
 939                 iter->skip = 0;
 940                 iter->walker->resize = false;
 941                 return -EAGAIN;
 942         }
 943
 944         return 0;
 945 }
 946 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 947
 948 /**
 949  * rhashtable_walk_next - Return the next object and advance the iterator
 950  * @iter:       Hash table iterator
 951  *
 952  * Note that you must call rhashtable_walk_stop when you are finished
 953  * with the walk.
 954  *
 955  * Returns the next object or NULL when the end of the table is reached.
 956  *
 957  * Returns -EAGAIN if resize event occured.  Note that the iterator
 958  * will rewind back to the beginning and you may continue to use it.
 959  */
 960 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 961 {
 962         const struct bucket_table *tbl;
 963         struct rhashtable *ht = iter->ht;
 964         struct rhash_head *p = iter->p;
 965         void *obj = NULL;
 966
 967         tbl = rht_dereference_rcu(ht->tbl, ht);
 968
 969         if (p) {
 970                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 971                 goto next;
 972         }
 973
 974         for (; iter->slot < tbl->size; iter->slot++) {
 975                 int skip = iter->skip;
 976
 977                 rht_for_each_rcu(p, tbl, iter->slot) {
 978                         if (!skip)
 979                                 break;
 980                         skip--;
 981                 }
 982
 983 next:
 984                 if (!rht_is_a_nulls(p)) {
 985                         iter->skip++;
 986                         iter->p = p;
 987                         obj = rht_obj(ht, p);
 988                         goto out;
 989                 }
 990
 991                 iter->skip = 0;
 992         }
 993
 994         iter->p = NULL;
 995
 996 out:
 997         if (iter->walker->resize) {
 998                 iter->p = NULL;
 999                 iter->slot = 0;
1000                 iter->skip = 0;
1001                 iter->walker->resize = false;
1002                 return ERR_PTR(-EAGAIN);
1003         }
1004
1005         return obj;
1006 }
1007 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
1008
1009 /**
1010  * rhashtable_walk_stop - Finish a hash table walk
1011  * @iter:       Hash table iterator
1012  *
1013  * Finish a hash table walk.
1014  */
1015 void rhashtable_walk_stop(struct rhashtable_iter *iter)
1016 {
1017         rcu_read_unlock();
1018         iter->p = NULL;
1019 }
1020 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
1021
1022 static size_t rounded_hashtable_size(struct rhashtable_params *params)
1023 {
1024         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
1025                    1UL << params->min_shift);
1026 }
1027
1028 /**
1029  * rhashtable_init - initialize a new hash table
1030  * @ht:         hash table to be initialized
1031  * @params:     configuration parameters
1032  *
1033  * Initializes a new hash table based on the provided configuration
1034  * parameters. A table can be configured either with a variable or
1035  * fixed length key:
1036  *
1037  * Configuration Example 1: Fixed length keys
1038  * struct test_obj {
1039  *      int                     key;
1040  *      void *                  my_member;
1041  *      struct rhash_head       node;
1042  * };
1043  *
1044  * struct rhashtable_params params = {
1045  *      .head_offset = offsetof(struct test_obj, node),
1046  *      .key_offset = offsetof(struct test_obj, key),
1047  *      .key_len = sizeof(int),
1048  *      .hashfn = jhash,
1049  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1050  * };
1051  *
1052  * Configuration Example 2: Variable length keys
1053  * struct test_obj {
1054  *      [...]
1055  *      struct rhash_head       node;
1056  * };
1057  *
1058  * u32 my_hash_fn(const void *data, u32 seed)
1059  * {
1060  *      struct test_obj *obj = data;
1061  *
1062  *      return [... hash ...];
1063  * }
1064  *
1065  * struct rhashtable_params params = {
1066  *      .head_offset = offsetof(struct test_obj, node),
1067  *      .hashfn = jhash,
1068  *      .obj_hashfn = my_hash_fn,
1069  * };
1070  */
1071 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1072 {
1073         struct bucket_table *tbl;
1074         size_t size;
1075
1076         size = HASH_DEFAULT_SIZE;
1077
1078         if ((params->key_len && !params->hashfn) ||
1079             (!params->key_len && !params->obj_hashfn))
1080                 return -EINVAL;
1081
1082         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1083                 return -EINVAL;
1084
1085         params->min_shift = max_t(size_t, params->min_shift,
1086                                   ilog2(HASH_MIN_SIZE));
1087
1088         if (params->nelem_hint)
1089                 size = rounded_hashtable_size(params);
1090
1091         memset(ht, 0, sizeof(*ht));
1092         mutex_init(&ht->mutex);
1093         memcpy(&ht->p, params, sizeof(*params));
1094         INIT_LIST_HEAD(&ht->walkers);
1095
1096         if (params->locks_mul)
1097                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1098         else
1099                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1100
1101         tbl = bucket_table_alloc(ht, size);
1102         if (tbl == NULL)
1103                 return -ENOMEM;
1104
1105         atomic_set(&ht->nelems, 0);
1106         atomic_set(&ht->shift, ilog2(tbl->size));
1107         RCU_INIT_POINTER(ht->tbl, tbl);
1108         RCU_INIT_POINTER(ht->future_tbl, tbl);
1109
1110         if (!ht->p.hash_rnd)
1111                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1112
1113         if (ht->p.grow_decision || ht->p.shrink_decision)
1114                 INIT_WORK(&ht->run_work, rht_deferred_worker);
1115
1116         return 0;
1117 }
1118 EXPORT_SYMBOL_GPL(rhashtable_init);
1119
1120 /**
1121  * rhashtable_destroy - destroy hash table
1122  * @ht:         the hash table to destroy
1123  *
1124  * Frees the bucket array. This function is not rcu safe, therefore the caller
1125  * has to make sure that no resizing may happen by unpublishing the hashtable
1126  * and waiting for the quiescent cycle before releasing the bucket array.
1127  */
1128 void rhashtable_destroy(struct rhashtable *ht)
1129 {
1130         ht->being_destroyed = true;
1131
1132         if (ht->p.grow_decision || ht->p.shrink_decision)
1133                 cancel_work_sync(&ht->run_work);
1134
1135         mutex_lock(&ht->mutex);
1136         bucket_table_free(rht_dereference(ht->tbl, ht));
1137         mutex_unlock(&ht->mutex);
1138 }
1139 EXPORT_SYMBOL_GPL(rhashtable_destroy);