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 #include <linux/err.h>
  27
  28 #define HASH_DEFAULT_SIZE       64UL
  29 #define HASH_MIN_SIZE           4UL
  30 #define BUCKET_LOCKS_PER_CPU   128UL
  31
  32 /* Base bits plus 1 bit for nulls marker */
  33 #define HASH_RESERVED_SPACE     (RHT_BASE_BITS + 1)
  34
  35 enum {
  36         RHT_LOCK_NORMAL,
  37         RHT_LOCK_NESTED,
  38 };
  39
  40 /* The bucket lock is selected based on the hash and protects mutations
  41  * on a group of hash buckets.
  42  *
  43  * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
  44  * a single lock always covers both buckets which may both contains
  45  * entries which link to the same bucket of the old table during resizing.
  46  * This allows to simplify the locking as locking the bucket in both
  47  * tables during resize always guarantee protection.
  48  *
  49  * IMPORTANT: When holding the bucket lock of both the old and new table
  50  * during expansions and shrinking, the old bucket lock must always be
  51  * acquired first.
  52  */
  53 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  54 {
  55         return &tbl->locks[hash & tbl->locks_mask];
  56 }
  57
  58 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  59 {
  60         return (void *) he - ht->p.head_offset;
  61 }
  62
  63 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  64 {
  65         return hash & (tbl->size - 1);
  66 }
  67
  68 static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr)
  69 {
  70         u32 hash;
  71
  72         if (unlikely(!ht->p.key_len))
  73                 hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
  74         else
  75                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  76                                     ht->p.hash_rnd);
  77
  78         return hash >> HASH_RESERVED_SPACE;
  79 }
  80
  81 static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len)
  82 {
  83         return ht->p.hashfn(key, len, ht->p.hash_rnd) >> HASH_RESERVED_SPACE;
  84 }
  85
  86 static u32 head_hashfn(const struct rhashtable *ht,
  87                        const struct bucket_table *tbl,
  88                        const struct rhash_head *he)
  89 {
  90         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
  91 }
  92
  93 #ifdef CONFIG_PROVE_LOCKING
  94 static void debug_dump_buckets(const struct rhashtable *ht,
  95                                const struct bucket_table *tbl)
  96 {
  97         struct rhash_head *he;
  98         unsigned int i, hash;
  99
 100         for (i = 0; i < tbl->size; i++) {
 101                 pr_warn(" [Bucket %d] ", i);
 102                 rht_for_each_rcu(he, tbl, i) {
 103                         hash = head_hashfn(ht, tbl, he);
 104                         pr_cont("[hash = %#x, lock = %p] ",
 105                                 hash, bucket_lock(tbl, hash));
 106                 }
 107                 pr_cont("\n");
 108         }
 109
 110 }
 111
 112 static void debug_dump_table(struct rhashtable *ht,
 113                              const struct bucket_table *tbl,
 114                              unsigned int hash)
 115 {
 116         struct bucket_table *old_tbl, *future_tbl;
 117
 118         pr_emerg("BUG: lock for hash %#x in table %p not held\n",
 119                  hash, tbl);
 120
 121         rcu_read_lock();
 122         future_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 123         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 124         if (future_tbl != old_tbl) {
 125                 pr_warn("Future table %p (size: %zd)\n",
 126                         future_tbl, future_tbl->size);
 127                 debug_dump_buckets(ht, future_tbl);
 128         }
 129
 130         pr_warn("Table %p (size: %zd)\n", old_tbl, old_tbl->size);
 131         debug_dump_buckets(ht, old_tbl);
 132
 133         rcu_read_unlock();
 134 }
 135
 136 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
 137 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)                               \
 138         do {                                                            \
 139                 if (unlikely(!lockdep_rht_bucket_is_held(TBL, HASH))) { \
 140                         debug_dump_table(HT, TBL, HASH);                \
 141                         BUG();                                          \
 142                 }                                                       \
 143         } while (0)
 144
 145 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
 146 {
 147         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
 148 }
 149 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
 150
 151 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
 152 {
 153         spinlock_t *lock = bucket_lock(tbl, hash);
 154
 155         return (debug_locks) ? lockdep_is_held(lock) : 1;
 156 }
 157 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
 158 #else
 159 #define ASSERT_RHT_MUTEX(HT)
 160 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)
 161 #endif
 162
 163
 164 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 165 {
 166         struct rhash_head __rcu **pprev;
 167
 168         for (pprev = &tbl->buckets[n];
 169              !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n));
 170              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 171                 ;
 172
 173         return pprev;
 174 }
 175
 176 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 177 {
 178         unsigned int i, size;
 179 #if defined(CONFIG_PROVE_LOCKING)
 180         unsigned int nr_pcpus = 2;
 181 #else
 182         unsigned int nr_pcpus = num_possible_cpus();
 183 #endif
 184
 185         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 186         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 187
 188         /* Never allocate more than 0.5 locks per bucket */
 189         size = min_t(unsigned int, size, tbl->size >> 1);
 190
 191         if (sizeof(spinlock_t) != 0) {
 192 #ifdef CONFIG_NUMA
 193                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 194                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 195                 else
 196 #endif
 197                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 198                                            GFP_KERNEL);
 199                 if (!tbl->locks)
 200                         return -ENOMEM;
 201                 for (i = 0; i < size; i++)
 202                         spin_lock_init(&tbl->locks[i]);
 203         }
 204         tbl->locks_mask = size - 1;
 205
 206         return 0;
 207 }
 208
 209 static void bucket_table_free(const struct bucket_table *tbl)
 210 {
 211         if (tbl)
 212                 kvfree(tbl->locks);
 213
 214         kvfree(tbl);
 215 }
 216
 217 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 218                                                size_t nbuckets)
 219 {
 220         struct bucket_table *tbl = NULL;
 221         size_t size;
 222         int i;
 223
 224         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 225         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER))
 226                 tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
 227         if (tbl == NULL)
 228                 tbl = vzalloc(size);
 229         if (tbl == NULL)
 230                 return NULL;
 231
 232         tbl->size = nbuckets;
 233
 234         if (alloc_bucket_locks(ht, tbl) < 0) {
 235                 bucket_table_free(tbl);
 236                 return NULL;
 237         }
 238
 239         for (i = 0; i < nbuckets; i++)
 240                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 241
 242         return tbl;
 243 }
 244
 245 /**
 246  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 247  * @ht:         hash table
 248  * @new_size:   new table size
 249  */
 250 static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 251 {
 252         /* Expand table when exceeding 75% load */
 253         return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
 254                (!ht->p.max_shift || atomic_read(&ht->shift) < ht->p.max_shift);
 255 }
 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 static 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
 269 static void lock_buckets(struct bucket_table *new_tbl,
 270                          struct bucket_table *old_tbl, unsigned int hash)
 271         __acquires(old_bucket_lock)
 272 {
 273         spin_lock_bh(bucket_lock(old_tbl, hash));
 274         if (new_tbl != old_tbl)
 275                 spin_lock_bh_nested(bucket_lock(new_tbl, hash),
 276                                     RHT_LOCK_NESTED);
 277 }
 278
 279 static void unlock_buckets(struct bucket_table *new_tbl,
 280                            struct bucket_table *old_tbl, unsigned int hash)
 281         __releases(old_bucket_lock)
 282 {
 283         if (new_tbl != old_tbl)
 284                 spin_unlock_bh(bucket_lock(new_tbl, hash));
 285         spin_unlock_bh(bucket_lock(old_tbl, hash));
 286 }
 287
 288 /**
 289  * Unlink entries on bucket which hash to different bucket.
 290  *
 291  * Returns true if no more work needs to be performed on the bucket.
 292  */
 293 static bool hashtable_chain_unzip(struct rhashtable *ht,
 294                                   const struct bucket_table *new_tbl,
 295                                   struct bucket_table *old_tbl,
 296                                   size_t old_hash)
 297 {
 298         struct rhash_head *he, *p, *next;
 299         unsigned int new_hash, new_hash2;
 300
 301         ASSERT_BUCKET_LOCK(ht, old_tbl, old_hash);
 302
 303         /* Old bucket empty, no work needed. */
 304         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 305                                    old_hash);
 306         if (rht_is_a_nulls(p))
 307                 return false;
 308
 309         new_hash = head_hashfn(ht, new_tbl, p);
 310         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 311
 312         /* Advance the old bucket pointer one or more times until it
 313          * reaches a node that doesn't hash to the same bucket as the
 314          * previous node p. Call the previous node p;
 315          */
 316         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 317                 new_hash2 = head_hashfn(ht, new_tbl, he);
 318                 ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash2);
 319
 320                 if (new_hash != new_hash2)
 321                         break;
 322                 p = he;
 323         }
 324         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 325
 326         /* Find the subsequent node which does hash to the same
 327          * bucket as node P, or NULL if no such node exists.
 328          */
 329         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 330         if (!rht_is_a_nulls(he)) {
 331                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 332                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 333                                 next = he;
 334                                 break;
 335                         }
 336                 }
 337         }
 338
 339         /* Set p's next pointer to that subsequent node pointer,
 340          * bypassing the nodes which do not hash to p's bucket
 341          */
 342         rcu_assign_pointer(p->next, next);
 343
 344         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 345                                    old_hash);
 346
 347         return !rht_is_a_nulls(p);
 348 }
 349
 350 static void link_old_to_new(struct rhashtable *ht, struct bucket_table *new_tbl,
 351                             unsigned int new_hash, struct rhash_head *entry)
 352 {
 353         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 354
 355         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 356 }
 357
 358 /**
 359  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 360  * @ht:         the hash table to expand
 361  *
 362  * A secondary bucket array is allocated and the hash entries are migrated
 363  * while keeping them on both lists until the end of the RCU grace period.
 364  *
 365  * This function may only be called in a context where it is safe to call
 366  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 367  *
 368  * The caller must ensure that no concurrent resizing occurs by holding
 369  * ht->mutex.
 370  *
 371  * It is valid to have concurrent insertions and deletions protected by per
 372  * bucket locks or concurrent RCU protected lookups and traversals.
 373  */
 374 int rhashtable_expand(struct rhashtable *ht)
 375 {
 376         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 377         struct rhash_head *he;
 378         unsigned int new_hash, old_hash;
 379         bool complete = false;
 380
 381         ASSERT_RHT_MUTEX(ht);
 382
 383         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 384         if (new_tbl == NULL)
 385                 return -ENOMEM;
 386
 387         atomic_inc(&ht->shift);
 388
 389         /* Make insertions go into the new, empty table right away. Deletions
 390          * and lookups will be attempted in both tables until we synchronize.
 391          * The synchronize_rcu() guarantees for the new table to be picked up
 392          * so no new additions go into the old table while we relink.
 393          */
 394         rcu_assign_pointer(ht->future_tbl, new_tbl);
 395         synchronize_rcu();
 396
 397         /* For each new bucket, search the corresponding old bucket for the
 398          * first entry that hashes to the new bucket, and link the end of
 399          * newly formed bucket chain (containing entries added to future
 400          * table) to that entry. Since all the entries which will end up in
 401          * the new bucket appear in the same old bucket, this constructs an
 402          * entirely valid new hash table, but with multiple buckets
 403          * "zipped" together into a single imprecise chain.
 404          */
 405         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 406                 old_hash = rht_bucket_index(old_tbl, new_hash);
 407                 lock_buckets(new_tbl, old_tbl, new_hash);
 408                 rht_for_each(he, old_tbl, old_hash) {
 409                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 410                                 link_old_to_new(ht, new_tbl, new_hash, he);
 411                                 break;
 412                         }
 413                 }
 414                 unlock_buckets(new_tbl, old_tbl, new_hash);
 415         }
 416
 417         /* Unzip interleaved hash chains */
 418         while (!complete && !ht->being_destroyed) {
 419                 /* Wait for readers. All new readers will see the new
 420                  * table, and thus no references to the old table will
 421                  * remain.
 422                  */
 423                 synchronize_rcu();
 424
 425                 /* For each bucket in the old table (each of which
 426                  * contains items from multiple buckets of the new
 427                  * table): ...
 428                  */
 429                 complete = true;
 430                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 431                         lock_buckets(new_tbl, old_tbl, old_hash);
 432
 433                         if (hashtable_chain_unzip(ht, new_tbl, old_tbl,
 434                                                   old_hash))
 435                                 complete = false;
 436
 437                         unlock_buckets(new_tbl, old_tbl, old_hash);
 438                 }
 439         }
 440
 441         rcu_assign_pointer(ht->tbl, new_tbl);
 442         synchronize_rcu();
 443
 444         bucket_table_free(old_tbl);
 445         return 0;
 446 }
 447 EXPORT_SYMBOL_GPL(rhashtable_expand);
 448
 449 /**
 450  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 451  * @ht:         the hash table to shrink
 452  *
 453  * This function may only be called in a context where it is safe to call
 454  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 455  *
 456  * The caller must ensure that no concurrent resizing occurs by holding
 457  * ht->mutex.
 458  *
 459  * The caller must ensure that no concurrent table mutations take place.
 460  * It is however valid to have concurrent lookups if they are RCU protected.
 461  *
 462  * It is valid to have concurrent insertions and deletions protected by per
 463  * bucket locks or concurrent RCU protected lookups and traversals.
 464  */
 465 int rhashtable_shrink(struct rhashtable *ht)
 466 {
 467         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 468         unsigned int new_hash;
 469
 470         ASSERT_RHT_MUTEX(ht);
 471
 472         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 473         if (new_tbl == NULL)
 474                 return -ENOMEM;
 475
 476         rcu_assign_pointer(ht->future_tbl, new_tbl);
 477         synchronize_rcu();
 478
 479         /* Link the first entry in the old bucket to the end of the
 480          * bucket in the new table. As entries are concurrently being
 481          * added to the new table, lock down the new bucket. As we
 482          * always divide the size in half when shrinking, each bucket
 483          * in the new table maps to exactly two buckets in the old
 484          * table.
 485          */
 486         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 487                 lock_buckets(new_tbl, tbl, new_hash);
 488
 489                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 490                                    tbl->buckets[new_hash]);
 491                 ASSERT_BUCKET_LOCK(ht, tbl, new_hash + new_tbl->size);
 492                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 493                                    tbl->buckets[new_hash + new_tbl->size]);
 494
 495                 unlock_buckets(new_tbl, tbl, new_hash);
 496         }
 497
 498         /* Publish the new, valid hash table */
 499         rcu_assign_pointer(ht->tbl, new_tbl);
 500         atomic_dec(&ht->shift);
 501
 502         /* Wait for readers. No new readers will have references to the
 503          * old hash table.
 504          */
 505         synchronize_rcu();
 506
 507         bucket_table_free(tbl);
 508
 509         return 0;
 510 }
 511 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 512
 513 static void rht_deferred_worker(struct work_struct *work)
 514 {
 515         struct rhashtable *ht;
 516         struct bucket_table *tbl;
 517         struct rhashtable_walker *walker;
 518
 519         ht = container_of(work, struct rhashtable, run_work);
 520         mutex_lock(&ht->mutex);
 521         if (ht->being_destroyed)
 522                 goto unlock;
 523
 524         tbl = rht_dereference(ht->tbl, ht);
 525
 526         list_for_each_entry(walker, &ht->walkers, list)
 527                 walker->resize = true;
 528
 529         if (rht_grow_above_75(ht, tbl->size))
 530                 rhashtable_expand(ht);
 531         else if (rht_shrink_below_30(ht, tbl->size))
 532                 rhashtable_shrink(ht);
 533 unlock:
 534         mutex_unlock(&ht->mutex);
 535 }
 536
 537 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 538                                 struct bucket_table *tbl,
 539                                 const struct bucket_table *old_tbl, u32 hash)
 540 {
 541         bool no_resize_running = tbl == old_tbl;
 542         struct rhash_head *head;
 543
 544         hash = rht_bucket_index(tbl, hash);
 545         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 546
 547         ASSERT_BUCKET_LOCK(ht, tbl, hash);
 548
 549         if (rht_is_a_nulls(head))
 550                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 551         else
 552                 RCU_INIT_POINTER(obj->next, head);
 553
 554         rcu_assign_pointer(tbl->buckets[hash], obj);
 555
 556         atomic_inc(&ht->nelems);
 557         if (no_resize_running && rht_grow_above_75(ht, tbl->size))
 558                 schedule_work(&ht->run_work);
 559 }
 560
 561 /**
 562  * rhashtable_insert - insert object into hash table
 563  * @ht:         hash table
 564  * @obj:        pointer to hash head inside object
 565  *
 566  * Will take a per bucket spinlock to protect against mutual mutations
 567  * on the same bucket. Multiple insertions may occur in parallel unless
 568  * they map to the same bucket lock.
 569  *
 570  * It is safe to call this function from atomic context.
 571  *
 572  * Will trigger an automatic deferred table resizing if the size grows
 573  * beyond the watermark indicated by grow_decision() which can be passed
 574  * to rhashtable_init().
 575  */
 576 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 577 {
 578         struct bucket_table *tbl, *old_tbl;
 579         unsigned hash;
 580
 581         rcu_read_lock();
 582
 583         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 584         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 585         hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 586
 587         lock_buckets(tbl, old_tbl, hash);
 588         __rhashtable_insert(ht, obj, tbl, old_tbl, hash);
 589         unlock_buckets(tbl, old_tbl, hash);
 590
 591         rcu_read_unlock();
 592 }
 593 EXPORT_SYMBOL_GPL(rhashtable_insert);
 594
 595 /**
 596  * rhashtable_remove - remove object from hash table
 597  * @ht:         hash table
 598  * @obj:        pointer to hash head inside object
 599  *
 600  * Since the hash chain is single linked, the removal operation needs to
 601  * walk the bucket chain upon removal. The removal operation is thus
 602  * considerable slow if the hash table is not correctly sized.
 603  *
 604  * Will automatically shrink the table via rhashtable_expand() if the
 605  * shrink_decision function specified at rhashtable_init() returns true.
 606  *
 607  * The caller must ensure that no concurrent table mutations occur. It is
 608  * however valid to have concurrent lookups if they are RCU protected.
 609  */
 610 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 611 {
 612         struct bucket_table *tbl, *new_tbl, *old_tbl;
 613         struct rhash_head __rcu **pprev;
 614         struct rhash_head *he, *he2;
 615         unsigned int hash, new_hash;
 616         bool ret = false;
 617
 618         rcu_read_lock();
 619         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 620         tbl = new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 621         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 622
 623         lock_buckets(new_tbl, old_tbl, new_hash);
 624 restart:
 625         hash = rht_bucket_index(tbl, new_hash);
 626         pprev = &tbl->buckets[hash];
 627         rht_for_each(he, tbl, hash) {
 628                 if (he != obj) {
 629                         pprev = &he->next;
 630                         continue;
 631                 }
 632
 633                 ASSERT_BUCKET_LOCK(ht, tbl, hash);
 634
 635                 if (old_tbl->size > new_tbl->size && tbl == old_tbl &&
 636                     !rht_is_a_nulls(obj->next) &&
 637                     head_hashfn(ht, tbl, obj->next) != hash) {
 638                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 639                 } else if (unlikely(old_tbl->size < new_tbl->size && tbl == new_tbl)) {
 640                         rht_for_each_continue(he2, obj->next, tbl, hash) {
 641                                 if (head_hashfn(ht, tbl, he2) == hash) {
 642                                         rcu_assign_pointer(*pprev, he2);
 643                                         goto found;
 644                                 }
 645                         }
 646
 647                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 648                 } else {
 649                         rcu_assign_pointer(*pprev, obj->next);
 650                 }
 651
 652 found:
 653                 ret = true;
 654                 break;
 655         }
 656
 657         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 658          * 'future_tbl' only exists for a short period of time during
 659          * resizing. Thus traversing both is fine and the added cost is
 660          * very rare.
 661          */
 662         if (tbl != old_tbl) {
 663                 tbl = old_tbl;
 664                 goto restart;
 665         }
 666
 667         unlock_buckets(new_tbl, old_tbl, new_hash);
 668
 669         if (ret) {
 670                 bool no_resize_running = new_tbl == old_tbl;
 671
 672                 atomic_dec(&ht->nelems);
 673                 if (no_resize_running && rht_shrink_below_30(ht, new_tbl->size))
 674                         schedule_work(&ht->run_work);
 675         }
 676
 677         rcu_read_unlock();
 678
 679         return ret;
 680 }
 681 EXPORT_SYMBOL_GPL(rhashtable_remove);
 682
 683 struct rhashtable_compare_arg {
 684         struct rhashtable *ht;
 685         const void *key;
 686 };
 687
 688 static bool rhashtable_compare(void *ptr, void *arg)
 689 {
 690         struct rhashtable_compare_arg *x = arg;
 691         struct rhashtable *ht = x->ht;
 692
 693         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 694 }
 695
 696 /**
 697  * rhashtable_lookup - lookup key in hash table
 698  * @ht:         hash table
 699  * @key:        pointer to key
 700  *
 701  * Computes the hash value for the key and traverses the bucket chain looking
 702  * for a entry with an identical key. The first matching entry is returned.
 703  *
 704  * This lookup function may only be used for fixed key hash table (key_len
 705  * parameter set). It will BUG() if used inappropriately.
 706  *
 707  * Lookups may occur in parallel with hashtable mutations and resizing.
 708  */
 709 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 710 {
 711         struct rhashtable_compare_arg arg = {
 712                 .ht = ht,
 713                 .key = key,
 714         };
 715
 716         BUG_ON(!ht->p.key_len);
 717
 718         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 719 }
 720 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 721
 722 /**
 723  * rhashtable_lookup_compare - search hash table with compare function
 724  * @ht:         hash table
 725  * @key:        the pointer to the key
 726  * @compare:    compare function, must return true on match
 727  * @arg:        argument passed on to compare function
 728  *
 729  * Traverses the bucket chain behind the provided hash value and calls the
 730  * specified compare function for each entry.
 731  *
 732  * Lookups may occur in parallel with hashtable mutations and resizing.
 733  *
 734  * Returns the first entry on which the compare function returned true.
 735  */
 736 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 737                                 bool (*compare)(void *, void *), void *arg)
 738 {
 739         const struct bucket_table *tbl, *old_tbl;
 740         struct rhash_head *he;
 741         u32 hash;
 742
 743         rcu_read_lock();
 744
 745         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 746         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 747         hash = key_hashfn(ht, key, ht->p.key_len);
 748 restart:
 749         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 750                 if (!compare(rht_obj(ht, he), arg))
 751                         continue;
 752                 rcu_read_unlock();
 753                 return rht_obj(ht, he);
 754         }
 755
 756         if (unlikely(tbl != old_tbl)) {
 757                 tbl = old_tbl;
 758                 goto restart;
 759         }
 760         rcu_read_unlock();
 761
 762         return NULL;
 763 }
 764 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 765
 766 /**
 767  * rhashtable_lookup_insert - lookup and insert object into hash table
 768  * @ht:         hash table
 769  * @obj:        pointer to hash head inside object
 770  *
 771  * Locks down the bucket chain in both the old and new table if a resize
 772  * is in progress to ensure that writers can't remove from the old table
 773  * and can't insert to the new table during the atomic operation of search
 774  * and insertion. Searches for duplicates in both the old and new table if
 775  * a resize is in progress.
 776  *
 777  * This lookup function may only be used for fixed key hash table (key_len
 778  * parameter set). It will BUG() if used inappropriately.
 779  *
 780  * It is safe to call this function from atomic context.
 781  *
 782  * Will trigger an automatic deferred table resizing if the size grows
 783  * beyond the watermark indicated by grow_decision() which can be passed
 784  * to rhashtable_init().
 785  */
 786 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 787 {
 788         struct rhashtable_compare_arg arg = {
 789                 .ht = ht,
 790                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 791         };
 792
 793         BUG_ON(!ht->p.key_len);
 794
 795         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 796                                                 &arg);
 797 }
 798 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 799
 800 /**
 801  * rhashtable_lookup_compare_insert - search and insert object to hash table
 802  *                                    with compare function
 803  * @ht:         hash table
 804  * @obj:        pointer to hash head inside object
 805  * @compare:    compare function, must return true on match
 806  * @arg:        argument passed on to compare function
 807  *
 808  * Locks down the bucket chain in both the old and new table if a resize
 809  * is in progress to ensure that writers can't remove from the old table
 810  * and can't insert to the new table during the atomic operation of search
 811  * and insertion. Searches for duplicates in both the old and new table if
 812  * a resize is in progress.
 813  *
 814  * Lookups may occur in parallel with hashtable mutations and resizing.
 815  *
 816  * Will trigger an automatic deferred table resizing if the size grows
 817  * beyond the watermark indicated by grow_decision() which can be passed
 818  * to rhashtable_init().
 819  */
 820 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 821                                       struct rhash_head *obj,
 822                                       bool (*compare)(void *, void *),
 823                                       void *arg)
 824 {
 825         struct bucket_table *new_tbl, *old_tbl;
 826         u32 new_hash;
 827         bool success = true;
 828
 829         BUG_ON(!ht->p.key_len);
 830
 831         rcu_read_lock();
 832         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 833         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 834         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 835
 836         lock_buckets(new_tbl, old_tbl, new_hash);
 837
 838         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 839                                       compare, arg)) {
 840                 success = false;
 841                 goto exit;
 842         }
 843
 844         __rhashtable_insert(ht, obj, new_tbl, old_tbl, new_hash);
 845
 846 exit:
 847         unlock_buckets(new_tbl, old_tbl, new_hash);
 848         rcu_read_unlock();
 849
 850         return success;
 851 }
 852 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 853
 854 /**
 855  * rhashtable_walk_init - Initialise an iterator
 856  * @ht:         Table to walk over
 857  * @iter:       Hash table Iterator
 858  *
 859  * This function prepares a hash table walk.
 860  *
 861  * Note that if you restart a walk after rhashtable_walk_stop you
 862  * may see the same object twice.  Also, you may miss objects if
 863  * there are removals in between rhashtable_walk_stop and the next
 864  * call to rhashtable_walk_start.
 865  *
 866  * For a completely stable walk you should construct your own data
 867  * structure outside the hash table.
 868  *
 869  * This function may sleep so you must not call it from interrupt
 870  * context or with spin locks held.
 871  *
 872  * You must call rhashtable_walk_exit if this function returns
 873  * successfully.
 874  */
 875 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 876 {
 877         iter->ht = ht;
 878         iter->p = NULL;
 879         iter->slot = 0;
 880         iter->skip = 0;
 881
 882         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 883         if (!iter->walker)
 884                 return -ENOMEM;
 885
 886         INIT_LIST_HEAD(&iter->walker->list);
 887         iter->walker->resize = false;
 888
 889         mutex_lock(&ht->mutex);
 890         list_add(&iter->walker->list, &ht->walkers);
 891         mutex_unlock(&ht->mutex);
 892
 893         return 0;
 894 }
 895 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 896
 897 /**
 898  * rhashtable_walk_exit - Free an iterator
 899  * @iter:       Hash table Iterator
 900  *
 901  * This function frees resources allocated by rhashtable_walk_init.
 902  */
 903 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 904 {
 905         mutex_lock(&iter->ht->mutex);
 906         list_del(&iter->walker->list);
 907         mutex_unlock(&iter->ht->mutex);
 908         kfree(iter->walker);
 909 }
 910 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 911
 912 /**
 913  * rhashtable_walk_start - Start a hash table walk
 914  * @iter:       Hash table iterator
 915  *
 916  * Start a hash table walk.  Note that we take the RCU lock in all
 917  * cases including when we return an error.  So you must always call
 918  * rhashtable_walk_stop to clean up.
 919  *
 920  * Returns zero if successful.
 921  *
 922  * Returns -EAGAIN if resize event occured.  Note that the iterator
 923  * will rewind back to the beginning and you may use it immediately
 924  * by calling rhashtable_walk_next.
 925  */
 926 int rhashtable_walk_start(struct rhashtable_iter *iter)
 927 {
 928         rcu_read_lock();
 929
 930         if (iter->walker->resize) {
 931                 iter->slot = 0;
 932                 iter->skip = 0;
 933                 iter->walker->resize = false;
 934                 return -EAGAIN;
 935         }
 936
 937         return 0;
 938 }
 939 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 940
 941 /**
 942  * rhashtable_walk_next - Return the next object and advance the iterator
 943  * @iter:       Hash table iterator
 944  *
 945  * Note that you must call rhashtable_walk_stop when you are finished
 946  * with the walk.
 947  *
 948  * Returns the next object or NULL when the end of the table is reached.
 949  *
 950  * Returns -EAGAIN if resize event occured.  Note that the iterator
 951  * will rewind back to the beginning and you may continue to use it.
 952  */
 953 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 954 {
 955         const struct bucket_table *tbl;
 956         struct rhashtable *ht = iter->ht;
 957         struct rhash_head *p = iter->p;
 958         void *obj = NULL;
 959
 960         tbl = rht_dereference_rcu(ht->tbl, ht);
 961
 962         if (p) {
 963                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 964                 goto next;
 965         }
 966
 967         for (; iter->slot < tbl->size; iter->slot++) {
 968                 int skip = iter->skip;
 969
 970                 rht_for_each_rcu(p, tbl, iter->slot) {
 971                         if (!skip)
 972                                 break;
 973                         skip--;
 974                 }
 975
 976 next:
 977                 if (!rht_is_a_nulls(p)) {
 978                         iter->skip++;
 979                         iter->p = p;
 980                         obj = rht_obj(ht, p);
 981                         goto out;
 982                 }
 983
 984                 iter->skip = 0;
 985         }
 986
 987         iter->p = NULL;
 988
 989 out:
 990         if (iter->walker->resize) {
 991                 iter->p = NULL;
 992                 iter->slot = 0;
 993                 iter->skip = 0;
 994                 iter->walker->resize = false;
 995                 return ERR_PTR(-EAGAIN);
 996         }
 997
 998         return obj;
 999 }
1000 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
1001
1002 /**
1003  * rhashtable_walk_stop - Finish a hash table walk
1004  * @iter:       Hash table iterator
1005  *
1006  * Finish a hash table walk.
1007  */
1008 void rhashtable_walk_stop(struct rhashtable_iter *iter)
1009 {
1010         rcu_read_unlock();
1011         iter->p = NULL;
1012 }
1013 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
1014
1015 static size_t rounded_hashtable_size(struct rhashtable_params *params)
1016 {
1017         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
1018                    1UL << params->min_shift);
1019 }
1020
1021 /**
1022  * rhashtable_init - initialize a new hash table
1023  * @ht:         hash table to be initialized
1024  * @params:     configuration parameters
1025  *
1026  * Initializes a new hash table based on the provided configuration
1027  * parameters. A table can be configured either with a variable or
1028  * fixed length key:
1029  *
1030  * Configuration Example 1: Fixed length keys
1031  * struct test_obj {
1032  *      int                     key;
1033  *      void *                  my_member;
1034  *      struct rhash_head       node;
1035  * };
1036  *
1037  * struct rhashtable_params params = {
1038  *      .head_offset = offsetof(struct test_obj, node),
1039  *      .key_offset = offsetof(struct test_obj, key),
1040  *      .key_len = sizeof(int),
1041  *      .hashfn = jhash,
1042  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1043  * };
1044  *
1045  * Configuration Example 2: Variable length keys
1046  * struct test_obj {
1047  *      [...]
1048  *      struct rhash_head       node;
1049  * };
1050  *
1051  * u32 my_hash_fn(const void *data, u32 seed)
1052  * {
1053  *      struct test_obj *obj = data;
1054  *
1055  *      return [... hash ...];
1056  * }
1057  *
1058  * struct rhashtable_params params = {
1059  *      .head_offset = offsetof(struct test_obj, node),
1060  *      .hashfn = jhash,
1061  *      .obj_hashfn = my_hash_fn,
1062  * };
1063  */
1064 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1065 {
1066         struct bucket_table *tbl;
1067         size_t size;
1068
1069         size = HASH_DEFAULT_SIZE;
1070
1071         if ((params->key_len && !params->hashfn) ||
1072             (!params->key_len && !params->obj_hashfn))
1073                 return -EINVAL;
1074
1075         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1076                 return -EINVAL;
1077
1078         params->min_shift = max_t(size_t, params->min_shift,
1079                                   ilog2(HASH_MIN_SIZE));
1080
1081         if (params->nelem_hint)
1082                 size = rounded_hashtable_size(params);
1083
1084         memset(ht, 0, sizeof(*ht));
1085         mutex_init(&ht->mutex);
1086         memcpy(&ht->p, params, sizeof(*params));
1087         INIT_LIST_HEAD(&ht->walkers);
1088
1089         if (params->locks_mul)
1090                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1091         else
1092                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1093
1094         tbl = bucket_table_alloc(ht, size);
1095         if (tbl == NULL)
1096                 return -ENOMEM;
1097
1098         atomic_set(&ht->nelems, 0);
1099         atomic_set(&ht->shift, ilog2(tbl->size));
1100         RCU_INIT_POINTER(ht->tbl, tbl);
1101         RCU_INIT_POINTER(ht->future_tbl, tbl);
1102
1103         if (!ht->p.hash_rnd)
1104                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1105
1106         INIT_WORK(&ht->run_work, rht_deferred_worker);
1107
1108         return 0;
1109 }
1110 EXPORT_SYMBOL_GPL(rhashtable_init);
1111
1112 /**
1113  * rhashtable_destroy - destroy hash table
1114  * @ht:         the hash table to destroy
1115  *
1116  * Frees the bucket array. This function is not rcu safe, therefore the caller
1117  * has to make sure that no resizing may happen by unpublishing the hashtable
1118  * and waiting for the quiescent cycle before releasing the bucket array.
1119  */
1120 void rhashtable_destroy(struct rhashtable *ht)
1121 {
1122         ht->being_destroyed = true;
1123
1124         cancel_work_sync(&ht->run_work);
1125
1126         mutex_lock(&ht->mutex);
1127         bucket_table_free(rht_dereference(ht->tbl, ht));
1128         mutex_unlock(&ht->mutex);
1129 }
1130 EXPORT_SYMBOL_GPL(rhashtable_destroy);