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 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 EXPORT_SYMBOL_GPL(rht_grow_above_75);
 257
 258 /**
 259  * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
 260  * @ht:         hash table
 261  * @new_size:   new table size
 262  */
 263 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
 264 {
 265         /* Shrink table beneath 30% load */
 266         return atomic_read(&ht->nelems) < (new_size * 3 / 10) &&
 267                (atomic_read(&ht->shift) > ht->p.min_shift);
 268 }
 269 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
 270
 271 static void lock_buckets(struct bucket_table *new_tbl,
 272                          struct bucket_table *old_tbl, unsigned int hash)
 273         __acquires(old_bucket_lock)
 274 {
 275         spin_lock_bh(bucket_lock(old_tbl, hash));
 276         if (new_tbl != old_tbl)
 277                 spin_lock_bh_nested(bucket_lock(new_tbl, hash),
 278                                     RHT_LOCK_NESTED);
 279 }
 280
 281 static void unlock_buckets(struct bucket_table *new_tbl,
 282                            struct bucket_table *old_tbl, unsigned int hash)
 283         __releases(old_bucket_lock)
 284 {
 285         if (new_tbl != old_tbl)
 286                 spin_unlock_bh(bucket_lock(new_tbl, hash));
 287         spin_unlock_bh(bucket_lock(old_tbl, hash));
 288 }
 289
 290 /**
 291  * Unlink entries on bucket which hash to different bucket.
 292  *
 293  * Returns true if no more work needs to be performed on the bucket.
 294  */
 295 static bool hashtable_chain_unzip(struct rhashtable *ht,
 296                                   const struct bucket_table *new_tbl,
 297                                   struct bucket_table *old_tbl,
 298                                   size_t old_hash)
 299 {
 300         struct rhash_head *he, *p, *next;
 301         unsigned int new_hash, new_hash2;
 302
 303         ASSERT_BUCKET_LOCK(ht, old_tbl, old_hash);
 304
 305         /* Old bucket empty, no work needed. */
 306         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 307                                    old_hash);
 308         if (rht_is_a_nulls(p))
 309                 return false;
 310
 311         new_hash = head_hashfn(ht, new_tbl, p);
 312         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 313
 314         /* Advance the old bucket pointer one or more times until it
 315          * reaches a node that doesn't hash to the same bucket as the
 316          * previous node p. Call the previous node p;
 317          */
 318         rht_for_each_continue(he, p->next, old_tbl, old_hash) {
 319                 new_hash2 = head_hashfn(ht, new_tbl, he);
 320                 ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash2);
 321
 322                 if (new_hash != new_hash2)
 323                         break;
 324                 p = he;
 325         }
 326         rcu_assign_pointer(old_tbl->buckets[old_hash], p->next);
 327
 328         /* Find the subsequent node which does hash to the same
 329          * bucket as node P, or NULL if no such node exists.
 330          */
 331         INIT_RHT_NULLS_HEAD(next, ht, old_hash);
 332         if (!rht_is_a_nulls(he)) {
 333                 rht_for_each_continue(he, he->next, old_tbl, old_hash) {
 334                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 335                                 next = he;
 336                                 break;
 337                         }
 338                 }
 339         }
 340
 341         /* Set p's next pointer to that subsequent node pointer,
 342          * bypassing the nodes which do not hash to p's bucket
 343          */
 344         rcu_assign_pointer(p->next, next);
 345
 346         p = rht_dereference_bucket(old_tbl->buckets[old_hash], old_tbl,
 347                                    old_hash);
 348
 349         return !rht_is_a_nulls(p);
 350 }
 351
 352 static void link_old_to_new(struct rhashtable *ht, struct bucket_table *new_tbl,
 353                             unsigned int new_hash, struct rhash_head *entry)
 354 {
 355         ASSERT_BUCKET_LOCK(ht, new_tbl, new_hash);
 356
 357         rcu_assign_pointer(*bucket_tail(new_tbl, new_hash), entry);
 358 }
 359
 360 /**
 361  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 362  * @ht:         the hash table to expand
 363  *
 364  * A secondary bucket array is allocated and the hash entries are migrated
 365  * while keeping them on both lists until the end of the RCU grace period.
 366  *
 367  * This function may only be called in a context where it is safe to call
 368  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 369  *
 370  * The caller must ensure that no concurrent resizing occurs by holding
 371  * ht->mutex.
 372  *
 373  * It is valid to have concurrent insertions and deletions protected by per
 374  * bucket locks or concurrent RCU protected lookups and traversals.
 375  */
 376 int rhashtable_expand(struct rhashtable *ht)
 377 {
 378         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 379         struct rhash_head *he;
 380         unsigned int new_hash, old_hash;
 381         bool complete = false;
 382
 383         ASSERT_RHT_MUTEX(ht);
 384
 385         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2);
 386         if (new_tbl == NULL)
 387                 return -ENOMEM;
 388
 389         atomic_inc(&ht->shift);
 390
 391         /* Make insertions go into the new, empty table right away. Deletions
 392          * and lookups will be attempted in both tables until we synchronize.
 393          * The synchronize_rcu() guarantees for the new table to be picked up
 394          * so no new additions go into the old table while we relink.
 395          */
 396         rcu_assign_pointer(ht->future_tbl, new_tbl);
 397         synchronize_rcu();
 398
 399         /* For each new bucket, search the corresponding old bucket for the
 400          * first entry that hashes to the new bucket, and link the end of
 401          * newly formed bucket chain (containing entries added to future
 402          * table) to that entry. Since all the entries which will end up in
 403          * the new bucket appear in the same old bucket, this constructs an
 404          * entirely valid new hash table, but with multiple buckets
 405          * "zipped" together into a single imprecise chain.
 406          */
 407         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 408                 old_hash = rht_bucket_index(old_tbl, new_hash);
 409                 lock_buckets(new_tbl, old_tbl, new_hash);
 410                 rht_for_each(he, old_tbl, old_hash) {
 411                         if (head_hashfn(ht, new_tbl, he) == new_hash) {
 412                                 link_old_to_new(ht, new_tbl, new_hash, he);
 413                                 break;
 414                         }
 415                 }
 416                 unlock_buckets(new_tbl, old_tbl, new_hash);
 417         }
 418
 419         /* Unzip interleaved hash chains */
 420         while (!complete && !ht->being_destroyed) {
 421                 /* Wait for readers. All new readers will see the new
 422                  * table, and thus no references to the old table will
 423                  * remain.
 424                  */
 425                 synchronize_rcu();
 426
 427                 /* For each bucket in the old table (each of which
 428                  * contains items from multiple buckets of the new
 429                  * table): ...
 430                  */
 431                 complete = true;
 432                 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 433                         lock_buckets(new_tbl, old_tbl, old_hash);
 434
 435                         if (hashtable_chain_unzip(ht, new_tbl, old_tbl,
 436                                                   old_hash))
 437                                 complete = false;
 438
 439                         unlock_buckets(new_tbl, old_tbl, old_hash);
 440                 }
 441         }
 442
 443         rcu_assign_pointer(ht->tbl, new_tbl);
 444         synchronize_rcu();
 445
 446         bucket_table_free(old_tbl);
 447         return 0;
 448 }
 449 EXPORT_SYMBOL_GPL(rhashtable_expand);
 450
 451 /**
 452  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 453  * @ht:         the hash table to shrink
 454  *
 455  * This function may only be called in a context where it is safe to call
 456  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 457  *
 458  * The caller must ensure that no concurrent resizing occurs by holding
 459  * ht->mutex.
 460  *
 461  * The caller must ensure that no concurrent table mutations take place.
 462  * It is however valid to have concurrent lookups if they are RCU protected.
 463  *
 464  * It is valid to have concurrent insertions and deletions protected by per
 465  * bucket locks or concurrent RCU protected lookups and traversals.
 466  */
 467 int rhashtable_shrink(struct rhashtable *ht)
 468 {
 469         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 470         unsigned int new_hash;
 471
 472         ASSERT_RHT_MUTEX(ht);
 473
 474         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 475         if (new_tbl == NULL)
 476                 return -ENOMEM;
 477
 478         rcu_assign_pointer(ht->future_tbl, new_tbl);
 479         synchronize_rcu();
 480
 481         /* Link the first entry in the old bucket to the end of the
 482          * bucket in the new table. As entries are concurrently being
 483          * added to the new table, lock down the new bucket. As we
 484          * always divide the size in half when shrinking, each bucket
 485          * in the new table maps to exactly two buckets in the old
 486          * table.
 487          */
 488         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 489                 lock_buckets(new_tbl, tbl, new_hash);
 490
 491                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 492                                    tbl->buckets[new_hash]);
 493                 ASSERT_BUCKET_LOCK(ht, tbl, new_hash + new_tbl->size);
 494                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 495                                    tbl->buckets[new_hash + new_tbl->size]);
 496
 497                 unlock_buckets(new_tbl, tbl, new_hash);
 498         }
 499
 500         /* Publish the new, valid hash table */
 501         rcu_assign_pointer(ht->tbl, new_tbl);
 502         atomic_dec(&ht->shift);
 503
 504         /* Wait for readers. No new readers will have references to the
 505          * old hash table.
 506          */
 507         synchronize_rcu();
 508
 509         bucket_table_free(tbl);
 510
 511         return 0;
 512 }
 513 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 514
 515 static void rht_deferred_worker(struct work_struct *work)
 516 {
 517         struct rhashtable *ht;
 518         struct bucket_table *tbl;
 519         struct rhashtable_walker *walker;
 520
 521         ht = container_of(work, struct rhashtable, run_work);
 522         mutex_lock(&ht->mutex);
 523         if (ht->being_destroyed)
 524                 goto unlock;
 525
 526         tbl = rht_dereference(ht->tbl, ht);
 527
 528         list_for_each_entry(walker, &ht->walkers, list)
 529                 walker->resize = true;
 530
 531         if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
 532                 rhashtable_expand(ht);
 533         else if (ht->p.shrink_decision && ht->p.shrink_decision(ht, tbl->size))
 534                 rhashtable_shrink(ht);
 535
 536 unlock:
 537         mutex_unlock(&ht->mutex);
 538 }
 539
 540 static void rhashtable_probe_expand(struct rhashtable *ht)
 541 {
 542         const struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 543         const struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 544
 545         /* Only adjust the table if no resizing is currently in progress. */
 546         if (tbl == new_tbl && ht->p.grow_decision &&
 547             ht->p.grow_decision(ht, tbl->size))
 548                 schedule_work(&ht->run_work);
 549 }
 550
 551 static void rhashtable_probe_shrink(struct rhashtable *ht)
 552 {
 553         const struct bucket_table *new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 554         const struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
 555
 556         /* Only adjust the table if no resizing is currently in progress. */
 557         if (tbl == new_tbl && ht->p.shrink_decision &&
 558             ht->p.shrink_decision(ht, tbl->size))
 559                 schedule_work(&ht->run_work);
 560 }
 561
 562 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 563                                 struct bucket_table *tbl, u32 hash)
 564 {
 565         struct rhash_head *head;
 566
 567         hash = rht_bucket_index(tbl, hash);
 568         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 569
 570         ASSERT_BUCKET_LOCK(ht, tbl, hash);
 571
 572         if (rht_is_a_nulls(head))
 573                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 574         else
 575                 RCU_INIT_POINTER(obj->next, head);
 576
 577         rcu_assign_pointer(tbl->buckets[hash], obj);
 578
 579         atomic_inc(&ht->nelems);
 580
 581         rhashtable_probe_expand(ht);
 582 }
 583
 584 /**
 585  * rhashtable_insert - insert object into hash table
 586  * @ht:         hash table
 587  * @obj:        pointer to hash head inside object
 588  *
 589  * Will take a per bucket spinlock to protect against mutual mutations
 590  * on the same bucket. Multiple insertions may occur in parallel unless
 591  * they map to the same bucket lock.
 592  *
 593  * It is safe to call this function from atomic context.
 594  *
 595  * Will trigger an automatic deferred table resizing if the size grows
 596  * beyond the watermark indicated by grow_decision() which can be passed
 597  * to rhashtable_init().
 598  */
 599 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 600 {
 601         struct bucket_table *tbl, *old_tbl;
 602         unsigned hash;
 603
 604         rcu_read_lock();
 605
 606         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 607         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 608         hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 609
 610         lock_buckets(tbl, old_tbl, hash);
 611         __rhashtable_insert(ht, obj, tbl, hash);
 612         unlock_buckets(tbl, old_tbl, hash);
 613
 614         rcu_read_unlock();
 615 }
 616 EXPORT_SYMBOL_GPL(rhashtable_insert);
 617
 618 /**
 619  * rhashtable_remove - remove object from hash table
 620  * @ht:         hash table
 621  * @obj:        pointer to hash head inside object
 622  *
 623  * Since the hash chain is single linked, the removal operation needs to
 624  * walk the bucket chain upon removal. The removal operation is thus
 625  * considerable slow if the hash table is not correctly sized.
 626  *
 627  * Will automatically shrink the table via rhashtable_expand() if the
 628  * shrink_decision function specified at rhashtable_init() returns true.
 629  *
 630  * The caller must ensure that no concurrent table mutations occur. It is
 631  * however valid to have concurrent lookups if they are RCU protected.
 632  */
 633 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 634 {
 635         struct bucket_table *tbl, *new_tbl, *old_tbl;
 636         struct rhash_head __rcu **pprev;
 637         struct rhash_head *he, *he2;
 638         unsigned int hash, new_hash;
 639         bool ret = false;
 640
 641         rcu_read_lock();
 642         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 643         tbl = new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 644         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 645
 646         lock_buckets(new_tbl, old_tbl, new_hash);
 647 restart:
 648         hash = rht_bucket_index(tbl, new_hash);
 649         pprev = &tbl->buckets[hash];
 650         rht_for_each(he, tbl, hash) {
 651                 if (he != obj) {
 652                         pprev = &he->next;
 653                         continue;
 654                 }
 655
 656                 ASSERT_BUCKET_LOCK(ht, tbl, hash);
 657
 658                 if (old_tbl->size > new_tbl->size && tbl == old_tbl &&
 659                     !rht_is_a_nulls(obj->next) &&
 660                     head_hashfn(ht, tbl, obj->next) != hash) {
 661                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 662                 } else if (unlikely(old_tbl->size < new_tbl->size && tbl == new_tbl)) {
 663                         rht_for_each_continue(he2, obj->next, tbl, hash) {
 664                                 if (head_hashfn(ht, tbl, he2) == hash) {
 665                                         rcu_assign_pointer(*pprev, he2);
 666                                         goto found;
 667                                 }
 668                         }
 669
 670                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 671                 } else {
 672                         rcu_assign_pointer(*pprev, obj->next);
 673                 }
 674
 675 found:
 676                 ret = true;
 677                 break;
 678         }
 679
 680         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 681          * 'future_tbl' only exists for a short period of time during
 682          * resizing. Thus traversing both is fine and the added cost is
 683          * very rare.
 684          */
 685         if (tbl != old_tbl) {
 686                 tbl = old_tbl;
 687                 goto restart;
 688         }
 689
 690         unlock_buckets(new_tbl, old_tbl, new_hash);
 691
 692         if (ret) {
 693                 atomic_dec(&ht->nelems);
 694                 rhashtable_probe_shrink(ht);
 695         }
 696
 697         rcu_read_unlock();
 698
 699         return ret;
 700 }
 701 EXPORT_SYMBOL_GPL(rhashtable_remove);
 702
 703 struct rhashtable_compare_arg {
 704         struct rhashtable *ht;
 705         const void *key;
 706 };
 707
 708 static bool rhashtable_compare(void *ptr, void *arg)
 709 {
 710         struct rhashtable_compare_arg *x = arg;
 711         struct rhashtable *ht = x->ht;
 712
 713         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 714 }
 715
 716 /**
 717  * rhashtable_lookup - lookup key in hash table
 718  * @ht:         hash table
 719  * @key:        pointer to key
 720  *
 721  * Computes the hash value for the key and traverses the bucket chain looking
 722  * for a entry with an identical key. The first matching entry is returned.
 723  *
 724  * This lookup function may only be used for fixed key hash table (key_len
 725  * parameter set). It will BUG() if used inappropriately.
 726  *
 727  * Lookups may occur in parallel with hashtable mutations and resizing.
 728  */
 729 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 730 {
 731         struct rhashtable_compare_arg arg = {
 732                 .ht = ht,
 733                 .key = key,
 734         };
 735
 736         BUG_ON(!ht->p.key_len);
 737
 738         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 739 }
 740 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 741
 742 /**
 743  * rhashtable_lookup_compare - search hash table with compare function
 744  * @ht:         hash table
 745  * @key:        the pointer to the key
 746  * @compare:    compare function, must return true on match
 747  * @arg:        argument passed on to compare function
 748  *
 749  * Traverses the bucket chain behind the provided hash value and calls the
 750  * specified compare function for each entry.
 751  *
 752  * Lookups may occur in parallel with hashtable mutations and resizing.
 753  *
 754  * Returns the first entry on which the compare function returned true.
 755  */
 756 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 757                                 bool (*compare)(void *, void *), void *arg)
 758 {
 759         const struct bucket_table *tbl, *old_tbl;
 760         struct rhash_head *he;
 761         u32 hash;
 762
 763         rcu_read_lock();
 764
 765         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 766         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 767         hash = key_hashfn(ht, key, ht->p.key_len);
 768 restart:
 769         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 770                 if (!compare(rht_obj(ht, he), arg))
 771                         continue;
 772                 rcu_read_unlock();
 773                 return rht_obj(ht, he);
 774         }
 775
 776         if (unlikely(tbl != old_tbl)) {
 777                 tbl = old_tbl;
 778                 goto restart;
 779         }
 780         rcu_read_unlock();
 781
 782         return NULL;
 783 }
 784 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 785
 786 /**
 787  * rhashtable_lookup_insert - lookup and insert object into hash table
 788  * @ht:         hash table
 789  * @obj:        pointer to hash head inside object
 790  *
 791  * Locks down the bucket chain in both the old and new table if a resize
 792  * is in progress to ensure that writers can't remove from the old table
 793  * and can't insert to the new table during the atomic operation of search
 794  * and insertion. Searches for duplicates in both the old and new table if
 795  * a resize is in progress.
 796  *
 797  * This lookup function may only be used for fixed key hash table (key_len
 798  * parameter set). It will BUG() if used inappropriately.
 799  *
 800  * It is safe to call this function from atomic context.
 801  *
 802  * Will trigger an automatic deferred table resizing if the size grows
 803  * beyond the watermark indicated by grow_decision() which can be passed
 804  * to rhashtable_init().
 805  */
 806 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 807 {
 808         struct rhashtable_compare_arg arg = {
 809                 .ht = ht,
 810                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 811         };
 812
 813         BUG_ON(!ht->p.key_len);
 814
 815         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 816                                                 &arg);
 817 }
 818 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 819
 820 /**
 821  * rhashtable_lookup_compare_insert - search and insert object to hash table
 822  *                                    with compare function
 823  * @ht:         hash table
 824  * @obj:        pointer to hash head inside object
 825  * @compare:    compare function, must return true on match
 826  * @arg:        argument passed on to compare function
 827  *
 828  * Locks down the bucket chain in both the old and new table if a resize
 829  * is in progress to ensure that writers can't remove from the old table
 830  * and can't insert to the new table during the atomic operation of search
 831  * and insertion. Searches for duplicates in both the old and new table if
 832  * a resize is in progress.
 833  *
 834  * Lookups may occur in parallel with hashtable mutations and resizing.
 835  *
 836  * Will trigger an automatic deferred table resizing if the size grows
 837  * beyond the watermark indicated by grow_decision() which can be passed
 838  * to rhashtable_init().
 839  */
 840 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 841                                       struct rhash_head *obj,
 842                                       bool (*compare)(void *, void *),
 843                                       void *arg)
 844 {
 845         struct bucket_table *new_tbl, *old_tbl;
 846         u32 new_hash;
 847         bool success = true;
 848
 849         BUG_ON(!ht->p.key_len);
 850
 851         rcu_read_lock();
 852         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 853         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 854         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 855
 856         lock_buckets(new_tbl, old_tbl, new_hash);
 857
 858         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 859                                       compare, arg)) {
 860                 success = false;
 861                 goto exit;
 862         }
 863
 864         __rhashtable_insert(ht, obj, new_tbl, new_hash);
 865
 866 exit:
 867         unlock_buckets(new_tbl, old_tbl, new_hash);
 868         rcu_read_unlock();
 869
 870         return success;
 871 }
 872 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 873
 874 /**
 875  * rhashtable_walk_init - Initialise an iterator
 876  * @ht:         Table to walk over
 877  * @iter:       Hash table Iterator
 878  *
 879  * This function prepares a hash table walk.
 880  *
 881  * Note that if you restart a walk after rhashtable_walk_stop you
 882  * may see the same object twice.  Also, you may miss objects if
 883  * there are removals in between rhashtable_walk_stop and the next
 884  * call to rhashtable_walk_start.
 885  *
 886  * For a completely stable walk you should construct your own data
 887  * structure outside the hash table.
 888  *
 889  * This function may sleep so you must not call it from interrupt
 890  * context or with spin locks held.
 891  *
 892  * You must call rhashtable_walk_exit if this function returns
 893  * successfully.
 894  */
 895 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 896 {
 897         iter->ht = ht;
 898         iter->p = NULL;
 899         iter->slot = 0;
 900         iter->skip = 0;
 901
 902         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 903         if (!iter->walker)
 904                 return -ENOMEM;
 905
 906         mutex_lock(&ht->mutex);
 907         list_add(&iter->walker->list, &ht->walkers);
 908         mutex_unlock(&ht->mutex);
 909
 910         return 0;
 911 }
 912 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 913
 914 /**
 915  * rhashtable_walk_exit - Free an iterator
 916  * @iter:       Hash table Iterator
 917  *
 918  * This function frees resources allocated by rhashtable_walk_init.
 919  */
 920 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 921 {
 922         mutex_lock(&iter->ht->mutex);
 923         list_del(&iter->walker->list);
 924         mutex_unlock(&iter->ht->mutex);
 925         kfree(iter->walker);
 926 }
 927 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 928
 929 /**
 930  * rhashtable_walk_start - Start a hash table walk
 931  * @iter:       Hash table iterator
 932  *
 933  * Start a hash table walk.  Note that we take the RCU lock in all
 934  * cases including when we return an error.  So you must always call
 935  * rhashtable_walk_stop to clean up.
 936  *
 937  * Returns zero if successful.
 938  *
 939  * Returns -EAGAIN if resize event occured.  Note that the iterator
 940  * will rewind back to the beginning and you may use it immediately
 941  * by calling rhashtable_walk_next.
 942  */
 943 int rhashtable_walk_start(struct rhashtable_iter *iter)
 944 {
 945         rcu_read_lock();
 946
 947         if (iter->walker->resize) {
 948                 iter->slot = 0;
 949                 iter->skip = 0;
 950                 iter->walker->resize = false;
 951                 return -EAGAIN;
 952         }
 953
 954         return 0;
 955 }
 956 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 957
 958 /**
 959  * rhashtable_walk_next - Return the next object and advance the iterator
 960  * @iter:       Hash table iterator
 961  *
 962  * Note that you must call rhashtable_walk_stop when you are finished
 963  * with the walk.
 964  *
 965  * Returns the next object or NULL when the end of the table is reached.
 966  *
 967  * Returns -EAGAIN if resize event occured.  Note that the iterator
 968  * will rewind back to the beginning and you may continue to use it.
 969  */
 970 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 971 {
 972         const struct bucket_table *tbl;
 973         struct rhashtable *ht = iter->ht;
 974         struct rhash_head *p = iter->p;
 975         void *obj = NULL;
 976
 977         tbl = rht_dereference_rcu(ht->tbl, ht);
 978
 979         if (p) {
 980                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 981                 goto next;
 982         }
 983
 984         for (; iter->slot < tbl->size; iter->slot++) {
 985                 int skip = iter->skip;
 986
 987                 rht_for_each_rcu(p, tbl, iter->slot) {
 988                         if (!skip)
 989                                 break;
 990                         skip--;
 991                 }
 992
 993 next:
 994                 if (!rht_is_a_nulls(p)) {
 995                         iter->skip++;
 996                         iter->p = p;
 997                         obj = rht_obj(ht, p);
 998                         goto out;
 999                 }
1000
1001                 iter->skip = 0;
1002         }
1003
1004         iter->p = NULL;
1005
1006 out:
1007         if (iter->walker->resize) {
1008                 iter->p = NULL;
1009                 iter->slot = 0;
1010                 iter->skip = 0;
1011                 iter->walker->resize = false;
1012                 return ERR_PTR(-EAGAIN);
1013         }
1014
1015         return obj;
1016 }
1017 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
1018
1019 /**
1020  * rhashtable_walk_stop - Finish a hash table walk
1021  * @iter:       Hash table iterator
1022  *
1023  * Finish a hash table walk.
1024  */
1025 void rhashtable_walk_stop(struct rhashtable_iter *iter)
1026 {
1027         rcu_read_unlock();
1028         iter->p = NULL;
1029 }
1030 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
1031
1032 static size_t rounded_hashtable_size(struct rhashtable_params *params)
1033 {
1034         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
1035                    1UL << params->min_shift);
1036 }
1037
1038 /**
1039  * rhashtable_init - initialize a new hash table
1040  * @ht:         hash table to be initialized
1041  * @params:     configuration parameters
1042  *
1043  * Initializes a new hash table based on the provided configuration
1044  * parameters. A table can be configured either with a variable or
1045  * fixed length key:
1046  *
1047  * Configuration Example 1: Fixed length keys
1048  * struct test_obj {
1049  *      int                     key;
1050  *      void *                  my_member;
1051  *      struct rhash_head       node;
1052  * };
1053  *
1054  * struct rhashtable_params params = {
1055  *      .head_offset = offsetof(struct test_obj, node),
1056  *      .key_offset = offsetof(struct test_obj, key),
1057  *      .key_len = sizeof(int),
1058  *      .hashfn = jhash,
1059  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1060  * };
1061  *
1062  * Configuration Example 2: Variable length keys
1063  * struct test_obj {
1064  *      [...]
1065  *      struct rhash_head       node;
1066  * };
1067  *
1068  * u32 my_hash_fn(const void *data, u32 seed)
1069  * {
1070  *      struct test_obj *obj = data;
1071  *
1072  *      return [... hash ...];
1073  * }
1074  *
1075  * struct rhashtable_params params = {
1076  *      .head_offset = offsetof(struct test_obj, node),
1077  *      .hashfn = jhash,
1078  *      .obj_hashfn = my_hash_fn,
1079  * };
1080  */
1081 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1082 {
1083         struct bucket_table *tbl;
1084         size_t size;
1085
1086         size = HASH_DEFAULT_SIZE;
1087
1088         if ((params->key_len && !params->hashfn) ||
1089             (!params->key_len && !params->obj_hashfn))
1090                 return -EINVAL;
1091
1092         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1093                 return -EINVAL;
1094
1095         params->min_shift = max_t(size_t, params->min_shift,
1096                                   ilog2(HASH_MIN_SIZE));
1097
1098         if (params->nelem_hint)
1099                 size = rounded_hashtable_size(params);
1100
1101         memset(ht, 0, sizeof(*ht));
1102         mutex_init(&ht->mutex);
1103         memcpy(&ht->p, params, sizeof(*params));
1104         INIT_LIST_HEAD(&ht->walkers);
1105
1106         if (params->locks_mul)
1107                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1108         else
1109                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1110
1111         tbl = bucket_table_alloc(ht, size);
1112         if (tbl == NULL)
1113                 return -ENOMEM;
1114
1115         atomic_set(&ht->nelems, 0);
1116         atomic_set(&ht->shift, ilog2(tbl->size));
1117         RCU_INIT_POINTER(ht->tbl, tbl);
1118         RCU_INIT_POINTER(ht->future_tbl, tbl);
1119
1120         if (!ht->p.hash_rnd)
1121                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1122
1123         if (ht->p.grow_decision || ht->p.shrink_decision)
1124                 INIT_WORK(&ht->run_work, rht_deferred_worker);
1125
1126         return 0;
1127 }
1128 EXPORT_SYMBOL_GPL(rhashtable_init);
1129
1130 /**
1131  * rhashtable_destroy - destroy hash table
1132  * @ht:         the hash table to destroy
1133  *
1134  * Frees the bucket array. This function is not rcu safe, therefore the caller
1135  * has to make sure that no resizing may happen by unpublishing the hashtable
1136  * and waiting for the quiescent cycle before releasing the bucket array.
1137  */
1138 void rhashtable_destroy(struct rhashtable *ht)
1139 {
1140         ht->being_destroyed = true;
1141
1142         if (ht->p.grow_decision || ht->p.shrink_decision)
1143                 cancel_work_sync(&ht->run_work);
1144
1145         mutex_lock(&ht->mutex);
1146         bucket_table_free(rht_dereference(ht->tbl, ht));
1147         mutex_unlock(&ht->mutex);
1148 }
1149 EXPORT_SYMBOL_GPL(rhashtable_destroy);