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/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/vmalloc.h>
  23 #include <linux/mm.h>
  24 #include <linux/jhash.h>
  25 #include <linux/random.h>
  26 #include <linux/rhashtable.h>
  27 #include <linux/err.h>
  28
  29 #define HASH_DEFAULT_SIZE       64UL
  30 #define HASH_MIN_SIZE           4UL
  31 #define BUCKET_LOCKS_PER_CPU   128UL
  32
  33 /* Base bits plus 1 bit for nulls marker */
  34 #define HASH_RESERVED_SPACE     (RHT_BASE_BITS + 1)
  35
  36 enum {
  37         RHT_LOCK_NORMAL,
  38         RHT_LOCK_NESTED,
  39 };
  40
  41 /* The bucket lock is selected based on the hash and protects mutations
  42  * on a group of hash buckets.
  43  *
  44  * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
  45  * a single lock always covers both buckets which may both contains
  46  * entries which link to the same bucket of the old table during resizing.
  47  * This allows to simplify the locking as locking the bucket in both
  48  * tables during resize always guarantee protection.
  49  *
  50  * IMPORTANT: When holding the bucket lock of both the old and new table
  51  * during expansions and shrinking, the old bucket lock must always be
  52  * acquired first.
  53  */
  54 static spinlock_t *bucket_lock(const struct bucket_table *tbl, u32 hash)
  55 {
  56         return &tbl->locks[hash & tbl->locks_mask];
  57 }
  58
  59 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
  60 {
  61         return (void *) he - ht->p.head_offset;
  62 }
  63
  64 static u32 rht_bucket_index(const struct bucket_table *tbl, u32 hash)
  65 {
  66         return hash & (tbl->size - 1);
  67 }
  68
  69 static u32 obj_raw_hashfn(const struct rhashtable *ht, const void *ptr)
  70 {
  71         u32 hash;
  72
  73         if (unlikely(!ht->p.key_len))
  74                 hash = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
  75         else
  76                 hash = ht->p.hashfn(ptr + ht->p.key_offset, ht->p.key_len,
  77                                     ht->p.hash_rnd);
  78
  79         return hash >> HASH_RESERVED_SPACE;
  80 }
  81
  82 static u32 key_hashfn(struct rhashtable *ht, const void *key, u32 len)
  83 {
  84         return ht->p.hashfn(key, len, ht->p.hash_rnd) >> HASH_RESERVED_SPACE;
  85 }
  86
  87 static u32 head_hashfn(const struct rhashtable *ht,
  88                        const struct bucket_table *tbl,
  89                        const struct rhash_head *he)
  90 {
  91         return rht_bucket_index(tbl, obj_raw_hashfn(ht, rht_obj(ht, he)));
  92 }
  93
  94 #ifdef CONFIG_PROVE_LOCKING
  95 static void debug_dump_buckets(const struct rhashtable *ht,
  96                                const struct bucket_table *tbl)
  97 {
  98         struct rhash_head *he;
  99         unsigned int i, hash;
 100
 101         for (i = 0; i < tbl->size; i++) {
 102                 pr_warn(" [Bucket %d] ", i);
 103                 rht_for_each_rcu(he, tbl, i) {
 104                         hash = head_hashfn(ht, tbl, he);
 105                         pr_cont("[hash = %#x, lock = %p] ",
 106                                 hash, bucket_lock(tbl, hash));
 107                 }
 108                 pr_cont("\n");
 109         }
 110
 111 }
 112
 113 static void debug_dump_table(struct rhashtable *ht,
 114                              const struct bucket_table *tbl,
 115                              unsigned int hash)
 116 {
 117         struct bucket_table *old_tbl, *future_tbl;
 118
 119         pr_emerg("BUG: lock for hash %#x in table %p not held\n",
 120                  hash, tbl);
 121
 122         rcu_read_lock();
 123         future_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 124         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 125         if (future_tbl != old_tbl) {
 126                 pr_warn("Future table %p (size: %zd)\n",
 127                         future_tbl, future_tbl->size);
 128                 debug_dump_buckets(ht, future_tbl);
 129         }
 130
 131         pr_warn("Table %p (size: %zd)\n", old_tbl, old_tbl->size);
 132         debug_dump_buckets(ht, old_tbl);
 133
 134         rcu_read_unlock();
 135 }
 136
 137 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
 138 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)                               \
 139         do {                                                            \
 140                 if (unlikely(!lockdep_rht_bucket_is_held(TBL, HASH))) { \
 141                         debug_dump_table(HT, TBL, HASH);                \
 142                         BUG();                                          \
 143                 }                                                       \
 144         } while (0)
 145
 146 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
 147 {
 148         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
 149 }
 150 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
 151
 152 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
 153 {
 154         spinlock_t *lock = bucket_lock(tbl, hash);
 155
 156         return (debug_locks) ? lockdep_is_held(lock) : 1;
 157 }
 158 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
 159 #else
 160 #define ASSERT_RHT_MUTEX(HT)
 161 #define ASSERT_BUCKET_LOCK(HT, TBL, HASH)
 162 #endif
 163
 164
 165 static struct rhash_head __rcu **bucket_tail(struct bucket_table *tbl, u32 n)
 166 {
 167         struct rhash_head __rcu **pprev;
 168
 169         for (pprev = &tbl->buckets[n];
 170              !rht_is_a_nulls(rht_dereference_bucket(*pprev, tbl, n));
 171              pprev = &rht_dereference_bucket(*pprev, tbl, n)->next)
 172                 ;
 173
 174         return pprev;
 175 }
 176
 177 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl)
 178 {
 179         unsigned int i, size;
 180 #if defined(CONFIG_PROVE_LOCKING)
 181         unsigned int nr_pcpus = 2;
 182 #else
 183         unsigned int nr_pcpus = num_possible_cpus();
 184 #endif
 185
 186         nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
 187         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
 188
 189         /* Never allocate more than 0.5 locks per bucket */
 190         size = min_t(unsigned int, size, tbl->size >> 1);
 191
 192         if (sizeof(spinlock_t) != 0) {
 193 #ifdef CONFIG_NUMA
 194                 if (size * sizeof(spinlock_t) > PAGE_SIZE)
 195                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
 196                 else
 197 #endif
 198                 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
 199                                            GFP_KERNEL);
 200                 if (!tbl->locks)
 201                         return -ENOMEM;
 202                 for (i = 0; i < size; i++)
 203                         spin_lock_init(&tbl->locks[i]);
 204         }
 205         tbl->locks_mask = size - 1;
 206
 207         return 0;
 208 }
 209
 210 static void bucket_table_free(const struct bucket_table *tbl)
 211 {
 212         if (tbl)
 213                 kvfree(tbl->locks);
 214
 215         kvfree(tbl);
 216 }
 217
 218 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 219                                                size_t nbuckets)
 220 {
 221         struct bucket_table *tbl = NULL;
 222         size_t size;
 223         int i;
 224
 225         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 226         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER))
 227                 tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
 228         if (tbl == NULL)
 229                 tbl = vzalloc(size);
 230         if (tbl == NULL)
 231                 return NULL;
 232
 233         tbl->size = nbuckets;
 234
 235         if (alloc_bucket_locks(ht, tbl) < 0) {
 236                 bucket_table_free(tbl);
 237                 return NULL;
 238         }
 239
 240         for (i = 0; i < nbuckets; i++)
 241                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 242
 243         return tbl;
 244 }
 245
 246 /**
 247  * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
 248  * @ht:         hash table
 249  * @new_size:   new table size
 250  */
 251 static bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
 252 {
 253         /* Expand table when exceeding 75% load */
 254         return atomic_read(&ht->nelems) > (new_size / 4 * 3) &&
 255                (!ht->p.max_shift || atomic_read(&ht->shift) < ht->p.max_shift);
 256 }
 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 static 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
 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                 cond_resched();
 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                         cond_resched();
 441                 }
 442         }
 443
 444         rcu_assign_pointer(ht->tbl, new_tbl);
 445         synchronize_rcu();
 446
 447         bucket_table_free(old_tbl);
 448         return 0;
 449 }
 450 EXPORT_SYMBOL_GPL(rhashtable_expand);
 451
 452 /**
 453  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 454  * @ht:         the hash table to shrink
 455  *
 456  * This function may only be called in a context where it is safe to call
 457  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 458  *
 459  * The caller must ensure that no concurrent resizing occurs by holding
 460  * ht->mutex.
 461  *
 462  * The caller must ensure that no concurrent table mutations take place.
 463  * It is however valid to have concurrent lookups if they are RCU protected.
 464  *
 465  * It is valid to have concurrent insertions and deletions protected by per
 466  * bucket locks or concurrent RCU protected lookups and traversals.
 467  */
 468 int rhashtable_shrink(struct rhashtable *ht)
 469 {
 470         struct bucket_table *new_tbl, *tbl = rht_dereference(ht->tbl, ht);
 471         unsigned int new_hash;
 472
 473         ASSERT_RHT_MUTEX(ht);
 474
 475         new_tbl = bucket_table_alloc(ht, tbl->size / 2);
 476         if (new_tbl == NULL)
 477                 return -ENOMEM;
 478
 479         rcu_assign_pointer(ht->future_tbl, new_tbl);
 480         synchronize_rcu();
 481
 482         /* Link the first entry in the old bucket to the end of the
 483          * bucket in the new table. As entries are concurrently being
 484          * added to the new table, lock down the new bucket. As we
 485          * always divide the size in half when shrinking, each bucket
 486          * in the new table maps to exactly two buckets in the old
 487          * table.
 488          */
 489         for (new_hash = 0; new_hash < new_tbl->size; new_hash++) {
 490                 lock_buckets(new_tbl, tbl, new_hash);
 491
 492                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 493                                    tbl->buckets[new_hash]);
 494                 ASSERT_BUCKET_LOCK(ht, tbl, new_hash + new_tbl->size);
 495                 rcu_assign_pointer(*bucket_tail(new_tbl, new_hash),
 496                                    tbl->buckets[new_hash + new_tbl->size]);
 497
 498                 unlock_buckets(new_tbl, tbl, new_hash);
 499                 cond_resched();
 500         }
 501
 502         /* Publish the new, valid hash table */
 503         rcu_assign_pointer(ht->tbl, new_tbl);
 504         atomic_dec(&ht->shift);
 505
 506         /* Wait for readers. No new readers will have references to the
 507          * old hash table.
 508          */
 509         synchronize_rcu();
 510
 511         bucket_table_free(tbl);
 512
 513         return 0;
 514 }
 515 EXPORT_SYMBOL_GPL(rhashtable_shrink);
 516
 517 static void rht_deferred_worker(struct work_struct *work)
 518 {
 519         struct rhashtable *ht;
 520         struct bucket_table *tbl;
 521         struct rhashtable_walker *walker;
 522
 523         ht = container_of(work, struct rhashtable, run_work);
 524         mutex_lock(&ht->mutex);
 525         if (ht->being_destroyed)
 526                 goto unlock;
 527
 528         tbl = rht_dereference(ht->tbl, ht);
 529
 530         list_for_each_entry(walker, &ht->walkers, list)
 531                 walker->resize = true;
 532
 533         if (rht_grow_above_75(ht, tbl->size))
 534                 rhashtable_expand(ht);
 535         else if (rht_shrink_below_30(ht, tbl->size))
 536                 rhashtable_shrink(ht);
 537 unlock:
 538         mutex_unlock(&ht->mutex);
 539 }
 540
 541 static void __rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj,
 542                                 struct bucket_table *tbl,
 543                                 const struct bucket_table *old_tbl, u32 hash)
 544 {
 545         bool no_resize_running = tbl == old_tbl;
 546         struct rhash_head *head;
 547
 548         hash = rht_bucket_index(tbl, hash);
 549         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 550
 551         ASSERT_BUCKET_LOCK(ht, tbl, hash);
 552
 553         if (rht_is_a_nulls(head))
 554                 INIT_RHT_NULLS_HEAD(obj->next, ht, hash);
 555         else
 556                 RCU_INIT_POINTER(obj->next, head);
 557
 558         rcu_assign_pointer(tbl->buckets[hash], obj);
 559
 560         atomic_inc(&ht->nelems);
 561         if (no_resize_running && rht_grow_above_75(ht, tbl->size))
 562                 schedule_work(&ht->run_work);
 563 }
 564
 565 /**
 566  * rhashtable_insert - insert object into hash table
 567  * @ht:         hash table
 568  * @obj:        pointer to hash head inside object
 569  *
 570  * Will take a per bucket spinlock to protect against mutual mutations
 571  * on the same bucket. Multiple insertions may occur in parallel unless
 572  * they map to the same bucket lock.
 573  *
 574  * It is safe to call this function from atomic context.
 575  *
 576  * Will trigger an automatic deferred table resizing if the size grows
 577  * beyond the watermark indicated by grow_decision() which can be passed
 578  * to rhashtable_init().
 579  */
 580 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
 581 {
 582         struct bucket_table *tbl, *old_tbl;
 583         unsigned hash;
 584
 585         rcu_read_lock();
 586
 587         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 588         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 589         hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 590
 591         lock_buckets(tbl, old_tbl, hash);
 592         __rhashtable_insert(ht, obj, tbl, old_tbl, hash);
 593         unlock_buckets(tbl, old_tbl, hash);
 594
 595         rcu_read_unlock();
 596 }
 597 EXPORT_SYMBOL_GPL(rhashtable_insert);
 598
 599 /**
 600  * rhashtable_remove - remove object from hash table
 601  * @ht:         hash table
 602  * @obj:        pointer to hash head inside object
 603  *
 604  * Since the hash chain is single linked, the removal operation needs to
 605  * walk the bucket chain upon removal. The removal operation is thus
 606  * considerable slow if the hash table is not correctly sized.
 607  *
 608  * Will automatically shrink the table via rhashtable_expand() if the
 609  * shrink_decision function specified at rhashtable_init() returns true.
 610  *
 611  * The caller must ensure that no concurrent table mutations occur. It is
 612  * however valid to have concurrent lookups if they are RCU protected.
 613  */
 614 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
 615 {
 616         struct bucket_table *tbl, *new_tbl, *old_tbl;
 617         struct rhash_head __rcu **pprev;
 618         struct rhash_head *he, *he2;
 619         unsigned int hash, new_hash;
 620         bool ret = false;
 621
 622         rcu_read_lock();
 623         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 624         tbl = new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 625         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 626
 627         lock_buckets(new_tbl, old_tbl, new_hash);
 628 restart:
 629         hash = rht_bucket_index(tbl, new_hash);
 630         pprev = &tbl->buckets[hash];
 631         rht_for_each(he, tbl, hash) {
 632                 if (he != obj) {
 633                         pprev = &he->next;
 634                         continue;
 635                 }
 636
 637                 ASSERT_BUCKET_LOCK(ht, tbl, hash);
 638
 639                 if (old_tbl->size > new_tbl->size && tbl == old_tbl &&
 640                     !rht_is_a_nulls(obj->next) &&
 641                     head_hashfn(ht, tbl, obj->next) != hash) {
 642                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 643                 } else if (unlikely(old_tbl->size < new_tbl->size && tbl == new_tbl)) {
 644                         rht_for_each_continue(he2, obj->next, tbl, hash) {
 645                                 if (head_hashfn(ht, tbl, he2) == hash) {
 646                                         rcu_assign_pointer(*pprev, he2);
 647                                         goto found;
 648                                 }
 649                         }
 650
 651                         rcu_assign_pointer(*pprev, (struct rhash_head *) rht_marker(ht, hash));
 652                 } else {
 653                         rcu_assign_pointer(*pprev, obj->next);
 654                 }
 655
 656 found:
 657                 ret = true;
 658                 break;
 659         }
 660
 661         /* The entry may be linked in either 'tbl', 'future_tbl', or both.
 662          * 'future_tbl' only exists for a short period of time during
 663          * resizing. Thus traversing both is fine and the added cost is
 664          * very rare.
 665          */
 666         if (tbl != old_tbl) {
 667                 tbl = old_tbl;
 668                 goto restart;
 669         }
 670
 671         unlock_buckets(new_tbl, old_tbl, new_hash);
 672
 673         if (ret) {
 674                 bool no_resize_running = new_tbl == old_tbl;
 675
 676                 atomic_dec(&ht->nelems);
 677                 if (no_resize_running && rht_shrink_below_30(ht, new_tbl->size))
 678                         schedule_work(&ht->run_work);
 679         }
 680
 681         rcu_read_unlock();
 682
 683         return ret;
 684 }
 685 EXPORT_SYMBOL_GPL(rhashtable_remove);
 686
 687 struct rhashtable_compare_arg {
 688         struct rhashtable *ht;
 689         const void *key;
 690 };
 691
 692 static bool rhashtable_compare(void *ptr, void *arg)
 693 {
 694         struct rhashtable_compare_arg *x = arg;
 695         struct rhashtable *ht = x->ht;
 696
 697         return !memcmp(ptr + ht->p.key_offset, x->key, ht->p.key_len);
 698 }
 699
 700 /**
 701  * rhashtable_lookup - lookup key in hash table
 702  * @ht:         hash table
 703  * @key:        pointer to key
 704  *
 705  * Computes the hash value for the key and traverses the bucket chain looking
 706  * for a entry with an identical key. The first matching entry is returned.
 707  *
 708  * This lookup function may only be used for fixed key hash table (key_len
 709  * parameter set). It will BUG() if used inappropriately.
 710  *
 711  * Lookups may occur in parallel with hashtable mutations and resizing.
 712  */
 713 void *rhashtable_lookup(struct rhashtable *ht, const void *key)
 714 {
 715         struct rhashtable_compare_arg arg = {
 716                 .ht = ht,
 717                 .key = key,
 718         };
 719
 720         BUG_ON(!ht->p.key_len);
 721
 722         return rhashtable_lookup_compare(ht, key, &rhashtable_compare, &arg);
 723 }
 724 EXPORT_SYMBOL_GPL(rhashtable_lookup);
 725
 726 /**
 727  * rhashtable_lookup_compare - search hash table with compare function
 728  * @ht:         hash table
 729  * @key:        the pointer to the key
 730  * @compare:    compare function, must return true on match
 731  * @arg:        argument passed on to compare function
 732  *
 733  * Traverses the bucket chain behind the provided hash value and calls the
 734  * specified compare function for each entry.
 735  *
 736  * Lookups may occur in parallel with hashtable mutations and resizing.
 737  *
 738  * Returns the first entry on which the compare function returned true.
 739  */
 740 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
 741                                 bool (*compare)(void *, void *), void *arg)
 742 {
 743         const struct bucket_table *tbl, *old_tbl;
 744         struct rhash_head *he;
 745         u32 hash;
 746
 747         rcu_read_lock();
 748
 749         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 750         tbl = rht_dereference_rcu(ht->future_tbl, ht);
 751         hash = key_hashfn(ht, key, ht->p.key_len);
 752 restart:
 753         rht_for_each_rcu(he, tbl, rht_bucket_index(tbl, hash)) {
 754                 if (!compare(rht_obj(ht, he), arg))
 755                         continue;
 756                 rcu_read_unlock();
 757                 return rht_obj(ht, he);
 758         }
 759
 760         if (unlikely(tbl != old_tbl)) {
 761                 tbl = old_tbl;
 762                 goto restart;
 763         }
 764         rcu_read_unlock();
 765
 766         return NULL;
 767 }
 768 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
 769
 770 /**
 771  * rhashtable_lookup_insert - lookup and insert object into hash table
 772  * @ht:         hash table
 773  * @obj:        pointer to hash head inside object
 774  *
 775  * Locks down the bucket chain in both the old and new table if a resize
 776  * is in progress to ensure that writers can't remove from the old table
 777  * and can't insert to the new table during the atomic operation of search
 778  * and insertion. Searches for duplicates in both the old and new table if
 779  * a resize is in progress.
 780  *
 781  * This lookup function may only be used for fixed key hash table (key_len
 782  * parameter set). It will BUG() if used inappropriately.
 783  *
 784  * It is safe to call this function from atomic context.
 785  *
 786  * Will trigger an automatic deferred table resizing if the size grows
 787  * beyond the watermark indicated by grow_decision() which can be passed
 788  * to rhashtable_init().
 789  */
 790 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj)
 791 {
 792         struct rhashtable_compare_arg arg = {
 793                 .ht = ht,
 794                 .key = rht_obj(ht, obj) + ht->p.key_offset,
 795         };
 796
 797         BUG_ON(!ht->p.key_len);
 798
 799         return rhashtable_lookup_compare_insert(ht, obj, &rhashtable_compare,
 800                                                 &arg);
 801 }
 802 EXPORT_SYMBOL_GPL(rhashtable_lookup_insert);
 803
 804 /**
 805  * rhashtable_lookup_compare_insert - search and insert object to hash table
 806  *                                    with compare function
 807  * @ht:         hash table
 808  * @obj:        pointer to hash head inside object
 809  * @compare:    compare function, must return true on match
 810  * @arg:        argument passed on to compare function
 811  *
 812  * Locks down the bucket chain in both the old and new table if a resize
 813  * is in progress to ensure that writers can't remove from the old table
 814  * and can't insert to the new table during the atomic operation of search
 815  * and insertion. Searches for duplicates in both the old and new table if
 816  * a resize is in progress.
 817  *
 818  * Lookups may occur in parallel with hashtable mutations and resizing.
 819  *
 820  * Will trigger an automatic deferred table resizing if the size grows
 821  * beyond the watermark indicated by grow_decision() which can be passed
 822  * to rhashtable_init().
 823  */
 824 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
 825                                       struct rhash_head *obj,
 826                                       bool (*compare)(void *, void *),
 827                                       void *arg)
 828 {
 829         struct bucket_table *new_tbl, *old_tbl;
 830         u32 new_hash;
 831         bool success = true;
 832
 833         BUG_ON(!ht->p.key_len);
 834
 835         rcu_read_lock();
 836         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 837         new_tbl = rht_dereference_rcu(ht->future_tbl, ht);
 838         new_hash = obj_raw_hashfn(ht, rht_obj(ht, obj));
 839
 840         lock_buckets(new_tbl, old_tbl, new_hash);
 841
 842         if (rhashtable_lookup_compare(ht, rht_obj(ht, obj) + ht->p.key_offset,
 843                                       compare, arg)) {
 844                 success = false;
 845                 goto exit;
 846         }
 847
 848         __rhashtable_insert(ht, obj, new_tbl, old_tbl, new_hash);
 849
 850 exit:
 851         unlock_buckets(new_tbl, old_tbl, new_hash);
 852         rcu_read_unlock();
 853
 854         return success;
 855 }
 856 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare_insert);
 857
 858 /**
 859  * rhashtable_walk_init - Initialise an iterator
 860  * @ht:         Table to walk over
 861  * @iter:       Hash table Iterator
 862  *
 863  * This function prepares a hash table walk.
 864  *
 865  * Note that if you restart a walk after rhashtable_walk_stop you
 866  * may see the same object twice.  Also, you may miss objects if
 867  * there are removals in between rhashtable_walk_stop and the next
 868  * call to rhashtable_walk_start.
 869  *
 870  * For a completely stable walk you should construct your own data
 871  * structure outside the hash table.
 872  *
 873  * This function may sleep so you must not call it from interrupt
 874  * context or with spin locks held.
 875  *
 876  * You must call rhashtable_walk_exit if this function returns
 877  * successfully.
 878  */
 879 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
 880 {
 881         iter->ht = ht;
 882         iter->p = NULL;
 883         iter->slot = 0;
 884         iter->skip = 0;
 885
 886         iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
 887         if (!iter->walker)
 888                 return -ENOMEM;
 889
 890         INIT_LIST_HEAD(&iter->walker->list);
 891         iter->walker->resize = false;
 892
 893         mutex_lock(&ht->mutex);
 894         list_add(&iter->walker->list, &ht->walkers);
 895         mutex_unlock(&ht->mutex);
 896
 897         return 0;
 898 }
 899 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
 900
 901 /**
 902  * rhashtable_walk_exit - Free an iterator
 903  * @iter:       Hash table Iterator
 904  *
 905  * This function frees resources allocated by rhashtable_walk_init.
 906  */
 907 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 908 {
 909         mutex_lock(&iter->ht->mutex);
 910         list_del(&iter->walker->list);
 911         mutex_unlock(&iter->ht->mutex);
 912         kfree(iter->walker);
 913 }
 914 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 915
 916 /**
 917  * rhashtable_walk_start - Start a hash table walk
 918  * @iter:       Hash table iterator
 919  *
 920  * Start a hash table walk.  Note that we take the RCU lock in all
 921  * cases including when we return an error.  So you must always call
 922  * rhashtable_walk_stop to clean up.
 923  *
 924  * Returns zero if successful.
 925  *
 926  * Returns -EAGAIN if resize event occured.  Note that the iterator
 927  * will rewind back to the beginning and you may use it immediately
 928  * by calling rhashtable_walk_next.
 929  */
 930 int rhashtable_walk_start(struct rhashtable_iter *iter)
 931 {
 932         rcu_read_lock();
 933
 934         if (iter->walker->resize) {
 935                 iter->slot = 0;
 936                 iter->skip = 0;
 937                 iter->walker->resize = false;
 938                 return -EAGAIN;
 939         }
 940
 941         return 0;
 942 }
 943 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 944
 945 /**
 946  * rhashtable_walk_next - Return the next object and advance the iterator
 947  * @iter:       Hash table iterator
 948  *
 949  * Note that you must call rhashtable_walk_stop when you are finished
 950  * with the walk.
 951  *
 952  * Returns the next object or NULL when the end of the table is reached.
 953  *
 954  * Returns -EAGAIN if resize event occured.  Note that the iterator
 955  * will rewind back to the beginning and you may continue to use it.
 956  */
 957 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 958 {
 959         const struct bucket_table *tbl;
 960         struct rhashtable *ht = iter->ht;
 961         struct rhash_head *p = iter->p;
 962         void *obj = NULL;
 963
 964         tbl = rht_dereference_rcu(ht->tbl, ht);
 965
 966         if (p) {
 967                 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
 968                 goto next;
 969         }
 970
 971         for (; iter->slot < tbl->size; iter->slot++) {
 972                 int skip = iter->skip;
 973
 974                 rht_for_each_rcu(p, tbl, iter->slot) {
 975                         if (!skip)
 976                                 break;
 977                         skip--;
 978                 }
 979
 980 next:
 981                 if (!rht_is_a_nulls(p)) {
 982                         iter->skip++;
 983                         iter->p = p;
 984                         obj = rht_obj(ht, p);
 985                         goto out;
 986                 }
 987
 988                 iter->skip = 0;
 989         }
 990
 991         iter->p = NULL;
 992
 993 out:
 994         if (iter->walker->resize) {
 995                 iter->p = NULL;
 996                 iter->slot = 0;
 997                 iter->skip = 0;
 998                 iter->walker->resize = false;
 999                 return ERR_PTR(-EAGAIN);
1000         }
1001
1002         return obj;
1003 }
1004 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
1005
1006 /**
1007  * rhashtable_walk_stop - Finish a hash table walk
1008  * @iter:       Hash table iterator
1009  *
1010  * Finish a hash table walk.
1011  */
1012 void rhashtable_walk_stop(struct rhashtable_iter *iter)
1013 {
1014         rcu_read_unlock();
1015         iter->p = NULL;
1016 }
1017 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
1018
1019 static size_t rounded_hashtable_size(struct rhashtable_params *params)
1020 {
1021         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
1022                    1UL << params->min_shift);
1023 }
1024
1025 /**
1026  * rhashtable_init - initialize a new hash table
1027  * @ht:         hash table to be initialized
1028  * @params:     configuration parameters
1029  *
1030  * Initializes a new hash table based on the provided configuration
1031  * parameters. A table can be configured either with a variable or
1032  * fixed length key:
1033  *
1034  * Configuration Example 1: Fixed length keys
1035  * struct test_obj {
1036  *      int                     key;
1037  *      void *                  my_member;
1038  *      struct rhash_head       node;
1039  * };
1040  *
1041  * struct rhashtable_params params = {
1042  *      .head_offset = offsetof(struct test_obj, node),
1043  *      .key_offset = offsetof(struct test_obj, key),
1044  *      .key_len = sizeof(int),
1045  *      .hashfn = jhash,
1046  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1047  * };
1048  *
1049  * Configuration Example 2: Variable length keys
1050  * struct test_obj {
1051  *      [...]
1052  *      struct rhash_head       node;
1053  * };
1054  *
1055  * u32 my_hash_fn(const void *data, u32 seed)
1056  * {
1057  *      struct test_obj *obj = data;
1058  *
1059  *      return [... hash ...];
1060  * }
1061  *
1062  * struct rhashtable_params params = {
1063  *      .head_offset = offsetof(struct test_obj, node),
1064  *      .hashfn = jhash,
1065  *      .obj_hashfn = my_hash_fn,
1066  * };
1067  */
1068 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
1069 {
1070         struct bucket_table *tbl;
1071         size_t size;
1072
1073         size = HASH_DEFAULT_SIZE;
1074
1075         if ((params->key_len && !params->hashfn) ||
1076             (!params->key_len && !params->obj_hashfn))
1077                 return -EINVAL;
1078
1079         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1080                 return -EINVAL;
1081
1082         params->min_shift = max_t(size_t, params->min_shift,
1083                                   ilog2(HASH_MIN_SIZE));
1084
1085         if (params->nelem_hint)
1086                 size = rounded_hashtable_size(params);
1087
1088         memset(ht, 0, sizeof(*ht));
1089         mutex_init(&ht->mutex);
1090         memcpy(&ht->p, params, sizeof(*params));
1091         INIT_LIST_HEAD(&ht->walkers);
1092
1093         if (params->locks_mul)
1094                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1095         else
1096                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1097
1098         tbl = bucket_table_alloc(ht, size);
1099         if (tbl == NULL)
1100                 return -ENOMEM;
1101
1102         atomic_set(&ht->nelems, 0);
1103         atomic_set(&ht->shift, ilog2(tbl->size));
1104         RCU_INIT_POINTER(ht->tbl, tbl);
1105         RCU_INIT_POINTER(ht->future_tbl, tbl);
1106
1107         if (!ht->p.hash_rnd)
1108                 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
1109
1110         INIT_WORK(&ht->run_work, rht_deferred_worker);
1111
1112         return 0;
1113 }
1114 EXPORT_SYMBOL_GPL(rhashtable_init);
1115
1116 /**
1117  * rhashtable_destroy - destroy hash table
1118  * @ht:         the hash table to destroy
1119  *
1120  * Frees the bucket array. This function is not rcu safe, therefore the caller
1121  * has to make sure that no resizing may happen by unpublishing the hashtable
1122  * and waiting for the quiescent cycle before releasing the bucket array.
1123  */
1124 void rhashtable_destroy(struct rhashtable *ht)
1125 {
1126         ht->being_destroyed = true;
1127
1128         cancel_work_sync(&ht->run_work);
1129
1130         mutex_lock(&ht->mutex);
1131         bucket_table_free(rht_dereference(ht->tbl, ht));
1132         mutex_unlock(&ht->mutex);
1133 }
1134 EXPORT_SYMBOL_GPL(rhashtable_destroy);