lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
   5  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
   6  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   7  *
   8  * Code partially derived from nft_hash
   9  * Rewritten with rehash code from br_multicast plus single list
  10  * pointer as suggested by Josh Triplett
  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/atomic.h>
  18 #include <linux/kernel.h>
  19 #include <linux/init.h>
  20 #include <linux/log2.h>
  21 #include <linux/sched.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/mm.h>
  25 #include <linux/jhash.h>
  26 #include <linux/random.h>
  27 #include <linux/rhashtable.h>
  28 #include <linux/err.h>
  29 #include <linux/export.h>
  30
  31 #define HASH_DEFAULT_SIZE       64UL
  32 #define HASH_MIN_SIZE           4U
  33 #define BUCKET_LOCKS_PER_CPU    32UL
  34
  35 static u32 head_hashfn(struct rhashtable *ht,
  36                        const struct bucket_table *tbl,
  37                        const struct rhash_head *he)
  38 {
  39         return rht_head_hashfn(ht, tbl, he, ht->p);
  40 }
  41
  42 #ifdef CONFIG_PROVE_LOCKING
  43 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  44
  45 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  46 {
  47         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  48 }
  49 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  50
  51 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  52 {
  53         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  54
  55         return (debug_locks) ? lockdep_is_held(lock) : 1;
  56 }
  57 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  58 #else
  59 #define ASSERT_RHT_MUTEX(HT)
  60 #endif
  61
  62
  63 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  64                               gfp_t gfp)
  65 {
  66         unsigned int i, size;
  67 #if defined(CONFIG_PROVE_LOCKING)
  68         unsigned int nr_pcpus = 2;
  69 #else
  70         unsigned int nr_pcpus = num_possible_cpus();
  71 #endif
  72
  73         nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
  74         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  75
  76         /* Never allocate more than 0.5 locks per bucket */
  77         size = min_t(unsigned int, size, tbl->size >> 1);
  78
  79         if (sizeof(spinlock_t) != 0) {
  80                 tbl->locks = NULL;
  81 #ifdef CONFIG_NUMA
  82                 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
  83                     gfp == GFP_KERNEL)
  84                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
  85 #endif
  86                 if (gfp != GFP_KERNEL)
  87                         gfp |= __GFP_NOWARN | __GFP_NORETRY;
  88
  89                 if (!tbl->locks)
  90                         tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  91                                                    gfp);
  92                 if (!tbl->locks)
  93                         return -ENOMEM;
  94                 for (i = 0; i < size; i++)
  95                         spin_lock_init(&tbl->locks[i]);
  96         }
  97         tbl->locks_mask = size - 1;
  98
  99         return 0;
 100 }
 101
 102 static void bucket_table_free(const struct bucket_table *tbl)
 103 {
 104         if (tbl)
 105                 kvfree(tbl->locks);
 106
 107         kvfree(tbl);
 108 }
 109
 110 static void bucket_table_free_rcu(struct rcu_head *head)
 111 {
 112         bucket_table_free(container_of(head, struct bucket_table, rcu));
 113 }
 114
 115 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 116                                                size_t nbuckets,
 117                                                gfp_t gfp)
 118 {
 119         struct bucket_table *tbl = NULL;
 120         size_t size;
 121         int i;
 122
 123         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 124         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 125             gfp != GFP_KERNEL)
 126                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 127         if (tbl == NULL && gfp == GFP_KERNEL)
 128                 tbl = vzalloc(size);
 129         if (tbl == NULL)
 130                 return NULL;
 131
 132         tbl->size = nbuckets;
 133
 134         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 135                 bucket_table_free(tbl);
 136                 return NULL;
 137         }
 138
 139         INIT_LIST_HEAD(&tbl->walkers);
 140
 141         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 142
 143         for (i = 0; i < nbuckets; i++)
 144                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 145
 146         return tbl;
 147 }
 148
 149 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 150                                                   struct bucket_table *tbl)
 151 {
 152         struct bucket_table *new_tbl;
 153
 154         do {
 155                 new_tbl = tbl;
 156                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 157         } while (tbl);
 158
 159         return new_tbl;
 160 }
 161
 162 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 163 {
 164         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 165         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 166                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 167         struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
 168         int err = -ENOENT;
 169         struct rhash_head *head, *next, *entry;
 170         spinlock_t *new_bucket_lock;
 171         unsigned int new_hash;
 172
 173         rht_for_each(entry, old_tbl, old_hash) {
 174                 err = 0;
 175                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 176
 177                 if (rht_is_a_nulls(next))
 178                         break;
 179
 180                 pprev = &entry->next;
 181         }
 182
 183         if (err)
 184                 goto out;
 185
 186         new_hash = head_hashfn(ht, new_tbl, entry);
 187
 188         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 189
 190         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 191         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 192                                       new_tbl, new_hash);
 193
 194         RCU_INIT_POINTER(entry->next, head);
 195
 196         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 197         spin_unlock(new_bucket_lock);
 198
 199         rcu_assign_pointer(*pprev, next);
 200
 201 out:
 202         return err;
 203 }
 204
 205 static void rhashtable_rehash_chain(struct rhashtable *ht,
 206                                     unsigned int old_hash)
 207 {
 208         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 209         spinlock_t *old_bucket_lock;
 210
 211         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 212
 213         spin_lock_bh(old_bucket_lock);
 214         while (!rhashtable_rehash_one(ht, old_hash))
 215                 ;
 216         old_tbl->rehash++;
 217         spin_unlock_bh(old_bucket_lock);
 218 }
 219
 220 static int rhashtable_rehash_attach(struct rhashtable *ht,
 221                                     struct bucket_table *old_tbl,
 222                                     struct bucket_table *new_tbl)
 223 {
 224         /* Protect future_tbl using the first bucket lock. */
 225         spin_lock_bh(old_tbl->locks);
 226
 227         /* Did somebody beat us to it? */
 228         if (rcu_access_pointer(old_tbl->future_tbl)) {
 229                 spin_unlock_bh(old_tbl->locks);
 230                 return -EEXIST;
 231         }
 232
 233         /* Make insertions go into the new, empty table right away. Deletions
 234          * and lookups will be attempted in both tables until we synchronize.
 235          */
 236         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 237
 238         spin_unlock_bh(old_tbl->locks);
 239
 240         return 0;
 241 }
 242
 243 static int rhashtable_rehash_table(struct rhashtable *ht)
 244 {
 245         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 246         struct bucket_table *new_tbl;
 247         struct rhashtable_walker *walker;
 248         unsigned int old_hash;
 249
 250         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 251         if (!new_tbl)
 252                 return 0;
 253
 254         for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
 255                 rhashtable_rehash_chain(ht, old_hash);
 256
 257         /* Publish the new table pointer. */
 258         rcu_assign_pointer(ht->tbl, new_tbl);
 259
 260         spin_lock(&ht->lock);
 261         list_for_each_entry(walker, &old_tbl->walkers, list)
 262                 walker->tbl = NULL;
 263         spin_unlock(&ht->lock);
 264
 265         /* Wait for readers. All new readers will see the new
 266          * table, and thus no references to the old table will
 267          * remain.
 268          */
 269         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 270
 271         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 272 }
 273
 274 /**
 275  * rhashtable_expand - Expand hash table while allowing concurrent lookups
 276  * @ht:         the hash table to expand
 277  *
 278  * A secondary bucket array is allocated and the hash entries are migrated.
 279  *
 280  * This function may only be called in a context where it is safe to call
 281  * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
 282  *
 283  * The caller must ensure that no concurrent resizing occurs by holding
 284  * ht->mutex.
 285  *
 286  * It is valid to have concurrent insertions and deletions protected by per
 287  * bucket locks or concurrent RCU protected lookups and traversals.
 288  */
 289 static int rhashtable_expand(struct rhashtable *ht)
 290 {
 291         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 292         int err;
 293
 294         ASSERT_RHT_MUTEX(ht);
 295
 296         old_tbl = rhashtable_last_table(ht, old_tbl);
 297
 298         new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
 299         if (new_tbl == NULL)
 300                 return -ENOMEM;
 301
 302         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 303         if (err)
 304                 bucket_table_free(new_tbl);
 305
 306         return err;
 307 }
 308
 309 /**
 310  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 311  * @ht:         the hash table to shrink
 312  *
 313  * This function shrinks the hash table to fit, i.e., the smallest
 314  * size would not cause it to expand right away automatically.
 315  *
 316  * The caller must ensure that no concurrent resizing occurs by holding
 317  * ht->mutex.
 318  *
 319  * The caller must ensure that no concurrent table mutations take place.
 320  * It is however valid to have concurrent lookups if they are RCU protected.
 321  *
 322  * It is valid to have concurrent insertions and deletions protected by per
 323  * bucket locks or concurrent RCU protected lookups and traversals.
 324  */
 325 static int rhashtable_shrink(struct rhashtable *ht)
 326 {
 327         struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
 328         unsigned int nelems = atomic_read(&ht->nelems);
 329         unsigned int size = 0;
 330         int err;
 331
 332         ASSERT_RHT_MUTEX(ht);
 333
 334         if (nelems)
 335                 size = roundup_pow_of_two(nelems * 3 / 2);
 336         if (size < ht->p.min_size)
 337                 size = ht->p.min_size;
 338
 339         if (old_tbl->size <= size)
 340                 return 0;
 341
 342         if (rht_dereference(old_tbl->future_tbl, ht))
 343                 return -EEXIST;
 344
 345         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 346         if (new_tbl == NULL)
 347                 return -ENOMEM;
 348
 349         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 350         if (err)
 351                 bucket_table_free(new_tbl);
 352
 353         return err;
 354 }
 355
 356 static void rht_deferred_worker(struct work_struct *work)
 357 {
 358         struct rhashtable *ht;
 359         struct bucket_table *tbl;
 360         int err = 0;
 361
 362         ht = container_of(work, struct rhashtable, run_work);
 363         mutex_lock(&ht->mutex);
 364
 365         tbl = rht_dereference(ht->tbl, ht);
 366         tbl = rhashtable_last_table(ht, tbl);
 367
 368         if (rht_grow_above_75(ht, tbl))
 369                 rhashtable_expand(ht);
 370         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 371                 rhashtable_shrink(ht);
 372
 373         err = rhashtable_rehash_table(ht);
 374
 375         mutex_unlock(&ht->mutex);
 376
 377         if (err)
 378                 schedule_work(&ht->run_work);
 379 }
 380
 381 static int rhashtable_insert_rehash(struct rhashtable *ht,
 382                                     struct bucket_table *tbl)
 383 {
 384         struct bucket_table *old_tbl;
 385         struct bucket_table *new_tbl;
 386         unsigned int size;
 387         int err;
 388
 389         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 390
 391         size = tbl->size;
 392
 393         err = -EBUSY;
 394
 395         if (rht_grow_above_75(ht, tbl))
 396                 size *= 2;
 397         /* Do not schedule more than one rehash */
 398         else if (old_tbl != tbl)
 399                 goto fail;
 400
 401         err = -ENOMEM;
 402
 403         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 404         if (new_tbl == NULL)
 405                 goto fail;
 406
 407         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 408         if (err) {
 409                 bucket_table_free(new_tbl);
 410                 if (err == -EEXIST)
 411                         err = 0;
 412         } else
 413                 schedule_work(&ht->run_work);
 414
 415         return err;
 416
 417 fail:
 418         /* Do not fail the insert if someone else did a rehash. */
 419         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 420                 return 0;
 421
 422         /* Schedule async rehash to retry allocation in process context. */
 423         if (err == -ENOMEM)
 424                 schedule_work(&ht->run_work);
 425
 426         return err;
 427 }
 428
 429 static void *rhashtable_lookup_one(struct rhashtable *ht,
 430                                    struct bucket_table *tbl, unsigned int hash,
 431                                    const void *key, struct rhash_head *obj)
 432 {
 433         struct rhashtable_compare_arg arg = {
 434                 .ht = ht,
 435                 .key = key,
 436         };
 437         struct rhash_head __rcu **pprev;
 438         struct rhash_head *head;
 439         int elasticity;
 440
 441         elasticity = ht->elasticity;
 442         pprev = &tbl->buckets[hash];
 443         rht_for_each(head, tbl, hash) {
 444                 struct rhlist_head *list;
 445                 struct rhlist_head *plist;
 446
 447                 elasticity--;
 448                 if (!key ||
 449                     (ht->p.obj_cmpfn ?
 450                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 451                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 452                         pprev = &head->next;
 453                         continue;
 454                 }
 455
 456                 if (!ht->rhlist)
 457                         return rht_obj(ht, head);
 458
 459                 list = container_of(obj, struct rhlist_head, rhead);
 460                 plist = container_of(head, struct rhlist_head, rhead);
 461
 462                 RCU_INIT_POINTER(list->next, plist);
 463                 head = rht_dereference_bucket(head->next, tbl, hash);
 464                 RCU_INIT_POINTER(list->rhead.next, head);
 465                 rcu_assign_pointer(*pprev, obj);
 466
 467                 return NULL;
 468         }
 469
 470         if (elasticity <= 0)
 471                 return ERR_PTR(-EAGAIN);
 472
 473         return ERR_PTR(-ENOENT);
 474 }
 475
 476 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 477                                                   struct bucket_table *tbl,
 478                                                   unsigned int hash,
 479                                                   struct rhash_head *obj,
 480                                                   void *data)
 481 {
 482         struct bucket_table *new_tbl;
 483         struct rhash_head *head;
 484
 485         if (!IS_ERR_OR_NULL(data))
 486                 return ERR_PTR(-EEXIST);
 487
 488         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 489                 return ERR_CAST(data);
 490
 491         new_tbl = rcu_dereference(tbl->future_tbl);
 492         if (new_tbl)
 493                 return new_tbl;
 494
 495         if (PTR_ERR(data) != -ENOENT)
 496                 return ERR_CAST(data);
 497
 498         if (unlikely(rht_grow_above_max(ht, tbl)))
 499                 return ERR_PTR(-E2BIG);
 500
 501         if (unlikely(rht_grow_above_100(ht, tbl)))
 502                 return ERR_PTR(-EAGAIN);
 503
 504         head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
 505
 506         RCU_INIT_POINTER(obj->next, head);
 507         if (ht->rhlist) {
 508                 struct rhlist_head *list;
 509
 510                 list = container_of(obj, struct rhlist_head, rhead);
 511                 RCU_INIT_POINTER(list->next, NULL);
 512         }
 513
 514         rcu_assign_pointer(tbl->buckets[hash], obj);
 515
 516         atomic_inc(&ht->nelems);
 517         if (rht_grow_above_75(ht, tbl))
 518                 schedule_work(&ht->run_work);
 519
 520         return NULL;
 521 }
 522
 523 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 524                                    struct rhash_head *obj)
 525 {
 526         struct bucket_table *new_tbl;
 527         struct bucket_table *tbl;
 528         unsigned int hash;
 529         spinlock_t *lock;
 530         void *data;
 531
 532         tbl = rcu_dereference(ht->tbl);
 533
 534         /* All insertions must grab the oldest table containing
 535          * the hashed bucket that is yet to be rehashed.
 536          */
 537         for (;;) {
 538                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 539                 lock = rht_bucket_lock(tbl, hash);
 540                 spin_lock_bh(lock);
 541
 542                 if (tbl->rehash <= hash)
 543                         break;
 544
 545                 spin_unlock_bh(lock);
 546                 tbl = rcu_dereference(tbl->future_tbl);
 547         }
 548
 549         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 550         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 551         if (PTR_ERR(new_tbl) != -EEXIST)
 552                 data = ERR_CAST(new_tbl);
 553
 554         while (!IS_ERR_OR_NULL(new_tbl)) {
 555                 tbl = new_tbl;
 556                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 557                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 558                                  SINGLE_DEPTH_NESTING);
 559
 560                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 561                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 562                 if (PTR_ERR(new_tbl) != -EEXIST)
 563                         data = ERR_CAST(new_tbl);
 564
 565                 spin_unlock(rht_bucket_lock(tbl, hash));
 566         }
 567
 568         spin_unlock_bh(lock);
 569
 570         if (PTR_ERR(data) == -EAGAIN)
 571                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 572                                -EAGAIN);
 573
 574         return data;
 575 }
 576
 577 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 578                              struct rhash_head *obj)
 579 {
 580         void *data;
 581
 582         do {
 583                 rcu_read_lock();
 584                 data = rhashtable_try_insert(ht, key, obj);
 585                 rcu_read_unlock();
 586         } while (PTR_ERR(data) == -EAGAIN);
 587
 588         return data;
 589 }
 590 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 591
 592 /**
 593  * rhashtable_walk_enter - Initialise an iterator
 594  * @ht:         Table to walk over
 595  * @iter:       Hash table Iterator
 596  *
 597  * This function prepares a hash table walk.
 598  *
 599  * Note that if you restart a walk after rhashtable_walk_stop you
 600  * may see the same object twice.  Also, you may miss objects if
 601  * there are removals in between rhashtable_walk_stop and the next
 602  * call to rhashtable_walk_start.
 603  *
 604  * For a completely stable walk you should construct your own data
 605  * structure outside the hash table.
 606  *
 607  * This function may sleep so you must not call it from interrupt
 608  * context or with spin locks held.
 609  *
 610  * You must call rhashtable_walk_exit after this function returns.
 611  */
 612 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 613 {
 614         iter->ht = ht;
 615         iter->p = NULL;
 616         iter->slot = 0;
 617         iter->skip = 0;
 618
 619         spin_lock(&ht->lock);
 620         iter->walker.tbl =
 621                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 622         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 623         spin_unlock(&ht->lock);
 624 }
 625 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 626
 627 /**
 628  * rhashtable_walk_exit - Free an iterator
 629  * @iter:       Hash table Iterator
 630  *
 631  * This function frees resources allocated by rhashtable_walk_init.
 632  */
 633 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 634 {
 635         spin_lock(&iter->ht->lock);
 636         if (iter->walker.tbl)
 637                 list_del(&iter->walker.list);
 638         spin_unlock(&iter->ht->lock);
 639 }
 640 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 641
 642 /**
 643  * rhashtable_walk_start - Start a hash table walk
 644  * @iter:       Hash table iterator
 645  *
 646  * Start a hash table walk.  Note that we take the RCU lock in all
 647  * cases including when we return an error.  So you must always call
 648  * rhashtable_walk_stop to clean up.
 649  *
 650  * Returns zero if successful.
 651  *
 652  * Returns -EAGAIN if resize event occured.  Note that the iterator
 653  * will rewind back to the beginning and you may use it immediately
 654  * by calling rhashtable_walk_next.
 655  */
 656 int rhashtable_walk_start(struct rhashtable_iter *iter)
 657         __acquires(RCU)
 658 {
 659         struct rhashtable *ht = iter->ht;
 660
 661         rcu_read_lock();
 662
 663         spin_lock(&ht->lock);
 664         if (iter->walker.tbl)
 665                 list_del(&iter->walker.list);
 666         spin_unlock(&ht->lock);
 667
 668         if (!iter->walker.tbl) {
 669                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 670                 return -EAGAIN;
 671         }
 672
 673         return 0;
 674 }
 675 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 676
 677 /**
 678  * rhashtable_walk_next - Return the next object and advance the iterator
 679  * @iter:       Hash table iterator
 680  *
 681  * Note that you must call rhashtable_walk_stop when you are finished
 682  * with the walk.
 683  *
 684  * Returns the next object or NULL when the end of the table is reached.
 685  *
 686  * Returns -EAGAIN if resize event occured.  Note that the iterator
 687  * will rewind back to the beginning and you may continue to use it.
 688  */
 689 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 690 {
 691         struct bucket_table *tbl = iter->walker.tbl;
 692         struct rhlist_head *list = iter->list;
 693         struct rhashtable *ht = iter->ht;
 694         struct rhash_head *p = iter->p;
 695         bool rhlist = ht->rhlist;
 696
 697         if (p) {
 698                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 699                         p = rcu_dereference(p->next);
 700                         list = container_of(p, struct rhlist_head, rhead);
 701                 }
 702                 goto next;
 703         }
 704
 705         for (; iter->slot < tbl->size; iter->slot++) {
 706                 int skip = iter->skip;
 707
 708                 rht_for_each_rcu(p, tbl, iter->slot) {
 709                         if (rhlist) {
 710                                 list = container_of(p, struct rhlist_head,
 711                                                     rhead);
 712                                 do {
 713                                         if (!skip)
 714                                                 goto next;
 715                                         skip--;
 716                                         list = rcu_dereference(list->next);
 717                                 } while (list);
 718
 719                                 continue;
 720                         }
 721                         if (!skip)
 722                                 break;
 723                         skip--;
 724                 }
 725
 726 next:
 727                 if (!rht_is_a_nulls(p)) {
 728                         iter->skip++;
 729                         iter->p = p;
 730                         iter->list = list;
 731                         return rht_obj(ht, rhlist ? &list->rhead : p);
 732                 }
 733
 734                 iter->skip = 0;
 735         }
 736
 737         iter->p = NULL;
 738
 739         /* Ensure we see any new tables. */
 740         smp_rmb();
 741
 742         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 743         if (iter->walker.tbl) {
 744                 iter->slot = 0;
 745                 iter->skip = 0;
 746                 return ERR_PTR(-EAGAIN);
 747         }
 748
 749         return NULL;
 750 }
 751 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 752
 753 /**
 754  * rhashtable_walk_stop - Finish a hash table walk
 755  * @iter:       Hash table iterator
 756  *
 757  * Finish a hash table walk.
 758  */
 759 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 760         __releases(RCU)
 761 {
 762         struct rhashtable *ht;
 763         struct bucket_table *tbl = iter->walker.tbl;
 764
 765         if (!tbl)
 766                 goto out;
 767
 768         ht = iter->ht;
 769
 770         spin_lock(&ht->lock);
 771         if (tbl->rehash < tbl->size)
 772                 list_add(&iter->walker.list, &tbl->walkers);
 773         else
 774                 iter->walker.tbl = NULL;
 775         spin_unlock(&ht->lock);
 776
 777         iter->p = NULL;
 778
 779 out:
 780         rcu_read_unlock();
 781 }
 782 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 783
 784 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 785 {
 786         size_t retsize;
 787
 788         if (params->nelem_hint)
 789                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 790                               (unsigned long)params->min_size);
 791         else
 792                 retsize = max(HASH_DEFAULT_SIZE,
 793                               (unsigned long)params->min_size);
 794
 795         return retsize;
 796 }
 797
 798 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 799 {
 800         return jhash2(key, length, seed);
 801 }
 802
 803 /**
 804  * rhashtable_init - initialize a new hash table
 805  * @ht:         hash table to be initialized
 806  * @params:     configuration parameters
 807  *
 808  * Initializes a new hash table based on the provided configuration
 809  * parameters. A table can be configured either with a variable or
 810  * fixed length key:
 811  *
 812  * Configuration Example 1: Fixed length keys
 813  * struct test_obj {
 814  *      int                     key;
 815  *      void *                  my_member;
 816  *      struct rhash_head       node;
 817  * };
 818  *
 819  * struct rhashtable_params params = {
 820  *      .head_offset = offsetof(struct test_obj, node),
 821  *      .key_offset = offsetof(struct test_obj, key),
 822  *      .key_len = sizeof(int),
 823  *      .hashfn = jhash,
 824  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 825  * };
 826  *
 827  * Configuration Example 2: Variable length keys
 828  * struct test_obj {
 829  *      [...]
 830  *      struct rhash_head       node;
 831  * };
 832  *
 833  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 834  * {
 835  *      struct test_obj *obj = data;
 836  *
 837  *      return [... hash ...];
 838  * }
 839  *
 840  * struct rhashtable_params params = {
 841  *      .head_offset = offsetof(struct test_obj, node),
 842  *      .hashfn = jhash,
 843  *      .obj_hashfn = my_hash_fn,
 844  * };
 845  */
 846 int rhashtable_init(struct rhashtable *ht,
 847                     const struct rhashtable_params *params)
 848 {
 849         struct bucket_table *tbl;
 850         size_t size;
 851
 852         if ((!params->key_len && !params->obj_hashfn) ||
 853             (params->obj_hashfn && !params->obj_cmpfn))
 854                 return -EINVAL;
 855
 856         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 857                 return -EINVAL;
 858
 859         memset(ht, 0, sizeof(*ht));
 860         mutex_init(&ht->mutex);
 861         spin_lock_init(&ht->lock);
 862         memcpy(&ht->p, params, sizeof(*params));
 863
 864         if (params->min_size)
 865                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 866
 867         if (params->max_size)
 868                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 869
 870         if (params->insecure_max_entries)
 871                 ht->p.insecure_max_entries =
 872                         rounddown_pow_of_two(params->insecure_max_entries);
 873         else
 874                 ht->p.insecure_max_entries = ht->p.max_size * 2;
 875
 876         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 877
 878         size = rounded_hashtable_size(&ht->p);
 879
 880         /* The maximum (not average) chain length grows with the
 881          * size of the hash table, at a rate of (log N)/(log log N).
 882          * The value of 16 is selected so that even if the hash
 883          * table grew to 2^32 you would not expect the maximum
 884          * chain length to exceed it unless we are under attack
 885          * (or extremely unlucky).
 886          *
 887          * As this limit is only to detect attacks, we don't need
 888          * to set it to a lower value as you'd need the chain
 889          * length to vastly exceed 16 to have any real effect
 890          * on the system.
 891          */
 892         if (!params->insecure_elasticity)
 893                 ht->elasticity = 16;
 894
 895         if (params->locks_mul)
 896                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 897         else
 898                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 899
 900         ht->key_len = ht->p.key_len;
 901         if (!params->hashfn) {
 902                 ht->p.hashfn = jhash;
 903
 904                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 905                         ht->key_len /= sizeof(u32);
 906                         ht->p.hashfn = rhashtable_jhash2;
 907                 }
 908         }
 909
 910         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 911         if (tbl == NULL)
 912                 return -ENOMEM;
 913
 914         atomic_set(&ht->nelems, 0);
 915
 916         RCU_INIT_POINTER(ht->tbl, tbl);
 917
 918         INIT_WORK(&ht->run_work, rht_deferred_worker);
 919
 920         return 0;
 921 }
 922 EXPORT_SYMBOL_GPL(rhashtable_init);
 923
 924 /**
 925  * rhltable_init - initialize a new hash list table
 926  * @hlt:        hash list table to be initialized
 927  * @params:     configuration parameters
 928  *
 929  * Initializes a new hash list table.
 930  *
 931  * See documentation for rhashtable_init.
 932  */
 933 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
 934 {
 935         int err;
 936
 937         /* No rhlist NULLs marking for now. */
 938         if (params->nulls_base)
 939                 return -EINVAL;
 940
 941         err = rhashtable_init(&hlt->ht, params);
 942         hlt->ht.rhlist = true;
 943         return err;
 944 }
 945 EXPORT_SYMBOL_GPL(rhltable_init);
 946
 947 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
 948                                 void (*free_fn)(void *ptr, void *arg),
 949                                 void *arg)
 950 {
 951         struct rhlist_head *list;
 952
 953         if (!ht->rhlist) {
 954                 free_fn(rht_obj(ht, obj), arg);
 955                 return;
 956         }
 957
 958         list = container_of(obj, struct rhlist_head, rhead);
 959         do {
 960                 obj = &list->rhead;
 961                 list = rht_dereference(list->next, ht);
 962                 free_fn(rht_obj(ht, obj), arg);
 963         } while (list);
 964 }
 965
 966 /**
 967  * rhashtable_free_and_destroy - free elements and destroy hash table
 968  * @ht:         the hash table to destroy
 969  * @free_fn:    callback to release resources of element
 970  * @arg:        pointer passed to free_fn
 971  *
 972  * Stops an eventual async resize. If defined, invokes free_fn for each
 973  * element to releasal resources. Please note that RCU protected
 974  * readers may still be accessing the elements. Releasing of resources
 975  * must occur in a compatible manner. Then frees the bucket array.
 976  *
 977  * This function will eventually sleep to wait for an async resize
 978  * to complete. The caller is responsible that no further write operations
 979  * occurs in parallel.
 980  */
 981 void rhashtable_free_and_destroy(struct rhashtable *ht,
 982                                  void (*free_fn)(void *ptr, void *arg),
 983                                  void *arg)
 984 {
 985         const struct bucket_table *tbl;
 986         unsigned int i;
 987
 988         cancel_work_sync(&ht->run_work);
 989
 990         mutex_lock(&ht->mutex);
 991         tbl = rht_dereference(ht->tbl, ht);
 992         if (free_fn) {
 993                 for (i = 0; i < tbl->size; i++) {
 994                         struct rhash_head *pos, *next;
 995
 996                         for (pos = rht_dereference(tbl->buckets[i], ht),
 997                              next = !rht_is_a_nulls(pos) ?
 998                                         rht_dereference(pos->next, ht) : NULL;
 999                              !rht_is_a_nulls(pos);
1000                              pos = next,
1001                              next = !rht_is_a_nulls(pos) ?
1002                                         rht_dereference(pos->next, ht) : NULL)
1003                                 rhashtable_free_one(ht, pos, free_fn, arg);
1004                 }
1005         }
1006
1007         bucket_table_free(tbl);
1008         mutex_unlock(&ht->mutex);
1009 }
1010 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1011
1012 void rhashtable_destroy(struct rhashtable *ht)
1013 {
1014         return rhashtable_free_and_destroy(ht, NULL, NULL);
1015 }
1016 EXPORT_SYMBOL_GPL(rhashtable_destroy);