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/rculist.h>
  23 #include <linux/slab.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/mm.h>
  26 #include <linux/jhash.h>
  27 #include <linux/random.h>
  28 #include <linux/rhashtable.h>
  29 #include <linux/err.h>
  30 #include <linux/export.h>
  31
  32 #define HASH_DEFAULT_SIZE       64UL
  33 #define HASH_MIN_SIZE           4U
  34 #define BUCKET_LOCKS_PER_CPU    32UL
  35
  36 union nested_table {
  37         union nested_table __rcu *table;
  38         struct rhash_head __rcu *bucket;
  39 };
  40
  41 static u32 head_hashfn(struct rhashtable *ht,
  42                        const struct bucket_table *tbl,
  43                        const struct rhash_head *he)
  44 {
  45         return rht_head_hashfn(ht, tbl, he, ht->p);
  46 }
  47
  48 #ifdef CONFIG_PROVE_LOCKING
  49 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  50
  51 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  52 {
  53         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  54 }
  55 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  56
  57 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  58 {
  59         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  60
  61         return (debug_locks) ? lockdep_is_held(lock) : 1;
  62 }
  63 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  64 #else
  65 #define ASSERT_RHT_MUTEX(HT)
  66 #endif
  67
  68 static void nested_table_free(union nested_table *ntbl, unsigned int size)
  69 {
  70         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
  71         const unsigned int len = 1 << shift;
  72         unsigned int i;
  73
  74         ntbl = rcu_dereference_raw(ntbl->table);
  75         if (!ntbl)
  76                 return;
  77
  78         if (size > len) {
  79                 size >>= shift;
  80                 for (i = 0; i < len; i++)
  81                         nested_table_free(ntbl + i, size);
  82         }
  83
  84         kfree(ntbl);
  85 }
  86
  87 static void nested_bucket_table_free(const struct bucket_table *tbl)
  88 {
  89         unsigned int size = tbl->size >> tbl->nest;
  90         unsigned int len = 1 << tbl->nest;
  91         union nested_table *ntbl;
  92         unsigned int i;
  93
  94         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
  95
  96         for (i = 0; i < len; i++)
  97                 nested_table_free(ntbl + i, size);
  98
  99         kfree(ntbl);
 100 }
 101
 102 static void bucket_table_free(const struct bucket_table *tbl)
 103 {
 104         if (tbl->nest)
 105                 nested_bucket_table_free(tbl);
 106
 107         free_bucket_spinlocks(tbl->locks);
 108         kvfree(tbl);
 109 }
 110
 111 static void bucket_table_free_rcu(struct rcu_head *head)
 112 {
 113         bucket_table_free(container_of(head, struct bucket_table, rcu));
 114 }
 115
 116 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 117                                               union nested_table __rcu **prev,
 118                                               bool leaf)
 119 {
 120         union nested_table *ntbl;
 121         int i;
 122
 123         ntbl = rcu_dereference(*prev);
 124         if (ntbl)
 125                 return ntbl;
 126
 127         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 128
 129         if (ntbl && leaf) {
 130                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
 131                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
 132         }
 133
 134         rcu_assign_pointer(*prev, ntbl);
 135
 136         return ntbl;
 137 }
 138
 139 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 140                                                       size_t nbuckets,
 141                                                       gfp_t gfp)
 142 {
 143         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 144         struct bucket_table *tbl;
 145         size_t size;
 146
 147         if (nbuckets < (1 << (shift + 1)))
 148                 return NULL;
 149
 150         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 151
 152         tbl = kzalloc(size, gfp);
 153         if (!tbl)
 154                 return NULL;
 155
 156         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 157                                 false)) {
 158                 kfree(tbl);
 159                 return NULL;
 160         }
 161
 162         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 163
 164         return tbl;
 165 }
 166
 167 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 168                                                size_t nbuckets,
 169                                                gfp_t gfp)
 170 {
 171         struct bucket_table *tbl = NULL;
 172         size_t size, max_locks;
 173         int i;
 174
 175         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 176         tbl = kvzalloc(size, gfp);
 177
 178         size = nbuckets;
 179
 180         if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
 181                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 182                 nbuckets = 0;
 183         }
 184
 185         if (tbl == NULL)
 186                 return NULL;
 187
 188         tbl->size = size;
 189
 190         max_locks = size >> 1;
 191         if (tbl->nest)
 192                 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
 193
 194         if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
 195                                    ht->p.locks_mul, gfp) < 0) {
 196                 bucket_table_free(tbl);
 197                 return NULL;
 198         }
 199
 200         INIT_LIST_HEAD(&tbl->walkers);
 201
 202         tbl->hash_rnd = get_random_u32();
 203
 204         for (i = 0; i < nbuckets; i++)
 205                 INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
 206
 207         return tbl;
 208 }
 209
 210 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 211                                                   struct bucket_table *tbl)
 212 {
 213         struct bucket_table *new_tbl;
 214
 215         do {
 216                 new_tbl = tbl;
 217                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 218         } while (tbl);
 219
 220         return new_tbl;
 221 }
 222
 223 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 224 {
 225         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 226         struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
 227         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 228         int err = -EAGAIN;
 229         struct rhash_head *head, *next, *entry;
 230         spinlock_t *new_bucket_lock;
 231         unsigned int new_hash;
 232
 233         if (new_tbl->nest)
 234                 goto out;
 235
 236         err = -ENOENT;
 237
 238         rht_for_each(entry, old_tbl, old_hash) {
 239                 err = 0;
 240                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 241
 242                 if (rht_is_a_nulls(next))
 243                         break;
 244
 245                 pprev = &entry->next;
 246         }
 247
 248         if (err)
 249                 goto out;
 250
 251         new_hash = head_hashfn(ht, new_tbl, entry);
 252
 253         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 254
 255         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 256         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 257                                       new_tbl, new_hash);
 258
 259         RCU_INIT_POINTER(entry->next, head);
 260
 261         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 262         spin_unlock(new_bucket_lock);
 263
 264         rcu_assign_pointer(*pprev, next);
 265
 266 out:
 267         return err;
 268 }
 269
 270 static int rhashtable_rehash_chain(struct rhashtable *ht,
 271                                     unsigned int old_hash)
 272 {
 273         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 274         spinlock_t *old_bucket_lock;
 275         int err;
 276
 277         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 278
 279         spin_lock_bh(old_bucket_lock);
 280         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 281                 ;
 282
 283         if (err == -ENOENT) {
 284                 old_tbl->rehash++;
 285                 err = 0;
 286         }
 287         spin_unlock_bh(old_bucket_lock);
 288
 289         return err;
 290 }
 291
 292 static int rhashtable_rehash_attach(struct rhashtable *ht,
 293                                     struct bucket_table *old_tbl,
 294                                     struct bucket_table *new_tbl)
 295 {
 296         /* Make insertions go into the new, empty table right away. Deletions
 297          * and lookups will be attempted in both tables until we synchronize.
 298          * As cmpxchg() provides strong barriers, we do not need
 299          * rcu_assign_pointer().
 300          */
 301
 302         if (cmpxchg(&old_tbl->future_tbl, NULL, new_tbl) != NULL)
 303                 return -EEXIST;
 304
 305         return 0;
 306 }
 307
 308 static int rhashtable_rehash_table(struct rhashtable *ht)
 309 {
 310         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 311         struct bucket_table *new_tbl;
 312         struct rhashtable_walker *walker;
 313         unsigned int old_hash;
 314         int err;
 315
 316         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 317         if (!new_tbl)
 318                 return 0;
 319
 320         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 321                 err = rhashtable_rehash_chain(ht, old_hash);
 322                 if (err)
 323                         return err;
 324                 cond_resched();
 325         }
 326
 327         /* Publish the new table pointer. */
 328         rcu_assign_pointer(ht->tbl, new_tbl);
 329
 330         spin_lock(&ht->lock);
 331         list_for_each_entry(walker, &old_tbl->walkers, list)
 332                 walker->tbl = NULL;
 333         spin_unlock(&ht->lock);
 334
 335         /* Wait for readers. All new readers will see the new
 336          * table, and thus no references to the old table will
 337          * remain.
 338          */
 339         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 340
 341         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 342 }
 343
 344 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 345                                    struct bucket_table *old_tbl,
 346                                    unsigned int size)
 347 {
 348         struct bucket_table *new_tbl;
 349         int err;
 350
 351         ASSERT_RHT_MUTEX(ht);
 352
 353         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 354         if (new_tbl == NULL)
 355                 return -ENOMEM;
 356
 357         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 358         if (err)
 359                 bucket_table_free(new_tbl);
 360
 361         return err;
 362 }
 363
 364 /**
 365  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 366  * @ht:         the hash table to shrink
 367  *
 368  * This function shrinks the hash table to fit, i.e., the smallest
 369  * size would not cause it to expand right away automatically.
 370  *
 371  * The caller must ensure that no concurrent resizing occurs by holding
 372  * ht->mutex.
 373  *
 374  * The caller must ensure that no concurrent table mutations take place.
 375  * It is however valid to have concurrent lookups if they are RCU protected.
 376  *
 377  * It is valid to have concurrent insertions and deletions protected by per
 378  * bucket locks or concurrent RCU protected lookups and traversals.
 379  */
 380 static int rhashtable_shrink(struct rhashtable *ht)
 381 {
 382         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 383         unsigned int nelems = atomic_read(&ht->nelems);
 384         unsigned int size = 0;
 385
 386         if (nelems)
 387                 size = roundup_pow_of_two(nelems * 3 / 2);
 388         if (size < ht->p.min_size)
 389                 size = ht->p.min_size;
 390
 391         if (old_tbl->size <= size)
 392                 return 0;
 393
 394         if (rht_dereference(old_tbl->future_tbl, ht))
 395                 return -EEXIST;
 396
 397         return rhashtable_rehash_alloc(ht, old_tbl, size);
 398 }
 399
 400 static void rht_deferred_worker(struct work_struct *work)
 401 {
 402         struct rhashtable *ht;
 403         struct bucket_table *tbl;
 404         int err = 0;
 405
 406         ht = container_of(work, struct rhashtable, run_work);
 407         mutex_lock(&ht->mutex);
 408
 409         tbl = rht_dereference(ht->tbl, ht);
 410         tbl = rhashtable_last_table(ht, tbl);
 411
 412         if (rht_grow_above_75(ht, tbl))
 413                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 414         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 415                 err = rhashtable_shrink(ht);
 416         else if (tbl->nest)
 417                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 418
 419         if (!err || err == -EEXIST) {
 420                 int nerr;
 421
 422                 nerr = rhashtable_rehash_table(ht);
 423                 err = err ?: nerr;
 424         }
 425
 426         mutex_unlock(&ht->mutex);
 427
 428         if (err)
 429                 schedule_work(&ht->run_work);
 430 }
 431
 432 static int rhashtable_insert_rehash(struct rhashtable *ht,
 433                                     struct bucket_table *tbl)
 434 {
 435         struct bucket_table *old_tbl;
 436         struct bucket_table *new_tbl;
 437         unsigned int size;
 438         int err;
 439
 440         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 441
 442         size = tbl->size;
 443
 444         err = -EBUSY;
 445
 446         if (rht_grow_above_75(ht, tbl))
 447                 size *= 2;
 448         /* Do not schedule more than one rehash */
 449         else if (old_tbl != tbl)
 450                 goto fail;
 451
 452         err = -ENOMEM;
 453
 454         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 455         if (new_tbl == NULL)
 456                 goto fail;
 457
 458         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 459         if (err) {
 460                 bucket_table_free(new_tbl);
 461                 if (err == -EEXIST)
 462                         err = 0;
 463         } else
 464                 schedule_work(&ht->run_work);
 465
 466         return err;
 467
 468 fail:
 469         /* Do not fail the insert if someone else did a rehash. */
 470         if (likely(rcu_access_pointer(tbl->future_tbl)))
 471                 return 0;
 472
 473         /* Schedule async rehash to retry allocation in process context. */
 474         if (err == -ENOMEM)
 475                 schedule_work(&ht->run_work);
 476
 477         return err;
 478 }
 479
 480 static void *rhashtable_lookup_one(struct rhashtable *ht,
 481                                    struct bucket_table *tbl, unsigned int hash,
 482                                    const void *key, struct rhash_head *obj)
 483 {
 484         struct rhashtable_compare_arg arg = {
 485                 .ht = ht,
 486                 .key = key,
 487         };
 488         struct rhash_head __rcu **pprev;
 489         struct rhash_head *head;
 490         int elasticity;
 491
 492         elasticity = RHT_ELASTICITY;
 493         pprev = rht_bucket_var(tbl, hash);
 494         rht_for_each_continue(head, *pprev, tbl, hash) {
 495                 struct rhlist_head *list;
 496                 struct rhlist_head *plist;
 497
 498                 elasticity--;
 499                 if (!key ||
 500                     (ht->p.obj_cmpfn ?
 501                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 502                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 503                         pprev = &head->next;
 504                         continue;
 505                 }
 506
 507                 if (!ht->rhlist)
 508                         return rht_obj(ht, head);
 509
 510                 list = container_of(obj, struct rhlist_head, rhead);
 511                 plist = container_of(head, struct rhlist_head, rhead);
 512
 513                 RCU_INIT_POINTER(list->next, plist);
 514                 head = rht_dereference_bucket(head->next, tbl, hash);
 515                 RCU_INIT_POINTER(list->rhead.next, head);
 516                 rcu_assign_pointer(*pprev, obj);
 517
 518                 return NULL;
 519         }
 520
 521         if (elasticity <= 0)
 522                 return ERR_PTR(-EAGAIN);
 523
 524         return ERR_PTR(-ENOENT);
 525 }
 526
 527 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 528                                                   struct bucket_table *tbl,
 529                                                   unsigned int hash,
 530                                                   struct rhash_head *obj,
 531                                                   void *data)
 532 {
 533         struct rhash_head __rcu **pprev;
 534         struct bucket_table *new_tbl;
 535         struct rhash_head *head;
 536
 537         if (!IS_ERR_OR_NULL(data))
 538                 return ERR_PTR(-EEXIST);
 539
 540         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 541                 return ERR_CAST(data);
 542
 543         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 544         if (new_tbl)
 545                 return new_tbl;
 546
 547         if (PTR_ERR(data) != -ENOENT)
 548                 return ERR_CAST(data);
 549
 550         if (unlikely(rht_grow_above_max(ht, tbl)))
 551                 return ERR_PTR(-E2BIG);
 552
 553         if (unlikely(rht_grow_above_100(ht, tbl)))
 554                 return ERR_PTR(-EAGAIN);
 555
 556         pprev = rht_bucket_insert(ht, tbl, hash);
 557         if (!pprev)
 558                 return ERR_PTR(-ENOMEM);
 559
 560         head = rht_dereference_bucket(*pprev, tbl, hash);
 561
 562         RCU_INIT_POINTER(obj->next, head);
 563         if (ht->rhlist) {
 564                 struct rhlist_head *list;
 565
 566                 list = container_of(obj, struct rhlist_head, rhead);
 567                 RCU_INIT_POINTER(list->next, NULL);
 568         }
 569
 570         rcu_assign_pointer(*pprev, obj);
 571
 572         atomic_inc(&ht->nelems);
 573         if (rht_grow_above_75(ht, tbl))
 574                 schedule_work(&ht->run_work);
 575
 576         return NULL;
 577 }
 578
 579 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 580                                    struct rhash_head *obj)
 581 {
 582         struct bucket_table *new_tbl;
 583         struct bucket_table *tbl;
 584         unsigned int hash;
 585         spinlock_t *lock;
 586         void *data;
 587
 588         tbl = rcu_dereference(ht->tbl);
 589
 590         /* All insertions must grab the oldest table containing
 591          * the hashed bucket that is yet to be rehashed.
 592          */
 593         for (;;) {
 594                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 595                 lock = rht_bucket_lock(tbl, hash);
 596                 spin_lock_bh(lock);
 597
 598                 if (tbl->rehash <= hash)
 599                         break;
 600
 601                 spin_unlock_bh(lock);
 602                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 603         }
 604
 605         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 606         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 607         if (PTR_ERR(new_tbl) != -EEXIST)
 608                 data = ERR_CAST(new_tbl);
 609
 610         while (!IS_ERR_OR_NULL(new_tbl)) {
 611                 tbl = new_tbl;
 612                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 613                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 614                                  SINGLE_DEPTH_NESTING);
 615
 616                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 617                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 618                 if (PTR_ERR(new_tbl) != -EEXIST)
 619                         data = ERR_CAST(new_tbl);
 620
 621                 spin_unlock(rht_bucket_lock(tbl, hash));
 622         }
 623
 624         spin_unlock_bh(lock);
 625
 626         if (PTR_ERR(data) == -EAGAIN)
 627                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 628                                -EAGAIN);
 629
 630         return data;
 631 }
 632
 633 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 634                              struct rhash_head *obj)
 635 {
 636         void *data;
 637
 638         do {
 639                 rcu_read_lock();
 640                 data = rhashtable_try_insert(ht, key, obj);
 641                 rcu_read_unlock();
 642         } while (PTR_ERR(data) == -EAGAIN);
 643
 644         return data;
 645 }
 646 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 647
 648 /**
 649  * rhashtable_walk_enter - Initialise an iterator
 650  * @ht:         Table to walk over
 651  * @iter:       Hash table Iterator
 652  *
 653  * This function prepares a hash table walk.
 654  *
 655  * Note that if you restart a walk after rhashtable_walk_stop you
 656  * may see the same object twice.  Also, you may miss objects if
 657  * there are removals in between rhashtable_walk_stop and the next
 658  * call to rhashtable_walk_start.
 659  *
 660  * For a completely stable walk you should construct your own data
 661  * structure outside the hash table.
 662  *
 663  * This function may be called from any process context, including
 664  * non-preemptable context, but cannot be called from softirq or
 665  * hardirq context.
 666  *
 667  * You must call rhashtable_walk_exit after this function returns.
 668  */
 669 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 670 {
 671         iter->ht = ht;
 672         iter->p = NULL;
 673         iter->slot = 0;
 674         iter->skip = 0;
 675         iter->end_of_table = 0;
 676
 677         spin_lock(&ht->lock);
 678         iter->walker.tbl =
 679                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 680         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 681         spin_unlock(&ht->lock);
 682 }
 683 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 684
 685 /**
 686  * rhashtable_walk_exit - Free an iterator
 687  * @iter:       Hash table Iterator
 688  *
 689  * This function frees resources allocated by rhashtable_walk_enter.
 690  */
 691 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 692 {
 693         spin_lock(&iter->ht->lock);
 694         if (iter->walker.tbl)
 695                 list_del(&iter->walker.list);
 696         spin_unlock(&iter->ht->lock);
 697 }
 698 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 699
 700 /**
 701  * rhashtable_walk_start_check - Start a hash table walk
 702  * @iter:       Hash table iterator
 703  *
 704  * Start a hash table walk at the current iterator position.  Note that we take
 705  * the RCU lock in all cases including when we return an error.  So you must
 706  * always call rhashtable_walk_stop to clean up.
 707  *
 708  * Returns zero if successful.
 709  *
 710  * Returns -EAGAIN if resize event occured.  Note that the iterator
 711  * will rewind back to the beginning and you may use it immediately
 712  * by calling rhashtable_walk_next.
 713  *
 714  * rhashtable_walk_start is defined as an inline variant that returns
 715  * void. This is preferred in cases where the caller would ignore
 716  * resize events and always continue.
 717  */
 718 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 719         __acquires(RCU)
 720 {
 721         struct rhashtable *ht = iter->ht;
 722         bool rhlist = ht->rhlist;
 723
 724         rcu_read_lock();
 725
 726         spin_lock(&ht->lock);
 727         if (iter->walker.tbl)
 728                 list_del(&iter->walker.list);
 729         spin_unlock(&ht->lock);
 730
 731         if (iter->end_of_table)
 732                 return 0;
 733         if (!iter->walker.tbl) {
 734                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 735                 iter->slot = 0;
 736                 iter->skip = 0;
 737                 return -EAGAIN;
 738         }
 739
 740         if (iter->p && !rhlist) {
 741                 /*
 742                  * We need to validate that 'p' is still in the table, and
 743                  * if so, update 'skip'
 744                  */
 745                 struct rhash_head *p;
 746                 int skip = 0;
 747                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 748                         skip++;
 749                         if (p == iter->p) {
 750                                 iter->skip = skip;
 751                                 goto found;
 752                         }
 753                 }
 754                 iter->p = NULL;
 755         } else if (iter->p && rhlist) {
 756                 /* Need to validate that 'list' is still in the table, and
 757                  * if so, update 'skip' and 'p'.
 758                  */
 759                 struct rhash_head *p;
 760                 struct rhlist_head *list;
 761                 int skip = 0;
 762                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 763                         for (list = container_of(p, struct rhlist_head, rhead);
 764                              list;
 765                              list = rcu_dereference(list->next)) {
 766                                 skip++;
 767                                 if (list == iter->list) {
 768                                         iter->p = p;
 769                                         iter->skip = skip;
 770                                         goto found;
 771                                 }
 772                         }
 773                 }
 774                 iter->p = NULL;
 775         }
 776 found:
 777         return 0;
 778 }
 779 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 780
 781 /**
 782  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 783  * one in case of a new walk).
 784  *
 785  * @iter:       Hash table iterator
 786  *
 787  * Returns the found object or NULL when the end of the table is reached.
 788  *
 789  * Returns -EAGAIN if resize event occurred.
 790  */
 791 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 792 {
 793         struct bucket_table *tbl = iter->walker.tbl;
 794         struct rhlist_head *list = iter->list;
 795         struct rhashtable *ht = iter->ht;
 796         struct rhash_head *p = iter->p;
 797         bool rhlist = ht->rhlist;
 798
 799         if (!tbl)
 800                 return NULL;
 801
 802         for (; iter->slot < tbl->size; iter->slot++) {
 803                 int skip = iter->skip;
 804
 805                 rht_for_each_rcu(p, tbl, iter->slot) {
 806                         if (rhlist) {
 807                                 list = container_of(p, struct rhlist_head,
 808                                                     rhead);
 809                                 do {
 810                                         if (!skip)
 811                                                 goto next;
 812                                         skip--;
 813                                         list = rcu_dereference(list->next);
 814                                 } while (list);
 815
 816                                 continue;
 817                         }
 818                         if (!skip)
 819                                 break;
 820                         skip--;
 821                 }
 822
 823 next:
 824                 if (!rht_is_a_nulls(p)) {
 825                         iter->skip++;
 826                         iter->p = p;
 827                         iter->list = list;
 828                         return rht_obj(ht, rhlist ? &list->rhead : p);
 829                 }
 830
 831                 iter->skip = 0;
 832         }
 833
 834         iter->p = NULL;
 835
 836         /* Ensure we see any new tables. */
 837         smp_rmb();
 838
 839         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 840         if (iter->walker.tbl) {
 841                 iter->slot = 0;
 842                 iter->skip = 0;
 843                 return ERR_PTR(-EAGAIN);
 844         } else {
 845                 iter->end_of_table = true;
 846         }
 847
 848         return NULL;
 849 }
 850
 851 /**
 852  * rhashtable_walk_next - Return the next object and advance the iterator
 853  * @iter:       Hash table iterator
 854  *
 855  * Note that you must call rhashtable_walk_stop when you are finished
 856  * with the walk.
 857  *
 858  * Returns the next object or NULL when the end of the table is reached.
 859  *
 860  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 861  * will rewind back to the beginning and you may continue to use it.
 862  */
 863 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 864 {
 865         struct rhlist_head *list = iter->list;
 866         struct rhashtable *ht = iter->ht;
 867         struct rhash_head *p = iter->p;
 868         bool rhlist = ht->rhlist;
 869
 870         if (p) {
 871                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 872                         p = rcu_dereference(p->next);
 873                         list = container_of(p, struct rhlist_head, rhead);
 874                 }
 875                 if (!rht_is_a_nulls(p)) {
 876                         iter->skip++;
 877                         iter->p = p;
 878                         iter->list = list;
 879                         return rht_obj(ht, rhlist ? &list->rhead : p);
 880                 }
 881
 882                 /* At the end of this slot, switch to next one and then find
 883                  * next entry from that point.
 884                  */
 885                 iter->skip = 0;
 886                 iter->slot++;
 887         }
 888
 889         return __rhashtable_walk_find_next(iter);
 890 }
 891 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 892
 893 /**
 894  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 895  * @iter:       Hash table iterator
 896  *
 897  * Returns the next object or NULL when the end of the table is reached.
 898  *
 899  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 900  * will rewind back to the beginning and you may continue to use it.
 901  */
 902 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 903 {
 904         struct rhlist_head *list = iter->list;
 905         struct rhashtable *ht = iter->ht;
 906         struct rhash_head *p = iter->p;
 907
 908         if (p)
 909                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 910
 911         /* No object found in current iter, find next one in the table. */
 912
 913         if (iter->skip) {
 914                 /* A nonzero skip value points to the next entry in the table
 915                  * beyond that last one that was found. Decrement skip so
 916                  * we find the current value. __rhashtable_walk_find_next
 917                  * will restore the original value of skip assuming that
 918                  * the table hasn't changed.
 919                  */
 920                 iter->skip--;
 921         }
 922
 923         return __rhashtable_walk_find_next(iter);
 924 }
 925 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 926
 927 /**
 928  * rhashtable_walk_stop - Finish a hash table walk
 929  * @iter:       Hash table iterator
 930  *
 931  * Finish a hash table walk.  Does not reset the iterator to the start of the
 932  * hash table.
 933  */
 934 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 935         __releases(RCU)
 936 {
 937         struct rhashtable *ht;
 938         struct bucket_table *tbl = iter->walker.tbl;
 939
 940         if (!tbl)
 941                 goto out;
 942
 943         ht = iter->ht;
 944
 945         spin_lock(&ht->lock);
 946         if (tbl->rehash < tbl->size)
 947                 list_add(&iter->walker.list, &tbl->walkers);
 948         else
 949                 iter->walker.tbl = NULL;
 950         spin_unlock(&ht->lock);
 951
 952 out:
 953         rcu_read_unlock();
 954 }
 955 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 956
 957 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 958 {
 959         size_t retsize;
 960
 961         if (params->nelem_hint)
 962                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 963                               (unsigned long)params->min_size);
 964         else
 965                 retsize = max(HASH_DEFAULT_SIZE,
 966                               (unsigned long)params->min_size);
 967
 968         return retsize;
 969 }
 970
 971 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 972 {
 973         return jhash2(key, length, seed);
 974 }
 975
 976 /**
 977  * rhashtable_init - initialize a new hash table
 978  * @ht:         hash table to be initialized
 979  * @params:     configuration parameters
 980  *
 981  * Initializes a new hash table based on the provided configuration
 982  * parameters. A table can be configured either with a variable or
 983  * fixed length key:
 984  *
 985  * Configuration Example 1: Fixed length keys
 986  * struct test_obj {
 987  *      int                     key;
 988  *      void *                  my_member;
 989  *      struct rhash_head       node;
 990  * };
 991  *
 992  * struct rhashtable_params params = {
 993  *      .head_offset = offsetof(struct test_obj, node),
 994  *      .key_offset = offsetof(struct test_obj, key),
 995  *      .key_len = sizeof(int),
 996  *      .hashfn = jhash,
 997  * };
 998  *
 999  * Configuration Example 2: Variable length keys
1000  * struct test_obj {
1001  *      [...]
1002  *      struct rhash_head       node;
1003  * };
1004  *
1005  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1006  * {
1007  *      struct test_obj *obj = data;
1008  *
1009  *      return [... hash ...];
1010  * }
1011  *
1012  * struct rhashtable_params params = {
1013  *      .head_offset = offsetof(struct test_obj, node),
1014  *      .hashfn = jhash,
1015  *      .obj_hashfn = my_hash_fn,
1016  * };
1017  */
1018 int rhashtable_init(struct rhashtable *ht,
1019                     const struct rhashtable_params *params)
1020 {
1021         struct bucket_table *tbl;
1022         size_t size;
1023
1024         if ((!params->key_len && !params->obj_hashfn) ||
1025             (params->obj_hashfn && !params->obj_cmpfn))
1026                 return -EINVAL;
1027
1028         memset(ht, 0, sizeof(*ht));
1029         mutex_init(&ht->mutex);
1030         spin_lock_init(&ht->lock);
1031         memcpy(&ht->p, params, sizeof(*params));
1032
1033         if (params->min_size)
1034                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1035
1036         /* Cap total entries at 2^31 to avoid nelems overflow. */
1037         ht->max_elems = 1u << 31;
1038
1039         if (params->max_size) {
1040                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1041                 if (ht->p.max_size < ht->max_elems / 2)
1042                         ht->max_elems = ht->p.max_size * 2;
1043         }
1044
1045         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1046
1047         size = rounded_hashtable_size(&ht->p);
1048
1049         if (params->locks_mul)
1050                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1051         else
1052                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1053
1054         ht->key_len = ht->p.key_len;
1055         if (!params->hashfn) {
1056                 ht->p.hashfn = jhash;
1057
1058                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1059                         ht->key_len /= sizeof(u32);
1060                         ht->p.hashfn = rhashtable_jhash2;
1061                 }
1062         }
1063
1064         /*
1065          * This is api initialization and thus we need to guarantee the
1066          * initial rhashtable allocation. Upon failure, retry with the
1067          * smallest possible size with __GFP_NOFAIL semantics.
1068          */
1069         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1070         if (unlikely(tbl == NULL)) {
1071                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1072                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1073         }
1074
1075         atomic_set(&ht->nelems, 0);
1076
1077         RCU_INIT_POINTER(ht->tbl, tbl);
1078
1079         INIT_WORK(&ht->run_work, rht_deferred_worker);
1080
1081         return 0;
1082 }
1083 EXPORT_SYMBOL_GPL(rhashtable_init);
1084
1085 /**
1086  * rhltable_init - initialize a new hash list table
1087  * @hlt:        hash list table to be initialized
1088  * @params:     configuration parameters
1089  *
1090  * Initializes a new hash list table.
1091  *
1092  * See documentation for rhashtable_init.
1093  */
1094 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1095 {
1096         int err;
1097
1098         err = rhashtable_init(&hlt->ht, params);
1099         hlt->ht.rhlist = true;
1100         return err;
1101 }
1102 EXPORT_SYMBOL_GPL(rhltable_init);
1103
1104 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1105                                 void (*free_fn)(void *ptr, void *arg),
1106                                 void *arg)
1107 {
1108         struct rhlist_head *list;
1109
1110         if (!ht->rhlist) {
1111                 free_fn(rht_obj(ht, obj), arg);
1112                 return;
1113         }
1114
1115         list = container_of(obj, struct rhlist_head, rhead);
1116         do {
1117                 obj = &list->rhead;
1118                 list = rht_dereference(list->next, ht);
1119                 free_fn(rht_obj(ht, obj), arg);
1120         } while (list);
1121 }
1122
1123 /**
1124  * rhashtable_free_and_destroy - free elements and destroy hash table
1125  * @ht:         the hash table to destroy
1126  * @free_fn:    callback to release resources of element
1127  * @arg:        pointer passed to free_fn
1128  *
1129  * Stops an eventual async resize. If defined, invokes free_fn for each
1130  * element to releasal resources. Please note that RCU protected
1131  * readers may still be accessing the elements. Releasing of resources
1132  * must occur in a compatible manner. Then frees the bucket array.
1133  *
1134  * This function will eventually sleep to wait for an async resize
1135  * to complete. The caller is responsible that no further write operations
1136  * occurs in parallel.
1137  */
1138 void rhashtable_free_and_destroy(struct rhashtable *ht,
1139                                  void (*free_fn)(void *ptr, void *arg),
1140                                  void *arg)
1141 {
1142         struct bucket_table *tbl, *next_tbl;
1143         unsigned int i;
1144
1145         cancel_work_sync(&ht->run_work);
1146
1147         mutex_lock(&ht->mutex);
1148         tbl = rht_dereference(ht->tbl, ht);
1149 restart:
1150         if (free_fn) {
1151                 for (i = 0; i < tbl->size; i++) {
1152                         struct rhash_head *pos, *next;
1153
1154                         cond_resched();
1155                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1156                              next = !rht_is_a_nulls(pos) ?
1157                                         rht_dereference(pos->next, ht) : NULL;
1158                              !rht_is_a_nulls(pos);
1159                              pos = next,
1160                              next = !rht_is_a_nulls(pos) ?
1161                                         rht_dereference(pos->next, ht) : NULL)
1162                                 rhashtable_free_one(ht, pos, free_fn, arg);
1163                 }
1164         }
1165
1166         next_tbl = rht_dereference(tbl->future_tbl, ht);
1167         bucket_table_free(tbl);
1168         if (next_tbl) {
1169                 tbl = next_tbl;
1170                 goto restart;
1171         }
1172         mutex_unlock(&ht->mutex);
1173 }
1174 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1175
1176 void rhashtable_destroy(struct rhashtable *ht)
1177 {
1178         return rhashtable_free_and_destroy(ht, NULL, NULL);
1179 }
1180 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1181
1182 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1183                                             unsigned int hash)
1184 {
1185         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1186         static struct rhash_head __rcu *rhnull;
1187         unsigned int index = hash & ((1 << tbl->nest) - 1);
1188         unsigned int size = tbl->size >> tbl->nest;
1189         unsigned int subhash = hash;
1190         union nested_table *ntbl;
1191
1192         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1193         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1194         subhash >>= tbl->nest;
1195
1196         while (ntbl && size > (1 << shift)) {
1197                 index = subhash & ((1 << shift) - 1);
1198                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1199                                                   tbl, hash);
1200                 size >>= shift;
1201                 subhash >>= shift;
1202         }
1203
1204         if (!ntbl) {
1205                 if (!rhnull)
1206                         INIT_RHT_NULLS_HEAD(rhnull);
1207                 return &rhnull;
1208         }
1209
1210         return &ntbl[subhash].bucket;
1211
1212 }
1213 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1214
1215 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1216                                                    struct bucket_table *tbl,
1217                                                    unsigned int hash)
1218 {
1219         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1220         unsigned int index = hash & ((1 << tbl->nest) - 1);
1221         unsigned int size = tbl->size >> tbl->nest;
1222         union nested_table *ntbl;
1223
1224         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1225         hash >>= tbl->nest;
1226         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1227                                   size <= (1 << shift));
1228
1229         while (ntbl && size > (1 << shift)) {
1230                 index = hash & ((1 << shift) - 1);
1231                 size >>= shift;
1232                 hash >>= shift;
1233                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1234                                           size <= (1 << shift));
1235         }
1236
1237         if (!ntbl)
1238                 return NULL;
1239
1240         return &ntbl[hash].bucket;
1241
1242 }
1243 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);