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