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 union nested_table {
  36         union nested_table __rcu *table;
  37         struct rhash_head __rcu *bucket;
  38 };
  39
  40 static u32 head_hashfn(struct rhashtable *ht,
  41                        const struct bucket_table *tbl,
  42                        const struct rhash_head *he)
  43 {
  44         return rht_head_hashfn(ht, tbl, he, ht->p);
  45 }
  46
  47 #ifdef CONFIG_PROVE_LOCKING
  48 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  49
  50 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  51 {
  52         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  53 }
  54 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  55
  56 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  57 {
  58         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  59
  60         return (debug_locks) ? lockdep_is_held(lock) : 1;
  61 }
  62 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  63 #else
  64 #define ASSERT_RHT_MUTEX(HT)
  65 #endif
  66
  67
  68 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  69                               gfp_t gfp)
  70 {
  71         unsigned int i, size;
  72 #if defined(CONFIG_PROVE_LOCKING)
  73         unsigned int nr_pcpus = 2;
  74 #else
  75         unsigned int nr_pcpus = num_possible_cpus();
  76 #endif
  77
  78         nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
  79         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  80
  81         /* Never allocate more than 0.5 locks per bucket */
  82         size = min_t(unsigned int, size, tbl->size >> 1);
  83
  84         if (tbl->nest)
  85                 size = min(size, 1U << tbl->nest);
  86
  87         if (sizeof(spinlock_t) != 0) {
  88                 tbl->locks = NULL;
  89 #ifdef CONFIG_NUMA
  90                 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
  91                     gfp == GFP_KERNEL)
  92                         tbl->locks = vmalloc(size * sizeof(spinlock_t));
  93 #endif
  94                 if (gfp != GFP_KERNEL)
  95                         gfp |= __GFP_NOWARN | __GFP_NORETRY;
  96
  97                 if (!tbl->locks)
  98                         tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  99                                                    gfp);
 100                 if (!tbl->locks)
 101                         return -ENOMEM;
 102                 for (i = 0; i < size; i++)
 103                         spin_lock_init(&tbl->locks[i]);
 104         }
 105         tbl->locks_mask = size - 1;
 106
 107         return 0;
 108 }
 109
 110 static void nested_table_free(union nested_table *ntbl, unsigned int size)
 111 {
 112         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 113         const unsigned int len = 1 << shift;
 114         unsigned int i;
 115
 116         ntbl = rcu_dereference_raw(ntbl->table);
 117         if (!ntbl)
 118                 return;
 119
 120         if (size > len) {
 121                 size >>= shift;
 122                 for (i = 0; i < len; i++)
 123                         nested_table_free(ntbl + i, size);
 124         }
 125
 126         kfree(ntbl);
 127 }
 128
 129 static void nested_bucket_table_free(const struct bucket_table *tbl)
 130 {
 131         unsigned int size = tbl->size >> tbl->nest;
 132         unsigned int len = 1 << tbl->nest;
 133         union nested_table *ntbl;
 134         unsigned int i;
 135
 136         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
 137
 138         for (i = 0; i < len; i++)
 139                 nested_table_free(ntbl + i, size);
 140
 141         kfree(ntbl);
 142 }
 143
 144 static void bucket_table_free(const struct bucket_table *tbl)
 145 {
 146         if (tbl->nest)
 147                 nested_bucket_table_free(tbl);
 148
 149         if (tbl)
 150                 kvfree(tbl->locks);
 151
 152         kvfree(tbl);
 153 }
 154
 155 static void bucket_table_free_rcu(struct rcu_head *head)
 156 {
 157         bucket_table_free(container_of(head, struct bucket_table, rcu));
 158 }
 159
 160 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 161                                               union nested_table __rcu **prev,
 162                                               unsigned int shifted,
 163                                               unsigned int nhash)
 164 {
 165         union nested_table *ntbl;
 166         int i;
 167
 168         ntbl = rcu_dereference(*prev);
 169         if (ntbl)
 170                 return ntbl;
 171
 172         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 173
 174         if (ntbl && shifted) {
 175                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
 176                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
 177                                             (i << shifted) | nhash);
 178         }
 179
 180         rcu_assign_pointer(*prev, ntbl);
 181
 182         return ntbl;
 183 }
 184
 185 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 186                                                       size_t nbuckets,
 187                                                       gfp_t gfp)
 188 {
 189         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 190         struct bucket_table *tbl;
 191         size_t size;
 192
 193         if (nbuckets < (1 << (shift + 1)))
 194                 return NULL;
 195
 196         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 197
 198         tbl = kzalloc(size, gfp);
 199         if (!tbl)
 200                 return NULL;
 201
 202         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 203                                 0, 0)) {
 204                 kfree(tbl);
 205                 return NULL;
 206         }
 207
 208         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 209
 210         return tbl;
 211 }
 212
 213 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 214                                                size_t nbuckets,
 215                                                gfp_t gfp)
 216 {
 217         struct bucket_table *tbl = NULL;
 218         size_t size;
 219         int i;
 220
 221         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 222         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 223             gfp != GFP_KERNEL)
 224                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 225         if (tbl == NULL && gfp == GFP_KERNEL)
 226                 tbl = vzalloc(size);
 227
 228         size = nbuckets;
 229
 230         if (tbl == NULL && gfp != GFP_KERNEL) {
 231                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 232                 nbuckets = 0;
 233         }
 234         if (tbl == NULL)
 235                 return NULL;
 236
 237         tbl->size = size;
 238
 239         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 240                 bucket_table_free(tbl);
 241                 return NULL;
 242         }
 243
 244         INIT_LIST_HEAD(&tbl->walkers);
 245
 246         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 247
 248         for (i = 0; i < nbuckets; i++)
 249                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 250
 251         return tbl;
 252 }
 253
 254 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 255                                                   struct bucket_table *tbl)
 256 {
 257         struct bucket_table *new_tbl;
 258
 259         do {
 260                 new_tbl = tbl;
 261                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 262         } while (tbl);
 263
 264         return new_tbl;
 265 }
 266
 267 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 268 {
 269         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 270         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 271                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 272         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 273         int err = -EAGAIN;
 274         struct rhash_head *head, *next, *entry;
 275         spinlock_t *new_bucket_lock;
 276         unsigned int new_hash;
 277
 278         if (new_tbl->nest)
 279                 goto out;
 280
 281         err = -ENOENT;
 282
 283         rht_for_each(entry, old_tbl, old_hash) {
 284                 err = 0;
 285                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 286
 287                 if (rht_is_a_nulls(next))
 288                         break;
 289
 290                 pprev = &entry->next;
 291         }
 292
 293         if (err)
 294                 goto out;
 295
 296         new_hash = head_hashfn(ht, new_tbl, entry);
 297
 298         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 299
 300         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 301         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 302                                       new_tbl, new_hash);
 303
 304         RCU_INIT_POINTER(entry->next, head);
 305
 306         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 307         spin_unlock(new_bucket_lock);
 308
 309         rcu_assign_pointer(*pprev, next);
 310
 311 out:
 312         return err;
 313 }
 314
 315 static int rhashtable_rehash_chain(struct rhashtable *ht,
 316                                     unsigned int old_hash)
 317 {
 318         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 319         spinlock_t *old_bucket_lock;
 320         int err;
 321
 322         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 323
 324         spin_lock_bh(old_bucket_lock);
 325         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 326                 ;
 327
 328         if (err == -ENOENT) {
 329                 old_tbl->rehash++;
 330                 err = 0;
 331         }
 332         spin_unlock_bh(old_bucket_lock);
 333
 334         return err;
 335 }
 336
 337 static int rhashtable_rehash_attach(struct rhashtable *ht,
 338                                     struct bucket_table *old_tbl,
 339                                     struct bucket_table *new_tbl)
 340 {
 341         /* Protect future_tbl using the first bucket lock. */
 342         spin_lock_bh(old_tbl->locks);
 343
 344         /* Did somebody beat us to it? */
 345         if (rcu_access_pointer(old_tbl->future_tbl)) {
 346                 spin_unlock_bh(old_tbl->locks);
 347                 return -EEXIST;
 348         }
 349
 350         /* Make insertions go into the new, empty table right away. Deletions
 351          * and lookups will be attempted in both tables until we synchronize.
 352          */
 353         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 354
 355         spin_unlock_bh(old_tbl->locks);
 356
 357         return 0;
 358 }
 359
 360 static int rhashtable_rehash_table(struct rhashtable *ht)
 361 {
 362         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 363         struct bucket_table *new_tbl;
 364         struct rhashtable_walker *walker;
 365         unsigned int old_hash;
 366         int err;
 367
 368         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 369         if (!new_tbl)
 370                 return 0;
 371
 372         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 373                 err = rhashtable_rehash_chain(ht, old_hash);
 374                 if (err)
 375                         return err;
 376         }
 377
 378         /* Publish the new table pointer. */
 379         rcu_assign_pointer(ht->tbl, new_tbl);
 380
 381         spin_lock(&ht->lock);
 382         list_for_each_entry(walker, &old_tbl->walkers, list)
 383                 walker->tbl = NULL;
 384         spin_unlock(&ht->lock);
 385
 386         /* Wait for readers. All new readers will see the new
 387          * table, and thus no references to the old table will
 388          * remain.
 389          */
 390         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 391
 392         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 393 }
 394
 395 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 396                                    struct bucket_table *old_tbl,
 397                                    unsigned int size)
 398 {
 399         struct bucket_table *new_tbl;
 400         int err;
 401
 402         ASSERT_RHT_MUTEX(ht);
 403
 404         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 405         if (new_tbl == NULL)
 406                 return -ENOMEM;
 407
 408         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 409         if (err)
 410                 bucket_table_free(new_tbl);
 411
 412         return err;
 413 }
 414
 415 /**
 416  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 417  * @ht:         the hash table to shrink
 418  *
 419  * This function shrinks the hash table to fit, i.e., the smallest
 420  * size would not cause it to expand right away automatically.
 421  *
 422  * The caller must ensure that no concurrent resizing occurs by holding
 423  * ht->mutex.
 424  *
 425  * The caller must ensure that no concurrent table mutations take place.
 426  * It is however valid to have concurrent lookups if they are RCU protected.
 427  *
 428  * It is valid to have concurrent insertions and deletions protected by per
 429  * bucket locks or concurrent RCU protected lookups and traversals.
 430  */
 431 static int rhashtable_shrink(struct rhashtable *ht)
 432 {
 433         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 434         unsigned int nelems = atomic_read(&ht->nelems);
 435         unsigned int size = 0;
 436
 437         if (nelems)
 438                 size = roundup_pow_of_two(nelems * 3 / 2);
 439         if (size < ht->p.min_size)
 440                 size = ht->p.min_size;
 441
 442         if (old_tbl->size <= size)
 443                 return 0;
 444
 445         if (rht_dereference(old_tbl->future_tbl, ht))
 446                 return -EEXIST;
 447
 448         return rhashtable_rehash_alloc(ht, old_tbl, size);
 449 }
 450
 451 static void rht_deferred_worker(struct work_struct *work)
 452 {
 453         struct rhashtable *ht;
 454         struct bucket_table *tbl;
 455         int err = 0;
 456
 457         ht = container_of(work, struct rhashtable, run_work);
 458         mutex_lock(&ht->mutex);
 459
 460         tbl = rht_dereference(ht->tbl, ht);
 461         tbl = rhashtable_last_table(ht, tbl);
 462
 463         if (rht_grow_above_75(ht, tbl))
 464                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 465         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 466                 err = rhashtable_shrink(ht);
 467         else if (tbl->nest)
 468                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 469
 470         if (!err)
 471                 err = rhashtable_rehash_table(ht);
 472
 473         mutex_unlock(&ht->mutex);
 474
 475         if (err)
 476                 schedule_work(&ht->run_work);
 477 }
 478
 479 static int rhashtable_insert_rehash(struct rhashtable *ht,
 480                                     struct bucket_table *tbl)
 481 {
 482         struct bucket_table *old_tbl;
 483         struct bucket_table *new_tbl;
 484         unsigned int size;
 485         int err;
 486
 487         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 488
 489         size = tbl->size;
 490
 491         err = -EBUSY;
 492
 493         if (rht_grow_above_75(ht, tbl))
 494                 size *= 2;
 495         /* Do not schedule more than one rehash */
 496         else if (old_tbl != tbl)
 497                 goto fail;
 498
 499         err = -ENOMEM;
 500
 501         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 502         if (new_tbl == NULL)
 503                 goto fail;
 504
 505         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 506         if (err) {
 507                 bucket_table_free(new_tbl);
 508                 if (err == -EEXIST)
 509                         err = 0;
 510         } else
 511                 schedule_work(&ht->run_work);
 512
 513         return err;
 514
 515 fail:
 516         /* Do not fail the insert if someone else did a rehash. */
 517         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 518                 return 0;
 519
 520         /* Schedule async rehash to retry allocation in process context. */
 521         if (err == -ENOMEM)
 522                 schedule_work(&ht->run_work);
 523
 524         return err;
 525 }
 526
 527 static void *rhashtable_lookup_one(struct rhashtable *ht,
 528                                    struct bucket_table *tbl, unsigned int hash,
 529                                    const void *key, struct rhash_head *obj)
 530 {
 531         struct rhashtable_compare_arg arg = {
 532                 .ht = ht,
 533                 .key = key,
 534         };
 535         struct rhash_head __rcu **pprev;
 536         struct rhash_head *head;
 537         int elasticity;
 538
 539         elasticity = ht->elasticity;
 540         pprev = rht_bucket_var(tbl, hash);
 541         rht_for_each_continue(head, *pprev, tbl, hash) {
 542                 struct rhlist_head *list;
 543                 struct rhlist_head *plist;
 544
 545                 elasticity--;
 546                 if (!key ||
 547                     (ht->p.obj_cmpfn ?
 548                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 549                      rhashtable_compare(&arg, rht_obj(ht, head))))
 550                         continue;
 551
 552                 if (!ht->rhlist)
 553                         return rht_obj(ht, head);
 554
 555                 list = container_of(obj, struct rhlist_head, rhead);
 556                 plist = container_of(head, struct rhlist_head, rhead);
 557
 558                 RCU_INIT_POINTER(list->next, plist);
 559                 head = rht_dereference_bucket(head->next, tbl, hash);
 560                 RCU_INIT_POINTER(list->rhead.next, head);
 561                 rcu_assign_pointer(*pprev, obj);
 562
 563                 return NULL;
 564         }
 565
 566         if (elasticity <= 0)
 567                 return ERR_PTR(-EAGAIN);
 568
 569         return ERR_PTR(-ENOENT);
 570 }
 571
 572 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 573                                                   struct bucket_table *tbl,
 574                                                   unsigned int hash,
 575                                                   struct rhash_head *obj,
 576                                                   void *data)
 577 {
 578         struct rhash_head __rcu **pprev;
 579         struct bucket_table *new_tbl;
 580         struct rhash_head *head;
 581
 582         if (!IS_ERR_OR_NULL(data))
 583                 return ERR_PTR(-EEXIST);
 584
 585         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 586                 return ERR_CAST(data);
 587
 588         new_tbl = rcu_dereference(tbl->future_tbl);
 589         if (new_tbl)
 590                 return new_tbl;
 591
 592         if (PTR_ERR(data) != -ENOENT)
 593                 return ERR_CAST(data);
 594
 595         if (unlikely(rht_grow_above_max(ht, tbl)))
 596                 return ERR_PTR(-E2BIG);
 597
 598         if (unlikely(rht_grow_above_100(ht, tbl)))
 599                 return ERR_PTR(-EAGAIN);
 600
 601         pprev = rht_bucket_insert(ht, tbl, hash);
 602         if (!pprev)
 603                 return ERR_PTR(-ENOMEM);
 604
 605         head = rht_dereference_bucket(*pprev, tbl, hash);
 606
 607         RCU_INIT_POINTER(obj->next, head);
 608         if (ht->rhlist) {
 609                 struct rhlist_head *list;
 610
 611                 list = container_of(obj, struct rhlist_head, rhead);
 612                 RCU_INIT_POINTER(list->next, NULL);
 613         }
 614
 615         rcu_assign_pointer(*pprev, obj);
 616
 617         atomic_inc(&ht->nelems);
 618         if (rht_grow_above_75(ht, tbl))
 619                 schedule_work(&ht->run_work);
 620
 621         return NULL;
 622 }
 623
 624 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 625                                    struct rhash_head *obj)
 626 {
 627         struct bucket_table *new_tbl;
 628         struct bucket_table *tbl;
 629         unsigned int hash;
 630         spinlock_t *lock;
 631         void *data;
 632
 633         tbl = rcu_dereference(ht->tbl);
 634
 635         /* All insertions must grab the oldest table containing
 636          * the hashed bucket that is yet to be rehashed.
 637          */
 638         for (;;) {
 639                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 640                 lock = rht_bucket_lock(tbl, hash);
 641                 spin_lock_bh(lock);
 642
 643                 if (tbl->rehash <= hash)
 644                         break;
 645
 646                 spin_unlock_bh(lock);
 647                 tbl = rcu_dereference(tbl->future_tbl);
 648         }
 649
 650         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 651         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 652         if (PTR_ERR(new_tbl) != -EEXIST)
 653                 data = ERR_CAST(new_tbl);
 654
 655         while (!IS_ERR_OR_NULL(new_tbl)) {
 656                 tbl = new_tbl;
 657                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 658                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 659                                  SINGLE_DEPTH_NESTING);
 660
 661                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 662                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 663                 if (PTR_ERR(new_tbl) != -EEXIST)
 664                         data = ERR_CAST(new_tbl);
 665
 666                 spin_unlock(rht_bucket_lock(tbl, hash));
 667         }
 668
 669         spin_unlock_bh(lock);
 670
 671         if (PTR_ERR(data) == -EAGAIN)
 672                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 673                                -EAGAIN);
 674
 675         return data;
 676 }
 677
 678 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 679                              struct rhash_head *obj)
 680 {
 681         void *data;
 682
 683         do {
 684                 rcu_read_lock();
 685                 data = rhashtable_try_insert(ht, key, obj);
 686                 rcu_read_unlock();
 687         } while (PTR_ERR(data) == -EAGAIN);
 688
 689         return data;
 690 }
 691 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 692
 693 /**
 694  * rhashtable_walk_enter - Initialise an iterator
 695  * @ht:         Table to walk over
 696  * @iter:       Hash table Iterator
 697  *
 698  * This function prepares a hash table walk.
 699  *
 700  * Note that if you restart a walk after rhashtable_walk_stop you
 701  * may see the same object twice.  Also, you may miss objects if
 702  * there are removals in between rhashtable_walk_stop and the next
 703  * call to rhashtable_walk_start.
 704  *
 705  * For a completely stable walk you should construct your own data
 706  * structure outside the hash table.
 707  *
 708  * This function may sleep so you must not call it from interrupt
 709  * context or with spin locks held.
 710  *
 711  * You must call rhashtable_walk_exit after this function returns.
 712  */
 713 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 714 {
 715         iter->ht = ht;
 716         iter->p = NULL;
 717         iter->slot = 0;
 718         iter->skip = 0;
 719
 720         spin_lock(&ht->lock);
 721         iter->walker.tbl =
 722                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 723         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 724         spin_unlock(&ht->lock);
 725 }
 726 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 727
 728 /**
 729  * rhashtable_walk_exit - Free an iterator
 730  * @iter:       Hash table Iterator
 731  *
 732  * This function frees resources allocated by rhashtable_walk_init.
 733  */
 734 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 735 {
 736         spin_lock(&iter->ht->lock);
 737         if (iter->walker.tbl)
 738                 list_del(&iter->walker.list);
 739         spin_unlock(&iter->ht->lock);
 740 }
 741 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 742
 743 /**
 744  * rhashtable_walk_start - Start a hash table walk
 745  * @iter:       Hash table iterator
 746  *
 747  * Start a hash table walk.  Note that we take the RCU lock in all
 748  * cases including when we return an error.  So you must always call
 749  * rhashtable_walk_stop to clean up.
 750  *
 751  * Returns zero if successful.
 752  *
 753  * Returns -EAGAIN if resize event occured.  Note that the iterator
 754  * will rewind back to the beginning and you may use it immediately
 755  * by calling rhashtable_walk_next.
 756  */
 757 int rhashtable_walk_start(struct rhashtable_iter *iter)
 758         __acquires(RCU)
 759 {
 760         struct rhashtable *ht = iter->ht;
 761
 762         rcu_read_lock();
 763
 764         spin_lock(&ht->lock);
 765         if (iter->walker.tbl)
 766                 list_del(&iter->walker.list);
 767         spin_unlock(&ht->lock);
 768
 769         if (!iter->walker.tbl) {
 770                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 771                 return -EAGAIN;
 772         }
 773
 774         return 0;
 775 }
 776 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 777
 778 /**
 779  * rhashtable_walk_next - Return the next object and advance the iterator
 780  * @iter:       Hash table iterator
 781  *
 782  * Note that you must call rhashtable_walk_stop when you are finished
 783  * with the walk.
 784  *
 785  * Returns the next object or NULL when the end of the table is reached.
 786  *
 787  * Returns -EAGAIN if resize event occured.  Note that the iterator
 788  * will rewind back to the beginning and you may continue to use it.
 789  */
 790 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 791 {
 792         struct bucket_table *tbl = iter->walker.tbl;
 793         struct rhlist_head *list = iter->list;
 794         struct rhashtable *ht = iter->ht;
 795         struct rhash_head *p = iter->p;
 796         bool rhlist = ht->rhlist;
 797
 798         if (p) {
 799                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 800                         p = rcu_dereference(p->next);
 801                         list = container_of(p, struct rhlist_head, rhead);
 802                 }
 803                 goto next;
 804         }
 805
 806         for (; iter->slot < tbl->size; iter->slot++) {
 807                 int skip = iter->skip;
 808
 809                 rht_for_each_rcu(p, tbl, iter->slot) {
 810                         if (rhlist) {
 811                                 list = container_of(p, struct rhlist_head,
 812                                                     rhead);
 813                                 do {
 814                                         if (!skip)
 815                                                 goto next;
 816                                         skip--;
 817                                         list = rcu_dereference(list->next);
 818                                 } while (list);
 819
 820                                 continue;
 821                         }
 822                         if (!skip)
 823                                 break;
 824                         skip--;
 825                 }
 826
 827 next:
 828                 if (!rht_is_a_nulls(p)) {
 829                         iter->skip++;
 830                         iter->p = p;
 831                         iter->list = list;
 832                         return rht_obj(ht, rhlist ? &list->rhead : p);
 833                 }
 834
 835                 iter->skip = 0;
 836         }
 837
 838         iter->p = NULL;
 839
 840         /* Ensure we see any new tables. */
 841         smp_rmb();
 842
 843         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 844         if (iter->walker.tbl) {
 845                 iter->slot = 0;
 846                 iter->skip = 0;
 847                 return ERR_PTR(-EAGAIN);
 848         }
 849
 850         return NULL;
 851 }
 852 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 853
 854 /**
 855  * rhashtable_walk_stop - Finish a hash table walk
 856  * @iter:       Hash table iterator
 857  *
 858  * Finish a hash table walk.
 859  */
 860 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 861         __releases(RCU)
 862 {
 863         struct rhashtable *ht;
 864         struct bucket_table *tbl = iter->walker.tbl;
 865
 866         if (!tbl)
 867                 goto out;
 868
 869         ht = iter->ht;
 870
 871         spin_lock(&ht->lock);
 872         if (tbl->rehash < tbl->size)
 873                 list_add(&iter->walker.list, &tbl->walkers);
 874         else
 875                 iter->walker.tbl = NULL;
 876         spin_unlock(&ht->lock);
 877
 878         iter->p = NULL;
 879
 880 out:
 881         rcu_read_unlock();
 882 }
 883 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 884
 885 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 886 {
 887         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 888                    (unsigned long)params->min_size);
 889 }
 890
 891 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 892 {
 893         return jhash2(key, length, seed);
 894 }
 895
 896 /**
 897  * rhashtable_init - initialize a new hash table
 898  * @ht:         hash table to be initialized
 899  * @params:     configuration parameters
 900  *
 901  * Initializes a new hash table based on the provided configuration
 902  * parameters. A table can be configured either with a variable or
 903  * fixed length key:
 904  *
 905  * Configuration Example 1: Fixed length keys
 906  * struct test_obj {
 907  *      int                     key;
 908  *      void *                  my_member;
 909  *      struct rhash_head       node;
 910  * };
 911  *
 912  * struct rhashtable_params params = {
 913  *      .head_offset = offsetof(struct test_obj, node),
 914  *      .key_offset = offsetof(struct test_obj, key),
 915  *      .key_len = sizeof(int),
 916  *      .hashfn = jhash,
 917  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 918  * };
 919  *
 920  * Configuration Example 2: Variable length keys
 921  * struct test_obj {
 922  *      [...]
 923  *      struct rhash_head       node;
 924  * };
 925  *
 926  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 927  * {
 928  *      struct test_obj *obj = data;
 929  *
 930  *      return [... hash ...];
 931  * }
 932  *
 933  * struct rhashtable_params params = {
 934  *      .head_offset = offsetof(struct test_obj, node),
 935  *      .hashfn = jhash,
 936  *      .obj_hashfn = my_hash_fn,
 937  * };
 938  */
 939 int rhashtable_init(struct rhashtable *ht,
 940                     const struct rhashtable_params *params)
 941 {
 942         struct bucket_table *tbl;
 943         size_t size;
 944
 945         size = HASH_DEFAULT_SIZE;
 946
 947         if ((!params->key_len && !params->obj_hashfn) ||
 948             (params->obj_hashfn && !params->obj_cmpfn))
 949                 return -EINVAL;
 950
 951         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 952                 return -EINVAL;
 953
 954         memset(ht, 0, sizeof(*ht));
 955         mutex_init(&ht->mutex);
 956         spin_lock_init(&ht->lock);
 957         memcpy(&ht->p, params, sizeof(*params));
 958
 959         if (params->min_size)
 960                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 961
 962         if (params->max_size)
 963                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 964
 965         if (params->insecure_max_entries)
 966                 ht->p.insecure_max_entries =
 967                         rounddown_pow_of_two(params->insecure_max_entries);
 968         else
 969                 ht->p.insecure_max_entries = ht->p.max_size * 2;
 970
 971         ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
 972
 973         if (params->nelem_hint)
 974                 size = rounded_hashtable_size(&ht->p);
 975
 976         /* The maximum (not average) chain length grows with the
 977          * size of the hash table, at a rate of (log N)/(log log N).
 978          * The value of 16 is selected so that even if the hash
 979          * table grew to 2^32 you would not expect the maximum
 980          * chain length to exceed it unless we are under attack
 981          * (or extremely unlucky).
 982          *
 983          * As this limit is only to detect attacks, we don't need
 984          * to set it to a lower value as you'd need the chain
 985          * length to vastly exceed 16 to have any real effect
 986          * on the system.
 987          */
 988         if (!params->insecure_elasticity)
 989                 ht->elasticity = 16;
 990
 991         if (params->locks_mul)
 992                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 993         else
 994                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 995
 996         ht->key_len = ht->p.key_len;
 997         if (!params->hashfn) {
 998                 ht->p.hashfn = jhash;
 999
1000                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1001                         ht->key_len /= sizeof(u32);
1002                         ht->p.hashfn = rhashtable_jhash2;
1003                 }
1004         }
1005
1006         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1007         if (tbl == NULL)
1008                 return -ENOMEM;
1009
1010         atomic_set(&ht->nelems, 0);
1011
1012         RCU_INIT_POINTER(ht->tbl, tbl);
1013
1014         INIT_WORK(&ht->run_work, rht_deferred_worker);
1015
1016         return 0;
1017 }
1018 EXPORT_SYMBOL_GPL(rhashtable_init);
1019
1020 /**
1021  * rhltable_init - initialize a new hash list table
1022  * @hlt:        hash list table to be initialized
1023  * @params:     configuration parameters
1024  *
1025  * Initializes a new hash list table.
1026  *
1027  * See documentation for rhashtable_init.
1028  */
1029 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1030 {
1031         int err;
1032
1033         /* No rhlist NULLs marking for now. */
1034         if (params->nulls_base)
1035                 return -EINVAL;
1036
1037         err = rhashtable_init(&hlt->ht, params);
1038         hlt->ht.rhlist = true;
1039         return err;
1040 }
1041 EXPORT_SYMBOL_GPL(rhltable_init);
1042
1043 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1044                                 void (*free_fn)(void *ptr, void *arg),
1045                                 void *arg)
1046 {
1047         struct rhlist_head *list;
1048
1049         if (!ht->rhlist) {
1050                 free_fn(rht_obj(ht, obj), arg);
1051                 return;
1052         }
1053
1054         list = container_of(obj, struct rhlist_head, rhead);
1055         do {
1056                 obj = &list->rhead;
1057                 list = rht_dereference(list->next, ht);
1058                 free_fn(rht_obj(ht, obj), arg);
1059         } while (list);
1060 }
1061
1062 /**
1063  * rhashtable_free_and_destroy - free elements and destroy hash table
1064  * @ht:         the hash table to destroy
1065  * @free_fn:    callback to release resources of element
1066  * @arg:        pointer passed to free_fn
1067  *
1068  * Stops an eventual async resize. If defined, invokes free_fn for each
1069  * element to releasal resources. Please note that RCU protected
1070  * readers may still be accessing the elements. Releasing of resources
1071  * must occur in a compatible manner. Then frees the bucket array.
1072  *
1073  * This function will eventually sleep to wait for an async resize
1074  * to complete. The caller is responsible that no further write operations
1075  * occurs in parallel.
1076  */
1077 void rhashtable_free_and_destroy(struct rhashtable *ht,
1078                                  void (*free_fn)(void *ptr, void *arg),
1079                                  void *arg)
1080 {
1081         struct bucket_table *tbl;
1082         unsigned int i;
1083
1084         cancel_work_sync(&ht->run_work);
1085
1086         mutex_lock(&ht->mutex);
1087         tbl = rht_dereference(ht->tbl, ht);
1088         if (free_fn) {
1089                 for (i = 0; i < tbl->size; i++) {
1090                         struct rhash_head *pos, *next;
1091
1092                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1093                              next = !rht_is_a_nulls(pos) ?
1094                                         rht_dereference(pos->next, ht) : NULL;
1095                              !rht_is_a_nulls(pos);
1096                              pos = next,
1097                              next = !rht_is_a_nulls(pos) ?
1098                                         rht_dereference(pos->next, ht) : NULL)
1099                                 rhashtable_free_one(ht, pos, free_fn, arg);
1100                 }
1101         }
1102
1103         bucket_table_free(tbl);
1104         mutex_unlock(&ht->mutex);
1105 }
1106 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1107
1108 void rhashtable_destroy(struct rhashtable *ht)
1109 {
1110         return rhashtable_free_and_destroy(ht, NULL, NULL);
1111 }
1112 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1113
1114 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1115                                             unsigned int hash)
1116 {
1117         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1118         static struct rhash_head __rcu *rhnull =
1119                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1120         unsigned int index = hash & ((1 << tbl->nest) - 1);
1121         unsigned int size = tbl->size >> tbl->nest;
1122         unsigned int subhash = hash;
1123         union nested_table *ntbl;
1124
1125         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1126         ntbl = rht_dereference_bucket(ntbl[index].table, tbl, hash);
1127         subhash >>= tbl->nest;
1128
1129         while (ntbl && size > (1 << shift)) {
1130                 index = subhash & ((1 << shift) - 1);
1131                 ntbl = rht_dereference_bucket(ntbl[index].table, tbl, hash);
1132                 size >>= shift;
1133                 subhash >>= shift;
1134         }
1135
1136         if (!ntbl)
1137                 return &rhnull;
1138
1139         return &ntbl[subhash].bucket;
1140
1141 }
1142 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1143
1144 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1145                                                    struct bucket_table *tbl,
1146                                                    unsigned int hash)
1147 {
1148         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1149         unsigned int index = hash & ((1 << tbl->nest) - 1);
1150         unsigned int size = tbl->size >> tbl->nest;
1151         union nested_table *ntbl;
1152         unsigned int shifted;
1153         unsigned int nhash;
1154
1155         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1156         hash >>= tbl->nest;
1157         nhash = index;
1158         shifted = tbl->nest;
1159         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1160                                   size <= (1 << shift) ? shifted : 0, nhash);
1161
1162         while (ntbl && size > (1 << shift)) {
1163                 index = hash & ((1 << shift) - 1);
1164                 size >>= shift;
1165                 hash >>= shift;
1166                 nhash |= index << shifted;
1167                 shifted += shift;
1168                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1169                                           size <= (1 << shift) ? shifted : 0,
1170                                           nhash);
1171         }
1172
1173         if (!ntbl)
1174                 return NULL;
1175
1176         return &ntbl[hash].bucket;
1177
1178 }
1179 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);