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