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
  69 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  70                               gfp_t gfp)
  71 {
  72         unsigned int i, size;
  73 #if defined(CONFIG_PROVE_LOCKING)
  74         unsigned int nr_pcpus = 2;
  75 #else
  76         unsigned int nr_pcpus = num_possible_cpus();
  77 #endif
  78
  79         nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
  80         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  81
  82         /* Never allocate more than 0.5 locks per bucket */
  83         size = min_t(unsigned int, size, tbl->size >> 1);
  84
  85         if (tbl->nest)
  86                 size = min(size, 1U << tbl->nest);
  87
  88         if (sizeof(spinlock_t) != 0) {
  89                 if (gfpflags_allow_blocking(gfp))
  90                         tbl->locks = kvmalloc(size * sizeof(spinlock_t), gfp);
  91                 else
  92                         tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  93                                                    gfp);
  94                 if (!tbl->locks)
  95                         return -ENOMEM;
  96                 for (i = 0; i < size; i++)
  97                         spin_lock_init(&tbl->locks[i]);
  98         }
  99         tbl->locks_mask = size - 1;
 100
 101         return 0;
 102 }
 103
 104 static void nested_table_free(union nested_table *ntbl, unsigned int size)
 105 {
 106         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 107         const unsigned int len = 1 << shift;
 108         unsigned int i;
 109
 110         ntbl = rcu_dereference_raw(ntbl->table);
 111         if (!ntbl)
 112                 return;
 113
 114         if (size > len) {
 115                 size >>= shift;
 116                 for (i = 0; i < len; i++)
 117                         nested_table_free(ntbl + i, size);
 118         }
 119
 120         kfree(ntbl);
 121 }
 122
 123 static void nested_bucket_table_free(const struct bucket_table *tbl)
 124 {
 125         unsigned int size = tbl->size >> tbl->nest;
 126         unsigned int len = 1 << tbl->nest;
 127         union nested_table *ntbl;
 128         unsigned int i;
 129
 130         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
 131
 132         for (i = 0; i < len; i++)
 133                 nested_table_free(ntbl + i, size);
 134
 135         kfree(ntbl);
 136 }
 137
 138 static void bucket_table_free(const struct bucket_table *tbl)
 139 {
 140         if (tbl->nest)
 141                 nested_bucket_table_free(tbl);
 142
 143         kvfree(tbl->locks);
 144         kvfree(tbl);
 145 }
 146
 147 static void bucket_table_free_rcu(struct rcu_head *head)
 148 {
 149         bucket_table_free(container_of(head, struct bucket_table, rcu));
 150 }
 151
 152 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 153                                               union nested_table __rcu **prev,
 154                                               unsigned int shifted,
 155                                               unsigned int nhash)
 156 {
 157         union nested_table *ntbl;
 158         int i;
 159
 160         ntbl = rcu_dereference(*prev);
 161         if (ntbl)
 162                 return ntbl;
 163
 164         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 165
 166         if (ntbl && shifted) {
 167                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
 168                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
 169                                             (i << shifted) | nhash);
 170         }
 171
 172         rcu_assign_pointer(*prev, ntbl);
 173
 174         return ntbl;
 175 }
 176
 177 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 178                                                       size_t nbuckets,
 179                                                       gfp_t gfp)
 180 {
 181         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 182         struct bucket_table *tbl;
 183         size_t size;
 184
 185         if (nbuckets < (1 << (shift + 1)))
 186                 return NULL;
 187
 188         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 189
 190         tbl = kzalloc(size, gfp);
 191         if (!tbl)
 192                 return NULL;
 193
 194         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 195                                 0, 0)) {
 196                 kfree(tbl);
 197                 return NULL;
 198         }
 199
 200         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 201
 202         return tbl;
 203 }
 204
 205 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 206                                                size_t nbuckets,
 207                                                gfp_t gfp)
 208 {
 209         struct bucket_table *tbl = NULL;
 210         size_t size;
 211         int i;
 212
 213         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 214         if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
 215             gfp != GFP_KERNEL)
 216                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 217         if (tbl == NULL && gfp == GFP_KERNEL)
 218                 tbl = vzalloc(size);
 219
 220         size = nbuckets;
 221
 222         if (tbl == NULL && gfp != GFP_KERNEL) {
 223                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 224                 nbuckets = 0;
 225         }
 226         if (tbl == NULL)
 227                 return NULL;
 228
 229         tbl->size = size;
 230
 231         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 232                 bucket_table_free(tbl);
 233                 return NULL;
 234         }
 235
 236         INIT_LIST_HEAD(&tbl->walkers);
 237
 238         get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
 239
 240         for (i = 0; i < nbuckets; i++)
 241                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 242
 243         return tbl;
 244 }
 245
 246 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 247                                                   struct bucket_table *tbl)
 248 {
 249         struct bucket_table *new_tbl;
 250
 251         do {
 252                 new_tbl = tbl;
 253                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 254         } while (tbl);
 255
 256         return new_tbl;
 257 }
 258
 259 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 260 {
 261         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 262         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 263                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 264         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 265         int err = -EAGAIN;
 266         struct rhash_head *head, *next, *entry;
 267         spinlock_t *new_bucket_lock;
 268         unsigned int new_hash;
 269
 270         if (new_tbl->nest)
 271                 goto out;
 272
 273         err = -ENOENT;
 274
 275         rht_for_each(entry, old_tbl, old_hash) {
 276                 err = 0;
 277                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 278
 279                 if (rht_is_a_nulls(next))
 280                         break;
 281
 282                 pprev = &entry->next;
 283         }
 284
 285         if (err)
 286                 goto out;
 287
 288         new_hash = head_hashfn(ht, new_tbl, entry);
 289
 290         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 291
 292         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 293         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 294                                       new_tbl, new_hash);
 295
 296         RCU_INIT_POINTER(entry->next, head);
 297
 298         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 299         spin_unlock(new_bucket_lock);
 300
 301         rcu_assign_pointer(*pprev, next);
 302
 303 out:
 304         return err;
 305 }
 306
 307 static int rhashtable_rehash_chain(struct rhashtable *ht,
 308                                     unsigned int old_hash)
 309 {
 310         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 311         spinlock_t *old_bucket_lock;
 312         int err;
 313
 314         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 315
 316         spin_lock_bh(old_bucket_lock);
 317         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 318                 ;
 319
 320         if (err == -ENOENT) {
 321                 old_tbl->rehash++;
 322                 err = 0;
 323         }
 324         spin_unlock_bh(old_bucket_lock);
 325
 326         return err;
 327 }
 328
 329 static int rhashtable_rehash_attach(struct rhashtable *ht,
 330                                     struct bucket_table *old_tbl,
 331                                     struct bucket_table *new_tbl)
 332 {
 333         /* Protect future_tbl using the first bucket lock. */
 334         spin_lock_bh(old_tbl->locks);
 335
 336         /* Did somebody beat us to it? */
 337         if (rcu_access_pointer(old_tbl->future_tbl)) {
 338                 spin_unlock_bh(old_tbl->locks);
 339                 return -EEXIST;
 340         }
 341
 342         /* Make insertions go into the new, empty table right away. Deletions
 343          * and lookups will be attempted in both tables until we synchronize.
 344          */
 345         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 346
 347         spin_unlock_bh(old_tbl->locks);
 348
 349         return 0;
 350 }
 351
 352 static int rhashtable_rehash_table(struct rhashtable *ht)
 353 {
 354         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 355         struct bucket_table *new_tbl;
 356         struct rhashtable_walker *walker;
 357         unsigned int old_hash;
 358         int err;
 359
 360         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 361         if (!new_tbl)
 362                 return 0;
 363
 364         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 365                 err = rhashtable_rehash_chain(ht, old_hash);
 366                 if (err)
 367                         return err;
 368         }
 369
 370         /* Publish the new table pointer. */
 371         rcu_assign_pointer(ht->tbl, new_tbl);
 372
 373         spin_lock(&ht->lock);
 374         list_for_each_entry(walker, &old_tbl->walkers, list)
 375                 walker->tbl = NULL;
 376         spin_unlock(&ht->lock);
 377
 378         /* Wait for readers. All new readers will see the new
 379          * table, and thus no references to the old table will
 380          * remain.
 381          */
 382         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 383
 384         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 385 }
 386
 387 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 388                                    struct bucket_table *old_tbl,
 389                                    unsigned int size)
 390 {
 391         struct bucket_table *new_tbl;
 392         int err;
 393
 394         ASSERT_RHT_MUTEX(ht);
 395
 396         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 397         if (new_tbl == NULL)
 398                 return -ENOMEM;
 399
 400         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 401         if (err)
 402                 bucket_table_free(new_tbl);
 403
 404         return err;
 405 }
 406
 407 /**
 408  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 409  * @ht:         the hash table to shrink
 410  *
 411  * This function shrinks the hash table to fit, i.e., the smallest
 412  * size would not cause it to expand right away automatically.
 413  *
 414  * The caller must ensure that no concurrent resizing occurs by holding
 415  * ht->mutex.
 416  *
 417  * The caller must ensure that no concurrent table mutations take place.
 418  * It is however valid to have concurrent lookups if they are RCU protected.
 419  *
 420  * It is valid to have concurrent insertions and deletions protected by per
 421  * bucket locks or concurrent RCU protected lookups and traversals.
 422  */
 423 static int rhashtable_shrink(struct rhashtable *ht)
 424 {
 425         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 426         unsigned int nelems = atomic_read(&ht->nelems);
 427         unsigned int size = 0;
 428
 429         if (nelems)
 430                 size = roundup_pow_of_two(nelems * 3 / 2);
 431         if (size < ht->p.min_size)
 432                 size = ht->p.min_size;
 433
 434         if (old_tbl->size <= size)
 435                 return 0;
 436
 437         if (rht_dereference(old_tbl->future_tbl, ht))
 438                 return -EEXIST;
 439
 440         return rhashtable_rehash_alloc(ht, old_tbl, size);
 441 }
 442
 443 static void rht_deferred_worker(struct work_struct *work)
 444 {
 445         struct rhashtable *ht;
 446         struct bucket_table *tbl;
 447         int err = 0;
 448
 449         ht = container_of(work, struct rhashtable, run_work);
 450         mutex_lock(&ht->mutex);
 451
 452         tbl = rht_dereference(ht->tbl, ht);
 453         tbl = rhashtable_last_table(ht, tbl);
 454
 455         if (rht_grow_above_75(ht, tbl))
 456                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 457         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 458                 err = rhashtable_shrink(ht);
 459         else if (tbl->nest)
 460                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 461
 462         if (!err)
 463                 err = rhashtable_rehash_table(ht);
 464
 465         mutex_unlock(&ht->mutex);
 466
 467         if (err)
 468                 schedule_work(&ht->run_work);
 469 }
 470
 471 static int rhashtable_insert_rehash(struct rhashtable *ht,
 472                                     struct bucket_table *tbl)
 473 {
 474         struct bucket_table *old_tbl;
 475         struct bucket_table *new_tbl;
 476         unsigned int size;
 477         int err;
 478
 479         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 480
 481         size = tbl->size;
 482
 483         err = -EBUSY;
 484
 485         if (rht_grow_above_75(ht, tbl))
 486                 size *= 2;
 487         /* Do not schedule more than one rehash */
 488         else if (old_tbl != tbl)
 489                 goto fail;
 490
 491         err = -ENOMEM;
 492
 493         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 494         if (new_tbl == NULL)
 495                 goto fail;
 496
 497         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 498         if (err) {
 499                 bucket_table_free(new_tbl);
 500                 if (err == -EEXIST)
 501                         err = 0;
 502         } else
 503                 schedule_work(&ht->run_work);
 504
 505         return err;
 506
 507 fail:
 508         /* Do not fail the insert if someone else did a rehash. */
 509         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 510                 return 0;
 511
 512         /* Schedule async rehash to retry allocation in process context. */
 513         if (err == -ENOMEM)
 514                 schedule_work(&ht->run_work);
 515
 516         return err;
 517 }
 518
 519 static void *rhashtable_lookup_one(struct rhashtable *ht,
 520                                    struct bucket_table *tbl, unsigned int hash,
 521                                    const void *key, struct rhash_head *obj)
 522 {
 523         struct rhashtable_compare_arg arg = {
 524                 .ht = ht,
 525                 .key = key,
 526         };
 527         struct rhash_head __rcu **pprev;
 528         struct rhash_head *head;
 529         int elasticity;
 530
 531         elasticity = RHT_ELASTICITY;
 532         pprev = rht_bucket_var(tbl, hash);
 533         rht_for_each_continue(head, *pprev, tbl, hash) {
 534                 struct rhlist_head *list;
 535                 struct rhlist_head *plist;
 536
 537                 elasticity--;
 538                 if (!key ||
 539                     (ht->p.obj_cmpfn ?
 540                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 541                      rhashtable_compare(&arg, rht_obj(ht, head))))
 542                         continue;
 543
 544                 if (!ht->rhlist)
 545                         return rht_obj(ht, head);
 546
 547                 list = container_of(obj, struct rhlist_head, rhead);
 548                 plist = container_of(head, struct rhlist_head, rhead);
 549
 550                 RCU_INIT_POINTER(list->next, plist);
 551                 head = rht_dereference_bucket(head->next, tbl, hash);
 552                 RCU_INIT_POINTER(list->rhead.next, head);
 553                 rcu_assign_pointer(*pprev, obj);
 554
 555                 return NULL;
 556         }
 557
 558         if (elasticity <= 0)
 559                 return ERR_PTR(-EAGAIN);
 560
 561         return ERR_PTR(-ENOENT);
 562 }
 563
 564 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 565                                                   struct bucket_table *tbl,
 566                                                   unsigned int hash,
 567                                                   struct rhash_head *obj,
 568                                                   void *data)
 569 {
 570         struct rhash_head __rcu **pprev;
 571         struct bucket_table *new_tbl;
 572         struct rhash_head *head;
 573
 574         if (!IS_ERR_OR_NULL(data))
 575                 return ERR_PTR(-EEXIST);
 576
 577         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 578                 return ERR_CAST(data);
 579
 580         new_tbl = rcu_dereference(tbl->future_tbl);
 581         if (new_tbl)
 582                 return new_tbl;
 583
 584         if (PTR_ERR(data) != -ENOENT)
 585                 return ERR_CAST(data);
 586
 587         if (unlikely(rht_grow_above_max(ht, tbl)))
 588                 return ERR_PTR(-E2BIG);
 589
 590         if (unlikely(rht_grow_above_100(ht, tbl)))
 591                 return ERR_PTR(-EAGAIN);
 592
 593         pprev = rht_bucket_insert(ht, tbl, hash);
 594         if (!pprev)
 595                 return ERR_PTR(-ENOMEM);
 596
 597         head = rht_dereference_bucket(*pprev, tbl, hash);
 598
 599         RCU_INIT_POINTER(obj->next, head);
 600         if (ht->rhlist) {
 601                 struct rhlist_head *list;
 602
 603                 list = container_of(obj, struct rhlist_head, rhead);
 604                 RCU_INIT_POINTER(list->next, NULL);
 605         }
 606
 607         rcu_assign_pointer(*pprev, obj);
 608
 609         atomic_inc(&ht->nelems);
 610         if (rht_grow_above_75(ht, tbl))
 611                 schedule_work(&ht->run_work);
 612
 613         return NULL;
 614 }
 615
 616 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 617                                    struct rhash_head *obj)
 618 {
 619         struct bucket_table *new_tbl;
 620         struct bucket_table *tbl;
 621         unsigned int hash;
 622         spinlock_t *lock;
 623         void *data;
 624
 625         tbl = rcu_dereference(ht->tbl);
 626
 627         /* All insertions must grab the oldest table containing
 628          * the hashed bucket that is yet to be rehashed.
 629          */
 630         for (;;) {
 631                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 632                 lock = rht_bucket_lock(tbl, hash);
 633                 spin_lock_bh(lock);
 634
 635                 if (tbl->rehash <= hash)
 636                         break;
 637
 638                 spin_unlock_bh(lock);
 639                 tbl = rcu_dereference(tbl->future_tbl);
 640         }
 641
 642         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 643         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 644         if (PTR_ERR(new_tbl) != -EEXIST)
 645                 data = ERR_CAST(new_tbl);
 646
 647         while (!IS_ERR_OR_NULL(new_tbl)) {
 648                 tbl = new_tbl;
 649                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 650                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 651                                  SINGLE_DEPTH_NESTING);
 652
 653                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 654                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 655                 if (PTR_ERR(new_tbl) != -EEXIST)
 656                         data = ERR_CAST(new_tbl);
 657
 658                 spin_unlock(rht_bucket_lock(tbl, hash));
 659         }
 660
 661         spin_unlock_bh(lock);
 662
 663         if (PTR_ERR(data) == -EAGAIN)
 664                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 665                                -EAGAIN);
 666
 667         return data;
 668 }
 669
 670 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 671                              struct rhash_head *obj)
 672 {
 673         void *data;
 674
 675         do {
 676                 rcu_read_lock();
 677                 data = rhashtable_try_insert(ht, key, obj);
 678                 rcu_read_unlock();
 679         } while (PTR_ERR(data) == -EAGAIN);
 680
 681         return data;
 682 }
 683 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 684
 685 /**
 686  * rhashtable_walk_enter - Initialise an iterator
 687  * @ht:         Table to walk over
 688  * @iter:       Hash table Iterator
 689  *
 690  * This function prepares a hash table walk.
 691  *
 692  * Note that if you restart a walk after rhashtable_walk_stop you
 693  * may see the same object twice.  Also, you may miss objects if
 694  * there are removals in between rhashtable_walk_stop and the next
 695  * call to rhashtable_walk_start.
 696  *
 697  * For a completely stable walk you should construct your own data
 698  * structure outside the hash table.
 699  *
 700  * This function may sleep so you must not call it from interrupt
 701  * context or with spin locks held.
 702  *
 703  * You must call rhashtable_walk_exit after this function returns.
 704  */
 705 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 706 {
 707         iter->ht = ht;
 708         iter->p = NULL;
 709         iter->slot = 0;
 710         iter->skip = 0;
 711
 712         spin_lock(&ht->lock);
 713         iter->walker.tbl =
 714                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 715         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 716         spin_unlock(&ht->lock);
 717 }
 718 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 719
 720 /**
 721  * rhashtable_walk_exit - Free an iterator
 722  * @iter:       Hash table Iterator
 723  *
 724  * This function frees resources allocated by rhashtable_walk_init.
 725  */
 726 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 727 {
 728         spin_lock(&iter->ht->lock);
 729         if (iter->walker.tbl)
 730                 list_del(&iter->walker.list);
 731         spin_unlock(&iter->ht->lock);
 732 }
 733 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 734
 735 /**
 736  * rhashtable_walk_start - Start a hash table walk
 737  * @iter:       Hash table iterator
 738  *
 739  * Start a hash table walk.  Note that we take the RCU lock in all
 740  * cases including when we return an error.  So you must always call
 741  * rhashtable_walk_stop to clean up.
 742  *
 743  * Returns zero if successful.
 744  *
 745  * Returns -EAGAIN if resize event occured.  Note that the iterator
 746  * will rewind back to the beginning and you may use it immediately
 747  * by calling rhashtable_walk_next.
 748  */
 749 int rhashtable_walk_start(struct rhashtable_iter *iter)
 750         __acquires(RCU)
 751 {
 752         struct rhashtable *ht = iter->ht;
 753
 754         rcu_read_lock();
 755
 756         spin_lock(&ht->lock);
 757         if (iter->walker.tbl)
 758                 list_del(&iter->walker.list);
 759         spin_unlock(&ht->lock);
 760
 761         if (!iter->walker.tbl) {
 762                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 763                 return -EAGAIN;
 764         }
 765
 766         return 0;
 767 }
 768 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 769
 770 /**
 771  * rhashtable_walk_next - Return the next object and advance the iterator
 772  * @iter:       Hash table iterator
 773  *
 774  * Note that you must call rhashtable_walk_stop when you are finished
 775  * with the walk.
 776  *
 777  * Returns the next object or NULL when the end of the table is reached.
 778  *
 779  * Returns -EAGAIN if resize event occured.  Note that the iterator
 780  * will rewind back to the beginning and you may continue to use it.
 781  */
 782 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 783 {
 784         struct bucket_table *tbl = iter->walker.tbl;
 785         struct rhlist_head *list = iter->list;
 786         struct rhashtable *ht = iter->ht;
 787         struct rhash_head *p = iter->p;
 788         bool rhlist = ht->rhlist;
 789
 790         if (p) {
 791                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 792                         p = rcu_dereference(p->next);
 793                         list = container_of(p, struct rhlist_head, rhead);
 794                 }
 795                 goto next;
 796         }
 797
 798         for (; iter->slot < tbl->size; iter->slot++) {
 799                 int skip = iter->skip;
 800
 801                 rht_for_each_rcu(p, tbl, iter->slot) {
 802                         if (rhlist) {
 803                                 list = container_of(p, struct rhlist_head,
 804                                                     rhead);
 805                                 do {
 806                                         if (!skip)
 807                                                 goto next;
 808                                         skip--;
 809                                         list = rcu_dereference(list->next);
 810                                 } while (list);
 811
 812                                 continue;
 813                         }
 814                         if (!skip)
 815                                 break;
 816                         skip--;
 817                 }
 818
 819 next:
 820                 if (!rht_is_a_nulls(p)) {
 821                         iter->skip++;
 822                         iter->p = p;
 823                         iter->list = list;
 824                         return rht_obj(ht, rhlist ? &list->rhead : p);
 825                 }
 826
 827                 iter->skip = 0;
 828         }
 829
 830         iter->p = NULL;
 831
 832         /* Ensure we see any new tables. */
 833         smp_rmb();
 834
 835         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 836         if (iter->walker.tbl) {
 837                 iter->slot = 0;
 838                 iter->skip = 0;
 839                 return ERR_PTR(-EAGAIN);
 840         }
 841
 842         return NULL;
 843 }
 844 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 845
 846 /**
 847  * rhashtable_walk_stop - Finish a hash table walk
 848  * @iter:       Hash table iterator
 849  *
 850  * Finish a hash table walk.
 851  */
 852 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 853         __releases(RCU)
 854 {
 855         struct rhashtable *ht;
 856         struct bucket_table *tbl = iter->walker.tbl;
 857
 858         if (!tbl)
 859                 goto out;
 860
 861         ht = iter->ht;
 862
 863         spin_lock(&ht->lock);
 864         if (tbl->rehash < tbl->size)
 865                 list_add(&iter->walker.list, &tbl->walkers);
 866         else
 867                 iter->walker.tbl = NULL;
 868         spin_unlock(&ht->lock);
 869
 870         iter->p = NULL;
 871
 872 out:
 873         rcu_read_unlock();
 874 }
 875 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 876
 877 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 878 {
 879         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 880                    (unsigned long)params->min_size);
 881 }
 882
 883 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 884 {
 885         return jhash2(key, length, seed);
 886 }
 887
 888 /**
 889  * rhashtable_init - initialize a new hash table
 890  * @ht:         hash table to be initialized
 891  * @params:     configuration parameters
 892  *
 893  * Initializes a new hash table based on the provided configuration
 894  * parameters. A table can be configured either with a variable or
 895  * fixed length key:
 896  *
 897  * Configuration Example 1: Fixed length keys
 898  * struct test_obj {
 899  *      int                     key;
 900  *      void *                  my_member;
 901  *      struct rhash_head       node;
 902  * };
 903  *
 904  * struct rhashtable_params params = {
 905  *      .head_offset = offsetof(struct test_obj, node),
 906  *      .key_offset = offsetof(struct test_obj, key),
 907  *      .key_len = sizeof(int),
 908  *      .hashfn = jhash,
 909  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 910  * };
 911  *
 912  * Configuration Example 2: Variable length keys
 913  * struct test_obj {
 914  *      [...]
 915  *      struct rhash_head       node;
 916  * };
 917  *
 918  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 919  * {
 920  *      struct test_obj *obj = data;
 921  *
 922  *      return [... hash ...];
 923  * }
 924  *
 925  * struct rhashtable_params params = {
 926  *      .head_offset = offsetof(struct test_obj, node),
 927  *      .hashfn = jhash,
 928  *      .obj_hashfn = my_hash_fn,
 929  * };
 930  */
 931 int rhashtable_init(struct rhashtable *ht,
 932                     const struct rhashtable_params *params)
 933 {
 934         struct bucket_table *tbl;
 935         size_t size;
 936
 937         size = HASH_DEFAULT_SIZE;
 938
 939         if ((!params->key_len && !params->obj_hashfn) ||
 940             (params->obj_hashfn && !params->obj_cmpfn))
 941                 return -EINVAL;
 942
 943         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 944                 return -EINVAL;
 945
 946         memset(ht, 0, sizeof(*ht));
 947         mutex_init(&ht->mutex);
 948         spin_lock_init(&ht->lock);
 949         memcpy(&ht->p, params, sizeof(*params));
 950
 951         if (params->min_size)
 952                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 953
 954         /* Cap total entries at 2^31 to avoid nelems overflow. */
 955         ht->max_elems = 1u << 31;
 956
 957         if (params->max_size) {
 958                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 959                 if (ht->p.max_size < ht->max_elems / 2)
 960                         ht->max_elems = ht->p.max_size * 2;
 961         }
 962
 963         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
 964
 965         if (params->nelem_hint)
 966                 size = rounded_hashtable_size(&ht->p);
 967
 968         if (params->locks_mul)
 969                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 970         else
 971                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 972
 973         ht->key_len = ht->p.key_len;
 974         if (!params->hashfn) {
 975                 ht->p.hashfn = jhash;
 976
 977                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 978                         ht->key_len /= sizeof(u32);
 979                         ht->p.hashfn = rhashtable_jhash2;
 980                 }
 981         }
 982
 983         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 984         if (tbl == NULL)
 985                 return -ENOMEM;
 986
 987         atomic_set(&ht->nelems, 0);
 988
 989         RCU_INIT_POINTER(ht->tbl, tbl);
 990
 991         INIT_WORK(&ht->run_work, rht_deferred_worker);
 992
 993         return 0;
 994 }
 995 EXPORT_SYMBOL_GPL(rhashtable_init);
 996
 997 /**
 998  * rhltable_init - initialize a new hash list table
 999  * @hlt:        hash list table to be initialized
1000  * @params:     configuration parameters
1001  *
1002  * Initializes a new hash list table.
1003  *
1004  * See documentation for rhashtable_init.
1005  */
1006 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1007 {
1008         int err;
1009
1010         /* No rhlist NULLs marking for now. */
1011         if (params->nulls_base)
1012                 return -EINVAL;
1013
1014         err = rhashtable_init(&hlt->ht, params);
1015         hlt->ht.rhlist = true;
1016         return err;
1017 }
1018 EXPORT_SYMBOL_GPL(rhltable_init);
1019
1020 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1021                                 void (*free_fn)(void *ptr, void *arg),
1022                                 void *arg)
1023 {
1024         struct rhlist_head *list;
1025
1026         if (!ht->rhlist) {
1027                 free_fn(rht_obj(ht, obj), arg);
1028                 return;
1029         }
1030
1031         list = container_of(obj, struct rhlist_head, rhead);
1032         do {
1033                 obj = &list->rhead;
1034                 list = rht_dereference(list->next, ht);
1035                 free_fn(rht_obj(ht, obj), arg);
1036         } while (list);
1037 }
1038
1039 /**
1040  * rhashtable_free_and_destroy - free elements and destroy hash table
1041  * @ht:         the hash table to destroy
1042  * @free_fn:    callback to release resources of element
1043  * @arg:        pointer passed to free_fn
1044  *
1045  * Stops an eventual async resize. If defined, invokes free_fn for each
1046  * element to releasal resources. Please note that RCU protected
1047  * readers may still be accessing the elements. Releasing of resources
1048  * must occur in a compatible manner. Then frees the bucket array.
1049  *
1050  * This function will eventually sleep to wait for an async resize
1051  * to complete. The caller is responsible that no further write operations
1052  * occurs in parallel.
1053  */
1054 void rhashtable_free_and_destroy(struct rhashtable *ht,
1055                                  void (*free_fn)(void *ptr, void *arg),
1056                                  void *arg)
1057 {
1058         struct bucket_table *tbl;
1059         unsigned int i;
1060
1061         cancel_work_sync(&ht->run_work);
1062
1063         mutex_lock(&ht->mutex);
1064         tbl = rht_dereference(ht->tbl, ht);
1065         if (free_fn) {
1066                 for (i = 0; i < tbl->size; i++) {
1067                         struct rhash_head *pos, *next;
1068
1069                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1070                              next = !rht_is_a_nulls(pos) ?
1071                                         rht_dereference(pos->next, ht) : NULL;
1072                              !rht_is_a_nulls(pos);
1073                              pos = next,
1074                              next = !rht_is_a_nulls(pos) ?
1075                                         rht_dereference(pos->next, ht) : NULL)
1076                                 rhashtable_free_one(ht, pos, free_fn, arg);
1077                 }
1078         }
1079
1080         bucket_table_free(tbl);
1081         mutex_unlock(&ht->mutex);
1082 }
1083 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1084
1085 void rhashtable_destroy(struct rhashtable *ht)
1086 {
1087         return rhashtable_free_and_destroy(ht, NULL, NULL);
1088 }
1089 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1090
1091 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1092                                             unsigned int hash)
1093 {
1094         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1095         static struct rhash_head __rcu *rhnull =
1096                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1097         unsigned int index = hash & ((1 << tbl->nest) - 1);
1098         unsigned int size = tbl->size >> tbl->nest;
1099         unsigned int subhash = hash;
1100         union nested_table *ntbl;
1101
1102         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1103         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1104         subhash >>= tbl->nest;
1105
1106         while (ntbl && size > (1 << shift)) {
1107                 index = subhash & ((1 << shift) - 1);
1108                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1109                                                   tbl, hash);
1110                 size >>= shift;
1111                 subhash >>= shift;
1112         }
1113
1114         if (!ntbl)
1115                 return &rhnull;
1116
1117         return &ntbl[subhash].bucket;
1118
1119 }
1120 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1121
1122 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1123                                                    struct bucket_table *tbl,
1124                                                    unsigned int hash)
1125 {
1126         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1127         unsigned int index = hash & ((1 << tbl->nest) - 1);
1128         unsigned int size = tbl->size >> tbl->nest;
1129         union nested_table *ntbl;
1130         unsigned int shifted;
1131         unsigned int nhash;
1132
1133         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1134         hash >>= tbl->nest;
1135         nhash = index;
1136         shifted = tbl->nest;
1137         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1138                                   size <= (1 << shift) ? shifted : 0, nhash);
1139
1140         while (ntbl && size > (1 << shift)) {
1141                 index = hash & ((1 << shift) - 1);
1142                 size >>= shift;
1143                 hash >>= shift;
1144                 nhash |= index << shifted;
1145                 shifted += shift;
1146                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1147                                           size <= (1 << shift) ? shifted : 0,
1148                                           nhash);
1149         }
1150
1151         if (!ntbl)
1152                 return NULL;
1153
1154         return &ntbl[hash].bucket;
1155
1156 }
1157 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);