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 (gfp != GFP_KERNEL)
 215                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 216         else
 217                 tbl = kvzalloc(size, gfp);
 218
 219         size = nbuckets;
 220
 221         if (tbl == NULL && gfp != GFP_KERNEL) {
 222                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 223                 nbuckets = 0;
 224         }
 225         if (tbl == NULL)
 226                 return NULL;
 227
 228         tbl->size = size;
 229
 230         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 231                 bucket_table_free(tbl);
 232                 return NULL;
 233         }
 234
 235         INIT_LIST_HEAD(&tbl->walkers);
 236
 237         tbl->hash_rnd = get_random_u32();
 238
 239         for (i = 0; i < nbuckets; i++)
 240                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 241
 242         return tbl;
 243 }
 244
 245 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 246                                                   struct bucket_table *tbl)
 247 {
 248         struct bucket_table *new_tbl;
 249
 250         do {
 251                 new_tbl = tbl;
 252                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 253         } while (tbl);
 254
 255         return new_tbl;
 256 }
 257
 258 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 259 {
 260         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 261         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 262                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 263         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 264         int err = -EAGAIN;
 265         struct rhash_head *head, *next, *entry;
 266         spinlock_t *new_bucket_lock;
 267         unsigned int new_hash;
 268
 269         if (new_tbl->nest)
 270                 goto out;
 271
 272         err = -ENOENT;
 273
 274         rht_for_each(entry, old_tbl, old_hash) {
 275                 err = 0;
 276                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 277
 278                 if (rht_is_a_nulls(next))
 279                         break;
 280
 281                 pprev = &entry->next;
 282         }
 283
 284         if (err)
 285                 goto out;
 286
 287         new_hash = head_hashfn(ht, new_tbl, entry);
 288
 289         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 290
 291         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 292         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 293                                       new_tbl, new_hash);
 294
 295         RCU_INIT_POINTER(entry->next, head);
 296
 297         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 298         spin_unlock(new_bucket_lock);
 299
 300         rcu_assign_pointer(*pprev, next);
 301
 302 out:
 303         return err;
 304 }
 305
 306 static int rhashtable_rehash_chain(struct rhashtable *ht,
 307                                     unsigned int old_hash)
 308 {
 309         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 310         spinlock_t *old_bucket_lock;
 311         int err;
 312
 313         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 314
 315         spin_lock_bh(old_bucket_lock);
 316         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 317                 ;
 318
 319         if (err == -ENOENT) {
 320                 old_tbl->rehash++;
 321                 err = 0;
 322         }
 323         spin_unlock_bh(old_bucket_lock);
 324
 325         return err;
 326 }
 327
 328 static int rhashtable_rehash_attach(struct rhashtable *ht,
 329                                     struct bucket_table *old_tbl,
 330                                     struct bucket_table *new_tbl)
 331 {
 332         /* Protect future_tbl using the first bucket lock. */
 333         spin_lock_bh(old_tbl->locks);
 334
 335         /* Did somebody beat us to it? */
 336         if (rcu_access_pointer(old_tbl->future_tbl)) {
 337                 spin_unlock_bh(old_tbl->locks);
 338                 return -EEXIST;
 339         }
 340
 341         /* Make insertions go into the new, empty table right away. Deletions
 342          * and lookups will be attempted in both tables until we synchronize.
 343          */
 344         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 345
 346         spin_unlock_bh(old_tbl->locks);
 347
 348         return 0;
 349 }
 350
 351 static int rhashtable_rehash_table(struct rhashtable *ht)
 352 {
 353         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 354         struct bucket_table *new_tbl;
 355         struct rhashtable_walker *walker;
 356         unsigned int old_hash;
 357         int err;
 358
 359         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 360         if (!new_tbl)
 361                 return 0;
 362
 363         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 364                 err = rhashtable_rehash_chain(ht, old_hash);
 365                 if (err)
 366                         return err;
 367                 cond_resched();
 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 || err == -EEXIST) {
 463                 int nerr;
 464
 465                 nerr = rhashtable_rehash_table(ht);
 466                 err = err ?: nerr;
 467         }
 468
 469         mutex_unlock(&ht->mutex);
 470
 471         if (err)
 472                 schedule_work(&ht->run_work);
 473 }
 474
 475 static int rhashtable_insert_rehash(struct rhashtable *ht,
 476                                     struct bucket_table *tbl)
 477 {
 478         struct bucket_table *old_tbl;
 479         struct bucket_table *new_tbl;
 480         unsigned int size;
 481         int err;
 482
 483         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 484
 485         size = tbl->size;
 486
 487         err = -EBUSY;
 488
 489         if (rht_grow_above_75(ht, tbl))
 490                 size *= 2;
 491         /* Do not schedule more than one rehash */
 492         else if (old_tbl != tbl)
 493                 goto fail;
 494
 495         err = -ENOMEM;
 496
 497         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 498         if (new_tbl == NULL)
 499                 goto fail;
 500
 501         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 502         if (err) {
 503                 bucket_table_free(new_tbl);
 504                 if (err == -EEXIST)
 505                         err = 0;
 506         } else
 507                 schedule_work(&ht->run_work);
 508
 509         return err;
 510
 511 fail:
 512         /* Do not fail the insert if someone else did a rehash. */
 513         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 514                 return 0;
 515
 516         /* Schedule async rehash to retry allocation in process context. */
 517         if (err == -ENOMEM)
 518                 schedule_work(&ht->run_work);
 519
 520         return err;
 521 }
 522
 523 static void *rhashtable_lookup_one(struct rhashtable *ht,
 524                                    struct bucket_table *tbl, unsigned int hash,
 525                                    const void *key, struct rhash_head *obj)
 526 {
 527         struct rhashtable_compare_arg arg = {
 528                 .ht = ht,
 529                 .key = key,
 530         };
 531         struct rhash_head __rcu **pprev;
 532         struct rhash_head *head;
 533         int elasticity;
 534
 535         elasticity = RHT_ELASTICITY;
 536         pprev = rht_bucket_var(tbl, hash);
 537         rht_for_each_continue(head, *pprev, tbl, hash) {
 538                 struct rhlist_head *list;
 539                 struct rhlist_head *plist;
 540
 541                 elasticity--;
 542                 if (!key ||
 543                     (ht->p.obj_cmpfn ?
 544                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 545                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 546                         pprev = &head->next;
 547                         continue;
 548                 }
 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 at the current iterator position.  Note that we take
 746  * the RCU lock in all cases including when we return an error.  So you must
 747  * always call 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.  Does not reset the iterator to the start of the
 857  * hash table.
 858  */
 859 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 860         __releases(RCU)
 861 {
 862         struct rhashtable *ht;
 863         struct bucket_table *tbl = iter->walker.tbl;
 864
 865         if (!tbl)
 866                 goto out;
 867
 868         ht = iter->ht;
 869
 870         spin_lock(&ht->lock);
 871         if (tbl->rehash < tbl->size)
 872                 list_add(&iter->walker.list, &tbl->walkers);
 873         else
 874                 iter->walker.tbl = NULL;
 875         spin_unlock(&ht->lock);
 876
 877         iter->p = NULL;
 878
 879 out:
 880         rcu_read_unlock();
 881 }
 882 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 883
 884 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 885 {
 886         size_t retsize;
 887
 888         if (params->nelem_hint)
 889                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 890                               (unsigned long)params->min_size);
 891         else
 892                 retsize = max(HASH_DEFAULT_SIZE,
 893                               (unsigned long)params->min_size);
 894
 895         return retsize;
 896 }
 897
 898 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 899 {
 900         return jhash2(key, length, seed);
 901 }
 902
 903 /**
 904  * rhashtable_init - initialize a new hash table
 905  * @ht:         hash table to be initialized
 906  * @params:     configuration parameters
 907  *
 908  * Initializes a new hash table based on the provided configuration
 909  * parameters. A table can be configured either with a variable or
 910  * fixed length key:
 911  *
 912  * Configuration Example 1: Fixed length keys
 913  * struct test_obj {
 914  *      int                     key;
 915  *      void *                  my_member;
 916  *      struct rhash_head       node;
 917  * };
 918  *
 919  * struct rhashtable_params params = {
 920  *      .head_offset = offsetof(struct test_obj, node),
 921  *      .key_offset = offsetof(struct test_obj, key),
 922  *      .key_len = sizeof(int),
 923  *      .hashfn = jhash,
 924  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 925  * };
 926  *
 927  * Configuration Example 2: Variable length keys
 928  * struct test_obj {
 929  *      [...]
 930  *      struct rhash_head       node;
 931  * };
 932  *
 933  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 934  * {
 935  *      struct test_obj *obj = data;
 936  *
 937  *      return [... hash ...];
 938  * }
 939  *
 940  * struct rhashtable_params params = {
 941  *      .head_offset = offsetof(struct test_obj, node),
 942  *      .hashfn = jhash,
 943  *      .obj_hashfn = my_hash_fn,
 944  * };
 945  */
 946 int rhashtable_init(struct rhashtable *ht,
 947                     const struct rhashtable_params *params)
 948 {
 949         struct bucket_table *tbl;
 950         size_t size;
 951
 952         if ((!params->key_len && !params->obj_hashfn) ||
 953             (params->obj_hashfn && !params->obj_cmpfn))
 954                 return -EINVAL;
 955
 956         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 957                 return -EINVAL;
 958
 959         memset(ht, 0, sizeof(*ht));
 960         mutex_init(&ht->mutex);
 961         spin_lock_init(&ht->lock);
 962         memcpy(&ht->p, params, sizeof(*params));
 963
 964         if (params->min_size)
 965                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 966
 967         /* Cap total entries at 2^31 to avoid nelems overflow. */
 968         ht->max_elems = 1u << 31;
 969
 970         if (params->max_size) {
 971                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 972                 if (ht->p.max_size < ht->max_elems / 2)
 973                         ht->max_elems = ht->p.max_size * 2;
 974         }
 975
 976         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
 977
 978         size = rounded_hashtable_size(&ht->p);
 979
 980         if (params->locks_mul)
 981                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 982         else
 983                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 984
 985         ht->key_len = ht->p.key_len;
 986         if (!params->hashfn) {
 987                 ht->p.hashfn = jhash;
 988
 989                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 990                         ht->key_len /= sizeof(u32);
 991                         ht->p.hashfn = rhashtable_jhash2;
 992                 }
 993         }
 994
 995         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 996         if (tbl == NULL)
 997                 return -ENOMEM;
 998
 999         atomic_set(&ht->nelems, 0);
1000
1001         RCU_INIT_POINTER(ht->tbl, tbl);
1002
1003         INIT_WORK(&ht->run_work, rht_deferred_worker);
1004
1005         return 0;
1006 }
1007 EXPORT_SYMBOL_GPL(rhashtable_init);
1008
1009 /**
1010  * rhltable_init - initialize a new hash list table
1011  * @hlt:        hash list table to be initialized
1012  * @params:     configuration parameters
1013  *
1014  * Initializes a new hash list table.
1015  *
1016  * See documentation for rhashtable_init.
1017  */
1018 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1019 {
1020         int err;
1021
1022         /* No rhlist NULLs marking for now. */
1023         if (params->nulls_base)
1024                 return -EINVAL;
1025
1026         err = rhashtable_init(&hlt->ht, params);
1027         hlt->ht.rhlist = true;
1028         return err;
1029 }
1030 EXPORT_SYMBOL_GPL(rhltable_init);
1031
1032 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1033                                 void (*free_fn)(void *ptr, void *arg),
1034                                 void *arg)
1035 {
1036         struct rhlist_head *list;
1037
1038         if (!ht->rhlist) {
1039                 free_fn(rht_obj(ht, obj), arg);
1040                 return;
1041         }
1042
1043         list = container_of(obj, struct rhlist_head, rhead);
1044         do {
1045                 obj = &list->rhead;
1046                 list = rht_dereference(list->next, ht);
1047                 free_fn(rht_obj(ht, obj), arg);
1048         } while (list);
1049 }
1050
1051 /**
1052  * rhashtable_free_and_destroy - free elements and destroy hash table
1053  * @ht:         the hash table to destroy
1054  * @free_fn:    callback to release resources of element
1055  * @arg:        pointer passed to free_fn
1056  *
1057  * Stops an eventual async resize. If defined, invokes free_fn for each
1058  * element to releasal resources. Please note that RCU protected
1059  * readers may still be accessing the elements. Releasing of resources
1060  * must occur in a compatible manner. Then frees the bucket array.
1061  *
1062  * This function will eventually sleep to wait for an async resize
1063  * to complete. The caller is responsible that no further write operations
1064  * occurs in parallel.
1065  */
1066 void rhashtable_free_and_destroy(struct rhashtable *ht,
1067                                  void (*free_fn)(void *ptr, void *arg),
1068                                  void *arg)
1069 {
1070         struct bucket_table *tbl;
1071         unsigned int i;
1072
1073         cancel_work_sync(&ht->run_work);
1074
1075         mutex_lock(&ht->mutex);
1076         tbl = rht_dereference(ht->tbl, ht);
1077         if (free_fn) {
1078                 for (i = 0; i < tbl->size; i++) {
1079                         struct rhash_head *pos, *next;
1080
1081                         cond_resched();
1082                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1083                              next = !rht_is_a_nulls(pos) ?
1084                                         rht_dereference(pos->next, ht) : NULL;
1085                              !rht_is_a_nulls(pos);
1086                              pos = next,
1087                              next = !rht_is_a_nulls(pos) ?
1088                                         rht_dereference(pos->next, ht) : NULL)
1089                                 rhashtable_free_one(ht, pos, free_fn, arg);
1090                 }
1091         }
1092
1093         bucket_table_free(tbl);
1094         mutex_unlock(&ht->mutex);
1095 }
1096 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1097
1098 void rhashtable_destroy(struct rhashtable *ht)
1099 {
1100         return rhashtable_free_and_destroy(ht, NULL, NULL);
1101 }
1102 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1103
1104 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1105                                             unsigned int hash)
1106 {
1107         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1108         static struct rhash_head __rcu *rhnull =
1109                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1110         unsigned int index = hash & ((1 << tbl->nest) - 1);
1111         unsigned int size = tbl->size >> tbl->nest;
1112         unsigned int subhash = hash;
1113         union nested_table *ntbl;
1114
1115         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1116         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1117         subhash >>= tbl->nest;
1118
1119         while (ntbl && size > (1 << shift)) {
1120                 index = subhash & ((1 << shift) - 1);
1121                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1122                                                   tbl, hash);
1123                 size >>= shift;
1124                 subhash >>= shift;
1125         }
1126
1127         if (!ntbl)
1128                 return &rhnull;
1129
1130         return &ntbl[subhash].bucket;
1131
1132 }
1133 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1134
1135 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1136                                                    struct bucket_table *tbl,
1137                                                    unsigned int hash)
1138 {
1139         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1140         unsigned int index = hash & ((1 << tbl->nest) - 1);
1141         unsigned int size = tbl->size >> tbl->nest;
1142         union nested_table *ntbl;
1143         unsigned int shifted;
1144         unsigned int nhash;
1145
1146         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1147         hash >>= tbl->nest;
1148         nhash = index;
1149         shifted = tbl->nest;
1150         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1151                                   size <= (1 << shift) ? shifted : 0, nhash);
1152
1153         while (ntbl && size > (1 << shift)) {
1154                 index = hash & ((1 << shift) - 1);
1155                 size >>= shift;
1156                 hash >>= shift;
1157                 nhash |= index << shifted;
1158                 shifted += shift;
1159                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1160                                           size <= (1 << shift) ? shifted : 0,
1161                                           nhash);
1162         }
1163
1164         if (!ntbl)
1165                 return NULL;
1166
1167         return &ntbl[hash].bucket;
1168
1169 }
1170 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);