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)
 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                         pprev = &head->next;
 543                         continue;
 544                 }
 545
 546                 if (!ht->rhlist)
 547                         return rht_obj(ht, head);
 548
 549                 list = container_of(obj, struct rhlist_head, rhead);
 550                 plist = container_of(head, struct rhlist_head, rhead);
 551
 552                 RCU_INIT_POINTER(list->next, plist);
 553                 head = rht_dereference_bucket(head->next, tbl, hash);
 554                 RCU_INIT_POINTER(list->rhead.next, head);
 555                 rcu_assign_pointer(*pprev, obj);
 556
 557                 return NULL;
 558         }
 559
 560         if (elasticity <= 0)
 561                 return ERR_PTR(-EAGAIN);
 562
 563         return ERR_PTR(-ENOENT);
 564 }
 565
 566 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 567                                                   struct bucket_table *tbl,
 568                                                   unsigned int hash,
 569                                                   struct rhash_head *obj,
 570                                                   void *data)
 571 {
 572         struct rhash_head __rcu **pprev;
 573         struct bucket_table *new_tbl;
 574         struct rhash_head *head;
 575
 576         if (!IS_ERR_OR_NULL(data))
 577                 return ERR_PTR(-EEXIST);
 578
 579         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 580                 return ERR_CAST(data);
 581
 582         new_tbl = rcu_dereference(tbl->future_tbl);
 583         if (new_tbl)
 584                 return new_tbl;
 585
 586         if (PTR_ERR(data) != -ENOENT)
 587                 return ERR_CAST(data);
 588
 589         if (unlikely(rht_grow_above_max(ht, tbl)))
 590                 return ERR_PTR(-E2BIG);
 591
 592         if (unlikely(rht_grow_above_100(ht, tbl)))
 593                 return ERR_PTR(-EAGAIN);
 594
 595         pprev = rht_bucket_insert(ht, tbl, hash);
 596         if (!pprev)
 597                 return ERR_PTR(-ENOMEM);
 598
 599         head = rht_dereference_bucket(*pprev, tbl, hash);
 600
 601         RCU_INIT_POINTER(obj->next, head);
 602         if (ht->rhlist) {
 603                 struct rhlist_head *list;
 604
 605                 list = container_of(obj, struct rhlist_head, rhead);
 606                 RCU_INIT_POINTER(list->next, NULL);
 607         }
 608
 609         rcu_assign_pointer(*pprev, obj);
 610
 611         atomic_inc(&ht->nelems);
 612         if (rht_grow_above_75(ht, tbl))
 613                 schedule_work(&ht->run_work);
 614
 615         return NULL;
 616 }
 617
 618 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 619                                    struct rhash_head *obj)
 620 {
 621         struct bucket_table *new_tbl;
 622         struct bucket_table *tbl;
 623         unsigned int hash;
 624         spinlock_t *lock;
 625         void *data;
 626
 627         tbl = rcu_dereference(ht->tbl);
 628
 629         /* All insertions must grab the oldest table containing
 630          * the hashed bucket that is yet to be rehashed.
 631          */
 632         for (;;) {
 633                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 634                 lock = rht_bucket_lock(tbl, hash);
 635                 spin_lock_bh(lock);
 636
 637                 if (tbl->rehash <= hash)
 638                         break;
 639
 640                 spin_unlock_bh(lock);
 641                 tbl = rcu_dereference(tbl->future_tbl);
 642         }
 643
 644         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 645         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 646         if (PTR_ERR(new_tbl) != -EEXIST)
 647                 data = ERR_CAST(new_tbl);
 648
 649         while (!IS_ERR_OR_NULL(new_tbl)) {
 650                 tbl = new_tbl;
 651                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 652                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 653                                  SINGLE_DEPTH_NESTING);
 654
 655                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 656                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 657                 if (PTR_ERR(new_tbl) != -EEXIST)
 658                         data = ERR_CAST(new_tbl);
 659
 660                 spin_unlock(rht_bucket_lock(tbl, hash));
 661         }
 662
 663         spin_unlock_bh(lock);
 664
 665         if (PTR_ERR(data) == -EAGAIN)
 666                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 667                                -EAGAIN);
 668
 669         return data;
 670 }
 671
 672 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 673                              struct rhash_head *obj)
 674 {
 675         void *data;
 676
 677         do {
 678                 rcu_read_lock();
 679                 data = rhashtable_try_insert(ht, key, obj);
 680                 rcu_read_unlock();
 681         } while (PTR_ERR(data) == -EAGAIN);
 682
 683         return data;
 684 }
 685 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 686
 687 /**
 688  * rhashtable_walk_enter - Initialise an iterator
 689  * @ht:         Table to walk over
 690  * @iter:       Hash table Iterator
 691  *
 692  * This function prepares a hash table walk.
 693  *
 694  * Note that if you restart a walk after rhashtable_walk_stop you
 695  * may see the same object twice.  Also, you may miss objects if
 696  * there are removals in between rhashtable_walk_stop and the next
 697  * call to rhashtable_walk_start.
 698  *
 699  * For a completely stable walk you should construct your own data
 700  * structure outside the hash table.
 701  *
 702  * This function may sleep so you must not call it from interrupt
 703  * context or with spin locks held.
 704  *
 705  * You must call rhashtable_walk_exit after this function returns.
 706  */
 707 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 708 {
 709         iter->ht = ht;
 710         iter->p = NULL;
 711         iter->slot = 0;
 712         iter->skip = 0;
 713
 714         spin_lock(&ht->lock);
 715         iter->walker.tbl =
 716                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 717         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 718         spin_unlock(&ht->lock);
 719 }
 720 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 721
 722 /**
 723  * rhashtable_walk_exit - Free an iterator
 724  * @iter:       Hash table Iterator
 725  *
 726  * This function frees resources allocated by rhashtable_walk_init.
 727  */
 728 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 729 {
 730         spin_lock(&iter->ht->lock);
 731         if (iter->walker.tbl)
 732                 list_del(&iter->walker.list);
 733         spin_unlock(&iter->ht->lock);
 734 }
 735 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 736
 737 /**
 738  * rhashtable_walk_start - Start a hash table walk
 739  * @iter:       Hash table iterator
 740  *
 741  * Start a hash table walk at the current iterator position.  Note that we take
 742  * the RCU lock in all cases including when we return an error.  So you must
 743  * always call rhashtable_walk_stop to clean up.
 744  *
 745  * Returns zero if successful.
 746  *
 747  * Returns -EAGAIN if resize event occured.  Note that the iterator
 748  * will rewind back to the beginning and you may use it immediately
 749  * by calling rhashtable_walk_next.
 750  */
 751 int rhashtable_walk_start(struct rhashtable_iter *iter)
 752         __acquires(RCU)
 753 {
 754         struct rhashtable *ht = iter->ht;
 755
 756         rcu_read_lock();
 757
 758         spin_lock(&ht->lock);
 759         if (iter->walker.tbl)
 760                 list_del(&iter->walker.list);
 761         spin_unlock(&ht->lock);
 762
 763         if (!iter->walker.tbl) {
 764                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 765                 return -EAGAIN;
 766         }
 767
 768         return 0;
 769 }
 770 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 771
 772 /**
 773  * rhashtable_walk_next - Return the next object and advance the iterator
 774  * @iter:       Hash table iterator
 775  *
 776  * Note that you must call rhashtable_walk_stop when you are finished
 777  * with the walk.
 778  *
 779  * Returns the next object or NULL when the end of the table is reached.
 780  *
 781  * Returns -EAGAIN if resize event occured.  Note that the iterator
 782  * will rewind back to the beginning and you may continue to use it.
 783  */
 784 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 785 {
 786         struct bucket_table *tbl = iter->walker.tbl;
 787         struct rhlist_head *list = iter->list;
 788         struct rhashtable *ht = iter->ht;
 789         struct rhash_head *p = iter->p;
 790         bool rhlist = ht->rhlist;
 791
 792         if (p) {
 793                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 794                         p = rcu_dereference(p->next);
 795                         list = container_of(p, struct rhlist_head, rhead);
 796                 }
 797                 goto next;
 798         }
 799
 800         for (; iter->slot < tbl->size; iter->slot++) {
 801                 int skip = iter->skip;
 802
 803                 rht_for_each_rcu(p, tbl, iter->slot) {
 804                         if (rhlist) {
 805                                 list = container_of(p, struct rhlist_head,
 806                                                     rhead);
 807                                 do {
 808                                         if (!skip)
 809                                                 goto next;
 810                                         skip--;
 811                                         list = rcu_dereference(list->next);
 812                                 } while (list);
 813
 814                                 continue;
 815                         }
 816                         if (!skip)
 817                                 break;
 818                         skip--;
 819                 }
 820
 821 next:
 822                 if (!rht_is_a_nulls(p)) {
 823                         iter->skip++;
 824                         iter->p = p;
 825                         iter->list = list;
 826                         return rht_obj(ht, rhlist ? &list->rhead : p);
 827                 }
 828
 829                 iter->skip = 0;
 830         }
 831
 832         iter->p = NULL;
 833
 834         /* Ensure we see any new tables. */
 835         smp_rmb();
 836
 837         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 838         if (iter->walker.tbl) {
 839                 iter->slot = 0;
 840                 iter->skip = 0;
 841                 return ERR_PTR(-EAGAIN);
 842         }
 843
 844         return NULL;
 845 }
 846 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 847
 848 /**
 849  * rhashtable_walk_stop - Finish a hash table walk
 850  * @iter:       Hash table iterator
 851  *
 852  * Finish a hash table walk.  Does not reset the iterator to the start of the
 853  * hash table.
 854  */
 855 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 856         __releases(RCU)
 857 {
 858         struct rhashtable *ht;
 859         struct bucket_table *tbl = iter->walker.tbl;
 860
 861         if (!tbl)
 862                 goto out;
 863
 864         ht = iter->ht;
 865
 866         spin_lock(&ht->lock);
 867         if (tbl->rehash < tbl->size)
 868                 list_add(&iter->walker.list, &tbl->walkers);
 869         else
 870                 iter->walker.tbl = NULL;
 871         spin_unlock(&ht->lock);
 872
 873         iter->p = NULL;
 874
 875 out:
 876         rcu_read_unlock();
 877 }
 878 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 879
 880 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 881 {
 882         size_t retsize;
 883
 884         if (params->nelem_hint)
 885                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 886                               (unsigned long)params->min_size);
 887         else
 888                 retsize = max(HASH_DEFAULT_SIZE,
 889                               (unsigned long)params->min_size);
 890
 891         return retsize;
 892 }
 893
 894 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 895 {
 896         return jhash2(key, length, seed);
 897 }
 898
 899 /**
 900  * rhashtable_init - initialize a new hash table
 901  * @ht:         hash table to be initialized
 902  * @params:     configuration parameters
 903  *
 904  * Initializes a new hash table based on the provided configuration
 905  * parameters. A table can be configured either with a variable or
 906  * fixed length key:
 907  *
 908  * Configuration Example 1: Fixed length keys
 909  * struct test_obj {
 910  *      int                     key;
 911  *      void *                  my_member;
 912  *      struct rhash_head       node;
 913  * };
 914  *
 915  * struct rhashtable_params params = {
 916  *      .head_offset = offsetof(struct test_obj, node),
 917  *      .key_offset = offsetof(struct test_obj, key),
 918  *      .key_len = sizeof(int),
 919  *      .hashfn = jhash,
 920  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 921  * };
 922  *
 923  * Configuration Example 2: Variable length keys
 924  * struct test_obj {
 925  *      [...]
 926  *      struct rhash_head       node;
 927  * };
 928  *
 929  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 930  * {
 931  *      struct test_obj *obj = data;
 932  *
 933  *      return [... hash ...];
 934  * }
 935  *
 936  * struct rhashtable_params params = {
 937  *      .head_offset = offsetof(struct test_obj, node),
 938  *      .hashfn = jhash,
 939  *      .obj_hashfn = my_hash_fn,
 940  * };
 941  */
 942 int rhashtable_init(struct rhashtable *ht,
 943                     const struct rhashtable_params *params)
 944 {
 945         struct bucket_table *tbl;
 946         size_t size;
 947
 948         if ((!params->key_len && !params->obj_hashfn) ||
 949             (params->obj_hashfn && !params->obj_cmpfn))
 950                 return -EINVAL;
 951
 952         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 953                 return -EINVAL;
 954
 955         memset(ht, 0, sizeof(*ht));
 956         mutex_init(&ht->mutex);
 957         spin_lock_init(&ht->lock);
 958         memcpy(&ht->p, params, sizeof(*params));
 959
 960         if (params->min_size)
 961                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 962
 963         /* Cap total entries at 2^31 to avoid nelems overflow. */
 964         ht->max_elems = 1u << 31;
 965
 966         if (params->max_size) {
 967                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 968                 if (ht->p.max_size < ht->max_elems / 2)
 969                         ht->max_elems = ht->p.max_size * 2;
 970         }
 971
 972         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
 973
 974         size = rounded_hashtable_size(&ht->p);
 975
 976         if (params->locks_mul)
 977                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 978         else
 979                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 980
 981         ht->key_len = ht->p.key_len;
 982         if (!params->hashfn) {
 983                 ht->p.hashfn = jhash;
 984
 985                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 986                         ht->key_len /= sizeof(u32);
 987                         ht->p.hashfn = rhashtable_jhash2;
 988                 }
 989         }
 990
 991         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 992         if (tbl == NULL)
 993                 return -ENOMEM;
 994
 995         atomic_set(&ht->nelems, 0);
 996
 997         RCU_INIT_POINTER(ht->tbl, tbl);
 998
 999         INIT_WORK(&ht->run_work, rht_deferred_worker);
1000
1001         return 0;
1002 }
1003 EXPORT_SYMBOL_GPL(rhashtable_init);
1004
1005 /**
1006  * rhltable_init - initialize a new hash list table
1007  * @hlt:        hash list table to be initialized
1008  * @params:     configuration parameters
1009  *
1010  * Initializes a new hash list table.
1011  *
1012  * See documentation for rhashtable_init.
1013  */
1014 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1015 {
1016         int err;
1017
1018         /* No rhlist NULLs marking for now. */
1019         if (params->nulls_base)
1020                 return -EINVAL;
1021
1022         err = rhashtable_init(&hlt->ht, params);
1023         hlt->ht.rhlist = true;
1024         return err;
1025 }
1026 EXPORT_SYMBOL_GPL(rhltable_init);
1027
1028 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1029                                 void (*free_fn)(void *ptr, void *arg),
1030                                 void *arg)
1031 {
1032         struct rhlist_head *list;
1033
1034         if (!ht->rhlist) {
1035                 free_fn(rht_obj(ht, obj), arg);
1036                 return;
1037         }
1038
1039         list = container_of(obj, struct rhlist_head, rhead);
1040         do {
1041                 obj = &list->rhead;
1042                 list = rht_dereference(list->next, ht);
1043                 free_fn(rht_obj(ht, obj), arg);
1044         } while (list);
1045 }
1046
1047 /**
1048  * rhashtable_free_and_destroy - free elements and destroy hash table
1049  * @ht:         the hash table to destroy
1050  * @free_fn:    callback to release resources of element
1051  * @arg:        pointer passed to free_fn
1052  *
1053  * Stops an eventual async resize. If defined, invokes free_fn for each
1054  * element to releasal resources. Please note that RCU protected
1055  * readers may still be accessing the elements. Releasing of resources
1056  * must occur in a compatible manner. Then frees the bucket array.
1057  *
1058  * This function will eventually sleep to wait for an async resize
1059  * to complete. The caller is responsible that no further write operations
1060  * occurs in parallel.
1061  */
1062 void rhashtable_free_and_destroy(struct rhashtable *ht,
1063                                  void (*free_fn)(void *ptr, void *arg),
1064                                  void *arg)
1065 {
1066         struct bucket_table *tbl;
1067         unsigned int i;
1068
1069         cancel_work_sync(&ht->run_work);
1070
1071         mutex_lock(&ht->mutex);
1072         tbl = rht_dereference(ht->tbl, ht);
1073         if (free_fn) {
1074                 for (i = 0; i < tbl->size; i++) {
1075                         struct rhash_head *pos, *next;
1076
1077                         cond_resched();
1078                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1079                              next = !rht_is_a_nulls(pos) ?
1080                                         rht_dereference(pos->next, ht) : NULL;
1081                              !rht_is_a_nulls(pos);
1082                              pos = next,
1083                              next = !rht_is_a_nulls(pos) ?
1084                                         rht_dereference(pos->next, ht) : NULL)
1085                                 rhashtable_free_one(ht, pos, free_fn, arg);
1086                 }
1087         }
1088
1089         bucket_table_free(tbl);
1090         mutex_unlock(&ht->mutex);
1091 }
1092 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1093
1094 void rhashtable_destroy(struct rhashtable *ht)
1095 {
1096         return rhashtable_free_and_destroy(ht, NULL, NULL);
1097 }
1098 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1099
1100 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1101                                             unsigned int hash)
1102 {
1103         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1104         static struct rhash_head __rcu *rhnull =
1105                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1106         unsigned int index = hash & ((1 << tbl->nest) - 1);
1107         unsigned int size = tbl->size >> tbl->nest;
1108         unsigned int subhash = hash;
1109         union nested_table *ntbl;
1110
1111         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1112         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1113         subhash >>= tbl->nest;
1114
1115         while (ntbl && size > (1 << shift)) {
1116                 index = subhash & ((1 << shift) - 1);
1117                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1118                                                   tbl, hash);
1119                 size >>= shift;
1120                 subhash >>= shift;
1121         }
1122
1123         if (!ntbl)
1124                 return &rhnull;
1125
1126         return &ntbl[subhash].bucket;
1127
1128 }
1129 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1130
1131 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1132                                                    struct bucket_table *tbl,
1133                                                    unsigned int hash)
1134 {
1135         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1136         unsigned int index = hash & ((1 << tbl->nest) - 1);
1137         unsigned int size = tbl->size >> tbl->nest;
1138         union nested_table *ntbl;
1139         unsigned int shifted;
1140         unsigned int nhash;
1141
1142         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1143         hash >>= tbl->nest;
1144         nhash = index;
1145         shifted = tbl->nest;
1146         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1147                                   size <= (1 << shift) ? shifted : 0, nhash);
1148
1149         while (ntbl && size > (1 << shift)) {
1150                 index = hash & ((1 << shift) - 1);
1151                 size >>= shift;
1152                 hash >>= shift;
1153                 nhash |= index << shifted;
1154                 shifted += shift;
1155                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1156                                           size <= (1 << shift) ? shifted : 0,
1157                                           nhash);
1158         }
1159
1160         if (!ntbl)
1161                 return NULL;
1162
1163         return &ntbl[hash].bucket;
1164
1165 }
1166 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);