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 static void nested_table_free(union nested_table *ntbl, unsigned int size)
  69 {
  70         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
  71         const unsigned int len = 1 << shift;
  72         unsigned int i;
  73
  74         ntbl = rcu_dereference_raw(ntbl->table);
  75         if (!ntbl)
  76                 return;
  77
  78         if (size > len) {
  79                 size >>= shift;
  80                 for (i = 0; i < len; i++)
  81                         nested_table_free(ntbl + i, size);
  82         }
  83
  84         kfree(ntbl);
  85 }
  86
  87 static void nested_bucket_table_free(const struct bucket_table *tbl)
  88 {
  89         unsigned int size = tbl->size >> tbl->nest;
  90         unsigned int len = 1 << tbl->nest;
  91         union nested_table *ntbl;
  92         unsigned int i;
  93
  94         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
  95
  96         for (i = 0; i < len; i++)
  97                 nested_table_free(ntbl + i, size);
  98
  99         kfree(ntbl);
 100 }
 101
 102 static void bucket_table_free(const struct bucket_table *tbl)
 103 {
 104         if (tbl->nest)
 105                 nested_bucket_table_free(tbl);
 106
 107         free_bucket_spinlocks(tbl->locks);
 108         kvfree(tbl);
 109 }
 110
 111 static void bucket_table_free_rcu(struct rcu_head *head)
 112 {
 113         bucket_table_free(container_of(head, struct bucket_table, rcu));
 114 }
 115
 116 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 117                                               union nested_table __rcu **prev,
 118                                               unsigned int shifted,
 119                                               unsigned int nhash)
 120 {
 121         union nested_table *ntbl;
 122         int i;
 123
 124         ntbl = rcu_dereference(*prev);
 125         if (ntbl)
 126                 return ntbl;
 127
 128         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 129
 130         if (ntbl && shifted) {
 131                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
 132                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
 133                                             (i << shifted) | nhash);
 134         }
 135
 136         rcu_assign_pointer(*prev, ntbl);
 137
 138         return ntbl;
 139 }
 140
 141 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 142                                                       size_t nbuckets,
 143                                                       gfp_t gfp)
 144 {
 145         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 146         struct bucket_table *tbl;
 147         size_t size;
 148
 149         if (nbuckets < (1 << (shift + 1)))
 150                 return NULL;
 151
 152         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 153
 154         tbl = kzalloc(size, gfp);
 155         if (!tbl)
 156                 return NULL;
 157
 158         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 159                                 0, 0)) {
 160                 kfree(tbl);
 161                 return NULL;
 162         }
 163
 164         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 165
 166         return tbl;
 167 }
 168
 169 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 170                                                size_t nbuckets,
 171                                                gfp_t gfp)
 172 {
 173         struct bucket_table *tbl = NULL;
 174         size_t size, max_locks;
 175         int i;
 176
 177         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 178         if (gfp != GFP_KERNEL)
 179                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 180         else
 181                 tbl = kvzalloc(size, gfp);
 182
 183         size = nbuckets;
 184
 185         if (tbl == NULL && gfp != GFP_KERNEL) {
 186                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 187                 nbuckets = 0;
 188         }
 189         if (tbl == NULL)
 190                 return NULL;
 191
 192         tbl->size = size;
 193
 194         max_locks = size >> 1;
 195         if (tbl->nest)
 196                 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
 197
 198         if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
 199                                    ht->p.locks_mul, gfp) < 0) {
 200                 bucket_table_free(tbl);
 201                 return NULL;
 202         }
 203
 204         INIT_LIST_HEAD(&tbl->walkers);
 205
 206         tbl->hash_rnd = get_random_u32();
 207
 208         for (i = 0; i < nbuckets; i++)
 209                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 210
 211         return tbl;
 212 }
 213
 214 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 215                                                   struct bucket_table *tbl)
 216 {
 217         struct bucket_table *new_tbl;
 218
 219         do {
 220                 new_tbl = tbl;
 221                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 222         } while (tbl);
 223
 224         return new_tbl;
 225 }
 226
 227 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 228 {
 229         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 230         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 231                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 232         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 233         int err = -EAGAIN;
 234         struct rhash_head *head, *next, *entry;
 235         spinlock_t *new_bucket_lock;
 236         unsigned int new_hash;
 237
 238         if (new_tbl->nest)
 239                 goto out;
 240
 241         err = -ENOENT;
 242
 243         rht_for_each(entry, old_tbl, old_hash) {
 244                 err = 0;
 245                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 246
 247                 if (rht_is_a_nulls(next))
 248                         break;
 249
 250                 pprev = &entry->next;
 251         }
 252
 253         if (err)
 254                 goto out;
 255
 256         new_hash = head_hashfn(ht, new_tbl, entry);
 257
 258         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 259
 260         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 261         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 262                                       new_tbl, new_hash);
 263
 264         RCU_INIT_POINTER(entry->next, head);
 265
 266         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 267         spin_unlock(new_bucket_lock);
 268
 269         rcu_assign_pointer(*pprev, next);
 270
 271 out:
 272         return err;
 273 }
 274
 275 static int rhashtable_rehash_chain(struct rhashtable *ht,
 276                                     unsigned int old_hash)
 277 {
 278         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 279         spinlock_t *old_bucket_lock;
 280         int err;
 281
 282         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 283
 284         spin_lock_bh(old_bucket_lock);
 285         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 286                 ;
 287
 288         if (err == -ENOENT) {
 289                 old_tbl->rehash++;
 290                 err = 0;
 291         }
 292         spin_unlock_bh(old_bucket_lock);
 293
 294         return err;
 295 }
 296
 297 static int rhashtable_rehash_attach(struct rhashtable *ht,
 298                                     struct bucket_table *old_tbl,
 299                                     struct bucket_table *new_tbl)
 300 {
 301         /* Protect future_tbl using the first bucket lock. */
 302         spin_lock_bh(old_tbl->locks);
 303
 304         /* Did somebody beat us to it? */
 305         if (rcu_access_pointer(old_tbl->future_tbl)) {
 306                 spin_unlock_bh(old_tbl->locks);
 307                 return -EEXIST;
 308         }
 309
 310         /* Make insertions go into the new, empty table right away. Deletions
 311          * and lookups will be attempted in both tables until we synchronize.
 312          */
 313         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 314
 315         spin_unlock_bh(old_tbl->locks);
 316
 317         return 0;
 318 }
 319
 320 static int rhashtable_rehash_table(struct rhashtable *ht)
 321 {
 322         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 323         struct bucket_table *new_tbl;
 324         struct rhashtable_walker *walker;
 325         unsigned int old_hash;
 326         int err;
 327
 328         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 329         if (!new_tbl)
 330                 return 0;
 331
 332         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 333                 err = rhashtable_rehash_chain(ht, old_hash);
 334                 if (err)
 335                         return err;
 336                 cond_resched();
 337         }
 338
 339         /* Publish the new table pointer. */
 340         rcu_assign_pointer(ht->tbl, new_tbl);
 341
 342         spin_lock(&ht->lock);
 343         list_for_each_entry(walker, &old_tbl->walkers, list)
 344                 walker->tbl = NULL;
 345         spin_unlock(&ht->lock);
 346
 347         /* Wait for readers. All new readers will see the new
 348          * table, and thus no references to the old table will
 349          * remain.
 350          */
 351         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 352
 353         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 354 }
 355
 356 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 357                                    struct bucket_table *old_tbl,
 358                                    unsigned int size)
 359 {
 360         struct bucket_table *new_tbl;
 361         int err;
 362
 363         ASSERT_RHT_MUTEX(ht);
 364
 365         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 366         if (new_tbl == NULL)
 367                 return -ENOMEM;
 368
 369         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 370         if (err)
 371                 bucket_table_free(new_tbl);
 372
 373         return err;
 374 }
 375
 376 /**
 377  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 378  * @ht:         the hash table to shrink
 379  *
 380  * This function shrinks the hash table to fit, i.e., the smallest
 381  * size would not cause it to expand right away automatically.
 382  *
 383  * The caller must ensure that no concurrent resizing occurs by holding
 384  * ht->mutex.
 385  *
 386  * The caller must ensure that no concurrent table mutations take place.
 387  * It is however valid to have concurrent lookups if they are RCU protected.
 388  *
 389  * It is valid to have concurrent insertions and deletions protected by per
 390  * bucket locks or concurrent RCU protected lookups and traversals.
 391  */
 392 static int rhashtable_shrink(struct rhashtable *ht)
 393 {
 394         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 395         unsigned int nelems = atomic_read(&ht->nelems);
 396         unsigned int size = 0;
 397
 398         if (nelems)
 399                 size = roundup_pow_of_two(nelems * 3 / 2);
 400         if (size < ht->p.min_size)
 401                 size = ht->p.min_size;
 402
 403         if (old_tbl->size <= size)
 404                 return 0;
 405
 406         if (rht_dereference(old_tbl->future_tbl, ht))
 407                 return -EEXIST;
 408
 409         return rhashtable_rehash_alloc(ht, old_tbl, size);
 410 }
 411
 412 static void rht_deferred_worker(struct work_struct *work)
 413 {
 414         struct rhashtable *ht;
 415         struct bucket_table *tbl;
 416         int err = 0;
 417
 418         ht = container_of(work, struct rhashtable, run_work);
 419         mutex_lock(&ht->mutex);
 420
 421         tbl = rht_dereference(ht->tbl, ht);
 422         tbl = rhashtable_last_table(ht, tbl);
 423
 424         if (rht_grow_above_75(ht, tbl))
 425                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 426         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 427                 err = rhashtable_shrink(ht);
 428         else if (tbl->nest)
 429                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 430
 431         if (!err)
 432                 err = rhashtable_rehash_table(ht);
 433
 434         mutex_unlock(&ht->mutex);
 435
 436         if (err)
 437                 schedule_work(&ht->run_work);
 438 }
 439
 440 static int rhashtable_insert_rehash(struct rhashtable *ht,
 441                                     struct bucket_table *tbl)
 442 {
 443         struct bucket_table *old_tbl;
 444         struct bucket_table *new_tbl;
 445         unsigned int size;
 446         int err;
 447
 448         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 449
 450         size = tbl->size;
 451
 452         err = -EBUSY;
 453
 454         if (rht_grow_above_75(ht, tbl))
 455                 size *= 2;
 456         /* Do not schedule more than one rehash */
 457         else if (old_tbl != tbl)
 458                 goto fail;
 459
 460         err = -ENOMEM;
 461
 462         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 463         if (new_tbl == NULL)
 464                 goto fail;
 465
 466         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 467         if (err) {
 468                 bucket_table_free(new_tbl);
 469                 if (err == -EEXIST)
 470                         err = 0;
 471         } else
 472                 schedule_work(&ht->run_work);
 473
 474         return err;
 475
 476 fail:
 477         /* Do not fail the insert if someone else did a rehash. */
 478         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 479                 return 0;
 480
 481         /* Schedule async rehash to retry allocation in process context. */
 482         if (err == -ENOMEM)
 483                 schedule_work(&ht->run_work);
 484
 485         return err;
 486 }
 487
 488 static void *rhashtable_lookup_one(struct rhashtable *ht,
 489                                    struct bucket_table *tbl, unsigned int hash,
 490                                    const void *key, struct rhash_head *obj)
 491 {
 492         struct rhashtable_compare_arg arg = {
 493                 .ht = ht,
 494                 .key = key,
 495         };
 496         struct rhash_head __rcu **pprev;
 497         struct rhash_head *head;
 498         int elasticity;
 499
 500         elasticity = RHT_ELASTICITY;
 501         pprev = rht_bucket_var(tbl, hash);
 502         rht_for_each_continue(head, *pprev, tbl, hash) {
 503                 struct rhlist_head *list;
 504                 struct rhlist_head *plist;
 505
 506                 elasticity--;
 507                 if (!key ||
 508                     (ht->p.obj_cmpfn ?
 509                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 510                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 511                         pprev = &head->next;
 512                         continue;
 513                 }
 514
 515                 if (!ht->rhlist)
 516                         return rht_obj(ht, head);
 517
 518                 list = container_of(obj, struct rhlist_head, rhead);
 519                 plist = container_of(head, struct rhlist_head, rhead);
 520
 521                 RCU_INIT_POINTER(list->next, plist);
 522                 head = rht_dereference_bucket(head->next, tbl, hash);
 523                 RCU_INIT_POINTER(list->rhead.next, head);
 524                 rcu_assign_pointer(*pprev, obj);
 525
 526                 return NULL;
 527         }
 528
 529         if (elasticity <= 0)
 530                 return ERR_PTR(-EAGAIN);
 531
 532         return ERR_PTR(-ENOENT);
 533 }
 534
 535 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 536                                                   struct bucket_table *tbl,
 537                                                   unsigned int hash,
 538                                                   struct rhash_head *obj,
 539                                                   void *data)
 540 {
 541         struct rhash_head __rcu **pprev;
 542         struct bucket_table *new_tbl;
 543         struct rhash_head *head;
 544
 545         if (!IS_ERR_OR_NULL(data))
 546                 return ERR_PTR(-EEXIST);
 547
 548         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 549                 return ERR_CAST(data);
 550
 551         new_tbl = rcu_dereference(tbl->future_tbl);
 552         if (new_tbl)
 553                 return new_tbl;
 554
 555         if (PTR_ERR(data) != -ENOENT)
 556                 return ERR_CAST(data);
 557
 558         if (unlikely(rht_grow_above_max(ht, tbl)))
 559                 return ERR_PTR(-E2BIG);
 560
 561         if (unlikely(rht_grow_above_100(ht, tbl)))
 562                 return ERR_PTR(-EAGAIN);
 563
 564         pprev = rht_bucket_insert(ht, tbl, hash);
 565         if (!pprev)
 566                 return ERR_PTR(-ENOMEM);
 567
 568         head = rht_dereference_bucket(*pprev, tbl, hash);
 569
 570         RCU_INIT_POINTER(obj->next, head);
 571         if (ht->rhlist) {
 572                 struct rhlist_head *list;
 573
 574                 list = container_of(obj, struct rhlist_head, rhead);
 575                 RCU_INIT_POINTER(list->next, NULL);
 576         }
 577
 578         rcu_assign_pointer(*pprev, obj);
 579
 580         atomic_inc(&ht->nelems);
 581         if (rht_grow_above_75(ht, tbl))
 582                 schedule_work(&ht->run_work);
 583
 584         return NULL;
 585 }
 586
 587 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 588                                    struct rhash_head *obj)
 589 {
 590         struct bucket_table *new_tbl;
 591         struct bucket_table *tbl;
 592         unsigned int hash;
 593         spinlock_t *lock;
 594         void *data;
 595
 596         tbl = rcu_dereference(ht->tbl);
 597
 598         /* All insertions must grab the oldest table containing
 599          * the hashed bucket that is yet to be rehashed.
 600          */
 601         for (;;) {
 602                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 603                 lock = rht_bucket_lock(tbl, hash);
 604                 spin_lock_bh(lock);
 605
 606                 if (tbl->rehash <= hash)
 607                         break;
 608
 609                 spin_unlock_bh(lock);
 610                 tbl = rcu_dereference(tbl->future_tbl);
 611         }
 612
 613         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 614         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 615         if (PTR_ERR(new_tbl) != -EEXIST)
 616                 data = ERR_CAST(new_tbl);
 617
 618         while (!IS_ERR_OR_NULL(new_tbl)) {
 619                 tbl = new_tbl;
 620                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 621                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 622                                  SINGLE_DEPTH_NESTING);
 623
 624                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 625                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 626                 if (PTR_ERR(new_tbl) != -EEXIST)
 627                         data = ERR_CAST(new_tbl);
 628
 629                 spin_unlock(rht_bucket_lock(tbl, hash));
 630         }
 631
 632         spin_unlock_bh(lock);
 633
 634         if (PTR_ERR(data) == -EAGAIN)
 635                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 636                                -EAGAIN);
 637
 638         return data;
 639 }
 640
 641 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 642                              struct rhash_head *obj)
 643 {
 644         void *data;
 645
 646         do {
 647                 rcu_read_lock();
 648                 data = rhashtable_try_insert(ht, key, obj);
 649                 rcu_read_unlock();
 650         } while (PTR_ERR(data) == -EAGAIN);
 651
 652         return data;
 653 }
 654 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 655
 656 /**
 657  * rhashtable_walk_enter - Initialise an iterator
 658  * @ht:         Table to walk over
 659  * @iter:       Hash table Iterator
 660  *
 661  * This function prepares a hash table walk.
 662  *
 663  * Note that if you restart a walk after rhashtable_walk_stop you
 664  * may see the same object twice.  Also, you may miss objects if
 665  * there are removals in between rhashtable_walk_stop and the next
 666  * call to rhashtable_walk_start.
 667  *
 668  * For a completely stable walk you should construct your own data
 669  * structure outside the hash table.
 670  *
 671  * This function may be called from any process context, including
 672  * non-preemptable context, but cannot be called from softirq or
 673  * hardirq context.
 674  *
 675  * You must call rhashtable_walk_exit after this function returns.
 676  */
 677 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 678 {
 679         iter->ht = ht;
 680         iter->p = NULL;
 681         iter->slot = 0;
 682         iter->skip = 0;
 683         iter->end_of_table = 0;
 684
 685         spin_lock(&ht->lock);
 686         iter->walker.tbl =
 687                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 688         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 689         spin_unlock(&ht->lock);
 690 }
 691 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 692
 693 /**
 694  * rhashtable_walk_exit - Free an iterator
 695  * @iter:       Hash table Iterator
 696  *
 697  * This function frees resources allocated by rhashtable_walk_init.
 698  */
 699 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 700 {
 701         spin_lock(&iter->ht->lock);
 702         if (iter->walker.tbl)
 703                 list_del(&iter->walker.list);
 704         spin_unlock(&iter->ht->lock);
 705 }
 706 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 707
 708 /**
 709  * rhashtable_walk_start_check - Start a hash table walk
 710  * @iter:       Hash table iterator
 711  *
 712  * Start a hash table walk at the current iterator position.  Note that we take
 713  * the RCU lock in all cases including when we return an error.  So you must
 714  * always call rhashtable_walk_stop to clean up.
 715  *
 716  * Returns zero if successful.
 717  *
 718  * Returns -EAGAIN if resize event occured.  Note that the iterator
 719  * will rewind back to the beginning and you may use it immediately
 720  * by calling rhashtable_walk_next.
 721  *
 722  * rhashtable_walk_start is defined as an inline variant that returns
 723  * void. This is preferred in cases where the caller would ignore
 724  * resize events and always continue.
 725  */
 726 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 727         __acquires(RCU)
 728 {
 729         struct rhashtable *ht = iter->ht;
 730         bool rhlist = ht->rhlist;
 731
 732         rcu_read_lock();
 733
 734         spin_lock(&ht->lock);
 735         if (iter->walker.tbl)
 736                 list_del(&iter->walker.list);
 737         spin_unlock(&ht->lock);
 738
 739         if (iter->end_of_table)
 740                 return 0;
 741         if (!iter->walker.tbl) {
 742                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 743                 iter->slot = 0;
 744                 iter->skip = 0;
 745                 return -EAGAIN;
 746         }
 747
 748         if (iter->p && !rhlist) {
 749                 /*
 750                  * We need to validate that 'p' is still in the table, and
 751                  * if so, update 'skip'
 752                  */
 753                 struct rhash_head *p;
 754                 int skip = 0;
 755                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 756                         skip++;
 757                         if (p == iter->p) {
 758                                 iter->skip = skip;
 759                                 goto found;
 760                         }
 761                 }
 762                 iter->p = NULL;
 763         } else if (iter->p && rhlist) {
 764                 /* Need to validate that 'list' is still in the table, and
 765                  * if so, update 'skip' and 'p'.
 766                  */
 767                 struct rhash_head *p;
 768                 struct rhlist_head *list;
 769                 int skip = 0;
 770                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 771                         for (list = container_of(p, struct rhlist_head, rhead);
 772                              list;
 773                              list = rcu_dereference(list->next)) {
 774                                 skip++;
 775                                 if (list == iter->list) {
 776                                         iter->p = p;
 777                                         iter->skip = skip;
 778                                         goto found;
 779                                 }
 780                         }
 781                 }
 782                 iter->p = NULL;
 783         }
 784 found:
 785         return 0;
 786 }
 787 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 788
 789 /**
 790  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 791  * one in case of a new walk).
 792  *
 793  * @iter:       Hash table iterator
 794  *
 795  * Returns the found object or NULL when the end of the table is reached.
 796  *
 797  * Returns -EAGAIN if resize event occurred.
 798  */
 799 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 800 {
 801         struct bucket_table *tbl = iter->walker.tbl;
 802         struct rhlist_head *list = iter->list;
 803         struct rhashtable *ht = iter->ht;
 804         struct rhash_head *p = iter->p;
 805         bool rhlist = ht->rhlist;
 806
 807         if (!tbl)
 808                 return NULL;
 809
 810         for (; iter->slot < tbl->size; iter->slot++) {
 811                 int skip = iter->skip;
 812
 813                 rht_for_each_rcu(p, tbl, iter->slot) {
 814                         if (rhlist) {
 815                                 list = container_of(p, struct rhlist_head,
 816                                                     rhead);
 817                                 do {
 818                                         if (!skip)
 819                                                 goto next;
 820                                         skip--;
 821                                         list = rcu_dereference(list->next);
 822                                 } while (list);
 823
 824                                 continue;
 825                         }
 826                         if (!skip)
 827                                 break;
 828                         skip--;
 829                 }
 830
 831 next:
 832                 if (!rht_is_a_nulls(p)) {
 833                         iter->skip++;
 834                         iter->p = p;
 835                         iter->list = list;
 836                         return rht_obj(ht, rhlist ? &list->rhead : p);
 837                 }
 838
 839                 iter->skip = 0;
 840         }
 841
 842         iter->p = NULL;
 843
 844         /* Ensure we see any new tables. */
 845         smp_rmb();
 846
 847         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 848         if (iter->walker.tbl) {
 849                 iter->slot = 0;
 850                 iter->skip = 0;
 851                 return ERR_PTR(-EAGAIN);
 852         } else {
 853                 iter->end_of_table = true;
 854         }
 855
 856         return NULL;
 857 }
 858
 859 /**
 860  * rhashtable_walk_next - Return the next object and advance the iterator
 861  * @iter:       Hash table iterator
 862  *
 863  * Note that you must call rhashtable_walk_stop when you are finished
 864  * with the walk.
 865  *
 866  * Returns the next object or NULL when the end of the table is reached.
 867  *
 868  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 869  * will rewind back to the beginning and you may continue to use it.
 870  */
 871 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 872 {
 873         struct rhlist_head *list = iter->list;
 874         struct rhashtable *ht = iter->ht;
 875         struct rhash_head *p = iter->p;
 876         bool rhlist = ht->rhlist;
 877
 878         if (p) {
 879                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 880                         p = rcu_dereference(p->next);
 881                         list = container_of(p, struct rhlist_head, rhead);
 882                 }
 883                 if (!rht_is_a_nulls(p)) {
 884                         iter->skip++;
 885                         iter->p = p;
 886                         iter->list = list;
 887                         return rht_obj(ht, rhlist ? &list->rhead : p);
 888                 }
 889
 890                 /* At the end of this slot, switch to next one and then find
 891                  * next entry from that point.
 892                  */
 893                 iter->skip = 0;
 894                 iter->slot++;
 895         }
 896
 897         return __rhashtable_walk_find_next(iter);
 898 }
 899 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 900
 901 /**
 902  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 903  * @iter:       Hash table iterator
 904  *
 905  * Returns the next object or NULL when the end of the table is reached.
 906  *
 907  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 908  * will rewind back to the beginning and you may continue to use it.
 909  */
 910 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 911 {
 912         struct rhlist_head *list = iter->list;
 913         struct rhashtable *ht = iter->ht;
 914         struct rhash_head *p = iter->p;
 915
 916         if (p)
 917                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 918
 919         /* No object found in current iter, find next one in the table. */
 920
 921         if (iter->skip) {
 922                 /* A nonzero skip value points to the next entry in the table
 923                  * beyond that last one that was found. Decrement skip so
 924                  * we find the current value. __rhashtable_walk_find_next
 925                  * will restore the original value of skip assuming that
 926                  * the table hasn't changed.
 927                  */
 928                 iter->skip--;
 929         }
 930
 931         return __rhashtable_walk_find_next(iter);
 932 }
 933 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 934
 935 /**
 936  * rhashtable_walk_stop - Finish a hash table walk
 937  * @iter:       Hash table iterator
 938  *
 939  * Finish a hash table walk.  Does not reset the iterator to the start of the
 940  * hash table.
 941  */
 942 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 943         __releases(RCU)
 944 {
 945         struct rhashtable *ht;
 946         struct bucket_table *tbl = iter->walker.tbl;
 947
 948         if (!tbl)
 949                 goto out;
 950
 951         ht = iter->ht;
 952
 953         spin_lock(&ht->lock);
 954         if (tbl->rehash < tbl->size)
 955                 list_add(&iter->walker.list, &tbl->walkers);
 956         else
 957                 iter->walker.tbl = NULL;
 958         spin_unlock(&ht->lock);
 959
 960 out:
 961         rcu_read_unlock();
 962 }
 963 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 964
 965 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 966 {
 967         size_t retsize;
 968
 969         if (params->nelem_hint)
 970                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 971                               (unsigned long)params->min_size);
 972         else
 973                 retsize = max(HASH_DEFAULT_SIZE,
 974                               (unsigned long)params->min_size);
 975
 976         return retsize;
 977 }
 978
 979 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 980 {
 981         return jhash2(key, length, seed);
 982 }
 983
 984 /**
 985  * rhashtable_init - initialize a new hash table
 986  * @ht:         hash table to be initialized
 987  * @params:     configuration parameters
 988  *
 989  * Initializes a new hash table based on the provided configuration
 990  * parameters. A table can be configured either with a variable or
 991  * fixed length key:
 992  *
 993  * Configuration Example 1: Fixed length keys
 994  * struct test_obj {
 995  *      int                     key;
 996  *      void *                  my_member;
 997  *      struct rhash_head       node;
 998  * };
 999  *
1000  * struct rhashtable_params params = {
1001  *      .head_offset = offsetof(struct test_obj, node),
1002  *      .key_offset = offsetof(struct test_obj, key),
1003  *      .key_len = sizeof(int),
1004  *      .hashfn = jhash,
1005  *      .nulls_base = (1U << RHT_BASE_SHIFT),
1006  * };
1007  *
1008  * Configuration Example 2: Variable length keys
1009  * struct test_obj {
1010  *      [...]
1011  *      struct rhash_head       node;
1012  * };
1013  *
1014  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1015  * {
1016  *      struct test_obj *obj = data;
1017  *
1018  *      return [... hash ...];
1019  * }
1020  *
1021  * struct rhashtable_params params = {
1022  *      .head_offset = offsetof(struct test_obj, node),
1023  *      .hashfn = jhash,
1024  *      .obj_hashfn = my_hash_fn,
1025  * };
1026  */
1027 int rhashtable_init(struct rhashtable *ht,
1028                     const struct rhashtable_params *params)
1029 {
1030         struct bucket_table *tbl;
1031         size_t size;
1032
1033         if ((!params->key_len && !params->obj_hashfn) ||
1034             (params->obj_hashfn && !params->obj_cmpfn))
1035                 return -EINVAL;
1036
1037         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
1038                 return -EINVAL;
1039
1040         memset(ht, 0, sizeof(*ht));
1041         mutex_init(&ht->mutex);
1042         spin_lock_init(&ht->lock);
1043         memcpy(&ht->p, params, sizeof(*params));
1044
1045         if (params->min_size)
1046                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1047
1048         /* Cap total entries at 2^31 to avoid nelems overflow. */
1049         ht->max_elems = 1u << 31;
1050
1051         if (params->max_size) {
1052                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1053                 if (ht->p.max_size < ht->max_elems / 2)
1054                         ht->max_elems = ht->p.max_size * 2;
1055         }
1056
1057         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1058
1059         size = rounded_hashtable_size(&ht->p);
1060
1061         if (params->locks_mul)
1062                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1063         else
1064                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1065
1066         ht->key_len = ht->p.key_len;
1067         if (!params->hashfn) {
1068                 ht->p.hashfn = jhash;
1069
1070                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1071                         ht->key_len /= sizeof(u32);
1072                         ht->p.hashfn = rhashtable_jhash2;
1073                 }
1074         }
1075
1076         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1077         if (tbl == NULL)
1078                 return -ENOMEM;
1079
1080         atomic_set(&ht->nelems, 0);
1081
1082         RCU_INIT_POINTER(ht->tbl, tbl);
1083
1084         INIT_WORK(&ht->run_work, rht_deferred_worker);
1085
1086         return 0;
1087 }
1088 EXPORT_SYMBOL_GPL(rhashtable_init);
1089
1090 /**
1091  * rhltable_init - initialize a new hash list table
1092  * @hlt:        hash list table to be initialized
1093  * @params:     configuration parameters
1094  *
1095  * Initializes a new hash list table.
1096  *
1097  * See documentation for rhashtable_init.
1098  */
1099 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1100 {
1101         int err;
1102
1103         /* No rhlist NULLs marking for now. */
1104         if (params->nulls_base)
1105                 return -EINVAL;
1106
1107         err = rhashtable_init(&hlt->ht, params);
1108         hlt->ht.rhlist = true;
1109         return err;
1110 }
1111 EXPORT_SYMBOL_GPL(rhltable_init);
1112
1113 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1114                                 void (*free_fn)(void *ptr, void *arg),
1115                                 void *arg)
1116 {
1117         struct rhlist_head *list;
1118
1119         if (!ht->rhlist) {
1120                 free_fn(rht_obj(ht, obj), arg);
1121                 return;
1122         }
1123
1124         list = container_of(obj, struct rhlist_head, rhead);
1125         do {
1126                 obj = &list->rhead;
1127                 list = rht_dereference(list->next, ht);
1128                 free_fn(rht_obj(ht, obj), arg);
1129         } while (list);
1130 }
1131
1132 /**
1133  * rhashtable_free_and_destroy - free elements and destroy hash table
1134  * @ht:         the hash table to destroy
1135  * @free_fn:    callback to release resources of element
1136  * @arg:        pointer passed to free_fn
1137  *
1138  * Stops an eventual async resize. If defined, invokes free_fn for each
1139  * element to releasal resources. Please note that RCU protected
1140  * readers may still be accessing the elements. Releasing of resources
1141  * must occur in a compatible manner. Then frees the bucket array.
1142  *
1143  * This function will eventually sleep to wait for an async resize
1144  * to complete. The caller is responsible that no further write operations
1145  * occurs in parallel.
1146  */
1147 void rhashtable_free_and_destroy(struct rhashtable *ht,
1148                                  void (*free_fn)(void *ptr, void *arg),
1149                                  void *arg)
1150 {
1151         struct bucket_table *tbl, *next_tbl;
1152         unsigned int i;
1153
1154         cancel_work_sync(&ht->run_work);
1155
1156         mutex_lock(&ht->mutex);
1157         tbl = rht_dereference(ht->tbl, ht);
1158 restart:
1159         if (free_fn) {
1160                 for (i = 0; i < tbl->size; i++) {
1161                         struct rhash_head *pos, *next;
1162
1163                         cond_resched();
1164                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1165                              next = !rht_is_a_nulls(pos) ?
1166                                         rht_dereference(pos->next, ht) : NULL;
1167                              !rht_is_a_nulls(pos);
1168                              pos = next,
1169                              next = !rht_is_a_nulls(pos) ?
1170                                         rht_dereference(pos->next, ht) : NULL)
1171                                 rhashtable_free_one(ht, pos, free_fn, arg);
1172                 }
1173         }
1174
1175         next_tbl = rht_dereference(tbl->future_tbl, ht);
1176         bucket_table_free(tbl);
1177         if (next_tbl) {
1178                 tbl = next_tbl;
1179                 goto restart;
1180         }
1181         mutex_unlock(&ht->mutex);
1182 }
1183 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1184
1185 void rhashtable_destroy(struct rhashtable *ht)
1186 {
1187         return rhashtable_free_and_destroy(ht, NULL, NULL);
1188 }
1189 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1190
1191 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1192                                             unsigned int hash)
1193 {
1194         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1195         static struct rhash_head __rcu *rhnull =
1196                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1197         unsigned int index = hash & ((1 << tbl->nest) - 1);
1198         unsigned int size = tbl->size >> tbl->nest;
1199         unsigned int subhash = hash;
1200         union nested_table *ntbl;
1201
1202         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1203         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1204         subhash >>= tbl->nest;
1205
1206         while (ntbl && size > (1 << shift)) {
1207                 index = subhash & ((1 << shift) - 1);
1208                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1209                                                   tbl, hash);
1210                 size >>= shift;
1211                 subhash >>= shift;
1212         }
1213
1214         if (!ntbl)
1215                 return &rhnull;
1216
1217         return &ntbl[subhash].bucket;
1218
1219 }
1220 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1221
1222 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1223                                                    struct bucket_table *tbl,
1224                                                    unsigned int hash)
1225 {
1226         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1227         unsigned int index = hash & ((1 << tbl->nest) - 1);
1228         unsigned int size = tbl->size >> tbl->nest;
1229         union nested_table *ntbl;
1230         unsigned int shifted;
1231         unsigned int nhash;
1232
1233         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1234         hash >>= tbl->nest;
1235         nhash = index;
1236         shifted = tbl->nest;
1237         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1238                                   size <= (1 << shift) ? shifted : 0, nhash);
1239
1240         while (ntbl && size > (1 << shift)) {
1241                 index = hash & ((1 << shift) - 1);
1242                 size >>= shift;
1243                 hash >>= shift;
1244                 nhash |= index << shifted;
1245                 shifted += shift;
1246                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1247                                           size <= (1 << shift) ? shifted : 0,
1248                                           nhash);
1249         }
1250
1251         if (!ntbl)
1252                 return NULL;
1253
1254         return &ntbl[hash].bucket;
1255
1256 }
1257 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);