lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2015 Herbert Xu <[email protected]>
   5  * Copyright (c) 2014-2015 Thomas Graf <[email protected]>
   6  * Copyright (c) 2008-2014 Patrick McHardy <[email protected]>
   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                                               bool leaf)
 119 {
 120         union nested_table *ntbl;
 121         int i;
 122
 123         ntbl = rcu_dereference(*prev);
 124         if (ntbl)
 125                 return ntbl;
 126
 127         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 128
 129         if (ntbl && leaf) {
 130                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
 131                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
 132         }
 133
 134         rcu_assign_pointer(*prev, ntbl);
 135
 136         return ntbl;
 137 }
 138
 139 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 140                                                       size_t nbuckets,
 141                                                       gfp_t gfp)
 142 {
 143         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 144         struct bucket_table *tbl;
 145         size_t size;
 146
 147         if (nbuckets < (1 << (shift + 1)))
 148                 return NULL;
 149
 150         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 151
 152         tbl = kzalloc(size, gfp);
 153         if (!tbl)
 154                 return NULL;
 155
 156         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 157                                 false)) {
 158                 kfree(tbl);
 159                 return NULL;
 160         }
 161
 162         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 163
 164         return tbl;
 165 }
 166
 167 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 168                                                size_t nbuckets,
 169                                                gfp_t gfp)
 170 {
 171         struct bucket_table *tbl = NULL;
 172         size_t size, max_locks;
 173         int i;
 174
 175         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 176         tbl = kvzalloc(size, gfp);
 177
 178         size = nbuckets;
 179
 180         if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
 181                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 182                 nbuckets = 0;
 183         }
 184
 185         if (tbl == NULL)
 186                 return NULL;
 187
 188         tbl->size = size;
 189
 190         max_locks = size >> 1;
 191         if (tbl->nest)
 192                 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
 193
 194         if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
 195                                    ht->p.locks_mul, gfp) < 0) {
 196                 bucket_table_free(tbl);
 197                 return NULL;
 198         }
 199
 200         rcu_head_init(&tbl->rcu);
 201         INIT_LIST_HEAD(&tbl->walkers);
 202
 203         tbl->hash_rnd = get_random_u32();
 204
 205         for (i = 0; i < nbuckets; i++)
 206                 INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
 207
 208         return tbl;
 209 }
 210
 211 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 212                                                   struct bucket_table *tbl)
 213 {
 214         struct bucket_table *new_tbl;
 215
 216         do {
 217                 new_tbl = tbl;
 218                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 219         } while (tbl);
 220
 221         return new_tbl;
 222 }
 223
 224 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 225 {
 226         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 227         struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
 228         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 229         int err = -EAGAIN;
 230         struct rhash_head *head, *next, *entry;
 231         spinlock_t *new_bucket_lock;
 232         unsigned int new_hash;
 233
 234         if (new_tbl->nest)
 235                 goto out;
 236
 237         err = -ENOENT;
 238
 239         rht_for_each(entry, old_tbl, old_hash) {
 240                 err = 0;
 241                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 242
 243                 if (rht_is_a_nulls(next))
 244                         break;
 245
 246                 pprev = &entry->next;
 247         }
 248
 249         if (err)
 250                 goto out;
 251
 252         new_hash = head_hashfn(ht, new_tbl, entry);
 253
 254         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 255
 256         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 257         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 258                                       new_tbl, new_hash);
 259
 260         RCU_INIT_POINTER(entry->next, head);
 261
 262         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 263         spin_unlock(new_bucket_lock);
 264
 265         rcu_assign_pointer(*pprev, next);
 266
 267 out:
 268         return err;
 269 }
 270
 271 static int rhashtable_rehash_chain(struct rhashtable *ht,
 272                                     unsigned int old_hash)
 273 {
 274         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 275         spinlock_t *old_bucket_lock;
 276         int err;
 277
 278         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 279
 280         spin_lock_bh(old_bucket_lock);
 281         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 282                 ;
 283
 284         if (err == -ENOENT)
 285                 err = 0;
 286
 287         spin_unlock_bh(old_bucket_lock);
 288
 289         return err;
 290 }
 291
 292 static int rhashtable_rehash_attach(struct rhashtable *ht,
 293                                     struct bucket_table *old_tbl,
 294                                     struct bucket_table *new_tbl)
 295 {
 296         /* Make insertions go into the new, empty table right away. Deletions
 297          * and lookups will be attempted in both tables until we synchronize.
 298          * As cmpxchg() provides strong barriers, we do not need
 299          * rcu_assign_pointer().
 300          */
 301
 302         if (cmpxchg(&old_tbl->future_tbl, NULL, new_tbl) != NULL)
 303                 return -EEXIST;
 304
 305         return 0;
 306 }
 307
 308 static int rhashtable_rehash_table(struct rhashtable *ht)
 309 {
 310         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 311         struct bucket_table *new_tbl;
 312         struct rhashtable_walker *walker;
 313         unsigned int old_hash;
 314         int err;
 315
 316         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 317         if (!new_tbl)
 318                 return 0;
 319
 320         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 321                 err = rhashtable_rehash_chain(ht, old_hash);
 322                 if (err)
 323                         return err;
 324                 cond_resched();
 325         }
 326
 327         /* Publish the new table pointer. */
 328         rcu_assign_pointer(ht->tbl, new_tbl);
 329
 330         spin_lock(&ht->lock);
 331         list_for_each_entry(walker, &old_tbl->walkers, list)
 332                 walker->tbl = NULL;
 333
 334         /* Wait for readers. All new readers will see the new
 335          * table, and thus no references to the old table will
 336          * remain.
 337          * We do this inside the locked region so that
 338          * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
 339          * to check if it should not re-link the table.
 340          */
 341         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 342         spin_unlock(&ht->lock);
 343
 344         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 345 }
 346
 347 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 348                                    struct bucket_table *old_tbl,
 349                                    unsigned int size)
 350 {
 351         struct bucket_table *new_tbl;
 352         int err;
 353
 354         ASSERT_RHT_MUTEX(ht);
 355
 356         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 357         if (new_tbl == NULL)
 358                 return -ENOMEM;
 359
 360         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 361         if (err)
 362                 bucket_table_free(new_tbl);
 363
 364         return err;
 365 }
 366
 367 /**
 368  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 369  * @ht:         the hash table to shrink
 370  *
 371  * This function shrinks the hash table to fit, i.e., the smallest
 372  * size would not cause it to expand right away automatically.
 373  *
 374  * The caller must ensure that no concurrent resizing occurs by holding
 375  * ht->mutex.
 376  *
 377  * The caller must ensure that no concurrent table mutations take place.
 378  * It is however valid to have concurrent lookups if they are RCU protected.
 379  *
 380  * It is valid to have concurrent insertions and deletions protected by per
 381  * bucket locks or concurrent RCU protected lookups and traversals.
 382  */
 383 static int rhashtable_shrink(struct rhashtable *ht)
 384 {
 385         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 386         unsigned int nelems = atomic_read(&ht->nelems);
 387         unsigned int size = 0;
 388
 389         if (nelems)
 390                 size = roundup_pow_of_two(nelems * 3 / 2);
 391         if (size < ht->p.min_size)
 392                 size = ht->p.min_size;
 393
 394         if (old_tbl->size <= size)
 395                 return 0;
 396
 397         if (rht_dereference(old_tbl->future_tbl, ht))
 398                 return -EEXIST;
 399
 400         return rhashtable_rehash_alloc(ht, old_tbl, size);
 401 }
 402
 403 static void rht_deferred_worker(struct work_struct *work)
 404 {
 405         struct rhashtable *ht;
 406         struct bucket_table *tbl;
 407         int err = 0;
 408
 409         ht = container_of(work, struct rhashtable, run_work);
 410         mutex_lock(&ht->mutex);
 411
 412         tbl = rht_dereference(ht->tbl, ht);
 413         tbl = rhashtable_last_table(ht, tbl);
 414
 415         if (rht_grow_above_75(ht, tbl))
 416                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 417         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 418                 err = rhashtable_shrink(ht);
 419         else if (tbl->nest)
 420                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 421
 422         if (!err)
 423                 err = rhashtable_rehash_table(ht);
 424
 425         mutex_unlock(&ht->mutex);
 426
 427         if (err)
 428                 schedule_work(&ht->run_work);
 429 }
 430
 431 static int rhashtable_insert_rehash(struct rhashtable *ht,
 432                                     struct bucket_table *tbl)
 433 {
 434         struct bucket_table *old_tbl;
 435         struct bucket_table *new_tbl;
 436         unsigned int size;
 437         int err;
 438
 439         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 440
 441         size = tbl->size;
 442
 443         err = -EBUSY;
 444
 445         if (rht_grow_above_75(ht, tbl))
 446                 size *= 2;
 447         /* Do not schedule more than one rehash */
 448         else if (old_tbl != tbl)
 449                 goto fail;
 450
 451         err = -ENOMEM;
 452
 453         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 454         if (new_tbl == NULL)
 455                 goto fail;
 456
 457         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 458         if (err) {
 459                 bucket_table_free(new_tbl);
 460                 if (err == -EEXIST)
 461                         err = 0;
 462         } else
 463                 schedule_work(&ht->run_work);
 464
 465         return err;
 466
 467 fail:
 468         /* Do not fail the insert if someone else did a rehash. */
 469         if (likely(rcu_access_pointer(tbl->future_tbl)))
 470                 return 0;
 471
 472         /* Schedule async rehash to retry allocation in process context. */
 473         if (err == -ENOMEM)
 474                 schedule_work(&ht->run_work);
 475
 476         return err;
 477 }
 478
 479 static void *rhashtable_lookup_one(struct rhashtable *ht,
 480                                    struct bucket_table *tbl, unsigned int hash,
 481                                    const void *key, struct rhash_head *obj)
 482 {
 483         struct rhashtable_compare_arg arg = {
 484                 .ht = ht,
 485                 .key = key,
 486         };
 487         struct rhash_head __rcu **pprev;
 488         struct rhash_head *head;
 489         int elasticity;
 490
 491         elasticity = RHT_ELASTICITY;
 492         pprev = rht_bucket_var(tbl, hash);
 493         rht_for_each_from(head, *pprev, tbl, hash) {
 494                 struct rhlist_head *list;
 495                 struct rhlist_head *plist;
 496
 497                 elasticity--;
 498                 if (!key ||
 499                     (ht->p.obj_cmpfn ?
 500                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 501                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 502                         pprev = &head->next;
 503                         continue;
 504                 }
 505
 506                 if (!ht->rhlist)
 507                         return rht_obj(ht, head);
 508
 509                 list = container_of(obj, struct rhlist_head, rhead);
 510                 plist = container_of(head, struct rhlist_head, rhead);
 511
 512                 RCU_INIT_POINTER(list->next, plist);
 513                 head = rht_dereference_bucket(head->next, tbl, hash);
 514                 RCU_INIT_POINTER(list->rhead.next, head);
 515                 rcu_assign_pointer(*pprev, obj);
 516
 517                 return NULL;
 518         }
 519
 520         if (elasticity <= 0)
 521                 return ERR_PTR(-EAGAIN);
 522
 523         return ERR_PTR(-ENOENT);
 524 }
 525
 526 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 527                                                   struct bucket_table *tbl,
 528                                                   unsigned int hash,
 529                                                   struct rhash_head *obj,
 530                                                   void *data)
 531 {
 532         struct rhash_head __rcu **pprev;
 533         struct bucket_table *new_tbl;
 534         struct rhash_head *head;
 535
 536         if (!IS_ERR_OR_NULL(data))
 537                 return ERR_PTR(-EEXIST);
 538
 539         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 540                 return ERR_CAST(data);
 541
 542         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 543         if (new_tbl)
 544                 return new_tbl;
 545
 546         if (PTR_ERR(data) != -ENOENT)
 547                 return ERR_CAST(data);
 548
 549         if (unlikely(rht_grow_above_max(ht, tbl)))
 550                 return ERR_PTR(-E2BIG);
 551
 552         if (unlikely(rht_grow_above_100(ht, tbl)))
 553                 return ERR_PTR(-EAGAIN);
 554
 555         pprev = rht_bucket_insert(ht, tbl, hash);
 556         if (!pprev)
 557                 return ERR_PTR(-ENOMEM);
 558
 559         head = rht_dereference_bucket(*pprev, tbl, hash);
 560
 561         RCU_INIT_POINTER(obj->next, head);
 562         if (ht->rhlist) {
 563                 struct rhlist_head *list;
 564
 565                 list = container_of(obj, struct rhlist_head, rhead);
 566                 RCU_INIT_POINTER(list->next, NULL);
 567         }
 568
 569         rcu_assign_pointer(*pprev, obj);
 570
 571         atomic_inc(&ht->nelems);
 572         if (rht_grow_above_75(ht, tbl))
 573                 schedule_work(&ht->run_work);
 574
 575         return NULL;
 576 }
 577
 578 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 579                                    struct rhash_head *obj)
 580 {
 581         struct bucket_table *new_tbl;
 582         struct bucket_table *tbl;
 583         unsigned int hash;
 584         void *data;
 585
 586         new_tbl = rcu_dereference(ht->tbl);
 587
 588         do {
 589                 tbl = new_tbl;
 590                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 591                 spin_lock_bh(rht_bucket_lock(tbl, hash));
 592
 593                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 594                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 595                 if (PTR_ERR(new_tbl) != -EEXIST)
 596                         data = ERR_CAST(new_tbl);
 597
 598                 spin_unlock_bh(rht_bucket_lock(tbl, hash));
 599         } while (!IS_ERR_OR_NULL(new_tbl));
 600
 601         if (PTR_ERR(data) == -EAGAIN)
 602                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 603                                -EAGAIN);
 604
 605         return data;
 606 }
 607
 608 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 609                              struct rhash_head *obj)
 610 {
 611         void *data;
 612
 613         do {
 614                 rcu_read_lock();
 615                 data = rhashtable_try_insert(ht, key, obj);
 616                 rcu_read_unlock();
 617         } while (PTR_ERR(data) == -EAGAIN);
 618
 619         return data;
 620 }
 621 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 622
 623 /**
 624  * rhashtable_walk_enter - Initialise an iterator
 625  * @ht:         Table to walk over
 626  * @iter:       Hash table Iterator
 627  *
 628  * This function prepares a hash table walk.
 629  *
 630  * Note that if you restart a walk after rhashtable_walk_stop you
 631  * may see the same object twice.  Also, you may miss objects if
 632  * there are removals in between rhashtable_walk_stop and the next
 633  * call to rhashtable_walk_start.
 634  *
 635  * For a completely stable walk you should construct your own data
 636  * structure outside the hash table.
 637  *
 638  * This function may be called from any process context, including
 639  * non-preemptable context, but cannot be called from softirq or
 640  * hardirq context.
 641  *
 642  * You must call rhashtable_walk_exit after this function returns.
 643  */
 644 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 645 {
 646         iter->ht = ht;
 647         iter->p = NULL;
 648         iter->slot = 0;
 649         iter->skip = 0;
 650         iter->end_of_table = 0;
 651
 652         spin_lock(&ht->lock);
 653         iter->walker.tbl =
 654                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 655         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 656         spin_unlock(&ht->lock);
 657 }
 658 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 659
 660 /**
 661  * rhashtable_walk_exit - Free an iterator
 662  * @iter:       Hash table Iterator
 663  *
 664  * This function frees resources allocated by rhashtable_walk_enter.
 665  */
 666 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 667 {
 668         spin_lock(&iter->ht->lock);
 669         if (iter->walker.tbl)
 670                 list_del(&iter->walker.list);
 671         spin_unlock(&iter->ht->lock);
 672 }
 673 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 674
 675 /**
 676  * rhashtable_walk_start_check - Start a hash table walk
 677  * @iter:       Hash table iterator
 678  *
 679  * Start a hash table walk at the current iterator position.  Note that we take
 680  * the RCU lock in all cases including when we return an error.  So you must
 681  * always call rhashtable_walk_stop to clean up.
 682  *
 683  * Returns zero if successful.
 684  *
 685  * Returns -EAGAIN if resize event occured.  Note that the iterator
 686  * will rewind back to the beginning and you may use it immediately
 687  * by calling rhashtable_walk_next.
 688  *
 689  * rhashtable_walk_start is defined as an inline variant that returns
 690  * void. This is preferred in cases where the caller would ignore
 691  * resize events and always continue.
 692  */
 693 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 694         __acquires(RCU)
 695 {
 696         struct rhashtable *ht = iter->ht;
 697         bool rhlist = ht->rhlist;
 698
 699         rcu_read_lock();
 700
 701         spin_lock(&ht->lock);
 702         if (iter->walker.tbl)
 703                 list_del(&iter->walker.list);
 704         spin_unlock(&ht->lock);
 705
 706         if (iter->end_of_table)
 707                 return 0;
 708         if (!iter->walker.tbl) {
 709                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 710                 iter->slot = 0;
 711                 iter->skip = 0;
 712                 return -EAGAIN;
 713         }
 714
 715         if (iter->p && !rhlist) {
 716                 /*
 717                  * We need to validate that 'p' is still in the table, and
 718                  * if so, update 'skip'
 719                  */
 720                 struct rhash_head *p;
 721                 int skip = 0;
 722                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 723                         skip++;
 724                         if (p == iter->p) {
 725                                 iter->skip = skip;
 726                                 goto found;
 727                         }
 728                 }
 729                 iter->p = NULL;
 730         } else if (iter->p && rhlist) {
 731                 /* Need to validate that 'list' is still in the table, and
 732                  * if so, update 'skip' and 'p'.
 733                  */
 734                 struct rhash_head *p;
 735                 struct rhlist_head *list;
 736                 int skip = 0;
 737                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 738                         for (list = container_of(p, struct rhlist_head, rhead);
 739                              list;
 740                              list = rcu_dereference(list->next)) {
 741                                 skip++;
 742                                 if (list == iter->list) {
 743                                         iter->p = p;
 744                                         iter->skip = skip;
 745                                         goto found;
 746                                 }
 747                         }
 748                 }
 749                 iter->p = NULL;
 750         }
 751 found:
 752         return 0;
 753 }
 754 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 755
 756 /**
 757  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 758  * one in case of a new walk).
 759  *
 760  * @iter:       Hash table iterator
 761  *
 762  * Returns the found object or NULL when the end of the table is reached.
 763  *
 764  * Returns -EAGAIN if resize event occurred.
 765  */
 766 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 767 {
 768         struct bucket_table *tbl = iter->walker.tbl;
 769         struct rhlist_head *list = iter->list;
 770         struct rhashtable *ht = iter->ht;
 771         struct rhash_head *p = iter->p;
 772         bool rhlist = ht->rhlist;
 773
 774         if (!tbl)
 775                 return NULL;
 776
 777         for (; iter->slot < tbl->size; iter->slot++) {
 778                 int skip = iter->skip;
 779
 780                 rht_for_each_rcu(p, tbl, iter->slot) {
 781                         if (rhlist) {
 782                                 list = container_of(p, struct rhlist_head,
 783                                                     rhead);
 784                                 do {
 785                                         if (!skip)
 786                                                 goto next;
 787                                         skip--;
 788                                         list = rcu_dereference(list->next);
 789                                 } while (list);
 790
 791                                 continue;
 792                         }
 793                         if (!skip)
 794                                 break;
 795                         skip--;
 796                 }
 797
 798 next:
 799                 if (!rht_is_a_nulls(p)) {
 800                         iter->skip++;
 801                         iter->p = p;
 802                         iter->list = list;
 803                         return rht_obj(ht, rhlist ? &list->rhead : p);
 804                 }
 805
 806                 iter->skip = 0;
 807         }
 808
 809         iter->p = NULL;
 810
 811         /* Ensure we see any new tables. */
 812         smp_rmb();
 813
 814         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 815         if (iter->walker.tbl) {
 816                 iter->slot = 0;
 817                 iter->skip = 0;
 818                 return ERR_PTR(-EAGAIN);
 819         } else {
 820                 iter->end_of_table = true;
 821         }
 822
 823         return NULL;
 824 }
 825
 826 /**
 827  * rhashtable_walk_next - Return the next object and advance the iterator
 828  * @iter:       Hash table iterator
 829  *
 830  * Note that you must call rhashtable_walk_stop when you are finished
 831  * with the walk.
 832  *
 833  * Returns the next object or NULL when the end of the table is reached.
 834  *
 835  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 836  * will rewind back to the beginning and you may continue to use it.
 837  */
 838 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 839 {
 840         struct rhlist_head *list = iter->list;
 841         struct rhashtable *ht = iter->ht;
 842         struct rhash_head *p = iter->p;
 843         bool rhlist = ht->rhlist;
 844
 845         if (p) {
 846                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 847                         p = rcu_dereference(p->next);
 848                         list = container_of(p, struct rhlist_head, rhead);
 849                 }
 850                 if (!rht_is_a_nulls(p)) {
 851                         iter->skip++;
 852                         iter->p = p;
 853                         iter->list = list;
 854                         return rht_obj(ht, rhlist ? &list->rhead : p);
 855                 }
 856
 857                 /* At the end of this slot, switch to next one and then find
 858                  * next entry from that point.
 859                  */
 860                 iter->skip = 0;
 861                 iter->slot++;
 862         }
 863
 864         return __rhashtable_walk_find_next(iter);
 865 }
 866 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 867
 868 /**
 869  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 870  * @iter:       Hash table iterator
 871  *
 872  * Returns the next object or NULL when the end of the table is reached.
 873  *
 874  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 875  * will rewind back to the beginning and you may continue to use it.
 876  */
 877 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 878 {
 879         struct rhlist_head *list = iter->list;
 880         struct rhashtable *ht = iter->ht;
 881         struct rhash_head *p = iter->p;
 882
 883         if (p)
 884                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 885
 886         /* No object found in current iter, find next one in the table. */
 887
 888         if (iter->skip) {
 889                 /* A nonzero skip value points to the next entry in the table
 890                  * beyond that last one that was found. Decrement skip so
 891                  * we find the current value. __rhashtable_walk_find_next
 892                  * will restore the original value of skip assuming that
 893                  * the table hasn't changed.
 894                  */
 895                 iter->skip--;
 896         }
 897
 898         return __rhashtable_walk_find_next(iter);
 899 }
 900 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 901
 902 /**
 903  * rhashtable_walk_stop - Finish a hash table walk
 904  * @iter:       Hash table iterator
 905  *
 906  * Finish a hash table walk.  Does not reset the iterator to the start of the
 907  * hash table.
 908  */
 909 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 910         __releases(RCU)
 911 {
 912         struct rhashtable *ht;
 913         struct bucket_table *tbl = iter->walker.tbl;
 914
 915         if (!tbl)
 916                 goto out;
 917
 918         ht = iter->ht;
 919
 920         spin_lock(&ht->lock);
 921         if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
 922                 /* This bucket table is being freed, don't re-link it. */
 923                 iter->walker.tbl = NULL;
 924         else
 925                 list_add(&iter->walker.list, &tbl->walkers);
 926         spin_unlock(&ht->lock);
 927
 928 out:
 929         rcu_read_unlock();
 930 }
 931 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 932
 933 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 934 {
 935         size_t retsize;
 936
 937         if (params->nelem_hint)
 938                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 939                               (unsigned long)params->min_size);
 940         else
 941                 retsize = max(HASH_DEFAULT_SIZE,
 942                               (unsigned long)params->min_size);
 943
 944         return retsize;
 945 }
 946
 947 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 948 {
 949         return jhash2(key, length, seed);
 950 }
 951
 952 /**
 953  * rhashtable_init - initialize a new hash table
 954  * @ht:         hash table to be initialized
 955  * @params:     configuration parameters
 956  *
 957  * Initializes a new hash table based on the provided configuration
 958  * parameters. A table can be configured either with a variable or
 959  * fixed length key:
 960  *
 961  * Configuration Example 1: Fixed length keys
 962  * struct test_obj {
 963  *      int                     key;
 964  *      void *                  my_member;
 965  *      struct rhash_head       node;
 966  * };
 967  *
 968  * struct rhashtable_params params = {
 969  *      .head_offset = offsetof(struct test_obj, node),
 970  *      .key_offset = offsetof(struct test_obj, key),
 971  *      .key_len = sizeof(int),
 972  *      .hashfn = jhash,
 973  * };
 974  *
 975  * Configuration Example 2: Variable length keys
 976  * struct test_obj {
 977  *      [...]
 978  *      struct rhash_head       node;
 979  * };
 980  *
 981  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 982  * {
 983  *      struct test_obj *obj = data;
 984  *
 985  *      return [... hash ...];
 986  * }
 987  *
 988  * struct rhashtable_params params = {
 989  *      .head_offset = offsetof(struct test_obj, node),
 990  *      .hashfn = jhash,
 991  *      .obj_hashfn = my_hash_fn,
 992  * };
 993  */
 994 int rhashtable_init(struct rhashtable *ht,
 995                     const struct rhashtable_params *params)
 996 {
 997         struct bucket_table *tbl;
 998         size_t size;
 999
1000         if ((!params->key_len && !params->obj_hashfn) ||
1001             (params->obj_hashfn && !params->obj_cmpfn))
1002                 return -EINVAL;
1003
1004         memset(ht, 0, sizeof(*ht));
1005         mutex_init(&ht->mutex);
1006         spin_lock_init(&ht->lock);
1007         memcpy(&ht->p, params, sizeof(*params));
1008
1009         if (params->min_size)
1010                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1011
1012         /* Cap total entries at 2^31 to avoid nelems overflow. */
1013         ht->max_elems = 1u << 31;
1014
1015         if (params->max_size) {
1016                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1017                 if (ht->p.max_size < ht->max_elems / 2)
1018                         ht->max_elems = ht->p.max_size * 2;
1019         }
1020
1021         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1022
1023         size = rounded_hashtable_size(&ht->p);
1024
1025         if (params->locks_mul)
1026                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1027         else
1028                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1029
1030         ht->key_len = ht->p.key_len;
1031         if (!params->hashfn) {
1032                 ht->p.hashfn = jhash;
1033
1034                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1035                         ht->key_len /= sizeof(u32);
1036                         ht->p.hashfn = rhashtable_jhash2;
1037                 }
1038         }
1039
1040         /*
1041          * This is api initialization and thus we need to guarantee the
1042          * initial rhashtable allocation. Upon failure, retry with the
1043          * smallest possible size with __GFP_NOFAIL semantics.
1044          */
1045         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1046         if (unlikely(tbl == NULL)) {
1047                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1048                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1049         }
1050
1051         atomic_set(&ht->nelems, 0);
1052
1053         RCU_INIT_POINTER(ht->tbl, tbl);
1054
1055         INIT_WORK(&ht->run_work, rht_deferred_worker);
1056
1057         return 0;
1058 }
1059 EXPORT_SYMBOL_GPL(rhashtable_init);
1060
1061 /**
1062  * rhltable_init - initialize a new hash list table
1063  * @hlt:        hash list table to be initialized
1064  * @params:     configuration parameters
1065  *
1066  * Initializes a new hash list table.
1067  *
1068  * See documentation for rhashtable_init.
1069  */
1070 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1071 {
1072         int err;
1073
1074         err = rhashtable_init(&hlt->ht, params);
1075         hlt->ht.rhlist = true;
1076         return err;
1077 }
1078 EXPORT_SYMBOL_GPL(rhltable_init);
1079
1080 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1081                                 void (*free_fn)(void *ptr, void *arg),
1082                                 void *arg)
1083 {
1084         struct rhlist_head *list;
1085
1086         if (!ht->rhlist) {
1087                 free_fn(rht_obj(ht, obj), arg);
1088                 return;
1089         }
1090
1091         list = container_of(obj, struct rhlist_head, rhead);
1092         do {
1093                 obj = &list->rhead;
1094                 list = rht_dereference(list->next, ht);
1095                 free_fn(rht_obj(ht, obj), arg);
1096         } while (list);
1097 }
1098
1099 /**
1100  * rhashtable_free_and_destroy - free elements and destroy hash table
1101  * @ht:         the hash table to destroy
1102  * @free_fn:    callback to release resources of element
1103  * @arg:        pointer passed to free_fn
1104  *
1105  * Stops an eventual async resize. If defined, invokes free_fn for each
1106  * element to releasal resources. Please note that RCU protected
1107  * readers may still be accessing the elements. Releasing of resources
1108  * must occur in a compatible manner. Then frees the bucket array.
1109  *
1110  * This function will eventually sleep to wait for an async resize
1111  * to complete. The caller is responsible that no further write operations
1112  * occurs in parallel.
1113  */
1114 void rhashtable_free_and_destroy(struct rhashtable *ht,
1115                                  void (*free_fn)(void *ptr, void *arg),
1116                                  void *arg)
1117 {
1118         struct bucket_table *tbl, *next_tbl;
1119         unsigned int i;
1120
1121         cancel_work_sync(&ht->run_work);
1122
1123         mutex_lock(&ht->mutex);
1124         tbl = rht_dereference(ht->tbl, ht);
1125 restart:
1126         if (free_fn) {
1127                 for (i = 0; i < tbl->size; i++) {
1128                         struct rhash_head *pos, *next;
1129
1130                         cond_resched();
1131                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1132                              next = !rht_is_a_nulls(pos) ?
1133                                         rht_dereference(pos->next, ht) : NULL;
1134                              !rht_is_a_nulls(pos);
1135                              pos = next,
1136                              next = !rht_is_a_nulls(pos) ?
1137                                         rht_dereference(pos->next, ht) : NULL)
1138                                 rhashtable_free_one(ht, pos, free_fn, arg);
1139                 }
1140         }
1141
1142         next_tbl = rht_dereference(tbl->future_tbl, ht);
1143         bucket_table_free(tbl);
1144         if (next_tbl) {
1145                 tbl = next_tbl;
1146                 goto restart;
1147         }
1148         mutex_unlock(&ht->mutex);
1149 }
1150 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1151
1152 void rhashtable_destroy(struct rhashtable *ht)
1153 {
1154         return rhashtable_free_and_destroy(ht, NULL, NULL);
1155 }
1156 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1157
1158 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1159                                             unsigned int hash)
1160 {
1161         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1162         static struct rhash_head __rcu *rhnull;
1163         unsigned int index = hash & ((1 << tbl->nest) - 1);
1164         unsigned int size = tbl->size >> tbl->nest;
1165         unsigned int subhash = hash;
1166         union nested_table *ntbl;
1167
1168         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1169         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1170         subhash >>= tbl->nest;
1171
1172         while (ntbl && size > (1 << shift)) {
1173                 index = subhash & ((1 << shift) - 1);
1174                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1175                                                   tbl, hash);
1176                 size >>= shift;
1177                 subhash >>= shift;
1178         }
1179
1180         if (!ntbl) {
1181                 if (!rhnull)
1182                         INIT_RHT_NULLS_HEAD(rhnull);
1183                 return &rhnull;
1184         }
1185
1186         return &ntbl[subhash].bucket;
1187
1188 }
1189 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1190
1191 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1192                                                    struct bucket_table *tbl,
1193                                                    unsigned int hash)
1194 {
1195         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1196         unsigned int index = hash & ((1 << tbl->nest) - 1);
1197         unsigned int size = tbl->size >> tbl->nest;
1198         union nested_table *ntbl;
1199
1200         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1201         hash >>= tbl->nest;
1202         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1203                                   size <= (1 << shift));
1204
1205         while (ntbl && size > (1 << shift)) {
1206                 index = hash & ((1 << shift) - 1);
1207                 size >>= shift;
1208                 hash >>= shift;
1209                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1210                                           size <= (1 << shift));
1211         }
1212
1213         if (!ntbl)
1214                 return NULL;
1215
1216         return &ntbl[hash].bucket;
1217
1218 }
1219 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);