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 #include <linux/rhashtable.h>
  32
  33 #define HASH_DEFAULT_SIZE       64UL
  34 #define HASH_MIN_SIZE           4U
  35 #define BUCKET_LOCKS_PER_CPU    32UL
  36
  37 union nested_table {
  38         union nested_table __rcu *table;
  39         struct rhash_head __rcu *bucket;
  40 };
  41
  42 static u32 head_hashfn(struct rhashtable *ht,
  43                        const struct bucket_table *tbl,
  44                        const struct rhash_head *he)
  45 {
  46         return rht_head_hashfn(ht, tbl, he, ht->p);
  47 }
  48
  49 #ifdef CONFIG_PROVE_LOCKING
  50 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  51
  52 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  53 {
  54         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  55 }
  56 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  57
  58 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  59 {
  60         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  61
  62         return (debug_locks) ? lockdep_is_held(lock) : 1;
  63 }
  64 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  65 #else
  66 #define ASSERT_RHT_MUTEX(HT)
  67 #endif
  68
  69 static void nested_table_free(union nested_table *ntbl, unsigned int size)
  70 {
  71         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
  72         const unsigned int len = 1 << shift;
  73         unsigned int i;
  74
  75         ntbl = rcu_dereference_raw(ntbl->table);
  76         if (!ntbl)
  77                 return;
  78
  79         if (size > len) {
  80                 size >>= shift;
  81                 for (i = 0; i < len; i++)
  82                         nested_table_free(ntbl + i, size);
  83         }
  84
  85         kfree(ntbl);
  86 }
  87
  88 static void nested_bucket_table_free(const struct bucket_table *tbl)
  89 {
  90         unsigned int size = tbl->size >> tbl->nest;
  91         unsigned int len = 1 << tbl->nest;
  92         union nested_table *ntbl;
  93         unsigned int i;
  94
  95         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
  96
  97         for (i = 0; i < len; i++)
  98                 nested_table_free(ntbl + i, size);
  99
 100         kfree(ntbl);
 101 }
 102
 103 static void bucket_table_free(const struct bucket_table *tbl)
 104 {
 105         if (tbl->nest)
 106                 nested_bucket_table_free(tbl);
 107
 108         free_bucket_spinlocks(tbl->locks);
 109         kvfree(tbl);
 110 }
 111
 112 static void bucket_table_free_rcu(struct rcu_head *head)
 113 {
 114         bucket_table_free(container_of(head, struct bucket_table, rcu));
 115 }
 116
 117 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 118                                               union nested_table __rcu **prev,
 119                                               bool leaf)
 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 && leaf) {
 131                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
 132                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
 133         }
 134
 135         rcu_assign_pointer(*prev, ntbl);
 136
 137         return ntbl;
 138 }
 139
 140 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 141                                                       size_t nbuckets,
 142                                                       gfp_t gfp)
 143 {
 144         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 145         struct bucket_table *tbl;
 146         size_t size;
 147
 148         if (nbuckets < (1 << (shift + 1)))
 149                 return NULL;
 150
 151         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 152
 153         tbl = kzalloc(size, gfp);
 154         if (!tbl)
 155                 return NULL;
 156
 157         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 158                                 false)) {
 159                 kfree(tbl);
 160                 return NULL;
 161         }
 162
 163         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 164
 165         return tbl;
 166 }
 167
 168 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 169                                                size_t nbuckets,
 170                                                gfp_t gfp)
 171 {
 172         struct bucket_table *tbl = NULL;
 173         size_t size, max_locks;
 174         int i;
 175
 176         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 177         tbl = kvzalloc(size, gfp);
 178
 179         size = nbuckets;
 180
 181         if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
 182                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 183                 nbuckets = 0;
 184         }
 185
 186         if (tbl == NULL)
 187                 return NULL;
 188
 189         tbl->size = size;
 190
 191         max_locks = size >> 1;
 192         if (tbl->nest)
 193                 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
 194
 195         if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
 196                                    ht->p.locks_mul, gfp) < 0) {
 197                 bucket_table_free(tbl);
 198                 return NULL;
 199         }
 200
 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                 old_tbl->rehash++;
 286                 err = 0;
 287         }
 288         spin_unlock_bh(old_bucket_lock);
 289
 290         return err;
 291 }
 292
 293 static int rhashtable_rehash_attach(struct rhashtable *ht,
 294                                     struct bucket_table *old_tbl,
 295                                     struct bucket_table *new_tbl)
 296 {
 297         /* Make insertions go into the new, empty table right away. Deletions
 298          * and lookups will be attempted in both tables until we synchronize.
 299          * As cmpxchg() provides strong barriers, we do not need
 300          * rcu_assign_pointer().
 301          */
 302
 303         if (cmpxchg(&old_tbl->future_tbl, NULL, new_tbl) != NULL)
 304                 return -EEXIST;
 305
 306         return 0;
 307 }
 308
 309 static int rhashtable_rehash_table(struct rhashtable *ht)
 310 {
 311         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 312         struct bucket_table *new_tbl;
 313         struct rhashtable_walker *walker;
 314         unsigned int old_hash;
 315         int err;
 316
 317         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 318         if (!new_tbl)
 319                 return 0;
 320
 321         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 322                 err = rhashtable_rehash_chain(ht, old_hash);
 323                 if (err)
 324                         return err;
 325                 cond_resched();
 326         }
 327
 328         /* Publish the new table pointer. */
 329         rcu_assign_pointer(ht->tbl, new_tbl);
 330
 331         spin_lock(&ht->lock);
 332         list_for_each_entry(walker, &old_tbl->walkers, list)
 333                 walker->tbl = NULL;
 334         spin_unlock(&ht->lock);
 335
 336         /* Wait for readers. All new readers will see the new
 337          * table, and thus no references to the old table will
 338          * remain.
 339          */
 340         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 341
 342         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 343 }
 344
 345 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 346                                    struct bucket_table *old_tbl,
 347                                    unsigned int size)
 348 {
 349         struct bucket_table *new_tbl;
 350         int err;
 351
 352         ASSERT_RHT_MUTEX(ht);
 353
 354         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 355         if (new_tbl == NULL)
 356                 return -ENOMEM;
 357
 358         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 359         if (err)
 360                 bucket_table_free(new_tbl);
 361
 362         return err;
 363 }
 364
 365 /**
 366  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 367  * @ht:         the hash table to shrink
 368  *
 369  * This function shrinks the hash table to fit, i.e., the smallest
 370  * size would not cause it to expand right away automatically.
 371  *
 372  * The caller must ensure that no concurrent resizing occurs by holding
 373  * ht->mutex.
 374  *
 375  * The caller must ensure that no concurrent table mutations take place.
 376  * It is however valid to have concurrent lookups if they are RCU protected.
 377  *
 378  * It is valid to have concurrent insertions and deletions protected by per
 379  * bucket locks or concurrent RCU protected lookups and traversals.
 380  */
 381 static int rhashtable_shrink(struct rhashtable *ht)
 382 {
 383         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 384         unsigned int nelems = atomic_read(&ht->nelems);
 385         unsigned int size = 0;
 386
 387         if (nelems)
 388                 size = roundup_pow_of_two(nelems * 3 / 2);
 389         if (size < ht->p.min_size)
 390                 size = ht->p.min_size;
 391
 392         if (old_tbl->size <= size)
 393                 return 0;
 394
 395         if (rht_dereference(old_tbl->future_tbl, ht))
 396                 return -EEXIST;
 397
 398         return rhashtable_rehash_alloc(ht, old_tbl, size);
 399 }
 400
 401 static void rht_deferred_worker(struct work_struct *work)
 402 {
 403         struct rhashtable *ht;
 404         struct bucket_table *tbl;
 405         int err = 0;
 406
 407         ht = container_of(work, struct rhashtable, run_work);
 408         mutex_lock(&ht->mutex);
 409
 410         tbl = rht_dereference(ht->tbl, ht);
 411         tbl = rhashtable_last_table(ht, tbl);
 412
 413         if (rht_grow_above_75(ht, tbl))
 414                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 415         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 416                 err = rhashtable_shrink(ht);
 417         else if (tbl->nest)
 418                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 419
 420         if (!err)
 421                 err = rhashtable_rehash_table(ht);
 422
 423         mutex_unlock(&ht->mutex);
 424
 425         if (err)
 426                 schedule_work(&ht->run_work);
 427 }
 428
 429 static int rhashtable_insert_rehash(struct rhashtable *ht,
 430                                     struct bucket_table *tbl)
 431 {
 432         struct bucket_table *old_tbl;
 433         struct bucket_table *new_tbl;
 434         unsigned int size;
 435         int err;
 436
 437         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 438
 439         size = tbl->size;
 440
 441         err = -EBUSY;
 442
 443         if (rht_grow_above_75(ht, tbl))
 444                 size *= 2;
 445         /* Do not schedule more than one rehash */
 446         else if (old_tbl != tbl)
 447                 goto fail;
 448
 449         err = -ENOMEM;
 450
 451         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 452         if (new_tbl == NULL)
 453                 goto fail;
 454
 455         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 456         if (err) {
 457                 bucket_table_free(new_tbl);
 458                 if (err == -EEXIST)
 459                         err = 0;
 460         } else
 461                 schedule_work(&ht->run_work);
 462
 463         return err;
 464
 465 fail:
 466         /* Do not fail the insert if someone else did a rehash. */
 467         if (likely(rcu_access_pointer(tbl->future_tbl)))
 468                 return 0;
 469
 470         /* Schedule async rehash to retry allocation in process context. */
 471         if (err == -ENOMEM)
 472                 schedule_work(&ht->run_work);
 473
 474         return err;
 475 }
 476
 477 static void *rhashtable_lookup_one(struct rhashtable *ht,
 478                                    struct bucket_table *tbl, unsigned int hash,
 479                                    const void *key, struct rhash_head *obj)
 480 {
 481         struct rhashtable_compare_arg arg = {
 482                 .ht = ht,
 483                 .key = key,
 484         };
 485         struct rhash_head __rcu **pprev;
 486         struct rhash_head *head;
 487         int elasticity;
 488
 489         elasticity = RHT_ELASTICITY;
 490         pprev = rht_bucket_var(tbl, hash);
 491         rht_for_each_continue(head, *pprev, tbl, hash) {
 492                 struct rhlist_head *list;
 493                 struct rhlist_head *plist;
 494
 495                 elasticity--;
 496                 if (!key ||
 497                     (ht->p.obj_cmpfn ?
 498                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 499                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 500                         pprev = &head->next;
 501                         continue;
 502                 }
 503
 504                 if (!ht->rhlist)
 505                         return rht_obj(ht, head);
 506
 507                 list = container_of(obj, struct rhlist_head, rhead);
 508                 plist = container_of(head, struct rhlist_head, rhead);
 509
 510                 RCU_INIT_POINTER(list->next, plist);
 511                 head = rht_dereference_bucket(head->next, tbl, hash);
 512                 RCU_INIT_POINTER(list->rhead.next, head);
 513                 rcu_assign_pointer(*pprev, obj);
 514
 515                 return NULL;
 516         }
 517
 518         if (elasticity <= 0)
 519                 return ERR_PTR(-EAGAIN);
 520
 521         return ERR_PTR(-ENOENT);
 522 }
 523
 524 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 525                                                   struct bucket_table *tbl,
 526                                                   unsigned int hash,
 527                                                   struct rhash_head *obj,
 528                                                   void *data)
 529 {
 530         struct rhash_head __rcu **pprev;
 531         struct bucket_table *new_tbl;
 532         struct rhash_head *head;
 533
 534         if (!IS_ERR_OR_NULL(data))
 535                 return ERR_PTR(-EEXIST);
 536
 537         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 538                 return ERR_CAST(data);
 539
 540         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 541         if (new_tbl)
 542                 return new_tbl;
 543
 544         if (PTR_ERR(data) != -ENOENT)
 545                 return ERR_CAST(data);
 546
 547         if (unlikely(rht_grow_above_max(ht, tbl)))
 548                 return ERR_PTR(-E2BIG);
 549
 550         if (unlikely(rht_grow_above_100(ht, tbl)))
 551                 return ERR_PTR(-EAGAIN);
 552
 553         pprev = rht_bucket_insert(ht, tbl, hash);
 554         if (!pprev)
 555                 return ERR_PTR(-ENOMEM);
 556
 557         head = rht_dereference_bucket(*pprev, tbl, hash);
 558
 559         RCU_INIT_POINTER(obj->next, head);
 560         if (ht->rhlist) {
 561                 struct rhlist_head *list;
 562
 563                 list = container_of(obj, struct rhlist_head, rhead);
 564                 RCU_INIT_POINTER(list->next, NULL);
 565         }
 566
 567         rcu_assign_pointer(*pprev, obj);
 568
 569         atomic_inc(&ht->nelems);
 570         if (rht_grow_above_75(ht, tbl))
 571                 schedule_work(&ht->run_work);
 572
 573         return NULL;
 574 }
 575
 576 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 577                                    struct rhash_head *obj)
 578 {
 579         struct bucket_table *new_tbl;
 580         struct bucket_table *tbl;
 581         unsigned int hash;
 582         spinlock_t *lock;
 583         void *data;
 584
 585         tbl = rcu_dereference(ht->tbl);
 586
 587         /* All insertions must grab the oldest table containing
 588          * the hashed bucket that is yet to be rehashed.
 589          */
 590         for (;;) {
 591                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 592                 lock = rht_bucket_lock(tbl, hash);
 593                 spin_lock_bh(lock);
 594
 595                 if (tbl->rehash <= hash)
 596                         break;
 597
 598                 spin_unlock_bh(lock);
 599                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 600         }
 601
 602         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 603         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 604         if (PTR_ERR(new_tbl) != -EEXIST)
 605                 data = ERR_CAST(new_tbl);
 606
 607         while (!IS_ERR_OR_NULL(new_tbl)) {
 608                 tbl = new_tbl;
 609                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 610                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 611                                  SINGLE_DEPTH_NESTING);
 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                 spin_unlock(rht_bucket_lock(tbl, hash));
 619         }
 620
 621         spin_unlock_bh(lock);
 622
 623         if (PTR_ERR(data) == -EAGAIN)
 624                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 625                                -EAGAIN);
 626
 627         return data;
 628 }
 629
 630 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 631                              struct rhash_head *obj)
 632 {
 633         void *data;
 634
 635         do {
 636                 rcu_read_lock();
 637                 data = rhashtable_try_insert(ht, key, obj);
 638                 rcu_read_unlock();
 639         } while (PTR_ERR(data) == -EAGAIN);
 640
 641         return data;
 642 }
 643 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 644
 645 /**
 646  * rhashtable_walk_enter - Initialise an iterator
 647  * @ht:         Table to walk over
 648  * @iter:       Hash table Iterator
 649  *
 650  * This function prepares a hash table walk.
 651  *
 652  * Note that if you restart a walk after rhashtable_walk_stop you
 653  * may see the same object twice.  Also, you may miss objects if
 654  * there are removals in between rhashtable_walk_stop and the next
 655  * call to rhashtable_walk_start.
 656  *
 657  * For a completely stable walk you should construct your own data
 658  * structure outside the hash table.
 659  *
 660  * This function may be called from any process context, including
 661  * non-preemptable context, but cannot be called from softirq or
 662  * hardirq context.
 663  *
 664  * You must call rhashtable_walk_exit after this function returns.
 665  */
 666 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 667 {
 668         iter->ht = ht;
 669         iter->p = NULL;
 670         iter->slot = 0;
 671         iter->skip = 0;
 672         iter->end_of_table = 0;
 673
 674         spin_lock(&ht->lock);
 675         iter->walker.tbl =
 676                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 677         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 678         spin_unlock(&ht->lock);
 679 }
 680 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 681
 682 /**
 683  * rhashtable_walk_exit - Free an iterator
 684  * @iter:       Hash table Iterator
 685  *
 686  * This function frees resources allocated by rhashtable_walk_init.
 687  */
 688 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 689 {
 690         spin_lock(&iter->ht->lock);
 691         if (iter->walker.tbl)
 692                 list_del(&iter->walker.list);
 693         spin_unlock(&iter->ht->lock);
 694 }
 695 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 696
 697 /**
 698  * rhashtable_walk_start_check - Start a hash table walk
 699  * @iter:       Hash table iterator
 700  *
 701  * Start a hash table walk at the current iterator position.  Note that we take
 702  * the RCU lock in all cases including when we return an error.  So you must
 703  * always call rhashtable_walk_stop to clean up.
 704  *
 705  * Returns zero if successful.
 706  *
 707  * Returns -EAGAIN if resize event occured.  Note that the iterator
 708  * will rewind back to the beginning and you may use it immediately
 709  * by calling rhashtable_walk_next.
 710  *
 711  * rhashtable_walk_start is defined as an inline variant that returns
 712  * void. This is preferred in cases where the caller would ignore
 713  * resize events and always continue.
 714  */
 715 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 716         __acquires(RCU)
 717 {
 718         struct rhashtable *ht = iter->ht;
 719         bool rhlist = ht->rhlist;
 720
 721         rcu_read_lock();
 722
 723         spin_lock(&ht->lock);
 724         if (iter->walker.tbl)
 725                 list_del(&iter->walker.list);
 726         spin_unlock(&ht->lock);
 727
 728         if (iter->end_of_table)
 729                 return 0;
 730         if (!iter->walker.tbl) {
 731                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 732                 iter->slot = 0;
 733                 iter->skip = 0;
 734                 return -EAGAIN;
 735         }
 736
 737         if (iter->p && !rhlist) {
 738                 /*
 739                  * We need to validate that 'p' is still in the table, and
 740                  * if so, update 'skip'
 741                  */
 742                 struct rhash_head *p;
 743                 int skip = 0;
 744                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 745                         skip++;
 746                         if (p == iter->p) {
 747                                 iter->skip = skip;
 748                                 goto found;
 749                         }
 750                 }
 751                 iter->p = NULL;
 752         } else if (iter->p && rhlist) {
 753                 /* Need to validate that 'list' is still in the table, and
 754                  * if so, update 'skip' and 'p'.
 755                  */
 756                 struct rhash_head *p;
 757                 struct rhlist_head *list;
 758                 int skip = 0;
 759                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 760                         for (list = container_of(p, struct rhlist_head, rhead);
 761                              list;
 762                              list = rcu_dereference(list->next)) {
 763                                 skip++;
 764                                 if (list == iter->list) {
 765                                         iter->p = p;
 766                                         iter->skip = skip;
 767                                         goto found;
 768                                 }
 769                         }
 770                 }
 771                 iter->p = NULL;
 772         }
 773 found:
 774         return 0;
 775 }
 776 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 777
 778 /**
 779  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 780  * one in case of a new walk).
 781  *
 782  * @iter:       Hash table iterator
 783  *
 784  * Returns the found object or NULL when the end of the table is reached.
 785  *
 786  * Returns -EAGAIN if resize event occurred.
 787  */
 788 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 789 {
 790         struct bucket_table *tbl = iter->walker.tbl;
 791         struct rhlist_head *list = iter->list;
 792         struct rhashtable *ht = iter->ht;
 793         struct rhash_head *p = iter->p;
 794         bool rhlist = ht->rhlist;
 795
 796         if (!tbl)
 797                 return NULL;
 798
 799         for (; iter->slot < tbl->size; iter->slot++) {
 800                 int skip = iter->skip;
 801
 802                 rht_for_each_rcu(p, tbl, iter->slot) {
 803                         if (rhlist) {
 804                                 list = container_of(p, struct rhlist_head,
 805                                                     rhead);
 806                                 do {
 807                                         if (!skip)
 808                                                 goto next;
 809                                         skip--;
 810                                         list = rcu_dereference(list->next);
 811                                 } while (list);
 812
 813                                 continue;
 814                         }
 815                         if (!skip)
 816                                 break;
 817                         skip--;
 818                 }
 819
 820 next:
 821                 if (!rht_is_a_nulls(p)) {
 822                         iter->skip++;
 823                         iter->p = p;
 824                         iter->list = list;
 825                         return rht_obj(ht, rhlist ? &list->rhead : p);
 826                 }
 827
 828                 iter->skip = 0;
 829         }
 830
 831         iter->p = NULL;
 832
 833         /* Ensure we see any new tables. */
 834         smp_rmb();
 835
 836         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 837         if (iter->walker.tbl) {
 838                 iter->slot = 0;
 839                 iter->skip = 0;
 840                 return ERR_PTR(-EAGAIN);
 841         } else {
 842                 iter->end_of_table = true;
 843         }
 844
 845         return NULL;
 846 }
 847
 848 /**
 849  * rhashtable_walk_next - Return the next object and advance the iterator
 850  * @iter:       Hash table iterator
 851  *
 852  * Note that you must call rhashtable_walk_stop when you are finished
 853  * with the walk.
 854  *
 855  * Returns the next object or NULL when the end of the table is reached.
 856  *
 857  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 858  * will rewind back to the beginning and you may continue to use it.
 859  */
 860 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 861 {
 862         struct rhlist_head *list = iter->list;
 863         struct rhashtable *ht = iter->ht;
 864         struct rhash_head *p = iter->p;
 865         bool rhlist = ht->rhlist;
 866
 867         if (p) {
 868                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 869                         p = rcu_dereference(p->next);
 870                         list = container_of(p, struct rhlist_head, rhead);
 871                 }
 872                 if (!rht_is_a_nulls(p)) {
 873                         iter->skip++;
 874                         iter->p = p;
 875                         iter->list = list;
 876                         return rht_obj(ht, rhlist ? &list->rhead : p);
 877                 }
 878
 879                 /* At the end of this slot, switch to next one and then find
 880                  * next entry from that point.
 881                  */
 882                 iter->skip = 0;
 883                 iter->slot++;
 884         }
 885
 886         return __rhashtable_walk_find_next(iter);
 887 }
 888 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 889
 890 /**
 891  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 892  * @iter:       Hash table iterator
 893  *
 894  * Returns the next object or NULL when the end of the table is reached.
 895  *
 896  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 897  * will rewind back to the beginning and you may continue to use it.
 898  */
 899 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 900 {
 901         struct rhlist_head *list = iter->list;
 902         struct rhashtable *ht = iter->ht;
 903         struct rhash_head *p = iter->p;
 904
 905         if (p)
 906                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 907
 908         /* No object found in current iter, find next one in the table. */
 909
 910         if (iter->skip) {
 911                 /* A nonzero skip value points to the next entry in the table
 912                  * beyond that last one that was found. Decrement skip so
 913                  * we find the current value. __rhashtable_walk_find_next
 914                  * will restore the original value of skip assuming that
 915                  * the table hasn't changed.
 916                  */
 917                 iter->skip--;
 918         }
 919
 920         return __rhashtable_walk_find_next(iter);
 921 }
 922 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 923
 924 /**
 925  * rhashtable_walk_stop - Finish a hash table walk
 926  * @iter:       Hash table iterator
 927  *
 928  * Finish a hash table walk.  Does not reset the iterator to the start of the
 929  * hash table.
 930  */
 931 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 932         __releases(RCU)
 933 {
 934         struct rhashtable *ht;
 935         struct bucket_table *tbl = iter->walker.tbl;
 936
 937         if (!tbl)
 938                 goto out;
 939
 940         ht = iter->ht;
 941
 942         spin_lock(&ht->lock);
 943         if (tbl->rehash < tbl->size)
 944                 list_add(&iter->walker.list, &tbl->walkers);
 945         else
 946                 iter->walker.tbl = NULL;
 947         spin_unlock(&ht->lock);
 948
 949 out:
 950         rcu_read_unlock();
 951 }
 952 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 953
 954 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 955 {
 956         size_t retsize;
 957
 958         if (params->nelem_hint)
 959                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 960                               (unsigned long)params->min_size);
 961         else
 962                 retsize = max(HASH_DEFAULT_SIZE,
 963                               (unsigned long)params->min_size);
 964
 965         return retsize;
 966 }
 967
 968 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 969 {
 970         return jhash2(key, length, seed);
 971 }
 972
 973 /**
 974  * rhashtable_init - initialize a new hash table
 975  * @ht:         hash table to be initialized
 976  * @params:     configuration parameters
 977  *
 978  * Initializes a new hash table based on the provided configuration
 979  * parameters. A table can be configured either with a variable or
 980  * fixed length key:
 981  *
 982  * Configuration Example 1: Fixed length keys
 983  * struct test_obj {
 984  *      int                     key;
 985  *      void *                  my_member;
 986  *      struct rhash_head       node;
 987  * };
 988  *
 989  * struct rhashtable_params params = {
 990  *      .head_offset = offsetof(struct test_obj, node),
 991  *      .key_offset = offsetof(struct test_obj, key),
 992  *      .key_len = sizeof(int),
 993  *      .hashfn = jhash,
 994  * };
 995  *
 996  * Configuration Example 2: Variable length keys
 997  * struct test_obj {
 998  *      [...]
 999  *      struct rhash_head       node;
1000  * };
1001  *
1002  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1003  * {
1004  *      struct test_obj *obj = data;
1005  *
1006  *      return [... hash ...];
1007  * }
1008  *
1009  * struct rhashtable_params params = {
1010  *      .head_offset = offsetof(struct test_obj, node),
1011  *      .hashfn = jhash,
1012  *      .obj_hashfn = my_hash_fn,
1013  * };
1014  */
1015 int rhashtable_init(struct rhashtable *ht,
1016                     const struct rhashtable_params *params)
1017 {
1018         struct bucket_table *tbl;
1019         size_t size;
1020
1021         if ((!params->key_len && !params->obj_hashfn) ||
1022             (params->obj_hashfn && !params->obj_cmpfn))
1023                 return -EINVAL;
1024
1025         memset(ht, 0, sizeof(*ht));
1026         mutex_init(&ht->mutex);
1027         spin_lock_init(&ht->lock);
1028         memcpy(&ht->p, params, sizeof(*params));
1029
1030         if (params->min_size)
1031                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1032
1033         /* Cap total entries at 2^31 to avoid nelems overflow. */
1034         ht->max_elems = 1u << 31;
1035
1036         if (params->max_size) {
1037                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1038                 if (ht->p.max_size < ht->max_elems / 2)
1039                         ht->max_elems = ht->p.max_size * 2;
1040         }
1041
1042         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1043
1044         size = rounded_hashtable_size(&ht->p);
1045
1046         if (params->locks_mul)
1047                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1048         else
1049                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1050
1051         ht->key_len = ht->p.key_len;
1052         if (!params->hashfn) {
1053                 ht->p.hashfn = jhash;
1054
1055                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1056                         ht->key_len /= sizeof(u32);
1057                         ht->p.hashfn = rhashtable_jhash2;
1058                 }
1059         }
1060
1061         /*
1062          * This is api initialization and thus we need to guarantee the
1063          * initial rhashtable allocation. Upon failure, retry with the
1064          * smallest possible size with __GFP_NOFAIL semantics.
1065          */
1066         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1067         if (unlikely(tbl == NULL)) {
1068                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1069                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1070         }
1071
1072         atomic_set(&ht->nelems, 0);
1073
1074         RCU_INIT_POINTER(ht->tbl, tbl);
1075
1076         INIT_WORK(&ht->run_work, rht_deferred_worker);
1077
1078         return 0;
1079 }
1080 EXPORT_SYMBOL_GPL(rhashtable_init);
1081
1082 /**
1083  * rhltable_init - initialize a new hash list table
1084  * @hlt:        hash list table to be initialized
1085  * @params:     configuration parameters
1086  *
1087  * Initializes a new hash list table.
1088  *
1089  * See documentation for rhashtable_init.
1090  */
1091 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1092 {
1093         int err;
1094
1095         err = rhashtable_init(&hlt->ht, params);
1096         hlt->ht.rhlist = true;
1097         return err;
1098 }
1099 EXPORT_SYMBOL_GPL(rhltable_init);
1100
1101 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1102                                 void (*free_fn)(void *ptr, void *arg),
1103                                 void *arg)
1104 {
1105         struct rhlist_head *list;
1106
1107         if (!ht->rhlist) {
1108                 free_fn(rht_obj(ht, obj), arg);
1109                 return;
1110         }
1111
1112         list = container_of(obj, struct rhlist_head, rhead);
1113         do {
1114                 obj = &list->rhead;
1115                 list = rht_dereference(list->next, ht);
1116                 free_fn(rht_obj(ht, obj), arg);
1117         } while (list);
1118 }
1119
1120 /**
1121  * rhashtable_free_and_destroy - free elements and destroy hash table
1122  * @ht:         the hash table to destroy
1123  * @free_fn:    callback to release resources of element
1124  * @arg:        pointer passed to free_fn
1125  *
1126  * Stops an eventual async resize. If defined, invokes free_fn for each
1127  * element to releasal resources. Please note that RCU protected
1128  * readers may still be accessing the elements. Releasing of resources
1129  * must occur in a compatible manner. Then frees the bucket array.
1130  *
1131  * This function will eventually sleep to wait for an async resize
1132  * to complete. The caller is responsible that no further write operations
1133  * occurs in parallel.
1134  */
1135 void rhashtable_free_and_destroy(struct rhashtable *ht,
1136                                  void (*free_fn)(void *ptr, void *arg),
1137                                  void *arg)
1138 {
1139         struct bucket_table *tbl, *next_tbl;
1140         unsigned int i;
1141
1142         cancel_work_sync(&ht->run_work);
1143
1144         mutex_lock(&ht->mutex);
1145         tbl = rht_dereference(ht->tbl, ht);
1146 restart:
1147         if (free_fn) {
1148                 for (i = 0; i < tbl->size; i++) {
1149                         struct rhash_head *pos, *next;
1150
1151                         cond_resched();
1152                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1153                              next = !rht_is_a_nulls(pos) ?
1154                                         rht_dereference(pos->next, ht) : NULL;
1155                              !rht_is_a_nulls(pos);
1156                              pos = next,
1157                              next = !rht_is_a_nulls(pos) ?
1158                                         rht_dereference(pos->next, ht) : NULL)
1159                                 rhashtable_free_one(ht, pos, free_fn, arg);
1160                 }
1161         }
1162
1163         next_tbl = rht_dereference(tbl->future_tbl, ht);
1164         bucket_table_free(tbl);
1165         if (next_tbl) {
1166                 tbl = next_tbl;
1167                 goto restart;
1168         }
1169         mutex_unlock(&ht->mutex);
1170 }
1171 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1172
1173 void rhashtable_destroy(struct rhashtable *ht)
1174 {
1175         return rhashtable_free_and_destroy(ht, NULL, NULL);
1176 }
1177 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1178
1179 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1180                                             unsigned int hash)
1181 {
1182         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1183         static struct rhash_head __rcu *rhnull =
1184                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1185         unsigned int index = hash & ((1 << tbl->nest) - 1);
1186         unsigned int size = tbl->size >> tbl->nest;
1187         unsigned int subhash = hash;
1188         union nested_table *ntbl;
1189
1190         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1191         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1192         subhash >>= tbl->nest;
1193
1194         while (ntbl && size > (1 << shift)) {
1195                 index = subhash & ((1 << shift) - 1);
1196                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1197                                                   tbl, hash);
1198                 size >>= shift;
1199                 subhash >>= shift;
1200         }
1201
1202         if (!ntbl)
1203                 return &rhnull;
1204
1205         return &ntbl[subhash].bucket;
1206
1207 }
1208 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1209
1210 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1211                                                    struct bucket_table *tbl,
1212                                                    unsigned int hash)
1213 {
1214         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1215         unsigned int index = hash & ((1 << tbl->nest) - 1);
1216         unsigned int size = tbl->size >> tbl->nest;
1217         union nested_table *ntbl;
1218
1219         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1220         hash >>= tbl->nest;
1221         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1222                                   size <= (1 << shift));
1223
1224         while (ntbl && size > (1 << shift)) {
1225                 index = hash & ((1 << shift) - 1);
1226                 size >>= shift;
1227                 hash >>= shift;
1228                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1229                                           size <= (1 << shift));
1230         }
1231
1232         if (!ntbl)
1233                 return NULL;
1234
1235         return &ntbl[hash].bucket;
1236
1237 }
1238 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);