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
  35 union nested_table {
  36         union nested_table __rcu *table;
  37         struct rhash_lock_head __rcu *bucket;
  38 };
  39
  40 static u32 head_hashfn(struct rhashtable *ht,
  41                        const struct bucket_table *tbl,
  42                        const struct rhash_head *he)
  43 {
  44         return rht_head_hashfn(ht, tbl, he, ht->p);
  45 }
  46
  47 #ifdef CONFIG_PROVE_LOCKING
  48 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  49
  50 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  51 {
  52         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  53 }
  54 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  55
  56 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  57 {
  58         if (!debug_locks)
  59                 return 1;
  60         if (unlikely(tbl->nest))
  61                 return 1;
  62         return bit_spin_is_locked(0, (unsigned long *)&tbl->buckets[hash]);
  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         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         if (cmpxchg(prev, NULL, ntbl) == NULL)
 135                 return ntbl;
 136         /* Raced with another thread. */
 137         kfree(ntbl);
 138         return rcu_dereference(*prev);
 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                                 false)) {
 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;
 175         int i;
 176         static struct lock_class_key __key;
 177
 178         tbl = kvzalloc(struct_size(tbl, buckets, nbuckets), gfp);
 179
 180         size = nbuckets;
 181
 182         if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
 183                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 184                 nbuckets = 0;
 185         }
 186
 187         if (tbl == NULL)
 188                 return NULL;
 189
 190         lockdep_init_map(&tbl->dep_map, "rhashtable_bucket", &__key, 0);
 191
 192         tbl->size = size;
 193
 194         rcu_head_init(&tbl->rcu);
 195         INIT_LIST_HEAD(&tbl->walkers);
 196
 197         tbl->hash_rnd = get_random_u32();
 198
 199         for (i = 0; i < nbuckets; i++)
 200                 INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
 201
 202         return tbl;
 203 }
 204
 205 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 206                                                   struct bucket_table *tbl)
 207 {
 208         struct bucket_table *new_tbl;
 209
 210         do {
 211                 new_tbl = tbl;
 212                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 213         } while (tbl);
 214
 215         return new_tbl;
 216 }
 217
 218 static int rhashtable_rehash_one(struct rhashtable *ht,
 219                                  struct rhash_lock_head __rcu **bkt,
 220                                  unsigned int old_hash)
 221 {
 222         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 223         struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
 224         int err = -EAGAIN;
 225         struct rhash_head *head, *next, *entry;
 226         struct rhash_head __rcu **pprev = NULL;
 227         unsigned int new_hash;
 228
 229         if (new_tbl->nest)
 230                 goto out;
 231
 232         err = -ENOENT;
 233
 234         rht_for_each_from(entry, rht_ptr(bkt, old_tbl, old_hash),
 235                           old_tbl, old_hash) {
 236                 err = 0;
 237                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 238
 239                 if (rht_is_a_nulls(next))
 240                         break;
 241
 242                 pprev = &entry->next;
 243         }
 244
 245         if (err)
 246                 goto out;
 247
 248         new_hash = head_hashfn(ht, new_tbl, entry);
 249
 250         rht_lock_nested(new_tbl, &new_tbl->buckets[new_hash], SINGLE_DEPTH_NESTING);
 251
 252         head = rht_ptr(new_tbl->buckets + new_hash, new_tbl, new_hash);
 253
 254         RCU_INIT_POINTER(entry->next, head);
 255
 256         rht_assign_unlock(new_tbl, &new_tbl->buckets[new_hash], entry);
 257
 258         if (pprev)
 259                 rcu_assign_pointer(*pprev, next);
 260         else
 261                 /* Need to preserved the bit lock. */
 262                 rht_assign_locked(bkt, next);
 263
 264 out:
 265         return err;
 266 }
 267
 268 static int rhashtable_rehash_chain(struct rhashtable *ht,
 269                                     unsigned int old_hash)
 270 {
 271         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 272         struct rhash_lock_head __rcu **bkt = rht_bucket_var(old_tbl, old_hash);
 273         int err;
 274
 275         if (!bkt)
 276                 return 0;
 277         rht_lock(old_tbl, bkt);
 278
 279         while (!(err = rhashtable_rehash_one(ht, bkt, old_hash)))
 280                 ;
 281
 282         if (err == -ENOENT)
 283                 err = 0;
 284         rht_unlock(old_tbl, bkt);
 285
 286         return err;
 287 }
 288
 289 static int rhashtable_rehash_attach(struct rhashtable *ht,
 290                                     struct bucket_table *old_tbl,
 291                                     struct bucket_table *new_tbl)
 292 {
 293         /* Make insertions go into the new, empty table right away. Deletions
 294          * and lookups will be attempted in both tables until we synchronize.
 295          * As cmpxchg() provides strong barriers, we do not need
 296          * rcu_assign_pointer().
 297          */
 298
 299         if (cmpxchg(&old_tbl->future_tbl, NULL, new_tbl) != NULL)
 300                 return -EEXIST;
 301
 302         return 0;
 303 }
 304
 305 static int rhashtable_rehash_table(struct rhashtable *ht)
 306 {
 307         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 308         struct bucket_table *new_tbl;
 309         struct rhashtable_walker *walker;
 310         unsigned int old_hash;
 311         int err;
 312
 313         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 314         if (!new_tbl)
 315                 return 0;
 316
 317         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 318                 err = rhashtable_rehash_chain(ht, old_hash);
 319                 if (err)
 320                         return err;
 321                 cond_resched();
 322         }
 323
 324         /* Publish the new table pointer. */
 325         rcu_assign_pointer(ht->tbl, new_tbl);
 326
 327         spin_lock(&ht->lock);
 328         list_for_each_entry(walker, &old_tbl->walkers, list)
 329                 walker->tbl = NULL;
 330
 331         /* Wait for readers. All new readers will see the new
 332          * table, and thus no references to the old table will
 333          * remain.
 334          * We do this inside the locked region so that
 335          * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
 336          * to check if it should not re-link the table.
 337          */
 338         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 339         spin_unlock(&ht->lock);
 340
 341         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 342 }
 343
 344 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 345                                    struct bucket_table *old_tbl,
 346                                    unsigned int size)
 347 {
 348         struct bucket_table *new_tbl;
 349         int err;
 350
 351         ASSERT_RHT_MUTEX(ht);
 352
 353         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 354         if (new_tbl == NULL)
 355                 return -ENOMEM;
 356
 357         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 358         if (err)
 359                 bucket_table_free(new_tbl);
 360
 361         return err;
 362 }
 363
 364 /**
 365  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 366  * @ht:         the hash table to shrink
 367  *
 368  * This function shrinks the hash table to fit, i.e., the smallest
 369  * size would not cause it to expand right away automatically.
 370  *
 371  * The caller must ensure that no concurrent resizing occurs by holding
 372  * ht->mutex.
 373  *
 374  * The caller must ensure that no concurrent table mutations take place.
 375  * It is however valid to have concurrent lookups if they are RCU protected.
 376  *
 377  * It is valid to have concurrent insertions and deletions protected by per
 378  * bucket locks or concurrent RCU protected lookups and traversals.
 379  */
 380 static int rhashtable_shrink(struct rhashtable *ht)
 381 {
 382         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 383         unsigned int nelems = atomic_read(&ht->nelems);
 384         unsigned int size = 0;
 385
 386         if (nelems)
 387                 size = roundup_pow_of_two(nelems * 3 / 2);
 388         if (size < ht->p.min_size)
 389                 size = ht->p.min_size;
 390
 391         if (old_tbl->size <= size)
 392                 return 0;
 393
 394         if (rht_dereference(old_tbl->future_tbl, ht))
 395                 return -EEXIST;
 396
 397         return rhashtable_rehash_alloc(ht, old_tbl, size);
 398 }
 399
 400 static void rht_deferred_worker(struct work_struct *work)
 401 {
 402         struct rhashtable *ht;
 403         struct bucket_table *tbl;
 404         int err = 0;
 405
 406         ht = container_of(work, struct rhashtable, run_work);
 407         mutex_lock(&ht->mutex);
 408
 409         tbl = rht_dereference(ht->tbl, ht);
 410         tbl = rhashtable_last_table(ht, tbl);
 411
 412         if (rht_grow_above_75(ht, tbl))
 413                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 414         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 415                 err = rhashtable_shrink(ht);
 416         else if (tbl->nest)
 417                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 418
 419         if (!err || err == -EEXIST) {
 420                 int nerr;
 421
 422                 nerr = rhashtable_rehash_table(ht);
 423                 err = err ?: nerr;
 424         }
 425
 426         mutex_unlock(&ht->mutex);
 427
 428         if (err)
 429                 schedule_work(&ht->run_work);
 430 }
 431
 432 static int rhashtable_insert_rehash(struct rhashtable *ht,
 433                                     struct bucket_table *tbl)
 434 {
 435         struct bucket_table *old_tbl;
 436         struct bucket_table *new_tbl;
 437         unsigned int size;
 438         int err;
 439
 440         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 441
 442         size = tbl->size;
 443
 444         err = -EBUSY;
 445
 446         if (rht_grow_above_75(ht, tbl))
 447                 size *= 2;
 448         /* Do not schedule more than one rehash */
 449         else if (old_tbl != tbl)
 450                 goto fail;
 451
 452         err = -ENOMEM;
 453
 454         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 455         if (new_tbl == NULL)
 456                 goto fail;
 457
 458         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 459         if (err) {
 460                 bucket_table_free(new_tbl);
 461                 if (err == -EEXIST)
 462                         err = 0;
 463         } else
 464                 schedule_work(&ht->run_work);
 465
 466         return err;
 467
 468 fail:
 469         /* Do not fail the insert if someone else did a rehash. */
 470         if (likely(rcu_access_pointer(tbl->future_tbl)))
 471                 return 0;
 472
 473         /* Schedule async rehash to retry allocation in process context. */
 474         if (err == -ENOMEM)
 475                 schedule_work(&ht->run_work);
 476
 477         return err;
 478 }
 479
 480 static void *rhashtable_lookup_one(struct rhashtable *ht,
 481                                    struct rhash_lock_head __rcu **bkt,
 482                                    struct bucket_table *tbl, unsigned int hash,
 483                                    const void *key, struct rhash_head *obj)
 484 {
 485         struct rhashtable_compare_arg arg = {
 486                 .ht = ht,
 487                 .key = key,
 488         };
 489         struct rhash_head __rcu **pprev = NULL;
 490         struct rhash_head *head;
 491         int elasticity;
 492
 493         elasticity = RHT_ELASTICITY;
 494         rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
 495                 struct rhlist_head *list;
 496                 struct rhlist_head *plist;
 497
 498                 elasticity--;
 499                 if (!key ||
 500                     (ht->p.obj_cmpfn ?
 501                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 502                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 503                         pprev = &head->next;
 504                         continue;
 505                 }
 506
 507                 if (!ht->rhlist)
 508                         return rht_obj(ht, head);
 509
 510                 list = container_of(obj, struct rhlist_head, rhead);
 511                 plist = container_of(head, struct rhlist_head, rhead);
 512
 513                 RCU_INIT_POINTER(list->next, plist);
 514                 head = rht_dereference_bucket(head->next, tbl, hash);
 515                 RCU_INIT_POINTER(list->rhead.next, head);
 516                 if (pprev)
 517                         rcu_assign_pointer(*pprev, obj);
 518                 else
 519                         /* Need to preserve the bit lock */
 520                         rht_assign_locked(bkt, obj);
 521
 522                 return NULL;
 523         }
 524
 525         if (elasticity <= 0)
 526                 return ERR_PTR(-EAGAIN);
 527
 528         return ERR_PTR(-ENOENT);
 529 }
 530
 531 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 532                                                   struct rhash_lock_head __rcu **bkt,
 533                                                   struct bucket_table *tbl,
 534                                                   unsigned int hash,
 535                                                   struct rhash_head *obj,
 536                                                   void *data)
 537 {
 538         struct bucket_table *new_tbl;
 539         struct rhash_head *head;
 540
 541         if (!IS_ERR_OR_NULL(data))
 542                 return ERR_PTR(-EEXIST);
 543
 544         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 545                 return ERR_CAST(data);
 546
 547         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 548         if (new_tbl)
 549                 return new_tbl;
 550
 551         if (PTR_ERR(data) != -ENOENT)
 552                 return ERR_CAST(data);
 553
 554         if (unlikely(rht_grow_above_max(ht, tbl)))
 555                 return ERR_PTR(-E2BIG);
 556
 557         if (unlikely(rht_grow_above_100(ht, tbl)))
 558                 return ERR_PTR(-EAGAIN);
 559
 560         head = rht_ptr(bkt, tbl, hash);
 561
 562         RCU_INIT_POINTER(obj->next, head);
 563         if (ht->rhlist) {
 564                 struct rhlist_head *list;
 565
 566                 list = container_of(obj, struct rhlist_head, rhead);
 567                 RCU_INIT_POINTER(list->next, NULL);
 568         }
 569
 570         /* bkt is always the head of the list, so it holds
 571          * the lock, which we need to preserve
 572          */
 573         rht_assign_locked(bkt, obj);
 574
 575         atomic_inc(&ht->nelems);
 576         if (rht_grow_above_75(ht, tbl))
 577                 schedule_work(&ht->run_work);
 578
 579         return NULL;
 580 }
 581
 582 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 583                                    struct rhash_head *obj)
 584 {
 585         struct bucket_table *new_tbl;
 586         struct bucket_table *tbl;
 587         struct rhash_lock_head __rcu **bkt;
 588         unsigned int hash;
 589         void *data;
 590
 591         new_tbl = rcu_dereference(ht->tbl);
 592
 593         do {
 594                 tbl = new_tbl;
 595                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 596                 if (rcu_access_pointer(tbl->future_tbl))
 597                         /* Failure is OK */
 598                         bkt = rht_bucket_var(tbl, hash);
 599                 else
 600                         bkt = rht_bucket_insert(ht, tbl, hash);
 601                 if (bkt == NULL) {
 602                         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 603                         data = ERR_PTR(-EAGAIN);
 604                 } else {
 605                         rht_lock(tbl, bkt);
 606                         data = rhashtable_lookup_one(ht, bkt, tbl,
 607                                                      hash, key, obj);
 608                         new_tbl = rhashtable_insert_one(ht, bkt, tbl,
 609                                                         hash, obj, data);
 610                         if (PTR_ERR(new_tbl) != -EEXIST)
 611                                 data = ERR_CAST(new_tbl);
 612
 613                         rht_unlock(tbl, bkt);
 614                 }
 615         } while (!IS_ERR_OR_NULL(new_tbl));
 616
 617         if (PTR_ERR(data) == -EAGAIN)
 618                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 619                                -EAGAIN);
 620
 621         return data;
 622 }
 623
 624 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 625                              struct rhash_head *obj)
 626 {
 627         void *data;
 628
 629         do {
 630                 rcu_read_lock();
 631                 data = rhashtable_try_insert(ht, key, obj);
 632                 rcu_read_unlock();
 633         } while (PTR_ERR(data) == -EAGAIN);
 634
 635         return data;
 636 }
 637 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 638
 639 /**
 640  * rhashtable_walk_enter - Initialise an iterator
 641  * @ht:         Table to walk over
 642  * @iter:       Hash table Iterator
 643  *
 644  * This function prepares a hash table walk.
 645  *
 646  * Note that if you restart a walk after rhashtable_walk_stop you
 647  * may see the same object twice.  Also, you may miss objects if
 648  * there are removals in between rhashtable_walk_stop and the next
 649  * call to rhashtable_walk_start.
 650  *
 651  * For a completely stable walk you should construct your own data
 652  * structure outside the hash table.
 653  *
 654  * This function may be called from any process context, including
 655  * non-preemptable context, but cannot be called from softirq or
 656  * hardirq context.
 657  *
 658  * You must call rhashtable_walk_exit after this function returns.
 659  */
 660 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 661 {
 662         iter->ht = ht;
 663         iter->p = NULL;
 664         iter->slot = 0;
 665         iter->skip = 0;
 666         iter->end_of_table = 0;
 667
 668         spin_lock(&ht->lock);
 669         iter->walker.tbl =
 670                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 671         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 672         spin_unlock(&ht->lock);
 673 }
 674 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 675
 676 /**
 677  * rhashtable_walk_exit - Free an iterator
 678  * @iter:       Hash table Iterator
 679  *
 680  * This function frees resources allocated by rhashtable_walk_enter.
 681  */
 682 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 683 {
 684         spin_lock(&iter->ht->lock);
 685         if (iter->walker.tbl)
 686                 list_del(&iter->walker.list);
 687         spin_unlock(&iter->ht->lock);
 688 }
 689 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 690
 691 /**
 692  * rhashtable_walk_start_check - Start a hash table walk
 693  * @iter:       Hash table iterator
 694  *
 695  * Start a hash table walk at the current iterator position.  Note that we take
 696  * the RCU lock in all cases including when we return an error.  So you must
 697  * always call rhashtable_walk_stop to clean up.
 698  *
 699  * Returns zero if successful.
 700  *
 701  * Returns -EAGAIN if resize event occured.  Note that the iterator
 702  * will rewind back to the beginning and you may use it immediately
 703  * by calling rhashtable_walk_next.
 704  *
 705  * rhashtable_walk_start is defined as an inline variant that returns
 706  * void. This is preferred in cases where the caller would ignore
 707  * resize events and always continue.
 708  */
 709 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 710         __acquires(RCU)
 711 {
 712         struct rhashtable *ht = iter->ht;
 713         bool rhlist = ht->rhlist;
 714
 715         rcu_read_lock();
 716
 717         spin_lock(&ht->lock);
 718         if (iter->walker.tbl)
 719                 list_del(&iter->walker.list);
 720         spin_unlock(&ht->lock);
 721
 722         if (iter->end_of_table)
 723                 return 0;
 724         if (!iter->walker.tbl) {
 725                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 726                 iter->slot = 0;
 727                 iter->skip = 0;
 728                 return -EAGAIN;
 729         }
 730
 731         if (iter->p && !rhlist) {
 732                 /*
 733                  * We need to validate that 'p' is still in the table, and
 734                  * if so, update 'skip'
 735                  */
 736                 struct rhash_head *p;
 737                 int skip = 0;
 738                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 739                         skip++;
 740                         if (p == iter->p) {
 741                                 iter->skip = skip;
 742                                 goto found;
 743                         }
 744                 }
 745                 iter->p = NULL;
 746         } else if (iter->p && rhlist) {
 747                 /* Need to validate that 'list' is still in the table, and
 748                  * if so, update 'skip' and 'p'.
 749                  */
 750                 struct rhash_head *p;
 751                 struct rhlist_head *list;
 752                 int skip = 0;
 753                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 754                         for (list = container_of(p, struct rhlist_head, rhead);
 755                              list;
 756                              list = rcu_dereference(list->next)) {
 757                                 skip++;
 758                                 if (list == iter->list) {
 759                                         iter->p = p;
 760                                         iter->skip = skip;
 761                                         goto found;
 762                                 }
 763                         }
 764                 }
 765                 iter->p = NULL;
 766         }
 767 found:
 768         return 0;
 769 }
 770 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 771
 772 /**
 773  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 774  * one in case of a new walk).
 775  *
 776  * @iter:       Hash table iterator
 777  *
 778  * Returns the found object or NULL when the end of the table is reached.
 779  *
 780  * Returns -EAGAIN if resize event occurred.
 781  */
 782 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 783 {
 784         struct bucket_table *tbl = iter->walker.tbl;
 785         struct rhlist_head *list = iter->list;
 786         struct rhashtable *ht = iter->ht;
 787         struct rhash_head *p = iter->p;
 788         bool rhlist = ht->rhlist;
 789
 790         if (!tbl)
 791                 return NULL;
 792
 793         for (; iter->slot < tbl->size; iter->slot++) {
 794                 int skip = iter->skip;
 795
 796                 rht_for_each_rcu(p, tbl, iter->slot) {
 797                         if (rhlist) {
 798                                 list = container_of(p, struct rhlist_head,
 799                                                     rhead);
 800                                 do {
 801                                         if (!skip)
 802                                                 goto next;
 803                                         skip--;
 804                                         list = rcu_dereference(list->next);
 805                                 } while (list);
 806
 807                                 continue;
 808                         }
 809                         if (!skip)
 810                                 break;
 811                         skip--;
 812                 }
 813
 814 next:
 815                 if (!rht_is_a_nulls(p)) {
 816                         iter->skip++;
 817                         iter->p = p;
 818                         iter->list = list;
 819                         return rht_obj(ht, rhlist ? &list->rhead : p);
 820                 }
 821
 822                 iter->skip = 0;
 823         }
 824
 825         iter->p = NULL;
 826
 827         /* Ensure we see any new tables. */
 828         smp_rmb();
 829
 830         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 831         if (iter->walker.tbl) {
 832                 iter->slot = 0;
 833                 iter->skip = 0;
 834                 return ERR_PTR(-EAGAIN);
 835         } else {
 836                 iter->end_of_table = true;
 837         }
 838
 839         return NULL;
 840 }
 841
 842 /**
 843  * rhashtable_walk_next - Return the next object and advance the iterator
 844  * @iter:       Hash table iterator
 845  *
 846  * Note that you must call rhashtable_walk_stop when you are finished
 847  * with the walk.
 848  *
 849  * Returns the next object or NULL when the end of the table is reached.
 850  *
 851  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 852  * will rewind back to the beginning and you may continue to use it.
 853  */
 854 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 855 {
 856         struct rhlist_head *list = iter->list;
 857         struct rhashtable *ht = iter->ht;
 858         struct rhash_head *p = iter->p;
 859         bool rhlist = ht->rhlist;
 860
 861         if (p) {
 862                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 863                         p = rcu_dereference(p->next);
 864                         list = container_of(p, struct rhlist_head, rhead);
 865                 }
 866                 if (!rht_is_a_nulls(p)) {
 867                         iter->skip++;
 868                         iter->p = p;
 869                         iter->list = list;
 870                         return rht_obj(ht, rhlist ? &list->rhead : p);
 871                 }
 872
 873                 /* At the end of this slot, switch to next one and then find
 874                  * next entry from that point.
 875                  */
 876                 iter->skip = 0;
 877                 iter->slot++;
 878         }
 879
 880         return __rhashtable_walk_find_next(iter);
 881 }
 882 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 883
 884 /**
 885  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 886  * @iter:       Hash table iterator
 887  *
 888  * Returns the next object or NULL when the end of the table is reached.
 889  *
 890  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 891  * will rewind back to the beginning and you may continue to use it.
 892  */
 893 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 894 {
 895         struct rhlist_head *list = iter->list;
 896         struct rhashtable *ht = iter->ht;
 897         struct rhash_head *p = iter->p;
 898
 899         if (p)
 900                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 901
 902         /* No object found in current iter, find next one in the table. */
 903
 904         if (iter->skip) {
 905                 /* A nonzero skip value points to the next entry in the table
 906                  * beyond that last one that was found. Decrement skip so
 907                  * we find the current value. __rhashtable_walk_find_next
 908                  * will restore the original value of skip assuming that
 909                  * the table hasn't changed.
 910                  */
 911                 iter->skip--;
 912         }
 913
 914         return __rhashtable_walk_find_next(iter);
 915 }
 916 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 917
 918 /**
 919  * rhashtable_walk_stop - Finish a hash table walk
 920  * @iter:       Hash table iterator
 921  *
 922  * Finish a hash table walk.  Does not reset the iterator to the start of the
 923  * hash table.
 924  */
 925 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 926         __releases(RCU)
 927 {
 928         struct rhashtable *ht;
 929         struct bucket_table *tbl = iter->walker.tbl;
 930
 931         if (!tbl)
 932                 goto out;
 933
 934         ht = iter->ht;
 935
 936         spin_lock(&ht->lock);
 937         if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
 938                 /* This bucket table is being freed, don't re-link it. */
 939                 iter->walker.tbl = NULL;
 940         else
 941                 list_add(&iter->walker.list, &tbl->walkers);
 942         spin_unlock(&ht->lock);
 943
 944 out:
 945         rcu_read_unlock();
 946 }
 947 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 948
 949 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 950 {
 951         size_t retsize;
 952
 953         if (params->nelem_hint)
 954                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 955                               (unsigned long)params->min_size);
 956         else
 957                 retsize = max(HASH_DEFAULT_SIZE,
 958                               (unsigned long)params->min_size);
 959
 960         return retsize;
 961 }
 962
 963 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 964 {
 965         return jhash2(key, length, seed);
 966 }
 967
 968 /**
 969  * rhashtable_init - initialize a new hash table
 970  * @ht:         hash table to be initialized
 971  * @params:     configuration parameters
 972  *
 973  * Initializes a new hash table based on the provided configuration
 974  * parameters. A table can be configured either with a variable or
 975  * fixed length key:
 976  *
 977  * Configuration Example 1: Fixed length keys
 978  * struct test_obj {
 979  *      int                     key;
 980  *      void *                  my_member;
 981  *      struct rhash_head       node;
 982  * };
 983  *
 984  * struct rhashtable_params params = {
 985  *      .head_offset = offsetof(struct test_obj, node),
 986  *      .key_offset = offsetof(struct test_obj, key),
 987  *      .key_len = sizeof(int),
 988  *      .hashfn = jhash,
 989  * };
 990  *
 991  * Configuration Example 2: Variable length keys
 992  * struct test_obj {
 993  *      [...]
 994  *      struct rhash_head       node;
 995  * };
 996  *
 997  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 998  * {
 999  *      struct test_obj *obj = data;
1000  *
1001  *      return [... hash ...];
1002  * }
1003  *
1004  * struct rhashtable_params params = {
1005  *      .head_offset = offsetof(struct test_obj, node),
1006  *      .hashfn = jhash,
1007  *      .obj_hashfn = my_hash_fn,
1008  * };
1009  */
1010 int rhashtable_init(struct rhashtable *ht,
1011                     const struct rhashtable_params *params)
1012 {
1013         struct bucket_table *tbl;
1014         size_t size;
1015
1016         if ((!params->key_len && !params->obj_hashfn) ||
1017             (params->obj_hashfn && !params->obj_cmpfn))
1018                 return -EINVAL;
1019
1020         memset(ht, 0, sizeof(*ht));
1021         mutex_init(&ht->mutex);
1022         spin_lock_init(&ht->lock);
1023         memcpy(&ht->p, params, sizeof(*params));
1024
1025         if (params->min_size)
1026                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1027
1028         /* Cap total entries at 2^31 to avoid nelems overflow. */
1029         ht->max_elems = 1u << 31;
1030
1031         if (params->max_size) {
1032                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1033                 if (ht->p.max_size < ht->max_elems / 2)
1034                         ht->max_elems = ht->p.max_size * 2;
1035         }
1036
1037         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1038
1039         size = rounded_hashtable_size(&ht->p);
1040
1041         ht->key_len = ht->p.key_len;
1042         if (!params->hashfn) {
1043                 ht->p.hashfn = jhash;
1044
1045                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1046                         ht->key_len /= sizeof(u32);
1047                         ht->p.hashfn = rhashtable_jhash2;
1048                 }
1049         }
1050
1051         /*
1052          * This is api initialization and thus we need to guarantee the
1053          * initial rhashtable allocation. Upon failure, retry with the
1054          * smallest possible size with __GFP_NOFAIL semantics.
1055          */
1056         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1057         if (unlikely(tbl == NULL)) {
1058                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1059                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1060         }
1061
1062         atomic_set(&ht->nelems, 0);
1063
1064         RCU_INIT_POINTER(ht->tbl, tbl);
1065
1066         INIT_WORK(&ht->run_work, rht_deferred_worker);
1067
1068         return 0;
1069 }
1070 EXPORT_SYMBOL_GPL(rhashtable_init);
1071
1072 /**
1073  * rhltable_init - initialize a new hash list table
1074  * @hlt:        hash list table to be initialized
1075  * @params:     configuration parameters
1076  *
1077  * Initializes a new hash list table.
1078  *
1079  * See documentation for rhashtable_init.
1080  */
1081 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1082 {
1083         int err;
1084
1085         err = rhashtable_init(&hlt->ht, params);
1086         hlt->ht.rhlist = true;
1087         return err;
1088 }
1089 EXPORT_SYMBOL_GPL(rhltable_init);
1090
1091 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1092                                 void (*free_fn)(void *ptr, void *arg),
1093                                 void *arg)
1094 {
1095         struct rhlist_head *list;
1096
1097         if (!ht->rhlist) {
1098                 free_fn(rht_obj(ht, obj), arg);
1099                 return;
1100         }
1101
1102         list = container_of(obj, struct rhlist_head, rhead);
1103         do {
1104                 obj = &list->rhead;
1105                 list = rht_dereference(list->next, ht);
1106                 free_fn(rht_obj(ht, obj), arg);
1107         } while (list);
1108 }
1109
1110 /**
1111  * rhashtable_free_and_destroy - free elements and destroy hash table
1112  * @ht:         the hash table to destroy
1113  * @free_fn:    callback to release resources of element
1114  * @arg:        pointer passed to free_fn
1115  *
1116  * Stops an eventual async resize. If defined, invokes free_fn for each
1117  * element to releasal resources. Please note that RCU protected
1118  * readers may still be accessing the elements. Releasing of resources
1119  * must occur in a compatible manner. Then frees the bucket array.
1120  *
1121  * This function will eventually sleep to wait for an async resize
1122  * to complete. The caller is responsible that no further write operations
1123  * occurs in parallel.
1124  */
1125 void rhashtable_free_and_destroy(struct rhashtable *ht,
1126                                  void (*free_fn)(void *ptr, void *arg),
1127                                  void *arg)
1128 {
1129         struct bucket_table *tbl, *next_tbl;
1130         unsigned int i;
1131
1132         cancel_work_sync(&ht->run_work);
1133
1134         mutex_lock(&ht->mutex);
1135         tbl = rht_dereference(ht->tbl, ht);
1136 restart:
1137         if (free_fn) {
1138                 for (i = 0; i < tbl->size; i++) {
1139                         struct rhash_head *pos, *next;
1140
1141                         cond_resched();
1142                         for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1143                              next = !rht_is_a_nulls(pos) ?
1144                                         rht_dereference(pos->next, ht) : NULL;
1145                              !rht_is_a_nulls(pos);
1146                              pos = next,
1147                              next = !rht_is_a_nulls(pos) ?
1148                                         rht_dereference(pos->next, ht) : NULL)
1149                                 rhashtable_free_one(ht, pos, free_fn, arg);
1150                 }
1151         }
1152
1153         next_tbl = rht_dereference(tbl->future_tbl, ht);
1154         bucket_table_free(tbl);
1155         if (next_tbl) {
1156                 tbl = next_tbl;
1157                 goto restart;
1158         }
1159         mutex_unlock(&ht->mutex);
1160 }
1161 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1162
1163 void rhashtable_destroy(struct rhashtable *ht)
1164 {
1165         return rhashtable_free_and_destroy(ht, NULL, NULL);
1166 }
1167 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1168
1169 struct rhash_lock_head __rcu **__rht_bucket_nested(const struct bucket_table *tbl,
1170                                                    unsigned int hash)
1171 {
1172         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1173         unsigned int index = hash & ((1 << tbl->nest) - 1);
1174         unsigned int size = tbl->size >> tbl->nest;
1175         unsigned int subhash = hash;
1176         union nested_table *ntbl;
1177
1178         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1179         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1180         subhash >>= tbl->nest;
1181
1182         while (ntbl && size > (1 << shift)) {
1183                 index = subhash & ((1 << shift) - 1);
1184                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1185                                                   tbl, hash);
1186                 size >>= shift;
1187                 subhash >>= shift;
1188         }
1189
1190         if (!ntbl)
1191                 return NULL;
1192
1193         return &ntbl[subhash].bucket;
1194
1195 }
1196 EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1197
1198 struct rhash_lock_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1199                                                  unsigned int hash)
1200 {
1201         static struct rhash_lock_head __rcu *rhnull;
1202
1203         if (!rhnull)
1204                 INIT_RHT_NULLS_HEAD(rhnull);
1205         return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1206 }
1207 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1208
1209 struct rhash_lock_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1210                                                         struct bucket_table *tbl,
1211                                                         unsigned int hash)
1212 {
1213         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1214         unsigned int index = hash & ((1 << tbl->nest) - 1);
1215         unsigned int size = tbl->size >> tbl->nest;
1216         union nested_table *ntbl;
1217
1218         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1219         hash >>= tbl->nest;
1220         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1221                                   size <= (1 << shift));
1222
1223         while (ntbl && size > (1 << shift)) {
1224                 index = hash & ((1 << shift) - 1);
1225                 size >>= shift;
1226                 hash >>= shift;
1227                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1228                                           size <= (1 << shift));
1229         }
1230
1231         if (!ntbl)
1232                 return NULL;
1233
1234         return &ntbl[hash].bucket;
1235
1236 }
1237 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);