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 *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((union nested_table **)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 **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 **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((struct bucket_table **)&old_tbl->future_tbl, NULL,
 300                     new_tbl) != NULL)
 301                 return -EEXIST;
 302
 303         return 0;
 304 }
 305
 306 static int rhashtable_rehash_table(struct rhashtable *ht)
 307 {
 308         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 309         struct bucket_table *new_tbl;
 310         struct rhashtable_walker *walker;
 311         unsigned int old_hash;
 312         int err;
 313
 314         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 315         if (!new_tbl)
 316                 return 0;
 317
 318         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 319                 err = rhashtable_rehash_chain(ht, old_hash);
 320                 if (err)
 321                         return err;
 322                 cond_resched();
 323         }
 324
 325         /* Publish the new table pointer. */
 326         rcu_assign_pointer(ht->tbl, new_tbl);
 327
 328         spin_lock(&ht->lock);
 329         list_for_each_entry(walker, &old_tbl->walkers, list)
 330                 walker->tbl = NULL;
 331
 332         /* Wait for readers. All new readers will see the new
 333          * table, and thus no references to the old table will
 334          * remain.
 335          * We do this inside the locked region so that
 336          * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
 337          * to check if it should not re-link the table.
 338          */
 339         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 340         spin_unlock(&ht->lock);
 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 || err == -EEXIST) {
 421                 int nerr;
 422
 423                 nerr = rhashtable_rehash_table(ht);
 424                 err = err ?: nerr;
 425         }
 426
 427         mutex_unlock(&ht->mutex);
 428
 429         if (err)
 430                 schedule_work(&ht->run_work);
 431 }
 432
 433 static int rhashtable_insert_rehash(struct rhashtable *ht,
 434                                     struct bucket_table *tbl)
 435 {
 436         struct bucket_table *old_tbl;
 437         struct bucket_table *new_tbl;
 438         unsigned int size;
 439         int err;
 440
 441         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 442
 443         size = tbl->size;
 444
 445         err = -EBUSY;
 446
 447         if (rht_grow_above_75(ht, tbl))
 448                 size *= 2;
 449         /* Do not schedule more than one rehash */
 450         else if (old_tbl != tbl)
 451                 goto fail;
 452
 453         err = -ENOMEM;
 454
 455         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 456         if (new_tbl == NULL)
 457                 goto fail;
 458
 459         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 460         if (err) {
 461                 bucket_table_free(new_tbl);
 462                 if (err == -EEXIST)
 463                         err = 0;
 464         } else
 465                 schedule_work(&ht->run_work);
 466
 467         return err;
 468
 469 fail:
 470         /* Do not fail the insert if someone else did a rehash. */
 471         if (likely(rcu_access_pointer(tbl->future_tbl)))
 472                 return 0;
 473
 474         /* Schedule async rehash to retry allocation in process context. */
 475         if (err == -ENOMEM)
 476                 schedule_work(&ht->run_work);
 477
 478         return err;
 479 }
 480
 481 static void *rhashtable_lookup_one(struct rhashtable *ht,
 482                                    struct rhash_lock_head **bkt,
 483                                    struct bucket_table *tbl, unsigned int hash,
 484                                    const void *key, struct rhash_head *obj)
 485 {
 486         struct rhashtable_compare_arg arg = {
 487                 .ht = ht,
 488                 .key = key,
 489         };
 490         struct rhash_head __rcu **pprev = NULL;
 491         struct rhash_head *head;
 492         int elasticity;
 493
 494         elasticity = RHT_ELASTICITY;
 495         rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
 496                 struct rhlist_head *list;
 497                 struct rhlist_head *plist;
 498
 499                 elasticity--;
 500                 if (!key ||
 501                     (ht->p.obj_cmpfn ?
 502                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 503                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 504                         pprev = &head->next;
 505                         continue;
 506                 }
 507
 508                 if (!ht->rhlist)
 509                         return rht_obj(ht, head);
 510
 511                 list = container_of(obj, struct rhlist_head, rhead);
 512                 plist = container_of(head, struct rhlist_head, rhead);
 513
 514                 RCU_INIT_POINTER(list->next, plist);
 515                 head = rht_dereference_bucket(head->next, tbl, hash);
 516                 RCU_INIT_POINTER(list->rhead.next, head);
 517                 if (pprev)
 518                         rcu_assign_pointer(*pprev, obj);
 519                 else
 520                         /* Need to preserve the bit lock */
 521                         rht_assign_locked(bkt, obj);
 522
 523                 return NULL;
 524         }
 525
 526         if (elasticity <= 0)
 527                 return ERR_PTR(-EAGAIN);
 528
 529         return ERR_PTR(-ENOENT);
 530 }
 531
 532 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 533                                                   struct rhash_lock_head **bkt,
 534                                                   struct bucket_table *tbl,
 535                                                   unsigned int hash,
 536                                                   struct rhash_head *obj,
 537                                                   void *data)
 538 {
 539         struct bucket_table *new_tbl;
 540         struct rhash_head *head;
 541
 542         if (!IS_ERR_OR_NULL(data))
 543                 return ERR_PTR(-EEXIST);
 544
 545         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 546                 return ERR_CAST(data);
 547
 548         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 549         if (new_tbl)
 550                 return new_tbl;
 551
 552         if (PTR_ERR(data) != -ENOENT)
 553                 return ERR_CAST(data);
 554
 555         if (unlikely(rht_grow_above_max(ht, tbl)))
 556                 return ERR_PTR(-E2BIG);
 557
 558         if (unlikely(rht_grow_above_100(ht, tbl)))
 559                 return ERR_PTR(-EAGAIN);
 560
 561         head = rht_ptr(bkt, tbl, hash);
 562
 563         RCU_INIT_POINTER(obj->next, head);
 564         if (ht->rhlist) {
 565                 struct rhlist_head *list;
 566
 567                 list = container_of(obj, struct rhlist_head, rhead);
 568                 RCU_INIT_POINTER(list->next, NULL);
 569         }
 570
 571         /* bkt is always the head of the list, so it holds
 572          * the lock, which we need to preserve
 573          */
 574         rht_assign_locked(bkt, obj);
 575
 576         atomic_inc(&ht->nelems);
 577         if (rht_grow_above_75(ht, tbl))
 578                 schedule_work(&ht->run_work);
 579
 580         return NULL;
 581 }
 582
 583 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 584                                    struct rhash_head *obj)
 585 {
 586         struct bucket_table *new_tbl;
 587         struct bucket_table *tbl;
 588         struct rhash_lock_head **bkt;
 589         unsigned int hash;
 590         void *data;
 591
 592         new_tbl = rcu_dereference(ht->tbl);
 593
 594         do {
 595                 tbl = new_tbl;
 596                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 597                 if (rcu_access_pointer(tbl->future_tbl))
 598                         /* Failure is OK */
 599                         bkt = rht_bucket_var(tbl, hash);
 600                 else
 601                         bkt = rht_bucket_insert(ht, tbl, hash);
 602                 if (bkt == NULL) {
 603                         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 604                         data = ERR_PTR(-EAGAIN);
 605                 } else {
 606                         rht_lock(tbl, bkt);
 607                         data = rhashtable_lookup_one(ht, bkt, tbl,
 608                                                      hash, key, obj);
 609                         new_tbl = rhashtable_insert_one(ht, bkt, tbl,
 610                                                         hash, obj, data);
 611                         if (PTR_ERR(new_tbl) != -EEXIST)
 612                                 data = ERR_CAST(new_tbl);
 613
 614                         rht_unlock(tbl, bkt);
 615                 }
 616         } while (!IS_ERR_OR_NULL(new_tbl));
 617
 618         if (PTR_ERR(data) == -EAGAIN)
 619                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 620                                -EAGAIN);
 621
 622         return data;
 623 }
 624
 625 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 626                              struct rhash_head *obj)
 627 {
 628         void *data;
 629
 630         do {
 631                 rcu_read_lock();
 632                 data = rhashtable_try_insert(ht, key, obj);
 633                 rcu_read_unlock();
 634         } while (PTR_ERR(data) == -EAGAIN);
 635
 636         return data;
 637 }
 638 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 639
 640 /**
 641  * rhashtable_walk_enter - Initialise an iterator
 642  * @ht:         Table to walk over
 643  * @iter:       Hash table Iterator
 644  *
 645  * This function prepares a hash table walk.
 646  *
 647  * Note that if you restart a walk after rhashtable_walk_stop you
 648  * may see the same object twice.  Also, you may miss objects if
 649  * there are removals in between rhashtable_walk_stop and the next
 650  * call to rhashtable_walk_start.
 651  *
 652  * For a completely stable walk you should construct your own data
 653  * structure outside the hash table.
 654  *
 655  * This function may be called from any process context, including
 656  * non-preemptable context, but cannot be called from softirq or
 657  * hardirq context.
 658  *
 659  * You must call rhashtable_walk_exit after this function returns.
 660  */
 661 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 662 {
 663         iter->ht = ht;
 664         iter->p = NULL;
 665         iter->slot = 0;
 666         iter->skip = 0;
 667         iter->end_of_table = 0;
 668
 669         spin_lock(&ht->lock);
 670         iter->walker.tbl =
 671                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 672         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 673         spin_unlock(&ht->lock);
 674 }
 675 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 676
 677 /**
 678  * rhashtable_walk_exit - Free an iterator
 679  * @iter:       Hash table Iterator
 680  *
 681  * This function frees resources allocated by rhashtable_walk_enter.
 682  */
 683 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 684 {
 685         spin_lock(&iter->ht->lock);
 686         if (iter->walker.tbl)
 687                 list_del(&iter->walker.list);
 688         spin_unlock(&iter->ht->lock);
 689 }
 690 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 691
 692 /**
 693  * rhashtable_walk_start_check - Start a hash table walk
 694  * @iter:       Hash table iterator
 695  *
 696  * Start a hash table walk at the current iterator position.  Note that we take
 697  * the RCU lock in all cases including when we return an error.  So you must
 698  * always call rhashtable_walk_stop to clean up.
 699  *
 700  * Returns zero if successful.
 701  *
 702  * Returns -EAGAIN if resize event occured.  Note that the iterator
 703  * will rewind back to the beginning and you may use it immediately
 704  * by calling rhashtable_walk_next.
 705  *
 706  * rhashtable_walk_start is defined as an inline variant that returns
 707  * void. This is preferred in cases where the caller would ignore
 708  * resize events and always continue.
 709  */
 710 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 711         __acquires(RCU)
 712 {
 713         struct rhashtable *ht = iter->ht;
 714         bool rhlist = ht->rhlist;
 715
 716         rcu_read_lock();
 717
 718         spin_lock(&ht->lock);
 719         if (iter->walker.tbl)
 720                 list_del(&iter->walker.list);
 721         spin_unlock(&ht->lock);
 722
 723         if (iter->end_of_table)
 724                 return 0;
 725         if (!iter->walker.tbl) {
 726                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 727                 iter->slot = 0;
 728                 iter->skip = 0;
 729                 return -EAGAIN;
 730         }
 731
 732         if (iter->p && !rhlist) {
 733                 /*
 734                  * We need to validate that 'p' is still in the table, and
 735                  * if so, update 'skip'
 736                  */
 737                 struct rhash_head *p;
 738                 int skip = 0;
 739                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 740                         skip++;
 741                         if (p == iter->p) {
 742                                 iter->skip = skip;
 743                                 goto found;
 744                         }
 745                 }
 746                 iter->p = NULL;
 747         } else if (iter->p && rhlist) {
 748                 /* Need to validate that 'list' is still in the table, and
 749                  * if so, update 'skip' and 'p'.
 750                  */
 751                 struct rhash_head *p;
 752                 struct rhlist_head *list;
 753                 int skip = 0;
 754                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 755                         for (list = container_of(p, struct rhlist_head, rhead);
 756                              list;
 757                              list = rcu_dereference(list->next)) {
 758                                 skip++;
 759                                 if (list == iter->list) {
 760                                         iter->p = p;
 761                                         iter->skip = skip;
 762                                         goto found;
 763                                 }
 764                         }
 765                 }
 766                 iter->p = NULL;
 767         }
 768 found:
 769         return 0;
 770 }
 771 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 772
 773 /**
 774  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 775  * one in case of a new walk).
 776  *
 777  * @iter:       Hash table iterator
 778  *
 779  * Returns the found object or NULL when the end of the table is reached.
 780  *
 781  * Returns -EAGAIN if resize event occurred.
 782  */
 783 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 784 {
 785         struct bucket_table *tbl = iter->walker.tbl;
 786         struct rhlist_head *list = iter->list;
 787         struct rhashtable *ht = iter->ht;
 788         struct rhash_head *p = iter->p;
 789         bool rhlist = ht->rhlist;
 790
 791         if (!tbl)
 792                 return NULL;
 793
 794         for (; iter->slot < tbl->size; iter->slot++) {
 795                 int skip = iter->skip;
 796
 797                 rht_for_each_rcu(p, tbl, iter->slot) {
 798                         if (rhlist) {
 799                                 list = container_of(p, struct rhlist_head,
 800                                                     rhead);
 801                                 do {
 802                                         if (!skip)
 803                                                 goto next;
 804                                         skip--;
 805                                         list = rcu_dereference(list->next);
 806                                 } while (list);
 807
 808                                 continue;
 809                         }
 810                         if (!skip)
 811                                 break;
 812                         skip--;
 813                 }
 814
 815 next:
 816                 if (!rht_is_a_nulls(p)) {
 817                         iter->skip++;
 818                         iter->p = p;
 819                         iter->list = list;
 820                         return rht_obj(ht, rhlist ? &list->rhead : p);
 821                 }
 822
 823                 iter->skip = 0;
 824         }
 825
 826         iter->p = NULL;
 827
 828         /* Ensure we see any new tables. */
 829         smp_rmb();
 830
 831         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 832         if (iter->walker.tbl) {
 833                 iter->slot = 0;
 834                 iter->skip = 0;
 835                 return ERR_PTR(-EAGAIN);
 836         } else {
 837                 iter->end_of_table = true;
 838         }
 839
 840         return NULL;
 841 }
 842
 843 /**
 844  * rhashtable_walk_next - Return the next object and advance the iterator
 845  * @iter:       Hash table iterator
 846  *
 847  * Note that you must call rhashtable_walk_stop when you are finished
 848  * with the walk.
 849  *
 850  * Returns the next object or NULL when the end of the table is reached.
 851  *
 852  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 853  * will rewind back to the beginning and you may continue to use it.
 854  */
 855 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 856 {
 857         struct rhlist_head *list = iter->list;
 858         struct rhashtable *ht = iter->ht;
 859         struct rhash_head *p = iter->p;
 860         bool rhlist = ht->rhlist;
 861
 862         if (p) {
 863                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 864                         p = rcu_dereference(p->next);
 865                         list = container_of(p, struct rhlist_head, rhead);
 866                 }
 867                 if (!rht_is_a_nulls(p)) {
 868                         iter->skip++;
 869                         iter->p = p;
 870                         iter->list = list;
 871                         return rht_obj(ht, rhlist ? &list->rhead : p);
 872                 }
 873
 874                 /* At the end of this slot, switch to next one and then find
 875                  * next entry from that point.
 876                  */
 877                 iter->skip = 0;
 878                 iter->slot++;
 879         }
 880
 881         return __rhashtable_walk_find_next(iter);
 882 }
 883 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 884
 885 /**
 886  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 887  * @iter:       Hash table iterator
 888  *
 889  * Returns the next object or NULL when the end of the table is reached.
 890  *
 891  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 892  * will rewind back to the beginning and you may continue to use it.
 893  */
 894 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 895 {
 896         struct rhlist_head *list = iter->list;
 897         struct rhashtable *ht = iter->ht;
 898         struct rhash_head *p = iter->p;
 899
 900         if (p)
 901                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 902
 903         /* No object found in current iter, find next one in the table. */
 904
 905         if (iter->skip) {
 906                 /* A nonzero skip value points to the next entry in the table
 907                  * beyond that last one that was found. Decrement skip so
 908                  * we find the current value. __rhashtable_walk_find_next
 909                  * will restore the original value of skip assuming that
 910                  * the table hasn't changed.
 911                  */
 912                 iter->skip--;
 913         }
 914
 915         return __rhashtable_walk_find_next(iter);
 916 }
 917 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 918
 919 /**
 920  * rhashtable_walk_stop - Finish a hash table walk
 921  * @iter:       Hash table iterator
 922  *
 923  * Finish a hash table walk.  Does not reset the iterator to the start of the
 924  * hash table.
 925  */
 926 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 927         __releases(RCU)
 928 {
 929         struct rhashtable *ht;
 930         struct bucket_table *tbl = iter->walker.tbl;
 931
 932         if (!tbl)
 933                 goto out;
 934
 935         ht = iter->ht;
 936
 937         spin_lock(&ht->lock);
 938         if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
 939                 /* This bucket table is being freed, don't re-link it. */
 940                 iter->walker.tbl = NULL;
 941         else
 942                 list_add(&iter->walker.list, &tbl->walkers);
 943         spin_unlock(&ht->lock);
 944
 945 out:
 946         rcu_read_unlock();
 947 }
 948 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 949
 950 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 951 {
 952         size_t retsize;
 953
 954         if (params->nelem_hint)
 955                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 956                               (unsigned long)params->min_size);
 957         else
 958                 retsize = max(HASH_DEFAULT_SIZE,
 959                               (unsigned long)params->min_size);
 960
 961         return retsize;
 962 }
 963
 964 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 965 {
 966         return jhash2(key, length, seed);
 967 }
 968
 969 /**
 970  * rhashtable_init - initialize a new hash table
 971  * @ht:         hash table to be initialized
 972  * @params:     configuration parameters
 973  *
 974  * Initializes a new hash table based on the provided configuration
 975  * parameters. A table can be configured either with a variable or
 976  * fixed length key:
 977  *
 978  * Configuration Example 1: Fixed length keys
 979  * struct test_obj {
 980  *      int                     key;
 981  *      void *                  my_member;
 982  *      struct rhash_head       node;
 983  * };
 984  *
 985  * struct rhashtable_params params = {
 986  *      .head_offset = offsetof(struct test_obj, node),
 987  *      .key_offset = offsetof(struct test_obj, key),
 988  *      .key_len = sizeof(int),
 989  *      .hashfn = jhash,
 990  * };
 991  *
 992  * Configuration Example 2: Variable length keys
 993  * struct test_obj {
 994  *      [...]
 995  *      struct rhash_head       node;
 996  * };
 997  *
 998  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 999  * {
1000  *      struct test_obj *obj = data;
1001  *
1002  *      return [... hash ...];
1003  * }
1004  *
1005  * struct rhashtable_params params = {
1006  *      .head_offset = offsetof(struct test_obj, node),
1007  *      .hashfn = jhash,
1008  *      .obj_hashfn = my_hash_fn,
1009  * };
1010  */
1011 int rhashtable_init(struct rhashtable *ht,
1012                     const struct rhashtable_params *params)
1013 {
1014         struct bucket_table *tbl;
1015         size_t size;
1016
1017         if ((!params->key_len && !params->obj_hashfn) ||
1018             (params->obj_hashfn && !params->obj_cmpfn))
1019                 return -EINVAL;
1020
1021         memset(ht, 0, sizeof(*ht));
1022         mutex_init(&ht->mutex);
1023         spin_lock_init(&ht->lock);
1024         memcpy(&ht->p, params, sizeof(*params));
1025
1026         if (params->min_size)
1027                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1028
1029         /* Cap total entries at 2^31 to avoid nelems overflow. */
1030         ht->max_elems = 1u << 31;
1031
1032         if (params->max_size) {
1033                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1034                 if (ht->p.max_size < ht->max_elems / 2)
1035                         ht->max_elems = ht->p.max_size * 2;
1036         }
1037
1038         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1039
1040         size = rounded_hashtable_size(&ht->p);
1041
1042         ht->key_len = ht->p.key_len;
1043         if (!params->hashfn) {
1044                 ht->p.hashfn = jhash;
1045
1046                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1047                         ht->key_len /= sizeof(u32);
1048                         ht->p.hashfn = rhashtable_jhash2;
1049                 }
1050         }
1051
1052         /*
1053          * This is api initialization and thus we need to guarantee the
1054          * initial rhashtable allocation. Upon failure, retry with the
1055          * smallest possible size with __GFP_NOFAIL semantics.
1056          */
1057         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1058         if (unlikely(tbl == NULL)) {
1059                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1060                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1061         }
1062
1063         atomic_set(&ht->nelems, 0);
1064
1065         RCU_INIT_POINTER(ht->tbl, tbl);
1066
1067         INIT_WORK(&ht->run_work, rht_deferred_worker);
1068
1069         return 0;
1070 }
1071 EXPORT_SYMBOL_GPL(rhashtable_init);
1072
1073 /**
1074  * rhltable_init - initialize a new hash list table
1075  * @hlt:        hash list table to be initialized
1076  * @params:     configuration parameters
1077  *
1078  * Initializes a new hash list table.
1079  *
1080  * See documentation for rhashtable_init.
1081  */
1082 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1083 {
1084         int err;
1085
1086         err = rhashtable_init(&hlt->ht, params);
1087         hlt->ht.rhlist = true;
1088         return err;
1089 }
1090 EXPORT_SYMBOL_GPL(rhltable_init);
1091
1092 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1093                                 void (*free_fn)(void *ptr, void *arg),
1094                                 void *arg)
1095 {
1096         struct rhlist_head *list;
1097
1098         if (!ht->rhlist) {
1099                 free_fn(rht_obj(ht, obj), arg);
1100                 return;
1101         }
1102
1103         list = container_of(obj, struct rhlist_head, rhead);
1104         do {
1105                 obj = &list->rhead;
1106                 list = rht_dereference(list->next, ht);
1107                 free_fn(rht_obj(ht, obj), arg);
1108         } while (list);
1109 }
1110
1111 /**
1112  * rhashtable_free_and_destroy - free elements and destroy hash table
1113  * @ht:         the hash table to destroy
1114  * @free_fn:    callback to release resources of element
1115  * @arg:        pointer passed to free_fn
1116  *
1117  * Stops an eventual async resize. If defined, invokes free_fn for each
1118  * element to releasal resources. Please note that RCU protected
1119  * readers may still be accessing the elements. Releasing of resources
1120  * must occur in a compatible manner. Then frees the bucket array.
1121  *
1122  * This function will eventually sleep to wait for an async resize
1123  * to complete. The caller is responsible that no further write operations
1124  * occurs in parallel.
1125  */
1126 void rhashtable_free_and_destroy(struct rhashtable *ht,
1127                                  void (*free_fn)(void *ptr, void *arg),
1128                                  void *arg)
1129 {
1130         struct bucket_table *tbl, *next_tbl;
1131         unsigned int i;
1132
1133         cancel_work_sync(&ht->run_work);
1134
1135         mutex_lock(&ht->mutex);
1136         tbl = rht_dereference(ht->tbl, ht);
1137 restart:
1138         if (free_fn) {
1139                 for (i = 0; i < tbl->size; i++) {
1140                         struct rhash_head *pos, *next;
1141
1142                         cond_resched();
1143                         for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1144                              next = !rht_is_a_nulls(pos) ?
1145                                         rht_dereference(pos->next, ht) : NULL;
1146                              !rht_is_a_nulls(pos);
1147                              pos = next,
1148                              next = !rht_is_a_nulls(pos) ?
1149                                         rht_dereference(pos->next, ht) : NULL)
1150                                 rhashtable_free_one(ht, pos, free_fn, arg);
1151                 }
1152         }
1153
1154         next_tbl = rht_dereference(tbl->future_tbl, ht);
1155         bucket_table_free(tbl);
1156         if (next_tbl) {
1157                 tbl = next_tbl;
1158                 goto restart;
1159         }
1160         mutex_unlock(&ht->mutex);
1161 }
1162 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1163
1164 void rhashtable_destroy(struct rhashtable *ht)
1165 {
1166         return rhashtable_free_and_destroy(ht, NULL, NULL);
1167 }
1168 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1169
1170 struct rhash_lock_head **__rht_bucket_nested(const struct bucket_table *tbl,
1171                                              unsigned int hash)
1172 {
1173         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1174         unsigned int index = hash & ((1 << tbl->nest) - 1);
1175         unsigned int size = tbl->size >> tbl->nest;
1176         unsigned int subhash = hash;
1177         union nested_table *ntbl;
1178
1179         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1180         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1181         subhash >>= tbl->nest;
1182
1183         while (ntbl && size > (1 << shift)) {
1184                 index = subhash & ((1 << shift) - 1);
1185                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1186                                                   tbl, hash);
1187                 size >>= shift;
1188                 subhash >>= shift;
1189         }
1190
1191         if (!ntbl)
1192                 return NULL;
1193
1194         return &ntbl[subhash].bucket;
1195
1196 }
1197 EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1198
1199 struct rhash_lock_head **rht_bucket_nested(const struct bucket_table *tbl,
1200                                            unsigned int hash)
1201 {
1202         static struct rhash_lock_head *rhnull;
1203
1204         if (!rhnull)
1205                 INIT_RHT_NULLS_HEAD(rhnull);
1206         return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1207 }
1208 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1209
1210 struct rhash_lock_head **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);