lib/rhashtable.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Resizable, Scalable, Concurrent Hash Table
   4  *
   5  * Copyright (c) 2015 Herbert Xu <[email protected]>
   6  * Copyright (c) 2014-2015 Thomas Graf <[email protected]>
   7  * Copyright (c) 2008-2014 Patrick McHardy <[email protected]>
   8  *
   9  * Code partially derived from nft_hash
  10  * Rewritten with rehash code from br_multicast plus single list
  11  * pointer as suggested by Josh Triplett
  12  */
  13
  14 #include <linux/atomic.h>
  15 #include <linux/kernel.h>
  16 #include <linux/init.h>
  17 #include <linux/log2.h>
  18 #include <linux/sched.h>
  19 #include <linux/rculist.h>
  20 #include <linux/slab.h>
  21 #include <linux/vmalloc.h>
  22 #include <linux/mm.h>
  23 #include <linux/jhash.h>
  24 #include <linux/random.h>
  25 #include <linux/rhashtable.h>
  26 #include <linux/err.h>
  27 #include <linux/export.h>
  28
  29 #define HASH_DEFAULT_SIZE       64UL
  30 #define HASH_MIN_SIZE           4U
  31
  32 union nested_table {
  33         union nested_table __rcu *table;
  34         struct rhash_lock_head *bucket;
  35 };
  36
  37 static u32 head_hashfn(struct rhashtable *ht,
  38                        const struct bucket_table *tbl,
  39                        const struct rhash_head *he)
  40 {
  41         return rht_head_hashfn(ht, tbl, he, ht->p);
  42 }
  43
  44 #ifdef CONFIG_PROVE_LOCKING
  45 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  46
  47 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  48 {
  49         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  50 }
  51 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  52
  53 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  54 {
  55         if (!debug_locks)
  56                 return 1;
  57         if (unlikely(tbl->nest))
  58                 return 1;
  59         return bit_spin_is_locked(0, (unsigned long *)&tbl->buckets[hash]);
  60 }
  61 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  62 #else
  63 #define ASSERT_RHT_MUTEX(HT)
  64 #endif
  65
  66 static inline union nested_table *nested_table_top(
  67         const struct bucket_table *tbl)
  68 {
  69         /* The top-level bucket entry does not need RCU protection
  70          * because it's set at the same time as tbl->nest.
  71          */
  72         return (void *)rcu_dereference_protected(tbl->buckets[0], 1);
  73 }
  74
  75 static void nested_table_free(union nested_table *ntbl, unsigned int size)
  76 {
  77         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
  78         const unsigned int len = 1 << shift;
  79         unsigned int i;
  80
  81         ntbl = rcu_dereference_protected(ntbl->table, 1);
  82         if (!ntbl)
  83                 return;
  84
  85         if (size > len) {
  86                 size >>= shift;
  87                 for (i = 0; i < len; i++)
  88                         nested_table_free(ntbl + i, size);
  89         }
  90
  91         kfree(ntbl);
  92 }
  93
  94 static void nested_bucket_table_free(const struct bucket_table *tbl)
  95 {
  96         unsigned int size = tbl->size >> tbl->nest;
  97         unsigned int len = 1 << tbl->nest;
  98         union nested_table *ntbl;
  99         unsigned int i;
 100
 101         ntbl = nested_table_top(tbl);
 102
 103         for (i = 0; i < len; i++)
 104                 nested_table_free(ntbl + i, size);
 105
 106         kfree(ntbl);
 107 }
 108
 109 static void bucket_table_free(const struct bucket_table *tbl)
 110 {
 111         if (tbl->nest)
 112                 nested_bucket_table_free(tbl);
 113
 114         kvfree(tbl);
 115 }
 116
 117 static void bucket_table_free_rcu(struct rcu_head *head)
 118 {
 119         bucket_table_free(container_of(head, struct bucket_table, rcu));
 120 }
 121
 122 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 123                                               union nested_table __rcu **prev,
 124                                               bool leaf)
 125 {
 126         union nested_table *ntbl;
 127         int i;
 128
 129         ntbl = rcu_dereference(*prev);
 130         if (ntbl)
 131                 return ntbl;
 132
 133         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 134
 135         if (ntbl && leaf) {
 136                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0]); i++)
 137                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket);
 138         }
 139
 140         if (cmpxchg((union nested_table **)prev, NULL, ntbl) == NULL)
 141                 return ntbl;
 142         /* Raced with another thread. */
 143         kfree(ntbl);
 144         return rcu_dereference(*prev);
 145 }
 146
 147 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 148                                                       size_t nbuckets,
 149                                                       gfp_t gfp)
 150 {
 151         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 152         struct bucket_table *tbl;
 153         size_t size;
 154
 155         if (nbuckets < (1 << (shift + 1)))
 156                 return NULL;
 157
 158         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 159
 160         tbl = kzalloc(size, gfp);
 161         if (!tbl)
 162                 return NULL;
 163
 164         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 165                                 false)) {
 166                 kfree(tbl);
 167                 return NULL;
 168         }
 169
 170         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 171
 172         return tbl;
 173 }
 174
 175 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 176                                                size_t nbuckets,
 177                                                gfp_t gfp)
 178 {
 179         struct bucket_table *tbl = NULL;
 180         size_t size;
 181         int i;
 182         static struct lock_class_key __key;
 183
 184         tbl = kvzalloc(struct_size(tbl, buckets, nbuckets), gfp);
 185
 186         size = nbuckets;
 187
 188         if (tbl == NULL && (gfp & ~__GFP_NOFAIL) != GFP_KERNEL) {
 189                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 190                 nbuckets = 0;
 191         }
 192
 193         if (tbl == NULL)
 194                 return NULL;
 195
 196         lockdep_init_map(&tbl->dep_map, "rhashtable_bucket", &__key, 0);
 197
 198         tbl->size = size;
 199
 200         rcu_head_init(&tbl->rcu);
 201         INIT_LIST_HEAD(&tbl->walkers);
 202
 203         tbl->hash_rnd = get_random_u32();
 204
 205         for (i = 0; i < nbuckets; i++)
 206                 INIT_RHT_NULLS_HEAD(tbl->buckets[i]);
 207
 208         return tbl;
 209 }
 210
 211 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 212                                                   struct bucket_table *tbl)
 213 {
 214         struct bucket_table *new_tbl;
 215
 216         do {
 217                 new_tbl = tbl;
 218                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 219         } while (tbl);
 220
 221         return new_tbl;
 222 }
 223
 224 static int rhashtable_rehash_one(struct rhashtable *ht,
 225                                  struct rhash_lock_head **bkt,
 226                                  unsigned int old_hash)
 227 {
 228         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 229         struct bucket_table *new_tbl = rhashtable_last_table(ht, old_tbl);
 230         int err = -EAGAIN;
 231         struct rhash_head *head, *next, *entry;
 232         struct rhash_head __rcu **pprev = NULL;
 233         unsigned int new_hash;
 234
 235         if (new_tbl->nest)
 236                 goto out;
 237
 238         err = -ENOENT;
 239
 240         rht_for_each_from(entry, rht_ptr(bkt, old_tbl, old_hash),
 241                           old_tbl, old_hash) {
 242                 err = 0;
 243                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 244
 245                 if (rht_is_a_nulls(next))
 246                         break;
 247
 248                 pprev = &entry->next;
 249         }
 250
 251         if (err)
 252                 goto out;
 253
 254         new_hash = head_hashfn(ht, new_tbl, entry);
 255
 256         rht_lock_nested(new_tbl, &new_tbl->buckets[new_hash], SINGLE_DEPTH_NESTING);
 257
 258         head = rht_ptr(new_tbl->buckets + new_hash, new_tbl, new_hash);
 259
 260         RCU_INIT_POINTER(entry->next, head);
 261
 262         rht_assign_unlock(new_tbl, &new_tbl->buckets[new_hash], entry);
 263
 264         if (pprev)
 265                 rcu_assign_pointer(*pprev, next);
 266         else
 267                 /* Need to preserved the bit lock. */
 268                 rht_assign_locked(bkt, next);
 269
 270 out:
 271         return err;
 272 }
 273
 274 static int rhashtable_rehash_chain(struct rhashtable *ht,
 275                                     unsigned int old_hash)
 276 {
 277         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 278         struct rhash_lock_head **bkt = rht_bucket_var(old_tbl, old_hash);
 279         int err;
 280
 281         if (!bkt)
 282                 return 0;
 283         rht_lock(old_tbl, bkt);
 284
 285         while (!(err = rhashtable_rehash_one(ht, bkt, old_hash)))
 286                 ;
 287
 288         if (err == -ENOENT)
 289                 err = 0;
 290         rht_unlock(old_tbl, bkt);
 291
 292         return err;
 293 }
 294
 295 static int rhashtable_rehash_attach(struct rhashtable *ht,
 296                                     struct bucket_table *old_tbl,
 297                                     struct bucket_table *new_tbl)
 298 {
 299         /* Make insertions go into the new, empty table right away. Deletions
 300          * and lookups will be attempted in both tables until we synchronize.
 301          * As cmpxchg() provides strong barriers, we do not need
 302          * rcu_assign_pointer().
 303          */
 304
 305         if (cmpxchg((struct bucket_table **)&old_tbl->future_tbl, NULL,
 306                     new_tbl) != NULL)
 307                 return -EEXIST;
 308
 309         return 0;
 310 }
 311
 312 static int rhashtable_rehash_table(struct rhashtable *ht)
 313 {
 314         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 315         struct bucket_table *new_tbl;
 316         struct rhashtable_walker *walker;
 317         unsigned int old_hash;
 318         int err;
 319
 320         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 321         if (!new_tbl)
 322                 return 0;
 323
 324         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 325                 err = rhashtable_rehash_chain(ht, old_hash);
 326                 if (err)
 327                         return err;
 328                 cond_resched();
 329         }
 330
 331         /* Publish the new table pointer. */
 332         rcu_assign_pointer(ht->tbl, new_tbl);
 333
 334         spin_lock(&ht->lock);
 335         list_for_each_entry(walker, &old_tbl->walkers, list)
 336                 walker->tbl = NULL;
 337
 338         /* Wait for readers. All new readers will see the new
 339          * table, and thus no references to the old table will
 340          * remain.
 341          * We do this inside the locked region so that
 342          * rhashtable_walk_stop() can use rcu_head_after_call_rcu()
 343          * to check if it should not re-link the table.
 344          */
 345         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 346         spin_unlock(&ht->lock);
 347
 348         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 349 }
 350
 351 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 352                                    struct bucket_table *old_tbl,
 353                                    unsigned int size)
 354 {
 355         struct bucket_table *new_tbl;
 356         int err;
 357
 358         ASSERT_RHT_MUTEX(ht);
 359
 360         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 361         if (new_tbl == NULL)
 362                 return -ENOMEM;
 363
 364         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 365         if (err)
 366                 bucket_table_free(new_tbl);
 367
 368         return err;
 369 }
 370
 371 /**
 372  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 373  * @ht:         the hash table to shrink
 374  *
 375  * This function shrinks the hash table to fit, i.e., the smallest
 376  * size would not cause it to expand right away automatically.
 377  *
 378  * The caller must ensure that no concurrent resizing occurs by holding
 379  * ht->mutex.
 380  *
 381  * The caller must ensure that no concurrent table mutations take place.
 382  * It is however valid to have concurrent lookups if they are RCU protected.
 383  *
 384  * It is valid to have concurrent insertions and deletions protected by per
 385  * bucket locks or concurrent RCU protected lookups and traversals.
 386  */
 387 static int rhashtable_shrink(struct rhashtable *ht)
 388 {
 389         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 390         unsigned int nelems = atomic_read(&ht->nelems);
 391         unsigned int size = 0;
 392
 393         if (nelems)
 394                 size = roundup_pow_of_two(nelems * 3 / 2);
 395         if (size < ht->p.min_size)
 396                 size = ht->p.min_size;
 397
 398         if (old_tbl->size <= size)
 399                 return 0;
 400
 401         if (rht_dereference(old_tbl->future_tbl, ht))
 402                 return -EEXIST;
 403
 404         return rhashtable_rehash_alloc(ht, old_tbl, size);
 405 }
 406
 407 static void rht_deferred_worker(struct work_struct *work)
 408 {
 409         struct rhashtable *ht;
 410         struct bucket_table *tbl;
 411         int err = 0;
 412
 413         ht = container_of(work, struct rhashtable, run_work);
 414         mutex_lock(&ht->mutex);
 415
 416         tbl = rht_dereference(ht->tbl, ht);
 417         tbl = rhashtable_last_table(ht, tbl);
 418
 419         if (rht_grow_above_75(ht, tbl))
 420                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 421         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 422                 err = rhashtable_shrink(ht);
 423         else if (tbl->nest)
 424                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 425
 426         if (!err || err == -EEXIST) {
 427                 int nerr;
 428
 429                 nerr = rhashtable_rehash_table(ht);
 430                 err = err ?: nerr;
 431         }
 432
 433         mutex_unlock(&ht->mutex);
 434
 435         if (err)
 436                 schedule_work(&ht->run_work);
 437 }
 438
 439 static int rhashtable_insert_rehash(struct rhashtable *ht,
 440                                     struct bucket_table *tbl)
 441 {
 442         struct bucket_table *old_tbl;
 443         struct bucket_table *new_tbl;
 444         unsigned int size;
 445         int err;
 446
 447         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 448
 449         size = tbl->size;
 450
 451         err = -EBUSY;
 452
 453         if (rht_grow_above_75(ht, tbl))
 454                 size *= 2;
 455         /* Do not schedule more than one rehash */
 456         else if (old_tbl != tbl)
 457                 goto fail;
 458
 459         err = -ENOMEM;
 460
 461         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC | __GFP_NOWARN);
 462         if (new_tbl == NULL)
 463                 goto fail;
 464
 465         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 466         if (err) {
 467                 bucket_table_free(new_tbl);
 468                 if (err == -EEXIST)
 469                         err = 0;
 470         } else
 471                 schedule_work(&ht->run_work);
 472
 473         return err;
 474
 475 fail:
 476         /* Do not fail the insert if someone else did a rehash. */
 477         if (likely(rcu_access_pointer(tbl->future_tbl)))
 478                 return 0;
 479
 480         /* Schedule async rehash to retry allocation in process context. */
 481         if (err == -ENOMEM)
 482                 schedule_work(&ht->run_work);
 483
 484         return err;
 485 }
 486
 487 static void *rhashtable_lookup_one(struct rhashtable *ht,
 488                                    struct rhash_lock_head **bkt,
 489                                    struct bucket_table *tbl, unsigned int hash,
 490                                    const void *key, struct rhash_head *obj)
 491 {
 492         struct rhashtable_compare_arg arg = {
 493                 .ht = ht,
 494                 .key = key,
 495         };
 496         struct rhash_head __rcu **pprev = NULL;
 497         struct rhash_head *head;
 498         int elasticity;
 499
 500         elasticity = RHT_ELASTICITY;
 501         rht_for_each_from(head, rht_ptr(bkt, tbl, hash), tbl, hash) {
 502                 struct rhlist_head *list;
 503                 struct rhlist_head *plist;
 504
 505                 elasticity--;
 506                 if (!key ||
 507                     (ht->p.obj_cmpfn ?
 508                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 509                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 510                         pprev = &head->next;
 511                         continue;
 512                 }
 513
 514                 if (!ht->rhlist)
 515                         return rht_obj(ht, head);
 516
 517                 list = container_of(obj, struct rhlist_head, rhead);
 518                 plist = container_of(head, struct rhlist_head, rhead);
 519
 520                 RCU_INIT_POINTER(list->next, plist);
 521                 head = rht_dereference_bucket(head->next, tbl, hash);
 522                 RCU_INIT_POINTER(list->rhead.next, head);
 523                 if (pprev)
 524                         rcu_assign_pointer(*pprev, obj);
 525                 else
 526                         /* Need to preserve the bit lock */
 527                         rht_assign_locked(bkt, obj);
 528
 529                 return NULL;
 530         }
 531
 532         if (elasticity <= 0)
 533                 return ERR_PTR(-EAGAIN);
 534
 535         return ERR_PTR(-ENOENT);
 536 }
 537
 538 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 539                                                   struct rhash_lock_head **bkt,
 540                                                   struct bucket_table *tbl,
 541                                                   unsigned int hash,
 542                                                   struct rhash_head *obj,
 543                                                   void *data)
 544 {
 545         struct bucket_table *new_tbl;
 546         struct rhash_head *head;
 547
 548         if (!IS_ERR_OR_NULL(data))
 549                 return ERR_PTR(-EEXIST);
 550
 551         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 552                 return ERR_CAST(data);
 553
 554         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 555         if (new_tbl)
 556                 return new_tbl;
 557
 558         if (PTR_ERR(data) != -ENOENT)
 559                 return ERR_CAST(data);
 560
 561         if (unlikely(rht_grow_above_max(ht, tbl)))
 562                 return ERR_PTR(-E2BIG);
 563
 564         if (unlikely(rht_grow_above_100(ht, tbl)))
 565                 return ERR_PTR(-EAGAIN);
 566
 567         head = rht_ptr(bkt, tbl, hash);
 568
 569         RCU_INIT_POINTER(obj->next, head);
 570         if (ht->rhlist) {
 571                 struct rhlist_head *list;
 572
 573                 list = container_of(obj, struct rhlist_head, rhead);
 574                 RCU_INIT_POINTER(list->next, NULL);
 575         }
 576
 577         /* bkt is always the head of the list, so it holds
 578          * the lock, which we need to preserve
 579          */
 580         rht_assign_locked(bkt, obj);
 581
 582         atomic_inc(&ht->nelems);
 583         if (rht_grow_above_75(ht, tbl))
 584                 schedule_work(&ht->run_work);
 585
 586         return NULL;
 587 }
 588
 589 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 590                                    struct rhash_head *obj)
 591 {
 592         struct bucket_table *new_tbl;
 593         struct bucket_table *tbl;
 594         struct rhash_lock_head **bkt;
 595         unsigned int hash;
 596         void *data;
 597
 598         new_tbl = rcu_dereference(ht->tbl);
 599
 600         do {
 601                 tbl = new_tbl;
 602                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 603                 if (rcu_access_pointer(tbl->future_tbl))
 604                         /* Failure is OK */
 605                         bkt = rht_bucket_var(tbl, hash);
 606                 else
 607                         bkt = rht_bucket_insert(ht, tbl, hash);
 608                 if (bkt == NULL) {
 609                         new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 610                         data = ERR_PTR(-EAGAIN);
 611                 } else {
 612                         rht_lock(tbl, bkt);
 613                         data = rhashtable_lookup_one(ht, bkt, tbl,
 614                                                      hash, key, obj);
 615                         new_tbl = rhashtable_insert_one(ht, bkt, tbl,
 616                                                         hash, obj, data);
 617                         if (PTR_ERR(new_tbl) != -EEXIST)
 618                                 data = ERR_CAST(new_tbl);
 619
 620                         rht_unlock(tbl, bkt);
 621                 }
 622         } while (!IS_ERR_OR_NULL(new_tbl));
 623
 624         if (PTR_ERR(data) == -EAGAIN)
 625                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 626                                -EAGAIN);
 627
 628         return data;
 629 }
 630
 631 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 632                              struct rhash_head *obj)
 633 {
 634         void *data;
 635
 636         do {
 637                 rcu_read_lock();
 638                 data = rhashtable_try_insert(ht, key, obj);
 639                 rcu_read_unlock();
 640         } while (PTR_ERR(data) == -EAGAIN);
 641
 642         return data;
 643 }
 644 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 645
 646 /**
 647  * rhashtable_walk_enter - Initialise an iterator
 648  * @ht:         Table to walk over
 649  * @iter:       Hash table Iterator
 650  *
 651  * This function prepares a hash table walk.
 652  *
 653  * Note that if you restart a walk after rhashtable_walk_stop you
 654  * may see the same object twice.  Also, you may miss objects if
 655  * there are removals in between rhashtable_walk_stop and the next
 656  * call to rhashtable_walk_start.
 657  *
 658  * For a completely stable walk you should construct your own data
 659  * structure outside the hash table.
 660  *
 661  * This function may be called from any process context, including
 662  * non-preemptable context, but cannot be called from softirq or
 663  * hardirq context.
 664  *
 665  * You must call rhashtable_walk_exit after this function returns.
 666  */
 667 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 668 {
 669         iter->ht = ht;
 670         iter->p = NULL;
 671         iter->slot = 0;
 672         iter->skip = 0;
 673         iter->end_of_table = 0;
 674
 675         spin_lock(&ht->lock);
 676         iter->walker.tbl =
 677                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 678         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 679         spin_unlock(&ht->lock);
 680 }
 681 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 682
 683 /**
 684  * rhashtable_walk_exit - Free an iterator
 685  * @iter:       Hash table Iterator
 686  *
 687  * This function frees resources allocated by rhashtable_walk_enter.
 688  */
 689 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 690 {
 691         spin_lock(&iter->ht->lock);
 692         if (iter->walker.tbl)
 693                 list_del(&iter->walker.list);
 694         spin_unlock(&iter->ht->lock);
 695 }
 696 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 697
 698 /**
 699  * rhashtable_walk_start_check - Start a hash table walk
 700  * @iter:       Hash table iterator
 701  *
 702  * Start a hash table walk at the current iterator position.  Note that we take
 703  * the RCU lock in all cases including when we return an error.  So you must
 704  * always call rhashtable_walk_stop to clean up.
 705  *
 706  * Returns zero if successful.
 707  *
 708  * Returns -EAGAIN if resize event occured.  Note that the iterator
 709  * will rewind back to the beginning and you may use it immediately
 710  * by calling rhashtable_walk_next.
 711  *
 712  * rhashtable_walk_start is defined as an inline variant that returns
 713  * void. This is preferred in cases where the caller would ignore
 714  * resize events and always continue.
 715  */
 716 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
 717         __acquires(RCU)
 718 {
 719         struct rhashtable *ht = iter->ht;
 720         bool rhlist = ht->rhlist;
 721
 722         rcu_read_lock();
 723
 724         spin_lock(&ht->lock);
 725         if (iter->walker.tbl)
 726                 list_del(&iter->walker.list);
 727         spin_unlock(&ht->lock);
 728
 729         if (iter->end_of_table)
 730                 return 0;
 731         if (!iter->walker.tbl) {
 732                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 733                 iter->slot = 0;
 734                 iter->skip = 0;
 735                 return -EAGAIN;
 736         }
 737
 738         if (iter->p && !rhlist) {
 739                 /*
 740                  * We need to validate that 'p' is still in the table, and
 741                  * if so, update 'skip'
 742                  */
 743                 struct rhash_head *p;
 744                 int skip = 0;
 745                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 746                         skip++;
 747                         if (p == iter->p) {
 748                                 iter->skip = skip;
 749                                 goto found;
 750                         }
 751                 }
 752                 iter->p = NULL;
 753         } else if (iter->p && rhlist) {
 754                 /* Need to validate that 'list' is still in the table, and
 755                  * if so, update 'skip' and 'p'.
 756                  */
 757                 struct rhash_head *p;
 758                 struct rhlist_head *list;
 759                 int skip = 0;
 760                 rht_for_each_rcu(p, iter->walker.tbl, iter->slot) {
 761                         for (list = container_of(p, struct rhlist_head, rhead);
 762                              list;
 763                              list = rcu_dereference(list->next)) {
 764                                 skip++;
 765                                 if (list == iter->list) {
 766                                         iter->p = p;
 767                                         iter->skip = skip;
 768                                         goto found;
 769                                 }
 770                         }
 771                 }
 772                 iter->p = NULL;
 773         }
 774 found:
 775         return 0;
 776 }
 777 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
 778
 779 /**
 780  * __rhashtable_walk_find_next - Find the next element in a table (or the first
 781  * one in case of a new walk).
 782  *
 783  * @iter:       Hash table iterator
 784  *
 785  * Returns the found object or NULL when the end of the table is reached.
 786  *
 787  * Returns -EAGAIN if resize event occurred.
 788  */
 789 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
 790 {
 791         struct bucket_table *tbl = iter->walker.tbl;
 792         struct rhlist_head *list = iter->list;
 793         struct rhashtable *ht = iter->ht;
 794         struct rhash_head *p = iter->p;
 795         bool rhlist = ht->rhlist;
 796
 797         if (!tbl)
 798                 return NULL;
 799
 800         for (; iter->slot < tbl->size; iter->slot++) {
 801                 int skip = iter->skip;
 802
 803                 rht_for_each_rcu(p, tbl, iter->slot) {
 804                         if (rhlist) {
 805                                 list = container_of(p, struct rhlist_head,
 806                                                     rhead);
 807                                 do {
 808                                         if (!skip)
 809                                                 goto next;
 810                                         skip--;
 811                                         list = rcu_dereference(list->next);
 812                                 } while (list);
 813
 814                                 continue;
 815                         }
 816                         if (!skip)
 817                                 break;
 818                         skip--;
 819                 }
 820
 821 next:
 822                 if (!rht_is_a_nulls(p)) {
 823                         iter->skip++;
 824                         iter->p = p;
 825                         iter->list = list;
 826                         return rht_obj(ht, rhlist ? &list->rhead : p);
 827                 }
 828
 829                 iter->skip = 0;
 830         }
 831
 832         iter->p = NULL;
 833
 834         /* Ensure we see any new tables. */
 835         smp_rmb();
 836
 837         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 838         if (iter->walker.tbl) {
 839                 iter->slot = 0;
 840                 iter->skip = 0;
 841                 return ERR_PTR(-EAGAIN);
 842         } else {
 843                 iter->end_of_table = true;
 844         }
 845
 846         return NULL;
 847 }
 848
 849 /**
 850  * rhashtable_walk_next - Return the next object and advance the iterator
 851  * @iter:       Hash table iterator
 852  *
 853  * Note that you must call rhashtable_walk_stop when you are finished
 854  * with the walk.
 855  *
 856  * Returns the next object or NULL when the end of the table is reached.
 857  *
 858  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 859  * will rewind back to the beginning and you may continue to use it.
 860  */
 861 void *rhashtable_walk_next(struct rhashtable_iter *iter)
 862 {
 863         struct rhlist_head *list = iter->list;
 864         struct rhashtable *ht = iter->ht;
 865         struct rhash_head *p = iter->p;
 866         bool rhlist = ht->rhlist;
 867
 868         if (p) {
 869                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 870                         p = rcu_dereference(p->next);
 871                         list = container_of(p, struct rhlist_head, rhead);
 872                 }
 873                 if (!rht_is_a_nulls(p)) {
 874                         iter->skip++;
 875                         iter->p = p;
 876                         iter->list = list;
 877                         return rht_obj(ht, rhlist ? &list->rhead : p);
 878                 }
 879
 880                 /* At the end of this slot, switch to next one and then find
 881                  * next entry from that point.
 882                  */
 883                 iter->skip = 0;
 884                 iter->slot++;
 885         }
 886
 887         return __rhashtable_walk_find_next(iter);
 888 }
 889 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 890
 891 /**
 892  * rhashtable_walk_peek - Return the next object but don't advance the iterator
 893  * @iter:       Hash table iterator
 894  *
 895  * Returns the next object or NULL when the end of the table is reached.
 896  *
 897  * Returns -EAGAIN if resize event occurred.  Note that the iterator
 898  * will rewind back to the beginning and you may continue to use it.
 899  */
 900 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
 901 {
 902         struct rhlist_head *list = iter->list;
 903         struct rhashtable *ht = iter->ht;
 904         struct rhash_head *p = iter->p;
 905
 906         if (p)
 907                 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
 908
 909         /* No object found in current iter, find next one in the table. */
 910
 911         if (iter->skip) {
 912                 /* A nonzero skip value points to the next entry in the table
 913                  * beyond that last one that was found. Decrement skip so
 914                  * we find the current value. __rhashtable_walk_find_next
 915                  * will restore the original value of skip assuming that
 916                  * the table hasn't changed.
 917                  */
 918                 iter->skip--;
 919         }
 920
 921         return __rhashtable_walk_find_next(iter);
 922 }
 923 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
 924
 925 /**
 926  * rhashtable_walk_stop - Finish a hash table walk
 927  * @iter:       Hash table iterator
 928  *
 929  * Finish a hash table walk.  Does not reset the iterator to the start of the
 930  * hash table.
 931  */
 932 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 933         __releases(RCU)
 934 {
 935         struct rhashtable *ht;
 936         struct bucket_table *tbl = iter->walker.tbl;
 937
 938         if (!tbl)
 939                 goto out;
 940
 941         ht = iter->ht;
 942
 943         spin_lock(&ht->lock);
 944         if (rcu_head_after_call_rcu(&tbl->rcu, bucket_table_free_rcu))
 945                 /* This bucket table is being freed, don't re-link it. */
 946                 iter->walker.tbl = NULL;
 947         else
 948                 list_add(&iter->walker.list, &tbl->walkers);
 949         spin_unlock(&ht->lock);
 950
 951 out:
 952         rcu_read_unlock();
 953 }
 954 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 955
 956 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 957 {
 958         size_t retsize;
 959
 960         if (params->nelem_hint)
 961                 retsize = max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 962                               (unsigned long)params->min_size);
 963         else
 964                 retsize = max(HASH_DEFAULT_SIZE,
 965                               (unsigned long)params->min_size);
 966
 967         return retsize;
 968 }
 969
 970 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 971 {
 972         return jhash2(key, length, seed);
 973 }
 974
 975 /**
 976  * rhashtable_init - initialize a new hash table
 977  * @ht:         hash table to be initialized
 978  * @params:     configuration parameters
 979  *
 980  * Initializes a new hash table based on the provided configuration
 981  * parameters. A table can be configured either with a variable or
 982  * fixed length key:
 983  *
 984  * Configuration Example 1: Fixed length keys
 985  * struct test_obj {
 986  *      int                     key;
 987  *      void *                  my_member;
 988  *      struct rhash_head       node;
 989  * };
 990  *
 991  * struct rhashtable_params params = {
 992  *      .head_offset = offsetof(struct test_obj, node),
 993  *      .key_offset = offsetof(struct test_obj, key),
 994  *      .key_len = sizeof(int),
 995  *      .hashfn = jhash,
 996  * };
 997  *
 998  * Configuration Example 2: Variable length keys
 999  * struct test_obj {
1000  *      [...]
1001  *      struct rhash_head       node;
1002  * };
1003  *
1004  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
1005  * {
1006  *      struct test_obj *obj = data;
1007  *
1008  *      return [... hash ...];
1009  * }
1010  *
1011  * struct rhashtable_params params = {
1012  *      .head_offset = offsetof(struct test_obj, node),
1013  *      .hashfn = jhash,
1014  *      .obj_hashfn = my_hash_fn,
1015  * };
1016  */
1017 int rhashtable_init(struct rhashtable *ht,
1018                     const struct rhashtable_params *params)
1019 {
1020         struct bucket_table *tbl;
1021         size_t size;
1022
1023         if ((!params->key_len && !params->obj_hashfn) ||
1024             (params->obj_hashfn && !params->obj_cmpfn))
1025                 return -EINVAL;
1026
1027         memset(ht, 0, sizeof(*ht));
1028         mutex_init(&ht->mutex);
1029         spin_lock_init(&ht->lock);
1030         memcpy(&ht->p, params, sizeof(*params));
1031
1032         if (params->min_size)
1033                 ht->p.min_size = roundup_pow_of_two(params->min_size);
1034
1035         /* Cap total entries at 2^31 to avoid nelems overflow. */
1036         ht->max_elems = 1u << 31;
1037
1038         if (params->max_size) {
1039                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1040                 if (ht->p.max_size < ht->max_elems / 2)
1041                         ht->max_elems = ht->p.max_size * 2;
1042         }
1043
1044         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1045
1046         size = rounded_hashtable_size(&ht->p);
1047
1048         ht->key_len = ht->p.key_len;
1049         if (!params->hashfn) {
1050                 ht->p.hashfn = jhash;
1051
1052                 if (!(ht->key_len & (sizeof(u32) - 1))) {
1053                         ht->key_len /= sizeof(u32);
1054                         ht->p.hashfn = rhashtable_jhash2;
1055                 }
1056         }
1057
1058         /*
1059          * This is api initialization and thus we need to guarantee the
1060          * initial rhashtable allocation. Upon failure, retry with the
1061          * smallest possible size with __GFP_NOFAIL semantics.
1062          */
1063         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1064         if (unlikely(tbl == NULL)) {
1065                 size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1066                 tbl = bucket_table_alloc(ht, size, GFP_KERNEL | __GFP_NOFAIL);
1067         }
1068
1069         atomic_set(&ht->nelems, 0);
1070
1071         RCU_INIT_POINTER(ht->tbl, tbl);
1072
1073         INIT_WORK(&ht->run_work, rht_deferred_worker);
1074
1075         return 0;
1076 }
1077 EXPORT_SYMBOL_GPL(rhashtable_init);
1078
1079 /**
1080  * rhltable_init - initialize a new hash list table
1081  * @hlt:        hash list table to be initialized
1082  * @params:     configuration parameters
1083  *
1084  * Initializes a new hash list table.
1085  *
1086  * See documentation for rhashtable_init.
1087  */
1088 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1089 {
1090         int err;
1091
1092         err = rhashtable_init(&hlt->ht, params);
1093         hlt->ht.rhlist = true;
1094         return err;
1095 }
1096 EXPORT_SYMBOL_GPL(rhltable_init);
1097
1098 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1099                                 void (*free_fn)(void *ptr, void *arg),
1100                                 void *arg)
1101 {
1102         struct rhlist_head *list;
1103
1104         if (!ht->rhlist) {
1105                 free_fn(rht_obj(ht, obj), arg);
1106                 return;
1107         }
1108
1109         list = container_of(obj, struct rhlist_head, rhead);
1110         do {
1111                 obj = &list->rhead;
1112                 list = rht_dereference(list->next, ht);
1113                 free_fn(rht_obj(ht, obj), arg);
1114         } while (list);
1115 }
1116
1117 /**
1118  * rhashtable_free_and_destroy - free elements and destroy hash table
1119  * @ht:         the hash table to destroy
1120  * @free_fn:    callback to release resources of element
1121  * @arg:        pointer passed to free_fn
1122  *
1123  * Stops an eventual async resize. If defined, invokes free_fn for each
1124  * element to releasal resources. Please note that RCU protected
1125  * readers may still be accessing the elements. Releasing of resources
1126  * must occur in a compatible manner. Then frees the bucket array.
1127  *
1128  * This function will eventually sleep to wait for an async resize
1129  * to complete. The caller is responsible that no further write operations
1130  * occurs in parallel.
1131  */
1132 void rhashtable_free_and_destroy(struct rhashtable *ht,
1133                                  void (*free_fn)(void *ptr, void *arg),
1134                                  void *arg)
1135 {
1136         struct bucket_table *tbl, *next_tbl;
1137         unsigned int i;
1138
1139         cancel_work_sync(&ht->run_work);
1140
1141         mutex_lock(&ht->mutex);
1142         tbl = rht_dereference(ht->tbl, ht);
1143 restart:
1144         if (free_fn) {
1145                 for (i = 0; i < tbl->size; i++) {
1146                         struct rhash_head *pos, *next;
1147
1148                         cond_resched();
1149                         for (pos = rht_ptr_exclusive(rht_bucket(tbl, i)),
1150                              next = !rht_is_a_nulls(pos) ?
1151                                         rht_dereference(pos->next, ht) : NULL;
1152                              !rht_is_a_nulls(pos);
1153                              pos = next,
1154                              next = !rht_is_a_nulls(pos) ?
1155                                         rht_dereference(pos->next, ht) : NULL)
1156                                 rhashtable_free_one(ht, pos, free_fn, arg);
1157                 }
1158         }
1159
1160         next_tbl = rht_dereference(tbl->future_tbl, ht);
1161         bucket_table_free(tbl);
1162         if (next_tbl) {
1163                 tbl = next_tbl;
1164                 goto restart;
1165         }
1166         mutex_unlock(&ht->mutex);
1167 }
1168 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1169
1170 void rhashtable_destroy(struct rhashtable *ht)
1171 {
1172         return rhashtable_free_and_destroy(ht, NULL, NULL);
1173 }
1174 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1175
1176 struct rhash_lock_head **__rht_bucket_nested(const struct bucket_table *tbl,
1177                                              unsigned int hash)
1178 {
1179         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1180         unsigned int index = hash & ((1 << tbl->nest) - 1);
1181         unsigned int size = tbl->size >> tbl->nest;
1182         unsigned int subhash = hash;
1183         union nested_table *ntbl;
1184
1185         ntbl = nested_table_top(tbl);
1186         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1187         subhash >>= tbl->nest;
1188
1189         while (ntbl && size > (1 << shift)) {
1190                 index = subhash & ((1 << shift) - 1);
1191                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1192                                                   tbl, hash);
1193                 size >>= shift;
1194                 subhash >>= shift;
1195         }
1196
1197         if (!ntbl)
1198                 return NULL;
1199
1200         return &ntbl[subhash].bucket;
1201
1202 }
1203 EXPORT_SYMBOL_GPL(__rht_bucket_nested);
1204
1205 struct rhash_lock_head **rht_bucket_nested(const struct bucket_table *tbl,
1206                                            unsigned int hash)
1207 {
1208         static struct rhash_lock_head *rhnull;
1209
1210         if (!rhnull)
1211                 INIT_RHT_NULLS_HEAD(rhnull);
1212         return __rht_bucket_nested(tbl, hash) ?: &rhnull;
1213 }
1214 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1215
1216 struct rhash_lock_head **rht_bucket_nested_insert(struct rhashtable *ht,
1217                                                   struct bucket_table *tbl,
1218                                                   unsigned int hash)
1219 {
1220         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1221         unsigned int index = hash & ((1 << tbl->nest) - 1);
1222         unsigned int size = tbl->size >> tbl->nest;
1223         union nested_table *ntbl;
1224
1225         ntbl = nested_table_top(tbl);
1226         hash >>= tbl->nest;
1227         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1228                                   size <= (1 << shift));
1229
1230         while (ntbl && size > (1 << shift)) {
1231                 index = hash & ((1 << shift) - 1);
1232                 size >>= shift;
1233                 hash >>= shift;
1234                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1235                                           size <= (1 << shift));
1236         }
1237
1238         if (!ntbl)
1239                 return NULL;
1240
1241         return &ntbl[hash].bucket;
1242
1243 }
1244 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);