lib/rhashtable.c

   1 /*
   2  * Resizable, Scalable, Concurrent Hash Table
   3  *
   4  * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
   5  * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
   6  * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
   7  *
   8  * Code partially derived from nft_hash
   9  * Rewritten with rehash code from br_multicast plus single list
  10  * pointer as suggested by Josh Triplett
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2 as
  14  * published by the Free Software Foundation.
  15  */
  16
  17 #include <linux/atomic.h>
  18 #include <linux/kernel.h>
  19 #include <linux/init.h>
  20 #include <linux/log2.h>
  21 #include <linux/sched.h>
  22 #include <linux/rculist.h>
  23 #include <linux/slab.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/mm.h>
  26 #include <linux/jhash.h>
  27 #include <linux/random.h>
  28 #include <linux/rhashtable.h>
  29 #include <linux/err.h>
  30 #include <linux/export.h>
  31
  32 #define HASH_DEFAULT_SIZE       64UL
  33 #define HASH_MIN_SIZE           4U
  34 #define BUCKET_LOCKS_PER_CPU    32UL
  35
  36 union nested_table {
  37         union nested_table __rcu *table;
  38         struct rhash_head __rcu *bucket;
  39 };
  40
  41 static u32 head_hashfn(struct rhashtable *ht,
  42                        const struct bucket_table *tbl,
  43                        const struct rhash_head *he)
  44 {
  45         return rht_head_hashfn(ht, tbl, he, ht->p);
  46 }
  47
  48 #ifdef CONFIG_PROVE_LOCKING
  49 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
  50
  51 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
  52 {
  53         return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
  54 }
  55 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
  56
  57 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
  58 {
  59         spinlock_t *lock = rht_bucket_lock(tbl, hash);
  60
  61         return (debug_locks) ? lockdep_is_held(lock) : 1;
  62 }
  63 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
  64 #else
  65 #define ASSERT_RHT_MUTEX(HT)
  66 #endif
  67
  68
  69 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
  70                               gfp_t gfp)
  71 {
  72         unsigned int i, size;
  73 #if defined(CONFIG_PROVE_LOCKING)
  74         unsigned int nr_pcpus = 2;
  75 #else
  76         unsigned int nr_pcpus = num_possible_cpus();
  77 #endif
  78
  79         nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
  80         size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
  81
  82         /* Never allocate more than 0.5 locks per bucket */
  83         size = min_t(unsigned int, size, tbl->size >> 1);
  84
  85         if (tbl->nest)
  86                 size = min(size, 1U << tbl->nest);
  87
  88         if (sizeof(spinlock_t) != 0) {
  89                 if (gfpflags_allow_blocking(gfp))
  90                         tbl->locks = kvmalloc(size * sizeof(spinlock_t), gfp);
  91                 else
  92                         tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
  93                                                    gfp);
  94                 if (!tbl->locks)
  95                         return -ENOMEM;
  96                 for (i = 0; i < size; i++)
  97                         spin_lock_init(&tbl->locks[i]);
  98         }
  99         tbl->locks_mask = size - 1;
 100
 101         return 0;
 102 }
 103
 104 static void nested_table_free(union nested_table *ntbl, unsigned int size)
 105 {
 106         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 107         const unsigned int len = 1 << shift;
 108         unsigned int i;
 109
 110         ntbl = rcu_dereference_raw(ntbl->table);
 111         if (!ntbl)
 112                 return;
 113
 114         if (size > len) {
 115                 size >>= shift;
 116                 for (i = 0; i < len; i++)
 117                         nested_table_free(ntbl + i, size);
 118         }
 119
 120         kfree(ntbl);
 121 }
 122
 123 static void nested_bucket_table_free(const struct bucket_table *tbl)
 124 {
 125         unsigned int size = tbl->size >> tbl->nest;
 126         unsigned int len = 1 << tbl->nest;
 127         union nested_table *ntbl;
 128         unsigned int i;
 129
 130         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
 131
 132         for (i = 0; i < len; i++)
 133                 nested_table_free(ntbl + i, size);
 134
 135         kfree(ntbl);
 136 }
 137
 138 static void bucket_table_free(const struct bucket_table *tbl)
 139 {
 140         if (tbl->nest)
 141                 nested_bucket_table_free(tbl);
 142
 143         kvfree(tbl->locks);
 144         kvfree(tbl);
 145 }
 146
 147 static void bucket_table_free_rcu(struct rcu_head *head)
 148 {
 149         bucket_table_free(container_of(head, struct bucket_table, rcu));
 150 }
 151
 152 static union nested_table *nested_table_alloc(struct rhashtable *ht,
 153                                               union nested_table __rcu **prev,
 154                                               unsigned int shifted,
 155                                               unsigned int nhash)
 156 {
 157         union nested_table *ntbl;
 158         int i;
 159
 160         ntbl = rcu_dereference(*prev);
 161         if (ntbl)
 162                 return ntbl;
 163
 164         ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
 165
 166         if (ntbl && shifted) {
 167                 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
 168                         INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
 169                                             (i << shifted) | nhash);
 170         }
 171
 172         rcu_assign_pointer(*prev, ntbl);
 173
 174         return ntbl;
 175 }
 176
 177 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
 178                                                       size_t nbuckets,
 179                                                       gfp_t gfp)
 180 {
 181         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
 182         struct bucket_table *tbl;
 183         size_t size;
 184
 185         if (nbuckets < (1 << (shift + 1)))
 186                 return NULL;
 187
 188         size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
 189
 190         tbl = kzalloc(size, gfp);
 191         if (!tbl)
 192                 return NULL;
 193
 194         if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
 195                                 0, 0)) {
 196                 kfree(tbl);
 197                 return NULL;
 198         }
 199
 200         tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
 201
 202         return tbl;
 203 }
 204
 205 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
 206                                                size_t nbuckets,
 207                                                gfp_t gfp)
 208 {
 209         struct bucket_table *tbl = NULL;
 210         size_t size;
 211         int i;
 212
 213         size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
 214         if (gfp != GFP_KERNEL)
 215                 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
 216         else
 217                 tbl = kvzalloc(size, gfp);
 218
 219         size = nbuckets;
 220
 221         if (tbl == NULL && gfp != GFP_KERNEL) {
 222                 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
 223                 nbuckets = 0;
 224         }
 225         if (tbl == NULL)
 226                 return NULL;
 227
 228         tbl->size = size;
 229
 230         if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
 231                 bucket_table_free(tbl);
 232                 return NULL;
 233         }
 234
 235         INIT_LIST_HEAD(&tbl->walkers);
 236
 237         tbl->hash_rnd = get_random_u32();
 238
 239         for (i = 0; i < nbuckets; i++)
 240                 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
 241
 242         return tbl;
 243 }
 244
 245 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
 246                                                   struct bucket_table *tbl)
 247 {
 248         struct bucket_table *new_tbl;
 249
 250         do {
 251                 new_tbl = tbl;
 252                 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 253         } while (tbl);
 254
 255         return new_tbl;
 256 }
 257
 258 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
 259 {
 260         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 261         struct bucket_table *new_tbl = rhashtable_last_table(ht,
 262                 rht_dereference_rcu(old_tbl->future_tbl, ht));
 263         struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
 264         int err = -EAGAIN;
 265         struct rhash_head *head, *next, *entry;
 266         spinlock_t *new_bucket_lock;
 267         unsigned int new_hash;
 268
 269         if (new_tbl->nest)
 270                 goto out;
 271
 272         err = -ENOENT;
 273
 274         rht_for_each(entry, old_tbl, old_hash) {
 275                 err = 0;
 276                 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
 277
 278                 if (rht_is_a_nulls(next))
 279                         break;
 280
 281                 pprev = &entry->next;
 282         }
 283
 284         if (err)
 285                 goto out;
 286
 287         new_hash = head_hashfn(ht, new_tbl, entry);
 288
 289         new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
 290
 291         spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
 292         head = rht_dereference_bucket(new_tbl->buckets[new_hash],
 293                                       new_tbl, new_hash);
 294
 295         RCU_INIT_POINTER(entry->next, head);
 296
 297         rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
 298         spin_unlock(new_bucket_lock);
 299
 300         rcu_assign_pointer(*pprev, next);
 301
 302 out:
 303         return err;
 304 }
 305
 306 static int rhashtable_rehash_chain(struct rhashtable *ht,
 307                                     unsigned int old_hash)
 308 {
 309         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 310         spinlock_t *old_bucket_lock;
 311         int err;
 312
 313         old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
 314
 315         spin_lock_bh(old_bucket_lock);
 316         while (!(err = rhashtable_rehash_one(ht, old_hash)))
 317                 ;
 318
 319         if (err == -ENOENT) {
 320                 old_tbl->rehash++;
 321                 err = 0;
 322         }
 323         spin_unlock_bh(old_bucket_lock);
 324
 325         return err;
 326 }
 327
 328 static int rhashtable_rehash_attach(struct rhashtable *ht,
 329                                     struct bucket_table *old_tbl,
 330                                     struct bucket_table *new_tbl)
 331 {
 332         /* Protect future_tbl using the first bucket lock. */
 333         spin_lock_bh(old_tbl->locks);
 334
 335         /* Did somebody beat us to it? */
 336         if (rcu_access_pointer(old_tbl->future_tbl)) {
 337                 spin_unlock_bh(old_tbl->locks);
 338                 return -EEXIST;
 339         }
 340
 341         /* Make insertions go into the new, empty table right away. Deletions
 342          * and lookups will be attempted in both tables until we synchronize.
 343          */
 344         rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
 345
 346         spin_unlock_bh(old_tbl->locks);
 347
 348         return 0;
 349 }
 350
 351 static int rhashtable_rehash_table(struct rhashtable *ht)
 352 {
 353         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 354         struct bucket_table *new_tbl;
 355         struct rhashtable_walker *walker;
 356         unsigned int old_hash;
 357         int err;
 358
 359         new_tbl = rht_dereference(old_tbl->future_tbl, ht);
 360         if (!new_tbl)
 361                 return 0;
 362
 363         for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
 364                 err = rhashtable_rehash_chain(ht, old_hash);
 365                 if (err)
 366                         return err;
 367         }
 368
 369         /* Publish the new table pointer. */
 370         rcu_assign_pointer(ht->tbl, new_tbl);
 371
 372         spin_lock(&ht->lock);
 373         list_for_each_entry(walker, &old_tbl->walkers, list)
 374                 walker->tbl = NULL;
 375         spin_unlock(&ht->lock);
 376
 377         /* Wait for readers. All new readers will see the new
 378          * table, and thus no references to the old table will
 379          * remain.
 380          */
 381         call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
 382
 383         return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
 384 }
 385
 386 static int rhashtable_rehash_alloc(struct rhashtable *ht,
 387                                    struct bucket_table *old_tbl,
 388                                    unsigned int size)
 389 {
 390         struct bucket_table *new_tbl;
 391         int err;
 392
 393         ASSERT_RHT_MUTEX(ht);
 394
 395         new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 396         if (new_tbl == NULL)
 397                 return -ENOMEM;
 398
 399         err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
 400         if (err)
 401                 bucket_table_free(new_tbl);
 402
 403         return err;
 404 }
 405
 406 /**
 407  * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
 408  * @ht:         the hash table to shrink
 409  *
 410  * This function shrinks the hash table to fit, i.e., the smallest
 411  * size would not cause it to expand right away automatically.
 412  *
 413  * The caller must ensure that no concurrent resizing occurs by holding
 414  * ht->mutex.
 415  *
 416  * The caller must ensure that no concurrent table mutations take place.
 417  * It is however valid to have concurrent lookups if they are RCU protected.
 418  *
 419  * It is valid to have concurrent insertions and deletions protected by per
 420  * bucket locks or concurrent RCU protected lookups and traversals.
 421  */
 422 static int rhashtable_shrink(struct rhashtable *ht)
 423 {
 424         struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
 425         unsigned int nelems = atomic_read(&ht->nelems);
 426         unsigned int size = 0;
 427
 428         if (nelems)
 429                 size = roundup_pow_of_two(nelems * 3 / 2);
 430         if (size < ht->p.min_size)
 431                 size = ht->p.min_size;
 432
 433         if (old_tbl->size <= size)
 434                 return 0;
 435
 436         if (rht_dereference(old_tbl->future_tbl, ht))
 437                 return -EEXIST;
 438
 439         return rhashtable_rehash_alloc(ht, old_tbl, size);
 440 }
 441
 442 static void rht_deferred_worker(struct work_struct *work)
 443 {
 444         struct rhashtable *ht;
 445         struct bucket_table *tbl;
 446         int err = 0;
 447
 448         ht = container_of(work, struct rhashtable, run_work);
 449         mutex_lock(&ht->mutex);
 450
 451         tbl = rht_dereference(ht->tbl, ht);
 452         tbl = rhashtable_last_table(ht, tbl);
 453
 454         if (rht_grow_above_75(ht, tbl))
 455                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
 456         else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
 457                 err = rhashtable_shrink(ht);
 458         else if (tbl->nest)
 459                 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
 460
 461         if (!err)
 462                 err = rhashtable_rehash_table(ht);
 463
 464         mutex_unlock(&ht->mutex);
 465
 466         if (err)
 467                 schedule_work(&ht->run_work);
 468 }
 469
 470 static int rhashtable_insert_rehash(struct rhashtable *ht,
 471                                     struct bucket_table *tbl)
 472 {
 473         struct bucket_table *old_tbl;
 474         struct bucket_table *new_tbl;
 475         unsigned int size;
 476         int err;
 477
 478         old_tbl = rht_dereference_rcu(ht->tbl, ht);
 479
 480         size = tbl->size;
 481
 482         err = -EBUSY;
 483
 484         if (rht_grow_above_75(ht, tbl))
 485                 size *= 2;
 486         /* Do not schedule more than one rehash */
 487         else if (old_tbl != tbl)
 488                 goto fail;
 489
 490         err = -ENOMEM;
 491
 492         new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
 493         if (new_tbl == NULL)
 494                 goto fail;
 495
 496         err = rhashtable_rehash_attach(ht, tbl, new_tbl);
 497         if (err) {
 498                 bucket_table_free(new_tbl);
 499                 if (err == -EEXIST)
 500                         err = 0;
 501         } else
 502                 schedule_work(&ht->run_work);
 503
 504         return err;
 505
 506 fail:
 507         /* Do not fail the insert if someone else did a rehash. */
 508         if (likely(rcu_dereference_raw(tbl->future_tbl)))
 509                 return 0;
 510
 511         /* Schedule async rehash to retry allocation in process context. */
 512         if (err == -ENOMEM)
 513                 schedule_work(&ht->run_work);
 514
 515         return err;
 516 }
 517
 518 static void *rhashtable_lookup_one(struct rhashtable *ht,
 519                                    struct bucket_table *tbl, unsigned int hash,
 520                                    const void *key, struct rhash_head *obj)
 521 {
 522         struct rhashtable_compare_arg arg = {
 523                 .ht = ht,
 524                 .key = key,
 525         };
 526         struct rhash_head __rcu **pprev;
 527         struct rhash_head *head;
 528         int elasticity;
 529
 530         elasticity = RHT_ELASTICITY;
 531         pprev = rht_bucket_var(tbl, hash);
 532         rht_for_each_continue(head, *pprev, tbl, hash) {
 533                 struct rhlist_head *list;
 534                 struct rhlist_head *plist;
 535
 536                 elasticity--;
 537                 if (!key ||
 538                     (ht->p.obj_cmpfn ?
 539                      ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
 540                      rhashtable_compare(&arg, rht_obj(ht, head)))) {
 541                         pprev = &head->next;
 542                         continue;
 543                 }
 544
 545                 if (!ht->rhlist)
 546                         return rht_obj(ht, head);
 547
 548                 list = container_of(obj, struct rhlist_head, rhead);
 549                 plist = container_of(head, struct rhlist_head, rhead);
 550
 551                 RCU_INIT_POINTER(list->next, plist);
 552                 head = rht_dereference_bucket(head->next, tbl, hash);
 553                 RCU_INIT_POINTER(list->rhead.next, head);
 554                 rcu_assign_pointer(*pprev, obj);
 555
 556                 return NULL;
 557         }
 558
 559         if (elasticity <= 0)
 560                 return ERR_PTR(-EAGAIN);
 561
 562         return ERR_PTR(-ENOENT);
 563 }
 564
 565 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
 566                                                   struct bucket_table *tbl,
 567                                                   unsigned int hash,
 568                                                   struct rhash_head *obj,
 569                                                   void *data)
 570 {
 571         struct rhash_head __rcu **pprev;
 572         struct bucket_table *new_tbl;
 573         struct rhash_head *head;
 574
 575         if (!IS_ERR_OR_NULL(data))
 576                 return ERR_PTR(-EEXIST);
 577
 578         if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
 579                 return ERR_CAST(data);
 580
 581         new_tbl = rcu_dereference(tbl->future_tbl);
 582         if (new_tbl)
 583                 return new_tbl;
 584
 585         if (PTR_ERR(data) != -ENOENT)
 586                 return ERR_CAST(data);
 587
 588         if (unlikely(rht_grow_above_max(ht, tbl)))
 589                 return ERR_PTR(-E2BIG);
 590
 591         if (unlikely(rht_grow_above_100(ht, tbl)))
 592                 return ERR_PTR(-EAGAIN);
 593
 594         pprev = rht_bucket_insert(ht, tbl, hash);
 595         if (!pprev)
 596                 return ERR_PTR(-ENOMEM);
 597
 598         head = rht_dereference_bucket(*pprev, tbl, hash);
 599
 600         RCU_INIT_POINTER(obj->next, head);
 601         if (ht->rhlist) {
 602                 struct rhlist_head *list;
 603
 604                 list = container_of(obj, struct rhlist_head, rhead);
 605                 RCU_INIT_POINTER(list->next, NULL);
 606         }
 607
 608         rcu_assign_pointer(*pprev, obj);
 609
 610         atomic_inc(&ht->nelems);
 611         if (rht_grow_above_75(ht, tbl))
 612                 schedule_work(&ht->run_work);
 613
 614         return NULL;
 615 }
 616
 617 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
 618                                    struct rhash_head *obj)
 619 {
 620         struct bucket_table *new_tbl;
 621         struct bucket_table *tbl;
 622         unsigned int hash;
 623         spinlock_t *lock;
 624         void *data;
 625
 626         tbl = rcu_dereference(ht->tbl);
 627
 628         /* All insertions must grab the oldest table containing
 629          * the hashed bucket that is yet to be rehashed.
 630          */
 631         for (;;) {
 632                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 633                 lock = rht_bucket_lock(tbl, hash);
 634                 spin_lock_bh(lock);
 635
 636                 if (tbl->rehash <= hash)
 637                         break;
 638
 639                 spin_unlock_bh(lock);
 640                 tbl = rcu_dereference(tbl->future_tbl);
 641         }
 642
 643         data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 644         new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 645         if (PTR_ERR(new_tbl) != -EEXIST)
 646                 data = ERR_CAST(new_tbl);
 647
 648         while (!IS_ERR_OR_NULL(new_tbl)) {
 649                 tbl = new_tbl;
 650                 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
 651                 spin_lock_nested(rht_bucket_lock(tbl, hash),
 652                                  SINGLE_DEPTH_NESTING);
 653
 654                 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
 655                 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
 656                 if (PTR_ERR(new_tbl) != -EEXIST)
 657                         data = ERR_CAST(new_tbl);
 658
 659                 spin_unlock(rht_bucket_lock(tbl, hash));
 660         }
 661
 662         spin_unlock_bh(lock);
 663
 664         if (PTR_ERR(data) == -EAGAIN)
 665                 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
 666                                -EAGAIN);
 667
 668         return data;
 669 }
 670
 671 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
 672                              struct rhash_head *obj)
 673 {
 674         void *data;
 675
 676         do {
 677                 rcu_read_lock();
 678                 data = rhashtable_try_insert(ht, key, obj);
 679                 rcu_read_unlock();
 680         } while (PTR_ERR(data) == -EAGAIN);
 681
 682         return data;
 683 }
 684 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
 685
 686 /**
 687  * rhashtable_walk_enter - Initialise an iterator
 688  * @ht:         Table to walk over
 689  * @iter:       Hash table Iterator
 690  *
 691  * This function prepares a hash table walk.
 692  *
 693  * Note that if you restart a walk after rhashtable_walk_stop you
 694  * may see the same object twice.  Also, you may miss objects if
 695  * there are removals in between rhashtable_walk_stop and the next
 696  * call to rhashtable_walk_start.
 697  *
 698  * For a completely stable walk you should construct your own data
 699  * structure outside the hash table.
 700  *
 701  * This function may sleep so you must not call it from interrupt
 702  * context or with spin locks held.
 703  *
 704  * You must call rhashtable_walk_exit after this function returns.
 705  */
 706 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
 707 {
 708         iter->ht = ht;
 709         iter->p = NULL;
 710         iter->slot = 0;
 711         iter->skip = 0;
 712
 713         spin_lock(&ht->lock);
 714         iter->walker.tbl =
 715                 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
 716         list_add(&iter->walker.list, &iter->walker.tbl->walkers);
 717         spin_unlock(&ht->lock);
 718 }
 719 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
 720
 721 /**
 722  * rhashtable_walk_exit - Free an iterator
 723  * @iter:       Hash table Iterator
 724  *
 725  * This function frees resources allocated by rhashtable_walk_init.
 726  */
 727 void rhashtable_walk_exit(struct rhashtable_iter *iter)
 728 {
 729         spin_lock(&iter->ht->lock);
 730         if (iter->walker.tbl)
 731                 list_del(&iter->walker.list);
 732         spin_unlock(&iter->ht->lock);
 733 }
 734 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
 735
 736 /**
 737  * rhashtable_walk_start - Start a hash table walk
 738  * @iter:       Hash table iterator
 739  *
 740  * Start a hash table walk at the current iterator position.  Note that we take
 741  * the RCU lock in all cases including when we return an error.  So you must
 742  * always call rhashtable_walk_stop to clean up.
 743  *
 744  * Returns zero if successful.
 745  *
 746  * Returns -EAGAIN if resize event occured.  Note that the iterator
 747  * will rewind back to the beginning and you may use it immediately
 748  * by calling rhashtable_walk_next.
 749  */
 750 int rhashtable_walk_start(struct rhashtable_iter *iter)
 751         __acquires(RCU)
 752 {
 753         struct rhashtable *ht = iter->ht;
 754
 755         rcu_read_lock();
 756
 757         spin_lock(&ht->lock);
 758         if (iter->walker.tbl)
 759                 list_del(&iter->walker.list);
 760         spin_unlock(&ht->lock);
 761
 762         if (!iter->walker.tbl) {
 763                 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
 764                 return -EAGAIN;
 765         }
 766
 767         return 0;
 768 }
 769 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
 770
 771 /**
 772  * rhashtable_walk_next - Return the next object and advance the iterator
 773  * @iter:       Hash table iterator
 774  *
 775  * Note that you must call rhashtable_walk_stop when you are finished
 776  * with the walk.
 777  *
 778  * Returns the next object or NULL when the end of the table is reached.
 779  *
 780  * Returns -EAGAIN if resize event occured.  Note that the iterator
 781  * will rewind back to the beginning and you may continue to use it.
 782  */
 783 void *rhashtable_walk_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 (p) {
 792                 if (!rhlist || !(list = rcu_dereference(list->next))) {
 793                         p = rcu_dereference(p->next);
 794                         list = container_of(p, struct rhlist_head, rhead);
 795                 }
 796                 goto next;
 797         }
 798
 799         for (; iter->slot < tbl->size; iter->slot++) {
 800                 int skip = iter->skip;
 801
 802                 rht_for_each_rcu(p, tbl, iter->slot) {
 803                         if (rhlist) {
 804                                 list = container_of(p, struct rhlist_head,
 805                                                     rhead);
 806                                 do {
 807                                         if (!skip)
 808                                                 goto next;
 809                                         skip--;
 810                                         list = rcu_dereference(list->next);
 811                                 } while (list);
 812
 813                                 continue;
 814                         }
 815                         if (!skip)
 816                                 break;
 817                         skip--;
 818                 }
 819
 820 next:
 821                 if (!rht_is_a_nulls(p)) {
 822                         iter->skip++;
 823                         iter->p = p;
 824                         iter->list = list;
 825                         return rht_obj(ht, rhlist ? &list->rhead : p);
 826                 }
 827
 828                 iter->skip = 0;
 829         }
 830
 831         iter->p = NULL;
 832
 833         /* Ensure we see any new tables. */
 834         smp_rmb();
 835
 836         iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
 837         if (iter->walker.tbl) {
 838                 iter->slot = 0;
 839                 iter->skip = 0;
 840                 return ERR_PTR(-EAGAIN);
 841         }
 842
 843         return NULL;
 844 }
 845 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
 846
 847 /**
 848  * rhashtable_walk_stop - Finish a hash table walk
 849  * @iter:       Hash table iterator
 850  *
 851  * Finish a hash table walk.  Does not reset the iterator to the start of the
 852  * hash table.
 853  */
 854 void rhashtable_walk_stop(struct rhashtable_iter *iter)
 855         __releases(RCU)
 856 {
 857         struct rhashtable *ht;
 858         struct bucket_table *tbl = iter->walker.tbl;
 859
 860         if (!tbl)
 861                 goto out;
 862
 863         ht = iter->ht;
 864
 865         spin_lock(&ht->lock);
 866         if (tbl->rehash < tbl->size)
 867                 list_add(&iter->walker.list, &tbl->walkers);
 868         else
 869                 iter->walker.tbl = NULL;
 870         spin_unlock(&ht->lock);
 871
 872         iter->p = NULL;
 873
 874 out:
 875         rcu_read_unlock();
 876 }
 877 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
 878
 879 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
 880 {
 881         return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
 882                    (unsigned long)params->min_size);
 883 }
 884
 885 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
 886 {
 887         return jhash2(key, length, seed);
 888 }
 889
 890 /**
 891  * rhashtable_init - initialize a new hash table
 892  * @ht:         hash table to be initialized
 893  * @params:     configuration parameters
 894  *
 895  * Initializes a new hash table based on the provided configuration
 896  * parameters. A table can be configured either with a variable or
 897  * fixed length key:
 898  *
 899  * Configuration Example 1: Fixed length keys
 900  * struct test_obj {
 901  *      int                     key;
 902  *      void *                  my_member;
 903  *      struct rhash_head       node;
 904  * };
 905  *
 906  * struct rhashtable_params params = {
 907  *      .head_offset = offsetof(struct test_obj, node),
 908  *      .key_offset = offsetof(struct test_obj, key),
 909  *      .key_len = sizeof(int),
 910  *      .hashfn = jhash,
 911  *      .nulls_base = (1U << RHT_BASE_SHIFT),
 912  * };
 913  *
 914  * Configuration Example 2: Variable length keys
 915  * struct test_obj {
 916  *      [...]
 917  *      struct rhash_head       node;
 918  * };
 919  *
 920  * u32 my_hash_fn(const void *data, u32 len, u32 seed)
 921  * {
 922  *      struct test_obj *obj = data;
 923  *
 924  *      return [... hash ...];
 925  * }
 926  *
 927  * struct rhashtable_params params = {
 928  *      .head_offset = offsetof(struct test_obj, node),
 929  *      .hashfn = jhash,
 930  *      .obj_hashfn = my_hash_fn,
 931  * };
 932  */
 933 int rhashtable_init(struct rhashtable *ht,
 934                     const struct rhashtable_params *params)
 935 {
 936         struct bucket_table *tbl;
 937         size_t size;
 938
 939         size = HASH_DEFAULT_SIZE;
 940
 941         if ((!params->key_len && !params->obj_hashfn) ||
 942             (params->obj_hashfn && !params->obj_cmpfn))
 943                 return -EINVAL;
 944
 945         if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
 946                 return -EINVAL;
 947
 948         memset(ht, 0, sizeof(*ht));
 949         mutex_init(&ht->mutex);
 950         spin_lock_init(&ht->lock);
 951         memcpy(&ht->p, params, sizeof(*params));
 952
 953         if (params->min_size)
 954                 ht->p.min_size = roundup_pow_of_two(params->min_size);
 955
 956         /* Cap total entries at 2^31 to avoid nelems overflow. */
 957         ht->max_elems = 1u << 31;
 958
 959         if (params->max_size) {
 960                 ht->p.max_size = rounddown_pow_of_two(params->max_size);
 961                 if (ht->p.max_size < ht->max_elems / 2)
 962                         ht->max_elems = ht->p.max_size * 2;
 963         }
 964
 965         ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
 966
 967         if (params->nelem_hint)
 968                 size = rounded_hashtable_size(&ht->p);
 969
 970         if (params->locks_mul)
 971                 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
 972         else
 973                 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
 974
 975         ht->key_len = ht->p.key_len;
 976         if (!params->hashfn) {
 977                 ht->p.hashfn = jhash;
 978
 979                 if (!(ht->key_len & (sizeof(u32) - 1))) {
 980                         ht->key_len /= sizeof(u32);
 981                         ht->p.hashfn = rhashtable_jhash2;
 982                 }
 983         }
 984
 985         tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
 986         if (tbl == NULL)
 987                 return -ENOMEM;
 988
 989         atomic_set(&ht->nelems, 0);
 990
 991         RCU_INIT_POINTER(ht->tbl, tbl);
 992
 993         INIT_WORK(&ht->run_work, rht_deferred_worker);
 994
 995         return 0;
 996 }
 997 EXPORT_SYMBOL_GPL(rhashtable_init);
 998
 999 /**
1000  * rhltable_init - initialize a new hash list table
1001  * @hlt:        hash list table to be initialized
1002  * @params:     configuration parameters
1003  *
1004  * Initializes a new hash list table.
1005  *
1006  * See documentation for rhashtable_init.
1007  */
1008 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1009 {
1010         int err;
1011
1012         /* No rhlist NULLs marking for now. */
1013         if (params->nulls_base)
1014                 return -EINVAL;
1015
1016         err = rhashtable_init(&hlt->ht, params);
1017         hlt->ht.rhlist = true;
1018         return err;
1019 }
1020 EXPORT_SYMBOL_GPL(rhltable_init);
1021
1022 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1023                                 void (*free_fn)(void *ptr, void *arg),
1024                                 void *arg)
1025 {
1026         struct rhlist_head *list;
1027
1028         if (!ht->rhlist) {
1029                 free_fn(rht_obj(ht, obj), arg);
1030                 return;
1031         }
1032
1033         list = container_of(obj, struct rhlist_head, rhead);
1034         do {
1035                 obj = &list->rhead;
1036                 list = rht_dereference(list->next, ht);
1037                 free_fn(rht_obj(ht, obj), arg);
1038         } while (list);
1039 }
1040
1041 /**
1042  * rhashtable_free_and_destroy - free elements and destroy hash table
1043  * @ht:         the hash table to destroy
1044  * @free_fn:    callback to release resources of element
1045  * @arg:        pointer passed to free_fn
1046  *
1047  * Stops an eventual async resize. If defined, invokes free_fn for each
1048  * element to releasal resources. Please note that RCU protected
1049  * readers may still be accessing the elements. Releasing of resources
1050  * must occur in a compatible manner. Then frees the bucket array.
1051  *
1052  * This function will eventually sleep to wait for an async resize
1053  * to complete. The caller is responsible that no further write operations
1054  * occurs in parallel.
1055  */
1056 void rhashtable_free_and_destroy(struct rhashtable *ht,
1057                                  void (*free_fn)(void *ptr, void *arg),
1058                                  void *arg)
1059 {
1060         struct bucket_table *tbl;
1061         unsigned int i;
1062
1063         cancel_work_sync(&ht->run_work);
1064
1065         mutex_lock(&ht->mutex);
1066         tbl = rht_dereference(ht->tbl, ht);
1067         if (free_fn) {
1068                 for (i = 0; i < tbl->size; i++) {
1069                         struct rhash_head *pos, *next;
1070
1071                         for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1072                              next = !rht_is_a_nulls(pos) ?
1073                                         rht_dereference(pos->next, ht) : NULL;
1074                              !rht_is_a_nulls(pos);
1075                              pos = next,
1076                              next = !rht_is_a_nulls(pos) ?
1077                                         rht_dereference(pos->next, ht) : NULL)
1078                                 rhashtable_free_one(ht, pos, free_fn, arg);
1079                 }
1080         }
1081
1082         bucket_table_free(tbl);
1083         mutex_unlock(&ht->mutex);
1084 }
1085 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1086
1087 void rhashtable_destroy(struct rhashtable *ht)
1088 {
1089         return rhashtable_free_and_destroy(ht, NULL, NULL);
1090 }
1091 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1092
1093 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1094                                             unsigned int hash)
1095 {
1096         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1097         static struct rhash_head __rcu *rhnull =
1098                 (struct rhash_head __rcu *)NULLS_MARKER(0);
1099         unsigned int index = hash & ((1 << tbl->nest) - 1);
1100         unsigned int size = tbl->size >> tbl->nest;
1101         unsigned int subhash = hash;
1102         union nested_table *ntbl;
1103
1104         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1105         ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1106         subhash >>= tbl->nest;
1107
1108         while (ntbl && size > (1 << shift)) {
1109                 index = subhash & ((1 << shift) - 1);
1110                 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1111                                                   tbl, hash);
1112                 size >>= shift;
1113                 subhash >>= shift;
1114         }
1115
1116         if (!ntbl)
1117                 return &rhnull;
1118
1119         return &ntbl[subhash].bucket;
1120
1121 }
1122 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1123
1124 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1125                                                    struct bucket_table *tbl,
1126                                                    unsigned int hash)
1127 {
1128         const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1129         unsigned int index = hash & ((1 << tbl->nest) - 1);
1130         unsigned int size = tbl->size >> tbl->nest;
1131         union nested_table *ntbl;
1132         unsigned int shifted;
1133         unsigned int nhash;
1134
1135         ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1136         hash >>= tbl->nest;
1137         nhash = index;
1138         shifted = tbl->nest;
1139         ntbl = nested_table_alloc(ht, &ntbl[index].table,
1140                                   size <= (1 << shift) ? shifted : 0, nhash);
1141
1142         while (ntbl && size > (1 << shift)) {
1143                 index = hash & ((1 << shift) - 1);
1144                 size >>= shift;
1145                 hash >>= shift;
1146                 nhash |= index << shifted;
1147                 shifted += shift;
1148                 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1149                                           size <= (1 << shift) ? shifted : 0,
1150                                           nhash);
1151         }
1152
1153         if (!ntbl)
1154                 return NULL;
1155
1156         return &ntbl[hash].bucket;
1157
1158 }
1159 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);