libcfs/libcfs/hash.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * libcfs/libcfs/hash.c
  37  *
  38  * Implement a hash class for hash process in lustre system.
  39  *
  40  * Author: YuZhangyong <yzy@clusterfs.com>
  41  *
  42  * 2008-08-15: Brian Behlendorf <behlendorf1@llnl.gov>
  43  * - Simplified API and improved documentation
  44  * - Added per-hash feature flags:
  45  *   * CFS_HASH_DEBUG additional validation
  46  *   * CFS_HASH_REHASH dynamic rehashing
  47  * - Added per-hash statistics
  48  * - General performance enhancements
  49  *
  50  * 2009-07-31: Liang Zhen <zhen.liang@sun.com>
  51  * - move all stuff to libcfs
  52  * - don't allow cur_bits != max_bits without setting of CFS_HASH_REHASH
  53  * - ignore hs_rwlock if without CFS_HASH_REHASH setting
  54  * - buckets are allocated one by one(instead of contiguous memory),
  55  *   to avoid unnecessary cacheline conflict
  56  *
  57  * 2010-03-01: Liang Zhen <zhen.liang@sun.com>
  58  * - "bucket" is a group of hlist_head now, user can specify bucket size
  59  *   by bkt_bits of cfs_hash_create(), all hlist_heads in a bucket share
  60  *   one lock for reducing memory overhead.
  61  *
  62  * - support lockless hash, caller will take care of locks:
  63  *   avoid lock overhead for hash tables that are already protected
  64  *   by locking in the caller for another reason
  65  *
  66  * - support both spin_lock/rwlock for bucket:
  67  *   overhead of spinlock contention is lower than read/write
  68  *   contention of rwlock, so using spinlock to serialize operations on
  69  *   bucket is more reasonable for those frequently changed hash tables
  70  *
  71  * - support one-single lock mode:
  72  *   one lock to protect all hash operations to avoid overhead of
  73  *   multiple locks if hash table is always small
  74  *
  75  * - removed a lot of unnecessary addref & decref on hash element:
  76  *   addref & decref are atomic operations in many use-cases which
  77  *   are expensive.
  78  *
  79  * - support non-blocking cfs_hash_add() and cfs_hash_findadd():
  80  *   some lustre use-cases require these functions to be strictly
  81  *   non-blocking, we need to schedule required rehash on a different
  82  *   thread on those cases.
  83  *
  84  * - safer rehash on large hash table
  85  *   In old implementation, rehash function will exclusively lock the
  86  *   hash table and finish rehash in one batch, it's dangerous on SMP
  87  *   system because rehash millions of elements could take long time.
  88  *   New implemented rehash can release lock and relax CPU in middle
  89  *   of rehash, it's safe for another thread to search/change on the
  90  *   hash table even it's in rehasing.
  91  *
  92  * - support two different refcount modes
  93  *   . hash table has refcount on element
  94  *   . hash table doesn't change refcount on adding/removing element
  95  *
  96  * - support long name hash table (for param-tree)
  97  *
  98  * - fix a bug for cfs_hash_rehash_key:
  99  *   in old implementation, cfs_hash_rehash_key could screw up the
 100  *   hash-table because @key is overwritten without any protection.
 101  *   Now we need user to define hs_keycpy for those rehash enabled
 102  *   hash tables, cfs_hash_rehash_key will overwrite hash-key
 103  *   inside lock by calling hs_keycpy.
 104  *
 105  * - better hash iteration:
 106  *   Now we support both locked iteration & lockless iteration of hash
 107  *   table. Also, user can break the iteration by return 1 in callback.
 108  */
 109 #include <linux/seq_file.h>
 110
 111 #include <libcfs/libcfs.h>
 112
 113 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
 114 static unsigned int warn_on_depth = 8;
 115 module_param(warn_on_depth, uint, 0644);
 116 MODULE_PARM_DESC(warn_on_depth, "warning when hash depth is high.");
 117 #endif
 118
 119 struct cfs_wi_sched *cfs_sched_rehash;
 120
 121 static inline void
 122 cfs_hash_nl_lock(union cfs_hash_lock *lock, int exclusive) {}
 123
 124 static inline void
 125 cfs_hash_nl_unlock(union cfs_hash_lock *lock, int exclusive) {}
 126
 127 static inline void
 128 cfs_hash_spin_lock(union cfs_hash_lock *lock, int exclusive)
 129         __acquires(&lock->spin)
 130 {
 131         spin_lock(&lock->spin);
 132 }
 133
 134 static inline void
 135 cfs_hash_spin_unlock(union cfs_hash_lock *lock, int exclusive)
 136         __releases(&lock->spin)
 137 {
 138         spin_unlock(&lock->spin);
 139 }
 140
 141 static inline void
 142 cfs_hash_rw_lock(union cfs_hash_lock *lock, int exclusive)
 143         __acquires(&lock->rw)
 144 {
 145         if (!exclusive)
 146                 read_lock(&lock->rw);
 147         else
 148                 write_lock(&lock->rw);
 149 }
 150
 151 static inline void
 152 cfs_hash_rw_unlock(union cfs_hash_lock *lock, int exclusive)
 153         __releases(&lock->rw)
 154 {
 155         if (!exclusive)
 156                 read_unlock(&lock->rw);
 157         else
 158                 write_unlock(&lock->rw);
 159 }
 160
 161 /** No lock hash */
 162 static struct cfs_hash_lock_ops cfs_hash_nl_lops = {
 163         .hs_lock        = cfs_hash_nl_lock,
 164         .hs_unlock      = cfs_hash_nl_unlock,
 165         .hs_bkt_lock    = cfs_hash_nl_lock,
 166         .hs_bkt_unlock  = cfs_hash_nl_unlock,
 167 };
 168
 169 /** no bucket lock, one spinlock to protect everything */
 170 static struct cfs_hash_lock_ops cfs_hash_nbl_lops = {
 171         .hs_lock        = cfs_hash_spin_lock,
 172         .hs_unlock      = cfs_hash_spin_unlock,
 173         .hs_bkt_lock    = cfs_hash_nl_lock,
 174         .hs_bkt_unlock  = cfs_hash_nl_unlock,
 175 };
 176
 177 /** spin bucket lock, rehash is enabled */
 178 static struct cfs_hash_lock_ops cfs_hash_bkt_spin_lops = {
 179         .hs_lock        = cfs_hash_rw_lock,
 180         .hs_unlock      = cfs_hash_rw_unlock,
 181         .hs_bkt_lock    = cfs_hash_spin_lock,
 182         .hs_bkt_unlock  = cfs_hash_spin_unlock,
 183 };
 184
 185 /** rw bucket lock, rehash is enabled */
 186 static struct cfs_hash_lock_ops cfs_hash_bkt_rw_lops = {
 187         .hs_lock        = cfs_hash_rw_lock,
 188         .hs_unlock      = cfs_hash_rw_unlock,
 189         .hs_bkt_lock    = cfs_hash_rw_lock,
 190         .hs_bkt_unlock  = cfs_hash_rw_unlock,
 191 };
 192
 193 /** spin bucket lock, rehash is disabled */
 194 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_spin_lops = {
 195         .hs_lock        = cfs_hash_nl_lock,
 196         .hs_unlock      = cfs_hash_nl_unlock,
 197         .hs_bkt_lock    = cfs_hash_spin_lock,
 198         .hs_bkt_unlock  = cfs_hash_spin_unlock,
 199 };
 200
 201 /** rw bucket lock, rehash is disabled */
 202 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_rw_lops = {
 203         .hs_lock        = cfs_hash_nl_lock,
 204         .hs_unlock      = cfs_hash_nl_unlock,
 205         .hs_bkt_lock    = cfs_hash_rw_lock,
 206         .hs_bkt_unlock  = cfs_hash_rw_unlock,
 207 };
 208
 209 static void
 210 cfs_hash_lock_setup(struct cfs_hash *hs)
 211 {
 212         if (cfs_hash_with_no_lock(hs)) {
 213                 hs->hs_lops = &cfs_hash_nl_lops;
 214
 215         } else if (cfs_hash_with_no_bktlock(hs)) {
 216                 hs->hs_lops = &cfs_hash_nbl_lops;
 217                 spin_lock_init(&hs->hs_lock.spin);
 218
 219         } else if (cfs_hash_with_rehash(hs)) {
 220                 rwlock_init(&hs->hs_lock.rw);
 221
 222                 if (cfs_hash_with_rw_bktlock(hs))
 223                         hs->hs_lops = &cfs_hash_bkt_rw_lops;
 224                 else if (cfs_hash_with_spin_bktlock(hs))
 225                         hs->hs_lops = &cfs_hash_bkt_spin_lops;
 226                 else
 227                         LBUG();
 228         } else {
 229                 if (cfs_hash_with_rw_bktlock(hs))
 230                         hs->hs_lops = &cfs_hash_nr_bkt_rw_lops;
 231                 else if (cfs_hash_with_spin_bktlock(hs))
 232                         hs->hs_lops = &cfs_hash_nr_bkt_spin_lops;
 233                 else
 234                         LBUG();
 235         }
 236 }
 237
 238 /**
 239  * Simple hash head without depth tracking
 240  * new element is always added to head of hlist
 241  */
 242 struct cfs_hash_head {
 243         struct hlist_head       hh_head;        /**< entries list */
 244 };
 245
 246 static int
 247 cfs_hash_hh_hhead_size(struct cfs_hash *hs)
 248 {
 249         return sizeof(struct cfs_hash_head);
 250 }
 251
 252 static struct hlist_head *
 253 cfs_hash_hh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
 254 {
 255         struct cfs_hash_head *head;
 256
 257         head = (struct cfs_hash_head *)&bd->bd_bucket->hsb_head[0];
 258         return &head[bd->bd_offset].hh_head;
 259 }
 260
 261 static int
 262 cfs_hash_hh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 263                       struct hlist_node *hnode)
 264 {
 265         hlist_add_head(hnode, cfs_hash_hh_hhead(hs, bd));
 266         return -1; /* unknown depth */
 267 }
 268
 269 static int
 270 cfs_hash_hh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 271                       struct hlist_node *hnode)
 272 {
 273         hlist_del_init(hnode);
 274         return -1; /* unknown depth */
 275 }
 276
 277 /**
 278  * Simple hash head with depth tracking
 279  * new element is always added to head of hlist
 280  */
 281 struct cfs_hash_head_dep {
 282         struct hlist_head       hd_head;        /**< entries list */
 283         unsigned int            hd_depth;       /**< list length */
 284 };
 285
 286 static int
 287 cfs_hash_hd_hhead_size(struct cfs_hash *hs)
 288 {
 289         return sizeof(struct cfs_hash_head_dep);
 290 }
 291
 292 static struct hlist_head *
 293 cfs_hash_hd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
 294 {
 295         struct cfs_hash_head_dep   *head;
 296
 297         head = (struct cfs_hash_head_dep *)&bd->bd_bucket->hsb_head[0];
 298         return &head[bd->bd_offset].hd_head;
 299 }
 300
 301 static int
 302 cfs_hash_hd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 303                       struct hlist_node *hnode)
 304 {
 305         struct cfs_hash_head_dep *hh;
 306
 307         hh = container_of(cfs_hash_hd_hhead(hs, bd),
 308                           struct cfs_hash_head_dep, hd_head);
 309         hlist_add_head(hnode, &hh->hd_head);
 310         return ++hh->hd_depth;
 311 }
 312
 313 static int
 314 cfs_hash_hd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 315                       struct hlist_node *hnode)
 316 {
 317         struct cfs_hash_head_dep *hh;
 318
 319         hh = container_of(cfs_hash_hd_hhead(hs, bd),
 320                           struct cfs_hash_head_dep, hd_head);
 321         hlist_del_init(hnode);
 322         return --hh->hd_depth;
 323 }
 324
 325 /**
 326  * double links hash head without depth tracking
 327  * new element is always added to tail of hlist
 328  */
 329 struct cfs_hash_dhead {
 330         struct hlist_head       dh_head;        /**< entries list */
 331         struct hlist_node       *dh_tail;       /**< the last entry */
 332 };
 333
 334 static int
 335 cfs_hash_dh_hhead_size(struct cfs_hash *hs)
 336 {
 337         return sizeof(struct cfs_hash_dhead);
 338 }
 339
 340 static struct hlist_head *
 341 cfs_hash_dh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
 342 {
 343         struct cfs_hash_dhead *head;
 344
 345         head = (struct cfs_hash_dhead *)&bd->bd_bucket->hsb_head[0];
 346         return &head[bd->bd_offset].dh_head;
 347 }
 348
 349 static int
 350 cfs_hash_dh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 351                       struct hlist_node *hnode)
 352 {
 353         struct cfs_hash_dhead *dh;
 354
 355         dh = container_of(cfs_hash_dh_hhead(hs, bd),
 356                           struct cfs_hash_dhead, dh_head);
 357         if (dh->dh_tail != NULL) /* not empty */
 358                 hlist_add_behind(hnode, dh->dh_tail);
 359         else /* empty list */
 360                 hlist_add_head(hnode, &dh->dh_head);
 361         dh->dh_tail = hnode;
 362         return -1; /* unknown depth */
 363 }
 364
 365 static int
 366 cfs_hash_dh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 367                       struct hlist_node *hnd)
 368 {
 369         struct cfs_hash_dhead *dh;
 370
 371         dh = container_of(cfs_hash_dh_hhead(hs, bd),
 372                           struct cfs_hash_dhead, dh_head);
 373         if (hnd->next == NULL) { /* it's the tail */
 374                 dh->dh_tail = (hnd->pprev == &dh->dh_head.first) ? NULL :
 375                               container_of(hnd->pprev, struct hlist_node, next);
 376         }
 377         hlist_del_init(hnd);
 378         return -1; /* unknown depth */
 379 }
 380
 381 /**
 382  * double links hash head with depth tracking
 383  * new element is always added to tail of hlist
 384  */
 385 struct cfs_hash_dhead_dep {
 386         struct hlist_head       dd_head;        /**< entries list */
 387         struct hlist_node       *dd_tail;       /**< the last entry */
 388         unsigned int            dd_depth;       /**< list length */
 389 };
 390
 391 static int
 392 cfs_hash_dd_hhead_size(struct cfs_hash *hs)
 393 {
 394         return sizeof(struct cfs_hash_dhead_dep);
 395 }
 396
 397 static struct hlist_head *
 398 cfs_hash_dd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
 399 {
 400         struct cfs_hash_dhead_dep *head;
 401
 402         head = (struct cfs_hash_dhead_dep *)&bd->bd_bucket->hsb_head[0];
 403         return &head[bd->bd_offset].dd_head;
 404 }
 405
 406 static int
 407 cfs_hash_dd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 408                       struct hlist_node *hnode)
 409 {
 410         struct cfs_hash_dhead_dep *dh;
 411
 412         dh = container_of(cfs_hash_dd_hhead(hs, bd),
 413                           struct cfs_hash_dhead_dep, dd_head);
 414         if (dh->dd_tail != NULL) /* not empty */
 415                 hlist_add_behind(hnode, dh->dd_tail);
 416         else /* empty list */
 417                 hlist_add_head(hnode, &dh->dd_head);
 418         dh->dd_tail = hnode;
 419         return ++dh->dd_depth;
 420 }
 421
 422 static int
 423 cfs_hash_dd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 424                       struct hlist_node *hnd)
 425 {
 426         struct cfs_hash_dhead_dep *dh;
 427
 428         dh = container_of(cfs_hash_dd_hhead(hs, bd),
 429                           struct cfs_hash_dhead_dep, dd_head);
 430         if (hnd->next == NULL) { /* it's the tail */
 431                 dh->dd_tail = (hnd->pprev == &dh->dd_head.first) ? NULL :
 432                               container_of(hnd->pprev, struct hlist_node, next);
 433         }
 434         hlist_del_init(hnd);
 435         return --dh->dd_depth;
 436 }
 437
 438 static struct cfs_hash_hlist_ops cfs_hash_hh_hops = {
 439        .hop_hhead      = cfs_hash_hh_hhead,
 440        .hop_hhead_size = cfs_hash_hh_hhead_size,
 441        .hop_hnode_add  = cfs_hash_hh_hnode_add,
 442        .hop_hnode_del  = cfs_hash_hh_hnode_del,
 443 };
 444
 445 static struct cfs_hash_hlist_ops cfs_hash_hd_hops = {
 446        .hop_hhead      = cfs_hash_hd_hhead,
 447        .hop_hhead_size = cfs_hash_hd_hhead_size,
 448        .hop_hnode_add  = cfs_hash_hd_hnode_add,
 449        .hop_hnode_del  = cfs_hash_hd_hnode_del,
 450 };
 451
 452 static struct cfs_hash_hlist_ops cfs_hash_dh_hops = {
 453        .hop_hhead      = cfs_hash_dh_hhead,
 454        .hop_hhead_size = cfs_hash_dh_hhead_size,
 455        .hop_hnode_add  = cfs_hash_dh_hnode_add,
 456        .hop_hnode_del  = cfs_hash_dh_hnode_del,
 457 };
 458
 459 static struct cfs_hash_hlist_ops cfs_hash_dd_hops = {
 460        .hop_hhead      = cfs_hash_dd_hhead,
 461        .hop_hhead_size = cfs_hash_dd_hhead_size,
 462        .hop_hnode_add  = cfs_hash_dd_hnode_add,
 463        .hop_hnode_del  = cfs_hash_dd_hnode_del,
 464 };
 465
 466 static void
 467 cfs_hash_hlist_setup(struct cfs_hash *hs)
 468 {
 469         if (cfs_hash_with_add_tail(hs)) {
 470                 hs->hs_hops = cfs_hash_with_depth(hs) ?
 471                               &cfs_hash_dd_hops : &cfs_hash_dh_hops;
 472         } else {
 473                 hs->hs_hops = cfs_hash_with_depth(hs) ?
 474                               &cfs_hash_hd_hops : &cfs_hash_hh_hops;
 475         }
 476 }
 477
 478 static void
 479 cfs_hash_bd_from_key(struct cfs_hash *hs, struct cfs_hash_bucket **bkts,
 480                      unsigned int bits, const void *key, struct cfs_hash_bd *bd)
 481 {
 482         unsigned int index = cfs_hash_id(hs, key, (1U << bits) - 1);
 483
 484         LASSERT(bits == hs->hs_cur_bits || bits == hs->hs_rehash_bits);
 485
 486         bd->bd_bucket = bkts[index & ((1U << (bits - hs->hs_bkt_bits)) - 1)];
 487         bd->bd_offset = index >> (bits - hs->hs_bkt_bits);
 488 }
 489
 490 void
 491 cfs_hash_bd_get(struct cfs_hash *hs, const void *key, struct cfs_hash_bd *bd)
 492 {
 493         /* NB: caller should hold hs->hs_rwlock if REHASH is set */
 494         if (likely(hs->hs_rehash_buckets == NULL)) {
 495                 cfs_hash_bd_from_key(hs, hs->hs_buckets,
 496                                      hs->hs_cur_bits, key, bd);
 497         } else {
 498                 LASSERT(hs->hs_rehash_bits != 0);
 499                 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
 500                                      hs->hs_rehash_bits, key, bd);
 501         }
 502 }
 503 EXPORT_SYMBOL(cfs_hash_bd_get);
 504
 505 static inline void
 506 cfs_hash_bd_dep_record(struct cfs_hash *hs, struct cfs_hash_bd *bd, int dep_cur)
 507 {
 508         if (likely(dep_cur <= bd->bd_bucket->hsb_depmax))
 509                 return;
 510
 511         bd->bd_bucket->hsb_depmax = dep_cur;
 512 # if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
 513         if (likely(warn_on_depth == 0 ||
 514                    max(warn_on_depth, hs->hs_dep_max) >= dep_cur))
 515                 return;
 516
 517         spin_lock(&hs->hs_dep_lock);
 518         hs->hs_dep_max  = dep_cur;
 519         hs->hs_dep_bkt  = bd->bd_bucket->hsb_index;
 520         hs->hs_dep_off  = bd->bd_offset;
 521         hs->hs_dep_bits = hs->hs_cur_bits;
 522         spin_unlock(&hs->hs_dep_lock);
 523
 524         cfs_wi_schedule(cfs_sched_rehash, &hs->hs_dep_wi);
 525 # endif
 526 }
 527
 528 void
 529 cfs_hash_bd_add_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 530                         struct hlist_node *hnode)
 531 {
 532         int rc;
 533
 534         rc = hs->hs_hops->hop_hnode_add(hs, bd, hnode);
 535         cfs_hash_bd_dep_record(hs, bd, rc);
 536         bd->bd_bucket->hsb_version++;
 537         if (unlikely(bd->bd_bucket->hsb_version == 0))
 538                 bd->bd_bucket->hsb_version++;
 539         bd->bd_bucket->hsb_count++;
 540
 541         if (cfs_hash_with_counter(hs))
 542                 atomic_inc(&hs->hs_count);
 543         if (!cfs_hash_with_no_itemref(hs))
 544                 cfs_hash_get(hs, hnode);
 545 }
 546 EXPORT_SYMBOL(cfs_hash_bd_add_locked);
 547
 548 void
 549 cfs_hash_bd_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 550                        struct hlist_node *hnode)
 551 {
 552         hs->hs_hops->hop_hnode_del(hs, bd, hnode);
 553
 554         LASSERT(bd->bd_bucket->hsb_count > 0);
 555         bd->bd_bucket->hsb_count--;
 556         bd->bd_bucket->hsb_version++;
 557         if (unlikely(bd->bd_bucket->hsb_version == 0))
 558                 bd->bd_bucket->hsb_version++;
 559
 560         if (cfs_hash_with_counter(hs)) {
 561                 LASSERT(atomic_read(&hs->hs_count) > 0);
 562                 atomic_dec(&hs->hs_count);
 563         }
 564         if (!cfs_hash_with_no_itemref(hs))
 565                 cfs_hash_put_locked(hs, hnode);
 566 }
 567 EXPORT_SYMBOL(cfs_hash_bd_del_locked);
 568
 569 void
 570 cfs_hash_bd_move_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd_old,
 571                         struct cfs_hash_bd *bd_new, struct hlist_node *hnode)
 572 {
 573         struct cfs_hash_bucket *obkt = bd_old->bd_bucket;
 574         struct cfs_hash_bucket *nbkt = bd_new->bd_bucket;
 575         int                rc;
 576
 577         if (cfs_hash_bd_compare(bd_old, bd_new) == 0)
 578                 return;
 579
 580         /* use cfs_hash_bd_hnode_add/del, to avoid atomic & refcount ops
 581          * in cfs_hash_bd_del/add_locked */
 582         hs->hs_hops->hop_hnode_del(hs, bd_old, hnode);
 583         rc = hs->hs_hops->hop_hnode_add(hs, bd_new, hnode);
 584         cfs_hash_bd_dep_record(hs, bd_new, rc);
 585
 586         LASSERT(obkt->hsb_count > 0);
 587         obkt->hsb_count--;
 588         obkt->hsb_version++;
 589         if (unlikely(obkt->hsb_version == 0))
 590                 obkt->hsb_version++;
 591         nbkt->hsb_count++;
 592         nbkt->hsb_version++;
 593         if (unlikely(nbkt->hsb_version == 0))
 594                 nbkt->hsb_version++;
 595 }
 596
 597 enum {
 598         /** always set, for sanity (avoid ZERO intent) */
 599         CFS_HS_LOOKUP_MASK_FIND     = 1 << 0,
 600         /** return entry with a ref */
 601         CFS_HS_LOOKUP_MASK_REF      = 1 << 1,
 602         /** add entry if not existing */
 603         CFS_HS_LOOKUP_MASK_ADD      = 1 << 2,
 604         /** delete entry, ignore other masks */
 605         CFS_HS_LOOKUP_MASK_DEL      = 1 << 3,
 606 };
 607
 608 enum cfs_hash_lookup_intent {
 609         /** return item w/o refcount */
 610         CFS_HS_LOOKUP_IT_PEEK       = CFS_HS_LOOKUP_MASK_FIND,
 611         /** return item with refcount */
 612         CFS_HS_LOOKUP_IT_FIND       = (CFS_HS_LOOKUP_MASK_FIND |
 613                                        CFS_HS_LOOKUP_MASK_REF),
 614         /** return item w/o refcount if existed, otherwise add */
 615         CFS_HS_LOOKUP_IT_ADD        = (CFS_HS_LOOKUP_MASK_FIND |
 616                                        CFS_HS_LOOKUP_MASK_ADD),
 617         /** return item with refcount if existed, otherwise add */
 618         CFS_HS_LOOKUP_IT_FINDADD    = (CFS_HS_LOOKUP_IT_FIND |
 619                                        CFS_HS_LOOKUP_MASK_ADD),
 620         /** delete if existed */
 621         CFS_HS_LOOKUP_IT_FINDDEL    = (CFS_HS_LOOKUP_MASK_FIND |
 622                                        CFS_HS_LOOKUP_MASK_DEL)
 623 };
 624
 625 static struct hlist_node *
 626 cfs_hash_bd_lookup_intent(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 627                           const void *key, struct hlist_node *hnode,
 628                           enum cfs_hash_lookup_intent intent)
 629
 630 {
 631         struct hlist_head  *hhead = cfs_hash_bd_hhead(hs, bd);
 632         struct hlist_node  *ehnode;
 633         struct hlist_node  *match;
 634         int  intent_add = (intent & CFS_HS_LOOKUP_MASK_ADD) != 0;
 635
 636         /* with this function, we can avoid a lot of useless refcount ops,
 637          * which are expensive atomic operations most time. */
 638         match = intent_add ? NULL : hnode;
 639         hlist_for_each(ehnode, hhead) {
 640                 if (!cfs_hash_keycmp(hs, key, ehnode))
 641                         continue;
 642
 643                 if (match != NULL && match != ehnode) /* can't match */
 644                         continue;
 645
 646                 /* match and ... */
 647                 if ((intent & CFS_HS_LOOKUP_MASK_DEL) != 0) {
 648                         cfs_hash_bd_del_locked(hs, bd, ehnode);
 649                         return ehnode;
 650                 }
 651
 652                 /* caller wants refcount? */
 653                 if ((intent & CFS_HS_LOOKUP_MASK_REF) != 0)
 654                         cfs_hash_get(hs, ehnode);
 655                 return ehnode;
 656         }
 657         /* no match item */
 658         if (!intent_add)
 659                 return NULL;
 660
 661         LASSERT(hnode != NULL);
 662         cfs_hash_bd_add_locked(hs, bd, hnode);
 663         return hnode;
 664 }
 665
 666 struct hlist_node *
 667 cfs_hash_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 668                           const void *key)
 669 {
 670         return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
 671                                         CFS_HS_LOOKUP_IT_FIND);
 672 }
 673 EXPORT_SYMBOL(cfs_hash_bd_lookup_locked);
 674
 675 struct hlist_node *
 676 cfs_hash_bd_peek_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 677                         const void *key)
 678 {
 679         return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
 680                                         CFS_HS_LOOKUP_IT_PEEK);
 681 }
 682 EXPORT_SYMBOL(cfs_hash_bd_peek_locked);
 683
 684 struct hlist_node *
 685 cfs_hash_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 686                            const void *key, struct hlist_node *hnode,
 687                            int noref)
 688 {
 689         return cfs_hash_bd_lookup_intent(hs, bd, key, hnode,
 690                                         CFS_HS_LOOKUP_IT_ADD |
 691                                         (!noref * CFS_HS_LOOKUP_MASK_REF));
 692 }
 693
 694 struct hlist_node *
 695 cfs_hash_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 696                            const void *key, struct hlist_node *hnode)
 697 {
 698         /* hnode can be NULL, we find the first item with @key */
 699         return cfs_hash_bd_lookup_intent(hs, bd, key, hnode,
 700                                         CFS_HS_LOOKUP_IT_FINDDEL);
 701 }
 702
 703 static void
 704 cfs_hash_multi_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 705                        unsigned n, int excl)
 706 {
 707         struct cfs_hash_bucket *prev = NULL;
 708         int                i;
 709
 710         /**
 711          * bds must be ascendantly ordered by bd->bd_bucket->hsb_index.
 712          * NB: it's possible that several bds point to the same bucket but
 713          * have different bd::bd_offset, so need take care of deadlock.
 714          */
 715         cfs_hash_for_each_bd(bds, n, i) {
 716                 if (prev == bds[i].bd_bucket)
 717                         continue;
 718
 719                 LASSERT(prev == NULL ||
 720                         prev->hsb_index < bds[i].bd_bucket->hsb_index);
 721                 cfs_hash_bd_lock(hs, &bds[i], excl);
 722                 prev = bds[i].bd_bucket;
 723         }
 724 }
 725
 726 static void
 727 cfs_hash_multi_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 728                          unsigned n, int excl)
 729 {
 730         struct cfs_hash_bucket *prev = NULL;
 731         int                i;
 732
 733         cfs_hash_for_each_bd(bds, n, i) {
 734                 if (prev != bds[i].bd_bucket) {
 735                         cfs_hash_bd_unlock(hs, &bds[i], excl);
 736                         prev = bds[i].bd_bucket;
 737                 }
 738         }
 739 }
 740
 741 static struct hlist_node *
 742 cfs_hash_multi_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 743                                 unsigned n, const void *key)
 744 {
 745         struct hlist_node *ehnode;
 746         unsigned          i;
 747
 748         cfs_hash_for_each_bd(bds, n, i) {
 749                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, NULL,
 750                                                         CFS_HS_LOOKUP_IT_FIND);
 751                 if (ehnode != NULL)
 752                         return ehnode;
 753         }
 754         return NULL;
 755 }
 756
 757 static struct hlist_node *
 758 cfs_hash_multi_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 759                                  unsigned n, const void *key,
 760                                  struct hlist_node *hnode, int noref)
 761 {
 762         struct hlist_node *ehnode;
 763         int               intent;
 764         unsigned          i;
 765
 766         LASSERT(hnode != NULL);
 767         intent = CFS_HS_LOOKUP_IT_PEEK | (!noref * CFS_HS_LOOKUP_MASK_REF);
 768
 769         cfs_hash_for_each_bd(bds, n, i) {
 770                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key,
 771                                                    NULL, intent);
 772                 if (ehnode != NULL)
 773                         return ehnode;
 774         }
 775
 776         if (i == 1) { /* only one bucket */
 777                 cfs_hash_bd_add_locked(hs, &bds[0], hnode);
 778         } else {
 779                 struct cfs_hash_bd      mybd;
 780
 781                 cfs_hash_bd_get(hs, key, &mybd);
 782                 cfs_hash_bd_add_locked(hs, &mybd, hnode);
 783         }
 784
 785         return hnode;
 786 }
 787
 788 static struct hlist_node *
 789 cfs_hash_multi_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 790                                  unsigned n, const void *key,
 791                                  struct hlist_node *hnode)
 792 {
 793         struct hlist_node *ehnode;
 794         unsigned           i;
 795
 796         cfs_hash_for_each_bd(bds, n, i) {
 797                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, hnode,
 798                                                    CFS_HS_LOOKUP_IT_FINDDEL);
 799                 if (ehnode != NULL)
 800                         return ehnode;
 801         }
 802         return NULL;
 803 }
 804
 805 static void
 806 cfs_hash_bd_order(struct cfs_hash_bd *bd1, struct cfs_hash_bd *bd2)
 807 {
 808         int     rc;
 809
 810         if (bd2->bd_bucket == NULL)
 811                 return;
 812
 813         if (bd1->bd_bucket == NULL) {
 814                 *bd1 = *bd2;
 815                 bd2->bd_bucket = NULL;
 816                 return;
 817         }
 818
 819         rc = cfs_hash_bd_compare(bd1, bd2);
 820         if (rc == 0) {
 821                 bd2->bd_bucket = NULL;
 822
 823         } else if (rc > 0) { /* swab bd1 and bd2 */
 824                 struct cfs_hash_bd tmp;
 825
 826                 tmp = *bd2;
 827                 *bd2 = *bd1;
 828                 *bd1 = tmp;
 829         }
 830 }
 831
 832 void
 833 cfs_hash_dual_bd_get(struct cfs_hash *hs, const void *key,
 834                      struct cfs_hash_bd *bds)
 835 {
 836         /* NB: caller should hold hs_lock.rw if REHASH is set */
 837         cfs_hash_bd_from_key(hs, hs->hs_buckets,
 838                              hs->hs_cur_bits, key, &bds[0]);
 839         if (likely(hs->hs_rehash_buckets == NULL)) {
 840                 /* no rehash or not rehashing */
 841                 bds[1].bd_bucket = NULL;
 842                 return;
 843         }
 844
 845         LASSERT(hs->hs_rehash_bits != 0);
 846         cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
 847                              hs->hs_rehash_bits, key, &bds[1]);
 848
 849         cfs_hash_bd_order(&bds[0], &bds[1]);
 850 }
 851
 852 void
 853 cfs_hash_dual_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
 854 {
 855         cfs_hash_multi_bd_lock(hs, bds, 2, excl);
 856 }
 857
 858 void
 859 cfs_hash_dual_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
 860 {
 861         cfs_hash_multi_bd_unlock(hs, bds, 2, excl);
 862 }
 863
 864 struct hlist_node *
 865 cfs_hash_dual_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 866                                const void *key)
 867 {
 868         return cfs_hash_multi_bd_lookup_locked(hs, bds, 2, key);
 869 }
 870
 871 struct hlist_node *
 872 cfs_hash_dual_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 873                                 const void *key, struct hlist_node *hnode,
 874                                 int noref)
 875 {
 876         return cfs_hash_multi_bd_findadd_locked(hs, bds, 2, key,
 877                                                 hnode, noref);
 878 }
 879
 880 struct hlist_node *
 881 cfs_hash_dual_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 882                                 const void *key, struct hlist_node *hnode)
 883 {
 884         return cfs_hash_multi_bd_finddel_locked(hs, bds, 2, key, hnode);
 885 }
 886
 887 static void
 888 cfs_hash_buckets_free(struct cfs_hash_bucket **buckets,
 889                       int bkt_size, int prev_size, int size)
 890 {
 891         int     i;
 892
 893         for (i = prev_size; i < size; i++) {
 894                 if (buckets[i] != NULL)
 895                         LIBCFS_FREE(buckets[i], bkt_size);
 896         }
 897
 898         LIBCFS_FREE(buckets, sizeof(buckets[0]) * size);
 899 }
 900
 901 /*
 902  * Create or grow bucket memory. Return old_buckets if no allocation was
 903  * needed, the newly allocated buckets if allocation was needed and
 904  * successful, and NULL on error.
 905  */
 906 static struct cfs_hash_bucket **
 907 cfs_hash_buckets_realloc(struct cfs_hash *hs, struct cfs_hash_bucket **old_bkts,
 908                          unsigned int old_size, unsigned int new_size)
 909 {
 910         struct cfs_hash_bucket **new_bkts;
 911         int                 i;
 912
 913         LASSERT(old_size == 0 || old_bkts != NULL);
 914
 915         if (old_bkts != NULL && old_size == new_size)
 916                 return old_bkts;
 917
 918         LIBCFS_ALLOC(new_bkts, sizeof(new_bkts[0]) * new_size);
 919         if (new_bkts == NULL)
 920                 return NULL;
 921
 922         if (old_bkts != NULL) {
 923                 memcpy(new_bkts, old_bkts,
 924                        min(old_size, new_size) * sizeof(*old_bkts));
 925         }
 926
 927         for (i = old_size; i < new_size; i++) {
 928                 struct hlist_head *hhead;
 929                 struct cfs_hash_bd     bd;
 930
 931                 LIBCFS_ALLOC(new_bkts[i], cfs_hash_bkt_size(hs));
 932                 if (new_bkts[i] == NULL) {
 933                         cfs_hash_buckets_free(new_bkts, cfs_hash_bkt_size(hs),
 934                                               old_size, new_size);
 935                         return NULL;
 936                 }
 937
 938                 new_bkts[i]->hsb_index   = i;
 939                 new_bkts[i]->hsb_version = 1;  /* shouldn't be zero */
 940                 new_bkts[i]->hsb_depmax  = -1; /* unknown */
 941                 bd.bd_bucket = new_bkts[i];
 942                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead)
 943                         INIT_HLIST_HEAD(hhead);
 944
 945                 if (cfs_hash_with_no_lock(hs) ||
 946                     cfs_hash_with_no_bktlock(hs))
 947                         continue;
 948
 949                 if (cfs_hash_with_rw_bktlock(hs))
 950                         rwlock_init(&new_bkts[i]->hsb_lock.rw);
 951                 else if (cfs_hash_with_spin_bktlock(hs))
 952                         spin_lock_init(&new_bkts[i]->hsb_lock.spin);
 953                 else
 954                         LBUG(); /* invalid use-case */
 955         }
 956         return new_bkts;
 957 }
 958
 959 /**
 960  * Initialize new libcfs hash, where:
 961  * @name     - Descriptive hash name
 962  * @cur_bits - Initial hash table size, in bits
 963  * @max_bits - Maximum allowed hash table resize, in bits
 964  * @ops      - Registered hash table operations
 965  * @flags    - CFS_HASH_REHASH enable synamic hash resizing
 966  *           - CFS_HASH_SORT enable chained hash sort
 967  */
 968 static int cfs_hash_rehash_worker(cfs_workitem_t *wi);
 969
 970 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
 971 static int cfs_hash_dep_print(cfs_workitem_t *wi)
 972 {
 973         struct cfs_hash *hs = container_of(wi, struct cfs_hash, hs_dep_wi);
 974         int         dep;
 975         int         bkt;
 976         int         off;
 977         int         bits;
 978
 979         spin_lock(&hs->hs_dep_lock);
 980         dep  = hs->hs_dep_max;
 981         bkt  = hs->hs_dep_bkt;
 982         off  = hs->hs_dep_off;
 983         bits = hs->hs_dep_bits;
 984         spin_unlock(&hs->hs_dep_lock);
 985
 986         LCONSOLE_WARN("#### HASH %s (bits: %d): max depth %d at bucket %d/%d\n",
 987                       hs->hs_name, bits, dep, bkt, off);
 988         spin_lock(&hs->hs_dep_lock);
 989         hs->hs_dep_bits = 0; /* mark as workitem done */
 990         spin_unlock(&hs->hs_dep_lock);
 991         return 0;
 992 }
 993
 994 static void cfs_hash_depth_wi_init(struct cfs_hash *hs)
 995 {
 996         spin_lock_init(&hs->hs_dep_lock);
 997         cfs_wi_init(&hs->hs_dep_wi, hs, cfs_hash_dep_print);
 998 }
 999
1000 static void cfs_hash_depth_wi_cancel(struct cfs_hash *hs)
1001 {
1002         if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_dep_wi))
1003                 return;
1004
1005         spin_lock(&hs->hs_dep_lock);
1006         while (hs->hs_dep_bits != 0) {
1007                 spin_unlock(&hs->hs_dep_lock);
1008                 cond_resched();
1009                 spin_lock(&hs->hs_dep_lock);
1010         }
1011         spin_unlock(&hs->hs_dep_lock);
1012 }
1013
1014 #else /* CFS_HASH_DEBUG_LEVEL < CFS_HASH_DEBUG_1 */
1015
1016 static inline void cfs_hash_depth_wi_init(struct cfs_hash *hs) {}
1017 static inline void cfs_hash_depth_wi_cancel(struct cfs_hash *hs) {}
1018
1019 #endif /* CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1 */
1020
1021 struct cfs_hash *
1022 cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
1023                 unsigned bkt_bits, unsigned extra_bytes,
1024                 unsigned min_theta, unsigned max_theta,
1025                 struct cfs_hash_ops *ops, unsigned flags)
1026 {
1027         struct cfs_hash *hs;
1028         int         len;
1029
1030         ENTRY;
1031
1032         CLASSERT(CFS_HASH_THETA_BITS < 15);
1033
1034         LASSERT(name != NULL);
1035         LASSERT(ops != NULL);
1036         LASSERT(ops->hs_key);
1037         LASSERT(ops->hs_hash);
1038         LASSERT(ops->hs_object);
1039         LASSERT(ops->hs_keycmp);
1040         LASSERT(ops->hs_get != NULL);
1041         LASSERT(ops->hs_put_locked != NULL);
1042
1043         if ((flags & CFS_HASH_REHASH) != 0)
1044                 flags |= CFS_HASH_COUNTER; /* must have counter */
1045
1046         LASSERT(cur_bits > 0);
1047         LASSERT(cur_bits >= bkt_bits);
1048         LASSERT(max_bits >= cur_bits && max_bits < 31);
1049         LASSERT(ergo((flags & CFS_HASH_REHASH) == 0, cur_bits == max_bits));
1050         LASSERT(ergo((flags & CFS_HASH_REHASH) != 0,
1051                      (flags & CFS_HASH_NO_LOCK) == 0));
1052         LASSERT(ergo((flags & CFS_HASH_REHASH_KEY) != 0,
1053                       ops->hs_keycpy != NULL));
1054
1055         len = (flags & CFS_HASH_BIGNAME) == 0 ?
1056               CFS_HASH_NAME_LEN : CFS_HASH_BIGNAME_LEN;
1057         LIBCFS_ALLOC(hs, offsetof(struct cfs_hash, hs_name[len]));
1058         if (hs == NULL)
1059                 RETURN(NULL);
1060
1061         strlcpy(hs->hs_name, name, len);
1062         hs->hs_flags = flags;
1063
1064         atomic_set(&hs->hs_refcount, 1);
1065         atomic_set(&hs->hs_count, 0);
1066
1067         cfs_hash_lock_setup(hs);
1068         cfs_hash_hlist_setup(hs);
1069
1070         hs->hs_cur_bits = (__u8)cur_bits;
1071         hs->hs_min_bits = (__u8)cur_bits;
1072         hs->hs_max_bits = (__u8)max_bits;
1073         hs->hs_bkt_bits = (__u8)bkt_bits;
1074
1075         hs->hs_ops         = ops;
1076         hs->hs_extra_bytes = extra_bytes;
1077         hs->hs_rehash_bits = 0;
1078         cfs_wi_init(&hs->hs_rehash_wi, hs, cfs_hash_rehash_worker);
1079         cfs_hash_depth_wi_init(hs);
1080
1081         if (cfs_hash_with_rehash(hs))
1082                 __cfs_hash_set_theta(hs, min_theta, max_theta);
1083
1084         hs->hs_buckets = cfs_hash_buckets_realloc(hs, NULL, 0,
1085                                                   CFS_HASH_NBKT(hs));
1086         if (hs->hs_buckets != NULL)
1087                 return hs;
1088
1089         LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[len]));
1090         RETURN(NULL);
1091 }
1092 EXPORT_SYMBOL(cfs_hash_create);
1093
1094 /**
1095  * Cleanup libcfs hash @hs.
1096  */
1097 static void
1098 cfs_hash_destroy(struct cfs_hash *hs)
1099 {
1100         struct hlist_node     *hnode;
1101         struct hlist_node     *pos;
1102         struct cfs_hash_bd         bd;
1103         int                   i;
1104         ENTRY;
1105
1106         LASSERT(hs != NULL);
1107         LASSERT(!cfs_hash_is_exiting(hs) &&
1108                 !cfs_hash_is_iterating(hs));
1109
1110         /**
1111          * prohibit further rehashes, don't need any lock because
1112          * I'm the only (last) one can change it.
1113          */
1114         hs->hs_exiting = 1;
1115         if (cfs_hash_with_rehash(hs))
1116                 cfs_hash_rehash_cancel(hs);
1117
1118         cfs_hash_depth_wi_cancel(hs);
1119         /* rehash should be done/canceled */
1120         LASSERT(hs->hs_buckets != NULL &&
1121                 hs->hs_rehash_buckets == NULL);
1122
1123         cfs_hash_for_each_bucket(hs, &bd, i) {
1124                 struct hlist_head *hhead;
1125
1126                 LASSERT(bd.bd_bucket != NULL);
1127                 /* no need to take this lock, just for consistent code */
1128                 cfs_hash_bd_lock(hs, &bd, 1);
1129
1130                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1131                         hlist_for_each_safe(hnode, pos, hhead) {
1132                                         LASSERTF(!cfs_hash_with_assert_empty(hs),
1133                                         "hash %s bucket %u(%u) is not "
1134                                         " empty: %u items left\n",
1135                                         hs->hs_name, bd.bd_bucket->hsb_index,
1136                                         bd.bd_offset, bd.bd_bucket->hsb_count);
1137                                 /* can't assert key valicate, because we
1138                                  * can interrupt rehash */
1139                                 cfs_hash_bd_del_locked(hs, &bd, hnode);
1140                                 cfs_hash_exit(hs, hnode);
1141                         }
1142                 }
1143                 LASSERT(bd.bd_bucket->hsb_count == 0);
1144                 cfs_hash_bd_unlock(hs, &bd, 1);
1145                 cond_resched();
1146         }
1147
1148         LASSERT(atomic_read(&hs->hs_count) == 0);
1149
1150         cfs_hash_buckets_free(hs->hs_buckets, cfs_hash_bkt_size(hs),
1151                               0, CFS_HASH_NBKT(hs));
1152         i = cfs_hash_with_bigname(hs) ?
1153             CFS_HASH_BIGNAME_LEN : CFS_HASH_NAME_LEN;
1154         LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[i]));
1155
1156         EXIT;
1157 }
1158
1159 struct cfs_hash *cfs_hash_getref(struct cfs_hash *hs)
1160 {
1161         if (atomic_inc_not_zero(&hs->hs_refcount))
1162                 return hs;
1163         return NULL;
1164 }
1165 EXPORT_SYMBOL(cfs_hash_getref);
1166
1167 void cfs_hash_putref(struct cfs_hash *hs)
1168 {
1169         if (atomic_dec_and_test(&hs->hs_refcount))
1170                 cfs_hash_destroy(hs);
1171 }
1172 EXPORT_SYMBOL(cfs_hash_putref);
1173
1174 static inline int
1175 cfs_hash_rehash_bits(struct cfs_hash *hs)
1176 {
1177         if (cfs_hash_with_no_lock(hs) ||
1178             !cfs_hash_with_rehash(hs))
1179                 return -EOPNOTSUPP;
1180
1181         if (unlikely(cfs_hash_is_exiting(hs)))
1182                 return -ESRCH;
1183
1184         if (unlikely(cfs_hash_is_rehashing(hs)))
1185                 return -EALREADY;
1186
1187         if (unlikely(cfs_hash_is_iterating(hs)))
1188                 return -EAGAIN;
1189
1190         /* XXX: need to handle case with max_theta != 2.0
1191          *      and the case with min_theta != 0.5 */
1192         if ((hs->hs_cur_bits < hs->hs_max_bits) &&
1193             (__cfs_hash_theta(hs) > hs->hs_max_theta))
1194                 return hs->hs_cur_bits + 1;
1195
1196         if (!cfs_hash_with_shrink(hs))
1197                 return 0;
1198
1199         if ((hs->hs_cur_bits > hs->hs_min_bits) &&
1200             (__cfs_hash_theta(hs) < hs->hs_min_theta))
1201                 return hs->hs_cur_bits - 1;
1202
1203         return 0;
1204 }
1205
1206 /**
1207  * don't allow inline rehash if:
1208  * - user wants non-blocking change (add/del) on hash table
1209  * - too many elements
1210  */
1211 static inline int
1212 cfs_hash_rehash_inline(struct cfs_hash *hs)
1213 {
1214         return !cfs_hash_with_nblk_change(hs) &&
1215                atomic_read(&hs->hs_count) < CFS_HASH_LOOP_HOG;
1216 }
1217
1218 /**
1219  * Add item @hnode to libcfs hash @hs using @key.  The registered
1220  * ops->hs_get function will be called when the item is added.
1221  */
1222 void
1223 cfs_hash_add(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1224 {
1225         struct cfs_hash_bd   bd;
1226         int             bits;
1227
1228         LASSERT(hlist_unhashed(hnode));
1229
1230         cfs_hash_lock(hs, 0);
1231         cfs_hash_bd_get_and_lock(hs, key, &bd, 1);
1232
1233         cfs_hash_key_validate(hs, key, hnode);
1234         cfs_hash_bd_add_locked(hs, &bd, hnode);
1235
1236         cfs_hash_bd_unlock(hs, &bd, 1);
1237
1238         bits = cfs_hash_rehash_bits(hs);
1239         cfs_hash_unlock(hs, 0);
1240         if (bits > 0)
1241                 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1242 }
1243 EXPORT_SYMBOL(cfs_hash_add);
1244
1245 static struct hlist_node *
1246 cfs_hash_find_or_add(struct cfs_hash *hs, const void *key,
1247                      struct hlist_node *hnode, int noref)
1248 {
1249         struct hlist_node *ehnode;
1250         struct cfs_hash_bd     bds[2];
1251         int               bits = 0;
1252
1253         LASSERT(hlist_unhashed(hnode));
1254
1255         cfs_hash_lock(hs, 0);
1256         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1257
1258         cfs_hash_key_validate(hs, key, hnode);
1259         ehnode = cfs_hash_dual_bd_findadd_locked(hs, bds, key,
1260                                                  hnode, noref);
1261         cfs_hash_dual_bd_unlock(hs, bds, 1);
1262
1263         if (ehnode == hnode) /* new item added */
1264                 bits = cfs_hash_rehash_bits(hs);
1265         cfs_hash_unlock(hs, 0);
1266         if (bits > 0)
1267                 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1268
1269         return ehnode;
1270 }
1271
1272 /**
1273  * Add item @hnode to libcfs hash @hs using @key.  The registered
1274  * ops->hs_get function will be called if the item was added.
1275  * Returns 0 on success or -EALREADY on key collisions.
1276  */
1277 int
1278 cfs_hash_add_unique(struct cfs_hash *hs, const void *key,
1279                     struct hlist_node *hnode)
1280 {
1281         return cfs_hash_find_or_add(hs, key, hnode, 1) != hnode ?
1282                -EALREADY : 0;
1283 }
1284 EXPORT_SYMBOL(cfs_hash_add_unique);
1285
1286 /**
1287  * Add item @hnode to libcfs hash @hs using @key.  If this @key
1288  * already exists in the hash then ops->hs_get will be called on the
1289  * conflicting entry and that entry will be returned to the caller.
1290  * Otherwise ops->hs_get is called on the item which was added.
1291  */
1292 void *
1293 cfs_hash_findadd_unique(struct cfs_hash *hs, const void *key,
1294                         struct hlist_node *hnode)
1295 {
1296         hnode = cfs_hash_find_or_add(hs, key, hnode, 0);
1297
1298         return cfs_hash_object(hs, hnode);
1299 }
1300 EXPORT_SYMBOL(cfs_hash_findadd_unique);
1301
1302 /**
1303  * Delete item @hnode from the libcfs hash @hs using @key.  The @key
1304  * is required to ensure the correct hash bucket is locked since there
1305  * is no direct linkage from the item to the bucket.  The object
1306  * removed from the hash will be returned and obs->hs_put is called
1307  * on the removed object.
1308  */
1309 void *
1310 cfs_hash_del(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1311 {
1312         void           *obj  = NULL;
1313         int             bits = 0;
1314         struct cfs_hash_bd   bds[2];
1315
1316         cfs_hash_lock(hs, 0);
1317         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1318
1319         /* NB: do nothing if @hnode is not in hash table */
1320         if (hnode == NULL || !hlist_unhashed(hnode)) {
1321                 if (bds[1].bd_bucket == NULL && hnode != NULL) {
1322                         cfs_hash_bd_del_locked(hs, &bds[0], hnode);
1323                 } else {
1324                         hnode = cfs_hash_dual_bd_finddel_locked(hs, bds,
1325                                                                 key, hnode);
1326                 }
1327         }
1328
1329         if (hnode != NULL) {
1330                 obj  = cfs_hash_object(hs, hnode);
1331                 bits = cfs_hash_rehash_bits(hs);
1332         }
1333
1334         cfs_hash_dual_bd_unlock(hs, bds, 1);
1335         cfs_hash_unlock(hs, 0);
1336         if (bits > 0)
1337                 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1338
1339         return obj;
1340 }
1341 EXPORT_SYMBOL(cfs_hash_del);
1342
1343 /**
1344  * Delete item given @key in libcfs hash @hs.  The first @key found in
1345  * the hash will be removed, if the key exists multiple times in the hash
1346  * @hs this function must be called once per key.  The removed object
1347  * will be returned and ops->hs_put is called on the removed object.
1348  */
1349 void *
1350 cfs_hash_del_key(struct cfs_hash *hs, const void *key)
1351 {
1352         return cfs_hash_del(hs, key, NULL);
1353 }
1354 EXPORT_SYMBOL(cfs_hash_del_key);
1355
1356 /**
1357  * Lookup an item using @key in the libcfs hash @hs and return it.
1358  * If the @key is found in the hash hs->hs_get() is called and the
1359  * matching objects is returned.  It is the callers responsibility
1360  * to call the counterpart ops->hs_put using the cfs_hash_put() macro
1361  * when when finished with the object.  If the @key was not found
1362  * in the hash @hs NULL is returned.
1363  */
1364 void *
1365 cfs_hash_lookup(struct cfs_hash *hs, const void *key)
1366 {
1367         void                 *obj = NULL;
1368         struct hlist_node     *hnode;
1369         struct cfs_hash_bd         bds[2];
1370
1371         cfs_hash_lock(hs, 0);
1372         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1373
1374         hnode = cfs_hash_dual_bd_lookup_locked(hs, bds, key);
1375         if (hnode != NULL)
1376                 obj = cfs_hash_object(hs, hnode);
1377
1378         cfs_hash_dual_bd_unlock(hs, bds, 0);
1379         cfs_hash_unlock(hs, 0);
1380
1381         return obj;
1382 }
1383 EXPORT_SYMBOL(cfs_hash_lookup);
1384
1385 static void
1386 cfs_hash_for_each_enter(struct cfs_hash *hs)
1387 {
1388         LASSERT(!cfs_hash_is_exiting(hs));
1389
1390         if (!cfs_hash_with_rehash(hs))
1391                 return;
1392         /*
1393          * NB: it's race on cfs_has_t::hs_iterating, but doesn't matter
1394          * because it's just an unreliable signal to rehash-thread,
1395          * rehash-thread will try to finish rehash ASAP when seeing this.
1396          */
1397         hs->hs_iterating = 1;
1398
1399         cfs_hash_lock(hs, 1);
1400         hs->hs_iterators++;
1401
1402         /* NB: iteration is mostly called by service thread,
1403          * we tend to cancel pending rehash-request, instead of
1404          * blocking service thread, we will relaunch rehash request
1405          * after iteration */
1406         if (cfs_hash_is_rehashing(hs))
1407                 cfs_hash_rehash_cancel_locked(hs);
1408         cfs_hash_unlock(hs, 1);
1409 }
1410
1411 static void
1412 cfs_hash_for_each_exit(struct cfs_hash *hs)
1413 {
1414         int remained;
1415         int bits;
1416
1417         if (!cfs_hash_with_rehash(hs))
1418                 return;
1419         cfs_hash_lock(hs, 1);
1420         remained = --hs->hs_iterators;
1421         bits = cfs_hash_rehash_bits(hs);
1422         cfs_hash_unlock(hs, 1);
1423         /* NB: it's race on cfs_has_t::hs_iterating, see above */
1424         if (remained == 0)
1425                 hs->hs_iterating = 0;
1426         if (bits > 0) {
1427                 cfs_hash_rehash(hs, atomic_read(&hs->hs_count) <
1428                                     CFS_HASH_LOOP_HOG);
1429         }
1430 }
1431
1432 /**
1433  * For each item in the libcfs hash @hs call the passed callback @func
1434  * and pass to it as an argument each hash item and the private @data.
1435  *
1436  * a) the function may sleep!
1437  * b) during the callback:
1438  *    . the bucket lock is held so the callback must never sleep.
1439  *    . if @removal_safe is true, use can remove current item by
1440  *      cfs_hash_bd_del_locked
1441  */
1442 static __u64
1443 cfs_hash_for_each_tight(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1444                         void *data, int remove_safe)
1445 {
1446         struct hlist_node       *hnode;
1447         struct hlist_node       *pos;
1448         struct cfs_hash_bd      bd;
1449         __u64                   count = 0;
1450         int                     excl  = !!remove_safe;
1451         int                     loop  = 0;
1452         int                     i;
1453         ENTRY;
1454
1455         cfs_hash_for_each_enter(hs);
1456
1457         cfs_hash_lock(hs, 0);
1458         LASSERT(!cfs_hash_is_rehashing(hs));
1459
1460         cfs_hash_for_each_bucket(hs, &bd, i) {
1461                 struct hlist_head *hhead;
1462
1463                 cfs_hash_bd_lock(hs, &bd, excl);
1464                 if (func == NULL) { /* only glimpse size */
1465                         count += bd.bd_bucket->hsb_count;
1466                         cfs_hash_bd_unlock(hs, &bd, excl);
1467                         continue;
1468                 }
1469
1470                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1471                         hlist_for_each_safe(hnode, pos, hhead) {
1472                                 cfs_hash_bucket_validate(hs, &bd, hnode);
1473                                 count++;
1474                                 loop++;
1475                                 if (func(hs, &bd, hnode, data)) {
1476                                         cfs_hash_bd_unlock(hs, &bd, excl);
1477                                         goto out;
1478                                 }
1479                         }
1480                 }
1481                 cfs_hash_bd_unlock(hs, &bd, excl);
1482                 if (loop < CFS_HASH_LOOP_HOG)
1483                         continue;
1484                 loop = 0;
1485                 cfs_hash_unlock(hs, 0);
1486                 cond_resched();
1487                 cfs_hash_lock(hs, 0);
1488         }
1489  out:
1490         cfs_hash_unlock(hs, 0);
1491
1492         cfs_hash_for_each_exit(hs);
1493         RETURN(count);
1494 }
1495
1496 struct cfs_hash_cond_arg {
1497         cfs_hash_cond_opt_cb_t  func;
1498         void                   *arg;
1499 };
1500
1501 static int
1502 cfs_hash_cond_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1503                          struct hlist_node *hnode, void *data)
1504 {
1505         struct cfs_hash_cond_arg *cond = data;
1506
1507         if (cond->func(cfs_hash_object(hs, hnode), cond->arg))
1508                 cfs_hash_bd_del_locked(hs, bd, hnode);
1509         return 0;
1510 }
1511
1512 /**
1513  * Delete item from the libcfs hash @hs when @func return true.
1514  * The write lock being hold during loop for each bucket to avoid
1515  * any object be reference.
1516  */
1517 void
1518 cfs_hash_cond_del(struct cfs_hash *hs, cfs_hash_cond_opt_cb_t func, void *data)
1519 {
1520         struct cfs_hash_cond_arg arg = {
1521                 .func   = func,
1522                 .arg    = data,
1523         };
1524
1525         cfs_hash_for_each_tight(hs, cfs_hash_cond_del_locked, &arg, 1);
1526 }
1527 EXPORT_SYMBOL(cfs_hash_cond_del);
1528
1529 void
1530 cfs_hash_for_each(struct cfs_hash *hs,
1531                   cfs_hash_for_each_cb_t func, void *data)
1532 {
1533         cfs_hash_for_each_tight(hs, func, data, 0);
1534 }
1535 EXPORT_SYMBOL(cfs_hash_for_each);
1536
1537 void
1538 cfs_hash_for_each_safe(struct cfs_hash *hs,
1539                        cfs_hash_for_each_cb_t func, void *data)
1540 {
1541         cfs_hash_for_each_tight(hs, func, data, 1);
1542 }
1543 EXPORT_SYMBOL(cfs_hash_for_each_safe);
1544
1545 static int
1546 cfs_hash_peek(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1547               struct hlist_node *hnode, void *data)
1548 {
1549         *(int *)data = 0;
1550         return 1; /* return 1 to break the loop */
1551 }
1552
1553 int
1554 cfs_hash_is_empty(struct cfs_hash *hs)
1555 {
1556         int empty = 1;
1557
1558         cfs_hash_for_each_tight(hs, cfs_hash_peek, &empty, 0);
1559         return empty;
1560 }
1561 EXPORT_SYMBOL(cfs_hash_is_empty);
1562
1563 __u64
1564 cfs_hash_size_get(struct cfs_hash *hs)
1565 {
1566         return cfs_hash_with_counter(hs) ?
1567                atomic_read(&hs->hs_count) :
1568                cfs_hash_for_each_tight(hs, NULL, NULL, 0);
1569 }
1570 EXPORT_SYMBOL(cfs_hash_size_get);
1571
1572 /*
1573  * cfs_hash_for_each_relax:
1574  * Iterate the hash table and call @func on each item without
1575  * any lock. This function can't guarantee to finish iteration
1576  * if these features are enabled:
1577  *
1578  *  a. if rehash_key is enabled, an item can be moved from
1579  *     one bucket to another bucket
1580  *  b. user can remove non-zero-ref item from hash-table,
1581  *     so the item can be removed from hash-table, even worse,
1582  *     it's possible that user changed key and insert to another
1583  *     hash bucket.
1584  * there's no way for us to finish iteration correctly on previous
1585  * two cases, so iteration has to be stopped on change.
1586  */
1587 static int
1588 cfs_hash_for_each_relax(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1589                         void *data)
1590 {
1591         struct hlist_node       *hnode;
1592         struct hlist_node       *tmp;
1593         struct cfs_hash_bd      bd;
1594         __u32                   version;
1595         int                     count = 0;
1596         int                     stop_on_change;
1597         int                     rc;
1598         int                     i;
1599         ENTRY;
1600
1601         stop_on_change = cfs_hash_with_rehash_key(hs) ||
1602                          !cfs_hash_with_no_itemref(hs) ||
1603                          hs->hs_ops->hs_put_locked == NULL;
1604         cfs_hash_lock(hs, 0);
1605         LASSERT(!cfs_hash_is_rehashing(hs));
1606
1607         cfs_hash_for_each_bucket(hs, &bd, i) {
1608                 struct hlist_head *hhead;
1609
1610                 cfs_hash_bd_lock(hs, &bd, 0);
1611                 version = cfs_hash_bd_version_get(&bd);
1612
1613                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1614                         for (hnode = hhead->first; hnode != NULL;) {
1615                                 cfs_hash_bucket_validate(hs, &bd, hnode);
1616                                 cfs_hash_get(hs, hnode);
1617                                 cfs_hash_bd_unlock(hs, &bd, 0);
1618                                 cfs_hash_unlock(hs, 0);
1619
1620                                 rc = func(hs, &bd, hnode, data);
1621                                 if (stop_on_change)
1622                                         cfs_hash_put(hs, hnode);
1623                                 cond_resched();
1624                                 count++;
1625
1626                                 cfs_hash_lock(hs, 0);
1627                                 cfs_hash_bd_lock(hs, &bd, 0);
1628                                 if (!stop_on_change) {
1629                                         tmp = hnode->next;
1630                                         cfs_hash_put_locked(hs, hnode);
1631                                         hnode = tmp;
1632                                 } else { /* bucket changed? */
1633                                         if (version !=
1634                                             cfs_hash_bd_version_get(&bd))
1635                                                 break;
1636                                         /* safe to continue because no change */
1637                                         hnode = hnode->next;
1638                                 }
1639                                 if (rc) /* callback wants to break iteration */
1640                                         break;
1641                         }
1642                         if (rc) /* callback wants to break iteration */
1643                                 break;
1644                 }
1645                 cfs_hash_bd_unlock(hs, &bd, 0);
1646                 if (rc) /* callback wants to break iteration */
1647                         break;
1648         }
1649         cfs_hash_unlock(hs, 0);
1650
1651         return count;
1652 }
1653
1654 int
1655 cfs_hash_for_each_nolock(struct cfs_hash *hs,
1656                          cfs_hash_for_each_cb_t func, void *data)
1657 {
1658         ENTRY;
1659
1660         if (cfs_hash_with_no_lock(hs) ||
1661             cfs_hash_with_rehash_key(hs) ||
1662             !cfs_hash_with_no_itemref(hs))
1663                 RETURN(-EOPNOTSUPP);
1664
1665         if (hs->hs_ops->hs_get == NULL ||
1666            (hs->hs_ops->hs_put == NULL &&
1667             hs->hs_ops->hs_put_locked == NULL))
1668                 RETURN(-EOPNOTSUPP);
1669
1670         cfs_hash_for_each_enter(hs);
1671         cfs_hash_for_each_relax(hs, func, data);
1672         cfs_hash_for_each_exit(hs);
1673
1674         RETURN(0);
1675 }
1676 EXPORT_SYMBOL(cfs_hash_for_each_nolock);
1677
1678 /**
1679  * For each hash bucket in the libcfs hash @hs call the passed callback
1680  * @func until all the hash buckets are empty.  The passed callback @func
1681  * or the previously registered callback hs->hs_put must remove the item
1682  * from the hash.  You may either use the cfs_hash_del() or hlist_del()
1683  * functions.  No rwlocks will be held during the callback @func it is
1684  * safe to sleep if needed.  This function will not terminate until the
1685  * hash is empty.  Note it is still possible to concurrently add new
1686  * items in to the hash.  It is the callers responsibility to ensure
1687  * the required locking is in place to prevent concurrent insertions.
1688  */
1689 int
1690 cfs_hash_for_each_empty(struct cfs_hash *hs,
1691                         cfs_hash_for_each_cb_t func, void *data)
1692 {
1693         unsigned  i = 0;
1694         ENTRY;
1695
1696         if (cfs_hash_with_no_lock(hs))
1697                 return -EOPNOTSUPP;
1698
1699         if (hs->hs_ops->hs_get == NULL ||
1700            (hs->hs_ops->hs_put == NULL &&
1701             hs->hs_ops->hs_put_locked == NULL))
1702                 return -EOPNOTSUPP;
1703
1704         cfs_hash_for_each_enter(hs);
1705         while (cfs_hash_for_each_relax(hs, func, data)) {
1706                 CDEBUG(D_INFO, "Try to empty hash: %s, loop: %u\n",
1707                        hs->hs_name, i++);
1708         }
1709         cfs_hash_for_each_exit(hs);
1710         RETURN(0);
1711 }
1712 EXPORT_SYMBOL(cfs_hash_for_each_empty);
1713
1714 void
1715 cfs_hash_hlist_for_each(struct cfs_hash *hs, unsigned hindex,
1716                         cfs_hash_for_each_cb_t func, void *data)
1717 {
1718         struct hlist_head *hhead;
1719         struct hlist_node *hnode;
1720         struct cfs_hash_bd         bd;
1721
1722         cfs_hash_for_each_enter(hs);
1723         cfs_hash_lock(hs, 0);
1724         if (hindex >= CFS_HASH_NHLIST(hs))
1725                 goto out;
1726
1727         cfs_hash_bd_index_set(hs, hindex, &bd);
1728
1729         cfs_hash_bd_lock(hs, &bd, 0);
1730         hhead = cfs_hash_bd_hhead(hs, &bd);
1731         hlist_for_each(hnode, hhead) {
1732                 if (func(hs, &bd, hnode, data))
1733                         break;
1734         }
1735         cfs_hash_bd_unlock(hs, &bd, 0);
1736 out:
1737         cfs_hash_unlock(hs, 0);
1738         cfs_hash_for_each_exit(hs);
1739 }
1740
1741 EXPORT_SYMBOL(cfs_hash_hlist_for_each);
1742
1743 /*
1744  * For each item in the libcfs hash @hs which matches the @key call
1745  * the passed callback @func and pass to it as an argument each hash
1746  * item and the private @data. During the callback the bucket lock
1747  * is held so the callback must never sleep.
1748    */
1749 void
1750 cfs_hash_for_each_key(struct cfs_hash *hs, const void *key,
1751                         cfs_hash_for_each_cb_t func, void *data)
1752 {
1753         struct hlist_node *hnode;
1754         struct cfs_hash_bd         bds[2];
1755         unsigned           i;
1756
1757         cfs_hash_lock(hs, 0);
1758
1759         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1760
1761         cfs_hash_for_each_bd(bds, 2, i) {
1762                 struct hlist_head *hlist = cfs_hash_bd_hhead(hs, &bds[i]);
1763
1764                 hlist_for_each(hnode, hlist) {
1765                         cfs_hash_bucket_validate(hs, &bds[i], hnode);
1766
1767                         if (cfs_hash_keycmp(hs, key, hnode)) {
1768                                 if (func(hs, &bds[i], hnode, data))
1769                                         break;
1770                         }
1771                 }
1772         }
1773
1774         cfs_hash_dual_bd_unlock(hs, bds, 0);
1775         cfs_hash_unlock(hs, 0);
1776 }
1777 EXPORT_SYMBOL(cfs_hash_for_each_key);
1778
1779 /**
1780  * Rehash the libcfs hash @hs to the given @bits.  This can be used
1781  * to grow the hash size when excessive chaining is detected, or to
1782  * shrink the hash when it is larger than needed.  When the CFS_HASH_REHASH
1783  * flag is set in @hs the libcfs hash may be dynamically rehashed
1784  * during addition or removal if the hash's theta value exceeds
1785  * either the hs->hs_min_theta or hs->max_theta values.  By default
1786  * these values are tuned to keep the chained hash depth small, and
1787  * this approach assumes a reasonably uniform hashing function.  The
1788  * theta thresholds for @hs are tunable via cfs_hash_set_theta().
1789  */
1790 void
1791 cfs_hash_rehash_cancel_locked(struct cfs_hash *hs)
1792 {
1793         int     i;
1794
1795         /* need hold cfs_hash_lock(hs, 1) */
1796         LASSERT(cfs_hash_with_rehash(hs) &&
1797                 !cfs_hash_with_no_lock(hs));
1798
1799         if (!cfs_hash_is_rehashing(hs))
1800                 return;
1801
1802         if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_rehash_wi)) {
1803                 hs->hs_rehash_bits = 0;
1804                 return;
1805         }
1806
1807         for (i = 2; cfs_hash_is_rehashing(hs); i++) {
1808                 cfs_hash_unlock(hs, 1);
1809                 /* raise console warning while waiting too long */
1810                 CDEBUG(IS_PO2(i >> 3) ? D_WARNING : D_INFO,
1811                        "hash %s is still rehashing, rescheded %d\n",
1812                        hs->hs_name, i - 1);
1813                 cond_resched();
1814                 cfs_hash_lock(hs, 1);
1815         }
1816 }
1817
1818 void
1819 cfs_hash_rehash_cancel(struct cfs_hash *hs)
1820 {
1821         cfs_hash_lock(hs, 1);
1822         cfs_hash_rehash_cancel_locked(hs);
1823         cfs_hash_unlock(hs, 1);
1824 }
1825
1826 int
1827 cfs_hash_rehash(struct cfs_hash *hs, int do_rehash)
1828 {
1829         int     rc;
1830
1831         LASSERT(cfs_hash_with_rehash(hs) && !cfs_hash_with_no_lock(hs));
1832
1833         cfs_hash_lock(hs, 1);
1834
1835         rc = cfs_hash_rehash_bits(hs);
1836         if (rc <= 0) {
1837                 cfs_hash_unlock(hs, 1);
1838                 return rc;
1839         }
1840
1841         hs->hs_rehash_bits = rc;
1842         if (!do_rehash) {
1843                 /* launch and return */
1844                 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_rehash_wi);
1845                 cfs_hash_unlock(hs, 1);
1846                 return 0;
1847         }
1848
1849         /* rehash right now */
1850         cfs_hash_unlock(hs, 1);
1851
1852         return cfs_hash_rehash_worker(&hs->hs_rehash_wi);
1853 }
1854
1855 static int
1856 cfs_hash_rehash_bd(struct cfs_hash *hs, struct cfs_hash_bd *old)
1857 {
1858         struct cfs_hash_bd      new;
1859         struct hlist_head *hhead;
1860         struct hlist_node *hnode;
1861         struct hlist_node *pos;
1862         void              *key;
1863         int                c = 0;
1864
1865         /* hold cfs_hash_lock(hs, 1), so don't need any bucket lock */
1866         cfs_hash_bd_for_each_hlist(hs, old, hhead) {
1867                 hlist_for_each_safe(hnode, pos, hhead) {
1868                         key = cfs_hash_key(hs, hnode);
1869                         LASSERT(key != NULL);
1870                         /* Validate hnode is in the correct bucket. */
1871                         cfs_hash_bucket_validate(hs, old, hnode);
1872                         /*
1873                          * Delete from old hash bucket; move to new bucket.
1874                          * ops->hs_key must be defined.
1875                          */
1876                         cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
1877                                              hs->hs_rehash_bits, key, &new);
1878                         cfs_hash_bd_move_locked(hs, old, &new, hnode);
1879                         c++;
1880                 }
1881         }
1882         return c;
1883 }
1884
1885 static int
1886 cfs_hash_rehash_worker(cfs_workitem_t *wi)
1887 {
1888         struct cfs_hash         *hs =
1889                 container_of(wi, struct cfs_hash, hs_rehash_wi);
1890         struct cfs_hash_bucket **bkts;
1891         struct cfs_hash_bd      bd;
1892         unsigned int            old_size;
1893         unsigned int            new_size;
1894         int                     bsize;
1895         int                     count = 0;
1896         int                     rc = 0;
1897         int                     i;
1898
1899         LASSERT(hs != NULL && cfs_hash_with_rehash(hs));
1900
1901         cfs_hash_lock(hs, 0);
1902         LASSERT(cfs_hash_is_rehashing(hs));
1903
1904         old_size = CFS_HASH_NBKT(hs);
1905         new_size = CFS_HASH_RH_NBKT(hs);
1906
1907         cfs_hash_unlock(hs, 0);
1908
1909         /*
1910          * don't need hs::hs_rwlock for hs::hs_buckets,
1911          * because nobody can change bkt-table except me.
1912          */
1913         bkts = cfs_hash_buckets_realloc(hs, hs->hs_buckets,
1914                                         old_size, new_size);
1915         cfs_hash_lock(hs, 1);
1916         if (bkts == NULL) {
1917                 rc = -ENOMEM;
1918                 goto out;
1919         }
1920
1921         if (bkts == hs->hs_buckets) {
1922                 bkts = NULL; /* do nothing */
1923                 goto out;
1924         }
1925
1926         rc = __cfs_hash_theta(hs);
1927         if ((rc >= hs->hs_min_theta) && (rc <= hs->hs_max_theta)) {
1928                 /* free the new allocated bkt-table */
1929                 old_size = new_size;
1930                 new_size = CFS_HASH_NBKT(hs);
1931                 rc = -EALREADY;
1932                 goto out;
1933         }
1934
1935         LASSERT(hs->hs_rehash_buckets == NULL);
1936         hs->hs_rehash_buckets = bkts;
1937
1938         rc = 0;
1939         cfs_hash_for_each_bucket(hs, &bd, i) {
1940                 if (cfs_hash_is_exiting(hs)) {
1941                         rc = -ESRCH;
1942                         /* someone wants to destroy the hash, abort now */
1943                         if (old_size < new_size) /* OK to free old bkt-table */
1944                                 break;
1945                         /* it's shrinking, need free new bkt-table */
1946                         hs->hs_rehash_buckets = NULL;
1947                         old_size = new_size;
1948                         new_size = CFS_HASH_NBKT(hs);
1949                         goto out;
1950                 }
1951
1952                 count += cfs_hash_rehash_bd(hs, &bd);
1953                 if (count < CFS_HASH_LOOP_HOG ||
1954                     cfs_hash_is_iterating(hs)) { /* need to finish ASAP */
1955                         continue;
1956                 }
1957
1958                 count = 0;
1959                 cfs_hash_unlock(hs, 1);
1960                 cond_resched();
1961                 cfs_hash_lock(hs, 1);
1962         }
1963
1964         hs->hs_rehash_count++;
1965
1966         bkts = hs->hs_buckets;
1967         hs->hs_buckets = hs->hs_rehash_buckets;
1968         hs->hs_rehash_buckets = NULL;
1969
1970         hs->hs_cur_bits = hs->hs_rehash_bits;
1971  out:
1972         hs->hs_rehash_bits = 0;
1973         if (rc == -ESRCH) /* never be scheduled again */
1974                 cfs_wi_exit(cfs_sched_rehash, wi);
1975         bsize = cfs_hash_bkt_size(hs);
1976         cfs_hash_unlock(hs, 1);
1977         /* can't refer to @hs anymore because it could be destroyed */
1978         if (bkts != NULL)
1979                 cfs_hash_buckets_free(bkts, bsize, new_size, old_size);
1980         if (rc != 0)
1981                 CDEBUG(D_INFO, "early quit of rehashing: %d\n", rc);
1982         /* return 1 only if cfs_wi_exit is called */
1983         return rc == -ESRCH;
1984 }
1985
1986 /**
1987  * Rehash the object referenced by @hnode in the libcfs hash @hs.  The
1988  * @old_key must be provided to locate the objects previous location
1989  * in the hash, and the @new_key will be used to reinsert the object.
1990  * Use this function instead of a cfs_hash_add() + cfs_hash_del()
1991  * combo when it is critical that there is no window in time where the
1992  * object is missing from the hash.  When an object is being rehashed
1993  * the registered cfs_hash_get() and cfs_hash_put() functions will
1994  * not be called.
1995  */
1996 void cfs_hash_rehash_key(struct cfs_hash *hs, const void *old_key,
1997                          void *new_key, struct hlist_node *hnode)
1998 {
1999         struct cfs_hash_bd        bds[3];
2000         struct cfs_hash_bd        old_bds[2];
2001         struct cfs_hash_bd        new_bd;
2002
2003         LASSERT(!hlist_unhashed(hnode));
2004
2005         cfs_hash_lock(hs, 0);
2006
2007         cfs_hash_dual_bd_get(hs, old_key, old_bds);
2008         cfs_hash_bd_get(hs, new_key, &new_bd);
2009
2010         bds[0] = old_bds[0];
2011         bds[1] = old_bds[1];
2012         bds[2] = new_bd;
2013
2014         /* NB: bds[0] and bds[1] are ordered already */
2015         cfs_hash_bd_order(&bds[1], &bds[2]);
2016         cfs_hash_bd_order(&bds[0], &bds[1]);
2017
2018         cfs_hash_multi_bd_lock(hs, bds, 3, 1);
2019         if (likely(old_bds[1].bd_bucket == NULL)) {
2020                 cfs_hash_bd_move_locked(hs, &old_bds[0], &new_bd, hnode);
2021         } else {
2022                 cfs_hash_dual_bd_finddel_locked(hs, old_bds, old_key, hnode);
2023                 cfs_hash_bd_add_locked(hs, &new_bd, hnode);
2024         }
2025         /* overwrite key inside locks, otherwise may screw up with
2026          * other operations, i.e: rehash */
2027         cfs_hash_keycpy(hs, hnode, new_key);
2028
2029         cfs_hash_multi_bd_unlock(hs, bds, 3, 1);
2030         cfs_hash_unlock(hs, 0);
2031 }
2032 EXPORT_SYMBOL(cfs_hash_rehash_key);
2033
2034 int cfs_hash_debug_header(struct seq_file *m)
2035 {
2036         return seq_printf(m, "%-*s%6s%6s%6s%6s%6s%6s%6s%7s%8s%8s%8s%s\n",
2037                         CFS_HASH_BIGNAME_LEN,
2038                         "name", "cur", "min", "max", "theta", "t-min", "t-max",
2039                         "flags", "rehash", "count", "maxdep", "maxdepb",
2040                         " distribution");
2041 }
2042 EXPORT_SYMBOL(cfs_hash_debug_header);
2043
2044 static struct cfs_hash_bucket **
2045 cfs_hash_full_bkts(struct cfs_hash *hs)
2046 {
2047         /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2048         if (hs->hs_rehash_buckets == NULL)
2049                 return hs->hs_buckets;
2050
2051         LASSERT(hs->hs_rehash_bits != 0);
2052         return hs->hs_rehash_bits > hs->hs_cur_bits ?
2053                hs->hs_rehash_buckets : hs->hs_buckets;
2054 }
2055
2056 static unsigned int
2057 cfs_hash_full_nbkt(struct cfs_hash *hs)
2058 {
2059         /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2060         if (hs->hs_rehash_buckets == NULL)
2061                 return CFS_HASH_NBKT(hs);
2062
2063         LASSERT(hs->hs_rehash_bits != 0);
2064         return hs->hs_rehash_bits > hs->hs_cur_bits ?
2065                CFS_HASH_RH_NBKT(hs) : CFS_HASH_NBKT(hs);
2066 }
2067
2068 int cfs_hash_debug_str(struct cfs_hash *hs, struct seq_file *m)
2069 {
2070         int     dist[8] = { 0, };
2071         int     maxdep  = -1;
2072         int     maxdepb = -1;
2073         int     total   = 0;
2074         int     c       = 0;
2075         int     theta;
2076         int     i;
2077
2078         cfs_hash_lock(hs, 0);
2079         theta = __cfs_hash_theta(hs);
2080
2081         c += seq_printf(m, "%-*s ", CFS_HASH_BIGNAME_LEN, hs->hs_name);
2082         c += seq_printf(m, "%5d ",  1 << hs->hs_cur_bits);
2083         c += seq_printf(m, "%5d ",  1 << hs->hs_min_bits);
2084         c += seq_printf(m, "%5d ",  1 << hs->hs_max_bits);
2085         c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(theta),
2086                         __cfs_hash_theta_frac(theta));
2087         c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(hs->hs_min_theta),
2088                         __cfs_hash_theta_frac(hs->hs_min_theta));
2089         c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(hs->hs_max_theta),
2090                         __cfs_hash_theta_frac(hs->hs_max_theta));
2091         c += seq_printf(m, " 0x%02x ", hs->hs_flags);
2092         c += seq_printf(m, "%6d ", hs->hs_rehash_count);
2093
2094         /*
2095          * The distribution is a summary of the chained hash depth in
2096          * each of the libcfs hash buckets.  Each buckets hsb_count is
2097          * divided by the hash theta value and used to generate a
2098          * histogram of the hash distribution.  A uniform hash will
2099          * result in all hash buckets being close to the average thus
2100          * only the first few entries in the histogram will be non-zero.
2101          * If you hash function results in a non-uniform hash the will
2102          * be observable by outlier bucks in the distribution histogram.
2103          *
2104          * Uniform hash distribution:           128/128/0/0/0/0/0/0
2105          * Non-Uniform hash distribution:       128/125/0/0/0/0/2/1
2106          */
2107         for (i = 0; i < cfs_hash_full_nbkt(hs); i++) {
2108                 struct cfs_hash_bd bd;
2109
2110                 bd.bd_bucket = cfs_hash_full_bkts(hs)[i];
2111                 cfs_hash_bd_lock(hs, &bd, 0);
2112                 if (maxdep < bd.bd_bucket->hsb_depmax) {
2113                         maxdep  = bd.bd_bucket->hsb_depmax;
2114                         maxdepb = ffz(~maxdep);
2115                 }
2116                 total += bd.bd_bucket->hsb_count;
2117                 dist[min(fls(bd.bd_bucket->hsb_count/max(theta,1)),7)]++;
2118                 cfs_hash_bd_unlock(hs, &bd, 0);
2119         }
2120
2121         c += seq_printf(m, "%7d ", total);
2122         c += seq_printf(m, "%7d ", maxdep);
2123         c += seq_printf(m, "%7d ", maxdepb);
2124         for (i = 0; i < 8; i++)
2125                 c += seq_printf(m, "%d%c",  dist[i], (i == 7) ? '\n' : '/');
2126
2127         cfs_hash_unlock(hs, 0);
2128         return c;
2129 }
2130 EXPORT_SYMBOL(cfs_hash_debug_str);