drivers/staging/lustre/lustre/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, 2012, 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
 110 #include "../../include/linux/libcfs/libcfs.h"
 111 #include <linux/seq_file.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 EXPORT_SYMBOL(cfs_hash_bd_move_locked);
 597
 598 enum {
 599         /** always set, for sanity (avoid ZERO intent) */
 600         CFS_HS_LOOKUP_MASK_FIND     = BIT(0),
 601         /** return entry with a ref */
 602         CFS_HS_LOOKUP_MASK_REF      = BIT(1),
 603         /** add entry if not existing */
 604         CFS_HS_LOOKUP_MASK_ADD      = BIT(2),
 605         /** delete entry, ignore other masks */
 606         CFS_HS_LOOKUP_MASK_DEL      = BIT(3),
 607 };
 608
 609 enum cfs_hash_lookup_intent {
 610         /** return item w/o refcount */
 611         CFS_HS_LOOKUP_IT_PEEK       = CFS_HS_LOOKUP_MASK_FIND,
 612         /** return item with refcount */
 613         CFS_HS_LOOKUP_IT_FIND       = (CFS_HS_LOOKUP_MASK_FIND |
 614                                        CFS_HS_LOOKUP_MASK_REF),
 615         /** return item w/o refcount if existed, otherwise add */
 616         CFS_HS_LOOKUP_IT_ADD    = (CFS_HS_LOOKUP_MASK_FIND |
 617                                        CFS_HS_LOOKUP_MASK_ADD),
 618         /** return item with refcount if existed, otherwise add */
 619         CFS_HS_LOOKUP_IT_FINDADD    = (CFS_HS_LOOKUP_IT_FIND |
 620                                        CFS_HS_LOOKUP_MASK_ADD),
 621         /** delete if existed */
 622         CFS_HS_LOOKUP_IT_FINDDEL    = (CFS_HS_LOOKUP_MASK_FIND |
 623                                        CFS_HS_LOOKUP_MASK_DEL)
 624 };
 625
 626 static struct hlist_node *
 627 cfs_hash_bd_lookup_intent(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 628                           const void *key, struct hlist_node *hnode,
 629                           enum cfs_hash_lookup_intent intent)
 630
 631 {
 632         struct hlist_head  *hhead = cfs_hash_bd_hhead(hs, bd);
 633         struct hlist_node  *ehnode;
 634         struct hlist_node  *match;
 635         int  intent_add = (intent & CFS_HS_LOOKUP_MASK_ADD) != 0;
 636
 637         /* with this function, we can avoid a lot of useless refcount ops,
 638          * which are expensive atomic operations most time. */
 639         match = intent_add ? NULL : hnode;
 640         hlist_for_each(ehnode, hhead) {
 641                 if (!cfs_hash_keycmp(hs, key, ehnode))
 642                         continue;
 643
 644                 if (match != NULL && match != ehnode) /* can't match */
 645                         continue;
 646
 647                 /* match and ... */
 648                 if ((intent & CFS_HS_LOOKUP_MASK_DEL) != 0) {
 649                         cfs_hash_bd_del_locked(hs, bd, ehnode);
 650                         return ehnode;
 651                 }
 652
 653                 /* caller wants refcount? */
 654                 if ((intent & CFS_HS_LOOKUP_MASK_REF) != 0)
 655                         cfs_hash_get(hs, ehnode);
 656                 return ehnode;
 657         }
 658         /* no match item */
 659         if (!intent_add)
 660                 return NULL;
 661
 662         LASSERT(hnode != NULL);
 663         cfs_hash_bd_add_locked(hs, bd, hnode);
 664         return hnode;
 665 }
 666
 667 struct hlist_node *
 668 cfs_hash_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd, 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, const void *key)
 677 {
 678         return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
 679                                          CFS_HS_LOOKUP_IT_PEEK);
 680 }
 681 EXPORT_SYMBOL(cfs_hash_bd_peek_locked);
 682
 683 struct hlist_node *
 684 cfs_hash_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
 685                            const void *key, struct hlist_node *hnode,
 686                            int noref)
 687 {
 688         return cfs_hash_bd_lookup_intent(hs, bd, key, hnode,
 689                                          (!noref * CFS_HS_LOOKUP_MASK_REF) |
 690                                          CFS_HS_LOOKUP_IT_ADD);
 691 }
 692 EXPORT_SYMBOL(cfs_hash_bd_findadd_locked);
 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 EXPORT_SYMBOL(cfs_hash_bd_finddel_locked);
 703
 704 static void
 705 cfs_hash_multi_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 706                        unsigned n, int excl)
 707 {
 708         struct cfs_hash_bucket *prev = NULL;
 709         int             i;
 710
 711         /**
 712          * bds must be ascendantly ordered by bd->bd_bucket->hsb_index.
 713          * NB: it's possible that several bds point to the same bucket but
 714          * have different bd::bd_offset, so need take care of deadlock.
 715          */
 716         cfs_hash_for_each_bd(bds, n, i) {
 717                 if (prev == bds[i].bd_bucket)
 718                         continue;
 719
 720                 LASSERT(prev == NULL ||
 721                         prev->hsb_index < bds[i].bd_bucket->hsb_index);
 722                 cfs_hash_bd_lock(hs, &bds[i], excl);
 723                 prev = bds[i].bd_bucket;
 724         }
 725 }
 726
 727 static void
 728 cfs_hash_multi_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 729                          unsigned n, int excl)
 730 {
 731         struct cfs_hash_bucket *prev = NULL;
 732         int             i;
 733
 734         cfs_hash_for_each_bd(bds, n, i) {
 735                 if (prev != bds[i].bd_bucket) {
 736                         cfs_hash_bd_unlock(hs, &bds[i], excl);
 737                         prev = bds[i].bd_bucket;
 738                 }
 739         }
 740 }
 741
 742 static struct hlist_node *
 743 cfs_hash_multi_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 744                                 unsigned n, const void *key)
 745 {
 746         struct hlist_node  *ehnode;
 747         unsigned           i;
 748
 749         cfs_hash_for_each_bd(bds, n, i) {
 750                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, NULL,
 751                                                    CFS_HS_LOOKUP_IT_FIND);
 752                 if (ehnode != NULL)
 753                         return ehnode;
 754         }
 755         return NULL;
 756 }
 757
 758 static struct hlist_node *
 759 cfs_hash_multi_bd_findadd_locked(struct cfs_hash *hs,
 760                                  struct cfs_hash_bd *bds, unsigned n, const void *key,
 761                                  struct hlist_node *hnode, int noref)
 762 {
 763         struct hlist_node  *ehnode;
 764         int             intent;
 765         unsigned           i;
 766
 767         LASSERT(hnode != NULL);
 768         intent = (!noref * CFS_HS_LOOKUP_MASK_REF) | CFS_HS_LOOKUP_IT_PEEK;
 769
 770         cfs_hash_for_each_bd(bds, n, i) {
 771                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key,
 772                                                    NULL, intent);
 773                 if (ehnode != NULL)
 774                         return ehnode;
 775         }
 776
 777         if (i == 1) { /* only one bucket */
 778                 cfs_hash_bd_add_locked(hs, &bds[0], hnode);
 779         } else {
 780                 struct cfs_hash_bd      mybd;
 781
 782                 cfs_hash_bd_get(hs, key, &mybd);
 783                 cfs_hash_bd_add_locked(hs, &mybd, hnode);
 784         }
 785
 786         return hnode;
 787 }
 788
 789 static struct hlist_node *
 790 cfs_hash_multi_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 791                                  unsigned n, const void *key,
 792                                  struct hlist_node *hnode)
 793 {
 794         struct hlist_node  *ehnode;
 795         unsigned           i;
 796
 797         cfs_hash_for_each_bd(bds, n, i) {
 798                 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, hnode,
 799                                                    CFS_HS_LOOKUP_IT_FINDDEL);
 800                 if (ehnode != NULL)
 801                         return ehnode;
 802         }
 803         return NULL;
 804 }
 805
 806 static void
 807 cfs_hash_bd_order(struct cfs_hash_bd *bd1, struct cfs_hash_bd *bd2)
 808 {
 809         int     rc;
 810
 811         if (bd2->bd_bucket == NULL)
 812                 return;
 813
 814         if (bd1->bd_bucket == NULL) {
 815                 *bd1 = *bd2;
 816                 bd2->bd_bucket = NULL;
 817                 return;
 818         }
 819
 820         rc = cfs_hash_bd_compare(bd1, bd2);
 821         if (rc == 0) {
 822                 bd2->bd_bucket = NULL;
 823
 824         } else if (rc > 0) { /* swab bd1 and bd2 */
 825                 struct cfs_hash_bd tmp;
 826
 827                 tmp = *bd2;
 828                 *bd2 = *bd1;
 829                 *bd1 = tmp;
 830         }
 831 }
 832
 833 void
 834 cfs_hash_dual_bd_get(struct cfs_hash *hs, const void *key, 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 EXPORT_SYMBOL(cfs_hash_dual_bd_get);
 852
 853 void
 854 cfs_hash_dual_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
 855 {
 856         cfs_hash_multi_bd_lock(hs, bds, 2, excl);
 857 }
 858 EXPORT_SYMBOL(cfs_hash_dual_bd_lock);
 859
 860 void
 861 cfs_hash_dual_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
 862 {
 863         cfs_hash_multi_bd_unlock(hs, bds, 2, excl);
 864 }
 865 EXPORT_SYMBOL(cfs_hash_dual_bd_unlock);
 866
 867 struct hlist_node *
 868 cfs_hash_dual_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 869                                const void *key)
 870 {
 871         return cfs_hash_multi_bd_lookup_locked(hs, bds, 2, key);
 872 }
 873 EXPORT_SYMBOL(cfs_hash_dual_bd_lookup_locked);
 874
 875 struct hlist_node *
 876 cfs_hash_dual_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 877                                 const void *key, struct hlist_node *hnode,
 878                                 int noref)
 879 {
 880         return cfs_hash_multi_bd_findadd_locked(hs, bds, 2, key,
 881                                                 hnode, noref);
 882 }
 883 EXPORT_SYMBOL(cfs_hash_dual_bd_findadd_locked);
 884
 885 struct hlist_node *
 886 cfs_hash_dual_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
 887                                 const void *key, struct hlist_node *hnode)
 888 {
 889         return cfs_hash_multi_bd_finddel_locked(hs, bds, 2, key, hnode);
 890 }
 891 EXPORT_SYMBOL(cfs_hash_dual_bd_finddel_locked);
 892
 893 static void
 894 cfs_hash_buckets_free(struct cfs_hash_bucket **buckets,
 895                       int bkt_size, int prev_size, int size)
 896 {
 897         int     i;
 898
 899         for (i = prev_size; i < size; i++) {
 900                 if (buckets[i] != NULL)
 901                         LIBCFS_FREE(buckets[i], bkt_size);
 902         }
 903
 904         LIBCFS_FREE(buckets, sizeof(buckets[0]) * size);
 905 }
 906
 907 /*
 908  * Create or grow bucket memory. Return old_buckets if no allocation was
 909  * needed, the newly allocated buckets if allocation was needed and
 910  * successful, and NULL on error.
 911  */
 912 static struct cfs_hash_bucket **
 913 cfs_hash_buckets_realloc(struct cfs_hash *hs, struct cfs_hash_bucket **old_bkts,
 914                          unsigned int old_size, unsigned int new_size)
 915 {
 916         struct cfs_hash_bucket **new_bkts;
 917         int              i;
 918
 919         LASSERT(old_size == 0 || old_bkts != NULL);
 920
 921         if (old_bkts != NULL && old_size == new_size)
 922                 return old_bkts;
 923
 924         LIBCFS_ALLOC(new_bkts, sizeof(new_bkts[0]) * new_size);
 925         if (new_bkts == NULL)
 926                 return NULL;
 927
 928         if (old_bkts != NULL) {
 929                 memcpy(new_bkts, old_bkts,
 930                        min(old_size, new_size) * sizeof(*old_bkts));
 931         }
 932
 933         for (i = old_size; i < new_size; i++) {
 934                 struct hlist_head *hhead;
 935                 struct cfs_hash_bd     bd;
 936
 937                 LIBCFS_ALLOC(new_bkts[i], cfs_hash_bkt_size(hs));
 938                 if (new_bkts[i] == NULL) {
 939                         cfs_hash_buckets_free(new_bkts, cfs_hash_bkt_size(hs),
 940                                               old_size, new_size);
 941                         return NULL;
 942                 }
 943
 944                 new_bkts[i]->hsb_index   = i;
 945                 new_bkts[i]->hsb_version = 1;  /* shouldn't be zero */
 946                 new_bkts[i]->hsb_depmax  = -1; /* unknown */
 947                 bd.bd_bucket = new_bkts[i];
 948                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead)
 949                         INIT_HLIST_HEAD(hhead);
 950
 951                 if (cfs_hash_with_no_lock(hs) ||
 952                     cfs_hash_with_no_bktlock(hs))
 953                         continue;
 954
 955                 if (cfs_hash_with_rw_bktlock(hs))
 956                         rwlock_init(&new_bkts[i]->hsb_lock.rw);
 957                 else if (cfs_hash_with_spin_bktlock(hs))
 958                         spin_lock_init(&new_bkts[i]->hsb_lock.spin);
 959                 else
 960                         LBUG(); /* invalid use-case */
 961         }
 962         return new_bkts;
 963 }
 964
 965 /**
 966  * Initialize new libcfs hash, where:
 967  * @name     - Descriptive hash name
 968  * @cur_bits - Initial hash table size, in bits
 969  * @max_bits - Maximum allowed hash table resize, in bits
 970  * @ops      - Registered hash table operations
 971  * @flags    - CFS_HASH_REHASH enable synamic hash resizing
 972  *         - CFS_HASH_SORT enable chained hash sort
 973  */
 974 static int cfs_hash_rehash_worker(cfs_workitem_t *wi);
 975
 976 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
 977 static int cfs_hash_dep_print(cfs_workitem_t *wi)
 978 {
 979         struct cfs_hash *hs = container_of(wi, struct cfs_hash, hs_dep_wi);
 980         int      dep;
 981         int      bkt;
 982         int      off;
 983         int      bits;
 984
 985         spin_lock(&hs->hs_dep_lock);
 986         dep  = hs->hs_dep_max;
 987         bkt  = hs->hs_dep_bkt;
 988         off  = hs->hs_dep_off;
 989         bits = hs->hs_dep_bits;
 990         spin_unlock(&hs->hs_dep_lock);
 991
 992         LCONSOLE_WARN("#### HASH %s (bits: %d): max depth %d at bucket %d/%d\n",
 993                       hs->hs_name, bits, dep, bkt, off);
 994         spin_lock(&hs->hs_dep_lock);
 995         hs->hs_dep_bits = 0; /* mark as workitem done */
 996         spin_unlock(&hs->hs_dep_lock);
 997         return 0;
 998 }
 999
1000 static void cfs_hash_depth_wi_init(struct cfs_hash *hs)
1001 {
1002         spin_lock_init(&hs->hs_dep_lock);
1003         cfs_wi_init(&hs->hs_dep_wi, hs, cfs_hash_dep_print);
1004 }
1005
1006 static void cfs_hash_depth_wi_cancel(struct cfs_hash *hs)
1007 {
1008         if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_dep_wi))
1009                 return;
1010
1011         spin_lock(&hs->hs_dep_lock);
1012         while (hs->hs_dep_bits != 0) {
1013                 spin_unlock(&hs->hs_dep_lock);
1014                 cond_resched();
1015                 spin_lock(&hs->hs_dep_lock);
1016         }
1017         spin_unlock(&hs->hs_dep_lock);
1018 }
1019
1020 #else /* CFS_HASH_DEBUG_LEVEL < CFS_HASH_DEBUG_1 */
1021
1022 static inline void cfs_hash_depth_wi_init(struct cfs_hash *hs) {}
1023 static inline void cfs_hash_depth_wi_cancel(struct cfs_hash *hs) {}
1024
1025 #endif /* CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1 */
1026
1027 struct cfs_hash *
1028 cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
1029                 unsigned bkt_bits, unsigned extra_bytes,
1030                 unsigned min_theta, unsigned max_theta,
1031                 struct cfs_hash_ops *ops, unsigned flags)
1032 {
1033         struct cfs_hash *hs;
1034         int      len;
1035
1036         CLASSERT(CFS_HASH_THETA_BITS < 15);
1037
1038         LASSERT(name != NULL);
1039         LASSERT(ops != NULL);
1040         LASSERT(ops->hs_key);
1041         LASSERT(ops->hs_hash);
1042         LASSERT(ops->hs_object);
1043         LASSERT(ops->hs_keycmp);
1044         LASSERT(ops->hs_get != NULL);
1045         LASSERT(ops->hs_put_locked != NULL);
1046
1047         if ((flags & CFS_HASH_REHASH) != 0)
1048                 flags |= CFS_HASH_COUNTER; /* must have counter */
1049
1050         LASSERT(cur_bits > 0);
1051         LASSERT(cur_bits >= bkt_bits);
1052         LASSERT(max_bits >= cur_bits && max_bits < 31);
1053         LASSERT(ergo((flags & CFS_HASH_REHASH) == 0, cur_bits == max_bits));
1054         LASSERT(ergo((flags & CFS_HASH_REHASH) != 0,
1055                      (flags & CFS_HASH_NO_LOCK) == 0));
1056         LASSERT(ergo((flags & CFS_HASH_REHASH_KEY) != 0,
1057                       ops->hs_keycpy != NULL));
1058
1059         len = (flags & CFS_HASH_BIGNAME) == 0 ?
1060               CFS_HASH_NAME_LEN : CFS_HASH_BIGNAME_LEN;
1061         LIBCFS_ALLOC(hs, offsetof(struct cfs_hash, hs_name[len]));
1062         if (hs == NULL)
1063                 return NULL;
1064
1065         strlcpy(hs->hs_name, name, len);
1066         hs->hs_flags = flags;
1067
1068         atomic_set(&hs->hs_refcount, 1);
1069         atomic_set(&hs->hs_count, 0);
1070
1071         cfs_hash_lock_setup(hs);
1072         cfs_hash_hlist_setup(hs);
1073
1074         hs->hs_cur_bits = (__u8)cur_bits;
1075         hs->hs_min_bits = (__u8)cur_bits;
1076         hs->hs_max_bits = (__u8)max_bits;
1077         hs->hs_bkt_bits = (__u8)bkt_bits;
1078
1079         hs->hs_ops       = ops;
1080         hs->hs_extra_bytes = extra_bytes;
1081         hs->hs_rehash_bits = 0;
1082         cfs_wi_init(&hs->hs_rehash_wi, hs, cfs_hash_rehash_worker);
1083         cfs_hash_depth_wi_init(hs);
1084
1085         if (cfs_hash_with_rehash(hs))
1086                 __cfs_hash_set_theta(hs, min_theta, max_theta);
1087
1088         hs->hs_buckets = cfs_hash_buckets_realloc(hs, NULL, 0,
1089                                                   CFS_HASH_NBKT(hs));
1090         if (hs->hs_buckets != NULL)
1091                 return hs;
1092
1093         LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[len]));
1094         return NULL;
1095 }
1096 EXPORT_SYMBOL(cfs_hash_create);
1097
1098 /**
1099  * Cleanup libcfs hash @hs.
1100  */
1101 static void
1102 cfs_hash_destroy(struct cfs_hash *hs)
1103 {
1104         struct hlist_node     *hnode;
1105         struct hlist_node     *pos;
1106         struct cfs_hash_bd       bd;
1107         int                i;
1108
1109         LASSERT(hs != NULL);
1110         LASSERT(!cfs_hash_is_exiting(hs) &&
1111                 !cfs_hash_is_iterating(hs));
1112
1113         /**
1114          * prohibit further rehashes, don't need any lock because
1115          * I'm the only (last) one can change it.
1116          */
1117         hs->hs_exiting = 1;
1118         if (cfs_hash_with_rehash(hs))
1119                 cfs_hash_rehash_cancel(hs);
1120
1121         cfs_hash_depth_wi_cancel(hs);
1122         /* rehash should be done/canceled */
1123         LASSERT(hs->hs_buckets != NULL &&
1124                 hs->hs_rehash_buckets == NULL);
1125
1126         cfs_hash_for_each_bucket(hs, &bd, i) {
1127                 struct hlist_head *hhead;
1128
1129                 LASSERT(bd.bd_bucket != NULL);
1130                 /* no need to take this lock, just for consistent code */
1131                 cfs_hash_bd_lock(hs, &bd, 1);
1132
1133                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1134                         hlist_for_each_safe(hnode, pos, hhead) {
1135                                 LASSERTF(!cfs_hash_with_assert_empty(hs),
1136                                          "hash %s bucket %u(%u) is not empty: %u items left\n",
1137                                          hs->hs_name, bd.bd_bucket->hsb_index,
1138                                          bd.bd_offset, bd.bd_bucket->hsb_count);
1139                                 /* can't assert key valicate, because we
1140                                  * can interrupt rehash */
1141                                 cfs_hash_bd_del_locked(hs, &bd, hnode);
1142                                 cfs_hash_exit(hs, hnode);
1143                         }
1144                 }
1145                 LASSERT(bd.bd_bucket->hsb_count == 0);
1146                 cfs_hash_bd_unlock(hs, &bd, 1);
1147                 cond_resched();
1148         }
1149
1150         LASSERT(atomic_read(&hs->hs_count) == 0);
1151
1152         cfs_hash_buckets_free(hs->hs_buckets, cfs_hash_bkt_size(hs),
1153                               0, CFS_HASH_NBKT(hs));
1154         i = cfs_hash_with_bigname(hs) ?
1155             CFS_HASH_BIGNAME_LEN : CFS_HASH_NAME_LEN;
1156         LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[i]));
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, struct hlist_node *hnode)
1279 {
1280         return cfs_hash_find_or_add(hs, key, hnode, 1) != hnode ?
1281                -EALREADY : 0;
1282 }
1283 EXPORT_SYMBOL(cfs_hash_add_unique);
1284
1285 /**
1286  * Add item @hnode to libcfs hash @hs using @key.  If this @key
1287  * already exists in the hash then ops->hs_get will be called on the
1288  * conflicting entry and that entry will be returned to the caller.
1289  * Otherwise ops->hs_get is called on the item which was added.
1290  */
1291 void *
1292 cfs_hash_findadd_unique(struct cfs_hash *hs, const void *key,
1293                         struct hlist_node *hnode)
1294 {
1295         hnode = cfs_hash_find_or_add(hs, key, hnode, 0);
1296
1297         return cfs_hash_object(hs, hnode);
1298 }
1299 EXPORT_SYMBOL(cfs_hash_findadd_unique);
1300
1301 /**
1302  * Delete item @hnode from the libcfs hash @hs using @key.  The @key
1303  * is required to ensure the correct hash bucket is locked since there
1304  * is no direct linkage from the item to the bucket.  The object
1305  * removed from the hash will be returned and obs->hs_put is called
1306  * on the removed object.
1307  */
1308 void *
1309 cfs_hash_del(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1310 {
1311         void       *obj  = NULL;
1312         int          bits = 0;
1313         struct cfs_hash_bd   bds[2];
1314
1315         cfs_hash_lock(hs, 0);
1316         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1317
1318         /* NB: do nothing if @hnode is not in hash table */
1319         if (hnode == NULL || !hlist_unhashed(hnode)) {
1320                 if (bds[1].bd_bucket == NULL && hnode != NULL) {
1321                         cfs_hash_bd_del_locked(hs, &bds[0], hnode);
1322                 } else {
1323                         hnode = cfs_hash_dual_bd_finddel_locked(hs, bds,
1324                                                                 key, hnode);
1325                 }
1326         }
1327
1328         if (hnode != NULL) {
1329                 obj  = cfs_hash_object(hs, hnode);
1330                 bits = cfs_hash_rehash_bits(hs);
1331         }
1332
1333         cfs_hash_dual_bd_unlock(hs, bds, 1);
1334         cfs_hash_unlock(hs, 0);
1335         if (bits > 0)
1336                 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1337
1338         return obj;
1339 }
1340 EXPORT_SYMBOL(cfs_hash_del);
1341
1342 /**
1343  * Delete item given @key in libcfs hash @hs.  The first @key found in
1344  * the hash will be removed, if the key exists multiple times in the hash
1345  * @hs this function must be called once per key.  The removed object
1346  * will be returned and ops->hs_put is called on the removed object.
1347  */
1348 void *
1349 cfs_hash_del_key(struct cfs_hash *hs, const void *key)
1350 {
1351         return cfs_hash_del(hs, key, NULL);
1352 }
1353 EXPORT_SYMBOL(cfs_hash_del_key);
1354
1355 /**
1356  * Lookup an item using @key in the libcfs hash @hs and return it.
1357  * If the @key is found in the hash hs->hs_get() is called and the
1358  * matching objects is returned.  It is the callers responsibility
1359  * to call the counterpart ops->hs_put using the cfs_hash_put() macro
1360  * when when finished with the object.  If the @key was not found
1361  * in the hash @hs NULL is returned.
1362  */
1363 void *
1364 cfs_hash_lookup(struct cfs_hash *hs, const void *key)
1365 {
1366         void             *obj = NULL;
1367         struct hlist_node     *hnode;
1368         struct cfs_hash_bd       bds[2];
1369
1370         cfs_hash_lock(hs, 0);
1371         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1372
1373         hnode = cfs_hash_dual_bd_lookup_locked(hs, bds, key);
1374         if (hnode != NULL)
1375                 obj = cfs_hash_object(hs, hnode);
1376
1377         cfs_hash_dual_bd_unlock(hs, bds, 0);
1378         cfs_hash_unlock(hs, 0);
1379
1380         return obj;
1381 }
1382 EXPORT_SYMBOL(cfs_hash_lookup);
1383
1384 static void
1385 cfs_hash_for_each_enter(struct cfs_hash *hs) {
1386         LASSERT(!cfs_hash_is_exiting(hs));
1387
1388         if (!cfs_hash_with_rehash(hs))
1389                 return;
1390         /*
1391          * NB: it's race on cfs_has_t::hs_iterating, but doesn't matter
1392          * because it's just an unreliable signal to rehash-thread,
1393          * rehash-thread will try to finish rehash ASAP when seeing this.
1394          */
1395         hs->hs_iterating = 1;
1396
1397         cfs_hash_lock(hs, 1);
1398         hs->hs_iterators++;
1399
1400         /* NB: iteration is mostly called by service thread,
1401          * we tend to cancel pending rehash-request, instead of
1402          * blocking service thread, we will relaunch rehash request
1403          * after iteration */
1404         if (cfs_hash_is_rehashing(hs))
1405                 cfs_hash_rehash_cancel_locked(hs);
1406         cfs_hash_unlock(hs, 1);
1407 }
1408
1409 static void
1410 cfs_hash_for_each_exit(struct cfs_hash *hs) {
1411         int remained;
1412         int bits;
1413
1414         if (!cfs_hash_with_rehash(hs))
1415                 return;
1416         cfs_hash_lock(hs, 1);
1417         remained = --hs->hs_iterators;
1418         bits = cfs_hash_rehash_bits(hs);
1419         cfs_hash_unlock(hs, 1);
1420         /* NB: it's race on cfs_has_t::hs_iterating, see above */
1421         if (remained == 0)
1422                 hs->hs_iterating = 0;
1423         if (bits > 0) {
1424                 cfs_hash_rehash(hs, atomic_read(&hs->hs_count) <
1425                                     CFS_HASH_LOOP_HOG);
1426         }
1427 }
1428
1429 /**
1430  * For each item in the libcfs hash @hs call the passed callback @func
1431  * and pass to it as an argument each hash item and the private @data.
1432  *
1433  * a) the function may sleep!
1434  * b) during the callback:
1435  *    . the bucket lock is held so the callback must never sleep.
1436  *    . if @removal_safe is true, use can remove current item by
1437  *      cfs_hash_bd_del_locked
1438  */
1439 static __u64
1440 cfs_hash_for_each_tight(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1441                         void *data, int remove_safe) {
1442         struct hlist_node     *hnode;
1443         struct hlist_node     *pos;
1444         struct cfs_hash_bd       bd;
1445         __u64            count = 0;
1446         int                excl  = !!remove_safe;
1447         int                loop  = 0;
1448         int                i;
1449
1450         cfs_hash_for_each_enter(hs);
1451
1452         cfs_hash_lock(hs, 0);
1453         LASSERT(!cfs_hash_is_rehashing(hs));
1454
1455         cfs_hash_for_each_bucket(hs, &bd, i) {
1456                 struct hlist_head *hhead;
1457
1458                 cfs_hash_bd_lock(hs, &bd, excl);
1459                 if (func == NULL) { /* only glimpse size */
1460                         count += bd.bd_bucket->hsb_count;
1461                         cfs_hash_bd_unlock(hs, &bd, excl);
1462                         continue;
1463                 }
1464
1465                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1466                         hlist_for_each_safe(hnode, pos, hhead) {
1467                                 cfs_hash_bucket_validate(hs, &bd, hnode);
1468                                 count++;
1469                                 loop++;
1470                                 if (func(hs, &bd, hnode, data)) {
1471                                         cfs_hash_bd_unlock(hs, &bd, excl);
1472                                         goto out;
1473                                 }
1474                         }
1475                 }
1476                 cfs_hash_bd_unlock(hs, &bd, excl);
1477                 if (loop < CFS_HASH_LOOP_HOG)
1478                         continue;
1479                 loop = 0;
1480                 cfs_hash_unlock(hs, 0);
1481                 cond_resched();
1482                 cfs_hash_lock(hs, 0);
1483         }
1484  out:
1485         cfs_hash_unlock(hs, 0);
1486
1487         cfs_hash_for_each_exit(hs);
1488         return count;
1489 }
1490
1491 struct cfs_hash_cond_arg {
1492         cfs_hash_cond_opt_cb_t  func;
1493         void                    *arg;
1494 };
1495
1496 static int
1497 cfs_hash_cond_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1498                          struct hlist_node *hnode, void *data)
1499 {
1500         struct cfs_hash_cond_arg *cond = data;
1501
1502         if (cond->func(cfs_hash_object(hs, hnode), cond->arg))
1503                 cfs_hash_bd_del_locked(hs, bd, hnode);
1504         return 0;
1505 }
1506
1507 /**
1508  * Delete item from the libcfs hash @hs when @func return true.
1509  * The write lock being hold during loop for each bucket to avoid
1510  * any object be reference.
1511  */
1512 void
1513 cfs_hash_cond_del(struct cfs_hash *hs, cfs_hash_cond_opt_cb_t func, void *data)
1514 {
1515         struct cfs_hash_cond_arg arg = {
1516                 .func   = func,
1517                 .arg    = data,
1518         };
1519
1520         cfs_hash_for_each_tight(hs, cfs_hash_cond_del_locked, &arg, 1);
1521 }
1522 EXPORT_SYMBOL(cfs_hash_cond_del);
1523
1524 void
1525 cfs_hash_for_each(struct cfs_hash *hs,
1526                   cfs_hash_for_each_cb_t func, void *data)
1527 {
1528         cfs_hash_for_each_tight(hs, func, data, 0);
1529 }
1530 EXPORT_SYMBOL(cfs_hash_for_each);
1531
1532 void
1533 cfs_hash_for_each_safe(struct cfs_hash *hs,
1534                        cfs_hash_for_each_cb_t func, void *data) {
1535         cfs_hash_for_each_tight(hs, func, data, 1);
1536 }
1537 EXPORT_SYMBOL(cfs_hash_for_each_safe);
1538
1539 static int
1540 cfs_hash_peek(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1541               struct hlist_node *hnode, void *data)
1542 {
1543         *(int *)data = 0;
1544         return 1; /* return 1 to break the loop */
1545 }
1546
1547 int
1548 cfs_hash_is_empty(struct cfs_hash *hs)
1549 {
1550         int empty = 1;
1551
1552         cfs_hash_for_each_tight(hs, cfs_hash_peek, &empty, 0);
1553         return empty;
1554 }
1555 EXPORT_SYMBOL(cfs_hash_is_empty);
1556
1557 __u64
1558 cfs_hash_size_get(struct cfs_hash *hs)
1559 {
1560         return cfs_hash_with_counter(hs) ?
1561                atomic_read(&hs->hs_count) :
1562                cfs_hash_for_each_tight(hs, NULL, NULL, 0);
1563 }
1564 EXPORT_SYMBOL(cfs_hash_size_get);
1565
1566 /*
1567  * cfs_hash_for_each_relax:
1568  * Iterate the hash table and call @func on each item without
1569  * any lock. This function can't guarantee to finish iteration
1570  * if these features are enabled:
1571  *
1572  *  a. if rehash_key is enabled, an item can be moved from
1573  *     one bucket to another bucket
1574  *  b. user can remove non-zero-ref item from hash-table,
1575  *     so the item can be removed from hash-table, even worse,
1576  *     it's possible that user changed key and insert to another
1577  *     hash bucket.
1578  * there's no way for us to finish iteration correctly on previous
1579  * two cases, so iteration has to be stopped on change.
1580  */
1581 static int
1582 cfs_hash_for_each_relax(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1583                         void *data) {
1584         struct hlist_node *hnode;
1585         struct hlist_node *tmp;
1586         struct cfs_hash_bd     bd;
1587         __u32        version;
1588         int            count = 0;
1589         int            stop_on_change;
1590         int            rc;
1591         int            i;
1592
1593         stop_on_change = cfs_hash_with_rehash_key(hs) ||
1594                          !cfs_hash_with_no_itemref(hs) ||
1595                          hs->hs_ops->hs_put_locked == NULL;
1596         cfs_hash_lock(hs, 0);
1597         LASSERT(!cfs_hash_is_rehashing(hs));
1598
1599         cfs_hash_for_each_bucket(hs, &bd, i) {
1600                 struct hlist_head *hhead;
1601
1602                 cfs_hash_bd_lock(hs, &bd, 0);
1603                 version = cfs_hash_bd_version_get(&bd);
1604
1605                 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1606                         for (hnode = hhead->first; hnode != NULL;) {
1607                                 cfs_hash_bucket_validate(hs, &bd, hnode);
1608                                 cfs_hash_get(hs, hnode);
1609                                 cfs_hash_bd_unlock(hs, &bd, 0);
1610                                 cfs_hash_unlock(hs, 0);
1611
1612                                 rc = func(hs, &bd, hnode, data);
1613                                 if (stop_on_change)
1614                                         cfs_hash_put(hs, hnode);
1615                                 cond_resched();
1616                                 count++;
1617
1618                                 cfs_hash_lock(hs, 0);
1619                                 cfs_hash_bd_lock(hs, &bd, 0);
1620                                 if (!stop_on_change) {
1621                                         tmp = hnode->next;
1622                                         cfs_hash_put_locked(hs, hnode);
1623                                         hnode = tmp;
1624                                 } else { /* bucket changed? */
1625                                         if (version !=
1626                                             cfs_hash_bd_version_get(&bd))
1627                                                 break;
1628                                         /* safe to continue because no change */
1629                                         hnode = hnode->next;
1630                                 }
1631                                 if (rc) /* callback wants to break iteration */
1632                                         break;
1633                         }
1634                         if (rc) /* callback wants to break iteration */
1635                                 break;
1636                 }
1637                 cfs_hash_bd_unlock(hs, &bd, 0);
1638                 if (rc) /* callback wants to break iteration */
1639                         break;
1640         }
1641         cfs_hash_unlock(hs, 0);
1642
1643         return count;
1644 }
1645
1646 int
1647 cfs_hash_for_each_nolock(struct cfs_hash *hs,
1648                          cfs_hash_for_each_cb_t func, void *data) {
1649         if (cfs_hash_with_no_lock(hs) ||
1650             cfs_hash_with_rehash_key(hs) ||
1651             !cfs_hash_with_no_itemref(hs))
1652                 return -EOPNOTSUPP;
1653
1654         if (hs->hs_ops->hs_get == NULL ||
1655             (hs->hs_ops->hs_put == NULL &&
1656              hs->hs_ops->hs_put_locked == NULL))
1657                 return -EOPNOTSUPP;
1658
1659         cfs_hash_for_each_enter(hs);
1660         cfs_hash_for_each_relax(hs, func, data);
1661         cfs_hash_for_each_exit(hs);
1662
1663         return 0;
1664 }
1665 EXPORT_SYMBOL(cfs_hash_for_each_nolock);
1666
1667 /**
1668  * For each hash bucket in the libcfs hash @hs call the passed callback
1669  * @func until all the hash buckets are empty.  The passed callback @func
1670  * or the previously registered callback hs->hs_put must remove the item
1671  * from the hash.  You may either use the cfs_hash_del() or hlist_del()
1672  * functions.  No rwlocks will be held during the callback @func it is
1673  * safe to sleep if needed.  This function will not terminate until the
1674  * hash is empty.  Note it is still possible to concurrently add new
1675  * items in to the hash.  It is the callers responsibility to ensure
1676  * the required locking is in place to prevent concurrent insertions.
1677  */
1678 int
1679 cfs_hash_for_each_empty(struct cfs_hash *hs,
1680                         cfs_hash_for_each_cb_t func, void *data) {
1681         unsigned  i = 0;
1682
1683         if (cfs_hash_with_no_lock(hs))
1684                 return -EOPNOTSUPP;
1685
1686         if (hs->hs_ops->hs_get == NULL ||
1687             (hs->hs_ops->hs_put == NULL &&
1688              hs->hs_ops->hs_put_locked == NULL))
1689                 return -EOPNOTSUPP;
1690
1691         cfs_hash_for_each_enter(hs);
1692         while (cfs_hash_for_each_relax(hs, func, data)) {
1693                 CDEBUG(D_INFO, "Try to empty hash: %s, loop: %u\n",
1694                        hs->hs_name, i++);
1695         }
1696         cfs_hash_for_each_exit(hs);
1697         return 0;
1698 }
1699 EXPORT_SYMBOL(cfs_hash_for_each_empty);
1700
1701 void
1702 cfs_hash_hlist_for_each(struct cfs_hash *hs, unsigned hindex,
1703                         cfs_hash_for_each_cb_t func, void *data)
1704 {
1705         struct hlist_head   *hhead;
1706         struct hlist_node   *hnode;
1707         struct cfs_hash_bd       bd;
1708
1709         cfs_hash_for_each_enter(hs);
1710         cfs_hash_lock(hs, 0);
1711         if (hindex >= CFS_HASH_NHLIST(hs))
1712                 goto out;
1713
1714         cfs_hash_bd_index_set(hs, hindex, &bd);
1715
1716         cfs_hash_bd_lock(hs, &bd, 0);
1717         hhead = cfs_hash_bd_hhead(hs, &bd);
1718         hlist_for_each(hnode, hhead) {
1719                 if (func(hs, &bd, hnode, data))
1720                         break;
1721         }
1722         cfs_hash_bd_unlock(hs, &bd, 0);
1723  out:
1724         cfs_hash_unlock(hs, 0);
1725         cfs_hash_for_each_exit(hs);
1726 }
1727
1728 EXPORT_SYMBOL(cfs_hash_hlist_for_each);
1729
1730 /*
1731  * For each item in the libcfs hash @hs which matches the @key call
1732  * the passed callback @func and pass to it as an argument each hash
1733  * item and the private @data. During the callback the bucket lock
1734  * is held so the callback must never sleep.
1735    */
1736 void
1737 cfs_hash_for_each_key(struct cfs_hash *hs, const void *key,
1738                       cfs_hash_for_each_cb_t func, void *data) {
1739         struct hlist_node   *hnode;
1740         struct cfs_hash_bd       bds[2];
1741         unsigned            i;
1742
1743         cfs_hash_lock(hs, 0);
1744
1745         cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1746
1747         cfs_hash_for_each_bd(bds, 2, i) {
1748                 struct hlist_head *hlist = cfs_hash_bd_hhead(hs, &bds[i]);
1749
1750                 hlist_for_each(hnode, hlist) {
1751                         cfs_hash_bucket_validate(hs, &bds[i], hnode);
1752
1753                         if (cfs_hash_keycmp(hs, key, hnode)) {
1754                                 if (func(hs, &bds[i], hnode, data))
1755                                         break;
1756                         }
1757                 }
1758         }
1759
1760         cfs_hash_dual_bd_unlock(hs, bds, 0);
1761         cfs_hash_unlock(hs, 0);
1762 }
1763 EXPORT_SYMBOL(cfs_hash_for_each_key);
1764
1765 /**
1766  * Rehash the libcfs hash @hs to the given @bits.  This can be used
1767  * to grow the hash size when excessive chaining is detected, or to
1768  * shrink the hash when it is larger than needed.  When the CFS_HASH_REHASH
1769  * flag is set in @hs the libcfs hash may be dynamically rehashed
1770  * during addition or removal if the hash's theta value exceeds
1771  * either the hs->hs_min_theta or hs->max_theta values.  By default
1772  * these values are tuned to keep the chained hash depth small, and
1773  * this approach assumes a reasonably uniform hashing function.  The
1774  * theta thresholds for @hs are tunable via cfs_hash_set_theta().
1775  */
1776 void
1777 cfs_hash_rehash_cancel_locked(struct cfs_hash *hs)
1778 {
1779         int     i;
1780
1781         /* need hold cfs_hash_lock(hs, 1) */
1782         LASSERT(cfs_hash_with_rehash(hs) &&
1783                 !cfs_hash_with_no_lock(hs));
1784
1785         if (!cfs_hash_is_rehashing(hs))
1786                 return;
1787
1788         if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_rehash_wi)) {
1789                 hs->hs_rehash_bits = 0;
1790                 return;
1791         }
1792
1793         for (i = 2; cfs_hash_is_rehashing(hs); i++) {
1794                 cfs_hash_unlock(hs, 1);
1795                 /* raise console warning while waiting too long */
1796                 CDEBUG(IS_PO2(i >> 3) ? D_WARNING : D_INFO,
1797                        "hash %s is still rehashing, rescheded %d\n",
1798                        hs->hs_name, i - 1);
1799                 cond_resched();
1800                 cfs_hash_lock(hs, 1);
1801         }
1802 }
1803 EXPORT_SYMBOL(cfs_hash_rehash_cancel_locked);
1804
1805 void
1806 cfs_hash_rehash_cancel(struct cfs_hash *hs)
1807 {
1808         cfs_hash_lock(hs, 1);
1809         cfs_hash_rehash_cancel_locked(hs);
1810         cfs_hash_unlock(hs, 1);
1811 }
1812 EXPORT_SYMBOL(cfs_hash_rehash_cancel);
1813
1814 int
1815 cfs_hash_rehash(struct cfs_hash *hs, int do_rehash)
1816 {
1817         int     rc;
1818
1819         LASSERT(cfs_hash_with_rehash(hs) && !cfs_hash_with_no_lock(hs));
1820
1821         cfs_hash_lock(hs, 1);
1822
1823         rc = cfs_hash_rehash_bits(hs);
1824         if (rc <= 0) {
1825                 cfs_hash_unlock(hs, 1);
1826                 return rc;
1827         }
1828
1829         hs->hs_rehash_bits = rc;
1830         if (!do_rehash) {
1831                 /* launch and return */
1832                 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_rehash_wi);
1833                 cfs_hash_unlock(hs, 1);
1834                 return 0;
1835         }
1836
1837         /* rehash right now */
1838         cfs_hash_unlock(hs, 1);
1839
1840         return cfs_hash_rehash_worker(&hs->hs_rehash_wi);
1841 }
1842 EXPORT_SYMBOL(cfs_hash_rehash);
1843
1844 static int
1845 cfs_hash_rehash_bd(struct cfs_hash *hs, struct cfs_hash_bd *old)
1846 {
1847         struct cfs_hash_bd      new;
1848         struct hlist_head  *hhead;
1849         struct hlist_node  *hnode;
1850         struct hlist_node  *pos;
1851         void          *key;
1852         int             c = 0;
1853
1854         /* hold cfs_hash_lock(hs, 1), so don't need any bucket lock */
1855         cfs_hash_bd_for_each_hlist(hs, old, hhead) {
1856                 hlist_for_each_safe(hnode, pos, hhead) {
1857                         key = cfs_hash_key(hs, hnode);
1858                         LASSERT(key != NULL);
1859                         /* Validate hnode is in the correct bucket. */
1860                         cfs_hash_bucket_validate(hs, old, hnode);
1861                         /*
1862                          * Delete from old hash bucket; move to new bucket.
1863                          * ops->hs_key must be defined.
1864                          */
1865                         cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
1866                                              hs->hs_rehash_bits, key, &new);
1867                         cfs_hash_bd_move_locked(hs, old, &new, hnode);
1868                         c++;
1869                 }
1870         }
1871
1872         return c;
1873 }
1874
1875 static int
1876 cfs_hash_rehash_worker(cfs_workitem_t *wi)
1877 {
1878         struct cfs_hash  *hs = container_of(wi, struct cfs_hash, hs_rehash_wi);
1879         struct cfs_hash_bucket **bkts;
1880         struct cfs_hash_bd       bd;
1881         unsigned int    old_size;
1882         unsigned int    new_size;
1883         int              bsize;
1884         int              count = 0;
1885         int              rc = 0;
1886         int              i;
1887
1888         LASSERT (hs != NULL && cfs_hash_with_rehash(hs));
1889
1890         cfs_hash_lock(hs, 0);
1891         LASSERT(cfs_hash_is_rehashing(hs));
1892
1893         old_size = CFS_HASH_NBKT(hs);
1894         new_size = CFS_HASH_RH_NBKT(hs);
1895
1896         cfs_hash_unlock(hs, 0);
1897
1898         /*
1899          * don't need hs::hs_rwlock for hs::hs_buckets,
1900          * because nobody can change bkt-table except me.
1901          */
1902         bkts = cfs_hash_buckets_realloc(hs, hs->hs_buckets,
1903                                         old_size, new_size);
1904         cfs_hash_lock(hs, 1);
1905         if (bkts == NULL) {
1906                 rc = -ENOMEM;
1907                 goto out;
1908         }
1909
1910         if (bkts == hs->hs_buckets) {
1911                 bkts = NULL; /* do nothing */
1912                 goto out;
1913         }
1914
1915         rc = __cfs_hash_theta(hs);
1916         if ((rc >= hs->hs_min_theta) && (rc <= hs->hs_max_theta)) {
1917                 /* free the new allocated bkt-table */
1918                 old_size = new_size;
1919                 new_size = CFS_HASH_NBKT(hs);
1920                 rc = -EALREADY;
1921                 goto out;
1922         }
1923
1924         LASSERT(hs->hs_rehash_buckets == NULL);
1925         hs->hs_rehash_buckets = bkts;
1926
1927         rc = 0;
1928         cfs_hash_for_each_bucket(hs, &bd, i) {
1929                 if (cfs_hash_is_exiting(hs)) {
1930                         rc = -ESRCH;
1931                         /* someone wants to destroy the hash, abort now */
1932                         if (old_size < new_size) /* OK to free old bkt-table */
1933                                 break;
1934                         /* it's shrinking, need free new bkt-table */
1935                         hs->hs_rehash_buckets = NULL;
1936                         old_size = new_size;
1937                         new_size = CFS_HASH_NBKT(hs);
1938                         goto out;
1939                 }
1940
1941                 count += cfs_hash_rehash_bd(hs, &bd);
1942                 if (count < CFS_HASH_LOOP_HOG ||
1943                     cfs_hash_is_iterating(hs)) { /* need to finish ASAP */
1944                         continue;
1945                 }
1946
1947                 count = 0;
1948                 cfs_hash_unlock(hs, 1);
1949                 cond_resched();
1950                 cfs_hash_lock(hs, 1);
1951         }
1952
1953         hs->hs_rehash_count++;
1954
1955         bkts = hs->hs_buckets;
1956         hs->hs_buckets = hs->hs_rehash_buckets;
1957         hs->hs_rehash_buckets = NULL;
1958
1959         hs->hs_cur_bits = hs->hs_rehash_bits;
1960  out:
1961         hs->hs_rehash_bits = 0;
1962         if (rc == -ESRCH) /* never be scheduled again */
1963                 cfs_wi_exit(cfs_sched_rehash, wi);
1964         bsize = cfs_hash_bkt_size(hs);
1965         cfs_hash_unlock(hs, 1);
1966         /* can't refer to @hs anymore because it could be destroyed */
1967         if (bkts != NULL)
1968                 cfs_hash_buckets_free(bkts, bsize, new_size, old_size);
1969         if (rc != 0)
1970                 CDEBUG(D_INFO, "early quit of rehashing: %d\n", rc);
1971         /* return 1 only if cfs_wi_exit is called */
1972         return rc == -ESRCH;
1973 }
1974
1975 /**
1976  * Rehash the object referenced by @hnode in the libcfs hash @hs.  The
1977  * @old_key must be provided to locate the objects previous location
1978  * in the hash, and the @new_key will be used to reinsert the object.
1979  * Use this function instead of a cfs_hash_add() + cfs_hash_del()
1980  * combo when it is critical that there is no window in time where the
1981  * object is missing from the hash.  When an object is being rehashed
1982  * the registered cfs_hash_get() and cfs_hash_put() functions will
1983  * not be called.
1984  */
1985 void cfs_hash_rehash_key(struct cfs_hash *hs, const void *old_key,
1986                          void *new_key, struct hlist_node *hnode)
1987 {
1988         struct cfs_hash_bd      bds[3];
1989         struct cfs_hash_bd      old_bds[2];
1990         struct cfs_hash_bd      new_bd;
1991
1992         LASSERT(!hlist_unhashed(hnode));
1993
1994         cfs_hash_lock(hs, 0);
1995
1996         cfs_hash_dual_bd_get(hs, old_key, old_bds);
1997         cfs_hash_bd_get(hs, new_key, &new_bd);
1998
1999         bds[0] = old_bds[0];
2000         bds[1] = old_bds[1];
2001         bds[2] = new_bd;
2002
2003         /* NB: bds[0] and bds[1] are ordered already */
2004         cfs_hash_bd_order(&bds[1], &bds[2]);
2005         cfs_hash_bd_order(&bds[0], &bds[1]);
2006
2007         cfs_hash_multi_bd_lock(hs, bds, 3, 1);
2008         if (likely(old_bds[1].bd_bucket == NULL)) {
2009                 cfs_hash_bd_move_locked(hs, &old_bds[0], &new_bd, hnode);
2010         } else {
2011                 cfs_hash_dual_bd_finddel_locked(hs, old_bds, old_key, hnode);
2012                 cfs_hash_bd_add_locked(hs, &new_bd, hnode);
2013         }
2014         /* overwrite key inside locks, otherwise may screw up with
2015          * other operations, i.e: rehash */
2016         cfs_hash_keycpy(hs, new_key, hnode);
2017
2018         cfs_hash_multi_bd_unlock(hs, bds, 3, 1);
2019         cfs_hash_unlock(hs, 0);
2020 }
2021 EXPORT_SYMBOL(cfs_hash_rehash_key);
2022
2023 void cfs_hash_debug_header(struct seq_file *m)
2024 {
2025         seq_printf(m, "%-*s   cur   min   max theta t-min t-max flags rehash   count  maxdep maxdepb distribution\n",
2026                    CFS_HASH_BIGNAME_LEN, "name");
2027 }
2028 EXPORT_SYMBOL(cfs_hash_debug_header);
2029
2030 static struct cfs_hash_bucket **
2031 cfs_hash_full_bkts(struct cfs_hash *hs)
2032 {
2033         /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2034         if (hs->hs_rehash_buckets == NULL)
2035                 return hs->hs_buckets;
2036
2037         LASSERT(hs->hs_rehash_bits != 0);
2038         return hs->hs_rehash_bits > hs->hs_cur_bits ?
2039                hs->hs_rehash_buckets : hs->hs_buckets;
2040 }
2041
2042 static unsigned int
2043 cfs_hash_full_nbkt(struct cfs_hash *hs)
2044 {
2045         /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2046         if (hs->hs_rehash_buckets == NULL)
2047                 return CFS_HASH_NBKT(hs);
2048
2049         LASSERT(hs->hs_rehash_bits != 0);
2050         return hs->hs_rehash_bits > hs->hs_cur_bits ?
2051                CFS_HASH_RH_NBKT(hs) : CFS_HASH_NBKT(hs);
2052 }
2053
2054 void cfs_hash_debug_str(struct cfs_hash *hs, struct seq_file *m)
2055 {
2056         int                 dist[8] = { 0, };
2057         int                 maxdep  = -1;
2058         int                 maxdepb = -1;
2059         int                 total   = 0;
2060         int                 theta;
2061         int                 i;
2062
2063         cfs_hash_lock(hs, 0);
2064         theta = __cfs_hash_theta(hs);
2065
2066         seq_printf(m, "%-*s %5d %5d %5d %d.%03d %d.%03d %d.%03d  0x%02x %6d ",
2067                       CFS_HASH_BIGNAME_LEN, hs->hs_name,
2068                       1 << hs->hs_cur_bits, 1 << hs->hs_min_bits,
2069                       1 << hs->hs_max_bits,
2070                       __cfs_hash_theta_int(theta), __cfs_hash_theta_frac(theta),
2071                       __cfs_hash_theta_int(hs->hs_min_theta),
2072                       __cfs_hash_theta_frac(hs->hs_min_theta),
2073                       __cfs_hash_theta_int(hs->hs_max_theta),
2074                       __cfs_hash_theta_frac(hs->hs_max_theta),
2075                       hs->hs_flags, hs->hs_rehash_count);
2076
2077         /*
2078          * The distribution is a summary of the chained hash depth in
2079          * each of the libcfs hash buckets.  Each buckets hsb_count is
2080          * divided by the hash theta value and used to generate a
2081          * histogram of the hash distribution.  A uniform hash will
2082          * result in all hash buckets being close to the average thus
2083          * only the first few entries in the histogram will be non-zero.
2084          * If you hash function results in a non-uniform hash the will
2085          * be observable by outlier bucks in the distribution histogram.
2086          *
2087          * Uniform hash distribution:      128/128/0/0/0/0/0/0
2088          * Non-Uniform hash distribution:  128/125/0/0/0/0/2/1
2089          */
2090         for (i = 0; i < cfs_hash_full_nbkt(hs); i++) {
2091                 struct cfs_hash_bd  bd;
2092
2093                 bd.bd_bucket = cfs_hash_full_bkts(hs)[i];
2094                 cfs_hash_bd_lock(hs, &bd, 0);
2095                 if (maxdep < bd.bd_bucket->hsb_depmax) {
2096                         maxdep  = bd.bd_bucket->hsb_depmax;
2097                         maxdepb = ffz(~maxdep);
2098                 }
2099                 total += bd.bd_bucket->hsb_count;
2100                 dist[min(fls(bd.bd_bucket->hsb_count / max(theta, 1)), 7)]++;
2101                 cfs_hash_bd_unlock(hs, &bd, 0);
2102         }
2103
2104         seq_printf(m, "%7d %7d %7d ", total, maxdep, maxdepb);
2105         for (i = 0; i < 8; i++)
2106                 seq_printf(m, "%d%c",  dist[i], (i == 7) ? '\n' : '/');
2107
2108         cfs_hash_unlock(hs, 0);
2109 }
2110 EXPORT_SYMBOL(cfs_hash_debug_str);