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