libcfs/libcfs/hash.c

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