libcfs/libcfs/hash.c

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