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