4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
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
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * libcfs/libcfs/hash.c
38 * Implement a hash class for hash process in lustre system.
40 * Author: YuZhangyong <yzy@clusterfs.com>
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
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
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.
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
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
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
75 * - removed a lot of unnecessary addref & decref on hash element:
76 * addref & decref are atomic operations in many use-cases which
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.
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.
92 * - support two different refcount modes
93 * . hash table has refcount on element
94 * . hash table doesn't change refcount on adding/removing element
96 * - support long name hash table (for param-tree)
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.
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.
109 #include <linux/seq_file.h>
111 #include <libcfs/libcfs.h>
113 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
114 static unsigned int warn_on_depth = 8;
115 CFS_MODULE_PARM(warn_on_depth, "i", uint, 0644,
116 "warning when hash depth is high.");
119 struct cfs_wi_sched *cfs_sched_rehash;
122 cfs_hash_nl_lock(cfs_hash_lock_t *lock, int exclusive) {}
125 cfs_hash_nl_unlock(cfs_hash_lock_t *lock, int exclusive) {}
128 cfs_hash_spin_lock(cfs_hash_lock_t *lock, int exclusive)
129 __acquires(&lock->spin)
131 spin_lock(&lock->spin);
135 cfs_hash_spin_unlock(cfs_hash_lock_t *lock, int exclusive)
136 __releases(&lock->spin)
138 spin_unlock(&lock->spin);
142 cfs_hash_rw_lock(cfs_hash_lock_t *lock, int exclusive)
143 __acquires(&lock->rw)
146 read_lock(&lock->rw);
148 write_lock(&lock->rw);
152 cfs_hash_rw_unlock(cfs_hash_lock_t *lock, int exclusive)
153 __releases(&lock->rw)
156 read_unlock(&lock->rw);
158 write_unlock(&lock->rw);
162 static cfs_hash_lock_ops_t cfs_hash_nl_lops =
164 .hs_lock = cfs_hash_nl_lock,
165 .hs_unlock = cfs_hash_nl_unlock,
166 .hs_bkt_lock = cfs_hash_nl_lock,
167 .hs_bkt_unlock = cfs_hash_nl_unlock,
170 /** no bucket lock, one spinlock to protect everything */
171 static cfs_hash_lock_ops_t cfs_hash_nbl_lops =
173 .hs_lock = cfs_hash_spin_lock,
174 .hs_unlock = cfs_hash_spin_unlock,
175 .hs_bkt_lock = cfs_hash_nl_lock,
176 .hs_bkt_unlock = cfs_hash_nl_unlock,
179 /** spin bucket lock, rehash is enabled */
180 static cfs_hash_lock_ops_t cfs_hash_bkt_spin_lops =
182 .hs_lock = cfs_hash_rw_lock,
183 .hs_unlock = cfs_hash_rw_unlock,
184 .hs_bkt_lock = cfs_hash_spin_lock,
185 .hs_bkt_unlock = cfs_hash_spin_unlock,
188 /** rw bucket lock, rehash is enabled */
189 static cfs_hash_lock_ops_t cfs_hash_bkt_rw_lops =
191 .hs_lock = cfs_hash_rw_lock,
192 .hs_unlock = cfs_hash_rw_unlock,
193 .hs_bkt_lock = cfs_hash_rw_lock,
194 .hs_bkt_unlock = cfs_hash_rw_unlock,
197 /** spin bucket lock, rehash is disabled */
198 static cfs_hash_lock_ops_t cfs_hash_nr_bkt_spin_lops =
200 .hs_lock = cfs_hash_nl_lock,
201 .hs_unlock = cfs_hash_nl_unlock,
202 .hs_bkt_lock = cfs_hash_spin_lock,
203 .hs_bkt_unlock = cfs_hash_spin_unlock,
206 /** rw bucket lock, rehash is disabled */
207 static cfs_hash_lock_ops_t cfs_hash_nr_bkt_rw_lops =
209 .hs_lock = cfs_hash_nl_lock,
210 .hs_unlock = cfs_hash_nl_unlock,
211 .hs_bkt_lock = cfs_hash_rw_lock,
212 .hs_bkt_unlock = cfs_hash_rw_unlock,
216 cfs_hash_lock_setup(cfs_hash_t *hs)
218 if (cfs_hash_with_no_lock(hs)) {
219 hs->hs_lops = &cfs_hash_nl_lops;
221 } else if (cfs_hash_with_no_bktlock(hs)) {
222 hs->hs_lops = &cfs_hash_nbl_lops;
223 spin_lock_init(&hs->hs_lock.spin);
225 } else if (cfs_hash_with_rehash(hs)) {
226 rwlock_init(&hs->hs_lock.rw);
228 if (cfs_hash_with_rw_bktlock(hs))
229 hs->hs_lops = &cfs_hash_bkt_rw_lops;
230 else if (cfs_hash_with_spin_bktlock(hs))
231 hs->hs_lops = &cfs_hash_bkt_spin_lops;
235 if (cfs_hash_with_rw_bktlock(hs))
236 hs->hs_lops = &cfs_hash_nr_bkt_rw_lops;
237 else if (cfs_hash_with_spin_bktlock(hs))
238 hs->hs_lops = &cfs_hash_nr_bkt_spin_lops;
245 * Simple hash head without depth tracking
246 * new element is always added to head of hlist
249 struct hlist_head hh_head; /**< entries list */
253 cfs_hash_hh_hhead_size(cfs_hash_t *hs)
255 return sizeof(cfs_hash_head_t);
258 static struct hlist_head *
259 cfs_hash_hh_hhead(cfs_hash_t *hs, cfs_hash_bd_t *bd)
261 cfs_hash_head_t *head = (cfs_hash_head_t *)&bd->bd_bucket->hsb_head[0];
263 return &head[bd->bd_offset].hh_head;
267 cfs_hash_hh_hnode_add(cfs_hash_t *hs, cfs_hash_bd_t *bd,
268 struct hlist_node *hnode)
270 hlist_add_head(hnode, cfs_hash_hh_hhead(hs, bd));
271 return -1; /* unknown depth */
275 cfs_hash_hh_hnode_del(cfs_hash_t *hs, cfs_hash_bd_t *bd,
276 struct hlist_node *hnode)
278 hlist_del_init(hnode);
279 return -1; /* unknown depth */
283 * Simple hash head with depth tracking
284 * new element is always added to head of hlist
287 struct hlist_head hd_head; /**< entries list */
288 unsigned int hd_depth; /**< list length */
289 } cfs_hash_head_dep_t;
292 cfs_hash_hd_hhead_size(cfs_hash_t *hs)
294 return sizeof(cfs_hash_head_dep_t);
297 static struct hlist_head *
298 cfs_hash_hd_hhead(cfs_hash_t *hs, cfs_hash_bd_t *bd)
300 cfs_hash_head_dep_t *head;
302 head = (cfs_hash_head_dep_t *)&bd->bd_bucket->hsb_head[0];
303 return &head[bd->bd_offset].hd_head;
307 cfs_hash_hd_hnode_add(cfs_hash_t *hs, cfs_hash_bd_t *bd,
308 struct hlist_node *hnode)
310 cfs_hash_head_dep_t *hh = container_of(cfs_hash_hd_hhead(hs, bd),
311 cfs_hash_head_dep_t, hd_head);
312 hlist_add_head(hnode, &hh->hd_head);
313 return ++hh->hd_depth;
317 cfs_hash_hd_hnode_del(cfs_hash_t *hs, cfs_hash_bd_t *bd,
318 struct hlist_node *hnode)
320 cfs_hash_head_dep_t *hh = container_of(cfs_hash_hd_hhead(hs, bd),
321 cfs_hash_head_dep_t, hd_head);
322 hlist_del_init(hnode);
323 return --hh->hd_depth;
327 * double links hash head without depth tracking
328 * new element is always added to tail of hlist
331 struct hlist_head dh_head; /**< entries list */
332 struct hlist_node *dh_tail; /**< the last entry */
336 cfs_hash_dh_hhead_size(cfs_hash_t *hs)
338 return sizeof(cfs_hash_dhead_t);
341 static struct hlist_head *
342 cfs_hash_dh_hhead(cfs_hash_t *hs, cfs_hash_bd_t *bd)
344 cfs_hash_dhead_t *head;
346 head = (cfs_hash_dhead_t *)&bd->bd_bucket->hsb_head[0];
347 return &head[bd->bd_offset].dh_head;
351 cfs_hash_dh_hnode_add(cfs_hash_t *hs, cfs_hash_bd_t *bd,
352 struct hlist_node *hnode)
354 cfs_hash_dhead_t *dh = container_of(cfs_hash_dh_hhead(hs, bd),
355 cfs_hash_dhead_t, dh_head);
357 if (dh->dh_tail != NULL) /* not empty */
358 hlist_add_after(dh->dh_tail, hnode);
359 else /* empty list */
360 hlist_add_head(hnode, &dh->dh_head);
362 return -1; /* unknown depth */
366 cfs_hash_dh_hnode_del(cfs_hash_t *hs, cfs_hash_bd_t *bd,
367 struct hlist_node *hnd)
369 cfs_hash_dhead_t *dh = container_of(cfs_hash_dh_hhead(hs, bd),
370 cfs_hash_dhead_t, dh_head);
372 if (hnd->next == NULL) { /* it's the tail */
373 dh->dh_tail = (hnd->pprev == &dh->dh_head.first) ? NULL :
374 container_of(hnd->pprev, struct hlist_node, next);
377 return -1; /* unknown depth */
381 * double links hash head with depth tracking
382 * new element is always added to tail of hlist
385 struct hlist_head dd_head; /**< entries list */
386 struct hlist_node *dd_tail; /**< the last entry */
387 unsigned int dd_depth; /**< list length */
388 } cfs_hash_dhead_dep_t;
391 cfs_hash_dd_hhead_size(cfs_hash_t *hs)
393 return sizeof(cfs_hash_dhead_dep_t);
396 static struct hlist_head *
397 cfs_hash_dd_hhead(cfs_hash_t *hs, cfs_hash_bd_t *bd)
399 cfs_hash_dhead_dep_t *head;
401 head = (cfs_hash_dhead_dep_t *)&bd->bd_bucket->hsb_head[0];
402 return &head[bd->bd_offset].dd_head;
406 cfs_hash_dd_hnode_add(cfs_hash_t *hs, cfs_hash_bd_t *bd,
407 struct hlist_node *hnode)
409 cfs_hash_dhead_dep_t *dh = container_of(cfs_hash_dd_hhead(hs, bd),
410 cfs_hash_dhead_dep_t, dd_head);
412 if (dh->dd_tail != NULL) /* not empty */
413 hlist_add_after(dh->dd_tail, hnode);
414 else /* empty list */
415 hlist_add_head(hnode, &dh->dd_head);
417 return ++dh->dd_depth;
421 cfs_hash_dd_hnode_del(cfs_hash_t *hs, cfs_hash_bd_t *bd,
422 struct hlist_node *hnd)
424 cfs_hash_dhead_dep_t *dh = container_of(cfs_hash_dd_hhead(hs, bd),
425 cfs_hash_dhead_dep_t, 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);
432 return --dh->dd_depth;
435 static cfs_hash_hlist_ops_t 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,
442 static cfs_hash_hlist_ops_t 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,
449 static cfs_hash_hlist_ops_t 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,
456 static cfs_hash_hlist_ops_t 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,
464 cfs_hash_hlist_setup(cfs_hash_t *hs)
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;
470 hs->hs_hops = cfs_hash_with_depth(hs) ?
471 &cfs_hash_hd_hops : &cfs_hash_hh_hops;
476 cfs_hash_bd_from_key(cfs_hash_t *hs, cfs_hash_bucket_t **bkts,
477 unsigned int bits, const void *key, cfs_hash_bd_t *bd)
479 unsigned int index = cfs_hash_id(hs, key, (1U << bits) - 1);
481 LASSERT(bits == hs->hs_cur_bits || bits == hs->hs_rehash_bits);
483 bd->bd_bucket = bkts[index & ((1U << (bits - hs->hs_bkt_bits)) - 1)];
484 bd->bd_offset = index >> (bits - hs->hs_bkt_bits);
488 cfs_hash_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bd)
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);
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);
500 EXPORT_SYMBOL(cfs_hash_bd_get);
503 cfs_hash_bd_dep_record(cfs_hash_t *hs, cfs_hash_bd_t *bd, int dep_cur)
505 if (likely(dep_cur <= bd->bd_bucket->hsb_depmax))
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))
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);
521 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_dep_wi);
526 cfs_hash_bd_add_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
527 struct hlist_node *hnode)
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++;
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);
543 EXPORT_SYMBOL(cfs_hash_bd_add_locked);
546 cfs_hash_bd_del_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
547 struct hlist_node *hnode)
549 hs->hs_hops->hop_hnode_del(hs, bd, hnode);
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++;
557 if (cfs_hash_with_counter(hs)) {
558 LASSERT(atomic_read(&hs->hs_count) > 0);
559 atomic_dec(&hs->hs_count);
561 if (!cfs_hash_with_no_itemref(hs))
562 cfs_hash_put_locked(hs, hnode);
564 EXPORT_SYMBOL(cfs_hash_bd_del_locked);
567 cfs_hash_bd_move_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd_old,
568 cfs_hash_bd_t *bd_new, struct hlist_node *hnode)
570 cfs_hash_bucket_t *obkt = bd_old->bd_bucket;
571 cfs_hash_bucket_t *nbkt = bd_new->bd_bucket;
574 if (cfs_hash_bd_compare(bd_old, bd_new) == 0)
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);
583 LASSERT(obkt->hsb_count > 0);
586 if (unlikely(obkt->hsb_version == 0))
590 if (unlikely(nbkt->hsb_version == 0))
593 EXPORT_SYMBOL(cfs_hash_bd_move_locked);
596 /** always set, for sanity (avoid ZERO intent) */
597 CFS_HS_LOOKUP_MASK_FIND = 1 << 0,
598 /** return entry with a ref */
599 CFS_HS_LOOKUP_MASK_REF = 1 << 1,
600 /** add entry if not existing */
601 CFS_HS_LOOKUP_MASK_ADD = 1 << 2,
602 /** delete entry, ignore other masks */
603 CFS_HS_LOOKUP_MASK_DEL = 1 << 3,
606 typedef enum cfs_hash_lookup_intent {
607 /** return item w/o refcount */
608 CFS_HS_LOOKUP_IT_PEEK = CFS_HS_LOOKUP_MASK_FIND,
609 /** return item with refcount */
610 CFS_HS_LOOKUP_IT_FIND = (CFS_HS_LOOKUP_MASK_FIND |
611 CFS_HS_LOOKUP_MASK_REF),
612 /** return item w/o refcount if existed, otherwise add */
613 CFS_HS_LOOKUP_IT_ADD = (CFS_HS_LOOKUP_MASK_FIND |
614 CFS_HS_LOOKUP_MASK_ADD),
615 /** return item with refcount if existed, otherwise add */
616 CFS_HS_LOOKUP_IT_FINDADD = (CFS_HS_LOOKUP_IT_FIND |
617 CFS_HS_LOOKUP_MASK_ADD),
618 /** delete if existed */
619 CFS_HS_LOOKUP_IT_FINDDEL = (CFS_HS_LOOKUP_MASK_FIND |
620 CFS_HS_LOOKUP_MASK_DEL)
621 } cfs_hash_lookup_intent_t;
623 static struct hlist_node *
624 cfs_hash_bd_lookup_intent(cfs_hash_t *hs, cfs_hash_bd_t *bd,
625 const void *key, struct hlist_node *hnode,
626 cfs_hash_lookup_intent_t intent)
629 struct hlist_head *hhead = cfs_hash_bd_hhead(hs, bd);
630 struct hlist_node *ehnode;
631 struct hlist_node *match;
632 int intent_add = (intent & CFS_HS_LOOKUP_MASK_ADD) != 0;
634 /* with this function, we can avoid a lot of useless refcount ops,
635 * which are expensive atomic operations most time. */
636 match = intent_add ? NULL : hnode;
637 hlist_for_each(ehnode, hhead) {
638 if (!cfs_hash_keycmp(hs, key, ehnode))
641 if (match != NULL && match != ehnode) /* can't match */
645 if ((intent & CFS_HS_LOOKUP_MASK_DEL) != 0) {
646 cfs_hash_bd_del_locked(hs, bd, ehnode);
650 /* caller wants refcount? */
651 if ((intent & CFS_HS_LOOKUP_MASK_REF) != 0)
652 cfs_hash_get(hs, ehnode);
659 LASSERT(hnode != NULL);
660 cfs_hash_bd_add_locked(hs, bd, hnode);
665 cfs_hash_bd_lookup_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd, const void *key)
667 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
668 CFS_HS_LOOKUP_IT_FIND);
670 EXPORT_SYMBOL(cfs_hash_bd_lookup_locked);
673 cfs_hash_bd_peek_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd, const void *key)
675 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
676 CFS_HS_LOOKUP_IT_PEEK);
678 EXPORT_SYMBOL(cfs_hash_bd_peek_locked);
681 cfs_hash_bd_findadd_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
682 const void *key, struct hlist_node *hnode,
685 return cfs_hash_bd_lookup_intent(hs, bd, key, hnode,
686 CFS_HS_LOOKUP_IT_ADD |
687 (!noref * CFS_HS_LOOKUP_MASK_REF));
689 EXPORT_SYMBOL(cfs_hash_bd_findadd_locked);
692 cfs_hash_bd_finddel_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
693 const void *key, struct hlist_node *hnode)
695 /* hnode can be NULL, we find the first item with @key */
696 return cfs_hash_bd_lookup_intent(hs, bd, key, hnode,
697 CFS_HS_LOOKUP_IT_FINDDEL);
699 EXPORT_SYMBOL(cfs_hash_bd_finddel_locked);
702 cfs_hash_multi_bd_lock(cfs_hash_t *hs, cfs_hash_bd_t *bds,
703 unsigned n, int excl)
705 cfs_hash_bucket_t *prev = NULL;
709 * bds must be ascendantly ordered by bd->bd_bucket->hsb_index.
710 * NB: it's possible that several bds point to the same bucket but
711 * have different bd::bd_offset, so need take care of deadlock.
713 cfs_hash_for_each_bd(bds, n, i) {
714 if (prev == bds[i].bd_bucket)
717 LASSERT(prev == NULL ||
718 prev->hsb_index < bds[i].bd_bucket->hsb_index);
719 cfs_hash_bd_lock(hs, &bds[i], excl);
720 prev = bds[i].bd_bucket;
725 cfs_hash_multi_bd_unlock(cfs_hash_t *hs, cfs_hash_bd_t *bds,
726 unsigned n, int excl)
728 cfs_hash_bucket_t *prev = NULL;
731 cfs_hash_for_each_bd(bds, n, i) {
732 if (prev != bds[i].bd_bucket) {
733 cfs_hash_bd_unlock(hs, &bds[i], excl);
734 prev = bds[i].bd_bucket;
739 static struct hlist_node *
740 cfs_hash_multi_bd_lookup_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
741 unsigned n, const void *key)
743 struct hlist_node *ehnode;
746 cfs_hash_for_each_bd(bds, n, i) {
747 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, NULL,
748 CFS_HS_LOOKUP_IT_FIND);
755 static struct hlist_node *
756 cfs_hash_multi_bd_findadd_locked(cfs_hash_t *hs,
757 cfs_hash_bd_t *bds, unsigned n, const void *key,
758 struct hlist_node *hnode, int noref)
760 struct hlist_node *ehnode;
764 LASSERT(hnode != NULL);
765 intent = CFS_HS_LOOKUP_IT_PEEK | (!noref * CFS_HS_LOOKUP_MASK_REF);
767 cfs_hash_for_each_bd(bds, n, i) {
768 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key,
774 if (i == 1) { /* only one bucket */
775 cfs_hash_bd_add_locked(hs, &bds[0], hnode);
779 cfs_hash_bd_get(hs, key, &mybd);
780 cfs_hash_bd_add_locked(hs, &mybd, hnode);
786 static struct hlist_node *
787 cfs_hash_multi_bd_finddel_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
788 unsigned n, const void *key,
789 struct hlist_node *hnode)
791 struct hlist_node *ehnode;
794 cfs_hash_for_each_bd(bds, n, i) {
795 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, hnode,
796 CFS_HS_LOOKUP_IT_FINDDEL);
804 cfs_hash_bd_order(cfs_hash_bd_t *bd1, cfs_hash_bd_t *bd2)
808 if (bd2->bd_bucket == NULL)
811 if (bd1->bd_bucket == NULL) {
813 bd2->bd_bucket = NULL;
817 rc = cfs_hash_bd_compare(bd1, bd2);
819 bd2->bd_bucket = NULL;
821 } else if (rc > 0) { /* swab bd1 and bd2 */
831 cfs_hash_dual_bd_get(cfs_hash_t *hs, const void *key, cfs_hash_bd_t *bds)
833 /* NB: caller should hold hs_lock.rw if REHASH is set */
834 cfs_hash_bd_from_key(hs, hs->hs_buckets,
835 hs->hs_cur_bits, key, &bds[0]);
836 if (likely(hs->hs_rehash_buckets == NULL)) {
837 /* no rehash or not rehashing */
838 bds[1].bd_bucket = NULL;
842 LASSERT(hs->hs_rehash_bits != 0);
843 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
844 hs->hs_rehash_bits, key, &bds[1]);
846 cfs_hash_bd_order(&bds[0], &bds[1]);
848 EXPORT_SYMBOL(cfs_hash_dual_bd_get);
851 cfs_hash_dual_bd_lock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl)
853 cfs_hash_multi_bd_lock(hs, bds, 2, excl);
855 EXPORT_SYMBOL(cfs_hash_dual_bd_lock);
858 cfs_hash_dual_bd_unlock(cfs_hash_t *hs, cfs_hash_bd_t *bds, int excl)
860 cfs_hash_multi_bd_unlock(hs, bds, 2, excl);
862 EXPORT_SYMBOL(cfs_hash_dual_bd_unlock);
865 cfs_hash_dual_bd_lookup_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
868 return cfs_hash_multi_bd_lookup_locked(hs, bds, 2, key);
870 EXPORT_SYMBOL(cfs_hash_dual_bd_lookup_locked);
873 cfs_hash_dual_bd_findadd_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
874 const void *key, struct hlist_node *hnode,
877 return cfs_hash_multi_bd_findadd_locked(hs, bds, 2, key,
880 EXPORT_SYMBOL(cfs_hash_dual_bd_findadd_locked);
883 cfs_hash_dual_bd_finddel_locked(cfs_hash_t *hs, cfs_hash_bd_t *bds,
884 const void *key, struct hlist_node *hnode)
886 return cfs_hash_multi_bd_finddel_locked(hs, bds, 2, key, hnode);
888 EXPORT_SYMBOL(cfs_hash_dual_bd_finddel_locked);
891 cfs_hash_buckets_free(cfs_hash_bucket_t **buckets,
892 int bkt_size, int prev_size, int size)
896 for (i = prev_size; i < size; i++) {
897 if (buckets[i] != NULL)
898 LIBCFS_FREE(buckets[i], bkt_size);
901 LIBCFS_FREE(buckets, sizeof(buckets[0]) * size);
905 * Create or grow bucket memory. Return old_buckets if no allocation was
906 * needed, the newly allocated buckets if allocation was needed and
907 * successful, and NULL on error.
909 static cfs_hash_bucket_t **
910 cfs_hash_buckets_realloc(cfs_hash_t *hs, cfs_hash_bucket_t **old_bkts,
911 unsigned int old_size, unsigned int new_size)
913 cfs_hash_bucket_t **new_bkts;
916 LASSERT(old_size == 0 || old_bkts != NULL);
918 if (old_bkts != NULL && old_size == new_size)
921 LIBCFS_ALLOC(new_bkts, sizeof(new_bkts[0]) * new_size);
922 if (new_bkts == NULL)
925 if (old_bkts != NULL) {
926 memcpy(new_bkts, old_bkts,
927 min(old_size, new_size) * sizeof(*old_bkts));
930 for (i = old_size; i < new_size; i++) {
931 struct hlist_head *hhead;
934 LIBCFS_ALLOC(new_bkts[i], cfs_hash_bkt_size(hs));
935 if (new_bkts[i] == NULL) {
936 cfs_hash_buckets_free(new_bkts, cfs_hash_bkt_size(hs),
941 new_bkts[i]->hsb_index = i;
942 new_bkts[i]->hsb_version = 1; /* shouldn't be zero */
943 new_bkts[i]->hsb_depmax = -1; /* unknown */
944 bd.bd_bucket = new_bkts[i];
945 cfs_hash_bd_for_each_hlist(hs, &bd, hhead)
946 INIT_HLIST_HEAD(hhead);
948 if (cfs_hash_with_no_lock(hs) ||
949 cfs_hash_with_no_bktlock(hs))
952 if (cfs_hash_with_rw_bktlock(hs))
953 rwlock_init(&new_bkts[i]->hsb_lock.rw);
954 else if (cfs_hash_with_spin_bktlock(hs))
955 spin_lock_init(&new_bkts[i]->hsb_lock.spin);
957 LBUG(); /* invalid use-case */
963 * Initialize new libcfs hash, where:
964 * @name - Descriptive hash name
965 * @cur_bits - Initial hash table size, in bits
966 * @max_bits - Maximum allowed hash table resize, in bits
967 * @ops - Registered hash table operations
968 * @flags - CFS_HASH_REHASH enable synamic hash resizing
969 * - CFS_HASH_SORT enable chained hash sort
971 static int cfs_hash_rehash_worker(cfs_workitem_t *wi);
973 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
974 static int cfs_hash_dep_print(cfs_workitem_t *wi)
976 cfs_hash_t *hs = container_of(wi, cfs_hash_t, hs_dep_wi);
982 spin_lock(&hs->hs_dep_lock);
983 dep = hs->hs_dep_max;
984 bkt = hs->hs_dep_bkt;
985 off = hs->hs_dep_off;
986 bits = hs->hs_dep_bits;
987 spin_unlock(&hs->hs_dep_lock);
989 LCONSOLE_WARN("#### HASH %s (bits: %d): max depth %d at bucket %d/%d\n",
990 hs->hs_name, bits, dep, bkt, off);
991 spin_lock(&hs->hs_dep_lock);
992 hs->hs_dep_bits = 0; /* mark as workitem done */
993 spin_unlock(&hs->hs_dep_lock);
997 static void cfs_hash_depth_wi_init(cfs_hash_t *hs)
999 spin_lock_init(&hs->hs_dep_lock);
1000 cfs_wi_init(&hs->hs_dep_wi, hs, cfs_hash_dep_print);
1003 static void cfs_hash_depth_wi_cancel(cfs_hash_t *hs)
1005 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_dep_wi))
1008 spin_lock(&hs->hs_dep_lock);
1009 while (hs->hs_dep_bits != 0) {
1010 spin_unlock(&hs->hs_dep_lock);
1012 spin_lock(&hs->hs_dep_lock);
1014 spin_unlock(&hs->hs_dep_lock);
1017 #else /* CFS_HASH_DEBUG_LEVEL < CFS_HASH_DEBUG_1 */
1019 static inline void cfs_hash_depth_wi_init(cfs_hash_t *hs) {}
1020 static inline void cfs_hash_depth_wi_cancel(cfs_hash_t *hs) {}
1022 #endif /* CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1 */
1025 cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
1026 unsigned bkt_bits, unsigned extra_bytes,
1027 unsigned min_theta, unsigned max_theta,
1028 cfs_hash_ops_t *ops, unsigned flags)
1035 CLASSERT(CFS_HASH_THETA_BITS < 15);
1037 LASSERT(name != NULL);
1038 LASSERT(ops != NULL);
1039 LASSERT(ops->hs_key);
1040 LASSERT(ops->hs_hash);
1041 LASSERT(ops->hs_object);
1042 LASSERT(ops->hs_keycmp);
1043 LASSERT(ops->hs_get != NULL);
1044 LASSERT(ops->hs_put_locked != NULL);
1046 if ((flags & CFS_HASH_REHASH) != 0)
1047 flags |= CFS_HASH_COUNTER; /* must have counter */
1049 LASSERT(cur_bits > 0);
1050 LASSERT(cur_bits >= bkt_bits);
1051 LASSERT(max_bits >= cur_bits && max_bits < 31);
1052 LASSERT(ergo((flags & CFS_HASH_REHASH) == 0, cur_bits == max_bits));
1053 LASSERT(ergo((flags & CFS_HASH_REHASH) != 0,
1054 (flags & CFS_HASH_NO_LOCK) == 0));
1055 LASSERT(ergo((flags & CFS_HASH_REHASH_KEY) != 0,
1056 ops->hs_keycpy != NULL));
1058 len = (flags & CFS_HASH_BIGNAME) == 0 ?
1059 CFS_HASH_NAME_LEN : CFS_HASH_BIGNAME_LEN;
1060 LIBCFS_ALLOC(hs, offsetof(cfs_hash_t, hs_name[len]));
1064 strlcpy(hs->hs_name, name, len);
1065 hs->hs_flags = flags;
1067 atomic_set(&hs->hs_refcount, 1);
1068 atomic_set(&hs->hs_count, 0);
1070 cfs_hash_lock_setup(hs);
1071 cfs_hash_hlist_setup(hs);
1073 hs->hs_cur_bits = (__u8)cur_bits;
1074 hs->hs_min_bits = (__u8)cur_bits;
1075 hs->hs_max_bits = (__u8)max_bits;
1076 hs->hs_bkt_bits = (__u8)bkt_bits;
1079 hs->hs_extra_bytes = extra_bytes;
1080 hs->hs_rehash_bits = 0;
1081 cfs_wi_init(&hs->hs_rehash_wi, hs, cfs_hash_rehash_worker);
1082 cfs_hash_depth_wi_init(hs);
1084 if (cfs_hash_with_rehash(hs))
1085 __cfs_hash_set_theta(hs, min_theta, max_theta);
1087 hs->hs_buckets = cfs_hash_buckets_realloc(hs, NULL, 0,
1089 if (hs->hs_buckets != NULL)
1092 LIBCFS_FREE(hs, offsetof(cfs_hash_t, hs_name[len]));
1095 EXPORT_SYMBOL(cfs_hash_create);
1098 * Cleanup libcfs hash @hs.
1101 cfs_hash_destroy(cfs_hash_t *hs)
1103 struct hlist_node *hnode;
1104 struct hlist_node *pos;
1109 LASSERT(hs != NULL);
1110 LASSERT(!cfs_hash_is_exiting(hs) &&
1111 !cfs_hash_is_iterating(hs));
1114 * prohibit further rehashes, don't need any lock because
1115 * I'm the only (last) one can change it.
1118 if (cfs_hash_with_rehash(hs))
1119 cfs_hash_rehash_cancel(hs);
1121 cfs_hash_depth_wi_cancel(hs);
1122 /* rehash should be done/canceled */
1123 LASSERT(hs->hs_buckets != NULL &&
1124 hs->hs_rehash_buckets == NULL);
1126 cfs_hash_for_each_bucket(hs, &bd, i) {
1127 struct hlist_head *hhead;
1129 LASSERT(bd.bd_bucket != NULL);
1130 /* no need to take this lock, just for consistent code */
1131 cfs_hash_bd_lock(hs, &bd, 1);
1133 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1134 hlist_for_each_safe(hnode, pos, hhead) {
1135 LASSERTF(!cfs_hash_with_assert_empty(hs),
1136 "hash %s bucket %u(%u) is not "
1137 " empty: %u items left\n",
1138 hs->hs_name, bd.bd_bucket->hsb_index,
1139 bd.bd_offset, bd.bd_bucket->hsb_count);
1140 /* can't assert key valicate, because we
1141 * can interrupt rehash */
1142 cfs_hash_bd_del_locked(hs, &bd, hnode);
1143 cfs_hash_exit(hs, hnode);
1146 LASSERT(bd.bd_bucket->hsb_count == 0);
1147 cfs_hash_bd_unlock(hs, &bd, 1);
1151 LASSERT(atomic_read(&hs->hs_count) == 0);
1153 cfs_hash_buckets_free(hs->hs_buckets, cfs_hash_bkt_size(hs),
1154 0, CFS_HASH_NBKT(hs));
1155 i = cfs_hash_with_bigname(hs) ?
1156 CFS_HASH_BIGNAME_LEN : CFS_HASH_NAME_LEN;
1157 LIBCFS_FREE(hs, offsetof(cfs_hash_t, hs_name[i]));
1162 cfs_hash_t *cfs_hash_getref(cfs_hash_t *hs)
1164 if (atomic_inc_not_zero(&hs->hs_refcount))
1168 EXPORT_SYMBOL(cfs_hash_getref);
1170 void cfs_hash_putref(cfs_hash_t *hs)
1172 if (atomic_dec_and_test(&hs->hs_refcount))
1173 cfs_hash_destroy(hs);
1175 EXPORT_SYMBOL(cfs_hash_putref);
1178 cfs_hash_rehash_bits(cfs_hash_t *hs)
1180 if (cfs_hash_with_no_lock(hs) ||
1181 !cfs_hash_with_rehash(hs))
1184 if (unlikely(cfs_hash_is_exiting(hs)))
1187 if (unlikely(cfs_hash_is_rehashing(hs)))
1190 if (unlikely(cfs_hash_is_iterating(hs)))
1193 /* XXX: need to handle case with max_theta != 2.0
1194 * and the case with min_theta != 0.5 */
1195 if ((hs->hs_cur_bits < hs->hs_max_bits) &&
1196 (__cfs_hash_theta(hs) > hs->hs_max_theta))
1197 return hs->hs_cur_bits + 1;
1199 if (!cfs_hash_with_shrink(hs))
1202 if ((hs->hs_cur_bits > hs->hs_min_bits) &&
1203 (__cfs_hash_theta(hs) < hs->hs_min_theta))
1204 return hs->hs_cur_bits - 1;
1210 * don't allow inline rehash if:
1211 * - user wants non-blocking change (add/del) on hash table
1212 * - too many elements
1215 cfs_hash_rehash_inline(cfs_hash_t *hs)
1217 return !cfs_hash_with_nblk_change(hs) &&
1218 atomic_read(&hs->hs_count) < CFS_HASH_LOOP_HOG;
1222 * Add item @hnode to libcfs hash @hs using @key. The registered
1223 * ops->hs_get function will be called when the item is added.
1226 cfs_hash_add(cfs_hash_t *hs, const void *key, struct hlist_node *hnode)
1231 LASSERT(hlist_unhashed(hnode));
1233 cfs_hash_lock(hs, 0);
1234 cfs_hash_bd_get_and_lock(hs, key, &bd, 1);
1236 cfs_hash_key_validate(hs, key, hnode);
1237 cfs_hash_bd_add_locked(hs, &bd, hnode);
1239 cfs_hash_bd_unlock(hs, &bd, 1);
1241 bits = cfs_hash_rehash_bits(hs);
1242 cfs_hash_unlock(hs, 0);
1244 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1246 EXPORT_SYMBOL(cfs_hash_add);
1248 static struct hlist_node *
1249 cfs_hash_find_or_add(cfs_hash_t *hs, const void *key,
1250 struct hlist_node *hnode, int noref)
1252 struct hlist_node *ehnode;
1253 cfs_hash_bd_t bds[2];
1256 LASSERT(hlist_unhashed(hnode));
1258 cfs_hash_lock(hs, 0);
1259 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1261 cfs_hash_key_validate(hs, key, hnode);
1262 ehnode = cfs_hash_dual_bd_findadd_locked(hs, bds, key,
1264 cfs_hash_dual_bd_unlock(hs, bds, 1);
1266 if (ehnode == hnode) /* new item added */
1267 bits = cfs_hash_rehash_bits(hs);
1268 cfs_hash_unlock(hs, 0);
1270 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1276 * Add item @hnode to libcfs hash @hs using @key. The registered
1277 * ops->hs_get function will be called if the item was added.
1278 * Returns 0 on success or -EALREADY on key collisions.
1281 cfs_hash_add_unique(cfs_hash_t *hs, const void *key, struct hlist_node *hnode)
1283 return cfs_hash_find_or_add(hs, key, hnode, 1) != hnode ?
1286 EXPORT_SYMBOL(cfs_hash_add_unique);
1289 * Add item @hnode to libcfs hash @hs using @key. If this @key
1290 * already exists in the hash then ops->hs_get will be called on the
1291 * conflicting entry and that entry will be returned to the caller.
1292 * Otherwise ops->hs_get is called on the item which was added.
1295 cfs_hash_findadd_unique(cfs_hash_t *hs, const void *key,
1296 struct hlist_node *hnode)
1298 hnode = cfs_hash_find_or_add(hs, key, hnode, 0);
1300 return cfs_hash_object(hs, hnode);
1302 EXPORT_SYMBOL(cfs_hash_findadd_unique);
1305 * Delete item @hnode from the libcfs hash @hs using @key. The @key
1306 * is required to ensure the correct hash bucket is locked since there
1307 * is no direct linkage from the item to the bucket. The object
1308 * removed from the hash will be returned and obs->hs_put is called
1309 * on the removed object.
1312 cfs_hash_del(cfs_hash_t *hs, const void *key, struct hlist_node *hnode)
1316 cfs_hash_bd_t bds[2];
1318 cfs_hash_lock(hs, 0);
1319 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1321 /* NB: do nothing if @hnode is not in hash table */
1322 if (hnode == NULL || !hlist_unhashed(hnode)) {
1323 if (bds[1].bd_bucket == NULL && hnode != NULL) {
1324 cfs_hash_bd_del_locked(hs, &bds[0], hnode);
1326 hnode = cfs_hash_dual_bd_finddel_locked(hs, bds,
1331 if (hnode != NULL) {
1332 obj = cfs_hash_object(hs, hnode);
1333 bits = cfs_hash_rehash_bits(hs);
1336 cfs_hash_dual_bd_unlock(hs, bds, 1);
1337 cfs_hash_unlock(hs, 0);
1339 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1343 EXPORT_SYMBOL(cfs_hash_del);
1346 * Delete item given @key in libcfs hash @hs. The first @key found in
1347 * the hash will be removed, if the key exists multiple times in the hash
1348 * @hs this function must be called once per key. The removed object
1349 * will be returned and ops->hs_put is called on the removed object.
1352 cfs_hash_del_key(cfs_hash_t *hs, const void *key)
1354 return cfs_hash_del(hs, key, NULL);
1356 EXPORT_SYMBOL(cfs_hash_del_key);
1359 * Lookup an item using @key in the libcfs hash @hs and return it.
1360 * If the @key is found in the hash hs->hs_get() is called and the
1361 * matching objects is returned. It is the callers responsibility
1362 * to call the counterpart ops->hs_put using the cfs_hash_put() macro
1363 * when when finished with the object. If the @key was not found
1364 * in the hash @hs NULL is returned.
1367 cfs_hash_lookup(cfs_hash_t *hs, const void *key)
1370 struct hlist_node *hnode;
1371 cfs_hash_bd_t bds[2];
1373 cfs_hash_lock(hs, 0);
1374 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1376 hnode = cfs_hash_dual_bd_lookup_locked(hs, bds, key);
1378 obj = cfs_hash_object(hs, hnode);
1380 cfs_hash_dual_bd_unlock(hs, bds, 0);
1381 cfs_hash_unlock(hs, 0);
1385 EXPORT_SYMBOL(cfs_hash_lookup);
1388 cfs_hash_for_each_enter(cfs_hash_t *hs)
1390 LASSERT(!cfs_hash_is_exiting(hs));
1392 if (!cfs_hash_with_rehash(hs))
1395 * NB: it's race on cfs_has_t::hs_iterating, but doesn't matter
1396 * because it's just an unreliable signal to rehash-thread,
1397 * rehash-thread will try to finsih rehash ASAP when seeing this.
1399 hs->hs_iterating = 1;
1401 cfs_hash_lock(hs, 1);
1404 /* NB: iteration is mostly called by service thread,
1405 * we tend to cancel pending rehash-requst, instead of
1406 * blocking service thread, we will relaunch rehash request
1407 * after iteration */
1408 if (cfs_hash_is_rehashing(hs))
1409 cfs_hash_rehash_cancel_locked(hs);
1410 cfs_hash_unlock(hs, 1);
1414 cfs_hash_for_each_exit(cfs_hash_t *hs)
1419 if (!cfs_hash_with_rehash(hs))
1421 cfs_hash_lock(hs, 1);
1422 remained = --hs->hs_iterators;
1423 bits = cfs_hash_rehash_bits(hs);
1424 cfs_hash_unlock(hs, 1);
1425 /* NB: it's race on cfs_has_t::hs_iterating, see above */
1427 hs->hs_iterating = 0;
1429 cfs_hash_rehash(hs, atomic_read(&hs->hs_count) <
1435 * For each item in the libcfs hash @hs call the passed callback @func
1436 * and pass to it as an argument each hash item and the private @data.
1438 * a) the function may sleep!
1439 * b) during the callback:
1440 * . the bucket lock is held so the callback must never sleep.
1441 * . if @removal_safe is true, use can remove current item by
1442 * cfs_hash_bd_del_locked
1445 cfs_hash_for_each_tight(cfs_hash_t *hs, cfs_hash_for_each_cb_t func,
1446 void *data, int remove_safe)
1448 struct hlist_node *hnode;
1449 struct hlist_node *pos;
1452 int excl = !!remove_safe;
1457 cfs_hash_for_each_enter(hs);
1459 cfs_hash_lock(hs, 0);
1460 LASSERT(!cfs_hash_is_rehashing(hs));
1462 cfs_hash_for_each_bucket(hs, &bd, i) {
1463 struct hlist_head *hhead;
1465 cfs_hash_bd_lock(hs, &bd, excl);
1466 if (func == NULL) { /* only glimpse size */
1467 count += bd.bd_bucket->hsb_count;
1468 cfs_hash_bd_unlock(hs, &bd, excl);
1472 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1473 hlist_for_each_safe(hnode, pos, hhead) {
1474 cfs_hash_bucket_validate(hs, &bd, hnode);
1477 if (func(hs, &bd, hnode, data)) {
1478 cfs_hash_bd_unlock(hs, &bd, excl);
1483 cfs_hash_bd_unlock(hs, &bd, excl);
1484 if (loop < CFS_HASH_LOOP_HOG)
1487 cfs_hash_unlock(hs, 0);
1489 cfs_hash_lock(hs, 0);
1492 cfs_hash_unlock(hs, 0);
1494 cfs_hash_for_each_exit(hs);
1499 cfs_hash_cond_opt_cb_t func;
1501 } cfs_hash_cond_arg_t;
1504 cfs_hash_cond_del_locked(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1505 struct hlist_node *hnode, void *data)
1507 cfs_hash_cond_arg_t *cond = data;
1509 if (cond->func(cfs_hash_object(hs, hnode), cond->arg))
1510 cfs_hash_bd_del_locked(hs, bd, hnode);
1515 * Delete item from the libcfs hash @hs when @func return true.
1516 * The write lock being hold during loop for each bucket to avoid
1517 * any object be reference.
1520 cfs_hash_cond_del(cfs_hash_t *hs, cfs_hash_cond_opt_cb_t func, void *data)
1522 cfs_hash_cond_arg_t arg = {
1527 cfs_hash_for_each_tight(hs, cfs_hash_cond_del_locked, &arg, 1);
1529 EXPORT_SYMBOL(cfs_hash_cond_del);
1532 cfs_hash_for_each(cfs_hash_t *hs,
1533 cfs_hash_for_each_cb_t func, void *data)
1535 cfs_hash_for_each_tight(hs, func, data, 0);
1537 EXPORT_SYMBOL(cfs_hash_for_each);
1540 cfs_hash_for_each_safe(cfs_hash_t *hs,
1541 cfs_hash_for_each_cb_t func, void *data)
1543 cfs_hash_for_each_tight(hs, func, data, 1);
1545 EXPORT_SYMBOL(cfs_hash_for_each_safe);
1548 cfs_hash_peek(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1549 struct hlist_node *hnode, void *data)
1552 return 1; /* return 1 to break the loop */
1556 cfs_hash_is_empty(cfs_hash_t *hs)
1560 cfs_hash_for_each_tight(hs, cfs_hash_peek, &empty, 0);
1563 EXPORT_SYMBOL(cfs_hash_is_empty);
1566 cfs_hash_size_get(cfs_hash_t *hs)
1568 return cfs_hash_with_counter(hs) ?
1569 atomic_read(&hs->hs_count) :
1570 cfs_hash_for_each_tight(hs, NULL, NULL, 0);
1572 EXPORT_SYMBOL(cfs_hash_size_get);
1575 * cfs_hash_for_each_relax:
1576 * Iterate the hash table and call @func on each item without
1577 * any lock. This function can't guarantee to finish iteration
1578 * if these features are enabled:
1580 * a. if rehash_key is enabled, an item can be moved from
1581 * one bucket to another bucket
1582 * b. user can remove non-zero-ref item from hash-table,
1583 * so the item can be removed from hash-table, even worse,
1584 * it's possible that user changed key and insert to another
1586 * there's no way for us to finish iteration correctly on previous
1587 * two cases, so iteration has to be stopped on change.
1590 cfs_hash_for_each_relax(cfs_hash_t *hs, cfs_hash_for_each_cb_t func, void *data)
1592 struct hlist_node *hnode;
1593 struct hlist_node *tmp;
1602 stop_on_change = cfs_hash_with_rehash_key(hs) ||
1603 !cfs_hash_with_no_itemref(hs) ||
1604 CFS_HOP(hs, put_locked) == NULL;
1605 cfs_hash_lock(hs, 0);
1606 LASSERT(!cfs_hash_is_rehashing(hs));
1608 cfs_hash_for_each_bucket(hs, &bd, i) {
1609 struct hlist_head *hhead;
1611 cfs_hash_bd_lock(hs, &bd, 0);
1612 version = cfs_hash_bd_version_get(&bd);
1614 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1615 for (hnode = hhead->first; hnode != NULL;) {
1616 cfs_hash_bucket_validate(hs, &bd, hnode);
1617 cfs_hash_get(hs, hnode);
1618 cfs_hash_bd_unlock(hs, &bd, 0);
1619 cfs_hash_unlock(hs, 0);
1621 rc = func(hs, &bd, hnode, data);
1623 cfs_hash_put(hs, hnode);
1627 cfs_hash_lock(hs, 0);
1628 cfs_hash_bd_lock(hs, &bd, 0);
1629 if (!stop_on_change) {
1631 cfs_hash_put_locked(hs, hnode);
1633 } else { /* bucket changed? */
1635 cfs_hash_bd_version_get(&bd))
1637 /* safe to continue because no change */
1638 hnode = hnode->next;
1640 if (rc) /* callback wants to break iteration */
1644 cfs_hash_bd_unlock(hs, &bd, 0);
1646 cfs_hash_unlock(hs, 0);
1652 cfs_hash_for_each_nolock(cfs_hash_t *hs,
1653 cfs_hash_for_each_cb_t func, void *data)
1657 if (cfs_hash_with_no_lock(hs) ||
1658 cfs_hash_with_rehash_key(hs) ||
1659 !cfs_hash_with_no_itemref(hs))
1660 RETURN(-EOPNOTSUPP);
1662 if (CFS_HOP(hs, get) == NULL ||
1663 (CFS_HOP(hs, put) == NULL &&
1664 CFS_HOP(hs, put_locked) == NULL))
1665 RETURN(-EOPNOTSUPP);
1667 cfs_hash_for_each_enter(hs);
1668 cfs_hash_for_each_relax(hs, func, data);
1669 cfs_hash_for_each_exit(hs);
1673 EXPORT_SYMBOL(cfs_hash_for_each_nolock);
1676 * For each hash bucket in the libcfs hash @hs call the passed callback
1677 * @func until all the hash buckets are empty. The passed callback @func
1678 * or the previously registered callback hs->hs_put must remove the item
1679 * from the hash. You may either use the cfs_hash_del() or hlist_del()
1680 * functions. No rwlocks will be held during the callback @func it is
1681 * safe to sleep if needed. This function will not terminate until the
1682 * hash is empty. Note it is still possible to concurrently add new
1683 * items in to the hash. It is the callers responsibility to ensure
1684 * the required locking is in place to prevent concurrent insertions.
1687 cfs_hash_for_each_empty(cfs_hash_t *hs,
1688 cfs_hash_for_each_cb_t func, void *data)
1693 if (cfs_hash_with_no_lock(hs))
1696 if (CFS_HOP(hs, get) == NULL ||
1697 (CFS_HOP(hs, put) == NULL &&
1698 CFS_HOP(hs, put_locked) == NULL))
1701 cfs_hash_for_each_enter(hs);
1702 while (cfs_hash_for_each_relax(hs, func, data)) {
1703 CDEBUG(D_INFO, "Try to empty hash: %s, loop: %u\n",
1706 cfs_hash_for_each_exit(hs);
1709 EXPORT_SYMBOL(cfs_hash_for_each_empty);
1712 cfs_hash_hlist_for_each(cfs_hash_t *hs, unsigned hindex,
1713 cfs_hash_for_each_cb_t func, void *data)
1715 struct hlist_head *hhead;
1716 struct hlist_node *hnode;
1719 cfs_hash_for_each_enter(hs);
1720 cfs_hash_lock(hs, 0);
1721 if (hindex >= CFS_HASH_NHLIST(hs))
1724 cfs_hash_bd_index_set(hs, hindex, &bd);
1726 cfs_hash_bd_lock(hs, &bd, 0);
1727 hhead = cfs_hash_bd_hhead(hs, &bd);
1728 hlist_for_each(hnode, hhead) {
1729 if (func(hs, &bd, hnode, data))
1732 cfs_hash_bd_unlock(hs, &bd, 0);
1734 cfs_hash_unlock(hs, 0);
1735 cfs_hash_for_each_exit(hs);
1738 EXPORT_SYMBOL(cfs_hash_hlist_for_each);
1741 * For each item in the libcfs hash @hs which matches the @key call
1742 * the passed callback @func and pass to it as an argument each hash
1743 * item and the private @data. During the callback the bucket lock
1744 * is held so the callback must never sleep.
1747 cfs_hash_for_each_key(cfs_hash_t *hs, const void *key,
1748 cfs_hash_for_each_cb_t func, void *data)
1750 struct hlist_node *hnode;
1751 cfs_hash_bd_t bds[2];
1754 cfs_hash_lock(hs, 0);
1756 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1758 cfs_hash_for_each_bd(bds, 2, i) {
1759 struct hlist_head *hlist = cfs_hash_bd_hhead(hs, &bds[i]);
1761 hlist_for_each(hnode, hlist) {
1762 cfs_hash_bucket_validate(hs, &bds[i], hnode);
1764 if (cfs_hash_keycmp(hs, key, hnode)) {
1765 if (func(hs, &bds[i], hnode, data))
1771 cfs_hash_dual_bd_unlock(hs, bds, 0);
1772 cfs_hash_unlock(hs, 0);
1774 EXPORT_SYMBOL(cfs_hash_for_each_key);
1777 * Rehash the libcfs hash @hs to the given @bits. This can be used
1778 * to grow the hash size when excessive chaining is detected, or to
1779 * shrink the hash when it is larger than needed. When the CFS_HASH_REHASH
1780 * flag is set in @hs the libcfs hash may be dynamically rehashed
1781 * during addition or removal if the hash's theta value exceeds
1782 * either the hs->hs_min_theta or hs->max_theta values. By default
1783 * these values are tuned to keep the chained hash depth small, and
1784 * this approach assumes a reasonably uniform hashing function. The
1785 * theta thresholds for @hs are tunable via cfs_hash_set_theta().
1788 cfs_hash_rehash_cancel_locked(cfs_hash_t *hs)
1792 /* need hold cfs_hash_lock(hs, 1) */
1793 LASSERT(cfs_hash_with_rehash(hs) &&
1794 !cfs_hash_with_no_lock(hs));
1796 if (!cfs_hash_is_rehashing(hs))
1799 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_rehash_wi)) {
1800 hs->hs_rehash_bits = 0;
1804 for (i = 2; cfs_hash_is_rehashing(hs); i++) {
1805 cfs_hash_unlock(hs, 1);
1806 /* raise console warning while waiting too long */
1807 CDEBUG(IS_PO2(i >> 3) ? D_WARNING : D_INFO,
1808 "hash %s is still rehashing, rescheded %d\n",
1809 hs->hs_name, i - 1);
1811 cfs_hash_lock(hs, 1);
1814 EXPORT_SYMBOL(cfs_hash_rehash_cancel_locked);
1817 cfs_hash_rehash_cancel(cfs_hash_t *hs)
1819 cfs_hash_lock(hs, 1);
1820 cfs_hash_rehash_cancel_locked(hs);
1821 cfs_hash_unlock(hs, 1);
1823 EXPORT_SYMBOL(cfs_hash_rehash_cancel);
1826 cfs_hash_rehash(cfs_hash_t *hs, int do_rehash)
1830 LASSERT(cfs_hash_with_rehash(hs) && !cfs_hash_with_no_lock(hs));
1832 cfs_hash_lock(hs, 1);
1834 rc = cfs_hash_rehash_bits(hs);
1836 cfs_hash_unlock(hs, 1);
1840 hs->hs_rehash_bits = rc;
1842 /* launch and return */
1843 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_rehash_wi);
1844 cfs_hash_unlock(hs, 1);
1848 /* rehash right now */
1849 cfs_hash_unlock(hs, 1);
1851 return cfs_hash_rehash_worker(&hs->hs_rehash_wi);
1853 EXPORT_SYMBOL(cfs_hash_rehash);
1856 cfs_hash_rehash_bd(cfs_hash_t *hs, cfs_hash_bd_t *old)
1859 struct hlist_head *hhead;
1860 struct hlist_node *hnode;
1861 struct hlist_node *pos;
1865 /* hold cfs_hash_lock(hs, 1), so don't need any bucket lock */
1866 cfs_hash_bd_for_each_hlist(hs, old, hhead) {
1867 hlist_for_each_safe(hnode, pos, hhead) {
1868 key = cfs_hash_key(hs, hnode);
1869 LASSERT(key != NULL);
1870 /* Validate hnode is in the correct bucket. */
1871 cfs_hash_bucket_validate(hs, old, hnode);
1873 * Delete from old hash bucket; move to new bucket.
1874 * ops->hs_key must be defined.
1876 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
1877 hs->hs_rehash_bits, key, &new);
1878 cfs_hash_bd_move_locked(hs, old, &new, hnode);
1886 cfs_hash_rehash_worker(cfs_workitem_t *wi)
1888 cfs_hash_t *hs = container_of(wi, cfs_hash_t, hs_rehash_wi);
1889 cfs_hash_bucket_t **bkts;
1891 unsigned int old_size;
1892 unsigned int new_size;
1898 LASSERT (hs != NULL && cfs_hash_with_rehash(hs));
1900 cfs_hash_lock(hs, 0);
1901 LASSERT(cfs_hash_is_rehashing(hs));
1903 old_size = CFS_HASH_NBKT(hs);
1904 new_size = CFS_HASH_RH_NBKT(hs);
1906 cfs_hash_unlock(hs, 0);
1909 * don't need hs::hs_rwlock for hs::hs_buckets,
1910 * because nobody can change bkt-table except me.
1912 bkts = cfs_hash_buckets_realloc(hs, hs->hs_buckets,
1913 old_size, new_size);
1914 cfs_hash_lock(hs, 1);
1920 if (bkts == hs->hs_buckets) {
1921 bkts = NULL; /* do nothing */
1925 rc = __cfs_hash_theta(hs);
1926 if ((rc >= hs->hs_min_theta) && (rc <= hs->hs_max_theta)) {
1927 /* free the new allocated bkt-table */
1928 old_size = new_size;
1929 new_size = CFS_HASH_NBKT(hs);
1934 LASSERT(hs->hs_rehash_buckets == NULL);
1935 hs->hs_rehash_buckets = bkts;
1938 cfs_hash_for_each_bucket(hs, &bd, i) {
1939 if (cfs_hash_is_exiting(hs)) {
1941 /* someone wants to destroy the hash, abort now */
1942 if (old_size < new_size) /* OK to free old bkt-table */
1944 /* it's shrinking, need free new bkt-table */
1945 hs->hs_rehash_buckets = NULL;
1946 old_size = new_size;
1947 new_size = CFS_HASH_NBKT(hs);
1951 count += cfs_hash_rehash_bd(hs, &bd);
1952 if (count < CFS_HASH_LOOP_HOG ||
1953 cfs_hash_is_iterating(hs)) { /* need to finish ASAP */
1958 cfs_hash_unlock(hs, 1);
1960 cfs_hash_lock(hs, 1);
1963 hs->hs_rehash_count++;
1965 bkts = hs->hs_buckets;
1966 hs->hs_buckets = hs->hs_rehash_buckets;
1967 hs->hs_rehash_buckets = NULL;
1969 hs->hs_cur_bits = hs->hs_rehash_bits;
1971 hs->hs_rehash_bits = 0;
1972 if (rc == -ESRCH) /* never be scheduled again */
1973 cfs_wi_exit(cfs_sched_rehash, wi);
1974 bsize = cfs_hash_bkt_size(hs);
1975 cfs_hash_unlock(hs, 1);
1976 /* can't refer to @hs anymore because it could be destroyed */
1978 cfs_hash_buckets_free(bkts, bsize, new_size, old_size);
1980 CDEBUG(D_INFO, "early quit of of rehashing: %d\n", rc);
1981 /* return 1 only if cfs_wi_exit is called */
1982 return rc == -ESRCH;
1986 * Rehash the object referenced by @hnode in the libcfs hash @hs. The
1987 * @old_key must be provided to locate the objects previous location
1988 * in the hash, and the @new_key will be used to reinsert the object.
1989 * Use this function instead of a cfs_hash_add() + cfs_hash_del()
1990 * combo when it is critical that there is no window in time where the
1991 * object is missing from the hash. When an object is being rehashed
1992 * the registered cfs_hash_get() and cfs_hash_put() functions will
1995 void cfs_hash_rehash_key(cfs_hash_t *hs, const void *old_key,
1996 void *new_key, struct hlist_node *hnode)
1998 cfs_hash_bd_t bds[3];
1999 cfs_hash_bd_t old_bds[2];
2000 cfs_hash_bd_t new_bd;
2002 LASSERT(!hlist_unhashed(hnode));
2004 cfs_hash_lock(hs, 0);
2006 cfs_hash_dual_bd_get(hs, old_key, old_bds);
2007 cfs_hash_bd_get(hs, new_key, &new_bd);
2009 bds[0] = old_bds[0];
2010 bds[1] = old_bds[1];
2013 /* NB: bds[0] and bds[1] are ordered already */
2014 cfs_hash_bd_order(&bds[1], &bds[2]);
2015 cfs_hash_bd_order(&bds[0], &bds[1]);
2017 cfs_hash_multi_bd_lock(hs, bds, 3, 1);
2018 if (likely(old_bds[1].bd_bucket == NULL)) {
2019 cfs_hash_bd_move_locked(hs, &old_bds[0], &new_bd, hnode);
2021 cfs_hash_dual_bd_finddel_locked(hs, old_bds, old_key, hnode);
2022 cfs_hash_bd_add_locked(hs, &new_bd, hnode);
2024 /* overwrite key inside locks, otherwise may screw up with
2025 * other operations, i.e: rehash */
2026 cfs_hash_keycpy(hs, hnode, new_key);
2028 cfs_hash_multi_bd_unlock(hs, bds, 3, 1);
2029 cfs_hash_unlock(hs, 0);
2031 EXPORT_SYMBOL(cfs_hash_rehash_key);
2033 int cfs_hash_debug_header(struct seq_file *m)
2035 return seq_printf(m, "%-*s%6s%6s%6s%6s%6s%6s%6s%7s%8s%8s%8s%s\n",
2036 CFS_HASH_BIGNAME_LEN,
2037 "name", "cur", "min", "max", "theta", "t-min", "t-max",
2038 "flags", "rehash", "count", "maxdep", "maxdepb",
2041 EXPORT_SYMBOL(cfs_hash_debug_header);
2043 static cfs_hash_bucket_t **
2044 cfs_hash_full_bkts(cfs_hash_t *hs)
2046 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2047 if (hs->hs_rehash_buckets == NULL)
2048 return hs->hs_buckets;
2050 LASSERT(hs->hs_rehash_bits != 0);
2051 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2052 hs->hs_rehash_buckets : hs->hs_buckets;
2056 cfs_hash_full_nbkt(cfs_hash_t *hs)
2058 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2059 if (hs->hs_rehash_buckets == NULL)
2060 return CFS_HASH_NBKT(hs);
2062 LASSERT(hs->hs_rehash_bits != 0);
2063 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2064 CFS_HASH_RH_NBKT(hs) : CFS_HASH_NBKT(hs);
2067 int cfs_hash_debug_str(cfs_hash_t *hs, struct seq_file *m)
2069 int dist[8] = { 0, };
2077 cfs_hash_lock(hs, 0);
2078 theta = __cfs_hash_theta(hs);
2080 c += seq_printf(m, "%-*s ", CFS_HASH_BIGNAME_LEN, hs->hs_name);
2081 c += seq_printf(m, "%5d ", 1 << hs->hs_cur_bits);
2082 c += seq_printf(m, "%5d ", 1 << hs->hs_min_bits);
2083 c += seq_printf(m, "%5d ", 1 << hs->hs_max_bits);
2084 c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(theta),
2085 __cfs_hash_theta_frac(theta));
2086 c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(hs->hs_min_theta),
2087 __cfs_hash_theta_frac(hs->hs_min_theta));
2088 c += seq_printf(m, "%d.%03d ", __cfs_hash_theta_int(hs->hs_max_theta),
2089 __cfs_hash_theta_frac(hs->hs_max_theta));
2090 c += seq_printf(m, " 0x%02x ", hs->hs_flags);
2091 c += seq_printf(m, "%6d ", hs->hs_rehash_count);
2094 * The distribution is a summary of the chained hash depth in
2095 * each of the libcfs hash buckets. Each buckets hsb_count is
2096 * divided by the hash theta value and used to generate a
2097 * histogram of the hash distribution. A uniform hash will
2098 * result in all hash buckets being close to the average thus
2099 * only the first few entries in the histogram will be non-zero.
2100 * If you hash function results in a non-uniform hash the will
2101 * be observable by outlier bucks in the distribution histogram.
2103 * Uniform hash distribution: 128/128/0/0/0/0/0/0
2104 * Non-Uniform hash distribution: 128/125/0/0/0/0/2/1
2106 for (i = 0; i < cfs_hash_full_nbkt(hs); i++) {
2109 bd.bd_bucket = cfs_hash_full_bkts(hs)[i];
2110 cfs_hash_bd_lock(hs, &bd, 0);
2111 if (maxdep < bd.bd_bucket->hsb_depmax) {
2112 maxdep = bd.bd_bucket->hsb_depmax;
2114 maxdepb = ffz(~maxdep);
2117 total += bd.bd_bucket->hsb_count;
2118 dist[min(fls(bd.bd_bucket->hsb_count/max(theta,1)),7)]++;
2119 cfs_hash_bd_unlock(hs, &bd, 0);
2122 c += seq_printf(m, "%7d ", total);
2123 c += seq_printf(m, "%7d ", maxdep);
2124 c += seq_printf(m, "%7d ", maxdepb);
2125 for (i = 0; i < 8; i++)
2126 c += seq_printf(m, "%d%c", dist[i], (i == 7) ? '\n' : '/');
2128 cfs_hash_unlock(hs, 0);
2131 EXPORT_SYMBOL(cfs_hash_debug_str);