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, 2015, 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(instead 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 specify bucket size
59 * by bkt_bits of cfs_hash_create(), all hlist_heads in a bucket share
60 * one lock for reducing memory overhead.
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/linux/linux-list.h>
112 #include <libcfs/libcfs.h>
114 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
115 static unsigned int warn_on_depth = 8;
116 module_param(warn_on_depth, uint, 0644);
117 MODULE_PARM_DESC(warn_on_depth, "warning when hash depth is high.");
120 struct cfs_wi_sched *cfs_sched_rehash;
123 cfs_hash_nl_lock(union cfs_hash_lock *lock, int exclusive) {}
126 cfs_hash_nl_unlock(union cfs_hash_lock *lock, int exclusive) {}
129 cfs_hash_spin_lock(union cfs_hash_lock *lock, int exclusive)
130 __acquires(&lock->spin)
132 spin_lock(&lock->spin);
136 cfs_hash_spin_unlock(union cfs_hash_lock *lock, int exclusive)
137 __releases(&lock->spin)
139 spin_unlock(&lock->spin);
143 cfs_hash_rw_lock(union cfs_hash_lock *lock, int exclusive)
144 __acquires(&lock->rw)
147 read_lock(&lock->rw);
149 write_lock(&lock->rw);
153 cfs_hash_rw_unlock(union cfs_hash_lock *lock, int exclusive)
154 __releases(&lock->rw)
157 read_unlock(&lock->rw);
159 write_unlock(&lock->rw);
163 static struct cfs_hash_lock_ops 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 struct cfs_hash_lock_ops cfs_hash_nbl_lops = {
172 .hs_lock = cfs_hash_spin_lock,
173 .hs_unlock = cfs_hash_spin_unlock,
174 .hs_bkt_lock = cfs_hash_nl_lock,
175 .hs_bkt_unlock = cfs_hash_nl_unlock,
178 /** spin bucket lock, rehash is enabled */
179 static struct cfs_hash_lock_ops cfs_hash_bkt_spin_lops = {
180 .hs_lock = cfs_hash_rw_lock,
181 .hs_unlock = cfs_hash_rw_unlock,
182 .hs_bkt_lock = cfs_hash_spin_lock,
183 .hs_bkt_unlock = cfs_hash_spin_unlock,
186 /** rw bucket lock, rehash is enabled */
187 static struct cfs_hash_lock_ops cfs_hash_bkt_rw_lops = {
188 .hs_lock = cfs_hash_rw_lock,
189 .hs_unlock = cfs_hash_rw_unlock,
190 .hs_bkt_lock = cfs_hash_rw_lock,
191 .hs_bkt_unlock = cfs_hash_rw_unlock,
194 /** spin bucket lock, rehash is disabled */
195 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_spin_lops = {
196 .hs_lock = cfs_hash_nl_lock,
197 .hs_unlock = cfs_hash_nl_unlock,
198 .hs_bkt_lock = cfs_hash_spin_lock,
199 .hs_bkt_unlock = cfs_hash_spin_unlock,
202 /** rw bucket lock, rehash is disabled */
203 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_rw_lops = {
204 .hs_lock = cfs_hash_nl_lock,
205 .hs_unlock = cfs_hash_nl_unlock,
206 .hs_bkt_lock = cfs_hash_rw_lock,
207 .hs_bkt_unlock = cfs_hash_rw_unlock,
211 cfs_hash_lock_setup(struct cfs_hash *hs)
213 if (cfs_hash_with_no_lock(hs)) {
214 hs->hs_lops = &cfs_hash_nl_lops;
216 } else if (cfs_hash_with_no_bktlock(hs)) {
217 hs->hs_lops = &cfs_hash_nbl_lops;
218 spin_lock_init(&hs->hs_lock.spin);
220 } else if (cfs_hash_with_rehash(hs)) {
221 rwlock_init(&hs->hs_lock.rw);
223 if (cfs_hash_with_rw_bktlock(hs))
224 hs->hs_lops = &cfs_hash_bkt_rw_lops;
225 else if (cfs_hash_with_spin_bktlock(hs))
226 hs->hs_lops = &cfs_hash_bkt_spin_lops;
230 if (cfs_hash_with_rw_bktlock(hs))
231 hs->hs_lops = &cfs_hash_nr_bkt_rw_lops;
232 else if (cfs_hash_with_spin_bktlock(hs))
233 hs->hs_lops = &cfs_hash_nr_bkt_spin_lops;
240 * Simple hash head without depth tracking
241 * new element is always added to head of hlist
243 struct cfs_hash_head {
244 struct hlist_head hh_head; /**< entries list */
248 cfs_hash_hh_hhead_size(struct cfs_hash *hs)
250 return sizeof(struct cfs_hash_head);
253 static struct hlist_head *
254 cfs_hash_hh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
256 struct cfs_hash_head *head;
258 head = (struct cfs_hash_head *)&bd->bd_bucket->hsb_head[0];
259 return &head[bd->bd_offset].hh_head;
263 cfs_hash_hh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
264 struct hlist_node *hnode)
266 hlist_add_head(hnode, cfs_hash_hh_hhead(hs, bd));
267 return -1; /* unknown depth */
271 cfs_hash_hh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
272 struct hlist_node *hnode)
274 hlist_del_init(hnode);
275 return -1; /* unknown depth */
279 * Simple hash head with depth tracking
280 * new element is always added to head of hlist
282 struct cfs_hash_head_dep {
283 struct hlist_head hd_head; /**< entries list */
284 unsigned int hd_depth; /**< list length */
288 cfs_hash_hd_hhead_size(struct cfs_hash *hs)
290 return sizeof(struct cfs_hash_head_dep);
293 static struct hlist_head *
294 cfs_hash_hd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
296 struct cfs_hash_head_dep *head;
298 head = (struct cfs_hash_head_dep *)&bd->bd_bucket->hsb_head[0];
299 return &head[bd->bd_offset].hd_head;
303 cfs_hash_hd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
304 struct hlist_node *hnode)
306 struct cfs_hash_head_dep *hh;
308 hh = container_of(cfs_hash_hd_hhead(hs, bd),
309 struct cfs_hash_head_dep, hd_head);
310 hlist_add_head(hnode, &hh->hd_head);
311 return ++hh->hd_depth;
315 cfs_hash_hd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
316 struct hlist_node *hnode)
318 struct cfs_hash_head_dep *hh;
320 hh = container_of(cfs_hash_hd_hhead(hs, bd),
321 struct cfs_hash_head_dep, 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
330 struct cfs_hash_dhead {
331 struct hlist_head dh_head; /**< entries list */
332 struct hlist_node *dh_tail; /**< the last entry */
336 cfs_hash_dh_hhead_size(struct cfs_hash *hs)
338 return sizeof(struct cfs_hash_dhead);
341 static struct hlist_head *
342 cfs_hash_dh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
344 struct cfs_hash_dhead *head;
346 head = (struct cfs_hash_dhead *)&bd->bd_bucket->hsb_head[0];
347 return &head[bd->bd_offset].dh_head;
351 cfs_hash_dh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
352 struct hlist_node *hnode)
354 struct cfs_hash_dhead *dh;
356 dh = container_of(cfs_hash_dh_hhead(hs, bd),
357 struct cfs_hash_dhead, dh_head);
358 if (dh->dh_tail != NULL) /* not empty */
359 hlist_add_behind(hnode, dh->dh_tail);
360 else /* empty list */
361 hlist_add_head(hnode, &dh->dh_head);
363 return -1; /* unknown depth */
367 cfs_hash_dh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
368 struct hlist_node *hnd)
370 struct cfs_hash_dhead *dh;
372 dh = container_of(cfs_hash_dh_hhead(hs, bd),
373 struct cfs_hash_dhead, dh_head);
374 if (hnd->next == NULL) { /* it's the tail */
375 dh->dh_tail = (hnd->pprev == &dh->dh_head.first) ? NULL :
376 container_of(hnd->pprev, struct hlist_node, next);
379 return -1; /* unknown depth */
383 * double links hash head with depth tracking
384 * new element is always added to tail of hlist
386 struct cfs_hash_dhead_dep {
387 struct hlist_head dd_head; /**< entries list */
388 struct hlist_node *dd_tail; /**< the last entry */
389 unsigned int dd_depth; /**< list length */
393 cfs_hash_dd_hhead_size(struct cfs_hash *hs)
395 return sizeof(struct cfs_hash_dhead_dep);
398 static struct hlist_head *
399 cfs_hash_dd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
401 struct cfs_hash_dhead_dep *head;
403 head = (struct cfs_hash_dhead_dep *)&bd->bd_bucket->hsb_head[0];
404 return &head[bd->bd_offset].dd_head;
408 cfs_hash_dd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
409 struct hlist_node *hnode)
411 struct cfs_hash_dhead_dep *dh;
413 dh = container_of(cfs_hash_dd_hhead(hs, bd),
414 struct cfs_hash_dhead_dep, dd_head);
415 if (dh->dd_tail != NULL) /* not empty */
416 hlist_add_behind(hnode, dh->dd_tail);
417 else /* empty list */
418 hlist_add_head(hnode, &dh->dd_head);
420 return ++dh->dd_depth;
424 cfs_hash_dd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
425 struct hlist_node *hnd)
427 struct cfs_hash_dhead_dep *dh;
429 dh = container_of(cfs_hash_dd_hhead(hs, bd),
430 struct cfs_hash_dhead_dep, dd_head);
431 if (hnd->next == NULL) { /* it's the tail */
432 dh->dd_tail = (hnd->pprev == &dh->dd_head.first) ? NULL :
433 container_of(hnd->pprev, struct hlist_node, next);
436 return --dh->dd_depth;
439 static struct cfs_hash_hlist_ops cfs_hash_hh_hops = {
440 .hop_hhead = cfs_hash_hh_hhead,
441 .hop_hhead_size = cfs_hash_hh_hhead_size,
442 .hop_hnode_add = cfs_hash_hh_hnode_add,
443 .hop_hnode_del = cfs_hash_hh_hnode_del,
446 static struct cfs_hash_hlist_ops cfs_hash_hd_hops = {
447 .hop_hhead = cfs_hash_hd_hhead,
448 .hop_hhead_size = cfs_hash_hd_hhead_size,
449 .hop_hnode_add = cfs_hash_hd_hnode_add,
450 .hop_hnode_del = cfs_hash_hd_hnode_del,
453 static struct cfs_hash_hlist_ops cfs_hash_dh_hops = {
454 .hop_hhead = cfs_hash_dh_hhead,
455 .hop_hhead_size = cfs_hash_dh_hhead_size,
456 .hop_hnode_add = cfs_hash_dh_hnode_add,
457 .hop_hnode_del = cfs_hash_dh_hnode_del,
460 static struct cfs_hash_hlist_ops cfs_hash_dd_hops = {
461 .hop_hhead = cfs_hash_dd_hhead,
462 .hop_hhead_size = cfs_hash_dd_hhead_size,
463 .hop_hnode_add = cfs_hash_dd_hnode_add,
464 .hop_hnode_del = cfs_hash_dd_hnode_del,
468 cfs_hash_hlist_setup(struct cfs_hash *hs)
470 if (cfs_hash_with_add_tail(hs)) {
471 hs->hs_hops = cfs_hash_with_depth(hs) ?
472 &cfs_hash_dd_hops : &cfs_hash_dh_hops;
474 hs->hs_hops = cfs_hash_with_depth(hs) ?
475 &cfs_hash_hd_hops : &cfs_hash_hh_hops;
480 cfs_hash_bd_from_key(struct cfs_hash *hs, struct cfs_hash_bucket **bkts,
481 unsigned int bits, const void *key, struct cfs_hash_bd *bd)
483 unsigned int index = cfs_hash_id(hs, key, (1U << bits) - 1);
485 LASSERT(bits == hs->hs_cur_bits || bits == hs->hs_rehash_bits);
487 bd->bd_bucket = bkts[index & ((1U << (bits - hs->hs_bkt_bits)) - 1)];
488 bd->bd_offset = index >> (bits - hs->hs_bkt_bits);
492 cfs_hash_bd_get(struct cfs_hash *hs, const void *key, struct cfs_hash_bd *bd)
494 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
495 if (likely(hs->hs_rehash_buckets == NULL)) {
496 cfs_hash_bd_from_key(hs, hs->hs_buckets,
497 hs->hs_cur_bits, key, bd);
499 LASSERT(hs->hs_rehash_bits != 0);
500 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
501 hs->hs_rehash_bits, key, bd);
504 EXPORT_SYMBOL(cfs_hash_bd_get);
507 cfs_hash_bd_dep_record(struct cfs_hash *hs, struct cfs_hash_bd *bd, int dep_cur)
509 if (likely(dep_cur <= bd->bd_bucket->hsb_depmax))
512 bd->bd_bucket->hsb_depmax = dep_cur;
513 # if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
514 if (likely(warn_on_depth == 0 ||
515 max(warn_on_depth, hs->hs_dep_max) >= dep_cur))
518 spin_lock(&hs->hs_dep_lock);
519 hs->hs_dep_max = dep_cur;
520 hs->hs_dep_bkt = bd->bd_bucket->hsb_index;
521 hs->hs_dep_off = bd->bd_offset;
522 hs->hs_dep_bits = hs->hs_cur_bits;
523 spin_unlock(&hs->hs_dep_lock);
525 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_dep_wi);
530 cfs_hash_bd_add_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
531 struct hlist_node *hnode)
535 rc = hs->hs_hops->hop_hnode_add(hs, bd, hnode);
536 cfs_hash_bd_dep_record(hs, bd, rc);
537 bd->bd_bucket->hsb_version++;
538 if (unlikely(bd->bd_bucket->hsb_version == 0))
539 bd->bd_bucket->hsb_version++;
540 bd->bd_bucket->hsb_count++;
542 if (cfs_hash_with_counter(hs))
543 atomic_inc(&hs->hs_count);
544 if (!cfs_hash_with_no_itemref(hs))
545 cfs_hash_get(hs, hnode);
547 EXPORT_SYMBOL(cfs_hash_bd_add_locked);
550 cfs_hash_bd_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
551 struct hlist_node *hnode)
553 hs->hs_hops->hop_hnode_del(hs, bd, hnode);
555 LASSERT(bd->bd_bucket->hsb_count > 0);
556 bd->bd_bucket->hsb_count--;
557 bd->bd_bucket->hsb_version++;
558 if (unlikely(bd->bd_bucket->hsb_version == 0))
559 bd->bd_bucket->hsb_version++;
561 if (cfs_hash_with_counter(hs)) {
562 LASSERT(atomic_read(&hs->hs_count) > 0);
563 atomic_dec(&hs->hs_count);
565 if (!cfs_hash_with_no_itemref(hs))
566 cfs_hash_put_locked(hs, hnode);
568 EXPORT_SYMBOL(cfs_hash_bd_del_locked);
571 cfs_hash_bd_move_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd_old,
572 struct cfs_hash_bd *bd_new, struct hlist_node *hnode)
574 struct cfs_hash_bucket *obkt = bd_old->bd_bucket;
575 struct cfs_hash_bucket *nbkt = bd_new->bd_bucket;
578 if (cfs_hash_bd_compare(bd_old, bd_new) == 0)
581 /* use cfs_hash_bd_hnode_add/del, to avoid atomic & refcount ops
582 * in cfs_hash_bd_del/add_locked */
583 hs->hs_hops->hop_hnode_del(hs, bd_old, hnode);
584 rc = hs->hs_hops->hop_hnode_add(hs, bd_new, hnode);
585 cfs_hash_bd_dep_record(hs, bd_new, rc);
587 LASSERT(obkt->hsb_count > 0);
590 if (unlikely(obkt->hsb_version == 0))
594 if (unlikely(nbkt->hsb_version == 0))
599 /** always set, for sanity (avoid ZERO intent) */
600 CFS_HS_LOOKUP_MASK_FIND = 1 << 0,
601 /** return entry with a ref */
602 CFS_HS_LOOKUP_MASK_REF = 1 << 1,
603 /** add entry if not existing */
604 CFS_HS_LOOKUP_MASK_ADD = 1 << 2,
605 /** delete entry, ignore other masks */
606 CFS_HS_LOOKUP_MASK_DEL = 1 << 3,
609 enum cfs_hash_lookup_intent {
610 /** return item w/o refcount */
611 CFS_HS_LOOKUP_IT_PEEK = CFS_HS_LOOKUP_MASK_FIND,
612 /** return item with refcount */
613 CFS_HS_LOOKUP_IT_FIND = (CFS_HS_LOOKUP_MASK_FIND |
614 CFS_HS_LOOKUP_MASK_REF),
615 /** return item w/o refcount if existed, otherwise add */
616 CFS_HS_LOOKUP_IT_ADD = (CFS_HS_LOOKUP_MASK_FIND |
617 CFS_HS_LOOKUP_MASK_ADD),
618 /** return item with refcount if existed, otherwise add */
619 CFS_HS_LOOKUP_IT_FINDADD = (CFS_HS_LOOKUP_IT_FIND |
620 CFS_HS_LOOKUP_MASK_ADD),
621 /** delete if existed */
622 CFS_HS_LOOKUP_IT_FINDDEL = (CFS_HS_LOOKUP_MASK_FIND |
623 CFS_HS_LOOKUP_MASK_DEL)
626 static struct hlist_node *
627 cfs_hash_bd_lookup_intent(struct cfs_hash *hs, struct cfs_hash_bd *bd,
628 const void *key, struct hlist_node *hnode,
629 enum cfs_hash_lookup_intent intent)
632 struct hlist_head *hhead = cfs_hash_bd_hhead(hs, bd);
633 struct hlist_node *ehnode;
634 struct hlist_node *match;
635 int intent_add = (intent & CFS_HS_LOOKUP_MASK_ADD) != 0;
637 /* with this function, we can avoid a lot of useless refcount ops,
638 * which are expensive atomic operations most time. */
639 match = intent_add ? NULL : hnode;
640 hlist_for_each(ehnode, hhead) {
641 if (!cfs_hash_keycmp(hs, key, ehnode))
644 if (match != NULL && match != ehnode) /* can't match */
648 if ((intent & CFS_HS_LOOKUP_MASK_DEL) != 0) {
649 cfs_hash_bd_del_locked(hs, bd, ehnode);
653 /* caller wants refcount? */
654 if ((intent & CFS_HS_LOOKUP_MASK_REF) != 0)
655 cfs_hash_get(hs, ehnode);
662 LASSERT(hnode != NULL);
663 cfs_hash_bd_add_locked(hs, bd, hnode);
668 cfs_hash_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
671 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
672 CFS_HS_LOOKUP_IT_FIND);
674 EXPORT_SYMBOL(cfs_hash_bd_lookup_locked);
677 cfs_hash_bd_peek_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
680 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
681 CFS_HS_LOOKUP_IT_PEEK);
683 EXPORT_SYMBOL(cfs_hash_bd_peek_locked);
686 cfs_hash_multi_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
687 unsigned n, int excl)
689 struct cfs_hash_bucket *prev = NULL;
693 * bds must be ascendantly ordered by bd->bd_bucket->hsb_index.
694 * NB: it's possible that several bds point to the same bucket but
695 * have different bd::bd_offset, so need take care of deadlock.
697 cfs_hash_for_each_bd(bds, n, i) {
698 if (prev == bds[i].bd_bucket)
701 LASSERT(prev == NULL ||
702 prev->hsb_index < bds[i].bd_bucket->hsb_index);
703 cfs_hash_bd_lock(hs, &bds[i], excl);
704 prev = bds[i].bd_bucket;
709 cfs_hash_multi_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
710 unsigned n, int excl)
712 struct cfs_hash_bucket *prev = NULL;
715 cfs_hash_for_each_bd(bds, n, i) {
716 if (prev != bds[i].bd_bucket) {
717 cfs_hash_bd_unlock(hs, &bds[i], excl);
718 prev = bds[i].bd_bucket;
723 static struct hlist_node *
724 cfs_hash_multi_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
725 unsigned n, const void *key)
727 struct hlist_node *ehnode;
730 cfs_hash_for_each_bd(bds, n, i) {
731 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, NULL,
732 CFS_HS_LOOKUP_IT_FIND);
739 static struct hlist_node *
740 cfs_hash_multi_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
741 unsigned n, const void *key,
742 struct hlist_node *hnode, int noref)
744 struct hlist_node *ehnode;
748 LASSERT(hnode != NULL);
749 intent = CFS_HS_LOOKUP_IT_PEEK | (!noref * CFS_HS_LOOKUP_MASK_REF);
751 cfs_hash_for_each_bd(bds, n, i) {
752 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key,
758 if (i == 1) { /* only one bucket */
759 cfs_hash_bd_add_locked(hs, &bds[0], hnode);
761 struct cfs_hash_bd mybd;
763 cfs_hash_bd_get(hs, key, &mybd);
764 cfs_hash_bd_add_locked(hs, &mybd, hnode);
770 static struct hlist_node *
771 cfs_hash_multi_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
772 unsigned n, const void *key,
773 struct hlist_node *hnode)
775 struct hlist_node *ehnode;
778 cfs_hash_for_each_bd(bds, n, i) {
779 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, hnode,
780 CFS_HS_LOOKUP_IT_FINDDEL);
788 cfs_hash_bd_order(struct cfs_hash_bd *bd1, struct cfs_hash_bd *bd2)
792 if (bd2->bd_bucket == NULL)
795 if (bd1->bd_bucket == NULL) {
797 bd2->bd_bucket = NULL;
801 rc = cfs_hash_bd_compare(bd1, bd2);
803 bd2->bd_bucket = NULL;
805 } else if (rc > 0) { /* swab bd1 and bd2 */
806 struct cfs_hash_bd tmp;
815 cfs_hash_dual_bd_get(struct cfs_hash *hs, const void *key,
816 struct cfs_hash_bd *bds)
818 /* NB: caller should hold hs_lock.rw if REHASH is set */
819 cfs_hash_bd_from_key(hs, hs->hs_buckets,
820 hs->hs_cur_bits, key, &bds[0]);
821 if (likely(hs->hs_rehash_buckets == NULL)) {
822 /* no rehash or not rehashing */
823 bds[1].bd_bucket = NULL;
827 LASSERT(hs->hs_rehash_bits != 0);
828 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
829 hs->hs_rehash_bits, key, &bds[1]);
831 cfs_hash_bd_order(&bds[0], &bds[1]);
835 cfs_hash_dual_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
837 cfs_hash_multi_bd_lock(hs, bds, 2, excl);
841 cfs_hash_dual_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
843 cfs_hash_multi_bd_unlock(hs, bds, 2, excl);
847 cfs_hash_dual_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
850 return cfs_hash_multi_bd_lookup_locked(hs, bds, 2, key);
854 cfs_hash_dual_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
855 const void *key, struct hlist_node *hnode,
858 return cfs_hash_multi_bd_findadd_locked(hs, bds, 2, key,
863 cfs_hash_dual_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
864 const void *key, struct hlist_node *hnode)
866 return cfs_hash_multi_bd_finddel_locked(hs, bds, 2, key, hnode);
870 cfs_hash_buckets_free(struct cfs_hash_bucket **buckets,
871 int bkt_size, int prev_size, int size)
875 for (i = prev_size; i < size; i++) {
876 if (buckets[i] != NULL)
877 LIBCFS_FREE(buckets[i], bkt_size);
880 LIBCFS_FREE(buckets, sizeof(buckets[0]) * size);
884 * Create or grow bucket memory. Return old_buckets if no allocation was
885 * needed, the newly allocated buckets if allocation was needed and
886 * successful, and NULL on error.
888 static struct cfs_hash_bucket **
889 cfs_hash_buckets_realloc(struct cfs_hash *hs, struct cfs_hash_bucket **old_bkts,
890 unsigned int old_size, unsigned int new_size)
892 struct cfs_hash_bucket **new_bkts;
895 LASSERT(old_size == 0 || old_bkts != NULL);
897 if (old_bkts != NULL && old_size == new_size)
900 LIBCFS_ALLOC(new_bkts, sizeof(new_bkts[0]) * new_size);
901 if (new_bkts == NULL)
904 if (old_bkts != NULL) {
905 memcpy(new_bkts, old_bkts,
906 min(old_size, new_size) * sizeof(*old_bkts));
909 for (i = old_size; i < new_size; i++) {
910 struct hlist_head *hhead;
911 struct cfs_hash_bd bd;
913 LIBCFS_ALLOC(new_bkts[i], cfs_hash_bkt_size(hs));
914 if (new_bkts[i] == NULL) {
915 cfs_hash_buckets_free(new_bkts, cfs_hash_bkt_size(hs),
920 new_bkts[i]->hsb_index = i;
921 new_bkts[i]->hsb_version = 1; /* shouldn't be zero */
922 new_bkts[i]->hsb_depmax = -1; /* unknown */
923 bd.bd_bucket = new_bkts[i];
924 cfs_hash_bd_for_each_hlist(hs, &bd, hhead)
925 INIT_HLIST_HEAD(hhead);
927 if (cfs_hash_with_no_lock(hs) ||
928 cfs_hash_with_no_bktlock(hs))
931 if (cfs_hash_with_rw_bktlock(hs))
932 rwlock_init(&new_bkts[i]->hsb_lock.rw);
933 else if (cfs_hash_with_spin_bktlock(hs))
934 spin_lock_init(&new_bkts[i]->hsb_lock.spin);
936 LBUG(); /* invalid use-case */
942 * Initialize new libcfs hash, where:
943 * @name - Descriptive hash name
944 * @cur_bits - Initial hash table size, in bits
945 * @max_bits - Maximum allowed hash table resize, in bits
946 * @ops - Registered hash table operations
947 * @flags - CFS_HASH_REHASH enable synamic hash resizing
948 * - CFS_HASH_SORT enable chained hash sort
950 static int cfs_hash_rehash_worker(struct cfs_workitem *wi);
952 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
953 static int cfs_hash_dep_print(struct cfs_workitem *wi)
955 struct cfs_hash *hs = container_of(wi, struct cfs_hash, hs_dep_wi);
961 spin_lock(&hs->hs_dep_lock);
962 dep = hs->hs_dep_max;
963 bkt = hs->hs_dep_bkt;
964 off = hs->hs_dep_off;
965 bits = hs->hs_dep_bits;
966 spin_unlock(&hs->hs_dep_lock);
968 LCONSOLE_WARN("#### HASH %s (bits: %d): max depth %d at bucket %d/%d\n",
969 hs->hs_name, bits, dep, bkt, off);
970 spin_lock(&hs->hs_dep_lock);
971 hs->hs_dep_bits = 0; /* mark as workitem done */
972 spin_unlock(&hs->hs_dep_lock);
976 static void cfs_hash_depth_wi_init(struct cfs_hash *hs)
978 spin_lock_init(&hs->hs_dep_lock);
979 cfs_wi_init(&hs->hs_dep_wi, hs, cfs_hash_dep_print);
982 static void cfs_hash_depth_wi_cancel(struct cfs_hash *hs)
984 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_dep_wi))
987 spin_lock(&hs->hs_dep_lock);
988 while (hs->hs_dep_bits != 0) {
989 spin_unlock(&hs->hs_dep_lock);
991 spin_lock(&hs->hs_dep_lock);
993 spin_unlock(&hs->hs_dep_lock);
996 #else /* CFS_HASH_DEBUG_LEVEL < CFS_HASH_DEBUG_1 */
998 static inline void cfs_hash_depth_wi_init(struct cfs_hash *hs) {}
999 static inline void cfs_hash_depth_wi_cancel(struct cfs_hash *hs) {}
1001 #endif /* CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1 */
1004 cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
1005 unsigned bkt_bits, unsigned extra_bytes,
1006 unsigned min_theta, unsigned max_theta,
1007 struct cfs_hash_ops *ops, unsigned flags)
1009 struct cfs_hash *hs;
1014 CLASSERT(CFS_HASH_THETA_BITS < 15);
1016 LASSERT(name != NULL);
1017 LASSERT(ops != NULL);
1018 LASSERT(ops->hs_key);
1019 LASSERT(ops->hs_hash);
1020 LASSERT(ops->hs_object);
1021 LASSERT(ops->hs_keycmp);
1022 LASSERT(ops->hs_get != NULL);
1023 LASSERT(ops->hs_put != NULL || ops->hs_put_locked != NULL);
1025 if ((flags & CFS_HASH_REHASH) != 0)
1026 flags |= CFS_HASH_COUNTER; /* must have counter */
1028 LASSERT(cur_bits > 0);
1029 LASSERT(cur_bits >= bkt_bits);
1030 LASSERT(max_bits >= cur_bits && max_bits < 31);
1031 LASSERT(ergo((flags & CFS_HASH_REHASH) == 0, cur_bits == max_bits));
1032 LASSERT(ergo((flags & CFS_HASH_REHASH) != 0,
1033 (flags & CFS_HASH_NO_LOCK) == 0));
1034 LASSERT(ergo((flags & CFS_HASH_REHASH_KEY) != 0,
1035 ops->hs_keycpy != NULL));
1037 len = (flags & CFS_HASH_BIGNAME) == 0 ?
1038 CFS_HASH_NAME_LEN : CFS_HASH_BIGNAME_LEN;
1039 LIBCFS_ALLOC(hs, offsetof(struct cfs_hash, hs_name[len]));
1043 strlcpy(hs->hs_name, name, len);
1044 hs->hs_flags = flags;
1046 atomic_set(&hs->hs_refcount, 1);
1047 atomic_set(&hs->hs_count, 0);
1049 cfs_hash_lock_setup(hs);
1050 cfs_hash_hlist_setup(hs);
1052 hs->hs_cur_bits = (__u8)cur_bits;
1053 hs->hs_min_bits = (__u8)cur_bits;
1054 hs->hs_max_bits = (__u8)max_bits;
1055 hs->hs_bkt_bits = (__u8)bkt_bits;
1058 hs->hs_extra_bytes = extra_bytes;
1059 hs->hs_rehash_bits = 0;
1060 cfs_wi_init(&hs->hs_rehash_wi, hs, cfs_hash_rehash_worker);
1061 cfs_hash_depth_wi_init(hs);
1063 if (cfs_hash_with_rehash(hs))
1064 __cfs_hash_set_theta(hs, min_theta, max_theta);
1066 hs->hs_buckets = cfs_hash_buckets_realloc(hs, NULL, 0,
1068 if (hs->hs_buckets != NULL)
1071 LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[len]));
1074 EXPORT_SYMBOL(cfs_hash_create);
1077 * Cleanup libcfs hash @hs.
1080 cfs_hash_destroy(struct cfs_hash *hs)
1082 struct hlist_node *hnode;
1083 struct hlist_node *pos;
1084 struct cfs_hash_bd bd;
1088 LASSERT(hs != NULL);
1089 LASSERT(!cfs_hash_is_exiting(hs) &&
1090 !cfs_hash_is_iterating(hs));
1093 * prohibit further rehashes, don't need any lock because
1094 * I'm the only (last) one can change it.
1097 if (cfs_hash_with_rehash(hs))
1098 cfs_hash_rehash_cancel(hs);
1100 cfs_hash_depth_wi_cancel(hs);
1101 /* rehash should be done/canceled */
1102 LASSERT(hs->hs_buckets != NULL &&
1103 hs->hs_rehash_buckets == NULL);
1105 cfs_hash_for_each_bucket(hs, &bd, i) {
1106 struct hlist_head *hhead;
1108 LASSERT(bd.bd_bucket != NULL);
1109 /* no need to take this lock, just for consistent code */
1110 cfs_hash_bd_lock(hs, &bd, 1);
1112 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1113 hlist_for_each_safe(hnode, pos, hhead) {
1114 LASSERTF(!cfs_hash_with_assert_empty(hs),
1115 "hash %s bucket %u(%u) is not "
1116 " empty: %u items left\n",
1117 hs->hs_name, bd.bd_bucket->hsb_index,
1118 bd.bd_offset, bd.bd_bucket->hsb_count);
1119 /* can't assert key valicate, because we
1120 * can interrupt rehash */
1121 cfs_hash_bd_del_locked(hs, &bd, hnode);
1122 cfs_hash_exit(hs, hnode);
1125 LASSERT(bd.bd_bucket->hsb_count == 0);
1126 cfs_hash_bd_unlock(hs, &bd, 1);
1130 LASSERT(atomic_read(&hs->hs_count) == 0);
1132 cfs_hash_buckets_free(hs->hs_buckets, cfs_hash_bkt_size(hs),
1133 0, CFS_HASH_NBKT(hs));
1134 i = cfs_hash_with_bigname(hs) ?
1135 CFS_HASH_BIGNAME_LEN : CFS_HASH_NAME_LEN;
1136 LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[i]));
1141 struct cfs_hash *cfs_hash_getref(struct cfs_hash *hs)
1143 if (atomic_inc_not_zero(&hs->hs_refcount))
1147 EXPORT_SYMBOL(cfs_hash_getref);
1149 void cfs_hash_putref(struct cfs_hash *hs)
1151 if (atomic_dec_and_test(&hs->hs_refcount))
1152 cfs_hash_destroy(hs);
1154 EXPORT_SYMBOL(cfs_hash_putref);
1157 cfs_hash_rehash_bits(struct cfs_hash *hs)
1159 if (cfs_hash_with_no_lock(hs) ||
1160 !cfs_hash_with_rehash(hs))
1163 if (unlikely(cfs_hash_is_exiting(hs)))
1166 if (unlikely(cfs_hash_is_rehashing(hs)))
1169 if (unlikely(cfs_hash_is_iterating(hs)))
1172 /* XXX: need to handle case with max_theta != 2.0
1173 * and the case with min_theta != 0.5 */
1174 if ((hs->hs_cur_bits < hs->hs_max_bits) &&
1175 (__cfs_hash_theta(hs) > hs->hs_max_theta))
1176 return hs->hs_cur_bits + 1;
1178 if (!cfs_hash_with_shrink(hs))
1181 if ((hs->hs_cur_bits > hs->hs_min_bits) &&
1182 (__cfs_hash_theta(hs) < hs->hs_min_theta))
1183 return hs->hs_cur_bits - 1;
1189 * don't allow inline rehash if:
1190 * - user wants non-blocking change (add/del) on hash table
1191 * - too many elements
1194 cfs_hash_rehash_inline(struct cfs_hash *hs)
1196 return !cfs_hash_with_nblk_change(hs) &&
1197 atomic_read(&hs->hs_count) < CFS_HASH_LOOP_HOG;
1201 * Add item @hnode to libcfs hash @hs using @key. The registered
1202 * ops->hs_get function will be called when the item is added.
1205 cfs_hash_add(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1207 struct cfs_hash_bd bd;
1210 LASSERT(hlist_unhashed(hnode));
1212 cfs_hash_lock(hs, 0);
1213 cfs_hash_bd_get_and_lock(hs, key, &bd, 1);
1215 cfs_hash_key_validate(hs, key, hnode);
1216 cfs_hash_bd_add_locked(hs, &bd, hnode);
1218 cfs_hash_bd_unlock(hs, &bd, 1);
1220 bits = cfs_hash_rehash_bits(hs);
1221 cfs_hash_unlock(hs, 0);
1223 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1225 EXPORT_SYMBOL(cfs_hash_add);
1227 static struct hlist_node *
1228 cfs_hash_find_or_add(struct cfs_hash *hs, const void *key,
1229 struct hlist_node *hnode, int noref)
1231 struct hlist_node *ehnode;
1232 struct cfs_hash_bd bds[2];
1235 LASSERTF(hlist_unhashed(hnode), "hnode = %p\n", hnode);
1237 cfs_hash_lock(hs, 0);
1238 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1240 cfs_hash_key_validate(hs, key, hnode);
1241 ehnode = cfs_hash_dual_bd_findadd_locked(hs, bds, key,
1243 cfs_hash_dual_bd_unlock(hs, bds, 1);
1245 if (ehnode == hnode) /* new item added */
1246 bits = cfs_hash_rehash_bits(hs);
1247 cfs_hash_unlock(hs, 0);
1249 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1255 * Add item @hnode to libcfs hash @hs using @key. The registered
1256 * ops->hs_get function will be called if the item was added.
1257 * Returns 0 on success or -EALREADY on key collisions.
1260 cfs_hash_add_unique(struct cfs_hash *hs, const void *key,
1261 struct hlist_node *hnode)
1263 return cfs_hash_find_or_add(hs, key, hnode, 1) != hnode ?
1266 EXPORT_SYMBOL(cfs_hash_add_unique);
1269 * Add item @hnode to libcfs hash @hs using @key. If this @key
1270 * already exists in the hash then ops->hs_get will be called on the
1271 * conflicting entry and that entry will be returned to the caller.
1272 * Otherwise ops->hs_get is called on the item which was added.
1275 cfs_hash_findadd_unique(struct cfs_hash *hs, const void *key,
1276 struct hlist_node *hnode)
1278 hnode = cfs_hash_find_or_add(hs, key, hnode, 0);
1280 return cfs_hash_object(hs, hnode);
1282 EXPORT_SYMBOL(cfs_hash_findadd_unique);
1285 * Delete item @hnode from the libcfs hash @hs using @key. The @key
1286 * is required to ensure the correct hash bucket is locked since there
1287 * is no direct linkage from the item to the bucket. The object
1288 * removed from the hash will be returned and obs->hs_put is called
1289 * on the removed object.
1292 cfs_hash_del(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1296 struct cfs_hash_bd bds[2];
1298 cfs_hash_lock(hs, 0);
1299 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1301 /* NB: do nothing if @hnode is not in hash table */
1302 if (hnode == NULL || !hlist_unhashed(hnode)) {
1303 if (bds[1].bd_bucket == NULL && hnode != NULL) {
1304 cfs_hash_bd_del_locked(hs, &bds[0], hnode);
1306 hnode = cfs_hash_dual_bd_finddel_locked(hs, bds,
1311 if (hnode != NULL) {
1312 obj = cfs_hash_object(hs, hnode);
1313 bits = cfs_hash_rehash_bits(hs);
1316 cfs_hash_dual_bd_unlock(hs, bds, 1);
1317 cfs_hash_unlock(hs, 0);
1319 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1323 EXPORT_SYMBOL(cfs_hash_del);
1326 * Delete item given @key in libcfs hash @hs. The first @key found in
1327 * the hash will be removed, if the key exists multiple times in the hash
1328 * @hs this function must be called once per key. The removed object
1329 * will be returned and ops->hs_put is called on the removed object.
1332 cfs_hash_del_key(struct cfs_hash *hs, const void *key)
1334 return cfs_hash_del(hs, key, NULL);
1336 EXPORT_SYMBOL(cfs_hash_del_key);
1339 * Lookup an item using @key in the libcfs hash @hs and return it.
1340 * If the @key is found in the hash hs->hs_get() is called and the
1341 * matching objects is returned. It is the callers responsibility
1342 * to call the counterpart ops->hs_put using the cfs_hash_put() macro
1343 * when when finished with the object. If the @key was not found
1344 * in the hash @hs NULL is returned.
1347 cfs_hash_lookup(struct cfs_hash *hs, const void *key)
1350 struct hlist_node *hnode;
1351 struct cfs_hash_bd bds[2];
1353 cfs_hash_lock(hs, 0);
1354 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1356 hnode = cfs_hash_dual_bd_lookup_locked(hs, bds, key);
1358 obj = cfs_hash_object(hs, hnode);
1360 cfs_hash_dual_bd_unlock(hs, bds, 0);
1361 cfs_hash_unlock(hs, 0);
1365 EXPORT_SYMBOL(cfs_hash_lookup);
1368 cfs_hash_for_each_enter(struct cfs_hash *hs)
1370 LASSERT(!cfs_hash_is_exiting(hs));
1372 if (!cfs_hash_with_rehash(hs))
1375 * NB: it's race on cfs_has_t::hs_iterating, but doesn't matter
1376 * because it's just an unreliable signal to rehash-thread,
1377 * rehash-thread will try to finish rehash ASAP when seeing this.
1379 hs->hs_iterating = 1;
1381 cfs_hash_lock(hs, 1);
1384 /* NB: iteration is mostly called by service thread,
1385 * we tend to cancel pending rehash-request, instead of
1386 * blocking service thread, we will relaunch rehash request
1387 * after iteration */
1388 if (cfs_hash_is_rehashing(hs))
1389 cfs_hash_rehash_cancel_locked(hs);
1390 cfs_hash_unlock(hs, 1);
1394 cfs_hash_for_each_exit(struct cfs_hash *hs)
1399 if (!cfs_hash_with_rehash(hs))
1401 cfs_hash_lock(hs, 1);
1402 remained = --hs->hs_iterators;
1403 bits = cfs_hash_rehash_bits(hs);
1404 cfs_hash_unlock(hs, 1);
1405 /* NB: it's race on cfs_has_t::hs_iterating, see above */
1407 hs->hs_iterating = 0;
1409 cfs_hash_rehash(hs, atomic_read(&hs->hs_count) <
1415 * For each item in the libcfs hash @hs call the passed callback @func
1416 * and pass to it as an argument each hash item and the private @data.
1418 * a) the function may sleep!
1419 * b) during the callback:
1420 * . the bucket lock is held so the callback must never sleep.
1421 * . if @removal_safe is true, use can remove current item by
1422 * cfs_hash_bd_del_locked
1425 cfs_hash_for_each_tight(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1426 void *data, int remove_safe)
1428 struct hlist_node *hnode;
1429 struct hlist_node *pos;
1430 struct cfs_hash_bd bd;
1432 int excl = !!remove_safe;
1437 cfs_hash_for_each_enter(hs);
1439 cfs_hash_lock(hs, 0);
1440 LASSERT(!cfs_hash_is_rehashing(hs));
1442 cfs_hash_for_each_bucket(hs, &bd, i) {
1443 struct hlist_head *hhead;
1445 cfs_hash_bd_lock(hs, &bd, excl);
1446 if (func == NULL) { /* only glimpse size */
1447 count += bd.bd_bucket->hsb_count;
1448 cfs_hash_bd_unlock(hs, &bd, excl);
1452 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1453 hlist_for_each_safe(hnode, pos, hhead) {
1454 cfs_hash_bucket_validate(hs, &bd, hnode);
1457 if (func(hs, &bd, hnode, data)) {
1458 cfs_hash_bd_unlock(hs, &bd, excl);
1463 cfs_hash_bd_unlock(hs, &bd, excl);
1464 if (loop < CFS_HASH_LOOP_HOG)
1467 cfs_hash_unlock(hs, 0);
1469 cfs_hash_lock(hs, 0);
1472 cfs_hash_unlock(hs, 0);
1474 cfs_hash_for_each_exit(hs);
1478 struct cfs_hash_cond_arg {
1479 cfs_hash_cond_opt_cb_t func;
1484 cfs_hash_cond_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1485 struct hlist_node *hnode, void *data)
1487 struct cfs_hash_cond_arg *cond = data;
1489 if (cond->func(cfs_hash_object(hs, hnode), cond->arg))
1490 cfs_hash_bd_del_locked(hs, bd, hnode);
1495 * Delete item from the libcfs hash @hs when @func return true.
1496 * The write lock being hold during loop for each bucket to avoid
1497 * any object be reference.
1500 cfs_hash_cond_del(struct cfs_hash *hs, cfs_hash_cond_opt_cb_t func, void *data)
1502 struct cfs_hash_cond_arg arg = {
1507 cfs_hash_for_each_tight(hs, cfs_hash_cond_del_locked, &arg, 1);
1509 EXPORT_SYMBOL(cfs_hash_cond_del);
1512 cfs_hash_for_each(struct cfs_hash *hs,
1513 cfs_hash_for_each_cb_t func, void *data)
1515 cfs_hash_for_each_tight(hs, func, data, 0);
1517 EXPORT_SYMBOL(cfs_hash_for_each);
1520 cfs_hash_for_each_safe(struct cfs_hash *hs,
1521 cfs_hash_for_each_cb_t func, void *data)
1523 cfs_hash_for_each_tight(hs, func, data, 1);
1525 EXPORT_SYMBOL(cfs_hash_for_each_safe);
1528 cfs_hash_peek(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1529 struct hlist_node *hnode, void *data)
1532 return 1; /* return 1 to break the loop */
1536 cfs_hash_is_empty(struct cfs_hash *hs)
1540 cfs_hash_for_each_tight(hs, cfs_hash_peek, &empty, 0);
1543 EXPORT_SYMBOL(cfs_hash_is_empty);
1546 cfs_hash_size_get(struct cfs_hash *hs)
1548 return cfs_hash_with_counter(hs) ?
1549 atomic_read(&hs->hs_count) :
1550 cfs_hash_for_each_tight(hs, NULL, NULL, 0);
1552 EXPORT_SYMBOL(cfs_hash_size_get);
1555 * cfs_hash_for_each_relax:
1556 * Iterate the hash table and call @func on each item without
1557 * any lock. This function can't guarantee to finish iteration
1558 * if these features are enabled:
1560 * a. if rehash_key is enabled, an item can be moved from
1561 * one bucket to another bucket
1562 * b. user can remove non-zero-ref item from hash-table,
1563 * so the item can be removed from hash-table, even worse,
1564 * it's possible that user changed key and insert to another
1566 * there's no way for us to finish iteration correctly on previous
1567 * two cases, so iteration has to be stopped on change.
1570 cfs_hash_for_each_relax(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1571 void *data, int start)
1573 struct hlist_node *hnode;
1574 struct hlist_node *next = NULL;
1575 struct cfs_hash_bd bd;
1584 stop_on_change = cfs_hash_with_rehash_key(hs) ||
1585 !cfs_hash_with_no_itemref(hs);
1586 has_put_locked = hs->hs_ops->hs_put_locked != NULL;
1587 cfs_hash_lock(hs, 0);
1589 LASSERT(!cfs_hash_is_rehashing(hs));
1591 cfs_hash_for_each_bucket(hs, &bd, i) {
1592 struct hlist_head *hhead;
1596 else if (end > 0 && i >= end)
1599 cfs_hash_bd_lock(hs, &bd, 0);
1600 version = cfs_hash_bd_version_get(&bd);
1602 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1603 hnode = hhead->first;
1606 cfs_hash_get(hs, hnode);
1607 for (; hnode != NULL; hnode = next) {
1608 cfs_hash_bucket_validate(hs, &bd, hnode);
1611 cfs_hash_get(hs, next);
1612 cfs_hash_bd_unlock(hs, &bd, 0);
1613 cfs_hash_unlock(hs, 0);
1615 rc = func(hs, &bd, hnode, data);
1616 if (stop_on_change || !has_put_locked)
1617 cfs_hash_put(hs, hnode);
1622 cfs_hash_lock(hs, 0);
1623 cfs_hash_bd_lock(hs, &bd, 0);
1624 if (stop_on_change) {
1626 cfs_hash_bd_version_get(&bd))
1628 } else if (has_put_locked) {
1629 cfs_hash_put_locked(hs, hnode);
1631 if (rc) /* callback wants to break iteration */
1635 if (has_put_locked) {
1636 cfs_hash_put_locked(hs, next);
1640 } else if (rc != 0) {
1644 cfs_hash_bd_unlock(hs, &bd, 0);
1645 if (next != NULL && !has_put_locked) {
1646 cfs_hash_put(hs, next);
1649 if (rc) /* callback wants to break iteration */
1653 if (start > 0 && rc == 0) {
1659 cfs_hash_unlock(hs, 0);
1664 cfs_hash_for_each_nolock(struct cfs_hash *hs,
1665 cfs_hash_for_each_cb_t func, void *data, int start)
1669 if (cfs_hash_with_no_lock(hs) ||
1670 cfs_hash_with_rehash_key(hs) ||
1671 !cfs_hash_with_no_itemref(hs))
1672 RETURN(-EOPNOTSUPP);
1674 if (hs->hs_ops->hs_get == NULL ||
1675 (hs->hs_ops->hs_put == NULL &&
1676 hs->hs_ops->hs_put_locked == NULL))
1677 RETURN(-EOPNOTSUPP);
1679 cfs_hash_for_each_enter(hs);
1680 cfs_hash_for_each_relax(hs, func, data, start);
1681 cfs_hash_for_each_exit(hs);
1685 EXPORT_SYMBOL(cfs_hash_for_each_nolock);
1688 * For each hash bucket in the libcfs hash @hs call the passed callback
1689 * @func until all the hash buckets are empty. The passed callback @func
1690 * or the previously registered callback hs->hs_put must remove the item
1691 * from the hash. You may either use the cfs_hash_del() or hlist_del()
1692 * functions. No rwlocks will be held during the callback @func it is
1693 * safe to sleep if needed. This function will not terminate until the
1694 * hash is empty. Note it is still possible to concurrently add new
1695 * items in to the hash. It is the callers responsibility to ensure
1696 * the required locking is in place to prevent concurrent insertions.
1699 cfs_hash_for_each_empty(struct cfs_hash *hs,
1700 cfs_hash_for_each_cb_t func, void *data)
1705 if (cfs_hash_with_no_lock(hs))
1708 if (hs->hs_ops->hs_get == NULL ||
1709 (hs->hs_ops->hs_put == NULL &&
1710 hs->hs_ops->hs_put_locked == NULL))
1713 cfs_hash_for_each_enter(hs);
1714 while (cfs_hash_for_each_relax(hs, func, data, 0)) {
1715 CDEBUG(D_INFO, "Try to empty hash: %s, loop: %u\n",
1718 cfs_hash_for_each_exit(hs);
1721 EXPORT_SYMBOL(cfs_hash_for_each_empty);
1724 cfs_hash_hlist_for_each(struct cfs_hash *hs, unsigned hindex,
1725 cfs_hash_for_each_cb_t func, void *data)
1727 struct hlist_head *hhead;
1728 struct hlist_node *hnode;
1729 struct cfs_hash_bd bd;
1731 cfs_hash_for_each_enter(hs);
1732 cfs_hash_lock(hs, 0);
1733 if (hindex >= CFS_HASH_NHLIST(hs))
1736 cfs_hash_bd_index_set(hs, hindex, &bd);
1738 cfs_hash_bd_lock(hs, &bd, 0);
1739 hhead = cfs_hash_bd_hhead(hs, &bd);
1740 hlist_for_each(hnode, hhead) {
1741 if (func(hs, &bd, hnode, data))
1744 cfs_hash_bd_unlock(hs, &bd, 0);
1746 cfs_hash_unlock(hs, 0);
1747 cfs_hash_for_each_exit(hs);
1750 EXPORT_SYMBOL(cfs_hash_hlist_for_each);
1753 * For each item in the libcfs hash @hs which matches the @key call
1754 * the passed callback @func and pass to it as an argument each hash
1755 * item and the private @data. During the callback the bucket lock
1756 * is held so the callback must never sleep.
1759 cfs_hash_for_each_key(struct cfs_hash *hs, const void *key,
1760 cfs_hash_for_each_cb_t func, void *data)
1762 struct hlist_node *hnode;
1763 struct cfs_hash_bd bds[2];
1766 cfs_hash_lock(hs, 0);
1768 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1770 cfs_hash_for_each_bd(bds, 2, i) {
1771 struct hlist_head *hlist = cfs_hash_bd_hhead(hs, &bds[i]);
1773 hlist_for_each(hnode, hlist) {
1774 cfs_hash_bucket_validate(hs, &bds[i], hnode);
1776 if (cfs_hash_keycmp(hs, key, hnode)) {
1777 if (func(hs, &bds[i], hnode, data))
1783 cfs_hash_dual_bd_unlock(hs, bds, 0);
1784 cfs_hash_unlock(hs, 0);
1786 EXPORT_SYMBOL(cfs_hash_for_each_key);
1789 * Rehash the libcfs hash @hs to the given @bits. This can be used
1790 * to grow the hash size when excessive chaining is detected, or to
1791 * shrink the hash when it is larger than needed. When the CFS_HASH_REHASH
1792 * flag is set in @hs the libcfs hash may be dynamically rehashed
1793 * during addition or removal if the hash's theta value exceeds
1794 * either the hs->hs_min_theta or hs->max_theta values. By default
1795 * these values are tuned to keep the chained hash depth small, and
1796 * this approach assumes a reasonably uniform hashing function. The
1797 * theta thresholds for @hs are tunable via cfs_hash_set_theta().
1800 cfs_hash_rehash_cancel_locked(struct cfs_hash *hs)
1804 /* need hold cfs_hash_lock(hs, 1) */
1805 LASSERT(cfs_hash_with_rehash(hs) &&
1806 !cfs_hash_with_no_lock(hs));
1808 if (!cfs_hash_is_rehashing(hs))
1811 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_rehash_wi)) {
1812 hs->hs_rehash_bits = 0;
1816 for (i = 2; cfs_hash_is_rehashing(hs); i++) {
1817 cfs_hash_unlock(hs, 1);
1818 /* raise console warning while waiting too long */
1819 CDEBUG(IS_PO2(i >> 3) ? D_WARNING : D_INFO,
1820 "hash %s is still rehashing, rescheded %d\n",
1821 hs->hs_name, i - 1);
1823 cfs_hash_lock(hs, 1);
1828 cfs_hash_rehash_cancel(struct cfs_hash *hs)
1830 cfs_hash_lock(hs, 1);
1831 cfs_hash_rehash_cancel_locked(hs);
1832 cfs_hash_unlock(hs, 1);
1836 cfs_hash_rehash(struct cfs_hash *hs, int do_rehash)
1840 LASSERT(cfs_hash_with_rehash(hs) && !cfs_hash_with_no_lock(hs));
1842 cfs_hash_lock(hs, 1);
1844 rc = cfs_hash_rehash_bits(hs);
1846 cfs_hash_unlock(hs, 1);
1850 hs->hs_rehash_bits = rc;
1852 /* launch and return */
1853 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_rehash_wi);
1854 cfs_hash_unlock(hs, 1);
1858 /* rehash right now */
1859 cfs_hash_unlock(hs, 1);
1861 return cfs_hash_rehash_worker(&hs->hs_rehash_wi);
1865 cfs_hash_rehash_bd(struct cfs_hash *hs, struct cfs_hash_bd *old)
1867 struct cfs_hash_bd new;
1868 struct hlist_head *hhead;
1869 struct hlist_node *hnode;
1870 struct hlist_node *pos;
1874 /* hold cfs_hash_lock(hs, 1), so don't need any bucket lock */
1875 cfs_hash_bd_for_each_hlist(hs, old, hhead) {
1876 hlist_for_each_safe(hnode, pos, hhead) {
1877 key = cfs_hash_key(hs, hnode);
1878 LASSERT(key != NULL);
1879 /* Validate hnode is in the correct bucket. */
1880 cfs_hash_bucket_validate(hs, old, hnode);
1882 * Delete from old hash bucket; move to new bucket.
1883 * ops->hs_key must be defined.
1885 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
1886 hs->hs_rehash_bits, key, &new);
1887 cfs_hash_bd_move_locked(hs, old, &new, hnode);
1895 cfs_hash_rehash_worker(struct cfs_workitem *wi)
1897 struct cfs_hash *hs =
1898 container_of(wi, struct cfs_hash, hs_rehash_wi);
1899 struct cfs_hash_bucket **bkts;
1900 struct cfs_hash_bd bd;
1901 unsigned int old_size;
1902 unsigned int new_size;
1908 LASSERT(hs != NULL && cfs_hash_with_rehash(hs));
1910 cfs_hash_lock(hs, 0);
1911 LASSERT(cfs_hash_is_rehashing(hs));
1913 old_size = CFS_HASH_NBKT(hs);
1914 new_size = CFS_HASH_RH_NBKT(hs);
1916 cfs_hash_unlock(hs, 0);
1919 * don't need hs::hs_rwlock for hs::hs_buckets,
1920 * because nobody can change bkt-table except me.
1922 bkts = cfs_hash_buckets_realloc(hs, hs->hs_buckets,
1923 old_size, new_size);
1924 cfs_hash_lock(hs, 1);
1930 if (bkts == hs->hs_buckets) {
1931 bkts = NULL; /* do nothing */
1935 rc = __cfs_hash_theta(hs);
1936 if ((rc >= hs->hs_min_theta) && (rc <= hs->hs_max_theta)) {
1937 /* free the new allocated bkt-table */
1938 old_size = new_size;
1939 new_size = CFS_HASH_NBKT(hs);
1944 LASSERT(hs->hs_rehash_buckets == NULL);
1945 hs->hs_rehash_buckets = bkts;
1948 cfs_hash_for_each_bucket(hs, &bd, i) {
1949 if (cfs_hash_is_exiting(hs)) {
1951 /* someone wants to destroy the hash, abort now */
1952 if (old_size < new_size) /* OK to free old bkt-table */
1954 /* it's shrinking, need free new bkt-table */
1955 hs->hs_rehash_buckets = NULL;
1956 old_size = new_size;
1957 new_size = CFS_HASH_NBKT(hs);
1961 count += cfs_hash_rehash_bd(hs, &bd);
1962 if (count < CFS_HASH_LOOP_HOG ||
1963 cfs_hash_is_iterating(hs)) { /* need to finish ASAP */
1968 cfs_hash_unlock(hs, 1);
1970 cfs_hash_lock(hs, 1);
1973 hs->hs_rehash_count++;
1975 bkts = hs->hs_buckets;
1976 hs->hs_buckets = hs->hs_rehash_buckets;
1977 hs->hs_rehash_buckets = NULL;
1979 hs->hs_cur_bits = hs->hs_rehash_bits;
1981 hs->hs_rehash_bits = 0;
1982 if (rc == -ESRCH) /* never be scheduled again */
1983 cfs_wi_exit(cfs_sched_rehash, wi);
1984 bsize = cfs_hash_bkt_size(hs);
1985 cfs_hash_unlock(hs, 1);
1986 /* can't refer to @hs anymore because it could be destroyed */
1988 cfs_hash_buckets_free(bkts, bsize, new_size, old_size);
1990 CDEBUG(D_INFO, "early quit of rehashing: %d\n", rc);
1991 /* return 1 only if cfs_wi_exit is called */
1992 return rc == -ESRCH;
1996 * Rehash the object referenced by @hnode in the libcfs hash @hs. The
1997 * @old_key must be provided to locate the objects previous location
1998 * in the hash, and the @new_key will be used to reinsert the object.
1999 * Use this function instead of a cfs_hash_add() + cfs_hash_del()
2000 * combo when it is critical that there is no window in time where the
2001 * object is missing from the hash. When an object is being rehashed
2002 * the registered cfs_hash_get() and cfs_hash_put() functions will
2005 void cfs_hash_rehash_key(struct cfs_hash *hs, const void *old_key,
2006 void *new_key, struct hlist_node *hnode)
2008 struct cfs_hash_bd bds[3];
2009 struct cfs_hash_bd old_bds[2];
2010 struct cfs_hash_bd new_bd;
2012 LASSERT(!hlist_unhashed(hnode));
2014 cfs_hash_lock(hs, 0);
2016 cfs_hash_dual_bd_get(hs, old_key, old_bds);
2017 cfs_hash_bd_get(hs, new_key, &new_bd);
2019 bds[0] = old_bds[0];
2020 bds[1] = old_bds[1];
2023 /* NB: bds[0] and bds[1] are ordered already */
2024 cfs_hash_bd_order(&bds[1], &bds[2]);
2025 cfs_hash_bd_order(&bds[0], &bds[1]);
2027 cfs_hash_multi_bd_lock(hs, bds, 3, 1);
2028 if (likely(old_bds[1].bd_bucket == NULL)) {
2029 cfs_hash_bd_move_locked(hs, &old_bds[0], &new_bd, hnode);
2031 cfs_hash_dual_bd_finddel_locked(hs, old_bds, old_key, hnode);
2032 cfs_hash_bd_add_locked(hs, &new_bd, hnode);
2034 /* overwrite key inside locks, otherwise may screw up with
2035 * other operations, i.e: rehash */
2036 cfs_hash_keycpy(hs, hnode, new_key);
2038 cfs_hash_multi_bd_unlock(hs, bds, 3, 1);
2039 cfs_hash_unlock(hs, 0);
2041 EXPORT_SYMBOL(cfs_hash_rehash_key);
2043 void cfs_hash_debug_header(struct seq_file *m)
2045 seq_printf(m, "%-*s cur min max theta t-min t-max flags rehash count maxdep maxdepb distribution\n",
2046 CFS_HASH_BIGNAME_LEN, "name");
2048 EXPORT_SYMBOL(cfs_hash_debug_header);
2050 static struct cfs_hash_bucket **
2051 cfs_hash_full_bkts(struct cfs_hash *hs)
2053 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2054 if (hs->hs_rehash_buckets == NULL)
2055 return hs->hs_buckets;
2057 LASSERT(hs->hs_rehash_bits != 0);
2058 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2059 hs->hs_rehash_buckets : hs->hs_buckets;
2063 cfs_hash_full_nbkt(struct cfs_hash *hs)
2065 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2066 if (hs->hs_rehash_buckets == NULL)
2067 return CFS_HASH_NBKT(hs);
2069 LASSERT(hs->hs_rehash_bits != 0);
2070 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2071 CFS_HASH_RH_NBKT(hs) : CFS_HASH_NBKT(hs);
2074 void cfs_hash_debug_str(struct cfs_hash *hs, struct seq_file *m)
2076 int dist[8] = { 0, };
2083 cfs_hash_lock(hs, 0);
2084 theta = __cfs_hash_theta(hs);
2086 seq_printf(m, "%-*s %5d %5d %5d %d.%03d %d.%03d %d.%03d 0x%02x %6d ",
2087 CFS_HASH_BIGNAME_LEN, hs->hs_name,
2088 1 << hs->hs_cur_bits, 1 << hs->hs_min_bits,
2089 1 << hs->hs_max_bits,
2090 __cfs_hash_theta_int(theta), __cfs_hash_theta_frac(theta),
2091 __cfs_hash_theta_int(hs->hs_min_theta),
2092 __cfs_hash_theta_frac(hs->hs_min_theta),
2093 __cfs_hash_theta_int(hs->hs_max_theta),
2094 __cfs_hash_theta_frac(hs->hs_max_theta),
2095 hs->hs_flags, hs->hs_rehash_count);
2098 * The distribution is a summary of the chained hash depth in
2099 * each of the libcfs hash buckets. Each buckets hsb_count is
2100 * divided by the hash theta value and used to generate a
2101 * histogram of the hash distribution. A uniform hash will
2102 * result in all hash buckets being close to the average thus
2103 * only the first few entries in the histogram will be non-zero.
2104 * If you hash function results in a non-uniform hash the will
2105 * be observable by outlier bucks in the distribution histogram.
2107 * Uniform hash distribution: 128/128/0/0/0/0/0/0
2108 * Non-Uniform hash distribution: 128/125/0/0/0/0/2/1
2110 for (i = 0; i < cfs_hash_full_nbkt(hs); i++) {
2111 struct cfs_hash_bd bd;
2113 bd.bd_bucket = cfs_hash_full_bkts(hs)[i];
2114 cfs_hash_bd_lock(hs, &bd, 0);
2115 if (maxdep < bd.bd_bucket->hsb_depmax) {
2116 maxdep = bd.bd_bucket->hsb_depmax;
2117 maxdepb = ffz(~maxdep);
2119 total += bd.bd_bucket->hsb_count;
2120 dist[min(fls(bd.bd_bucket->hsb_count / max(theta, 1)), 7)]++;
2121 cfs_hash_bd_unlock(hs, &bd, 0);
2124 seq_printf(m, "%7d %7d %7d ", total, maxdep, maxdepb);
2125 for (i = 0; i < 8; i++)
2126 seq_printf(m, "%d%c", dist[i], (i == 7) ? '\n' : '/');
2128 cfs_hash_unlock(hs, 0);
2130 EXPORT_SYMBOL(cfs_hash_debug_str);