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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2012, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * libcfs/libcfs/hash.c
34 * Implement a hash class for hash process in lustre system.
36 * Author: YuZhangyong <yzy@clusterfs.com>
38 * 2008-08-15: Brian Behlendorf <behlendorf1@llnl.gov>
39 * - Simplified API and improved documentation
40 * - Added per-hash feature flags:
41 * * CFS_HASH_DEBUG additional validation
42 * * CFS_HASH_REHASH dynamic rehashing
43 * - Added per-hash statistics
44 * - General performance enhancements
46 * 2009-07-31: Liang Zhen <zhen.liang@sun.com>
47 * - move all stuff to libcfs
48 * - don't allow cur_bits != max_bits without setting of CFS_HASH_REHASH
49 * - ignore hs_rwlock if without CFS_HASH_REHASH setting
50 * - buckets are allocated one by one(instead of contiguous memory),
51 * to avoid unnecessary cacheline conflict
53 * 2010-03-01: Liang Zhen <zhen.liang@sun.com>
54 * - "bucket" is a group of hlist_head now, user can specify bucket size
55 * by bkt_bits of cfs_hash_create(), all hlist_heads in a bucket share
56 * one lock for reducing memory overhead.
58 * - support lockless hash, caller will take care of locks:
59 * avoid lock overhead for hash tables that are already protected
60 * by locking in the caller for another reason
62 * - support both spin_lock/rwlock for bucket:
63 * overhead of spinlock contention is lower than read/write
64 * contention of rwlock, so using spinlock to serialize operations on
65 * bucket is more reasonable for those frequently changed hash tables
67 * - support one-single lock mode:
68 * one lock to protect all hash operations to avoid overhead of
69 * multiple locks if hash table is always small
71 * - removed a lot of unnecessary addref & decref on hash element:
72 * addref & decref are atomic operations in many use-cases which
75 * - support non-blocking cfs_hash_add() and cfs_hash_findadd():
76 * some lustre use-cases require these functions to be strictly
77 * non-blocking, we need to schedule required rehash on a different
78 * thread on those cases.
80 * - safer rehash on large hash table
81 * In old implementation, rehash function will exclusively lock the
82 * hash table and finish rehash in one batch, it's dangerous on SMP
83 * system because rehash millions of elements could take long time.
84 * New implemented rehash can release lock and relax CPU in middle
85 * of rehash, it's safe for another thread to search/change on the
86 * hash table even it's in rehasing.
88 * - support two different refcount modes
89 * . hash table has refcount on element
90 * . hash table doesn't change refcount on adding/removing element
92 * - support long name hash table (for param-tree)
94 * - fix a bug for cfs_hash_rehash_key:
95 * in old implementation, cfs_hash_rehash_key could screw up the
96 * hash-table because @key is overwritten without any protection.
97 * Now we need user to define hs_keycpy for those rehash enabled
98 * hash tables, cfs_hash_rehash_key will overwrite hash-key
99 * inside lock by calling hs_keycpy.
101 * - better hash iteration:
102 * Now we support both locked iteration & lockless iteration of hash
103 * table. Also, user can break the iteration by return 1 in callback.
105 #include <linux/seq_file.h>
106 #include <linux/log2.h>
108 #include "../../include/linux/libcfs/libcfs.h"
110 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
111 static unsigned int warn_on_depth = 8;
112 module_param(warn_on_depth, uint, 0644);
113 MODULE_PARM_DESC(warn_on_depth, "warning when hash depth is high.");
116 struct cfs_wi_sched *cfs_sched_rehash;
119 cfs_hash_nl_lock(union cfs_hash_lock *lock, int exclusive) {}
122 cfs_hash_nl_unlock(union cfs_hash_lock *lock, int exclusive) {}
125 cfs_hash_spin_lock(union cfs_hash_lock *lock, int exclusive)
126 __acquires(&lock->spin)
128 spin_lock(&lock->spin);
132 cfs_hash_spin_unlock(union cfs_hash_lock *lock, int exclusive)
133 __releases(&lock->spin)
135 spin_unlock(&lock->spin);
139 cfs_hash_rw_lock(union cfs_hash_lock *lock, int exclusive)
140 __acquires(&lock->rw)
143 read_lock(&lock->rw);
145 write_lock(&lock->rw);
149 cfs_hash_rw_unlock(union cfs_hash_lock *lock, int exclusive)
150 __releases(&lock->rw)
153 read_unlock(&lock->rw);
155 write_unlock(&lock->rw);
159 static struct cfs_hash_lock_ops cfs_hash_nl_lops = {
160 .hs_lock = cfs_hash_nl_lock,
161 .hs_unlock = cfs_hash_nl_unlock,
162 .hs_bkt_lock = cfs_hash_nl_lock,
163 .hs_bkt_unlock = cfs_hash_nl_unlock,
166 /** no bucket lock, one spinlock to protect everything */
167 static struct cfs_hash_lock_ops cfs_hash_nbl_lops = {
168 .hs_lock = cfs_hash_spin_lock,
169 .hs_unlock = cfs_hash_spin_unlock,
170 .hs_bkt_lock = cfs_hash_nl_lock,
171 .hs_bkt_unlock = cfs_hash_nl_unlock,
174 /** spin bucket lock, rehash is enabled */
175 static struct cfs_hash_lock_ops cfs_hash_bkt_spin_lops = {
176 .hs_lock = cfs_hash_rw_lock,
177 .hs_unlock = cfs_hash_rw_unlock,
178 .hs_bkt_lock = cfs_hash_spin_lock,
179 .hs_bkt_unlock = cfs_hash_spin_unlock,
182 /** rw bucket lock, rehash is enabled */
183 static struct cfs_hash_lock_ops cfs_hash_bkt_rw_lops = {
184 .hs_lock = cfs_hash_rw_lock,
185 .hs_unlock = cfs_hash_rw_unlock,
186 .hs_bkt_lock = cfs_hash_rw_lock,
187 .hs_bkt_unlock = cfs_hash_rw_unlock,
190 /** spin bucket lock, rehash is disabled */
191 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_spin_lops = {
192 .hs_lock = cfs_hash_nl_lock,
193 .hs_unlock = cfs_hash_nl_unlock,
194 .hs_bkt_lock = cfs_hash_spin_lock,
195 .hs_bkt_unlock = cfs_hash_spin_unlock,
198 /** rw bucket lock, rehash is disabled */
199 static struct cfs_hash_lock_ops cfs_hash_nr_bkt_rw_lops = {
200 .hs_lock = cfs_hash_nl_lock,
201 .hs_unlock = cfs_hash_nl_unlock,
202 .hs_bkt_lock = cfs_hash_rw_lock,
203 .hs_bkt_unlock = cfs_hash_rw_unlock,
207 cfs_hash_lock_setup(struct cfs_hash *hs)
209 if (cfs_hash_with_no_lock(hs)) {
210 hs->hs_lops = &cfs_hash_nl_lops;
212 } else if (cfs_hash_with_no_bktlock(hs)) {
213 hs->hs_lops = &cfs_hash_nbl_lops;
214 spin_lock_init(&hs->hs_lock.spin);
216 } else if (cfs_hash_with_rehash(hs)) {
217 rwlock_init(&hs->hs_lock.rw);
219 if (cfs_hash_with_rw_bktlock(hs))
220 hs->hs_lops = &cfs_hash_bkt_rw_lops;
221 else if (cfs_hash_with_spin_bktlock(hs))
222 hs->hs_lops = &cfs_hash_bkt_spin_lops;
226 if (cfs_hash_with_rw_bktlock(hs))
227 hs->hs_lops = &cfs_hash_nr_bkt_rw_lops;
228 else if (cfs_hash_with_spin_bktlock(hs))
229 hs->hs_lops = &cfs_hash_nr_bkt_spin_lops;
236 * Simple hash head without depth tracking
237 * new element is always added to head of hlist
239 struct cfs_hash_head {
240 struct hlist_head hh_head; /**< entries list */
244 cfs_hash_hh_hhead_size(struct cfs_hash *hs)
246 return sizeof(struct cfs_hash_head);
249 static struct hlist_head *
250 cfs_hash_hh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
252 struct cfs_hash_head *head;
254 head = (struct cfs_hash_head *)&bd->bd_bucket->hsb_head[0];
255 return &head[bd->bd_offset].hh_head;
259 cfs_hash_hh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
260 struct hlist_node *hnode)
262 hlist_add_head(hnode, cfs_hash_hh_hhead(hs, bd));
263 return -1; /* unknown depth */
267 cfs_hash_hh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
268 struct hlist_node *hnode)
270 hlist_del_init(hnode);
271 return -1; /* unknown depth */
275 * Simple hash head with depth tracking
276 * new element is always added to head of hlist
278 struct cfs_hash_head_dep {
279 struct hlist_head hd_head; /**< entries list */
280 unsigned int hd_depth; /**< list length */
284 cfs_hash_hd_hhead_size(struct cfs_hash *hs)
286 return sizeof(struct cfs_hash_head_dep);
289 static struct hlist_head *
290 cfs_hash_hd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
292 struct cfs_hash_head_dep *head;
294 head = (struct cfs_hash_head_dep *)&bd->bd_bucket->hsb_head[0];
295 return &head[bd->bd_offset].hd_head;
299 cfs_hash_hd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
300 struct hlist_node *hnode)
302 struct cfs_hash_head_dep *hh;
304 hh = container_of(cfs_hash_hd_hhead(hs, bd),
305 struct cfs_hash_head_dep, hd_head);
306 hlist_add_head(hnode, &hh->hd_head);
307 return ++hh->hd_depth;
311 cfs_hash_hd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
312 struct hlist_node *hnode)
314 struct cfs_hash_head_dep *hh;
316 hh = container_of(cfs_hash_hd_hhead(hs, bd),
317 struct cfs_hash_head_dep, hd_head);
318 hlist_del_init(hnode);
319 return --hh->hd_depth;
323 * double links hash head without depth tracking
324 * new element is always added to tail of hlist
326 struct cfs_hash_dhead {
327 struct hlist_head dh_head; /**< entries list */
328 struct hlist_node *dh_tail; /**< the last entry */
332 cfs_hash_dh_hhead_size(struct cfs_hash *hs)
334 return sizeof(struct cfs_hash_dhead);
337 static struct hlist_head *
338 cfs_hash_dh_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
340 struct cfs_hash_dhead *head;
342 head = (struct cfs_hash_dhead *)&bd->bd_bucket->hsb_head[0];
343 return &head[bd->bd_offset].dh_head;
347 cfs_hash_dh_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
348 struct hlist_node *hnode)
350 struct cfs_hash_dhead *dh;
352 dh = container_of(cfs_hash_dh_hhead(hs, bd),
353 struct cfs_hash_dhead, dh_head);
354 if (dh->dh_tail) /* not empty */
355 hlist_add_behind(hnode, dh->dh_tail);
356 else /* empty list */
357 hlist_add_head(hnode, &dh->dh_head);
359 return -1; /* unknown depth */
363 cfs_hash_dh_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
364 struct hlist_node *hnd)
366 struct cfs_hash_dhead *dh;
368 dh = container_of(cfs_hash_dh_hhead(hs, bd),
369 struct cfs_hash_dhead, dh_head);
370 if (!hnd->next) { /* it's the tail */
371 dh->dh_tail = (hnd->pprev == &dh->dh_head.first) ? NULL :
372 container_of(hnd->pprev, struct hlist_node, next);
375 return -1; /* unknown depth */
379 * double links hash head with depth tracking
380 * new element is always added to tail of hlist
382 struct cfs_hash_dhead_dep {
383 struct hlist_head dd_head; /**< entries list */
384 struct hlist_node *dd_tail; /**< the last entry */
385 unsigned int dd_depth; /**< list length */
389 cfs_hash_dd_hhead_size(struct cfs_hash *hs)
391 return sizeof(struct cfs_hash_dhead_dep);
394 static struct hlist_head *
395 cfs_hash_dd_hhead(struct cfs_hash *hs, struct cfs_hash_bd *bd)
397 struct cfs_hash_dhead_dep *head;
399 head = (struct cfs_hash_dhead_dep *)&bd->bd_bucket->hsb_head[0];
400 return &head[bd->bd_offset].dd_head;
404 cfs_hash_dd_hnode_add(struct cfs_hash *hs, struct cfs_hash_bd *bd,
405 struct hlist_node *hnode)
407 struct cfs_hash_dhead_dep *dh;
409 dh = container_of(cfs_hash_dd_hhead(hs, bd),
410 struct cfs_hash_dhead_dep, dd_head);
411 if (dh->dd_tail) /* not empty */
412 hlist_add_behind(hnode, dh->dd_tail);
413 else /* empty list */
414 hlist_add_head(hnode, &dh->dd_head);
416 return ++dh->dd_depth;
420 cfs_hash_dd_hnode_del(struct cfs_hash *hs, struct cfs_hash_bd *bd,
421 struct hlist_node *hnd)
423 struct cfs_hash_dhead_dep *dh;
425 dh = container_of(cfs_hash_dd_hhead(hs, bd),
426 struct cfs_hash_dhead_dep, dd_head);
427 if (!hnd->next) { /* 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 struct cfs_hash_hlist_ops cfs_hash_hh_hops = {
436 .hop_hhead = cfs_hash_hh_hhead,
437 .hop_hhead_size = cfs_hash_hh_hhead_size,
438 .hop_hnode_add = cfs_hash_hh_hnode_add,
439 .hop_hnode_del = cfs_hash_hh_hnode_del,
442 static struct cfs_hash_hlist_ops cfs_hash_hd_hops = {
443 .hop_hhead = cfs_hash_hd_hhead,
444 .hop_hhead_size = cfs_hash_hd_hhead_size,
445 .hop_hnode_add = cfs_hash_hd_hnode_add,
446 .hop_hnode_del = cfs_hash_hd_hnode_del,
449 static struct cfs_hash_hlist_ops cfs_hash_dh_hops = {
450 .hop_hhead = cfs_hash_dh_hhead,
451 .hop_hhead_size = cfs_hash_dh_hhead_size,
452 .hop_hnode_add = cfs_hash_dh_hnode_add,
453 .hop_hnode_del = cfs_hash_dh_hnode_del,
456 static struct cfs_hash_hlist_ops cfs_hash_dd_hops = {
457 .hop_hhead = cfs_hash_dd_hhead,
458 .hop_hhead_size = cfs_hash_dd_hhead_size,
459 .hop_hnode_add = cfs_hash_dd_hnode_add,
460 .hop_hnode_del = cfs_hash_dd_hnode_del,
464 cfs_hash_hlist_setup(struct cfs_hash *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(struct cfs_hash *hs, struct cfs_hash_bucket **bkts,
477 unsigned int bits, const void *key, struct cfs_hash_bd *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(struct cfs_hash *hs, const void *key, struct cfs_hash_bd *bd)
490 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
491 if (likely(!hs->hs_rehash_buckets)) {
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(struct cfs_hash *hs, struct cfs_hash_bd *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(struct cfs_hash *hs, struct cfs_hash_bd *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(struct cfs_hash *hs, struct cfs_hash_bd *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(struct cfs_hash *hs, struct cfs_hash_bd *bd_old,
568 struct cfs_hash_bd *bd_new, struct hlist_node *hnode)
570 struct cfs_hash_bucket *obkt = bd_old->bd_bucket;
571 struct cfs_hash_bucket *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
580 hs->hs_hops->hop_hnode_del(hs, bd_old, hnode);
581 rc = hs->hs_hops->hop_hnode_add(hs, bd_new, hnode);
582 cfs_hash_bd_dep_record(hs, bd_new, rc);
584 LASSERT(obkt->hsb_count > 0);
587 if (unlikely(obkt->hsb_version == 0))
591 if (unlikely(nbkt->hsb_version == 0))
596 /** always set, for sanity (avoid ZERO intent) */
597 CFS_HS_LOOKUP_MASK_FIND = BIT(0),
598 /** return entry with a ref */
599 CFS_HS_LOOKUP_MASK_REF = BIT(1),
600 /** add entry if not existing */
601 CFS_HS_LOOKUP_MASK_ADD = BIT(2),
602 /** delete entry, ignore other masks */
603 CFS_HS_LOOKUP_MASK_DEL = BIT(3),
606 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)
623 static struct hlist_node *
624 cfs_hash_bd_lookup_intent(struct cfs_hash *hs, struct cfs_hash_bd *bd,
625 const void *key, struct hlist_node *hnode,
626 enum cfs_hash_lookup_intent 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.
637 match = intent_add ? NULL : hnode;
638 hlist_for_each(ehnode, hhead) {
639 if (!cfs_hash_keycmp(hs, key, ehnode))
642 if (match && match != ehnode) /* can't match */
646 if ((intent & CFS_HS_LOOKUP_MASK_DEL) != 0) {
647 cfs_hash_bd_del_locked(hs, bd, ehnode);
651 /* caller wants refcount? */
652 if ((intent & CFS_HS_LOOKUP_MASK_REF) != 0)
653 cfs_hash_get(hs, ehnode);
661 cfs_hash_bd_add_locked(hs, bd, hnode);
666 cfs_hash_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
669 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
670 CFS_HS_LOOKUP_IT_FIND);
672 EXPORT_SYMBOL(cfs_hash_bd_lookup_locked);
675 cfs_hash_bd_peek_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
678 return cfs_hash_bd_lookup_intent(hs, bd, key, NULL,
679 CFS_HS_LOOKUP_IT_PEEK);
681 EXPORT_SYMBOL(cfs_hash_bd_peek_locked);
684 cfs_hash_multi_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
685 unsigned n, int excl)
687 struct cfs_hash_bucket *prev = NULL;
691 * bds must be ascendantly ordered by bd->bd_bucket->hsb_index.
692 * NB: it's possible that several bds point to the same bucket but
693 * have different bd::bd_offset, so need take care of deadlock.
695 cfs_hash_for_each_bd(bds, n, i) {
696 if (prev == bds[i].bd_bucket)
699 LASSERT(!prev || prev->hsb_index < bds[i].bd_bucket->hsb_index);
700 cfs_hash_bd_lock(hs, &bds[i], excl);
701 prev = bds[i].bd_bucket;
706 cfs_hash_multi_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds,
707 unsigned n, int excl)
709 struct cfs_hash_bucket *prev = NULL;
712 cfs_hash_for_each_bd(bds, n, i) {
713 if (prev != bds[i].bd_bucket) {
714 cfs_hash_bd_unlock(hs, &bds[i], excl);
715 prev = bds[i].bd_bucket;
720 static struct hlist_node *
721 cfs_hash_multi_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
722 unsigned n, const void *key)
724 struct hlist_node *ehnode;
727 cfs_hash_for_each_bd(bds, n, i) {
728 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, NULL,
729 CFS_HS_LOOKUP_IT_FIND);
736 static struct hlist_node *
737 cfs_hash_multi_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
738 unsigned n, const void *key,
739 struct hlist_node *hnode, int noref)
741 struct hlist_node *ehnode;
746 intent = (!noref * CFS_HS_LOOKUP_MASK_REF) | CFS_HS_LOOKUP_IT_PEEK;
748 cfs_hash_for_each_bd(bds, n, i) {
749 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key,
755 if (i == 1) { /* only one bucket */
756 cfs_hash_bd_add_locked(hs, &bds[0], hnode);
758 struct cfs_hash_bd mybd;
760 cfs_hash_bd_get(hs, key, &mybd);
761 cfs_hash_bd_add_locked(hs, &mybd, hnode);
767 static struct hlist_node *
768 cfs_hash_multi_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
769 unsigned n, const void *key,
770 struct hlist_node *hnode)
772 struct hlist_node *ehnode;
775 cfs_hash_for_each_bd(bds, n, i) {
776 ehnode = cfs_hash_bd_lookup_intent(hs, &bds[i], key, hnode,
777 CFS_HS_LOOKUP_IT_FINDDEL);
785 cfs_hash_bd_order(struct cfs_hash_bd *bd1, struct cfs_hash_bd *bd2)
792 if (!bd1->bd_bucket) {
794 bd2->bd_bucket = NULL;
798 rc = cfs_hash_bd_compare(bd1, bd2);
800 bd2->bd_bucket = NULL;
802 swap(*bd1, *bd2); /* swap bd1 and bd2 */
806 cfs_hash_dual_bd_get(struct cfs_hash *hs, const void *key,
807 struct cfs_hash_bd *bds)
809 /* NB: caller should hold hs_lock.rw if REHASH is set */
810 cfs_hash_bd_from_key(hs, hs->hs_buckets,
811 hs->hs_cur_bits, key, &bds[0]);
812 if (likely(!hs->hs_rehash_buckets)) {
813 /* no rehash or not rehashing */
814 bds[1].bd_bucket = NULL;
818 LASSERT(hs->hs_rehash_bits != 0);
819 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
820 hs->hs_rehash_bits, key, &bds[1]);
822 cfs_hash_bd_order(&bds[0], &bds[1]);
826 cfs_hash_dual_bd_lock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
828 cfs_hash_multi_bd_lock(hs, bds, 2, excl);
832 cfs_hash_dual_bd_unlock(struct cfs_hash *hs, struct cfs_hash_bd *bds, int excl)
834 cfs_hash_multi_bd_unlock(hs, bds, 2, excl);
838 cfs_hash_dual_bd_lookup_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
841 return cfs_hash_multi_bd_lookup_locked(hs, bds, 2, key);
845 cfs_hash_dual_bd_findadd_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
846 const void *key, struct hlist_node *hnode,
849 return cfs_hash_multi_bd_findadd_locked(hs, bds, 2, key,
854 cfs_hash_dual_bd_finddel_locked(struct cfs_hash *hs, struct cfs_hash_bd *bds,
855 const void *key, struct hlist_node *hnode)
857 return cfs_hash_multi_bd_finddel_locked(hs, bds, 2, key, hnode);
861 cfs_hash_buckets_free(struct cfs_hash_bucket **buckets,
862 int bkt_size, int prev_size, int size)
866 for (i = prev_size; i < size; i++) {
868 LIBCFS_FREE(buckets[i], bkt_size);
871 LIBCFS_FREE(buckets, sizeof(buckets[0]) * size);
875 * Create or grow bucket memory. Return old_buckets if no allocation was
876 * needed, the newly allocated buckets if allocation was needed and
877 * successful, and NULL on error.
879 static struct cfs_hash_bucket **
880 cfs_hash_buckets_realloc(struct cfs_hash *hs, struct cfs_hash_bucket **old_bkts,
881 unsigned int old_size, unsigned int new_size)
883 struct cfs_hash_bucket **new_bkts;
886 LASSERT(old_size == 0 || old_bkts);
888 if (old_bkts && old_size == new_size)
891 LIBCFS_ALLOC(new_bkts, sizeof(new_bkts[0]) * new_size);
896 memcpy(new_bkts, old_bkts,
897 min(old_size, new_size) * sizeof(*old_bkts));
900 for (i = old_size; i < new_size; i++) {
901 struct hlist_head *hhead;
902 struct cfs_hash_bd bd;
904 LIBCFS_ALLOC(new_bkts[i], cfs_hash_bkt_size(hs));
906 cfs_hash_buckets_free(new_bkts, cfs_hash_bkt_size(hs),
911 new_bkts[i]->hsb_index = i;
912 new_bkts[i]->hsb_version = 1; /* shouldn't be zero */
913 new_bkts[i]->hsb_depmax = -1; /* unknown */
914 bd.bd_bucket = new_bkts[i];
915 cfs_hash_bd_for_each_hlist(hs, &bd, hhead)
916 INIT_HLIST_HEAD(hhead);
918 if (cfs_hash_with_no_lock(hs) ||
919 cfs_hash_with_no_bktlock(hs))
922 if (cfs_hash_with_rw_bktlock(hs))
923 rwlock_init(&new_bkts[i]->hsb_lock.rw);
924 else if (cfs_hash_with_spin_bktlock(hs))
925 spin_lock_init(&new_bkts[i]->hsb_lock.spin);
927 LBUG(); /* invalid use-case */
933 * Initialize new libcfs hash, where:
934 * @name - Descriptive hash name
935 * @cur_bits - Initial hash table size, in bits
936 * @max_bits - Maximum allowed hash table resize, in bits
937 * @ops - Registered hash table operations
938 * @flags - CFS_HASH_REHASH enable synamic hash resizing
939 * - CFS_HASH_SORT enable chained hash sort
941 static int cfs_hash_rehash_worker(struct cfs_workitem *wi);
943 #if CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1
944 static int cfs_hash_dep_print(struct cfs_workitem *wi)
946 struct cfs_hash *hs = container_of(wi, struct cfs_hash, hs_dep_wi);
952 spin_lock(&hs->hs_dep_lock);
953 dep = hs->hs_dep_max;
954 bkt = hs->hs_dep_bkt;
955 off = hs->hs_dep_off;
956 bits = hs->hs_dep_bits;
957 spin_unlock(&hs->hs_dep_lock);
959 LCONSOLE_WARN("#### HASH %s (bits: %d): max depth %d at bucket %d/%d\n",
960 hs->hs_name, bits, dep, bkt, off);
961 spin_lock(&hs->hs_dep_lock);
962 hs->hs_dep_bits = 0; /* mark as workitem done */
963 spin_unlock(&hs->hs_dep_lock);
967 static void cfs_hash_depth_wi_init(struct cfs_hash *hs)
969 spin_lock_init(&hs->hs_dep_lock);
970 cfs_wi_init(&hs->hs_dep_wi, hs, cfs_hash_dep_print);
973 static void cfs_hash_depth_wi_cancel(struct cfs_hash *hs)
975 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_dep_wi))
978 spin_lock(&hs->hs_dep_lock);
979 while (hs->hs_dep_bits != 0) {
980 spin_unlock(&hs->hs_dep_lock);
982 spin_lock(&hs->hs_dep_lock);
984 spin_unlock(&hs->hs_dep_lock);
987 #else /* CFS_HASH_DEBUG_LEVEL < CFS_HASH_DEBUG_1 */
989 static inline void cfs_hash_depth_wi_init(struct cfs_hash *hs) {}
990 static inline void cfs_hash_depth_wi_cancel(struct cfs_hash *hs) {}
992 #endif /* CFS_HASH_DEBUG_LEVEL >= CFS_HASH_DEBUG_1 */
995 cfs_hash_create(char *name, unsigned cur_bits, unsigned max_bits,
996 unsigned bkt_bits, unsigned extra_bytes,
997 unsigned min_theta, unsigned max_theta,
998 struct cfs_hash_ops *ops, unsigned flags)
1000 struct cfs_hash *hs;
1003 CLASSERT(CFS_HASH_THETA_BITS < 15);
1006 LASSERT(ops->hs_key);
1007 LASSERT(ops->hs_hash);
1008 LASSERT(ops->hs_object);
1009 LASSERT(ops->hs_keycmp);
1010 LASSERT(ops->hs_get);
1011 LASSERT(ops->hs_put_locked);
1013 if ((flags & CFS_HASH_REHASH) != 0)
1014 flags |= CFS_HASH_COUNTER; /* must have counter */
1016 LASSERT(cur_bits > 0);
1017 LASSERT(cur_bits >= bkt_bits);
1018 LASSERT(max_bits >= cur_bits && max_bits < 31);
1019 LASSERT(ergo((flags & CFS_HASH_REHASH) == 0, cur_bits == max_bits));
1020 LASSERT(ergo((flags & CFS_HASH_REHASH) != 0,
1021 (flags & CFS_HASH_NO_LOCK) == 0));
1022 LASSERT(ergo((flags & CFS_HASH_REHASH_KEY) != 0, ops->hs_keycpy));
1024 len = (flags & CFS_HASH_BIGNAME) == 0 ?
1025 CFS_HASH_NAME_LEN : CFS_HASH_BIGNAME_LEN;
1026 LIBCFS_ALLOC(hs, offsetof(struct cfs_hash, hs_name[len]));
1030 strlcpy(hs->hs_name, name, len);
1031 hs->hs_flags = flags;
1033 atomic_set(&hs->hs_refcount, 1);
1034 atomic_set(&hs->hs_count, 0);
1036 cfs_hash_lock_setup(hs);
1037 cfs_hash_hlist_setup(hs);
1039 hs->hs_cur_bits = (__u8)cur_bits;
1040 hs->hs_min_bits = (__u8)cur_bits;
1041 hs->hs_max_bits = (__u8)max_bits;
1042 hs->hs_bkt_bits = (__u8)bkt_bits;
1045 hs->hs_extra_bytes = extra_bytes;
1046 hs->hs_rehash_bits = 0;
1047 cfs_wi_init(&hs->hs_rehash_wi, hs, cfs_hash_rehash_worker);
1048 cfs_hash_depth_wi_init(hs);
1050 if (cfs_hash_with_rehash(hs))
1051 __cfs_hash_set_theta(hs, min_theta, max_theta);
1053 hs->hs_buckets = cfs_hash_buckets_realloc(hs, NULL, 0,
1058 LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[len]));
1061 EXPORT_SYMBOL(cfs_hash_create);
1064 * Cleanup libcfs hash @hs.
1067 cfs_hash_destroy(struct cfs_hash *hs)
1069 struct hlist_node *hnode;
1070 struct hlist_node *pos;
1071 struct cfs_hash_bd bd;
1075 LASSERT(!cfs_hash_is_exiting(hs) &&
1076 !cfs_hash_is_iterating(hs));
1079 * prohibit further rehashes, don't need any lock because
1080 * I'm the only (last) one can change it.
1083 if (cfs_hash_with_rehash(hs))
1084 cfs_hash_rehash_cancel(hs);
1086 cfs_hash_depth_wi_cancel(hs);
1087 /* rehash should be done/canceled */
1088 LASSERT(hs->hs_buckets && !hs->hs_rehash_buckets);
1090 cfs_hash_for_each_bucket(hs, &bd, i) {
1091 struct hlist_head *hhead;
1093 LASSERT(bd.bd_bucket);
1094 /* no need to take this lock, just for consistent code */
1095 cfs_hash_bd_lock(hs, &bd, 1);
1097 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1098 hlist_for_each_safe(hnode, pos, hhead) {
1099 LASSERTF(!cfs_hash_with_assert_empty(hs),
1100 "hash %s bucket %u(%u) is not empty: %u items left\n",
1101 hs->hs_name, bd.bd_bucket->hsb_index,
1102 bd.bd_offset, bd.bd_bucket->hsb_count);
1103 /* can't assert key valicate, because we
1104 * can interrupt rehash
1106 cfs_hash_bd_del_locked(hs, &bd, hnode);
1107 cfs_hash_exit(hs, hnode);
1110 LASSERT(bd.bd_bucket->hsb_count == 0);
1111 cfs_hash_bd_unlock(hs, &bd, 1);
1115 LASSERT(atomic_read(&hs->hs_count) == 0);
1117 cfs_hash_buckets_free(hs->hs_buckets, cfs_hash_bkt_size(hs),
1118 0, CFS_HASH_NBKT(hs));
1119 i = cfs_hash_with_bigname(hs) ?
1120 CFS_HASH_BIGNAME_LEN : CFS_HASH_NAME_LEN;
1121 LIBCFS_FREE(hs, offsetof(struct cfs_hash, hs_name[i]));
1124 struct cfs_hash *cfs_hash_getref(struct cfs_hash *hs)
1126 if (atomic_inc_not_zero(&hs->hs_refcount))
1130 EXPORT_SYMBOL(cfs_hash_getref);
1132 void cfs_hash_putref(struct cfs_hash *hs)
1134 if (atomic_dec_and_test(&hs->hs_refcount))
1135 cfs_hash_destroy(hs);
1137 EXPORT_SYMBOL(cfs_hash_putref);
1140 cfs_hash_rehash_bits(struct cfs_hash *hs)
1142 if (cfs_hash_with_no_lock(hs) ||
1143 !cfs_hash_with_rehash(hs))
1146 if (unlikely(cfs_hash_is_exiting(hs)))
1149 if (unlikely(cfs_hash_is_rehashing(hs)))
1152 if (unlikely(cfs_hash_is_iterating(hs)))
1155 /* XXX: need to handle case with max_theta != 2.0
1156 * and the case with min_theta != 0.5
1158 if ((hs->hs_cur_bits < hs->hs_max_bits) &&
1159 (__cfs_hash_theta(hs) > hs->hs_max_theta))
1160 return hs->hs_cur_bits + 1;
1162 if (!cfs_hash_with_shrink(hs))
1165 if ((hs->hs_cur_bits > hs->hs_min_bits) &&
1166 (__cfs_hash_theta(hs) < hs->hs_min_theta))
1167 return hs->hs_cur_bits - 1;
1173 * don't allow inline rehash if:
1174 * - user wants non-blocking change (add/del) on hash table
1175 * - too many elements
1178 cfs_hash_rehash_inline(struct cfs_hash *hs)
1180 return !cfs_hash_with_nblk_change(hs) &&
1181 atomic_read(&hs->hs_count) < CFS_HASH_LOOP_HOG;
1185 * Add item @hnode to libcfs hash @hs using @key. The registered
1186 * ops->hs_get function will be called when the item is added.
1189 cfs_hash_add(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1191 struct cfs_hash_bd bd;
1194 LASSERT(hlist_unhashed(hnode));
1196 cfs_hash_lock(hs, 0);
1197 cfs_hash_bd_get_and_lock(hs, key, &bd, 1);
1199 cfs_hash_key_validate(hs, key, hnode);
1200 cfs_hash_bd_add_locked(hs, &bd, hnode);
1202 cfs_hash_bd_unlock(hs, &bd, 1);
1204 bits = cfs_hash_rehash_bits(hs);
1205 cfs_hash_unlock(hs, 0);
1207 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1209 EXPORT_SYMBOL(cfs_hash_add);
1211 static struct hlist_node *
1212 cfs_hash_find_or_add(struct cfs_hash *hs, const void *key,
1213 struct hlist_node *hnode, int noref)
1215 struct hlist_node *ehnode;
1216 struct cfs_hash_bd bds[2];
1219 LASSERT(hlist_unhashed(hnode));
1221 cfs_hash_lock(hs, 0);
1222 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1224 cfs_hash_key_validate(hs, key, hnode);
1225 ehnode = cfs_hash_dual_bd_findadd_locked(hs, bds, key,
1227 cfs_hash_dual_bd_unlock(hs, bds, 1);
1229 if (ehnode == hnode) /* new item added */
1230 bits = cfs_hash_rehash_bits(hs);
1231 cfs_hash_unlock(hs, 0);
1233 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1239 * Add item @hnode to libcfs hash @hs using @key. The registered
1240 * ops->hs_get function will be called if the item was added.
1241 * Returns 0 on success or -EALREADY on key collisions.
1244 cfs_hash_add_unique(struct cfs_hash *hs, const void *key,
1245 struct hlist_node *hnode)
1247 return cfs_hash_find_or_add(hs, key, hnode, 1) != hnode ?
1250 EXPORT_SYMBOL(cfs_hash_add_unique);
1253 * Add item @hnode to libcfs hash @hs using @key. If this @key
1254 * already exists in the hash then ops->hs_get will be called on the
1255 * conflicting entry and that entry will be returned to the caller.
1256 * Otherwise ops->hs_get is called on the item which was added.
1259 cfs_hash_findadd_unique(struct cfs_hash *hs, const void *key,
1260 struct hlist_node *hnode)
1262 hnode = cfs_hash_find_or_add(hs, key, hnode, 0);
1264 return cfs_hash_object(hs, hnode);
1266 EXPORT_SYMBOL(cfs_hash_findadd_unique);
1269 * Delete item @hnode from the libcfs hash @hs using @key. The @key
1270 * is required to ensure the correct hash bucket is locked since there
1271 * is no direct linkage from the item to the bucket. The object
1272 * removed from the hash will be returned and obs->hs_put is called
1273 * on the removed object.
1276 cfs_hash_del(struct cfs_hash *hs, const void *key, struct hlist_node *hnode)
1280 struct cfs_hash_bd bds[2];
1282 cfs_hash_lock(hs, 0);
1283 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 1);
1285 /* NB: do nothing if @hnode is not in hash table */
1286 if (!hnode || !hlist_unhashed(hnode)) {
1287 if (!bds[1].bd_bucket && hnode) {
1288 cfs_hash_bd_del_locked(hs, &bds[0], hnode);
1290 hnode = cfs_hash_dual_bd_finddel_locked(hs, bds,
1296 obj = cfs_hash_object(hs, hnode);
1297 bits = cfs_hash_rehash_bits(hs);
1300 cfs_hash_dual_bd_unlock(hs, bds, 1);
1301 cfs_hash_unlock(hs, 0);
1303 cfs_hash_rehash(hs, cfs_hash_rehash_inline(hs));
1307 EXPORT_SYMBOL(cfs_hash_del);
1310 * Delete item given @key in libcfs hash @hs. The first @key found in
1311 * the hash will be removed, if the key exists multiple times in the hash
1312 * @hs this function must be called once per key. The removed object
1313 * will be returned and ops->hs_put is called on the removed object.
1316 cfs_hash_del_key(struct cfs_hash *hs, const void *key)
1318 return cfs_hash_del(hs, key, NULL);
1320 EXPORT_SYMBOL(cfs_hash_del_key);
1323 * Lookup an item using @key in the libcfs hash @hs and return it.
1324 * If the @key is found in the hash hs->hs_get() is called and the
1325 * matching objects is returned. It is the callers responsibility
1326 * to call the counterpart ops->hs_put using the cfs_hash_put() macro
1327 * when when finished with the object. If the @key was not found
1328 * in the hash @hs NULL is returned.
1331 cfs_hash_lookup(struct cfs_hash *hs, const void *key)
1334 struct hlist_node *hnode;
1335 struct cfs_hash_bd bds[2];
1337 cfs_hash_lock(hs, 0);
1338 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1340 hnode = cfs_hash_dual_bd_lookup_locked(hs, bds, key);
1342 obj = cfs_hash_object(hs, hnode);
1344 cfs_hash_dual_bd_unlock(hs, bds, 0);
1345 cfs_hash_unlock(hs, 0);
1349 EXPORT_SYMBOL(cfs_hash_lookup);
1352 cfs_hash_for_each_enter(struct cfs_hash *hs)
1354 LASSERT(!cfs_hash_is_exiting(hs));
1356 if (!cfs_hash_with_rehash(hs))
1359 * NB: it's race on cfs_has_t::hs_iterating, but doesn't matter
1360 * because it's just an unreliable signal to rehash-thread,
1361 * rehash-thread will try to finish rehash ASAP when seeing this.
1363 hs->hs_iterating = 1;
1365 cfs_hash_lock(hs, 1);
1368 /* NB: iteration is mostly called by service thread,
1369 * we tend to cancel pending rehash-request, instead of
1370 * blocking service thread, we will relaunch rehash request
1373 if (cfs_hash_is_rehashing(hs))
1374 cfs_hash_rehash_cancel_locked(hs);
1375 cfs_hash_unlock(hs, 1);
1379 cfs_hash_for_each_exit(struct cfs_hash *hs)
1384 if (!cfs_hash_with_rehash(hs))
1386 cfs_hash_lock(hs, 1);
1387 remained = --hs->hs_iterators;
1388 bits = cfs_hash_rehash_bits(hs);
1389 cfs_hash_unlock(hs, 1);
1390 /* NB: it's race on cfs_has_t::hs_iterating, see above */
1392 hs->hs_iterating = 0;
1394 cfs_hash_rehash(hs, atomic_read(&hs->hs_count) <
1400 * For each item in the libcfs hash @hs call the passed callback @func
1401 * and pass to it as an argument each hash item and the private @data.
1403 * a) the function may sleep!
1404 * b) during the callback:
1405 * . the bucket lock is held so the callback must never sleep.
1406 * . if @removal_safe is true, use can remove current item by
1407 * cfs_hash_bd_del_locked
1410 cfs_hash_for_each_tight(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1411 void *data, int remove_safe)
1413 struct hlist_node *hnode;
1414 struct hlist_node *pos;
1415 struct cfs_hash_bd bd;
1417 int excl = !!remove_safe;
1421 cfs_hash_for_each_enter(hs);
1423 cfs_hash_lock(hs, 0);
1424 LASSERT(!cfs_hash_is_rehashing(hs));
1426 cfs_hash_for_each_bucket(hs, &bd, i) {
1427 struct hlist_head *hhead;
1429 cfs_hash_bd_lock(hs, &bd, excl);
1430 if (!func) { /* only glimpse size */
1431 count += bd.bd_bucket->hsb_count;
1432 cfs_hash_bd_unlock(hs, &bd, excl);
1436 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1437 hlist_for_each_safe(hnode, pos, hhead) {
1438 cfs_hash_bucket_validate(hs, &bd, hnode);
1441 if (func(hs, &bd, hnode, data)) {
1442 cfs_hash_bd_unlock(hs, &bd, excl);
1447 cfs_hash_bd_unlock(hs, &bd, excl);
1448 if (loop < CFS_HASH_LOOP_HOG)
1451 cfs_hash_unlock(hs, 0);
1453 cfs_hash_lock(hs, 0);
1456 cfs_hash_unlock(hs, 0);
1458 cfs_hash_for_each_exit(hs);
1462 struct cfs_hash_cond_arg {
1463 cfs_hash_cond_opt_cb_t func;
1468 cfs_hash_cond_del_locked(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1469 struct hlist_node *hnode, void *data)
1471 struct cfs_hash_cond_arg *cond = data;
1473 if (cond->func(cfs_hash_object(hs, hnode), cond->arg))
1474 cfs_hash_bd_del_locked(hs, bd, hnode);
1479 * Delete item from the libcfs hash @hs when @func return true.
1480 * The write lock being hold during loop for each bucket to avoid
1481 * any object be reference.
1484 cfs_hash_cond_del(struct cfs_hash *hs, cfs_hash_cond_opt_cb_t func, void *data)
1486 struct cfs_hash_cond_arg arg = {
1491 cfs_hash_for_each_tight(hs, cfs_hash_cond_del_locked, &arg, 1);
1493 EXPORT_SYMBOL(cfs_hash_cond_del);
1496 cfs_hash_for_each(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1499 cfs_hash_for_each_tight(hs, func, data, 0);
1501 EXPORT_SYMBOL(cfs_hash_for_each);
1504 cfs_hash_for_each_safe(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1507 cfs_hash_for_each_tight(hs, func, data, 1);
1509 EXPORT_SYMBOL(cfs_hash_for_each_safe);
1512 cfs_hash_peek(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1513 struct hlist_node *hnode, void *data)
1516 return 1; /* return 1 to break the loop */
1520 cfs_hash_is_empty(struct cfs_hash *hs)
1524 cfs_hash_for_each_tight(hs, cfs_hash_peek, &empty, 0);
1527 EXPORT_SYMBOL(cfs_hash_is_empty);
1530 cfs_hash_size_get(struct cfs_hash *hs)
1532 return cfs_hash_with_counter(hs) ?
1533 atomic_read(&hs->hs_count) :
1534 cfs_hash_for_each_tight(hs, NULL, NULL, 0);
1536 EXPORT_SYMBOL(cfs_hash_size_get);
1539 * cfs_hash_for_each_relax:
1540 * Iterate the hash table and call @func on each item without
1541 * any lock. This function can't guarantee to finish iteration
1542 * if these features are enabled:
1544 * a. if rehash_key is enabled, an item can be moved from
1545 * one bucket to another bucket
1546 * b. user can remove non-zero-ref item from hash-table,
1547 * so the item can be removed from hash-table, even worse,
1548 * it's possible that user changed key and insert to another
1550 * there's no way for us to finish iteration correctly on previous
1551 * two cases, so iteration has to be stopped on change.
1554 cfs_hash_for_each_relax(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1557 struct hlist_node *hnode;
1558 struct hlist_node *tmp;
1559 struct cfs_hash_bd bd;
1566 stop_on_change = cfs_hash_with_rehash_key(hs) ||
1567 !cfs_hash_with_no_itemref(hs) ||
1568 !hs->hs_ops->hs_put_locked;
1569 cfs_hash_lock(hs, 0);
1570 LASSERT(!cfs_hash_is_rehashing(hs));
1572 cfs_hash_for_each_bucket(hs, &bd, i) {
1573 struct hlist_head *hhead;
1575 cfs_hash_bd_lock(hs, &bd, 0);
1576 version = cfs_hash_bd_version_get(&bd);
1578 cfs_hash_bd_for_each_hlist(hs, &bd, hhead) {
1579 for (hnode = hhead->first; hnode;) {
1580 cfs_hash_bucket_validate(hs, &bd, hnode);
1581 cfs_hash_get(hs, hnode);
1582 cfs_hash_bd_unlock(hs, &bd, 0);
1583 cfs_hash_unlock(hs, 0);
1585 rc = func(hs, &bd, hnode, data);
1587 cfs_hash_put(hs, hnode);
1591 cfs_hash_lock(hs, 0);
1592 cfs_hash_bd_lock(hs, &bd, 0);
1593 if (!stop_on_change) {
1595 cfs_hash_put_locked(hs, hnode);
1597 } else { /* bucket changed? */
1599 cfs_hash_bd_version_get(&bd))
1601 /* safe to continue because no change */
1602 hnode = hnode->next;
1604 if (rc) /* callback wants to break iteration */
1607 if (rc) /* callback wants to break iteration */
1610 cfs_hash_bd_unlock(hs, &bd, 0);
1611 if (rc) /* callback wants to break iteration */
1614 cfs_hash_unlock(hs, 0);
1620 cfs_hash_for_each_nolock(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1623 if (cfs_hash_with_no_lock(hs) ||
1624 cfs_hash_with_rehash_key(hs) ||
1625 !cfs_hash_with_no_itemref(hs))
1628 if (!hs->hs_ops->hs_get ||
1629 (!hs->hs_ops->hs_put && !hs->hs_ops->hs_put_locked))
1632 cfs_hash_for_each_enter(hs);
1633 cfs_hash_for_each_relax(hs, func, data);
1634 cfs_hash_for_each_exit(hs);
1638 EXPORT_SYMBOL(cfs_hash_for_each_nolock);
1641 * For each hash bucket in the libcfs hash @hs call the passed callback
1642 * @func until all the hash buckets are empty. The passed callback @func
1643 * or the previously registered callback hs->hs_put must remove the item
1644 * from the hash. You may either use the cfs_hash_del() or hlist_del()
1645 * functions. No rwlocks will be held during the callback @func it is
1646 * safe to sleep if needed. This function will not terminate until the
1647 * hash is empty. Note it is still possible to concurrently add new
1648 * items in to the hash. It is the callers responsibility to ensure
1649 * the required locking is in place to prevent concurrent insertions.
1652 cfs_hash_for_each_empty(struct cfs_hash *hs, cfs_hash_for_each_cb_t func,
1657 if (cfs_hash_with_no_lock(hs))
1660 if (!hs->hs_ops->hs_get ||
1661 (!hs->hs_ops->hs_put && !hs->hs_ops->hs_put_locked))
1664 cfs_hash_for_each_enter(hs);
1665 while (cfs_hash_for_each_relax(hs, func, data)) {
1666 CDEBUG(D_INFO, "Try to empty hash: %s, loop: %u\n",
1669 cfs_hash_for_each_exit(hs);
1672 EXPORT_SYMBOL(cfs_hash_for_each_empty);
1675 cfs_hash_hlist_for_each(struct cfs_hash *hs, unsigned hindex,
1676 cfs_hash_for_each_cb_t func, void *data)
1678 struct hlist_head *hhead;
1679 struct hlist_node *hnode;
1680 struct cfs_hash_bd bd;
1682 cfs_hash_for_each_enter(hs);
1683 cfs_hash_lock(hs, 0);
1684 if (hindex >= CFS_HASH_NHLIST(hs))
1687 cfs_hash_bd_index_set(hs, hindex, &bd);
1689 cfs_hash_bd_lock(hs, &bd, 0);
1690 hhead = cfs_hash_bd_hhead(hs, &bd);
1691 hlist_for_each(hnode, hhead) {
1692 if (func(hs, &bd, hnode, data))
1695 cfs_hash_bd_unlock(hs, &bd, 0);
1697 cfs_hash_unlock(hs, 0);
1698 cfs_hash_for_each_exit(hs);
1700 EXPORT_SYMBOL(cfs_hash_hlist_for_each);
1703 * For each item in the libcfs hash @hs which matches the @key call
1704 * the passed callback @func and pass to it as an argument each hash
1705 * item and the private @data. During the callback the bucket lock
1706 * is held so the callback must never sleep.
1709 cfs_hash_for_each_key(struct cfs_hash *hs, const void *key,
1710 cfs_hash_for_each_cb_t func, void *data)
1712 struct hlist_node *hnode;
1713 struct cfs_hash_bd bds[2];
1716 cfs_hash_lock(hs, 0);
1718 cfs_hash_dual_bd_get_and_lock(hs, key, bds, 0);
1720 cfs_hash_for_each_bd(bds, 2, i) {
1721 struct hlist_head *hlist = cfs_hash_bd_hhead(hs, &bds[i]);
1723 hlist_for_each(hnode, hlist) {
1724 cfs_hash_bucket_validate(hs, &bds[i], hnode);
1726 if (cfs_hash_keycmp(hs, key, hnode)) {
1727 if (func(hs, &bds[i], hnode, data))
1733 cfs_hash_dual_bd_unlock(hs, bds, 0);
1734 cfs_hash_unlock(hs, 0);
1736 EXPORT_SYMBOL(cfs_hash_for_each_key);
1739 * Rehash the libcfs hash @hs to the given @bits. This can be used
1740 * to grow the hash size when excessive chaining is detected, or to
1741 * shrink the hash when it is larger than needed. When the CFS_HASH_REHASH
1742 * flag is set in @hs the libcfs hash may be dynamically rehashed
1743 * during addition or removal if the hash's theta value exceeds
1744 * either the hs->hs_min_theta or hs->max_theta values. By default
1745 * these values are tuned to keep the chained hash depth small, and
1746 * this approach assumes a reasonably uniform hashing function. The
1747 * theta thresholds for @hs are tunable via cfs_hash_set_theta().
1750 cfs_hash_rehash_cancel_locked(struct cfs_hash *hs)
1754 /* need hold cfs_hash_lock(hs, 1) */
1755 LASSERT(cfs_hash_with_rehash(hs) &&
1756 !cfs_hash_with_no_lock(hs));
1758 if (!cfs_hash_is_rehashing(hs))
1761 if (cfs_wi_deschedule(cfs_sched_rehash, &hs->hs_rehash_wi)) {
1762 hs->hs_rehash_bits = 0;
1766 for (i = 2; cfs_hash_is_rehashing(hs); i++) {
1767 cfs_hash_unlock(hs, 1);
1768 /* raise console warning while waiting too long */
1769 CDEBUG(is_power_of_2(i >> 3) ? D_WARNING : D_INFO,
1770 "hash %s is still rehashing, rescheded %d\n",
1771 hs->hs_name, i - 1);
1773 cfs_hash_lock(hs, 1);
1778 cfs_hash_rehash_cancel(struct cfs_hash *hs)
1780 cfs_hash_lock(hs, 1);
1781 cfs_hash_rehash_cancel_locked(hs);
1782 cfs_hash_unlock(hs, 1);
1786 cfs_hash_rehash(struct cfs_hash *hs, int do_rehash)
1790 LASSERT(cfs_hash_with_rehash(hs) && !cfs_hash_with_no_lock(hs));
1792 cfs_hash_lock(hs, 1);
1794 rc = cfs_hash_rehash_bits(hs);
1796 cfs_hash_unlock(hs, 1);
1800 hs->hs_rehash_bits = rc;
1802 /* launch and return */
1803 cfs_wi_schedule(cfs_sched_rehash, &hs->hs_rehash_wi);
1804 cfs_hash_unlock(hs, 1);
1808 /* rehash right now */
1809 cfs_hash_unlock(hs, 1);
1811 return cfs_hash_rehash_worker(&hs->hs_rehash_wi);
1815 cfs_hash_rehash_bd(struct cfs_hash *hs, struct cfs_hash_bd *old)
1817 struct cfs_hash_bd new;
1818 struct hlist_head *hhead;
1819 struct hlist_node *hnode;
1820 struct hlist_node *pos;
1824 /* hold cfs_hash_lock(hs, 1), so don't need any bucket lock */
1825 cfs_hash_bd_for_each_hlist(hs, old, hhead) {
1826 hlist_for_each_safe(hnode, pos, hhead) {
1827 key = cfs_hash_key(hs, hnode);
1829 /* Validate hnode is in the correct bucket. */
1830 cfs_hash_bucket_validate(hs, old, hnode);
1832 * Delete from old hash bucket; move to new bucket.
1833 * ops->hs_key must be defined.
1835 cfs_hash_bd_from_key(hs, hs->hs_rehash_buckets,
1836 hs->hs_rehash_bits, key, &new);
1837 cfs_hash_bd_move_locked(hs, old, &new, hnode);
1846 cfs_hash_rehash_worker(struct cfs_workitem *wi)
1848 struct cfs_hash *hs = container_of(wi, struct cfs_hash, hs_rehash_wi);
1849 struct cfs_hash_bucket **bkts;
1850 struct cfs_hash_bd bd;
1851 unsigned int old_size;
1852 unsigned int new_size;
1858 LASSERT(hs && cfs_hash_with_rehash(hs));
1860 cfs_hash_lock(hs, 0);
1861 LASSERT(cfs_hash_is_rehashing(hs));
1863 old_size = CFS_HASH_NBKT(hs);
1864 new_size = CFS_HASH_RH_NBKT(hs);
1866 cfs_hash_unlock(hs, 0);
1869 * don't need hs::hs_rwlock for hs::hs_buckets,
1870 * because nobody can change bkt-table except me.
1872 bkts = cfs_hash_buckets_realloc(hs, hs->hs_buckets,
1873 old_size, new_size);
1874 cfs_hash_lock(hs, 1);
1880 if (bkts == hs->hs_buckets) {
1881 bkts = NULL; /* do nothing */
1885 rc = __cfs_hash_theta(hs);
1886 if ((rc >= hs->hs_min_theta) && (rc <= hs->hs_max_theta)) {
1887 /* free the new allocated bkt-table */
1888 old_size = new_size;
1889 new_size = CFS_HASH_NBKT(hs);
1894 LASSERT(!hs->hs_rehash_buckets);
1895 hs->hs_rehash_buckets = bkts;
1898 cfs_hash_for_each_bucket(hs, &bd, i) {
1899 if (cfs_hash_is_exiting(hs)) {
1901 /* someone wants to destroy the hash, abort now */
1902 if (old_size < new_size) /* OK to free old bkt-table */
1904 /* it's shrinking, need free new bkt-table */
1905 hs->hs_rehash_buckets = NULL;
1906 old_size = new_size;
1907 new_size = CFS_HASH_NBKT(hs);
1911 count += cfs_hash_rehash_bd(hs, &bd);
1912 if (count < CFS_HASH_LOOP_HOG ||
1913 cfs_hash_is_iterating(hs)) { /* need to finish ASAP */
1918 cfs_hash_unlock(hs, 1);
1920 cfs_hash_lock(hs, 1);
1923 hs->hs_rehash_count++;
1925 bkts = hs->hs_buckets;
1926 hs->hs_buckets = hs->hs_rehash_buckets;
1927 hs->hs_rehash_buckets = NULL;
1929 hs->hs_cur_bits = hs->hs_rehash_bits;
1931 hs->hs_rehash_bits = 0;
1932 if (rc == -ESRCH) /* never be scheduled again */
1933 cfs_wi_exit(cfs_sched_rehash, wi);
1934 bsize = cfs_hash_bkt_size(hs);
1935 cfs_hash_unlock(hs, 1);
1936 /* can't refer to @hs anymore because it could be destroyed */
1938 cfs_hash_buckets_free(bkts, bsize, new_size, old_size);
1940 CDEBUG(D_INFO, "early quit of rehashing: %d\n", rc);
1941 /* return 1 only if cfs_wi_exit is called */
1942 return rc == -ESRCH;
1946 * Rehash the object referenced by @hnode in the libcfs hash @hs. The
1947 * @old_key must be provided to locate the objects previous location
1948 * in the hash, and the @new_key will be used to reinsert the object.
1949 * Use this function instead of a cfs_hash_add() + cfs_hash_del()
1950 * combo when it is critical that there is no window in time where the
1951 * object is missing from the hash. When an object is being rehashed
1952 * the registered cfs_hash_get() and cfs_hash_put() functions will
1955 void cfs_hash_rehash_key(struct cfs_hash *hs, const void *old_key,
1956 void *new_key, struct hlist_node *hnode)
1958 struct cfs_hash_bd bds[3];
1959 struct cfs_hash_bd old_bds[2];
1960 struct cfs_hash_bd new_bd;
1962 LASSERT(!hlist_unhashed(hnode));
1964 cfs_hash_lock(hs, 0);
1966 cfs_hash_dual_bd_get(hs, old_key, old_bds);
1967 cfs_hash_bd_get(hs, new_key, &new_bd);
1969 bds[0] = old_bds[0];
1970 bds[1] = old_bds[1];
1973 /* NB: bds[0] and bds[1] are ordered already */
1974 cfs_hash_bd_order(&bds[1], &bds[2]);
1975 cfs_hash_bd_order(&bds[0], &bds[1]);
1977 cfs_hash_multi_bd_lock(hs, bds, 3, 1);
1978 if (likely(!old_bds[1].bd_bucket)) {
1979 cfs_hash_bd_move_locked(hs, &old_bds[0], &new_bd, hnode);
1981 cfs_hash_dual_bd_finddel_locked(hs, old_bds, old_key, hnode);
1982 cfs_hash_bd_add_locked(hs, &new_bd, hnode);
1984 /* overwrite key inside locks, otherwise may screw up with
1985 * other operations, i.e: rehash
1987 cfs_hash_keycpy(hs, hnode, new_key);
1989 cfs_hash_multi_bd_unlock(hs, bds, 3, 1);
1990 cfs_hash_unlock(hs, 0);
1992 EXPORT_SYMBOL(cfs_hash_rehash_key);
1994 void cfs_hash_debug_header(struct seq_file *m)
1996 seq_printf(m, "%-*s cur min max theta t-min t-max flags rehash count maxdep maxdepb distribution\n",
1997 CFS_HASH_BIGNAME_LEN, "name");
1999 EXPORT_SYMBOL(cfs_hash_debug_header);
2001 static struct cfs_hash_bucket **
2002 cfs_hash_full_bkts(struct cfs_hash *hs)
2004 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2005 if (!hs->hs_rehash_buckets)
2006 return hs->hs_buckets;
2008 LASSERT(hs->hs_rehash_bits != 0);
2009 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2010 hs->hs_rehash_buckets : hs->hs_buckets;
2014 cfs_hash_full_nbkt(struct cfs_hash *hs)
2016 /* NB: caller should hold hs->hs_rwlock if REHASH is set */
2017 if (!hs->hs_rehash_buckets)
2018 return CFS_HASH_NBKT(hs);
2020 LASSERT(hs->hs_rehash_bits != 0);
2021 return hs->hs_rehash_bits > hs->hs_cur_bits ?
2022 CFS_HASH_RH_NBKT(hs) : CFS_HASH_NBKT(hs);
2025 void cfs_hash_debug_str(struct cfs_hash *hs, struct seq_file *m)
2027 int dist[8] = { 0, };
2034 cfs_hash_lock(hs, 0);
2035 theta = __cfs_hash_theta(hs);
2037 seq_printf(m, "%-*s %5d %5d %5d %d.%03d %d.%03d %d.%03d 0x%02x %6d ",
2038 CFS_HASH_BIGNAME_LEN, hs->hs_name,
2039 1 << hs->hs_cur_bits, 1 << hs->hs_min_bits,
2040 1 << hs->hs_max_bits,
2041 __cfs_hash_theta_int(theta), __cfs_hash_theta_frac(theta),
2042 __cfs_hash_theta_int(hs->hs_min_theta),
2043 __cfs_hash_theta_frac(hs->hs_min_theta),
2044 __cfs_hash_theta_int(hs->hs_max_theta),
2045 __cfs_hash_theta_frac(hs->hs_max_theta),
2046 hs->hs_flags, hs->hs_rehash_count);
2049 * The distribution is a summary of the chained hash depth in
2050 * each of the libcfs hash buckets. Each buckets hsb_count is
2051 * divided by the hash theta value and used to generate a
2052 * histogram of the hash distribution. A uniform hash will
2053 * result in all hash buckets being close to the average thus
2054 * only the first few entries in the histogram will be non-zero.
2055 * If you hash function results in a non-uniform hash the will
2056 * be observable by outlier bucks in the distribution histogram.
2058 * Uniform hash distribution: 128/128/0/0/0/0/0/0
2059 * Non-Uniform hash distribution: 128/125/0/0/0/0/2/1
2061 for (i = 0; i < cfs_hash_full_nbkt(hs); i++) {
2062 struct cfs_hash_bd bd;
2064 bd.bd_bucket = cfs_hash_full_bkts(hs)[i];
2065 cfs_hash_bd_lock(hs, &bd, 0);
2066 if (maxdep < bd.bd_bucket->hsb_depmax) {
2067 maxdep = bd.bd_bucket->hsb_depmax;
2068 maxdepb = ffz(~maxdep);
2070 total += bd.bd_bucket->hsb_count;
2071 dist[min(fls(bd.bd_bucket->hsb_count / max(theta, 1)), 7)]++;
2072 cfs_hash_bd_unlock(hs, &bd, 0);
2075 seq_printf(m, "%7d %7d %7d ", total, maxdep, maxdepb);
2076 for (i = 0; i < 8; i++)
2077 seq_printf(m, "%d%c", dist[i], (i == 7) ? '\n' : '/');
2079 cfs_hash_unlock(hs, 0);
2081 EXPORT_SYMBOL(cfs_hash_debug_str);