4 * Userspace RCU library - Lock-Free Expandable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
32 * Some specificities of this Lock-Free Expandable RCU Hash Table
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
51 * - The resize operation only allows expanding the hash table.
52 * It is triggered either through an API call or automatically by
53 * detecting long chains in the add operation.
54 * - Resize operation initiated by long chain detection is executed by a
55 * call_rcu thread, which keeps lock-freedom of add and remove.
56 * - Resize operations are protected by a mutex.
57 * - The removal operation is split in two parts: first, a "removed"
58 * flag is set in the next pointer within the node to remove. Then,
59 * a "garbage collection" is performed in the bucket containing the
60 * removed node (from the start of the bucket up to the removed node).
61 * All encountered nodes with "removed" flag set in their next
62 * pointers are removed from the linked-list. If the cmpxchg used for
63 * removal fails (due to concurrent garbage-collection or concurrent
64 * add), we retry from the beginning of the bucket. This ensures that
65 * the node with "removed" flag set is removed from the hash table
66 * (not visible to lookups anymore) before the RCU read-side critical
67 * section held across removal ends. Furthermore, this ensures that
68 * the node with "removed" flag set is removed from the linked-list
69 * before its memory is reclaimed. Only the thread which removal
70 * successfully set the "removed" flag (with a cmpxchg) into a node's
71 * next pointer is considered to have succeeded its removal (and thus
72 * owns the node to reclaim). Because we garbage-collect starting from
73 * an invariant node (the start-of-bucket dummy node) up to the
74 * "removed" node (or find a reverse-hash that is higher), we are sure
75 * that a successful traversal of the chain leads to a chain that is
76 * present in the linked-list (the start node is never removed) and
77 * that is does not contain the "removed" node anymore, even if
78 * concurrent delete/add operations are changing the structure of the
80 * - The add operation performs gargage collection of buckets if it
81 * encounters nodes with removed flag set in the bucket where it wants
82 * to add its new node. This ensures lock-freedom of add operation by
83 * helping the remover unlink nodes from the list rather than to wait
85 * - A RCU "order table" indexed by log2(hash index) is copied and
86 * expanded by the resize operation. This order table allows finding
87 * the "dummy node" tables.
88 * - There is one dummy node table per hash index order. The size of
89 * each dummy node table is half the number of hashes contained in
91 * - call_rcu is used to garbage-collect the old order table.
92 * - The per-order dummy node tables contain a compact version of the
93 * hash table nodes. These tables are invariant after they are
94 * populated into the hash table.
106 #include <urcu-call-rcu.h>
107 #include <urcu/arch.h>
108 #include <urcu/uatomic.h>
109 #include <urcu/jhash.h>
110 #include <urcu/compiler.h>
111 #include <urcu/rculfhash.h>
115 #ifndef CDS_HT_NOATOMIC
116 # define test_cmpxchg(ptr, old, _new) uatomic_cmpxchg(ptr, old, _new)
118 # define test_cmpxchg(ptr, old, _new) \
120 __typeof__(*(ptr)) __readptr = *(ptr); \
121 (__readptr == (old)) ? ({ \
129 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
131 #define dbg_printf(fmt, args...)
134 #define CHAIN_LEN_TARGET 4
135 #define CHAIN_LEN_RESIZE_THRESHOLD 8
138 #define max(a, b) ((a) > (b) ? (a) : (b))
142 * The removed flag needs to be updated atomically with the pointer.
143 * The dummy flag does not require to be updated atomically with the
144 * pointer, but it is added as a pointer low bit flag to save space.
146 #define REMOVED_FLAG (1UL << 0)
147 #define DUMMY_FLAG (1UL << 1)
148 #define FLAGS_MASK ((1UL << 2) - 1)
151 unsigned long size
; /* always a power of 2 */
152 unsigned long resize_target
;
153 int resize_initiated
;
154 struct rcu_head head
;
155 struct _cds_lfht_node
*tbl
[0];
159 struct rcu_table
*t
; /* shared */
160 cds_lfht_hash_fct hash_fct
;
161 cds_lfht_compare_fct compare_fct
;
162 unsigned long hash_seed
;
163 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
164 unsigned int in_progress_resize
, in_progress_destroy
;
165 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
166 void (*func
)(struct rcu_head
*head
));
169 struct rcu_resize_work
{
170 struct rcu_head head
;
175 * Algorithm to reverse bits in a word by lookup table, extended to
178 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
179 * Originally from Public Domain.
182 static const uint8_t BitReverseTable256
[256] =
184 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
185 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
186 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
187 R6(0), R6(2), R6(1), R6(3)
194 uint8_t bit_reverse_u8(uint8_t v
)
196 return BitReverseTable256
[v
];
199 static __attribute__((unused
))
200 uint32_t bit_reverse_u32(uint32_t v
)
202 return ((uint32_t) bit_reverse_u8(v
) << 24) |
203 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
204 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
205 ((uint32_t) bit_reverse_u8(v
>> 24));
208 static __attribute__((unused
))
209 uint64_t bit_reverse_u64(uint64_t v
)
211 return ((uint64_t) bit_reverse_u8(v
) << 56) |
212 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
213 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
214 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
215 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
216 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
217 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
218 ((uint64_t) bit_reverse_u8(v
>> 56));
222 unsigned long bit_reverse_ulong(unsigned long v
)
224 #if (CAA_BITS_PER_LONG == 32)
225 return bit_reverse_u32(v
);
227 return bit_reverse_u64(v
);
232 * fls: returns the position of the most significant bit.
233 * Returns 0 if no bit is set, else returns the position of the most
234 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
236 #if defined(__i386) || defined(__x86_64)
238 unsigned int fls_u32(uint32_t x
)
246 : "=r" (r
) : "rm" (x
));
252 #if defined(__x86_64)
254 unsigned int fls_u64(uint64_t x
)
262 : "=r" (r
) : "rm" (x
));
269 static __attribute__((unused
))
270 unsigned int fls_u64(uint64_t x
)
277 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
281 if (!(x
& 0xFFFF000000000000ULL
)) {
285 if (!(x
& 0xFF00000000000000ULL
)) {
289 if (!(x
& 0xF000000000000000ULL
)) {
293 if (!(x
& 0xC000000000000000ULL
)) {
297 if (!(x
& 0x8000000000000000ULL
)) {
306 static __attribute__((unused
))
307 unsigned int fls_u32(uint32_t x
)
313 if (!(x
& 0xFFFF0000U
)) {
317 if (!(x
& 0xFF000000U
)) {
321 if (!(x
& 0xF0000000U
)) {
325 if (!(x
& 0xC0000000U
)) {
329 if (!(x
& 0x80000000U
)) {
337 unsigned int fls_ulong(unsigned long x
)
339 #if (CAA_BITS_PER_lONG == 32)
346 int get_count_order_u32(uint32_t x
)
350 order
= fls_u32(x
) - 1;
356 int get_count_order_ulong(unsigned long x
)
360 order
= fls_ulong(x
) - 1;
367 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
);
370 void check_resize(struct cds_lfht
*ht
, struct rcu_table
*t
,
374 dbg_printf("WARNING: large chain length: %u.\n",
376 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
377 cds_lfht_resize_lazy(ht
, t
,
378 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
382 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
384 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
388 int is_removed(struct cds_lfht_node
*node
)
390 return ((unsigned long) node
) & REMOVED_FLAG
;
394 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
396 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
400 int is_dummy(struct cds_lfht_node
*node
)
402 return ((unsigned long) node
) & DUMMY_FLAG
;
406 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
408 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
412 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
414 unsigned long old1
, old2
;
416 old1
= uatomic_read(ptr
);
421 } while ((old1
= test_cmpxchg(ptr
, old2
, v
)) != old2
);
426 * Remove all logically deleted nodes from a bucket up to a certain node key.
429 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
431 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
435 /* We can always skip the dummy node initially */
436 iter
= rcu_dereference(iter_prev
->p
.next
);
437 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
439 if (unlikely(!clear_flag(iter
)))
441 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
443 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
444 if (likely(is_removed(next
)))
446 iter_prev
= clear_flag(iter
);
449 assert(!is_removed(iter
));
451 new_next
= flag_dummy(clear_flag(next
));
453 new_next
= clear_flag(next
);
454 (void) test_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
459 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
, struct rcu_table
*t
,
460 struct cds_lfht_node
*node
, int unique
, int dummy
)
462 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
464 struct _cds_lfht_node
*lookup
;
465 unsigned long hash
, index
, order
;
469 node
->p
.next
= flag_dummy(NULL
);
470 return node
; /* Initial first add (head) */
472 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
474 uint32_t chain_len
= 0;
477 * iter_prev points to the non-removed node prior to the
480 index
= hash
& (t
->size
- 1);
481 order
= get_count_order_ulong(index
+ 1);
482 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
483 iter_prev
= (struct cds_lfht_node
*) lookup
;
484 /* We can always skip the dummy node initially */
485 iter
= rcu_dereference(iter_prev
->p
.next
);
486 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
488 if (unlikely(!clear_flag(iter
)))
490 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
492 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
493 if (unlikely(is_removed(next
)))
497 && !ht
->compare_fct(node
->key
, node
->key_len
,
498 clear_flag(iter
)->key
,
499 clear_flag(iter
)->key_len
))
500 return clear_flag(iter
);
501 /* Only account for identical reverse hash once */
502 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
504 check_resize(ht
, t
, ++chain_len
);
505 iter_prev
= clear_flag(iter
);
509 assert(node
!= clear_flag(iter
));
510 assert(!is_removed(iter_prev
));
511 assert(iter_prev
!= node
);
513 node
->p
.next
= clear_flag(iter
);
515 node
->p
.next
= flag_dummy(clear_flag(iter
));
517 new_node
= flag_dummy(node
);
520 if (test_cmpxchg(&iter_prev
->p
.next
, iter
,
522 continue; /* retry */
526 assert(!is_removed(iter
));
528 new_next
= flag_dummy(clear_flag(next
));
530 new_next
= clear_flag(next
);
531 (void) test_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
535 /* Garbage collect logically removed nodes in the bucket */
536 index
= hash
& (t
->size
- 1);
537 order
= get_count_order_ulong(index
+ 1);
538 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
539 dummy_node
= (struct cds_lfht_node
*) lookup
;
540 _cds_lfht_gc_bucket(dummy_node
, node
);
545 int _cds_lfht_remove(struct cds_lfht
*ht
, struct rcu_table
*t
,
546 struct cds_lfht_node
*node
)
548 struct cds_lfht_node
*dummy
, *next
, *old
;
549 struct _cds_lfht_node
*lookup
;
551 unsigned long hash
, index
, order
;
553 /* logically delete the node */
554 old
= rcu_dereference(node
->p
.next
);
557 if (unlikely(is_removed(next
)))
559 assert(!is_dummy(next
));
560 old
= test_cmpxchg(&node
->p
.next
, next
,
562 } while (old
!= next
);
564 /* We performed the (logical) deletion. */
568 * Ensure that the node is not visible to readers anymore: lookup for
569 * the node, and remove it (along with any other logically removed node)
572 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
573 index
= hash
& (t
->size
- 1);
574 order
= get_count_order_ulong(index
+ 1);
575 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
576 dummy
= (struct cds_lfht_node
*) lookup
;
577 _cds_lfht_gc_bucket(dummy
, node
);
580 * Only the flagging action indicated that we (and no other)
581 * removed the node from the hash.
584 assert(is_removed(rcu_dereference(node
->p
.next
)));
591 void init_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
592 unsigned long first_order
, unsigned long len_order
)
594 unsigned long i
, end_order
;
596 dbg_printf("init table: first_order %lu end_order %lu\n",
597 first_order
, first_order
+ len_order
);
598 end_order
= first_order
+ len_order
;
599 t
->size
= !first_order
? 0 : (1UL << (first_order
- 1));
600 for (i
= first_order
; i
< end_order
; i
++) {
601 unsigned long j
, len
;
603 len
= !i
? 1 : 1UL << (i
- 1);
604 dbg_printf("init order %lu len: %lu\n", i
, len
);
605 t
->tbl
[i
] = calloc(len
, sizeof(struct _cds_lfht_node
));
606 for (j
= 0; j
< len
; j
++) {
607 dbg_printf("init entry: i %lu j %lu hash %lu\n",
608 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
609 struct cds_lfht_node
*new_node
=
610 (struct cds_lfht_node
*) &t
->tbl
[i
][j
];
611 new_node
->p
.reverse_hash
=
612 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
613 (void) _cds_lfht_add(ht
, t
, new_node
, 0, 1);
614 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
617 /* Update table size */
618 t
->size
= !i
? 1 : (1UL << i
);
619 dbg_printf("init new size: %lu\n", t
->size
);
620 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
623 t
->resize_target
= t
->size
;
624 t
->resize_initiated
= 0;
627 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
628 cds_lfht_compare_fct compare_fct
,
629 unsigned long hash_seed
,
630 unsigned long init_size
,
631 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
632 void (*func
)(struct rcu_head
*head
)))
637 /* init_size must be power of two */
638 if (init_size
&& (init_size
& (init_size
- 1)))
640 ht
= calloc(1, sizeof(struct cds_lfht
));
641 ht
->hash_fct
= hash_fct
;
642 ht
->compare_fct
= compare_fct
;
643 ht
->hash_seed
= hash_seed
;
644 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
645 ht
->in_progress_resize
= 0;
646 /* this mutex should not nest in read-side C.S. */
647 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
648 order
= get_count_order_ulong(max(init_size
, 1)) + 1;
649 ht
->t
= calloc(1, sizeof(struct cds_lfht
)
650 + (order
* sizeof(struct _cds_lfht_node
*)));
652 pthread_mutex_lock(&ht
->resize_mutex
);
653 init_table(ht
, ht
->t
, 0, order
);
654 pthread_mutex_unlock(&ht
->resize_mutex
);
658 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
661 struct cds_lfht_node
*node
, *next
;
662 struct _cds_lfht_node
*lookup
;
663 unsigned long hash
, reverse_hash
, index
, order
;
665 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
666 reverse_hash
= bit_reverse_ulong(hash
);
668 t
= rcu_dereference(ht
->t
);
669 index
= hash
& (t
->size
- 1);
670 order
= get_count_order_ulong(index
+ 1);
671 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
672 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
673 hash
, index
, order
, index
& ((1UL << (order
- 1)) - 1));
674 node
= (struct cds_lfht_node
*) lookup
;
678 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
682 next
= rcu_dereference(node
->p
.next
);
683 if (likely(!is_removed(next
))
685 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
688 node
= clear_flag(next
);
690 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
694 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
695 struct cds_lfht_node
*node
)
697 struct cds_lfht_node
*next
;
698 unsigned long reverse_hash
;
702 reverse_hash
= node
->p
.reverse_hash
;
704 key_len
= node
->key_len
;
705 next
= rcu_dereference(node
->p
.next
);
706 node
= clear_flag(next
);
711 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
715 next
= rcu_dereference(node
->p
.next
);
716 if (likely(!is_removed(next
))
718 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
721 node
= clear_flag(next
);
723 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
727 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
732 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
733 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
735 t
= rcu_dereference(ht
->t
);
736 (void) _cds_lfht_add(ht
, t
, node
, 0, 0);
739 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
740 struct cds_lfht_node
*node
)
745 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
746 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
748 t
= rcu_dereference(ht
->t
);
749 return _cds_lfht_add(ht
, t
, node
, 1, 0);
752 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
756 t
= rcu_dereference(ht
->t
);
757 return _cds_lfht_remove(ht
, t
, node
);
761 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
764 struct cds_lfht_node
*node
;
765 struct _cds_lfht_node
*lookup
;
766 unsigned long order
, i
;
769 /* Check that the table is empty */
770 lookup
= &t
->tbl
[0][0];
771 node
= (struct cds_lfht_node
*) lookup
;
773 node
= clear_flag(node
)->p
.next
;
776 assert(!is_removed(node
));
777 } while (clear_flag(node
));
778 /* Internal sanity check: all nodes left should be dummy */
779 for (order
= 0; order
< get_count_order_ulong(t
->size
) + 1; order
++) {
782 len
= !order
? 1 : 1UL << (order
- 1);
783 for (i
= 0; i
< len
; i
++) {
784 dbg_printf("delete order %lu i %lu hash %lu\n",
786 bit_reverse_ulong(t
->tbl
[order
][i
].reverse_hash
));
787 assert(is_dummy(t
->tbl
[order
][i
].next
));
795 * Should only be called when no more concurrent readers nor writers can
796 * possibly access the table.
798 int cds_lfht_destroy(struct cds_lfht
*ht
)
802 /* Wait for in-flight resize operations to complete */
803 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
804 while (uatomic_read(&ht
->in_progress_resize
))
805 poll(NULL
, 0, 100); /* wait for 100ms */
806 ret
= cds_lfht_delete_dummy(ht
);
814 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
815 unsigned long *count
,
816 unsigned long *removed
)
819 struct cds_lfht_node
*node
, *next
;
820 struct _cds_lfht_node
*lookup
;
821 unsigned long nr_dummy
= 0;
826 t
= rcu_dereference(ht
->t
);
827 /* Count non-dummy nodes in the table */
828 lookup
= &t
->tbl
[0][0];
829 node
= (struct cds_lfht_node
*) lookup
;
831 next
= rcu_dereference(node
->p
.next
);
832 if (is_removed(next
)) {
833 assert(!is_dummy(next
));
835 } else if (!is_dummy(next
))
839 node
= clear_flag(next
);
841 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
845 void cds_lfht_free_table_cb(struct rcu_head
*head
)
847 struct rcu_table
*t
=
848 caa_container_of(head
, struct rcu_table
, head
);
852 /* called with resize mutex held */
854 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
856 unsigned long new_size
, old_size
, old_order
, new_order
;
857 struct rcu_table
*new_t
, *old_t
;
860 old_size
= old_t
->size
;
861 old_order
= get_count_order_ulong(old_size
) + 1;
863 new_size
= CMM_LOAD_SHARED(old_t
->resize_target
);
864 if (old_size
== new_size
)
866 new_order
= get_count_order_ulong(new_size
) + 1;
867 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
868 old_size
, old_order
, new_size
, new_order
);
869 new_t
= malloc(sizeof(struct cds_lfht
)
870 + (new_order
* sizeof(struct _cds_lfht_node
*)));
871 assert(new_size
> old_size
);
872 memcpy(&new_t
->tbl
, &old_t
->tbl
,
873 old_order
* sizeof(struct _cds_lfht_node
*));
874 init_table(ht
, new_t
, old_order
, new_order
- old_order
);
875 /* Changing table and size atomically wrt lookups */
876 rcu_assign_pointer(ht
->t
, new_t
);
877 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
881 unsigned long resize_target_update(struct rcu_table
*t
,
884 return _uatomic_max(&t
->resize_target
,
885 t
->size
<< growth_order
);
888 void cds_lfht_resize(struct cds_lfht
*ht
, int growth
)
890 struct rcu_table
*t
= rcu_dereference(ht
->t
);
891 unsigned long target_size
;
895 * Silently refuse to shrink hash table. (not supported)
897 dbg_printf("shrinking hash table not supported.\n");
901 target_size
= resize_target_update(t
, growth
);
902 if (t
->size
< target_size
) {
903 CMM_STORE_SHARED(t
->resize_initiated
, 1);
904 pthread_mutex_lock(&ht
->resize_mutex
);
905 _do_cds_lfht_resize(ht
);
906 pthread_mutex_unlock(&ht
->resize_mutex
);
911 void do_resize_cb(struct rcu_head
*head
)
913 struct rcu_resize_work
*work
=
914 caa_container_of(head
, struct rcu_resize_work
, head
);
915 struct cds_lfht
*ht
= work
->ht
;
917 pthread_mutex_lock(&ht
->resize_mutex
);
918 _do_cds_lfht_resize(ht
);
919 pthread_mutex_unlock(&ht
->resize_mutex
);
921 cmm_smp_mb(); /* finish resize before decrement */
922 uatomic_dec(&ht
->in_progress_resize
);
926 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
)
928 struct rcu_resize_work
*work
;
929 unsigned long target_size
;
931 target_size
= resize_target_update(t
, growth
);
932 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
933 uatomic_inc(&ht
->in_progress_resize
);
934 cmm_smp_mb(); /* increment resize count before calling it */
935 work
= malloc(sizeof(*work
));
937 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
938 CMM_STORE_SHARED(t
->resize_initiated
, 1);