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.
107 #include <urcu-call-rcu.h>
108 #include <urcu/arch.h>
109 #include <urcu/uatomic.h>
110 #include <urcu/jhash.h>
111 #include <urcu/compiler.h>
112 #include <urcu/rculfhash.h>
117 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
119 #define dbg_printf(fmt, args...)
122 #define CHAIN_LEN_TARGET 4
123 #define CHAIN_LEN_RESIZE_THRESHOLD 8
125 /* Commit counter changes to global counter each 1024 steps */
126 #define COUNT_COMMIT_ORDER 10
129 #define max(a, b) ((a) > (b) ? (a) : (b))
133 * The removed flag needs to be updated atomically with the pointer.
134 * The dummy flag does not require to be updated atomically with the
135 * pointer, but it is added as a pointer low bit flag to save space.
137 #define REMOVED_FLAG (1UL << 0)
138 #define DUMMY_FLAG (1UL << 1)
139 #define FLAGS_MASK ((1UL << 2) - 1)
141 struct ht_items_count
{
143 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
146 unsigned long size
; /* always a power of 2 */
147 unsigned long resize_target
;
148 int resize_initiated
;
149 struct rcu_head head
;
150 struct _cds_lfht_node
*tbl
[0];
154 struct rcu_table
*t
; /* shared */
155 cds_lfht_hash_fct hash_fct
;
156 cds_lfht_compare_fct compare_fct
;
157 unsigned long hash_seed
;
158 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
159 unsigned int in_progress_resize
, in_progress_destroy
;
160 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
161 void (*func
)(struct rcu_head
*head
));
162 unsigned long count
; /* global approximate item count */
163 struct ht_items_count
*percpu_count
; /* per-cpu item count */
166 struct rcu_resize_work
{
167 struct rcu_head head
;
172 * Algorithm to reverse bits in a word by lookup table, extended to
175 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
176 * Originally from Public Domain.
179 static const uint8_t BitReverseTable256
[256] =
181 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
182 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
183 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
184 R6(0), R6(2), R6(1), R6(3)
191 uint8_t bit_reverse_u8(uint8_t v
)
193 return BitReverseTable256
[v
];
196 static __attribute__((unused
))
197 uint32_t bit_reverse_u32(uint32_t v
)
199 return ((uint32_t) bit_reverse_u8(v
) << 24) |
200 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
201 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
202 ((uint32_t) bit_reverse_u8(v
>> 24));
205 static __attribute__((unused
))
206 uint64_t bit_reverse_u64(uint64_t v
)
208 return ((uint64_t) bit_reverse_u8(v
) << 56) |
209 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
210 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
211 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
212 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
213 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
214 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
215 ((uint64_t) bit_reverse_u8(v
>> 56));
219 unsigned long bit_reverse_ulong(unsigned long v
)
221 #if (CAA_BITS_PER_LONG == 32)
222 return bit_reverse_u32(v
);
224 return bit_reverse_u64(v
);
229 * fls: returns the position of the most significant bit.
230 * Returns 0 if no bit is set, else returns the position of the most
231 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
233 #if defined(__i386) || defined(__x86_64)
235 unsigned int fls_u32(uint32_t x
)
243 : "=r" (r
) : "rm" (x
));
249 #if defined(__x86_64)
251 unsigned int fls_u64(uint64_t x
)
259 : "=r" (r
) : "rm" (x
));
266 static __attribute__((unused
))
267 unsigned int fls_u64(uint64_t x
)
274 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
278 if (!(x
& 0xFFFF000000000000ULL
)) {
282 if (!(x
& 0xFF00000000000000ULL
)) {
286 if (!(x
& 0xF000000000000000ULL
)) {
290 if (!(x
& 0xC000000000000000ULL
)) {
294 if (!(x
& 0x8000000000000000ULL
)) {
303 static __attribute__((unused
))
304 unsigned int fls_u32(uint32_t x
)
310 if (!(x
& 0xFFFF0000U
)) {
314 if (!(x
& 0xFF000000U
)) {
318 if (!(x
& 0xF0000000U
)) {
322 if (!(x
& 0xC0000000U
)) {
326 if (!(x
& 0x80000000U
)) {
334 unsigned int fls_ulong(unsigned long x
)
336 #if (CAA_BITS_PER_lONG == 32)
343 int get_count_order_u32(uint32_t x
)
347 order
= fls_u32(x
) - 1;
353 int get_count_order_ulong(unsigned long x
)
357 order
= fls_ulong(x
) - 1;
364 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
);
367 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
368 * available, then we support hash table item accounting.
369 * In the unfortunate event the number of CPUs reported would be
370 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
373 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
375 static long nr_cpus_mask
= -1;
378 struct ht_items_count
*alloc_per_cpu_items_count(void)
380 struct ht_items_count
*count
;
382 switch (nr_cpus_mask
) {
389 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
395 * round up number of CPUs to next power of two, so we
396 * can use & for modulo.
398 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
399 nr_cpus_mask
= maxcpus
- 1;
403 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
408 void free_per_cpu_items_count(struct ht_items_count
*count
)
418 assert(nr_cpus_mask
>= 0);
419 cpu
= sched_getcpu();
420 if (unlikely(cpu
< 0))
423 return cpu
& nr_cpus_mask
;
427 unsigned long ht_count_update(struct cds_lfht
*ht
, unsigned int value
)
431 if (unlikely(!ht
->percpu_count
))
434 if (unlikely(cpu
< 0))
436 return uatomic_add_return(&ht
->percpu_count
[cpu
].v
, 1);
440 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
442 unsigned long percpu_count
;
444 percpu_count
= ht_count_update(ht
, 1);
445 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
448 dbg_printf("add percpu %lu\n", percpu_count
);
449 count
= uatomic_add_return(&ht
->count
,
450 1UL << COUNT_COMMIT_ORDER
);
452 if (!(count
& (count
- 1))) {
453 dbg_printf("add global %lu\n", count
);
454 cds_lfht_resize_lazy(ht
, t
, 1);
460 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
462 unsigned long percpu_count
;
464 percpu_count
= ht_count_update(ht
, -1);
465 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
468 dbg_printf("remove percpu %lu\n", percpu_count
);
469 count
= uatomic_add_return(&ht
->count
,
470 1UL << COUNT_COMMIT_ORDER
);
472 if (!(count
& (count
- 1))) {
473 dbg_printf("remove global %lu\n", count
);
474 cds_lfht_resize_lazy(ht
, t
, -1);
479 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
481 static const long nr_cpus_mask
= -1;
484 struct ht_items_count
*alloc_per_cpu_items_count(void)
490 void free_per_cpu_items_count(struct ht_items_count
*count
)
495 void ht_count_add(struct cds_lfht
*ht
)
500 void ht_count_remove(struct cds_lfht
*ht
)
504 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
508 void check_resize(struct cds_lfht
*ht
, struct rcu_table
*t
,
513 dbg_printf("WARNING: large chain length: %u.\n",
515 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
516 cds_lfht_resize_lazy(ht
, t
,
517 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
521 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
523 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
527 int is_removed(struct cds_lfht_node
*node
)
529 return ((unsigned long) node
) & REMOVED_FLAG
;
533 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
535 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
539 int is_dummy(struct cds_lfht_node
*node
)
541 return ((unsigned long) node
) & DUMMY_FLAG
;
545 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
547 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
551 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
553 unsigned long old1
, old2
;
555 old1
= uatomic_read(ptr
);
560 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
565 * Remove all logically deleted nodes from a bucket up to a certain node key.
568 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
570 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
574 /* We can always skip the dummy node initially */
575 iter
= rcu_dereference(iter_prev
->p
.next
);
576 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
578 if (unlikely(!clear_flag(iter
)))
580 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
582 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
583 if (likely(is_removed(next
)))
585 iter_prev
= clear_flag(iter
);
588 assert(!is_removed(iter
));
590 new_next
= flag_dummy(clear_flag(next
));
592 new_next
= clear_flag(next
);
593 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
598 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
, struct rcu_table
*t
,
599 struct cds_lfht_node
*node
, int unique
, int dummy
)
601 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
603 struct _cds_lfht_node
*lookup
;
604 unsigned long hash
, index
, order
;
608 node
->p
.next
= flag_dummy(NULL
);
609 return node
; /* Initial first add (head) */
611 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
613 uint32_t chain_len
= 0;
616 * iter_prev points to the non-removed node prior to the
619 index
= hash
& (t
->size
- 1);
620 order
= get_count_order_ulong(index
+ 1);
621 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
622 iter_prev
= (struct cds_lfht_node
*) lookup
;
623 /* We can always skip the dummy node initially */
624 iter
= rcu_dereference(iter_prev
->p
.next
);
625 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
627 if (unlikely(!clear_flag(iter
)))
629 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
631 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
632 if (unlikely(is_removed(next
)))
636 && !ht
->compare_fct(node
->key
, node
->key_len
,
637 clear_flag(iter
)->key
,
638 clear_flag(iter
)->key_len
))
639 return clear_flag(iter
);
640 /* Only account for identical reverse hash once */
641 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
643 check_resize(ht
, t
, ++chain_len
);
644 iter_prev
= clear_flag(iter
);
648 assert(node
!= clear_flag(iter
));
649 assert(!is_removed(iter_prev
));
650 assert(iter_prev
!= node
);
652 node
->p
.next
= clear_flag(iter
);
654 node
->p
.next
= flag_dummy(clear_flag(iter
));
656 new_node
= flag_dummy(node
);
659 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
661 continue; /* retry */
665 assert(!is_removed(iter
));
667 new_next
= flag_dummy(clear_flag(next
));
669 new_next
= clear_flag(next
);
670 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
674 /* Garbage collect logically removed nodes in the bucket */
675 index
= hash
& (t
->size
- 1);
676 order
= get_count_order_ulong(index
+ 1);
677 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
678 dummy_node
= (struct cds_lfht_node
*) lookup
;
679 _cds_lfht_gc_bucket(dummy_node
, node
);
684 int _cds_lfht_remove(struct cds_lfht
*ht
, struct rcu_table
*t
,
685 struct cds_lfht_node
*node
)
687 struct cds_lfht_node
*dummy
, *next
, *old
;
688 struct _cds_lfht_node
*lookup
;
690 unsigned long hash
, index
, order
;
692 /* logically delete the node */
693 old
= rcu_dereference(node
->p
.next
);
696 if (unlikely(is_removed(next
)))
698 assert(!is_dummy(next
));
699 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
701 } while (old
!= next
);
703 /* We performed the (logical) deletion. */
707 * Ensure that the node is not visible to readers anymore: lookup for
708 * the node, and remove it (along with any other logically removed node)
711 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
712 index
= hash
& (t
->size
- 1);
713 order
= get_count_order_ulong(index
+ 1);
714 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
715 dummy
= (struct cds_lfht_node
*) lookup
;
716 _cds_lfht_gc_bucket(dummy
, node
);
719 * Only the flagging action indicated that we (and no other)
720 * removed the node from the hash.
723 assert(is_removed(rcu_dereference(node
->p
.next
)));
730 void init_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
731 unsigned long first_order
, unsigned long len_order
)
733 unsigned long i
, end_order
;
735 dbg_printf("init table: first_order %lu end_order %lu\n",
736 first_order
, first_order
+ len_order
);
737 end_order
= first_order
+ len_order
;
738 t
->size
= !first_order
? 0 : (1UL << (first_order
- 1));
739 for (i
= first_order
; i
< end_order
; i
++) {
740 unsigned long j
, len
;
742 len
= !i
? 1 : 1UL << (i
- 1);
743 dbg_printf("init order %lu len: %lu\n", i
, len
);
744 t
->tbl
[i
] = calloc(len
, sizeof(struct _cds_lfht_node
));
745 for (j
= 0; j
< len
; j
++) {
746 dbg_printf("init entry: i %lu j %lu hash %lu\n",
747 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
748 struct cds_lfht_node
*new_node
=
749 (struct cds_lfht_node
*) &t
->tbl
[i
][j
];
750 new_node
->p
.reverse_hash
=
751 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
752 (void) _cds_lfht_add(ht
, t
, new_node
, 0, 1);
753 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
756 /* Update table size */
757 t
->size
= !i
? 1 : (1UL << i
);
758 dbg_printf("init new size: %lu\n", t
->size
);
759 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
762 t
->resize_target
= t
->size
;
763 t
->resize_initiated
= 0;
766 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
767 cds_lfht_compare_fct compare_fct
,
768 unsigned long hash_seed
,
769 unsigned long init_size
,
770 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
771 void (*func
)(struct rcu_head
*head
)))
776 /* init_size must be power of two */
777 if (init_size
&& (init_size
& (init_size
- 1)))
779 ht
= calloc(1, sizeof(struct cds_lfht
));
780 ht
->hash_fct
= hash_fct
;
781 ht
->compare_fct
= compare_fct
;
782 ht
->hash_seed
= hash_seed
;
783 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
784 ht
->in_progress_resize
= 0;
785 ht
->percpu_count
= alloc_per_cpu_items_count();
786 /* this mutex should not nest in read-side C.S. */
787 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
788 order
= get_count_order_ulong(max(init_size
, 1)) + 1;
789 ht
->t
= calloc(1, sizeof(struct cds_lfht
)
790 + (order
* sizeof(struct _cds_lfht_node
*)));
792 pthread_mutex_lock(&ht
->resize_mutex
);
793 init_table(ht
, ht
->t
, 0, order
);
794 pthread_mutex_unlock(&ht
->resize_mutex
);
798 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
801 struct cds_lfht_node
*node
, *next
;
802 struct _cds_lfht_node
*lookup
;
803 unsigned long hash
, reverse_hash
, index
, order
;
805 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
806 reverse_hash
= bit_reverse_ulong(hash
);
808 t
= rcu_dereference(ht
->t
);
809 index
= hash
& (t
->size
- 1);
810 order
= get_count_order_ulong(index
+ 1);
811 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
812 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
813 hash
, index
, order
, index
& ((1UL << (order
- 1)) - 1));
814 node
= (struct cds_lfht_node
*) lookup
;
818 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
822 next
= rcu_dereference(node
->p
.next
);
823 if (likely(!is_removed(next
))
825 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
828 node
= clear_flag(next
);
830 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
834 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
835 struct cds_lfht_node
*node
)
837 struct cds_lfht_node
*next
;
838 unsigned long reverse_hash
;
842 reverse_hash
= node
->p
.reverse_hash
;
844 key_len
= node
->key_len
;
845 next
= rcu_dereference(node
->p
.next
);
846 node
= clear_flag(next
);
851 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
855 next
= rcu_dereference(node
->p
.next
);
856 if (likely(!is_removed(next
))
858 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
861 node
= clear_flag(next
);
863 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
867 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
872 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
873 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
875 t
= rcu_dereference(ht
->t
);
876 (void) _cds_lfht_add(ht
, t
, node
, 0, 0);
880 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
881 struct cds_lfht_node
*node
)
885 struct cds_lfht_node
*ret
;
887 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
888 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
890 t
= rcu_dereference(ht
->t
);
891 ret
= _cds_lfht_add(ht
, t
, node
, 1, 0);
897 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
902 t
= rcu_dereference(ht
->t
);
903 ret
= _cds_lfht_remove(ht
, t
, node
);
905 ht_count_remove(ht
, t
);
910 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
913 struct cds_lfht_node
*node
;
914 struct _cds_lfht_node
*lookup
;
915 unsigned long order
, i
;
918 /* Check that the table is empty */
919 lookup
= &t
->tbl
[0][0];
920 node
= (struct cds_lfht_node
*) lookup
;
922 node
= clear_flag(node
)->p
.next
;
925 assert(!is_removed(node
));
926 } while (clear_flag(node
));
927 /* Internal sanity check: all nodes left should be dummy */
928 for (order
= 0; order
< get_count_order_ulong(t
->size
) + 1; order
++) {
931 len
= !order
? 1 : 1UL << (order
- 1);
932 for (i
= 0; i
< len
; i
++) {
933 dbg_printf("delete order %lu i %lu hash %lu\n",
935 bit_reverse_ulong(t
->tbl
[order
][i
].reverse_hash
));
936 assert(is_dummy(t
->tbl
[order
][i
].next
));
944 * Should only be called when no more concurrent readers nor writers can
945 * possibly access the table.
947 int cds_lfht_destroy(struct cds_lfht
*ht
)
951 /* Wait for in-flight resize operations to complete */
952 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
953 while (uatomic_read(&ht
->in_progress_resize
))
954 poll(NULL
, 0, 100); /* wait for 100ms */
955 ret
= cds_lfht_delete_dummy(ht
);
959 free_per_cpu_items_count(ht
->percpu_count
);
964 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
965 unsigned long *count
,
966 unsigned long *removed
)
969 struct cds_lfht_node
*node
, *next
;
970 struct _cds_lfht_node
*lookup
;
971 unsigned long nr_dummy
= 0;
976 t
= rcu_dereference(ht
->t
);
977 /* Count non-dummy nodes in the table */
978 lookup
= &t
->tbl
[0][0];
979 node
= (struct cds_lfht_node
*) lookup
;
981 next
= rcu_dereference(node
->p
.next
);
982 if (is_removed(next
)) {
983 assert(!is_dummy(next
));
985 } else if (!is_dummy(next
))
989 node
= clear_flag(next
);
991 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
995 void cds_lfht_free_table_cb(struct rcu_head
*head
)
997 struct rcu_table
*t
=
998 caa_container_of(head
, struct rcu_table
, head
);
1002 /* called with resize mutex held */
1004 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1006 unsigned long new_size
, old_size
, old_order
, new_order
;
1007 struct rcu_table
*new_t
, *old_t
;
1010 old_size
= old_t
->size
;
1011 old_order
= get_count_order_ulong(old_size
) + 1;
1013 new_size
= CMM_LOAD_SHARED(old_t
->resize_target
);
1014 if (old_size
== new_size
)
1016 new_order
= get_count_order_ulong(new_size
) + 1;
1017 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1018 old_size
, old_order
, new_size
, new_order
);
1019 new_t
= malloc(sizeof(struct cds_lfht
)
1020 + (new_order
* sizeof(struct _cds_lfht_node
*)));
1021 assert(new_size
> old_size
);
1022 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1023 old_order
* sizeof(struct _cds_lfht_node
*));
1024 init_table(ht
, new_t
, old_order
, new_order
- old_order
);
1025 /* Changing table and size atomically wrt lookups */
1026 rcu_assign_pointer(ht
->t
, new_t
);
1027 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1031 unsigned long resize_target_update(struct rcu_table
*t
,
1034 return _uatomic_max(&t
->resize_target
,
1035 t
->size
<< growth_order
);
1038 void cds_lfht_resize(struct cds_lfht
*ht
, int growth
)
1040 struct rcu_table
*t
= rcu_dereference(ht
->t
);
1041 unsigned long target_size
;
1045 * Silently refuse to shrink hash table. (not supported)
1047 dbg_printf("shrinking hash table not supported.\n");
1051 target_size
= resize_target_update(t
, growth
);
1052 if (t
->size
< target_size
) {
1053 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1054 pthread_mutex_lock(&ht
->resize_mutex
);
1055 _do_cds_lfht_resize(ht
);
1056 pthread_mutex_unlock(&ht
->resize_mutex
);
1061 void do_resize_cb(struct rcu_head
*head
)
1063 struct rcu_resize_work
*work
=
1064 caa_container_of(head
, struct rcu_resize_work
, head
);
1065 struct cds_lfht
*ht
= work
->ht
;
1067 pthread_mutex_lock(&ht
->resize_mutex
);
1068 _do_cds_lfht_resize(ht
);
1069 pthread_mutex_unlock(&ht
->resize_mutex
);
1071 cmm_smp_mb(); /* finish resize before decrement */
1072 uatomic_dec(&ht
->in_progress_resize
);
1076 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
)
1078 struct rcu_resize_work
*work
;
1079 unsigned long target_size
;
1081 target_size
= resize_target_update(t
, growth
);
1082 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
1083 uatomic_inc(&ht
->in_progress_resize
);
1084 cmm_smp_mb(); /* increment resize count before calling it */
1085 work
= malloc(sizeof(*work
));
1087 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1088 CMM_STORE_SHARED(t
->resize_initiated
, 1);