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...)
123 * Per-CPU split-counters lazily update the global counter each 1024
124 * addition/removal. It automatically keeps track of resize required.
125 * We use the bucket length as indicator for need to expand for small
126 * tables and machines lacking per-cpu data suppport.
128 #define COUNT_COMMIT_ORDER 10
129 #define CHAIN_LEN_TARGET 1
130 #define CHAIN_LEN_RESIZE_THRESHOLD 3
133 #define max(a, b) ((a) > (b) ? (a) : (b))
137 * The removed flag needs to be updated atomically with the pointer.
138 * The dummy flag does not require to be updated atomically with the
139 * pointer, but it is added as a pointer low bit flag to save space.
141 #define REMOVED_FLAG (1UL << 0)
142 #define DUMMY_FLAG (1UL << 1)
143 #define FLAGS_MASK ((1UL << 2) - 1)
145 struct ht_items_count
{
146 unsigned long add
, remove
;
147 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
150 unsigned long size
; /* always a power of 2 */
151 unsigned long resize_target
;
152 int resize_initiated
;
153 struct rcu_head head
;
154 struct _cds_lfht_node
*tbl
[0];
158 struct rcu_table
*t
; /* shared */
159 cds_lfht_hash_fct hash_fct
;
160 cds_lfht_compare_fct compare_fct
;
161 unsigned long hash_seed
;
162 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
163 unsigned int in_progress_resize
, in_progress_destroy
;
164 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
165 void (*func
)(struct rcu_head
*head
));
166 unsigned long count
; /* global approximate item count */
167 struct ht_items_count
*percpu_count
; /* per-cpu item count */
170 struct rcu_resize_work
{
171 struct rcu_head head
;
176 * Algorithm to reverse bits in a word by lookup table, extended to
179 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
180 * Originally from Public Domain.
183 static const uint8_t BitReverseTable256
[256] =
185 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
186 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
187 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
188 R6(0), R6(2), R6(1), R6(3)
195 uint8_t bit_reverse_u8(uint8_t v
)
197 return BitReverseTable256
[v
];
200 static __attribute__((unused
))
201 uint32_t bit_reverse_u32(uint32_t v
)
203 return ((uint32_t) bit_reverse_u8(v
) << 24) |
204 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
205 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
206 ((uint32_t) bit_reverse_u8(v
>> 24));
209 static __attribute__((unused
))
210 uint64_t bit_reverse_u64(uint64_t v
)
212 return ((uint64_t) bit_reverse_u8(v
) << 56) |
213 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
214 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
215 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
216 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
217 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
218 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
219 ((uint64_t) bit_reverse_u8(v
>> 56));
223 unsigned long bit_reverse_ulong(unsigned long v
)
225 #if (CAA_BITS_PER_LONG == 32)
226 return bit_reverse_u32(v
);
228 return bit_reverse_u64(v
);
233 * fls: returns the position of the most significant bit.
234 * Returns 0 if no bit is set, else returns the position of the most
235 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
237 #if defined(__i386) || defined(__x86_64)
239 unsigned int fls_u32(uint32_t x
)
247 : "=r" (r
) : "rm" (x
));
253 #if defined(__x86_64)
255 unsigned int fls_u64(uint64_t x
)
263 : "=r" (r
) : "rm" (x
));
270 static __attribute__((unused
))
271 unsigned int fls_u64(uint64_t x
)
278 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
282 if (!(x
& 0xFFFF000000000000ULL
)) {
286 if (!(x
& 0xFF00000000000000ULL
)) {
290 if (!(x
& 0xF000000000000000ULL
)) {
294 if (!(x
& 0xC000000000000000ULL
)) {
298 if (!(x
& 0x8000000000000000ULL
)) {
307 static __attribute__((unused
))
308 unsigned int fls_u32(uint32_t x
)
314 if (!(x
& 0xFFFF0000U
)) {
318 if (!(x
& 0xFF000000U
)) {
322 if (!(x
& 0xF0000000U
)) {
326 if (!(x
& 0xC0000000U
)) {
330 if (!(x
& 0x80000000U
)) {
338 unsigned int fls_ulong(unsigned long x
)
340 #if (CAA_BITS_PER_lONG == 32)
347 int get_count_order_u32(uint32_t x
)
351 order
= fls_u32(x
) - 1;
357 int get_count_order_ulong(unsigned long x
)
361 order
= fls_ulong(x
) - 1;
368 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
);
371 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
372 * available, then we support hash table item accounting.
373 * In the unfortunate event the number of CPUs reported would be
374 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
376 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
379 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
380 unsigned long count
);
382 static long nr_cpus_mask
= -1;
385 struct ht_items_count
*alloc_per_cpu_items_count(void)
387 struct ht_items_count
*count
;
389 switch (nr_cpus_mask
) {
396 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
402 * round up number of CPUs to next power of two, so we
403 * can use & for modulo.
405 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
406 nr_cpus_mask
= maxcpus
- 1;
410 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
415 void free_per_cpu_items_count(struct ht_items_count
*count
)
425 assert(nr_cpus_mask
>= 0);
426 cpu
= sched_getcpu();
427 if (unlikely(cpu
< 0))
430 return cpu
& nr_cpus_mask
;
434 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
436 unsigned long percpu_count
;
439 if (unlikely(!ht
->percpu_count
))
442 if (unlikely(cpu
< 0))
444 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 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 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
456 dbg_printf("add set global %lu\n", count
);
457 cds_lfht_resize_lazy_count(ht
, t
,
458 count
>> (CHAIN_LEN_TARGET
- 1));
464 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
466 unsigned long percpu_count
;
469 if (unlikely(!ht
->percpu_count
))
472 if (unlikely(cpu
< 0))
474 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
475 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
478 dbg_printf("remove percpu %lu\n", percpu_count
);
479 count
= uatomic_add_return(&ht
->count
,
480 -(1UL << COUNT_COMMIT_ORDER
));
482 if (!(count
& (count
- 1))) {
483 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
486 dbg_printf("remove set global %lu\n", count
);
487 cds_lfht_resize_lazy_count(ht
, t
,
488 count
>> (CHAIN_LEN_TARGET
- 1));
493 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
495 static const long nr_cpus_mask
= -1;
498 struct ht_items_count
*alloc_per_cpu_items_count(void)
504 void free_per_cpu_items_count(struct ht_items_count
*count
)
509 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
514 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
518 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
522 void check_resize(struct cds_lfht
*ht
, struct rcu_table
*t
,
527 count
= uatomic_read(&ht
->count
);
529 * Use bucket-local length for small table expand and for
530 * environments lacking per-cpu data support.
532 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
535 dbg_printf("WARNING: large chain length: %u.\n",
537 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
538 cds_lfht_resize_lazy(ht
, t
,
539 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
543 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
545 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
549 int is_removed(struct cds_lfht_node
*node
)
551 return ((unsigned long) node
) & REMOVED_FLAG
;
555 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
557 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
561 int is_dummy(struct cds_lfht_node
*node
)
563 return ((unsigned long) node
) & DUMMY_FLAG
;
567 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
569 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
573 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
575 unsigned long old1
, old2
;
577 old1
= uatomic_read(ptr
);
582 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
587 * Remove all logically deleted nodes from a bucket up to a certain node key.
590 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
592 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
596 /* We can always skip the dummy node initially */
597 iter
= rcu_dereference(iter_prev
->p
.next
);
598 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
600 if (unlikely(!clear_flag(iter
)))
602 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
604 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
605 if (likely(is_removed(next
)))
607 iter_prev
= clear_flag(iter
);
610 assert(!is_removed(iter
));
612 new_next
= flag_dummy(clear_flag(next
));
614 new_next
= clear_flag(next
);
615 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
620 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
, struct rcu_table
*t
,
621 struct cds_lfht_node
*node
, int unique
, int dummy
)
623 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
625 struct _cds_lfht_node
*lookup
;
626 unsigned long hash
, index
, order
;
630 node
->p
.next
= flag_dummy(NULL
);
631 return node
; /* Initial first add (head) */
633 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
635 uint32_t chain_len
= 0;
638 * iter_prev points to the non-removed node prior to the
641 index
= hash
& (t
->size
- 1);
642 order
= get_count_order_ulong(index
+ 1);
643 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
644 iter_prev
= (struct cds_lfht_node
*) lookup
;
645 /* We can always skip the dummy node initially */
646 iter
= rcu_dereference(iter_prev
->p
.next
);
647 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
649 if (unlikely(!clear_flag(iter
)))
651 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
653 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
654 if (unlikely(is_removed(next
)))
658 && !ht
->compare_fct(node
->key
, node
->key_len
,
659 clear_flag(iter
)->key
,
660 clear_flag(iter
)->key_len
))
661 return clear_flag(iter
);
662 /* Only account for identical reverse hash once */
663 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
665 check_resize(ht
, t
, ++chain_len
);
666 iter_prev
= clear_flag(iter
);
670 assert(node
!= clear_flag(iter
));
671 assert(!is_removed(iter_prev
));
672 assert(iter_prev
!= node
);
674 node
->p
.next
= clear_flag(iter
);
676 node
->p
.next
= flag_dummy(clear_flag(iter
));
678 new_node
= flag_dummy(node
);
681 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
683 continue; /* retry */
687 assert(!is_removed(iter
));
689 new_next
= flag_dummy(clear_flag(next
));
691 new_next
= clear_flag(next
);
692 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
696 /* Garbage collect logically removed nodes in the bucket */
697 index
= hash
& (t
->size
- 1);
698 order
= get_count_order_ulong(index
+ 1);
699 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
700 dummy_node
= (struct cds_lfht_node
*) lookup
;
701 _cds_lfht_gc_bucket(dummy_node
, node
);
706 int _cds_lfht_remove(struct cds_lfht
*ht
, struct rcu_table
*t
,
707 struct cds_lfht_node
*node
)
709 struct cds_lfht_node
*dummy
, *next
, *old
;
710 struct _cds_lfht_node
*lookup
;
712 unsigned long hash
, index
, order
;
714 /* logically delete the node */
715 old
= rcu_dereference(node
->p
.next
);
718 if (unlikely(is_removed(next
)))
720 assert(!is_dummy(next
));
721 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
723 } while (old
!= next
);
725 /* We performed the (logical) deletion. */
729 * Ensure that the node is not visible to readers anymore: lookup for
730 * the node, and remove it (along with any other logically removed node)
733 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
734 index
= hash
& (t
->size
- 1);
735 order
= get_count_order_ulong(index
+ 1);
736 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
737 dummy
= (struct cds_lfht_node
*) lookup
;
738 _cds_lfht_gc_bucket(dummy
, node
);
741 * Only the flagging action indicated that we (and no other)
742 * removed the node from the hash.
745 assert(is_removed(rcu_dereference(node
->p
.next
)));
752 void init_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
753 unsigned long first_order
, unsigned long len_order
)
755 unsigned long i
, end_order
;
757 dbg_printf("init table: first_order %lu end_order %lu\n",
758 first_order
, first_order
+ len_order
);
759 end_order
= first_order
+ len_order
;
760 t
->size
= !first_order
? 0 : (1UL << (first_order
- 1));
761 for (i
= first_order
; i
< end_order
; i
++) {
762 unsigned long j
, len
;
764 len
= !i
? 1 : 1UL << (i
- 1);
765 dbg_printf("init order %lu len: %lu\n", i
, len
);
766 t
->tbl
[i
] = calloc(len
, sizeof(struct _cds_lfht_node
));
767 for (j
= 0; j
< len
; j
++) {
768 dbg_printf("init entry: i %lu j %lu hash %lu\n",
769 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
770 struct cds_lfht_node
*new_node
=
771 (struct cds_lfht_node
*) &t
->tbl
[i
][j
];
772 new_node
->p
.reverse_hash
=
773 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
774 (void) _cds_lfht_add(ht
, t
, new_node
, 0, 1);
775 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
778 /* Update table size */
779 t
->size
= !i
? 1 : (1UL << i
);
780 dbg_printf("init new size: %lu\n", t
->size
);
781 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
784 t
->resize_target
= t
->size
;
785 t
->resize_initiated
= 0;
788 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
789 cds_lfht_compare_fct compare_fct
,
790 unsigned long hash_seed
,
791 unsigned long init_size
,
792 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
793 void (*func
)(struct rcu_head
*head
)))
798 /* init_size must be power of two */
799 if (init_size
&& (init_size
& (init_size
- 1)))
801 ht
= calloc(1, sizeof(struct cds_lfht
));
802 ht
->hash_fct
= hash_fct
;
803 ht
->compare_fct
= compare_fct
;
804 ht
->hash_seed
= hash_seed
;
805 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
806 ht
->in_progress_resize
= 0;
807 ht
->percpu_count
= alloc_per_cpu_items_count();
808 /* this mutex should not nest in read-side C.S. */
809 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
810 order
= get_count_order_ulong(max(init_size
, 1)) + 1;
811 ht
->t
= calloc(1, sizeof(struct cds_lfht
)
812 + (order
* sizeof(struct _cds_lfht_node
*)));
814 pthread_mutex_lock(&ht
->resize_mutex
);
815 init_table(ht
, ht
->t
, 0, order
);
816 pthread_mutex_unlock(&ht
->resize_mutex
);
820 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
823 struct cds_lfht_node
*node
, *next
;
824 struct _cds_lfht_node
*lookup
;
825 unsigned long hash
, reverse_hash
, index
, order
;
827 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
828 reverse_hash
= bit_reverse_ulong(hash
);
830 t
= rcu_dereference(ht
->t
);
831 index
= hash
& (t
->size
- 1);
832 order
= get_count_order_ulong(index
+ 1);
833 lookup
= &t
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
834 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
835 hash
, index
, order
, index
& ((1UL << (order
- 1)) - 1));
836 node
= (struct cds_lfht_node
*) lookup
;
840 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
844 next
= rcu_dereference(node
->p
.next
);
845 if (likely(!is_removed(next
))
847 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
850 node
= clear_flag(next
);
852 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
856 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
857 struct cds_lfht_node
*node
)
859 struct cds_lfht_node
*next
;
860 unsigned long reverse_hash
;
864 reverse_hash
= node
->p
.reverse_hash
;
866 key_len
= node
->key_len
;
867 next
= rcu_dereference(node
->p
.next
);
868 node
= clear_flag(next
);
873 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
877 next
= rcu_dereference(node
->p
.next
);
878 if (likely(!is_removed(next
))
880 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
883 node
= clear_flag(next
);
885 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
889 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
894 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
895 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
897 t
= rcu_dereference(ht
->t
);
898 (void) _cds_lfht_add(ht
, t
, node
, 0, 0);
902 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
903 struct cds_lfht_node
*node
)
907 struct cds_lfht_node
*ret
;
909 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
910 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
912 t
= rcu_dereference(ht
->t
);
913 ret
= _cds_lfht_add(ht
, t
, node
, 1, 0);
919 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
924 t
= rcu_dereference(ht
->t
);
925 ret
= _cds_lfht_remove(ht
, t
, node
);
927 ht_count_remove(ht
, t
);
932 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
935 struct cds_lfht_node
*node
;
936 struct _cds_lfht_node
*lookup
;
937 unsigned long order
, i
;
940 /* Check that the table is empty */
941 lookup
= &t
->tbl
[0][0];
942 node
= (struct cds_lfht_node
*) lookup
;
944 node
= clear_flag(node
)->p
.next
;
947 assert(!is_removed(node
));
948 } while (clear_flag(node
));
949 /* Internal sanity check: all nodes left should be dummy */
950 for (order
= 0; order
< get_count_order_ulong(t
->size
) + 1; order
++) {
953 len
= !order
? 1 : 1UL << (order
- 1);
954 for (i
= 0; i
< len
; i
++) {
955 dbg_printf("delete order %lu i %lu hash %lu\n",
957 bit_reverse_ulong(t
->tbl
[order
][i
].reverse_hash
));
958 assert(is_dummy(t
->tbl
[order
][i
].next
));
966 * Should only be called when no more concurrent readers nor writers can
967 * possibly access the table.
969 int cds_lfht_destroy(struct cds_lfht
*ht
)
973 /* Wait for in-flight resize operations to complete */
974 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
975 while (uatomic_read(&ht
->in_progress_resize
))
976 poll(NULL
, 0, 100); /* wait for 100ms */
977 ret
= cds_lfht_delete_dummy(ht
);
981 free_per_cpu_items_count(ht
->percpu_count
);
986 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
987 unsigned long *count
,
988 unsigned long *removed
)
991 struct cds_lfht_node
*node
, *next
;
992 struct _cds_lfht_node
*lookup
;
993 unsigned long nr_dummy
= 0;
998 t
= rcu_dereference(ht
->t
);
999 /* Count non-dummy nodes in the table */
1000 lookup
= &t
->tbl
[0][0];
1001 node
= (struct cds_lfht_node
*) lookup
;
1003 next
= rcu_dereference(node
->p
.next
);
1004 if (is_removed(next
)) {
1005 assert(!is_dummy(next
));
1007 } else if (!is_dummy(next
))
1011 node
= clear_flag(next
);
1013 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1017 void cds_lfht_free_table_cb(struct rcu_head
*head
)
1019 struct rcu_table
*t
=
1020 caa_container_of(head
, struct rcu_table
, head
);
1024 /* called with resize mutex held */
1026 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1028 unsigned long new_size
, old_size
, old_order
, new_order
;
1029 struct rcu_table
*new_t
, *old_t
;
1032 old_size
= old_t
->size
;
1033 old_order
= get_count_order_ulong(old_size
) + 1;
1035 new_size
= CMM_LOAD_SHARED(old_t
->resize_target
);
1036 if (old_size
== new_size
)
1038 new_order
= get_count_order_ulong(new_size
) + 1;
1039 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1040 old_size
, old_order
, new_size
, new_order
);
1041 new_t
= malloc(sizeof(struct cds_lfht
)
1042 + (new_order
* sizeof(struct _cds_lfht_node
*)));
1043 assert(new_size
> old_size
);
1044 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1045 old_order
* sizeof(struct _cds_lfht_node
*));
1046 init_table(ht
, new_t
, old_order
, new_order
- old_order
);
1047 /* Changing table and size atomically wrt lookups */
1048 rcu_assign_pointer(ht
->t
, new_t
);
1049 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1053 unsigned long resize_target_update(struct rcu_table
*t
,
1056 return _uatomic_max(&t
->resize_target
,
1057 t
->size
<< growth_order
);
1060 void cds_lfht_resize(struct cds_lfht
*ht
, int growth
)
1062 struct rcu_table
*t
= rcu_dereference(ht
->t
);
1063 unsigned long target_size
;
1067 * Silently refuse to shrink hash table. (not supported)
1069 dbg_printf("shrinking hash table not supported.\n");
1073 target_size
= resize_target_update(t
, growth
);
1074 if (t
->size
< target_size
) {
1075 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1076 pthread_mutex_lock(&ht
->resize_mutex
);
1077 _do_cds_lfht_resize(ht
);
1078 pthread_mutex_unlock(&ht
->resize_mutex
);
1083 void do_resize_cb(struct rcu_head
*head
)
1085 struct rcu_resize_work
*work
=
1086 caa_container_of(head
, struct rcu_resize_work
, head
);
1087 struct cds_lfht
*ht
= work
->ht
;
1089 pthread_mutex_lock(&ht
->resize_mutex
);
1090 _do_cds_lfht_resize(ht
);
1091 pthread_mutex_unlock(&ht
->resize_mutex
);
1093 cmm_smp_mb(); /* finish resize before decrement */
1094 uatomic_dec(&ht
->in_progress_resize
);
1098 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
)
1100 struct rcu_resize_work
*work
;
1101 unsigned long target_size
;
1103 target_size
= resize_target_update(t
, growth
);
1104 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
1105 uatomic_inc(&ht
->in_progress_resize
);
1106 cmm_smp_mb(); /* increment resize count before calling it */
1107 work
= malloc(sizeof(*work
));
1109 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1110 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1114 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1117 unsigned long resize_target_update_count(struct rcu_table
*t
,
1118 unsigned long count
)
1120 return uatomic_set(&t
->resize_target
, count
);
1124 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
1125 unsigned long count
)
1127 struct rcu_resize_work
*work
;
1128 unsigned long target_size
;
1130 target_size
= resize_target_update_count(t
, count
);
1131 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
1132 uatomic_inc(&ht
->in_progress_resize
);
1133 cmm_smp_mb(); /* increment resize count before calling it */
1134 work
= malloc(sizeof(*work
));
1136 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1137 CMM_STORE_SHARED(t
->resize_initiated
, 1);