4 * Userspace RCU library - Lock-Free Resizable 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 Resizable 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 for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
100 * A bit of ascii art explanation:
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
105 * This shows the nodes for a small table ordered by reversed bits:
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
123 * 1 | 1 001 100 <- <-
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
128 * 3 -> | | | 4 100 001 | |
144 #include <urcu-call-rcu.h>
145 #include <urcu/arch.h>
146 #include <urcu/uatomic.h>
147 #include <urcu/jhash.h>
148 #include <urcu/compiler.h>
149 #include <urcu/rculfhash.h>
154 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
156 #define dbg_printf(fmt, args...)
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
165 #define COUNT_COMMIT_ORDER 10
166 #define CHAIN_LEN_TARGET 1
167 #define CHAIN_LEN_RESIZE_THRESHOLD 3
170 * Define the minimum table size.
172 #define MIN_TABLE_SIZE 1
174 #if (CAA_BITS_PER_LONG == 32)
175 #define MAX_TABLE_ORDER 32
177 #define MAX_TABLE_ORDER 64
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
183 #define MIN_PARTITION_PER_THREAD_ORDER 12
184 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
187 #define min(a, b) ((a) < (b) ? (a) : (b))
191 #define max(a, b) ((a) > (b) ? (a) : (b))
195 * The removed flag needs to be updated atomically with the pointer.
196 * It indicates that no node must attach to the node scheduled for
197 * removal, and that node garbage collection must be performed.
198 * The dummy flag does not require to be updated atomically with the
199 * pointer, but it is added as a pointer low bit flag to save space.
201 #define REMOVED_FLAG (1UL << 0)
202 #define DUMMY_FLAG (1UL << 1)
203 #define FLAGS_MASK ((1UL << 2) - 1)
205 /* Value of the end pointer. Should not interact with flags. */
206 #define END_VALUE NULL
208 struct ht_items_count
{
209 unsigned long add
, del
;
210 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
213 struct rcu_head head
;
214 struct _cds_lfht_node nodes
[0];
218 unsigned long size
; /* always a power of 2, shared (RCU) */
219 unsigned long resize_target
;
220 int resize_initiated
;
221 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
226 cds_lfht_hash_fct hash_fct
;
227 cds_lfht_compare_fct compare_fct
;
228 unsigned long hash_seed
;
231 * We need to put the work threads offline (QSBR) when taking this
232 * mutex, because we use synchronize_rcu within this mutex critical
233 * section, which waits on read-side critical sections, and could
234 * therefore cause grace-period deadlock if we hold off RCU G.P.
237 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
238 unsigned int in_progress_resize
, in_progress_destroy
;
239 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
240 void (*func
)(struct rcu_head
*head
));
241 void (*cds_lfht_synchronize_rcu
)(void);
242 void (*cds_lfht_rcu_read_lock
)(void);
243 void (*cds_lfht_rcu_read_unlock
)(void);
244 void (*cds_lfht_rcu_thread_offline
)(void);
245 void (*cds_lfht_rcu_thread_online
)(void);
246 void (*cds_lfht_rcu_register_thread
)(void);
247 void (*cds_lfht_rcu_unregister_thread
)(void);
248 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
249 long count
; /* global approximate item count */
250 struct ht_items_count
*percpu_count
; /* per-cpu item count */
253 struct rcu_resize_work
{
254 struct rcu_head head
;
258 struct partition_resize_work
{
259 struct rcu_head head
;
261 unsigned long i
, start
, len
;
262 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
263 unsigned long start
, unsigned long len
);
273 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
275 struct cds_lfht_node
*node
,
276 enum add_mode mode
, int dummy
);
279 * Algorithm to reverse bits in a word by lookup table, extended to
282 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
283 * Originally from Public Domain.
286 static const uint8_t BitReverseTable256
[256] =
288 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
289 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
290 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
291 R6(0), R6(2), R6(1), R6(3)
298 uint8_t bit_reverse_u8(uint8_t v
)
300 return BitReverseTable256
[v
];
303 static __attribute__((unused
))
304 uint32_t bit_reverse_u32(uint32_t v
)
306 return ((uint32_t) bit_reverse_u8(v
) << 24) |
307 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
308 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
309 ((uint32_t) bit_reverse_u8(v
>> 24));
312 static __attribute__((unused
))
313 uint64_t bit_reverse_u64(uint64_t v
)
315 return ((uint64_t) bit_reverse_u8(v
) << 56) |
316 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
317 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
318 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
319 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
320 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
321 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
322 ((uint64_t) bit_reverse_u8(v
>> 56));
326 unsigned long bit_reverse_ulong(unsigned long v
)
328 #if (CAA_BITS_PER_LONG == 32)
329 return bit_reverse_u32(v
);
331 return bit_reverse_u64(v
);
336 * fls: returns the position of the most significant bit.
337 * Returns 0 if no bit is set, else returns the position of the most
338 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
340 #if defined(__i386) || defined(__x86_64)
342 unsigned int fls_u32(uint32_t x
)
350 : "=r" (r
) : "rm" (x
));
356 #if defined(__x86_64)
358 unsigned int fls_u64(uint64_t x
)
366 : "=r" (r
) : "rm" (x
));
373 static __attribute__((unused
))
374 unsigned int fls_u64(uint64_t x
)
381 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
385 if (!(x
& 0xFFFF000000000000ULL
)) {
389 if (!(x
& 0xFF00000000000000ULL
)) {
393 if (!(x
& 0xF000000000000000ULL
)) {
397 if (!(x
& 0xC000000000000000ULL
)) {
401 if (!(x
& 0x8000000000000000ULL
)) {
410 static __attribute__((unused
))
411 unsigned int fls_u32(uint32_t x
)
417 if (!(x
& 0xFFFF0000U
)) {
421 if (!(x
& 0xFF000000U
)) {
425 if (!(x
& 0xF0000000U
)) {
429 if (!(x
& 0xC0000000U
)) {
433 if (!(x
& 0x80000000U
)) {
441 unsigned int fls_ulong(unsigned long x
)
443 #if (CAA_BITS_PER_lONG == 32)
450 int get_count_order_u32(uint32_t x
)
454 order
= fls_u32(x
) - 1;
460 int get_count_order_ulong(unsigned long x
)
464 order
= fls_ulong(x
) - 1;
471 #define poison_free(ptr) \
473 memset(ptr, 0x42, sizeof(*(ptr))); \
477 #define poison_free(ptr) free(ptr)
481 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
484 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
485 * available, then we support hash table item accounting.
486 * In the unfortunate event the number of CPUs reported would be
487 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
489 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
492 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
493 unsigned long count
);
495 static long nr_cpus_mask
= -1;
498 struct ht_items_count
*alloc_per_cpu_items_count(void)
500 struct ht_items_count
*count
;
502 switch (nr_cpus_mask
) {
509 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
515 * round up number of CPUs to next power of two, so we
516 * can use & for modulo.
518 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
519 nr_cpus_mask
= maxcpus
- 1;
523 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
528 void free_per_cpu_items_count(struct ht_items_count
*count
)
538 assert(nr_cpus_mask
>= 0);
539 cpu
= sched_getcpu();
540 if (unlikely(cpu
< 0))
543 return cpu
& nr_cpus_mask
;
547 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
549 unsigned long percpu_count
;
552 if (unlikely(!ht
->percpu_count
))
555 if (unlikely(cpu
< 0))
557 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
558 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
561 dbg_printf("add percpu %lu\n", percpu_count
);
562 count
= uatomic_add_return(&ht
->count
,
563 1UL << COUNT_COMMIT_ORDER
);
565 if (!(count
& (count
- 1))) {
566 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
568 dbg_printf("add set global %ld\n", count
);
570 * Don't resize table if the number of nodes is below a
573 if (count
< (1UL << COUNT_COMMIT_ORDER
))
575 cds_lfht_resize_lazy_count(ht
, size
,
576 count
>> (CHAIN_LEN_TARGET
- 1));
582 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
584 unsigned long percpu_count
;
587 if (unlikely(!ht
->percpu_count
))
590 if (unlikely(cpu
< 0))
592 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
593 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
596 dbg_printf("del percpu %lu\n", percpu_count
);
597 count
= uatomic_add_return(&ht
->count
,
598 -(1UL << COUNT_COMMIT_ORDER
));
600 if (!(count
& (count
- 1))) {
601 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
603 dbg_printf("del set global %ld\n", count
);
605 * Don't resize table if the number of nodes is below a
608 if (count
< (1UL << COUNT_COMMIT_ORDER
))
610 cds_lfht_resize_lazy_count(ht
, size
,
611 count
>> (CHAIN_LEN_TARGET
- 1));
616 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
618 static const long nr_cpus_mask
= -1;
621 struct ht_items_count
*alloc_per_cpu_items_count(void)
627 void free_per_cpu_items_count(struct ht_items_count
*count
)
632 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
637 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
641 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
645 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
649 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
651 count
= uatomic_read(&ht
->count
);
653 * Use bucket-local length for small table expand and for
654 * environments lacking per-cpu data support.
656 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
659 dbg_printf("WARNING: large chain length: %u.\n",
661 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
662 cds_lfht_resize_lazy(ht
, size
,
663 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
667 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
669 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
673 int is_removed(struct cds_lfht_node
*node
)
675 return ((unsigned long) node
) & REMOVED_FLAG
;
679 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
681 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
685 int is_dummy(struct cds_lfht_node
*node
)
687 return ((unsigned long) node
) & DUMMY_FLAG
;
691 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
693 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
697 struct cds_lfht_node
*get_end(void)
699 return (struct cds_lfht_node
*) END_VALUE
;
703 int is_end(struct cds_lfht_node
*node
)
705 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
709 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
711 unsigned long old1
, old2
;
713 old1
= uatomic_read(ptr
);
718 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
723 void cds_lfht_free_level(struct rcu_head
*head
)
725 struct rcu_level
*l
=
726 caa_container_of(head
, struct rcu_level
, head
);
731 * Remove all logically deleted nodes from a bucket up to a certain node key.
734 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
736 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
738 assert(!is_dummy(dummy
));
739 assert(!is_removed(dummy
));
740 assert(!is_dummy(node
));
741 assert(!is_removed(node
));
744 /* We can always skip the dummy node initially */
745 iter
= rcu_dereference(iter_prev
->p
.next
);
746 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
748 * We should never be called with dummy (start of chain)
749 * and logically removed node (end of path compression
750 * marker) being the actual same node. This would be a
751 * bug in the algorithm implementation.
753 assert(dummy
!= node
);
755 if (unlikely(is_end(iter
)))
757 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
759 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
760 if (likely(is_removed(next
)))
762 iter_prev
= clear_flag(iter
);
765 assert(!is_removed(iter
));
767 new_next
= flag_dummy(clear_flag(next
));
769 new_next
= clear_flag(next
);
770 if (is_removed(iter
))
771 new_next
= flag_removed(new_next
);
772 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
778 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
780 struct cds_lfht_node
*node
,
781 enum add_mode mode
, int dummy
)
783 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
784 *dummy_node
, *return_node
;
785 struct _cds_lfht_node
*lookup
;
786 unsigned long hash
, index
, order
;
788 assert(!is_dummy(node
));
789 assert(!is_removed(node
));
792 node
->p
.next
= flag_dummy(get_end());
793 return node
; /* Initial first add (head) */
795 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
797 uint32_t chain_len
= 0;
800 * iter_prev points to the non-removed node prior to the
803 index
= hash
& (size
- 1);
804 order
= get_count_order_ulong(index
+ 1);
805 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
806 iter_prev
= (struct cds_lfht_node
*) lookup
;
807 /* We can always skip the dummy node initially */
808 iter
= rcu_dereference(iter_prev
->p
.next
);
809 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
811 if (unlikely(is_end(iter
)))
813 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
815 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
816 if (unlikely(is_removed(next
)))
818 if ((mode
== ADD_UNIQUE
|| mode
== ADD_REPLACE
)
820 && !ht
->compare_fct(node
->key
, node
->key_len
,
821 clear_flag(iter
)->key
,
822 clear_flag(iter
)->key_len
)) {
823 if (mode
== ADD_UNIQUE
)
824 return clear_flag(iter
);
825 else /* mode == ADD_REPLACE */
828 /* Only account for identical reverse hash once */
829 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
831 check_resize(ht
, size
, ++chain_len
);
832 iter_prev
= clear_flag(iter
);
837 assert(node
!= clear_flag(iter
));
838 assert(!is_removed(iter_prev
));
839 assert(!is_removed(iter
));
840 assert(iter_prev
!= node
);
842 node
->p
.next
= clear_flag(iter
);
844 node
->p
.next
= flag_dummy(clear_flag(iter
));
846 new_node
= flag_dummy(node
);
849 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
851 continue; /* retry */
853 if (mode
== ADD_REPLACE
)
855 else /* ADD_DEFAULT and ADD_UNIQUE */
861 /* Insert after node to be replaced */
862 iter_prev
= clear_flag(iter
);
864 assert(node
!= clear_flag(iter
));
865 assert(!is_removed(iter_prev
));
866 assert(!is_removed(iter
));
867 assert(iter_prev
!= node
);
869 node
->p
.next
= clear_flag(iter
);
871 new_node
= flag_dummy(node
);
875 * Here is the whole trick for lock-free replace: we add
876 * the replacement node _after_ the node we want to
877 * replace by atomically setting its next pointer at the
878 * same time we set its removal flag. Given that
879 * the lookups/get next use an iterator aware of the
880 * next pointer, they will either skip the old node due
881 * to the removal flag and see the new node, or use
882 * the old node, but will not see the new one.
884 new_node
= flag_removed(new_node
);
885 if (uatomic_cmpxchg(&iter_prev
->p
.next
,
886 iter
, new_node
) != iter
) {
887 continue; /* retry */
889 return_node
= iter_prev
;
894 assert(!is_removed(iter
));
896 new_next
= flag_dummy(clear_flag(next
));
898 new_next
= clear_flag(next
);
899 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
903 /* Garbage collect logically removed nodes in the bucket */
904 index
= hash
& (size
- 1);
905 order
= get_count_order_ulong(index
+ 1);
906 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
907 dummy_node
= (struct cds_lfht_node
*) lookup
;
908 _cds_lfht_gc_bucket(dummy_node
, node
);
913 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
914 struct cds_lfht_node
*node
,
917 struct cds_lfht_node
*dummy
, *next
, *old
;
918 struct _cds_lfht_node
*lookup
;
920 unsigned long hash
, index
, order
;
922 /* logically delete the node */
923 assert(!is_dummy(node
));
924 assert(!is_removed(node
));
925 old
= rcu_dereference(node
->p
.next
);
927 struct cds_lfht_node
*new_next
;
930 if (unlikely(is_removed(next
)))
933 assert(is_dummy(next
));
935 assert(!is_dummy(next
));
936 new_next
= flag_removed(next
);
937 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
938 } while (old
!= next
);
940 /* We performed the (logical) deletion. */
944 * Ensure that the node is not visible to readers anymore: lookup for
945 * the node, and remove it (along with any other logically removed node)
948 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
950 index
= hash
& (size
- 1);
951 order
= get_count_order_ulong(index
+ 1);
952 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
953 dummy
= (struct cds_lfht_node
*) lookup
;
954 _cds_lfht_gc_bucket(dummy
, node
);
957 * Only the flagging action indicated that we (and no other)
958 * removed the node from the hash.
961 assert(is_removed(rcu_dereference(node
->p
.next
)));
968 void *partition_resize_thread(void *arg
)
970 struct partition_resize_work
*work
= arg
;
972 work
->ht
->cds_lfht_rcu_register_thread();
973 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
974 work
->ht
->cds_lfht_rcu_unregister_thread();
979 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
981 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
982 unsigned long start
, unsigned long len
))
984 unsigned long partition_len
;
985 struct partition_resize_work
*work
;
987 unsigned long nr_threads
;
988 pthread_t
*thread_id
;
991 * Note: nr_cpus_mask + 1 is always power of 2.
992 * We spawn just the number of threads we need to satisfy the minimum
993 * partition size, up to the number of CPUs in the system.
995 nr_threads
= min(nr_cpus_mask
+ 1,
996 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
997 partition_len
= len
>> get_count_order_ulong(nr_threads
);
998 work
= calloc(nr_threads
, sizeof(*work
));
999 thread_id
= calloc(nr_threads
, sizeof(*thread_id
));
1001 for (thread
= 0; thread
< nr_threads
; thread
++) {
1002 work
[thread
].ht
= ht
;
1004 work
[thread
].len
= partition_len
;
1005 work
[thread
].start
= thread
* partition_len
;
1006 work
[thread
].fct
= fct
;
1007 ret
= pthread_create(&thread_id
[thread
], ht
->resize_attr
,
1008 partition_resize_thread
, &work
[thread
]);
1011 for (thread
= 0; thread
< nr_threads
; thread
++) {
1012 ret
= pthread_join(thread_id
[thread
], NULL
);
1020 * Holding RCU read lock to protect _cds_lfht_add against memory
1021 * reclaim that could be performed by other call_rcu worker threads (ABA
1024 * When we reach a certain length, we can split this population phase over
1025 * many worker threads, based on the number of CPUs available in the system.
1026 * This should therefore take care of not having the expand lagging behind too
1027 * many concurrent insertion threads by using the scheduler's ability to
1028 * schedule dummy node population fairly with insertions.
1031 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1032 unsigned long start
, unsigned long len
)
1036 ht
->cds_lfht_rcu_read_lock();
1037 for (j
= start
; j
< start
+ len
; j
++) {
1038 struct cds_lfht_node
*new_node
=
1039 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1041 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1042 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1043 new_node
->p
.reverse_hash
=
1044 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1045 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1046 new_node
, ADD_DEFAULT
, 1);
1047 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1050 ht
->cds_lfht_rcu_read_unlock();
1054 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1057 assert(nr_cpus_mask
!= -1);
1058 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1059 ht
->cds_lfht_rcu_thread_online();
1060 init_table_populate_partition(ht
, i
, 0, len
);
1061 ht
->cds_lfht_rcu_thread_offline();
1064 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1068 void init_table(struct cds_lfht
*ht
,
1069 unsigned long first_order
, unsigned long len_order
)
1071 unsigned long i
, end_order
;
1073 dbg_printf("init table: first_order %lu end_order %lu\n",
1074 first_order
, first_order
+ len_order
);
1075 end_order
= first_order
+ len_order
;
1076 for (i
= first_order
; i
< end_order
; i
++) {
1079 len
= !i
? 1 : 1UL << (i
- 1);
1080 dbg_printf("init order %lu len: %lu\n", i
, len
);
1082 /* Stop expand if the resize target changes under us */
1083 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1086 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1087 + (len
* sizeof(struct _cds_lfht_node
)));
1088 assert(ht
->t
.tbl
[i
]);
1091 * Set all dummy nodes reverse hash values for a level and
1092 * link all dummy nodes into the table.
1094 init_table_populate(ht
, i
, len
);
1097 * Update table size.
1099 cmm_smp_wmb(); /* populate data before RCU size */
1100 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1102 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1103 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1109 * Holding RCU read lock to protect _cds_lfht_remove against memory
1110 * reclaim that could be performed by other call_rcu worker threads (ABA
1112 * For a single level, we logically remove and garbage collect each node.
1114 * As a design choice, we perform logical removal and garbage collection on a
1115 * node-per-node basis to simplify this algorithm. We also assume keeping good
1116 * cache locality of the operation would overweight possible performance gain
1117 * that could be achieved by batching garbage collection for multiple levels.
1118 * However, this would have to be justified by benchmarks.
1120 * Concurrent removal and add operations are helping us perform garbage
1121 * collection of logically removed nodes. We guarantee that all logically
1122 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1123 * invoked to free a hole level of dummy nodes (after a grace period).
1125 * Logical removal and garbage collection can therefore be done in batch or on a
1126 * node-per-node basis, as long as the guarantee above holds.
1128 * When we reach a certain length, we can split this removal over many worker
1129 * threads, based on the number of CPUs available in the system. This should
1130 * take care of not letting resize process lag behind too many concurrent
1131 * updater threads actively inserting into the hash table.
1134 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1135 unsigned long start
, unsigned long len
)
1139 ht
->cds_lfht_rcu_read_lock();
1140 for (j
= start
; j
< start
+ len
; j
++) {
1141 struct cds_lfht_node
*fini_node
=
1142 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1144 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1145 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1146 fini_node
->p
.reverse_hash
=
1147 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1148 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1150 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1153 ht
->cds_lfht_rcu_read_unlock();
1157 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1160 assert(nr_cpus_mask
!= -1);
1161 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1162 ht
->cds_lfht_rcu_thread_online();
1163 remove_table_partition(ht
, i
, 0, len
);
1164 ht
->cds_lfht_rcu_thread_offline();
1167 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1171 void fini_table(struct cds_lfht
*ht
,
1172 unsigned long first_order
, unsigned long len_order
)
1176 dbg_printf("fini table: first_order %lu end_order %lu\n",
1177 first_order
, first_order
+ len_order
);
1178 end_order
= first_order
+ len_order
;
1179 assert(first_order
> 0);
1180 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1183 len
= !i
? 1 : 1UL << (i
- 1);
1184 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1186 /* Stop shrink if the resize target changes under us */
1187 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1190 cmm_smp_wmb(); /* populate data before RCU size */
1191 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1194 * We need to wait for all add operations to reach Q.S. (and
1195 * thus use the new table for lookups) before we can start
1196 * releasing the old dummy nodes. Otherwise their lookup will
1197 * return a logically removed node as insert position.
1199 ht
->cds_lfht_synchronize_rcu();
1202 * Set "removed" flag in dummy nodes about to be removed.
1203 * Unlink all now-logically-removed dummy node pointers.
1204 * Concurrent add/remove operation are helping us doing
1207 remove_table(ht
, i
, len
);
1209 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1211 dbg_printf("fini new size: %lu\n", 1UL << i
);
1212 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1217 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1218 cds_lfht_compare_fct compare_fct
,
1219 unsigned long hash_seed
,
1220 unsigned long init_size
,
1222 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1223 void (*func
)(struct rcu_head
*head
)),
1224 void (*cds_lfht_synchronize_rcu
)(void),
1225 void (*cds_lfht_rcu_read_lock
)(void),
1226 void (*cds_lfht_rcu_read_unlock
)(void),
1227 void (*cds_lfht_rcu_thread_offline
)(void),
1228 void (*cds_lfht_rcu_thread_online
)(void),
1229 void (*cds_lfht_rcu_register_thread
)(void),
1230 void (*cds_lfht_rcu_unregister_thread
)(void),
1231 pthread_attr_t
*attr
)
1233 struct cds_lfht
*ht
;
1234 unsigned long order
;
1236 /* init_size must be power of two */
1237 if (init_size
&& (init_size
& (init_size
- 1)))
1239 ht
= calloc(1, sizeof(struct cds_lfht
));
1241 ht
->hash_fct
= hash_fct
;
1242 ht
->compare_fct
= compare_fct
;
1243 ht
->hash_seed
= hash_seed
;
1244 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1245 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1246 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1247 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1248 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1249 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1250 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1251 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1252 ht
->resize_attr
= attr
;
1253 ht
->percpu_count
= alloc_per_cpu_items_count();
1254 /* this mutex should not nest in read-side C.S. */
1255 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1256 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1258 ht
->cds_lfht_rcu_thread_offline();
1259 pthread_mutex_lock(&ht
->resize_mutex
);
1260 ht
->t
.resize_target
= 1UL << (order
- 1);
1261 init_table(ht
, 0, order
);
1262 pthread_mutex_unlock(&ht
->resize_mutex
);
1263 ht
->cds_lfht_rcu_thread_online();
1267 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1268 struct cds_lfht_iter
*iter
)
1270 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1271 struct _cds_lfht_node
*lookup
;
1272 unsigned long hash
, reverse_hash
, index
, order
, size
;
1274 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1275 reverse_hash
= bit_reverse_ulong(hash
);
1277 size
= rcu_dereference(ht
->t
.size
);
1278 index
= hash
& (size
- 1);
1279 order
= get_count_order_ulong(index
+ 1);
1280 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1281 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1282 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1283 dummy_node
= (struct cds_lfht_node
*) lookup
;
1284 /* We can always skip the dummy node initially */
1285 node
= rcu_dereference(dummy_node
->p
.next
);
1286 node
= clear_flag(node
);
1288 if (unlikely(is_end(node
))) {
1292 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1296 next
= rcu_dereference(node
->p
.next
);
1297 if (likely(!is_removed(next
))
1299 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1302 node
= clear_flag(next
);
1304 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1309 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1311 struct cds_lfht_node
*node
, *next
;
1312 unsigned long reverse_hash
;
1317 reverse_hash
= node
->p
.reverse_hash
;
1319 key_len
= node
->key_len
;
1321 node
= clear_flag(next
);
1324 if (unlikely(is_end(node
))) {
1328 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1332 next
= rcu_dereference(node
->p
.next
);
1333 if (likely(!is_removed(next
))
1335 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1338 node
= clear_flag(next
);
1340 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1345 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1347 unsigned long hash
, size
;
1349 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1350 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1352 size
= rcu_dereference(ht
->t
.size
);
1353 (void) _cds_lfht_add(ht
, size
, node
, ADD_DEFAULT
, 0);
1354 ht_count_add(ht
, size
);
1357 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1358 struct cds_lfht_node
*node
)
1360 unsigned long hash
, size
;
1361 struct cds_lfht_node
*ret
;
1363 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1364 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1366 size
= rcu_dereference(ht
->t
.size
);
1367 ret
= _cds_lfht_add(ht
, size
, node
, ADD_UNIQUE
, 0);
1369 ht_count_add(ht
, size
);
1373 struct cds_lfht_node
*cds_lfht_replace(struct cds_lfht
*ht
,
1374 struct cds_lfht_node
*node
)
1376 unsigned long hash
, size
;
1377 struct cds_lfht_node
*ret
;
1379 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1380 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1382 size
= rcu_dereference(ht
->t
.size
);
1383 ret
= _cds_lfht_add(ht
, size
, node
, ADD_REPLACE
, 0);
1385 ht_count_add(ht
, size
);
1389 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1394 size
= rcu_dereference(ht
->t
.size
);
1395 ret
= _cds_lfht_del(ht
, size
, node
, 0);
1397 ht_count_del(ht
, size
);
1402 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1404 struct cds_lfht_node
*node
;
1405 struct _cds_lfht_node
*lookup
;
1406 unsigned long order
, i
, size
;
1408 /* Check that the table is empty */
1409 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1410 node
= (struct cds_lfht_node
*) lookup
;
1412 node
= clear_flag(node
)->p
.next
;
1413 if (!is_dummy(node
))
1415 assert(!is_removed(node
));
1416 } while (!is_end(node
));
1418 * size accessed without rcu_dereference because hash table is
1422 /* Internal sanity check: all nodes left should be dummy */
1423 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1426 len
= !order
? 1 : 1UL << (order
- 1);
1427 for (i
= 0; i
< len
; i
++) {
1428 dbg_printf("delete order %lu i %lu hash %lu\n",
1430 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1431 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1433 poison_free(ht
->t
.tbl
[order
]);
1439 * Should only be called when no more concurrent readers nor writers can
1440 * possibly access the table.
1442 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1446 /* Wait for in-flight resize operations to complete */
1447 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1448 while (uatomic_read(&ht
->in_progress_resize
))
1449 poll(NULL
, 0, 100); /* wait for 100ms */
1450 ret
= cds_lfht_delete_dummy(ht
);
1453 free_per_cpu_items_count(ht
->percpu_count
);
1455 *attr
= ht
->resize_attr
;
1460 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1461 unsigned long *approx_before
,
1462 unsigned long *count
,
1463 unsigned long *removed
,
1464 unsigned long *approx_after
)
1466 struct cds_lfht_node
*node
, *next
;
1467 struct _cds_lfht_node
*lookup
;
1468 unsigned long nr_dummy
= 0;
1471 if (nr_cpus_mask
>= 0) {
1474 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1475 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1476 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1483 /* Count non-dummy nodes in the table */
1484 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1485 node
= (struct cds_lfht_node
*) lookup
;
1487 next
= rcu_dereference(node
->p
.next
);
1488 if (is_removed(next
)) {
1489 if (!is_dummy(next
))
1493 } else if (!is_dummy(next
))
1497 node
= clear_flag(next
);
1498 } while (!is_end(node
));
1499 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1501 if (nr_cpus_mask
>= 0) {
1504 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1505 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1506 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1511 /* called with resize mutex held */
1513 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1514 unsigned long old_size
, unsigned long new_size
)
1516 unsigned long old_order
, new_order
;
1518 old_order
= get_count_order_ulong(old_size
) + 1;
1519 new_order
= get_count_order_ulong(new_size
) + 1;
1520 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1521 old_size
, old_order
, new_size
, new_order
);
1522 assert(new_size
> old_size
);
1523 init_table(ht
, old_order
, new_order
- old_order
);
1526 /* called with resize mutex held */
1528 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1529 unsigned long old_size
, unsigned long new_size
)
1531 unsigned long old_order
, new_order
;
1533 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1534 old_order
= get_count_order_ulong(old_size
) + 1;
1535 new_order
= get_count_order_ulong(new_size
) + 1;
1536 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1537 old_size
, old_order
, new_size
, new_order
);
1538 assert(new_size
< old_size
);
1540 /* Remove and unlink all dummy nodes to remove. */
1541 fini_table(ht
, new_order
, old_order
- new_order
);
1545 /* called with resize mutex held */
1547 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1549 unsigned long new_size
, old_size
;
1552 * Resize table, re-do if the target size has changed under us.
1555 ht
->t
.resize_initiated
= 1;
1556 old_size
= ht
->t
.size
;
1557 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1558 if (old_size
< new_size
)
1559 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1560 else if (old_size
> new_size
)
1561 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1562 ht
->t
.resize_initiated
= 0;
1563 /* write resize_initiated before read resize_target */
1565 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1569 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1572 return _uatomic_max(&ht
->t
.resize_target
,
1573 size
<< growth_order
);
1577 void resize_target_update_count(struct cds_lfht
*ht
,
1578 unsigned long count
)
1580 count
= max(count
, MIN_TABLE_SIZE
);
1581 uatomic_set(&ht
->t
.resize_target
, count
);
1584 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1586 resize_target_update_count(ht
, new_size
);
1587 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1588 ht
->cds_lfht_rcu_thread_offline();
1589 pthread_mutex_lock(&ht
->resize_mutex
);
1590 _do_cds_lfht_resize(ht
);
1591 pthread_mutex_unlock(&ht
->resize_mutex
);
1592 ht
->cds_lfht_rcu_thread_online();
1596 void do_resize_cb(struct rcu_head
*head
)
1598 struct rcu_resize_work
*work
=
1599 caa_container_of(head
, struct rcu_resize_work
, head
);
1600 struct cds_lfht
*ht
= work
->ht
;
1602 ht
->cds_lfht_rcu_thread_offline();
1603 pthread_mutex_lock(&ht
->resize_mutex
);
1604 _do_cds_lfht_resize(ht
);
1605 pthread_mutex_unlock(&ht
->resize_mutex
);
1606 ht
->cds_lfht_rcu_thread_online();
1608 cmm_smp_mb(); /* finish resize before decrement */
1609 uatomic_dec(&ht
->in_progress_resize
);
1613 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1615 struct rcu_resize_work
*work
;
1616 unsigned long target_size
;
1618 target_size
= resize_target_update(ht
, size
, growth
);
1619 /* Store resize_target before read resize_initiated */
1621 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1622 uatomic_inc(&ht
->in_progress_resize
);
1623 cmm_smp_mb(); /* increment resize count before calling it */
1624 work
= malloc(sizeof(*work
));
1626 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1627 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1631 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1634 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1635 unsigned long count
)
1637 struct rcu_resize_work
*work
;
1639 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1641 resize_target_update_count(ht
, count
);
1642 /* Store resize_target before read resize_initiated */
1644 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1645 uatomic_inc(&ht
->in_progress_resize
);
1646 cmm_smp_mb(); /* increment resize count before calling it */
1647 work
= malloc(sizeof(*work
));
1649 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1650 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);