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 * 2 | | 2 010 010 <- |
124 * | | | 3 011 110 | <- |
125 * 3 -> | | | 4 100 001 | |
141 #include <urcu-call-rcu.h>
142 #include <urcu/arch.h>
143 #include <urcu/uatomic.h>
144 #include <urcu/compiler.h>
145 #include <urcu/rculfhash.h>
150 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
152 #define dbg_printf(fmt, args...)
156 * Per-CPU split-counters lazily update the global counter each 1024
157 * addition/removal. It automatically keeps track of resize required.
158 * We use the bucket length as indicator for need to expand for small
159 * tables and machines lacking per-cpu data suppport.
161 #define COUNT_COMMIT_ORDER 10
162 #define CHAIN_LEN_TARGET 1
163 #define CHAIN_LEN_RESIZE_THRESHOLD 3
166 * Define the minimum table size.
168 #define MIN_TABLE_SIZE 1
170 #if (CAA_BITS_PER_LONG == 32)
171 #define MAX_TABLE_ORDER 32
173 #define MAX_TABLE_ORDER 64
177 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
179 #define MIN_PARTITION_PER_THREAD_ORDER 12
180 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
183 #define min(a, b) ((a) < (b) ? (a) : (b))
187 #define max(a, b) ((a) > (b) ? (a) : (b))
191 * The removed flag needs to be updated atomically with the pointer.
192 * It indicates that no node must attach to the node scheduled for
193 * removal, and that node garbage collection must be performed.
194 * The dummy flag does not require to be updated atomically with the
195 * pointer, but it is added as a pointer low bit flag to save space.
197 #define REMOVED_FLAG (1UL << 0)
198 #define DUMMY_FLAG (1UL << 1)
199 #define FLAGS_MASK ((1UL << 2) - 1)
201 /* Value of the end pointer. Should not interact with flags. */
202 #define END_VALUE NULL
204 struct ht_items_count
{
205 unsigned long add
, del
;
206 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
209 /* Note: manually update allocation length when adding a field */
210 struct _cds_lfht_node nodes
[0];
214 unsigned long size
; /* always a power of 2, shared (RCU) */
215 unsigned long resize_target
;
216 int resize_initiated
;
217 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
222 cds_lfht_hash_fct hash_fct
;
223 cds_lfht_compare_fct compare_fct
;
224 unsigned long hash_seed
;
227 * We need to put the work threads offline (QSBR) when taking this
228 * mutex, because we use synchronize_rcu within this mutex critical
229 * section, which waits on read-side critical sections, and could
230 * therefore cause grace-period deadlock if we hold off RCU G.P.
233 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
234 unsigned int in_progress_resize
, in_progress_destroy
;
235 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
236 void (*func
)(struct rcu_head
*head
));
237 void (*cds_lfht_synchronize_rcu
)(void);
238 void (*cds_lfht_rcu_read_lock
)(void);
239 void (*cds_lfht_rcu_read_unlock
)(void);
240 void (*cds_lfht_rcu_thread_offline
)(void);
241 void (*cds_lfht_rcu_thread_online
)(void);
242 void (*cds_lfht_rcu_register_thread
)(void);
243 void (*cds_lfht_rcu_unregister_thread
)(void);
244 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
245 long count
; /* global approximate item count */
246 struct ht_items_count
*percpu_count
; /* per-cpu item count */
249 struct rcu_resize_work
{
250 struct rcu_head head
;
254 struct partition_resize_work
{
257 unsigned long i
, start
, len
;
258 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
259 unsigned long start
, unsigned long len
);
263 void _cds_lfht_add(struct cds_lfht
*ht
,
265 struct cds_lfht_node
*node
,
266 struct cds_lfht_iter
*unique_ret
,
270 * Algorithm to reverse bits in a word by lookup table, extended to
273 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
274 * Originally from Public Domain.
277 static const uint8_t BitReverseTable256
[256] =
279 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
280 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
281 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
282 R6(0), R6(2), R6(1), R6(3)
289 uint8_t bit_reverse_u8(uint8_t v
)
291 return BitReverseTable256
[v
];
294 static __attribute__((unused
))
295 uint32_t bit_reverse_u32(uint32_t v
)
297 return ((uint32_t) bit_reverse_u8(v
) << 24) |
298 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
299 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
300 ((uint32_t) bit_reverse_u8(v
>> 24));
303 static __attribute__((unused
))
304 uint64_t bit_reverse_u64(uint64_t v
)
306 return ((uint64_t) bit_reverse_u8(v
) << 56) |
307 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
308 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
309 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
310 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
311 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
312 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
313 ((uint64_t) bit_reverse_u8(v
>> 56));
317 unsigned long bit_reverse_ulong(unsigned long v
)
319 #if (CAA_BITS_PER_LONG == 32)
320 return bit_reverse_u32(v
);
322 return bit_reverse_u64(v
);
327 * fls: returns the position of the most significant bit.
328 * Returns 0 if no bit is set, else returns the position of the most
329 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
331 #if defined(__i386) || defined(__x86_64)
333 unsigned int fls_u32(uint32_t x
)
341 : "=r" (r
) : "rm" (x
));
347 #if defined(__x86_64)
349 unsigned int fls_u64(uint64_t x
)
357 : "=r" (r
) : "rm" (x
));
364 static __attribute__((unused
))
365 unsigned int fls_u64(uint64_t x
)
372 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
376 if (!(x
& 0xFFFF000000000000ULL
)) {
380 if (!(x
& 0xFF00000000000000ULL
)) {
384 if (!(x
& 0xF000000000000000ULL
)) {
388 if (!(x
& 0xC000000000000000ULL
)) {
392 if (!(x
& 0x8000000000000000ULL
)) {
401 static __attribute__((unused
))
402 unsigned int fls_u32(uint32_t x
)
408 if (!(x
& 0xFFFF0000U
)) {
412 if (!(x
& 0xFF000000U
)) {
416 if (!(x
& 0xF0000000U
)) {
420 if (!(x
& 0xC0000000U
)) {
424 if (!(x
& 0x80000000U
)) {
432 unsigned int fls_ulong(unsigned long x
)
434 #if (CAA_BITS_PER_lONG == 32)
442 * Return the minimum order for which x <= (1UL << order).
443 * Return -1 if x is 0.
445 int get_count_order_u32(uint32_t x
)
450 return fls_u32(x
- 1);
454 * Return the minimum order for which x <= (1UL << order).
455 * Return -1 if x is 0.
457 int get_count_order_ulong(unsigned long x
)
462 return fls_ulong(x
- 1);
466 #define poison_free(ptr) \
468 memset(ptr, 0x42, sizeof(*(ptr))); \
472 #define poison_free(ptr) free(ptr)
476 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
479 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
480 * available, then we support hash table item accounting.
481 * In the unfortunate event the number of CPUs reported would be
482 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
484 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
487 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
488 unsigned long count
);
490 static long nr_cpus_mask
= -1;
493 struct ht_items_count
*alloc_per_cpu_items_count(void)
495 struct ht_items_count
*count
;
497 switch (nr_cpus_mask
) {
504 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
510 * round up number of CPUs to next power of two, so we
511 * can use & for modulo.
513 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
514 nr_cpus_mask
= maxcpus
- 1;
518 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
523 void free_per_cpu_items_count(struct ht_items_count
*count
)
533 assert(nr_cpus_mask
>= 0);
534 cpu
= sched_getcpu();
535 if (unlikely(cpu
< 0))
538 return cpu
& nr_cpus_mask
;
542 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
544 unsigned long percpu_count
;
547 if (unlikely(!ht
->percpu_count
))
550 if (unlikely(cpu
< 0))
552 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
553 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
556 dbg_printf("add percpu %lu\n", percpu_count
);
557 count
= uatomic_add_return(&ht
->count
,
558 1UL << COUNT_COMMIT_ORDER
);
560 if (!(count
& (count
- 1))) {
561 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
563 dbg_printf("add set global %ld\n", count
);
564 cds_lfht_resize_lazy_count(ht
, size
,
565 count
>> (CHAIN_LEN_TARGET
- 1));
571 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
573 unsigned long percpu_count
;
576 if (unlikely(!ht
->percpu_count
))
579 if (unlikely(cpu
< 0))
581 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
582 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
585 dbg_printf("del percpu %lu\n", percpu_count
);
586 count
= uatomic_add_return(&ht
->count
,
587 -(1UL << COUNT_COMMIT_ORDER
));
589 if (!(count
& (count
- 1))) {
590 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
592 dbg_printf("del set global %ld\n", count
);
594 * Don't shrink table if the number of nodes is below a
597 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (nr_cpus_mask
+ 1))
599 cds_lfht_resize_lazy_count(ht
, size
,
600 count
>> (CHAIN_LEN_TARGET
- 1));
605 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
607 static const long nr_cpus_mask
= -2;
610 struct ht_items_count
*alloc_per_cpu_items_count(void)
616 void free_per_cpu_items_count(struct ht_items_count
*count
)
621 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
626 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
630 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
634 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
638 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
640 count
= uatomic_read(&ht
->count
);
642 * Use bucket-local length for small table expand and for
643 * environments lacking per-cpu data support.
645 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
648 dbg_printf("WARNING: large chain length: %u.\n",
650 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
651 cds_lfht_resize_lazy(ht
, size
,
652 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
656 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
658 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
662 int is_removed(struct cds_lfht_node
*node
)
664 return ((unsigned long) node
) & REMOVED_FLAG
;
668 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
670 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
674 int is_dummy(struct cds_lfht_node
*node
)
676 return ((unsigned long) node
) & DUMMY_FLAG
;
680 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
682 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
686 struct cds_lfht_node
*get_end(void)
688 return (struct cds_lfht_node
*) END_VALUE
;
692 int is_end(struct cds_lfht_node
*node
)
694 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
698 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
700 unsigned long old1
, old2
;
702 old1
= uatomic_read(ptr
);
707 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
712 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
715 unsigned long index
, order
;
718 index
= hash
& (size
- 1);
720 * equivalent to get_count_order_ulong(index + 1), but optimizes
721 * away the non-existing 0 special-case for
722 * get_count_order_ulong.
724 order
= fls_ulong(index
);
726 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
727 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
729 return &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
733 * Remove all logically deleted nodes from a bucket up to a certain node key.
736 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
738 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
740 assert(!is_dummy(dummy
));
741 assert(!is_removed(dummy
));
742 assert(!is_dummy(node
));
743 assert(!is_removed(node
));
746 /* We can always skip the dummy node initially */
747 iter
= rcu_dereference(iter_prev
->p
.next
);
748 assert(!is_removed(iter
));
749 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
751 * We should never be called with dummy (start of chain)
752 * and logically removed node (end of path compression
753 * marker) being the actual same node. This would be a
754 * bug in the algorithm implementation.
756 assert(dummy
!= node
);
758 if (unlikely(is_end(iter
)))
760 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
762 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
763 if (likely(is_removed(next
)))
765 iter_prev
= clear_flag(iter
);
768 assert(!is_removed(iter
));
770 new_next
= flag_dummy(clear_flag(next
));
772 new_next
= clear_flag(next
);
773 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
779 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
780 struct cds_lfht_node
*old_node
,
781 struct cds_lfht_node
*old_next
,
782 struct cds_lfht_node
*new_node
)
784 struct cds_lfht_node
*dummy
, *ret_next
;
785 struct _cds_lfht_node
*lookup
;
787 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
790 assert(!is_removed(old_node
));
791 assert(!is_dummy(old_node
));
792 assert(!is_removed(new_node
));
793 assert(!is_dummy(new_node
));
794 assert(new_node
!= old_node
);
796 /* Insert after node to be replaced */
797 if (is_removed(old_next
)) {
799 * Too late, the old node has been removed under us
800 * between lookup and replace. Fail.
804 assert(!is_dummy(old_next
));
805 assert(new_node
!= clear_flag(old_next
));
806 new_node
->p
.next
= clear_flag(old_next
);
808 * Here is the whole trick for lock-free replace: we add
809 * the replacement node _after_ the node we want to
810 * replace by atomically setting its next pointer at the
811 * same time we set its removal flag. Given that
812 * the lookups/get next use an iterator aware of the
813 * next pointer, they will either skip the old node due
814 * to the removal flag and see the new node, or use
815 * the old node, but will not see the new one.
817 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
818 old_next
, flag_removed(new_node
));
819 if (ret_next
== old_next
)
820 break; /* We performed the replacement. */
825 * Ensure that the old node is not visible to readers anymore:
826 * lookup for the node, and remove it (along with any other
827 * logically removed node) if found.
829 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
830 dummy
= (struct cds_lfht_node
*) lookup
;
831 _cds_lfht_gc_bucket(dummy
, new_node
);
833 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
838 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
839 * mode. A NULL unique_ret allows creation of duplicate keys.
842 void _cds_lfht_add(struct cds_lfht
*ht
,
844 struct cds_lfht_node
*node
,
845 struct cds_lfht_iter
*unique_ret
,
848 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
850 struct _cds_lfht_node
*lookup
;
852 assert(!is_dummy(node
));
853 assert(!is_removed(node
));
857 node
->p
.next
= flag_dummy(get_end());
858 return; /* Initial first add (head) */
860 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
862 uint32_t chain_len
= 0;
865 * iter_prev points to the non-removed node prior to the
868 iter_prev
= (struct cds_lfht_node
*) lookup
;
869 /* We can always skip the dummy node initially */
870 iter
= rcu_dereference(iter_prev
->p
.next
);
871 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
873 if (unlikely(is_end(iter
)))
875 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
877 /* dummy node is the first node of the identical-hash-value chain */
878 if (dummy
&& clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
)
880 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
881 if (unlikely(is_removed(next
)))
885 && clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
886 && !ht
->compare_fct(node
->key
, node
->key_len
,
887 clear_flag(iter
)->key
,
888 clear_flag(iter
)->key_len
)) {
889 unique_ret
->node
= clear_flag(iter
);
890 unique_ret
->next
= next
;
893 /* Only account for identical reverse hash once */
894 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
896 check_resize(ht
, size
, ++chain_len
);
897 iter_prev
= clear_flag(iter
);
902 assert(node
!= clear_flag(iter
));
903 assert(!is_removed(iter_prev
));
904 assert(!is_removed(iter
));
905 assert(iter_prev
!= node
);
907 node
->p
.next
= clear_flag(iter
);
909 node
->p
.next
= flag_dummy(clear_flag(iter
));
911 new_node
= flag_dummy(node
);
914 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
916 continue; /* retry */
923 assert(!is_removed(iter
));
925 new_next
= flag_dummy(clear_flag(next
));
927 new_next
= clear_flag(next
);
928 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
933 unique_ret
->node
= return_node
;
934 /* unique_ret->next left unset, never used. */
939 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
940 struct cds_lfht_node
*node
,
943 struct cds_lfht_node
*dummy
, *next
, *old
;
944 struct _cds_lfht_node
*lookup
;
946 if (!node
) /* Return -ENOENT if asked to delete NULL node */
949 /* logically delete the node */
950 assert(!is_dummy(node
));
951 assert(!is_removed(node
));
952 old
= rcu_dereference(node
->p
.next
);
954 struct cds_lfht_node
*new_next
;
957 if (unlikely(is_removed(next
)))
960 assert(is_dummy(next
));
962 assert(!is_dummy(next
));
963 new_next
= flag_removed(next
);
964 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
965 } while (old
!= next
);
966 /* We performed the (logical) deletion. */
969 * Ensure that the node is not visible to readers anymore: lookup for
970 * the node, and remove it (along with any other logically removed node)
973 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
974 dummy
= (struct cds_lfht_node
*) lookup
;
975 _cds_lfht_gc_bucket(dummy
, node
);
977 assert(is_removed(rcu_dereference(node
->p
.next
)));
982 void *partition_resize_thread(void *arg
)
984 struct partition_resize_work
*work
= arg
;
986 work
->ht
->cds_lfht_rcu_register_thread();
987 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
988 work
->ht
->cds_lfht_rcu_unregister_thread();
993 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
995 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
996 unsigned long start
, unsigned long len
))
998 unsigned long partition_len
;
999 struct partition_resize_work
*work
;
1001 unsigned long nr_threads
;
1004 * Note: nr_cpus_mask + 1 is always power of 2.
1005 * We spawn just the number of threads we need to satisfy the minimum
1006 * partition size, up to the number of CPUs in the system.
1008 if (nr_cpus_mask
> 0) {
1009 nr_threads
= min(nr_cpus_mask
+ 1,
1010 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1014 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1015 work
= calloc(nr_threads
, sizeof(*work
));
1017 for (thread
= 0; thread
< nr_threads
; thread
++) {
1018 work
[thread
].ht
= ht
;
1020 work
[thread
].len
= partition_len
;
1021 work
[thread
].start
= thread
* partition_len
;
1022 work
[thread
].fct
= fct
;
1023 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1024 partition_resize_thread
, &work
[thread
]);
1027 for (thread
= 0; thread
< nr_threads
; thread
++) {
1028 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1035 * Holding RCU read lock to protect _cds_lfht_add against memory
1036 * reclaim that could be performed by other call_rcu worker threads (ABA
1039 * When we reach a certain length, we can split this population phase over
1040 * many worker threads, based on the number of CPUs available in the system.
1041 * This should therefore take care of not having the expand lagging behind too
1042 * many concurrent insertion threads by using the scheduler's ability to
1043 * schedule dummy node population fairly with insertions.
1046 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1047 unsigned long start
, unsigned long len
)
1051 ht
->cds_lfht_rcu_read_lock();
1052 for (j
= start
; j
< start
+ len
; j
++) {
1053 struct cds_lfht_node
*new_node
=
1054 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1056 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1057 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1058 new_node
->p
.reverse_hash
=
1059 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1060 _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1063 ht
->cds_lfht_rcu_read_unlock();
1067 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1070 assert(nr_cpus_mask
!= -1);
1071 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1072 ht
->cds_lfht_rcu_thread_online();
1073 init_table_populate_partition(ht
, i
, 0, len
);
1074 ht
->cds_lfht_rcu_thread_offline();
1077 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1081 void init_table(struct cds_lfht
*ht
,
1082 unsigned long first_order
, unsigned long len_order
)
1084 unsigned long i
, end_order
;
1086 dbg_printf("init table: first_order %lu end_order %lu\n",
1087 first_order
, first_order
+ len_order
);
1088 end_order
= first_order
+ len_order
;
1089 for (i
= first_order
; i
< end_order
; i
++) {
1092 len
= !i
? 1 : 1UL << (i
- 1);
1093 dbg_printf("init order %lu len: %lu\n", i
, len
);
1095 /* Stop expand if the resize target changes under us */
1096 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1099 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1100 assert(ht
->t
.tbl
[i
]);
1103 * Set all dummy nodes reverse hash values for a level and
1104 * link all dummy nodes into the table.
1106 init_table_populate(ht
, i
, len
);
1109 * Update table size.
1111 cmm_smp_wmb(); /* populate data before RCU size */
1112 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1114 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1115 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1121 * Holding RCU read lock to protect _cds_lfht_remove against memory
1122 * reclaim that could be performed by other call_rcu worker threads (ABA
1124 * For a single level, we logically remove and garbage collect each node.
1126 * As a design choice, we perform logical removal and garbage collection on a
1127 * node-per-node basis to simplify this algorithm. We also assume keeping good
1128 * cache locality of the operation would overweight possible performance gain
1129 * that could be achieved by batching garbage collection for multiple levels.
1130 * However, this would have to be justified by benchmarks.
1132 * Concurrent removal and add operations are helping us perform garbage
1133 * collection of logically removed nodes. We guarantee that all logically
1134 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1135 * invoked to free a hole level of dummy nodes (after a grace period).
1137 * Logical removal and garbage collection can therefore be done in batch or on a
1138 * node-per-node basis, as long as the guarantee above holds.
1140 * When we reach a certain length, we can split this removal over many worker
1141 * threads, based on the number of CPUs available in the system. This should
1142 * take care of not letting resize process lag behind too many concurrent
1143 * updater threads actively inserting into the hash table.
1146 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1147 unsigned long start
, unsigned long len
)
1151 ht
->cds_lfht_rcu_read_lock();
1152 for (j
= start
; j
< start
+ len
; j
++) {
1153 struct cds_lfht_node
*fini_node
=
1154 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1156 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1157 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1158 fini_node
->p
.reverse_hash
=
1159 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1160 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1163 ht
->cds_lfht_rcu_read_unlock();
1167 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1170 assert(nr_cpus_mask
!= -1);
1171 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1172 ht
->cds_lfht_rcu_thread_online();
1173 remove_table_partition(ht
, i
, 0, len
);
1174 ht
->cds_lfht_rcu_thread_offline();
1177 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1181 void fini_table(struct cds_lfht
*ht
,
1182 unsigned long first_order
, unsigned long len_order
)
1185 void *free_by_rcu
= NULL
;
1187 dbg_printf("fini table: first_order %lu end_order %lu\n",
1188 first_order
, first_order
+ len_order
);
1189 end_order
= first_order
+ len_order
;
1190 assert(first_order
> 0);
1191 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1194 len
= !i
? 1 : 1UL << (i
- 1);
1195 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1197 /* Stop shrink if the resize target changes under us */
1198 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1201 cmm_smp_wmb(); /* populate data before RCU size */
1202 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1205 * We need to wait for all add operations to reach Q.S. (and
1206 * thus use the new table for lookups) before we can start
1207 * releasing the old dummy nodes. Otherwise their lookup will
1208 * return a logically removed node as insert position.
1210 ht
->cds_lfht_synchronize_rcu();
1215 * Set "removed" flag in dummy nodes about to be removed.
1216 * Unlink all now-logically-removed dummy node pointers.
1217 * Concurrent add/remove operation are helping us doing
1220 remove_table(ht
, i
, len
);
1222 free_by_rcu
= ht
->t
.tbl
[i
];
1224 dbg_printf("fini new size: %lu\n", 1UL << i
);
1225 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1230 ht
->cds_lfht_synchronize_rcu();
1235 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1236 cds_lfht_compare_fct compare_fct
,
1237 unsigned long hash_seed
,
1238 unsigned long init_size
,
1240 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1241 void (*func
)(struct rcu_head
*head
)),
1242 void (*cds_lfht_synchronize_rcu
)(void),
1243 void (*cds_lfht_rcu_read_lock
)(void),
1244 void (*cds_lfht_rcu_read_unlock
)(void),
1245 void (*cds_lfht_rcu_thread_offline
)(void),
1246 void (*cds_lfht_rcu_thread_online
)(void),
1247 void (*cds_lfht_rcu_register_thread
)(void),
1248 void (*cds_lfht_rcu_unregister_thread
)(void),
1249 pthread_attr_t
*attr
)
1251 struct cds_lfht
*ht
;
1252 unsigned long order
;
1254 /* init_size must be power of two */
1255 if (init_size
&& (init_size
& (init_size
- 1)))
1257 ht
= calloc(1, sizeof(struct cds_lfht
));
1259 ht
->hash_fct
= hash_fct
;
1260 ht
->compare_fct
= compare_fct
;
1261 ht
->hash_seed
= hash_seed
;
1262 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1263 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1264 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1265 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1266 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1267 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1268 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1269 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1270 ht
->resize_attr
= attr
;
1271 ht
->percpu_count
= alloc_per_cpu_items_count();
1272 /* this mutex should not nest in read-side C.S. */
1273 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1274 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1276 ht
->cds_lfht_rcu_thread_offline();
1277 pthread_mutex_lock(&ht
->resize_mutex
);
1278 ht
->t
.resize_target
= 1UL << (order
- 1);
1279 init_table(ht
, 0, order
);
1280 pthread_mutex_unlock(&ht
->resize_mutex
);
1281 ht
->cds_lfht_rcu_thread_online();
1285 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1286 struct cds_lfht_iter
*iter
)
1288 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1289 struct _cds_lfht_node
*lookup
;
1290 unsigned long hash
, reverse_hash
, size
;
1292 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1293 reverse_hash
= bit_reverse_ulong(hash
);
1295 size
= rcu_dereference(ht
->t
.size
);
1296 lookup
= lookup_bucket(ht
, size
, hash
);
1297 dummy_node
= (struct cds_lfht_node
*) lookup
;
1298 /* We can always skip the dummy node initially */
1299 node
= rcu_dereference(dummy_node
->p
.next
);
1300 node
= clear_flag(node
);
1302 if (unlikely(is_end(node
))) {
1306 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1310 next
= rcu_dereference(node
->p
.next
);
1311 if (likely(!is_removed(next
))
1313 && clear_flag(node
)->p
.reverse_hash
== reverse_hash
1314 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1317 node
= clear_flag(next
);
1319 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1324 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1326 struct cds_lfht_node
*node
, *next
;
1327 unsigned long reverse_hash
;
1332 reverse_hash
= node
->p
.reverse_hash
;
1334 key_len
= node
->key_len
;
1336 node
= clear_flag(next
);
1339 if (unlikely(is_end(node
))) {
1343 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1347 next
= rcu_dereference(node
->p
.next
);
1348 if (likely(!is_removed(next
))
1350 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1353 node
= clear_flag(next
);
1355 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1360 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1362 struct cds_lfht_node
*node
, *next
;
1364 node
= clear_flag(iter
->next
);
1366 if (unlikely(is_end(node
))) {
1370 next
= rcu_dereference(node
->p
.next
);
1371 if (likely(!is_removed(next
))
1372 && !is_dummy(next
)) {
1375 node
= clear_flag(next
);
1377 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1382 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1384 struct _cds_lfht_node
*lookup
;
1387 * Get next after first dummy node. The first dummy node is the
1388 * first node of the linked list.
1390 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1391 iter
->next
= lookup
->next
;
1392 cds_lfht_next(ht
, iter
);
1395 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1397 unsigned long hash
, size
;
1399 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1400 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1402 size
= rcu_dereference(ht
->t
.size
);
1403 _cds_lfht_add(ht
, size
, node
, NULL
, 0);
1404 ht_count_add(ht
, size
);
1407 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1408 struct cds_lfht_node
*node
)
1410 unsigned long hash
, size
;
1411 struct cds_lfht_iter iter
;
1413 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1414 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1416 size
= rcu_dereference(ht
->t
.size
);
1417 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1418 if (iter
.node
== node
)
1419 ht_count_add(ht
, size
);
1423 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1424 struct cds_lfht_node
*node
)
1426 unsigned long hash
, size
;
1427 struct cds_lfht_iter iter
;
1429 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1430 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1432 size
= rcu_dereference(ht
->t
.size
);
1434 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1435 if (iter
.node
== node
) {
1436 ht_count_add(ht
, size
);
1440 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1445 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1446 struct cds_lfht_node
*new_node
)
1450 size
= rcu_dereference(ht
->t
.size
);
1451 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1455 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1460 size
= rcu_dereference(ht
->t
.size
);
1461 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1463 ht_count_del(ht
, size
);
1468 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1470 struct cds_lfht_node
*node
;
1471 struct _cds_lfht_node
*lookup
;
1472 unsigned long order
, i
, size
;
1474 /* Check that the table is empty */
1475 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1476 node
= (struct cds_lfht_node
*) lookup
;
1478 node
= clear_flag(node
)->p
.next
;
1479 if (!is_dummy(node
))
1481 assert(!is_removed(node
));
1482 } while (!is_end(node
));
1484 * size accessed without rcu_dereference because hash table is
1488 /* Internal sanity check: all nodes left should be dummy */
1489 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1492 len
= !order
? 1 : 1UL << (order
- 1);
1493 for (i
= 0; i
< len
; i
++) {
1494 dbg_printf("delete order %lu i %lu hash %lu\n",
1496 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1497 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1499 poison_free(ht
->t
.tbl
[order
]);
1505 * Should only be called when no more concurrent readers nor writers can
1506 * possibly access the table.
1508 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1512 /* Wait for in-flight resize operations to complete */
1513 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1514 cmm_smp_mb(); /* Store destroy before load resize */
1515 while (uatomic_read(&ht
->in_progress_resize
))
1516 poll(NULL
, 0, 100); /* wait for 100ms */
1517 ret
= cds_lfht_delete_dummy(ht
);
1520 free_per_cpu_items_count(ht
->percpu_count
);
1522 *attr
= ht
->resize_attr
;
1527 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1528 long *approx_before
,
1529 unsigned long *count
,
1530 unsigned long *removed
,
1533 struct cds_lfht_node
*node
, *next
;
1534 struct _cds_lfht_node
*lookup
;
1535 unsigned long nr_dummy
= 0;
1538 if (nr_cpus_mask
>= 0) {
1541 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1542 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1543 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1550 /* Count non-dummy nodes in the table */
1551 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1552 node
= (struct cds_lfht_node
*) lookup
;
1554 next
= rcu_dereference(node
->p
.next
);
1555 if (is_removed(next
)) {
1556 if (!is_dummy(next
))
1560 } else if (!is_dummy(next
))
1564 node
= clear_flag(next
);
1565 } while (!is_end(node
));
1566 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1568 if (nr_cpus_mask
>= 0) {
1571 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1572 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1573 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1578 /* called with resize mutex held */
1580 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1581 unsigned long old_size
, unsigned long new_size
)
1583 unsigned long old_order
, new_order
;
1585 old_order
= get_count_order_ulong(old_size
) + 1;
1586 new_order
= get_count_order_ulong(new_size
) + 1;
1587 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1588 old_size
, old_order
, new_size
, new_order
);
1589 assert(new_size
> old_size
);
1590 init_table(ht
, old_order
, new_order
- old_order
);
1593 /* called with resize mutex held */
1595 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1596 unsigned long old_size
, unsigned long new_size
)
1598 unsigned long old_order
, new_order
;
1600 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1601 old_order
= get_count_order_ulong(old_size
) + 1;
1602 new_order
= get_count_order_ulong(new_size
) + 1;
1603 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1604 old_size
, old_order
, new_size
, new_order
);
1605 assert(new_size
< old_size
);
1607 /* Remove and unlink all dummy nodes to remove. */
1608 fini_table(ht
, new_order
, old_order
- new_order
);
1612 /* called with resize mutex held */
1614 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1616 unsigned long new_size
, old_size
;
1619 * Resize table, re-do if the target size has changed under us.
1622 assert(uatomic_read(&ht
->in_progress_resize
));
1623 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1625 ht
->t
.resize_initiated
= 1;
1626 old_size
= ht
->t
.size
;
1627 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1628 if (old_size
< new_size
)
1629 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1630 else if (old_size
> new_size
)
1631 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1632 ht
->t
.resize_initiated
= 0;
1633 /* write resize_initiated before read resize_target */
1635 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1639 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1642 return _uatomic_max(&ht
->t
.resize_target
,
1643 size
<< growth_order
);
1647 void resize_target_update_count(struct cds_lfht
*ht
,
1648 unsigned long count
)
1650 count
= max(count
, MIN_TABLE_SIZE
);
1651 uatomic_set(&ht
->t
.resize_target
, count
);
1654 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1656 resize_target_update_count(ht
, new_size
);
1657 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1658 ht
->cds_lfht_rcu_thread_offline();
1659 pthread_mutex_lock(&ht
->resize_mutex
);
1660 _do_cds_lfht_resize(ht
);
1661 pthread_mutex_unlock(&ht
->resize_mutex
);
1662 ht
->cds_lfht_rcu_thread_online();
1666 void do_resize_cb(struct rcu_head
*head
)
1668 struct rcu_resize_work
*work
=
1669 caa_container_of(head
, struct rcu_resize_work
, head
);
1670 struct cds_lfht
*ht
= work
->ht
;
1672 ht
->cds_lfht_rcu_thread_offline();
1673 pthread_mutex_lock(&ht
->resize_mutex
);
1674 _do_cds_lfht_resize(ht
);
1675 pthread_mutex_unlock(&ht
->resize_mutex
);
1676 ht
->cds_lfht_rcu_thread_online();
1678 cmm_smp_mb(); /* finish resize before decrement */
1679 uatomic_dec(&ht
->in_progress_resize
);
1683 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1685 struct rcu_resize_work
*work
;
1686 unsigned long target_size
;
1688 target_size
= resize_target_update(ht
, size
, growth
);
1689 /* Store resize_target before read resize_initiated */
1691 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1692 uatomic_inc(&ht
->in_progress_resize
);
1693 cmm_smp_mb(); /* increment resize count before load destroy */
1694 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1695 uatomic_dec(&ht
->in_progress_resize
);
1698 work
= malloc(sizeof(*work
));
1700 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1701 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1705 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1708 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1709 unsigned long count
)
1711 struct rcu_resize_work
*work
;
1713 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1715 resize_target_update_count(ht
, count
);
1716 /* Store resize_target before read resize_initiated */
1718 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1719 uatomic_inc(&ht
->in_progress_resize
);
1720 cmm_smp_mb(); /* increment resize count before load destroy */
1721 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1722 uatomic_dec(&ht
->in_progress_resize
);
1725 work
= malloc(sizeof(*work
));
1727 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1728 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);