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 | |
143 #include <urcu-call-rcu.h>
144 #include <urcu/arch.h>
145 #include <urcu/uatomic.h>
146 #include <urcu/compiler.h>
150 #include "rculfhash.h"
153 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
155 #define dbg_printf(fmt, args...)
159 * Per-CPU split-counters lazily update the global counter each 1024
160 * addition/removal. It automatically keeps track of resize required.
161 * We use the bucket length as indicator for need to expand for small
162 * tables and machines lacking per-cpu data suppport.
164 #define COUNT_COMMIT_ORDER 10
165 #define CHAIN_LEN_TARGET 1
166 #define CHAIN_LEN_RESIZE_THRESHOLD 3
169 * Define the minimum table size.
171 #define MIN_TABLE_SIZE 1
173 #if (CAA_BITS_PER_LONG == 32)
174 #define MAX_TABLE_ORDER 32
176 #define MAX_TABLE_ORDER 64
180 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 #define MIN_PARTITION_PER_THREAD_ORDER 12
183 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
186 #define min(a, b) ((a) < (b) ? (a) : (b))
190 #define max(a, b) ((a) > (b) ? (a) : (b))
194 * The removed flag needs to be updated atomically with the pointer.
195 * It indicates that no node must attach to the node scheduled for
196 * removal, and that node garbage collection must be performed.
197 * The dummy flag does not require to be updated atomically with the
198 * pointer, but it is added as a pointer low bit flag to save space.
200 #define REMOVED_FLAG (1UL << 0)
201 #define DUMMY_FLAG (1UL << 1)
202 #define FLAGS_MASK ((1UL << 2) - 1)
204 /* Value of the end pointer. Should not interact with flags. */
205 #define END_VALUE NULL
207 struct ht_items_count
{
208 unsigned long add
, del
;
209 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
212 struct rcu_head head
;
213 struct _cds_lfht_node nodes
[0];
217 unsigned long size
; /* always a power of 2, shared (RCU) */
218 unsigned long resize_target
;
219 int resize_initiated
;
220 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
225 cds_lfht_hash_fct hash_fct
;
226 cds_lfht_compare_fct compare_fct
;
227 unsigned long hash_seed
;
230 * We need to put the work threads offline (QSBR) when taking this
231 * mutex, because we use synchronize_rcu within this mutex critical
232 * section, which waits on read-side critical sections, and could
233 * therefore cause grace-period deadlock if we hold off RCU G.P.
236 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
237 unsigned int in_progress_resize
, in_progress_destroy
;
238 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
239 void (*func
)(struct rcu_head
*head
));
240 void (*cds_lfht_synchronize_rcu
)(void);
241 void (*cds_lfht_rcu_read_lock
)(void);
242 void (*cds_lfht_rcu_read_unlock
)(void);
243 void (*cds_lfht_rcu_thread_offline
)(void);
244 void (*cds_lfht_rcu_thread_online
)(void);
245 void (*cds_lfht_rcu_register_thread
)(void);
246 void (*cds_lfht_rcu_unregister_thread
)(void);
247 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
248 long count
; /* global approximate item count */
249 struct ht_items_count
*percpu_count
; /* per-cpu item count */
252 struct rcu_resize_work
{
253 struct rcu_head head
;
257 struct partition_resize_work
{
258 struct rcu_head head
;
260 unsigned long i
, start
, len
;
261 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
262 unsigned long start
, unsigned long len
);
272 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
274 struct cds_lfht_node
*node
,
275 enum add_mode mode
, int dummy
);
278 * Algorithm to reverse bits in a word by lookup table, extended to
281 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
282 * Originally from Public Domain.
285 static const uint8_t BitReverseTable256
[256] =
287 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
288 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
289 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
290 R6(0), R6(2), R6(1), R6(3)
297 uint8_t bit_reverse_u8(uint8_t v
)
299 return BitReverseTable256
[v
];
302 static __attribute__((unused
))
303 uint32_t bit_reverse_u32(uint32_t v
)
305 return ((uint32_t) bit_reverse_u8(v
) << 24) |
306 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
307 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
308 ((uint32_t) bit_reverse_u8(v
>> 24));
311 static __attribute__((unused
))
312 uint64_t bit_reverse_u64(uint64_t v
)
314 return ((uint64_t) bit_reverse_u8(v
) << 56) |
315 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
316 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
317 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
318 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
319 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
320 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
321 ((uint64_t) bit_reverse_u8(v
>> 56));
325 unsigned long bit_reverse_ulong(unsigned long v
)
327 #if (CAA_BITS_PER_LONG == 32)
328 return bit_reverse_u32(v
);
330 return bit_reverse_u64(v
);
335 * fls: returns the position of the most significant bit.
336 * Returns 0 if no bit is set, else returns the position of the most
337 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
339 #if defined(__i386) || defined(__x86_64)
341 unsigned int fls_u32(uint32_t x
)
349 : "=r" (r
) : "rm" (x
));
355 #if defined(__x86_64)
357 unsigned int fls_u64(uint64_t x
)
365 : "=r" (r
) : "rm" (x
));
372 static __attribute__((unused
))
373 unsigned int fls_u64(uint64_t x
)
380 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
384 if (!(x
& 0xFFFF000000000000ULL
)) {
388 if (!(x
& 0xFF00000000000000ULL
)) {
392 if (!(x
& 0xF000000000000000ULL
)) {
396 if (!(x
& 0xC000000000000000ULL
)) {
400 if (!(x
& 0x8000000000000000ULL
)) {
409 static __attribute__((unused
))
410 unsigned int fls_u32(uint32_t x
)
416 if (!(x
& 0xFFFF0000U
)) {
420 if (!(x
& 0xFF000000U
)) {
424 if (!(x
& 0xF0000000U
)) {
428 if (!(x
& 0xC0000000U
)) {
432 if (!(x
& 0x80000000U
)) {
440 unsigned int fls_ulong(unsigned long x
)
442 #if (CAA_BITS_PER_lONG == 32)
449 int get_count_order_u32(uint32_t x
)
453 order
= fls_u32(x
) - 1;
459 int get_count_order_ulong(unsigned long x
)
463 order
= fls_ulong(x
) - 1;
470 #define poison_free(ptr) \
472 memset(ptr, 0x42, sizeof(*(ptr))); \
476 #define poison_free(ptr) free(ptr)
480 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
483 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
484 * available, then we support hash table item accounting.
485 * In the unfortunate event the number of CPUs reported would be
486 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
488 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
491 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
492 unsigned long count
);
494 static long nr_cpus_mask
= -1;
497 struct ht_items_count
*alloc_per_cpu_items_count(void)
499 struct ht_items_count
*count
;
501 switch (nr_cpus_mask
) {
508 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
514 * round up number of CPUs to next power of two, so we
515 * can use & for modulo.
517 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
518 nr_cpus_mask
= maxcpus
- 1;
522 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
527 void free_per_cpu_items_count(struct ht_items_count
*count
)
537 assert(nr_cpus_mask
>= 0);
538 cpu
= sched_getcpu();
539 if (unlikely(cpu
< 0))
542 return cpu
& nr_cpus_mask
;
546 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
548 unsigned long percpu_count
;
551 if (unlikely(!ht
->percpu_count
))
554 if (unlikely(cpu
< 0))
556 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
557 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
560 dbg_printf("add percpu %lu\n", percpu_count
);
561 count
= uatomic_add_return(&ht
->count
,
562 1UL << COUNT_COMMIT_ORDER
);
564 if (!(count
& (count
- 1))) {
565 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
567 dbg_printf("add set global %ld\n", count
);
568 cds_lfht_resize_lazy_count(ht
, size
,
569 count
>> (CHAIN_LEN_TARGET
- 1));
575 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
577 unsigned long percpu_count
;
580 if (unlikely(!ht
->percpu_count
))
583 if (unlikely(cpu
< 0))
585 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
586 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
589 dbg_printf("del percpu %lu\n", percpu_count
);
590 count
= uatomic_add_return(&ht
->count
,
591 -(1UL << COUNT_COMMIT_ORDER
));
593 if (!(count
& (count
- 1))) {
594 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
596 dbg_printf("del set global %ld\n", count
);
598 * Don't shrink table if the number of nodes is below a
601 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (nr_cpus_mask
+ 1))
603 cds_lfht_resize_lazy_count(ht
, size
,
604 count
>> (CHAIN_LEN_TARGET
- 1));
609 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
611 static const long nr_cpus_mask
= -2;
614 struct ht_items_count
*alloc_per_cpu_items_count(void)
620 void free_per_cpu_items_count(struct ht_items_count
*count
)
625 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
630 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
634 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
638 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
642 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
644 count
= uatomic_read(&ht
->count
);
646 * Use bucket-local length for small table expand and for
647 * environments lacking per-cpu data support.
649 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
652 dbg_printf("WARNING: large chain length: %u.\n",
654 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
655 cds_lfht_resize_lazy(ht
, size
,
656 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
660 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
662 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
666 int is_removed(struct cds_lfht_node
*node
)
668 return ((unsigned long) node
) & REMOVED_FLAG
;
672 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
674 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
678 int is_dummy(struct cds_lfht_node
*node
)
680 return ((unsigned long) node
) & DUMMY_FLAG
;
684 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
686 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
690 struct cds_lfht_node
*get_end(void)
692 return (struct cds_lfht_node
*) END_VALUE
;
696 int is_end(struct cds_lfht_node
*node
)
698 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
702 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
704 unsigned long old1
, old2
;
706 old1
= uatomic_read(ptr
);
711 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
716 void cds_lfht_free_level(struct rcu_head
*head
)
718 struct rcu_level
*l
=
719 caa_container_of(head
, struct rcu_level
, head
);
724 * Remove all logically deleted nodes from a bucket up to a certain node key.
727 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
729 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
731 assert(!is_dummy(dummy
));
732 assert(!is_removed(dummy
));
733 assert(!is_dummy(node
));
734 assert(!is_removed(node
));
737 /* We can always skip the dummy node initially */
738 iter
= rcu_dereference(iter_prev
->p
.next
);
739 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
741 * We should never be called with dummy (start of chain)
742 * and logically removed node (end of path compression
743 * marker) being the actual same node. This would be a
744 * bug in the algorithm implementation.
746 assert(dummy
!= node
);
748 if (unlikely(is_end(iter
)))
750 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
752 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
753 if (likely(is_removed(next
)))
755 iter_prev
= clear_flag(iter
);
758 assert(!is_removed(iter
));
760 new_next
= flag_dummy(clear_flag(next
));
762 new_next
= clear_flag(next
);
763 if (is_removed(iter
))
764 new_next
= flag_removed(new_next
);
765 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
771 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
772 struct cds_lfht_node
*old_node
,
773 struct cds_lfht_node
*ret_next
,
774 struct cds_lfht_node
*new_node
)
776 struct cds_lfht_node
*dummy
, *old_next
;
777 struct _cds_lfht_node
*lookup
;
779 unsigned long hash
, index
, order
;
781 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
784 assert(!is_removed(old_node
));
785 assert(!is_dummy(old_node
));
786 assert(!is_removed(new_node
));
787 assert(!is_dummy(new_node
));
788 assert(new_node
!= old_node
);
790 /* Insert after node to be replaced */
792 if (is_removed(old_next
)) {
794 * Too late, the old node has been removed under us
795 * between lookup and replace. Fail.
799 assert(!is_dummy(old_next
));
800 assert(new_node
!= clear_flag(old_next
));
801 new_node
->p
.next
= clear_flag(old_next
);
803 * Here is the whole trick for lock-free replace: we add
804 * the replacement node _after_ the node we want to
805 * replace by atomically setting its next pointer at the
806 * same time we set its removal flag. Given that
807 * the lookups/get next use an iterator aware of the
808 * next pointer, they will either skip the old node due
809 * to the removal flag and see the new node, or use
810 * the old node, but will not see the new one.
812 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
813 old_next
, flag_removed(new_node
));
814 } while (ret_next
!= old_next
);
816 /* We performed the replacement. */
820 * Ensure that the old node is not visible to readers anymore:
821 * lookup for the node, and remove it (along with any other
822 * logically removed node) if found.
824 hash
= bit_reverse_ulong(old_node
->p
.reverse_hash
);
826 index
= hash
& (size
- 1);
827 order
= get_count_order_ulong(index
+ 1);
828 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
829 dummy
= (struct cds_lfht_node
*) lookup
;
830 _cds_lfht_gc_bucket(dummy
, new_node
);
833 * Only the flagging action indicated that we (and no other)
834 * replaced the node from the hash table.
837 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
845 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
847 struct cds_lfht_node
*node
,
848 enum add_mode mode
, int dummy
)
850 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
851 *dummy_node
, *return_node
;
852 struct _cds_lfht_node
*lookup
;
853 unsigned long hash
, index
, order
;
855 assert(!is_dummy(node
));
856 assert(!is_removed(node
));
859 node
->p
.next
= flag_dummy(get_end());
860 return node
; /* Initial first add (head) */
862 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
864 uint32_t chain_len
= 0;
867 * iter_prev points to the non-removed node prior to the
870 index
= hash
& (size
- 1);
871 order
= get_count_order_ulong(index
+ 1);
872 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
873 iter_prev
= (struct cds_lfht_node
*) lookup
;
874 /* We can always skip the dummy node initially */
875 iter
= rcu_dereference(iter_prev
->p
.next
);
876 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
878 if (unlikely(is_end(iter
)))
880 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
882 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
883 if (unlikely(is_removed(next
)))
885 if ((mode
== ADD_UNIQUE
|| mode
== ADD_REPLACE
)
887 && !ht
->compare_fct(node
->key
, node
->key_len
,
888 clear_flag(iter
)->key
,
889 clear_flag(iter
)->key_len
)) {
890 if (mode
== ADD_UNIQUE
)
891 return clear_flag(iter
);
892 else /* mode == ADD_REPLACE */
895 /* Only account for identical reverse hash once */
896 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
898 check_resize(ht
, size
, ++chain_len
);
899 iter_prev
= clear_flag(iter
);
904 assert(node
!= clear_flag(iter
));
905 assert(!is_removed(iter_prev
));
906 assert(!is_removed(iter
));
907 assert(iter_prev
!= node
);
909 node
->p
.next
= clear_flag(iter
);
911 node
->p
.next
= flag_dummy(clear_flag(iter
));
913 new_node
= flag_dummy(node
);
916 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
918 continue; /* retry */
920 if (mode
== ADD_REPLACE
)
922 else /* ADD_DEFAULT and ADD_UNIQUE */
929 if (!_cds_lfht_replace(ht
, size
, clear_flag(iter
), next
,
931 return_node
= clear_flag(iter
);
932 goto end
; /* gc already done */
934 continue; /* retry */
938 assert(!is_removed(iter
));
940 new_next
= flag_dummy(clear_flag(next
));
942 new_next
= clear_flag(next
);
943 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
947 /* Garbage collect logically removed nodes in the bucket */
948 index
= hash
& (size
- 1);
949 order
= get_count_order_ulong(index
+ 1);
950 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
951 dummy_node
= (struct cds_lfht_node
*) lookup
;
952 _cds_lfht_gc_bucket(dummy_node
, node
);
958 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
959 struct cds_lfht_node
*node
,
962 struct cds_lfht_node
*dummy
, *next
, *old
;
963 struct _cds_lfht_node
*lookup
;
965 unsigned long hash
, index
, order
;
967 if (!node
) /* Return -ENOENT if asked to delete NULL node */
970 /* logically delete the node */
971 assert(!is_dummy(node
));
972 assert(!is_removed(node
));
973 old
= rcu_dereference(node
->p
.next
);
975 struct cds_lfht_node
*new_next
;
978 if (unlikely(is_removed(next
)))
981 assert(is_dummy(next
));
983 assert(!is_dummy(next
));
984 new_next
= flag_removed(next
);
985 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
986 } while (old
!= next
);
988 /* We performed the (logical) deletion. */
992 * Ensure that the node is not visible to readers anymore: lookup for
993 * the node, and remove it (along with any other logically removed node)
996 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
998 index
= hash
& (size
- 1);
999 order
= get_count_order_ulong(index
+ 1);
1000 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
1001 dummy
= (struct cds_lfht_node
*) lookup
;
1002 _cds_lfht_gc_bucket(dummy
, node
);
1005 * Only the flagging action indicated that we (and no other)
1006 * removed the node from the hash.
1009 assert(is_removed(rcu_dereference(node
->p
.next
)));
1017 void *partition_resize_thread(void *arg
)
1019 struct partition_resize_work
*work
= arg
;
1021 work
->ht
->cds_lfht_rcu_register_thread();
1022 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1023 work
->ht
->cds_lfht_rcu_unregister_thread();
1028 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1030 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1031 unsigned long start
, unsigned long len
))
1033 unsigned long partition_len
;
1034 struct partition_resize_work
*work
;
1036 unsigned long nr_threads
;
1037 pthread_t
*thread_id
;
1040 * Note: nr_cpus_mask + 1 is always power of 2.
1041 * We spawn just the number of threads we need to satisfy the minimum
1042 * partition size, up to the number of CPUs in the system.
1044 if (nr_cpus_mask
> 0) {
1045 nr_threads
= min(nr_cpus_mask
+ 1,
1046 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1050 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1051 work
= calloc(nr_threads
, sizeof(*work
));
1052 thread_id
= calloc(nr_threads
, sizeof(*thread_id
));
1054 for (thread
= 0; thread
< nr_threads
; thread
++) {
1055 work
[thread
].ht
= ht
;
1057 work
[thread
].len
= partition_len
;
1058 work
[thread
].start
= thread
* partition_len
;
1059 work
[thread
].fct
= fct
;
1060 ret
= pthread_create(&thread_id
[thread
], ht
->resize_attr
,
1061 partition_resize_thread
, &work
[thread
]);
1064 for (thread
= 0; thread
< nr_threads
; thread
++) {
1065 ret
= pthread_join(thread_id
[thread
], NULL
);
1073 * Holding RCU read lock to protect _cds_lfht_add against memory
1074 * reclaim that could be performed by other call_rcu worker threads (ABA
1077 * When we reach a certain length, we can split this population phase over
1078 * many worker threads, based on the number of CPUs available in the system.
1079 * This should therefore take care of not having the expand lagging behind too
1080 * many concurrent insertion threads by using the scheduler's ability to
1081 * schedule dummy node population fairly with insertions.
1084 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1085 unsigned long start
, unsigned long len
)
1089 ht
->cds_lfht_rcu_read_lock();
1090 for (j
= start
; j
< start
+ len
; j
++) {
1091 struct cds_lfht_node
*new_node
=
1092 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1094 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1095 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1096 new_node
->p
.reverse_hash
=
1097 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1098 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1099 new_node
, ADD_DEFAULT
, 1);
1100 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1103 ht
->cds_lfht_rcu_read_unlock();
1107 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1110 assert(nr_cpus_mask
!= -1);
1111 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1112 ht
->cds_lfht_rcu_thread_online();
1113 init_table_populate_partition(ht
, i
, 0, len
);
1114 ht
->cds_lfht_rcu_thread_offline();
1117 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1121 void init_table(struct cds_lfht
*ht
,
1122 unsigned long first_order
, unsigned long len_order
)
1124 unsigned long i
, end_order
;
1126 dbg_printf("init table: first_order %lu end_order %lu\n",
1127 first_order
, first_order
+ len_order
);
1128 end_order
= first_order
+ len_order
;
1129 for (i
= first_order
; i
< end_order
; i
++) {
1132 len
= !i
? 1 : 1UL << (i
- 1);
1133 dbg_printf("init order %lu len: %lu\n", i
, len
);
1135 /* Stop expand if the resize target changes under us */
1136 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1139 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1140 + (len
* sizeof(struct _cds_lfht_node
)));
1141 assert(ht
->t
.tbl
[i
]);
1144 * Set all dummy nodes reverse hash values for a level and
1145 * link all dummy nodes into the table.
1147 init_table_populate(ht
, i
, len
);
1150 * Update table size.
1152 cmm_smp_wmb(); /* populate data before RCU size */
1153 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1155 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1156 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1162 * Holding RCU read lock to protect _cds_lfht_remove against memory
1163 * reclaim that could be performed by other call_rcu worker threads (ABA
1165 * For a single level, we logically remove and garbage collect each node.
1167 * As a design choice, we perform logical removal and garbage collection on a
1168 * node-per-node basis to simplify this algorithm. We also assume keeping good
1169 * cache locality of the operation would overweight possible performance gain
1170 * that could be achieved by batching garbage collection for multiple levels.
1171 * However, this would have to be justified by benchmarks.
1173 * Concurrent removal and add operations are helping us perform garbage
1174 * collection of logically removed nodes. We guarantee that all logically
1175 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1176 * invoked to free a hole level of dummy nodes (after a grace period).
1178 * Logical removal and garbage collection can therefore be done in batch or on a
1179 * node-per-node basis, as long as the guarantee above holds.
1181 * When we reach a certain length, we can split this removal over many worker
1182 * threads, based on the number of CPUs available in the system. This should
1183 * take care of not letting resize process lag behind too many concurrent
1184 * updater threads actively inserting into the hash table.
1187 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1188 unsigned long start
, unsigned long len
)
1192 ht
->cds_lfht_rcu_read_lock();
1193 for (j
= start
; j
< start
+ len
; j
++) {
1194 struct cds_lfht_node
*fini_node
=
1195 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1197 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1198 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1199 fini_node
->p
.reverse_hash
=
1200 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1201 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1203 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1206 ht
->cds_lfht_rcu_read_unlock();
1210 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1213 assert(nr_cpus_mask
!= -1);
1214 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1215 ht
->cds_lfht_rcu_thread_online();
1216 remove_table_partition(ht
, i
, 0, len
);
1217 ht
->cds_lfht_rcu_thread_offline();
1220 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1224 void fini_table(struct cds_lfht
*ht
,
1225 unsigned long first_order
, unsigned long len_order
)
1229 dbg_printf("fini table: first_order %lu end_order %lu\n",
1230 first_order
, first_order
+ len_order
);
1231 end_order
= first_order
+ len_order
;
1232 assert(first_order
> 0);
1233 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1236 len
= !i
? 1 : 1UL << (i
- 1);
1237 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1239 /* Stop shrink if the resize target changes under us */
1240 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1243 cmm_smp_wmb(); /* populate data before RCU size */
1244 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1247 * We need to wait for all add operations to reach Q.S. (and
1248 * thus use the new table for lookups) before we can start
1249 * releasing the old dummy nodes. Otherwise their lookup will
1250 * return a logically removed node as insert position.
1252 ht
->cds_lfht_synchronize_rcu();
1255 * Set "removed" flag in dummy nodes about to be removed.
1256 * Unlink all now-logically-removed dummy node pointers.
1257 * Concurrent add/remove operation are helping us doing
1260 remove_table(ht
, i
, len
);
1262 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1264 dbg_printf("fini new size: %lu\n", 1UL << i
);
1265 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1270 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1271 cds_lfht_compare_fct compare_fct
,
1272 unsigned long hash_seed
,
1273 unsigned long init_size
,
1275 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1276 void (*func
)(struct rcu_head
*head
)),
1277 void (*cds_lfht_synchronize_rcu
)(void),
1278 void (*cds_lfht_rcu_read_lock
)(void),
1279 void (*cds_lfht_rcu_read_unlock
)(void),
1280 void (*cds_lfht_rcu_thread_offline
)(void),
1281 void (*cds_lfht_rcu_thread_online
)(void),
1282 void (*cds_lfht_rcu_register_thread
)(void),
1283 void (*cds_lfht_rcu_unregister_thread
)(void),
1284 pthread_attr_t
*attr
)
1286 struct cds_lfht
*ht
;
1287 unsigned long order
;
1289 /* init_size must be power of two */
1290 if (init_size
&& (init_size
& (init_size
- 1)))
1292 ht
= calloc(1, sizeof(struct cds_lfht
));
1294 ht
->hash_fct
= hash_fct
;
1295 ht
->compare_fct
= compare_fct
;
1296 ht
->hash_seed
= hash_seed
;
1297 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1298 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1299 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1300 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1301 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1302 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1303 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1304 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1305 ht
->resize_attr
= attr
;
1306 ht
->percpu_count
= alloc_per_cpu_items_count();
1307 /* this mutex should not nest in read-side C.S. */
1308 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1309 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1311 ht
->cds_lfht_rcu_thread_offline();
1312 pthread_mutex_lock(&ht
->resize_mutex
);
1313 ht
->t
.resize_target
= 1UL << (order
- 1);
1314 init_table(ht
, 0, order
);
1315 pthread_mutex_unlock(&ht
->resize_mutex
);
1316 ht
->cds_lfht_rcu_thread_online();
1320 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1321 struct cds_lfht_iter
*iter
)
1323 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1324 struct _cds_lfht_node
*lookup
;
1325 unsigned long hash
, reverse_hash
, index
, order
, size
;
1327 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1328 reverse_hash
= bit_reverse_ulong(hash
);
1330 size
= rcu_dereference(ht
->t
.size
);
1331 index
= hash
& (size
- 1);
1332 order
= get_count_order_ulong(index
+ 1);
1333 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1334 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1335 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1336 dummy_node
= (struct cds_lfht_node
*) lookup
;
1337 /* We can always skip the dummy node initially */
1338 node
= rcu_dereference(dummy_node
->p
.next
);
1339 node
= clear_flag(node
);
1341 if (unlikely(is_end(node
))) {
1345 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1349 next
= rcu_dereference(node
->p
.next
);
1350 if (likely(!is_removed(next
))
1352 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1355 node
= clear_flag(next
);
1357 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1362 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1364 struct cds_lfht_node
*node
, *next
;
1365 unsigned long reverse_hash
;
1370 reverse_hash
= node
->p
.reverse_hash
;
1372 key_len
= node
->key_len
;
1374 node
= clear_flag(next
);
1377 if (unlikely(is_end(node
))) {
1381 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1385 next
= rcu_dereference(node
->p
.next
);
1386 if (likely(!is_removed(next
))
1388 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1391 node
= clear_flag(next
);
1393 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1398 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1400 struct cds_lfht_node
*node
, *next
;
1402 node
= clear_flag(iter
->next
);
1404 if (unlikely(is_end(node
))) {
1408 next
= rcu_dereference(node
->p
.next
);
1409 if (likely(!is_removed(next
))
1410 && !is_dummy(next
)) {
1413 node
= clear_flag(next
);
1415 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1420 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1422 struct _cds_lfht_node
*lookup
;
1425 * Get next after first dummy node. The first dummy node is the
1426 * first node of the linked list.
1428 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1429 iter
->next
= lookup
->next
;
1430 cds_lfht_next(ht
, iter
);
1433 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1435 unsigned long hash
, size
;
1437 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1438 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1440 size
= rcu_dereference(ht
->t
.size
);
1441 (void) _cds_lfht_add(ht
, size
, node
, ADD_DEFAULT
, 0);
1442 ht_count_add(ht
, size
);
1445 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1446 struct cds_lfht_node
*node
)
1448 unsigned long hash
, size
;
1449 struct cds_lfht_node
*ret
;
1451 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1452 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1454 size
= rcu_dereference(ht
->t
.size
);
1455 ret
= _cds_lfht_add(ht
, size
, node
, ADD_UNIQUE
, 0);
1457 ht_count_add(ht
, size
);
1461 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1462 struct cds_lfht_node
*node
)
1464 unsigned long hash
, size
;
1465 struct cds_lfht_node
*ret
;
1467 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1468 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1470 size
= rcu_dereference(ht
->t
.size
);
1471 ret
= _cds_lfht_add(ht
, size
, node
, ADD_REPLACE
, 0);
1473 ht_count_add(ht
, size
);
1477 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1478 struct cds_lfht_node
*new_node
)
1482 size
= rcu_dereference(ht
->t
.size
);
1483 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1487 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1492 size
= rcu_dereference(ht
->t
.size
);
1493 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1495 ht_count_del(ht
, size
);
1500 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1502 struct cds_lfht_node
*node
;
1503 struct _cds_lfht_node
*lookup
;
1504 unsigned long order
, i
, size
;
1506 /* Check that the table is empty */
1507 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1508 node
= (struct cds_lfht_node
*) lookup
;
1510 node
= clear_flag(node
)->p
.next
;
1511 if (!is_dummy(node
))
1513 assert(!is_removed(node
));
1514 } while (!is_end(node
));
1516 * size accessed without rcu_dereference because hash table is
1520 /* Internal sanity check: all nodes left should be dummy */
1521 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1524 len
= !order
? 1 : 1UL << (order
- 1);
1525 for (i
= 0; i
< len
; i
++) {
1526 dbg_printf("delete order %lu i %lu hash %lu\n",
1528 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1529 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1531 poison_free(ht
->t
.tbl
[order
]);
1537 * Should only be called when no more concurrent readers nor writers can
1538 * possibly access the table.
1540 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1544 /* Wait for in-flight resize operations to complete */
1545 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1546 while (uatomic_read(&ht
->in_progress_resize
))
1547 poll(NULL
, 0, 100); /* wait for 100ms */
1548 ret
= cds_lfht_delete_dummy(ht
);
1551 free_per_cpu_items_count(ht
->percpu_count
);
1553 *attr
= ht
->resize_attr
;
1558 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1559 long *approx_before
,
1560 unsigned long *count
,
1561 unsigned long *removed
,
1564 struct cds_lfht_node
*node
, *next
;
1565 struct _cds_lfht_node
*lookup
;
1566 unsigned long nr_dummy
= 0;
1569 if (nr_cpus_mask
>= 0) {
1572 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1573 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1574 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1581 /* Count non-dummy nodes in the table */
1582 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1583 node
= (struct cds_lfht_node
*) lookup
;
1585 next
= rcu_dereference(node
->p
.next
);
1586 if (is_removed(next
)) {
1587 if (!is_dummy(next
))
1591 } else if (!is_dummy(next
))
1595 node
= clear_flag(next
);
1596 } while (!is_end(node
));
1597 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1599 if (nr_cpus_mask
>= 0) {
1602 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1603 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1604 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1609 /* called with resize mutex held */
1611 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1612 unsigned long old_size
, unsigned long new_size
)
1614 unsigned long old_order
, new_order
;
1616 old_order
= get_count_order_ulong(old_size
) + 1;
1617 new_order
= get_count_order_ulong(new_size
) + 1;
1618 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1619 old_size
, old_order
, new_size
, new_order
);
1620 assert(new_size
> old_size
);
1621 init_table(ht
, old_order
, new_order
- old_order
);
1624 /* called with resize mutex held */
1626 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1627 unsigned long old_size
, unsigned long new_size
)
1629 unsigned long old_order
, new_order
;
1631 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1632 old_order
= get_count_order_ulong(old_size
) + 1;
1633 new_order
= get_count_order_ulong(new_size
) + 1;
1634 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1635 old_size
, old_order
, new_size
, new_order
);
1636 assert(new_size
< old_size
);
1638 /* Remove and unlink all dummy nodes to remove. */
1639 fini_table(ht
, new_order
, old_order
- new_order
);
1643 /* called with resize mutex held */
1645 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1647 unsigned long new_size
, old_size
;
1650 * Resize table, re-do if the target size has changed under us.
1653 ht
->t
.resize_initiated
= 1;
1654 old_size
= ht
->t
.size
;
1655 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1656 if (old_size
< new_size
)
1657 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1658 else if (old_size
> new_size
)
1659 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1660 ht
->t
.resize_initiated
= 0;
1661 /* write resize_initiated before read resize_target */
1663 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1667 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1670 return _uatomic_max(&ht
->t
.resize_target
,
1671 size
<< growth_order
);
1675 void resize_target_update_count(struct cds_lfht
*ht
,
1676 unsigned long count
)
1678 count
= max(count
, MIN_TABLE_SIZE
);
1679 uatomic_set(&ht
->t
.resize_target
, count
);
1682 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1684 resize_target_update_count(ht
, new_size
);
1685 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1686 ht
->cds_lfht_rcu_thread_offline();
1687 pthread_mutex_lock(&ht
->resize_mutex
);
1688 _do_cds_lfht_resize(ht
);
1689 pthread_mutex_unlock(&ht
->resize_mutex
);
1690 ht
->cds_lfht_rcu_thread_online();
1694 void do_resize_cb(struct rcu_head
*head
)
1696 struct rcu_resize_work
*work
=
1697 caa_container_of(head
, struct rcu_resize_work
, head
);
1698 struct cds_lfht
*ht
= work
->ht
;
1700 ht
->cds_lfht_rcu_thread_offline();
1701 pthread_mutex_lock(&ht
->resize_mutex
);
1702 _do_cds_lfht_resize(ht
);
1703 pthread_mutex_unlock(&ht
->resize_mutex
);
1704 ht
->cds_lfht_rcu_thread_online();
1706 cmm_smp_mb(); /* finish resize before decrement */
1707 uatomic_dec(&ht
->in_progress_resize
);
1711 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1713 struct rcu_resize_work
*work
;
1714 unsigned long target_size
;
1716 target_size
= resize_target_update(ht
, size
, growth
);
1717 /* Store resize_target before read resize_initiated */
1719 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1720 uatomic_inc(&ht
->in_progress_resize
);
1721 cmm_smp_mb(); /* increment resize count before calling it */
1722 work
= malloc(sizeof(*work
));
1724 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1725 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1729 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1732 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1733 unsigned long count
)
1735 struct rcu_resize_work
*work
;
1737 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1739 resize_target_update_count(ht
, count
);
1740 /* Store resize_target before read resize_initiated */
1742 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1743 uatomic_inc(&ht
->in_progress_resize
);
1744 cmm_smp_mb(); /* increment resize count before calling it */
1745 work
= malloc(sizeof(*work
));
1747 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1748 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);