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/compiler.h>
148 #include <urcu/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 /* Note: manually update allocation length when adding a field */
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
{
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)
450 * Return the minimum order for which x <= (1UL << order).
451 * Return -1 if x is 0.
453 int get_count_order_u32(uint32_t x
)
458 return fls_u32(x
- 1);
462 * Return the minimum order for which x <= (1UL << order).
463 * Return -1 if x is 0.
465 int get_count_order_ulong(unsigned long x
)
470 return fls_ulong(x
- 1);
474 #define poison_free(ptr) \
476 memset(ptr, 0x42, sizeof(*(ptr))); \
480 #define poison_free(ptr) free(ptr)
484 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
487 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
488 * available, then we support hash table item accounting.
489 * In the unfortunate event the number of CPUs reported would be
490 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
492 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
495 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
496 unsigned long count
);
498 static long nr_cpus_mask
= -1;
501 struct ht_items_count
*alloc_per_cpu_items_count(void)
503 struct ht_items_count
*count
;
505 switch (nr_cpus_mask
) {
512 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
518 * round up number of CPUs to next power of two, so we
519 * can use & for modulo.
521 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
522 nr_cpus_mask
= maxcpus
- 1;
526 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
531 void free_per_cpu_items_count(struct ht_items_count
*count
)
541 assert(nr_cpus_mask
>= 0);
542 cpu
= sched_getcpu();
543 if (unlikely(cpu
< 0))
546 return cpu
& nr_cpus_mask
;
550 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
552 unsigned long percpu_count
;
555 if (unlikely(!ht
->percpu_count
))
558 if (unlikely(cpu
< 0))
560 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
561 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
564 dbg_printf("add percpu %lu\n", percpu_count
);
565 count
= uatomic_add_return(&ht
->count
,
566 1UL << COUNT_COMMIT_ORDER
);
568 if (!(count
& (count
- 1))) {
569 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
571 dbg_printf("add set global %ld\n", count
);
572 cds_lfht_resize_lazy_count(ht
, size
,
573 count
>> (CHAIN_LEN_TARGET
- 1));
579 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
581 unsigned long percpu_count
;
584 if (unlikely(!ht
->percpu_count
))
587 if (unlikely(cpu
< 0))
589 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
590 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
593 dbg_printf("del percpu %lu\n", percpu_count
);
594 count
= uatomic_add_return(&ht
->count
,
595 -(1UL << COUNT_COMMIT_ORDER
));
597 if (!(count
& (count
- 1))) {
598 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
600 dbg_printf("del set global %ld\n", count
);
602 * Don't shrink table if the number of nodes is below a
605 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (nr_cpus_mask
+ 1))
607 cds_lfht_resize_lazy_count(ht
, size
,
608 count
>> (CHAIN_LEN_TARGET
- 1));
613 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
615 static const long nr_cpus_mask
= -2;
618 struct ht_items_count
*alloc_per_cpu_items_count(void)
624 void free_per_cpu_items_count(struct ht_items_count
*count
)
629 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
634 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
638 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
642 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
646 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
648 count
= uatomic_read(&ht
->count
);
650 * Use bucket-local length for small table expand and for
651 * environments lacking per-cpu data support.
653 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
656 dbg_printf("WARNING: large chain length: %u.\n",
658 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
659 cds_lfht_resize_lazy(ht
, size
,
660 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
664 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
666 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
670 int is_removed(struct cds_lfht_node
*node
)
672 return ((unsigned long) node
) & REMOVED_FLAG
;
676 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
678 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
682 int is_dummy(struct cds_lfht_node
*node
)
684 return ((unsigned long) node
) & DUMMY_FLAG
;
688 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
690 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
694 struct cds_lfht_node
*get_end(void)
696 return (struct cds_lfht_node
*) END_VALUE
;
700 int is_end(struct cds_lfht_node
*node
)
702 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
706 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
708 unsigned long old1
, old2
;
710 old1
= uatomic_read(ptr
);
715 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
720 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
723 unsigned long index
, order
;
726 index
= hash
& (size
- 1);
727 order
= get_count_order_ulong(index
+ 1);
729 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
730 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
732 return &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
736 * Remove all logically deleted nodes from a bucket up to a certain node key.
739 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
741 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
743 assert(!is_dummy(dummy
));
744 assert(!is_removed(dummy
));
745 assert(!is_dummy(node
));
746 assert(!is_removed(node
));
749 /* We can always skip the dummy node initially */
750 iter
= rcu_dereference(iter_prev
->p
.next
);
751 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
753 * We should never be called with dummy (start of chain)
754 * and logically removed node (end of path compression
755 * marker) being the actual same node. This would be a
756 * bug in the algorithm implementation.
758 assert(dummy
!= node
);
760 if (unlikely(is_end(iter
)))
762 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
764 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
765 if (likely(is_removed(next
)))
767 iter_prev
= clear_flag(iter
);
770 assert(!is_removed(iter
));
772 new_next
= flag_dummy(clear_flag(next
));
774 new_next
= clear_flag(next
);
775 if (is_removed(iter
))
776 new_next
= flag_removed(new_next
);
777 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
783 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
784 struct cds_lfht_node
*old_node
,
785 struct cds_lfht_node
*ret_next
,
786 struct cds_lfht_node
*new_node
)
788 struct cds_lfht_node
*dummy
, *old_next
;
789 struct _cds_lfht_node
*lookup
;
792 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
795 assert(!is_removed(old_node
));
796 assert(!is_dummy(old_node
));
797 assert(!is_removed(new_node
));
798 assert(!is_dummy(new_node
));
799 assert(new_node
!= old_node
);
801 /* Insert after node to be replaced */
803 if (is_removed(old_next
)) {
805 * Too late, the old node has been removed under us
806 * between lookup and replace. Fail.
810 assert(!is_dummy(old_next
));
811 assert(new_node
!= clear_flag(old_next
));
812 new_node
->p
.next
= clear_flag(old_next
);
814 * Here is the whole trick for lock-free replace: we add
815 * the replacement node _after_ the node we want to
816 * replace by atomically setting its next pointer at the
817 * same time we set its removal flag. Given that
818 * the lookups/get next use an iterator aware of the
819 * next pointer, they will either skip the old node due
820 * to the removal flag and see the new node, or use
821 * the old node, but will not see the new one.
823 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
824 old_next
, flag_removed(new_node
));
825 } while (ret_next
!= old_next
);
827 /* We performed the replacement. */
831 * Ensure that the old node is not visible to readers anymore:
832 * lookup for the node, and remove it (along with any other
833 * logically removed node) if found.
835 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
836 dummy
= (struct cds_lfht_node
*) lookup
;
837 _cds_lfht_gc_bucket(dummy
, new_node
);
840 * Only the flagging action indicated that we (and no other)
841 * replaced the node from the hash table.
844 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
852 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
854 struct cds_lfht_node
*node
,
855 enum add_mode mode
, int dummy
)
857 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
858 *dummy_node
, *return_node
;
859 struct _cds_lfht_node
*lookup
;
861 assert(!is_dummy(node
));
862 assert(!is_removed(node
));
865 node
->p
.next
= flag_dummy(get_end());
866 return node
; /* Initial first add (head) */
868 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
870 uint32_t chain_len
= 0;
873 * iter_prev points to the non-removed node prior to the
876 iter_prev
= (struct cds_lfht_node
*) lookup
;
877 /* We can always skip the dummy node initially */
878 iter
= rcu_dereference(iter_prev
->p
.next
);
879 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
881 if (unlikely(is_end(iter
)))
883 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
885 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
886 if (unlikely(is_removed(next
)))
888 if ((mode
== ADD_UNIQUE
|| mode
== ADD_REPLACE
)
890 && !ht
->compare_fct(node
->key
, node
->key_len
,
891 clear_flag(iter
)->key
,
892 clear_flag(iter
)->key_len
)) {
893 if (mode
== ADD_UNIQUE
)
894 return clear_flag(iter
);
895 else /* mode == ADD_REPLACE */
898 /* Only account for identical reverse hash once */
899 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
901 check_resize(ht
, size
, ++chain_len
);
902 iter_prev
= clear_flag(iter
);
907 assert(node
!= clear_flag(iter
));
908 assert(!is_removed(iter_prev
));
909 assert(!is_removed(iter
));
910 assert(iter_prev
!= node
);
912 node
->p
.next
= clear_flag(iter
);
914 node
->p
.next
= flag_dummy(clear_flag(iter
));
916 new_node
= flag_dummy(node
);
919 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
921 continue; /* retry */
923 if (mode
== ADD_REPLACE
)
925 else /* ADD_DEFAULT and ADD_UNIQUE */
932 if (!_cds_lfht_replace(ht
, size
, clear_flag(iter
), next
,
934 return_node
= clear_flag(iter
);
935 goto end
; /* gc already done */
937 continue; /* retry */
941 assert(!is_removed(iter
));
943 new_next
= flag_dummy(clear_flag(next
));
945 new_next
= clear_flag(next
);
946 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
950 /* Garbage collect logically removed nodes in the bucket */
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
;
966 if (!node
) /* Return -ENOENT if asked to delete NULL node */
969 /* logically delete the node */
970 assert(!is_dummy(node
));
971 assert(!is_removed(node
));
972 old
= rcu_dereference(node
->p
.next
);
974 struct cds_lfht_node
*new_next
;
977 if (unlikely(is_removed(next
)))
980 assert(is_dummy(next
));
982 assert(!is_dummy(next
));
983 new_next
= flag_removed(next
);
984 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
985 } while (old
!= next
);
987 /* We performed the (logical) deletion. */
991 * Ensure that the node is not visible to readers anymore: lookup for
992 * the node, and remove it (along with any other logically removed node)
995 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
996 dummy
= (struct cds_lfht_node
*) lookup
;
997 _cds_lfht_gc_bucket(dummy
, node
);
1000 * Only the flagging action indicated that we (and no other)
1001 * removed the node from the hash.
1004 assert(is_removed(rcu_dereference(node
->p
.next
)));
1012 void *partition_resize_thread(void *arg
)
1014 struct partition_resize_work
*work
= arg
;
1016 work
->ht
->cds_lfht_rcu_register_thread();
1017 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1018 work
->ht
->cds_lfht_rcu_unregister_thread();
1023 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1025 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1026 unsigned long start
, unsigned long len
))
1028 unsigned long partition_len
;
1029 struct partition_resize_work
*work
;
1031 unsigned long nr_threads
;
1034 * Note: nr_cpus_mask + 1 is always power of 2.
1035 * We spawn just the number of threads we need to satisfy the minimum
1036 * partition size, up to the number of CPUs in the system.
1038 if (nr_cpus_mask
> 0) {
1039 nr_threads
= min(nr_cpus_mask
+ 1,
1040 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1044 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1045 work
= calloc(nr_threads
, sizeof(*work
));
1047 for (thread
= 0; thread
< nr_threads
; thread
++) {
1048 work
[thread
].ht
= ht
;
1050 work
[thread
].len
= partition_len
;
1051 work
[thread
].start
= thread
* partition_len
;
1052 work
[thread
].fct
= fct
;
1053 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1054 partition_resize_thread
, &work
[thread
]);
1057 for (thread
= 0; thread
< nr_threads
; thread
++) {
1058 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1065 * Holding RCU read lock to protect _cds_lfht_add against memory
1066 * reclaim that could be performed by other call_rcu worker threads (ABA
1069 * When we reach a certain length, we can split this population phase over
1070 * many worker threads, based on the number of CPUs available in the system.
1071 * This should therefore take care of not having the expand lagging behind too
1072 * many concurrent insertion threads by using the scheduler's ability to
1073 * schedule dummy node population fairly with insertions.
1076 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1077 unsigned long start
, unsigned long len
)
1081 ht
->cds_lfht_rcu_read_lock();
1082 for (j
= start
; j
< start
+ len
; j
++) {
1083 struct cds_lfht_node
*new_node
=
1084 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1086 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1087 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1088 new_node
->p
.reverse_hash
=
1089 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1090 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1091 new_node
, ADD_DEFAULT
, 1);
1093 ht
->cds_lfht_rcu_read_unlock();
1097 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1100 assert(nr_cpus_mask
!= -1);
1101 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1102 ht
->cds_lfht_rcu_thread_online();
1103 init_table_populate_partition(ht
, i
, 0, len
);
1104 ht
->cds_lfht_rcu_thread_offline();
1107 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1111 void init_table(struct cds_lfht
*ht
,
1112 unsigned long first_order
, unsigned long len_order
)
1114 unsigned long i
, end_order
;
1116 dbg_printf("init table: first_order %lu end_order %lu\n",
1117 first_order
, first_order
+ len_order
);
1118 end_order
= first_order
+ len_order
;
1119 for (i
= first_order
; i
< end_order
; i
++) {
1122 len
= !i
? 1 : 1UL << (i
- 1);
1123 dbg_printf("init order %lu len: %lu\n", i
, len
);
1125 /* Stop expand if the resize target changes under us */
1126 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1129 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1130 assert(ht
->t
.tbl
[i
]);
1133 * Set all dummy nodes reverse hash values for a level and
1134 * link all dummy nodes into the table.
1136 init_table_populate(ht
, i
, len
);
1139 * Update table size.
1141 cmm_smp_wmb(); /* populate data before RCU size */
1142 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1144 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1145 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1151 * Holding RCU read lock to protect _cds_lfht_remove against memory
1152 * reclaim that could be performed by other call_rcu worker threads (ABA
1154 * For a single level, we logically remove and garbage collect each node.
1156 * As a design choice, we perform logical removal and garbage collection on a
1157 * node-per-node basis to simplify this algorithm. We also assume keeping good
1158 * cache locality of the operation would overweight possible performance gain
1159 * that could be achieved by batching garbage collection for multiple levels.
1160 * However, this would have to be justified by benchmarks.
1162 * Concurrent removal and add operations are helping us perform garbage
1163 * collection of logically removed nodes. We guarantee that all logically
1164 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1165 * invoked to free a hole level of dummy nodes (after a grace period).
1167 * Logical removal and garbage collection can therefore be done in batch or on a
1168 * node-per-node basis, as long as the guarantee above holds.
1170 * When we reach a certain length, we can split this removal over many worker
1171 * threads, based on the number of CPUs available in the system. This should
1172 * take care of not letting resize process lag behind too many concurrent
1173 * updater threads actively inserting into the hash table.
1176 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1177 unsigned long start
, unsigned long len
)
1181 ht
->cds_lfht_rcu_read_lock();
1182 for (j
= start
; j
< start
+ len
; j
++) {
1183 struct cds_lfht_node
*fini_node
=
1184 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1186 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1187 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1188 fini_node
->p
.reverse_hash
=
1189 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1190 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1193 ht
->cds_lfht_rcu_read_unlock();
1197 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1200 assert(nr_cpus_mask
!= -1);
1201 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1202 ht
->cds_lfht_rcu_thread_online();
1203 remove_table_partition(ht
, i
, 0, len
);
1204 ht
->cds_lfht_rcu_thread_offline();
1207 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1211 void fini_table(struct cds_lfht
*ht
,
1212 unsigned long first_order
, unsigned long len_order
)
1215 void *free_by_rcu
= NULL
;
1217 dbg_printf("fini table: first_order %lu end_order %lu\n",
1218 first_order
, first_order
+ len_order
);
1219 end_order
= first_order
+ len_order
;
1220 assert(first_order
> 0);
1221 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1224 len
= !i
? 1 : 1UL << (i
- 1);
1225 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1227 /* Stop shrink if the resize target changes under us */
1228 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1231 cmm_smp_wmb(); /* populate data before RCU size */
1232 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1235 * We need to wait for all add operations to reach Q.S. (and
1236 * thus use the new table for lookups) before we can start
1237 * releasing the old dummy nodes. Otherwise their lookup will
1238 * return a logically removed node as insert position.
1240 ht
->cds_lfht_synchronize_rcu();
1245 * Set "removed" flag in dummy nodes about to be removed.
1246 * Unlink all now-logically-removed dummy node pointers.
1247 * Concurrent add/remove operation are helping us doing
1250 remove_table(ht
, i
, len
);
1252 free_by_rcu
= ht
->t
.tbl
[i
];
1254 dbg_printf("fini new size: %lu\n", 1UL << i
);
1255 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1260 ht
->cds_lfht_synchronize_rcu();
1265 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1266 cds_lfht_compare_fct compare_fct
,
1267 unsigned long hash_seed
,
1268 unsigned long init_size
,
1270 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1271 void (*func
)(struct rcu_head
*head
)),
1272 void (*cds_lfht_synchronize_rcu
)(void),
1273 void (*cds_lfht_rcu_read_lock
)(void),
1274 void (*cds_lfht_rcu_read_unlock
)(void),
1275 void (*cds_lfht_rcu_thread_offline
)(void),
1276 void (*cds_lfht_rcu_thread_online
)(void),
1277 void (*cds_lfht_rcu_register_thread
)(void),
1278 void (*cds_lfht_rcu_unregister_thread
)(void),
1279 pthread_attr_t
*attr
)
1281 struct cds_lfht
*ht
;
1282 unsigned long order
;
1284 /* init_size must be power of two */
1285 if (init_size
&& (init_size
& (init_size
- 1)))
1287 ht
= calloc(1, sizeof(struct cds_lfht
));
1289 ht
->hash_fct
= hash_fct
;
1290 ht
->compare_fct
= compare_fct
;
1291 ht
->hash_seed
= hash_seed
;
1292 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1293 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1294 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1295 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1296 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1297 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1298 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1299 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1300 ht
->resize_attr
= attr
;
1301 ht
->percpu_count
= alloc_per_cpu_items_count();
1302 /* this mutex should not nest in read-side C.S. */
1303 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1304 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1306 ht
->cds_lfht_rcu_thread_offline();
1307 pthread_mutex_lock(&ht
->resize_mutex
);
1308 ht
->t
.resize_target
= 1UL << (order
- 1);
1309 init_table(ht
, 0, order
);
1310 pthread_mutex_unlock(&ht
->resize_mutex
);
1311 ht
->cds_lfht_rcu_thread_online();
1315 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1316 struct cds_lfht_iter
*iter
)
1318 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1319 struct _cds_lfht_node
*lookup
;
1320 unsigned long hash
, reverse_hash
, size
;
1322 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1323 reverse_hash
= bit_reverse_ulong(hash
);
1325 size
= rcu_dereference(ht
->t
.size
);
1326 lookup
= lookup_bucket(ht
, size
, hash
);
1327 dummy_node
= (struct cds_lfht_node
*) lookup
;
1328 /* We can always skip the dummy node initially */
1329 node
= rcu_dereference(dummy_node
->p
.next
);
1330 node
= clear_flag(node
);
1332 if (unlikely(is_end(node
))) {
1336 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1340 next
= rcu_dereference(node
->p
.next
);
1341 if (likely(!is_removed(next
))
1343 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1346 node
= clear_flag(next
);
1348 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1353 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1355 struct cds_lfht_node
*node
, *next
;
1356 unsigned long reverse_hash
;
1361 reverse_hash
= node
->p
.reverse_hash
;
1363 key_len
= node
->key_len
;
1365 node
= clear_flag(next
);
1368 if (unlikely(is_end(node
))) {
1372 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1376 next
= rcu_dereference(node
->p
.next
);
1377 if (likely(!is_removed(next
))
1379 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1382 node
= clear_flag(next
);
1384 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1389 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1391 struct cds_lfht_node
*node
, *next
;
1393 node
= clear_flag(iter
->next
);
1395 if (unlikely(is_end(node
))) {
1399 next
= rcu_dereference(node
->p
.next
);
1400 if (likely(!is_removed(next
))
1401 && !is_dummy(next
)) {
1404 node
= clear_flag(next
);
1406 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1411 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1413 struct _cds_lfht_node
*lookup
;
1416 * Get next after first dummy node. The first dummy node is the
1417 * first node of the linked list.
1419 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1420 iter
->next
= lookup
->next
;
1421 cds_lfht_next(ht
, iter
);
1424 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1426 unsigned long hash
, size
;
1428 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1429 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1431 size
= rcu_dereference(ht
->t
.size
);
1432 (void) _cds_lfht_add(ht
, size
, node
, ADD_DEFAULT
, 0);
1433 ht_count_add(ht
, size
);
1436 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1437 struct cds_lfht_node
*node
)
1439 unsigned long hash
, size
;
1440 struct cds_lfht_node
*ret
;
1442 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1443 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1445 size
= rcu_dereference(ht
->t
.size
);
1446 ret
= _cds_lfht_add(ht
, size
, node
, ADD_UNIQUE
, 0);
1448 ht_count_add(ht
, size
);
1452 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1453 struct cds_lfht_node
*node
)
1455 unsigned long hash
, size
;
1456 struct cds_lfht_node
*ret
;
1458 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1459 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1461 size
= rcu_dereference(ht
->t
.size
);
1462 ret
= _cds_lfht_add(ht
, size
, node
, ADD_REPLACE
, 0);
1464 ht_count_add(ht
, size
);
1468 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1469 struct cds_lfht_node
*new_node
)
1473 size
= rcu_dereference(ht
->t
.size
);
1474 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1478 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1483 size
= rcu_dereference(ht
->t
.size
);
1484 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1486 ht_count_del(ht
, size
);
1491 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1493 struct cds_lfht_node
*node
;
1494 struct _cds_lfht_node
*lookup
;
1495 unsigned long order
, i
, size
;
1497 /* Check that the table is empty */
1498 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1499 node
= (struct cds_lfht_node
*) lookup
;
1501 node
= clear_flag(node
)->p
.next
;
1502 if (!is_dummy(node
))
1504 assert(!is_removed(node
));
1505 } while (!is_end(node
));
1507 * size accessed without rcu_dereference because hash table is
1511 /* Internal sanity check: all nodes left should be dummy */
1512 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1515 len
= !order
? 1 : 1UL << (order
- 1);
1516 for (i
= 0; i
< len
; i
++) {
1517 dbg_printf("delete order %lu i %lu hash %lu\n",
1519 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1520 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1522 poison_free(ht
->t
.tbl
[order
]);
1528 * Should only be called when no more concurrent readers nor writers can
1529 * possibly access the table.
1531 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1535 /* Wait for in-flight resize operations to complete */
1536 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1537 cmm_smp_mb(); /* Store destroy before load resize */
1538 while (uatomic_read(&ht
->in_progress_resize
))
1539 poll(NULL
, 0, 100); /* wait for 100ms */
1540 ret
= cds_lfht_delete_dummy(ht
);
1543 free_per_cpu_items_count(ht
->percpu_count
);
1545 *attr
= ht
->resize_attr
;
1550 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1551 long *approx_before
,
1552 unsigned long *count
,
1553 unsigned long *removed
,
1556 struct cds_lfht_node
*node
, *next
;
1557 struct _cds_lfht_node
*lookup
;
1558 unsigned long nr_dummy
= 0;
1561 if (nr_cpus_mask
>= 0) {
1564 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1565 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1566 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1573 /* Count non-dummy nodes in the table */
1574 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1575 node
= (struct cds_lfht_node
*) lookup
;
1577 next
= rcu_dereference(node
->p
.next
);
1578 if (is_removed(next
)) {
1579 if (!is_dummy(next
))
1583 } else if (!is_dummy(next
))
1587 node
= clear_flag(next
);
1588 } while (!is_end(node
));
1589 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1591 if (nr_cpus_mask
>= 0) {
1594 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1595 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1596 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1601 /* called with resize mutex held */
1603 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1604 unsigned long old_size
, unsigned long new_size
)
1606 unsigned long old_order
, new_order
;
1608 old_order
= get_count_order_ulong(old_size
) + 1;
1609 new_order
= get_count_order_ulong(new_size
) + 1;
1610 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1611 old_size
, old_order
, new_size
, new_order
);
1612 assert(new_size
> old_size
);
1613 init_table(ht
, old_order
, new_order
- old_order
);
1616 /* called with resize mutex held */
1618 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1619 unsigned long old_size
, unsigned long new_size
)
1621 unsigned long old_order
, new_order
;
1623 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1624 old_order
= get_count_order_ulong(old_size
) + 1;
1625 new_order
= get_count_order_ulong(new_size
) + 1;
1626 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1627 old_size
, old_order
, new_size
, new_order
);
1628 assert(new_size
< old_size
);
1630 /* Remove and unlink all dummy nodes to remove. */
1631 fini_table(ht
, new_order
, old_order
- new_order
);
1635 /* called with resize mutex held */
1637 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1639 unsigned long new_size
, old_size
;
1642 * Resize table, re-do if the target size has changed under us.
1645 assert(uatomic_read(&ht
->in_progress_resize
));
1646 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1648 ht
->t
.resize_initiated
= 1;
1649 old_size
= ht
->t
.size
;
1650 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1651 if (old_size
< new_size
)
1652 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1653 else if (old_size
> new_size
)
1654 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1655 ht
->t
.resize_initiated
= 0;
1656 /* write resize_initiated before read resize_target */
1658 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1662 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1665 return _uatomic_max(&ht
->t
.resize_target
,
1666 size
<< growth_order
);
1670 void resize_target_update_count(struct cds_lfht
*ht
,
1671 unsigned long count
)
1673 count
= max(count
, MIN_TABLE_SIZE
);
1674 uatomic_set(&ht
->t
.resize_target
, count
);
1677 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1679 resize_target_update_count(ht
, new_size
);
1680 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1681 ht
->cds_lfht_rcu_thread_offline();
1682 pthread_mutex_lock(&ht
->resize_mutex
);
1683 _do_cds_lfht_resize(ht
);
1684 pthread_mutex_unlock(&ht
->resize_mutex
);
1685 ht
->cds_lfht_rcu_thread_online();
1689 void do_resize_cb(struct rcu_head
*head
)
1691 struct rcu_resize_work
*work
=
1692 caa_container_of(head
, struct rcu_resize_work
, head
);
1693 struct cds_lfht
*ht
= work
->ht
;
1695 ht
->cds_lfht_rcu_thread_offline();
1696 pthread_mutex_lock(&ht
->resize_mutex
);
1697 _do_cds_lfht_resize(ht
);
1698 pthread_mutex_unlock(&ht
->resize_mutex
);
1699 ht
->cds_lfht_rcu_thread_online();
1701 cmm_smp_mb(); /* finish resize before decrement */
1702 uatomic_dec(&ht
->in_progress_resize
);
1706 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1708 struct rcu_resize_work
*work
;
1709 unsigned long target_size
;
1711 target_size
= resize_target_update(ht
, size
, growth
);
1712 /* Store resize_target before read resize_initiated */
1714 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1715 uatomic_inc(&ht
->in_progress_resize
);
1716 cmm_smp_mb(); /* increment resize count before load destroy */
1717 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1718 uatomic_dec(&ht
->in_progress_resize
);
1721 work
= malloc(sizeof(*work
));
1723 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1724 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1728 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1731 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1732 unsigned long count
)
1734 struct rcu_resize_work
*work
;
1736 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1738 resize_target_update_count(ht
, count
);
1739 /* Store resize_target before read resize_initiated */
1741 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1742 uatomic_inc(&ht
->in_progress_resize
);
1743 cmm_smp_mb(); /* increment resize count before load destroy */
1744 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1745 uatomic_dec(&ht
->in_progress_resize
);
1748 work
= malloc(sizeof(*work
));
1750 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1751 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);