4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups and use the returned objects safely by delaying
38 * memory reclaim of a grace period.
39 * - Add and remove operations are lock-free, and do not need to
40 * allocate memory. They need to be executed within RCU read-side
41 * critical section to ensure the objects they read are valid and to
42 * deal with the cmpxchg ABA problem.
43 * - add and add_unique operations are supported. add_unique checks if
44 * the node key already exists in the hash table. It ensures no key
46 * - The resize operation executes concurrently with add/remove/lookup.
47 * - Hash table nodes are contained within a split-ordered list. This
48 * list is ordered by incrementing reversed-bits-hash value.
49 * - An index of bucket nodes is kept. These bucket nodes are the hash
50 * table "buckets", and they are also chained together in the
51 * split-ordered list, which allows recursive expansion.
52 * - The resize operation for small tables only allows expanding the hash table.
53 * It is triggered automatically by detecting long chains in the add
55 * - The resize operation for larger tables (and available through an
56 * API) allows both expanding and shrinking the hash table.
57 * - Split-counters are used to keep track of the number of
58 * nodes within the hash table for automatic resize triggering.
59 * - Resize operation initiated by long chain detection is executed by a
60 * call_rcu thread, which keeps lock-freedom of add and remove.
61 * - Resize operations are protected by a mutex.
62 * - The removal operation is split in two parts: first, a "removed"
63 * flag is set in the next pointer within the node to remove. Then,
64 * a "garbage collection" is performed in the bucket containing the
65 * removed node (from the start of the bucket up to the removed node).
66 * All encountered nodes with "removed" flag set in their next
67 * pointers are removed from the linked-list. If the cmpxchg used for
68 * removal fails (due to concurrent garbage-collection or concurrent
69 * add), we retry from the beginning of the bucket. This ensures that
70 * the node with "removed" flag set is removed from the hash table
71 * (not visible to lookups anymore) before the RCU read-side critical
72 * section held across removal ends. Furthermore, this ensures that
73 * the node with "removed" flag set is removed from the linked-list
74 * before its memory is reclaimed. Only the thread which removal
75 * successfully set the "removed" flag (with a cmpxchg) into a node's
76 * next pointer is considered to have succeeded its removal (and thus
77 * owns the node to reclaim). Because we garbage-collect starting from
78 * an invariant node (the start-of-bucket bucket node) up to the
79 * "removed" node (or find a reverse-hash that is higher), we are sure
80 * that a successful traversal of the chain leads to a chain that is
81 * present in the linked-list (the start node is never removed) and
82 * that is does not contain the "removed" node anymore, even if
83 * concurrent delete/add operations are changing the structure of the
85 * - The add operation performs gargage collection of buckets if it
86 * encounters nodes with removed flag set in the bucket where it wants
87 * to add its new node. This ensures lock-freedom of add operation by
88 * helping the remover unlink nodes from the list rather than to wait
90 * - A RCU "order table" indexed by log2(hash index) is copied and
91 * expanded by the resize operation. This order table allows finding
92 * the "bucket node" tables.
93 * - There is one bucket node table per hash index order. The size of
94 * each bucket node table is half the number of hashes contained in
95 * this order (except for order 0).
96 * - synchronzie_rcu is used to garbage-collect the old bucket node table.
97 * - The per-order bucket node tables contain a compact version of the
98 * hash table nodes. These tables are invariant after they are
99 * populated into the hash table.
101 * Bucket node tables:
103 * hash table hash table the last all bucket node tables
104 * order size bucket node 0 1 2 3 4 5 6(index)
111 * 5 32 16 1 1 2 4 8 16
112 * 6 64 32 1 1 2 4 8 16 32
114 * When growing/shrinking, we only focus on the last bucket node table
115 * which size is (!order ? 1 : (1 << (order -1))).
117 * Example for growing/shrinking:
118 * grow hash table from order 5 to 6: init the index=6 bucket node table
119 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
121 * A bit of ascii art explanation:
123 * Order index is the off-by-one compare to the actual power of 2 because
124 * we use index 0 to deal with the 0 special-case.
126 * This shows the nodes for a small table ordered by reversed bits:
138 * This shows the nodes in order of non-reversed bits, linked by
139 * reversed-bit order.
144 * 2 | | 2 010 010 <- |
145 * | | | 3 011 110 | <- |
146 * 3 -> | | | 4 100 001 | |
162 #include <urcu-call-rcu.h>
163 #include <urcu-flavor.h>
164 #include <urcu/arch.h>
165 #include <urcu/uatomic.h>
166 #include <urcu/compiler.h>
167 #include <urcu/rculfhash.h>
168 #include <rculfhash-internal.h>
173 * Split-counters lazily update the global counter each 1024
174 * addition/removal. It automatically keeps track of resize required.
175 * We use the bucket length as indicator for need to expand for small
176 * tables and machines lacking per-cpu data suppport.
178 #define COUNT_COMMIT_ORDER 10
179 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
180 #define CHAIN_LEN_TARGET 1
181 #define CHAIN_LEN_RESIZE_THRESHOLD 3
184 * Define the minimum table size.
186 #define MIN_TABLE_ORDER 0
187 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
190 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
192 #define MIN_PARTITION_PER_THREAD_ORDER 12
193 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
196 * The removed flag needs to be updated atomically with the pointer.
197 * It indicates that no node must attach to the node scheduled for
198 * removal, and that node garbage collection must be performed.
199 * The bucket flag does not require to be updated atomically with the
200 * pointer, but it is added as a pointer low bit flag to save space.
202 #define REMOVED_FLAG (1UL << 0)
203 #define BUCKET_FLAG (1UL << 1)
204 #define REMOVAL_OWNER_FLAG (1UL << 2)
205 #define FLAGS_MASK ((1UL << 3) - 1)
207 /* Value of the end pointer. Should not interact with flags. */
208 #define END_VALUE NULL
211 * ht_items_count: Split-counters counting the number of node addition
212 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
213 * is set at hash table creation.
215 * These are free-running counters, never reset to zero. They count the
216 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
217 * operations to update the global counter. We choose a power-of-2 value
218 * for the trigger to deal with 32 or 64-bit overflow of the counter.
220 struct ht_items_count
{
221 unsigned long add
, del
;
222 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
225 * rcu_resize_work: Contains arguments passed to RCU worker thread
226 * responsible for performing lazy resize.
228 struct rcu_resize_work
{
229 struct rcu_head head
;
234 * partition_resize_work: Contains arguments passed to worker threads
235 * executing the hash table resize on partitions of the hash table
236 * assigned to each processor's worker thread.
238 struct partition_resize_work
{
241 unsigned long i
, start
, len
;
242 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
243 unsigned long start
, unsigned long len
);
247 * Algorithm to reverse bits in a word by lookup table, extended to
250 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
251 * Originally from Public Domain.
254 static const uint8_t BitReverseTable256
[256] =
256 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
257 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
258 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
259 R6(0), R6(2), R6(1), R6(3)
266 uint8_t bit_reverse_u8(uint8_t v
)
268 return BitReverseTable256
[v
];
271 static __attribute__((unused
))
272 uint32_t bit_reverse_u32(uint32_t v
)
274 return ((uint32_t) bit_reverse_u8(v
) << 24) |
275 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
276 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
277 ((uint32_t) bit_reverse_u8(v
>> 24));
280 static __attribute__((unused
))
281 uint64_t bit_reverse_u64(uint64_t v
)
283 return ((uint64_t) bit_reverse_u8(v
) << 56) |
284 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
285 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
286 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
287 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
288 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
289 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
290 ((uint64_t) bit_reverse_u8(v
>> 56));
294 unsigned long bit_reverse_ulong(unsigned long v
)
296 #if (CAA_BITS_PER_LONG == 32)
297 return bit_reverse_u32(v
);
299 return bit_reverse_u64(v
);
304 * fls: returns the position of the most significant bit.
305 * Returns 0 if no bit is set, else returns the position of the most
306 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
308 #if defined(__i386) || defined(__x86_64)
310 unsigned int fls_u32(uint32_t x
)
318 : "=r" (r
) : "rm" (x
));
324 #if defined(__x86_64)
326 unsigned int fls_u64(uint64_t x
)
334 : "=r" (r
) : "rm" (x
));
341 static __attribute__((unused
))
342 unsigned int fls_u64(uint64_t x
)
349 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
353 if (!(x
& 0xFFFF000000000000ULL
)) {
357 if (!(x
& 0xFF00000000000000ULL
)) {
361 if (!(x
& 0xF000000000000000ULL
)) {
365 if (!(x
& 0xC000000000000000ULL
)) {
369 if (!(x
& 0x8000000000000000ULL
)) {
378 static __attribute__((unused
))
379 unsigned int fls_u32(uint32_t x
)
385 if (!(x
& 0xFFFF0000U
)) {
389 if (!(x
& 0xFF000000U
)) {
393 if (!(x
& 0xF0000000U
)) {
397 if (!(x
& 0xC0000000U
)) {
401 if (!(x
& 0x80000000U
)) {
409 unsigned int cds_lfht_fls_ulong(unsigned long x
)
411 #if (CAA_BITS_PER_LONG == 32)
419 * Return the minimum order for which x <= (1UL << order).
420 * Return -1 if x is 0.
422 int cds_lfht_get_count_order_u32(uint32_t x
)
427 return fls_u32(x
- 1);
431 * Return the minimum order for which x <= (1UL << order).
432 * Return -1 if x is 0.
434 int cds_lfht_get_count_order_ulong(unsigned long x
)
439 return cds_lfht_fls_ulong(x
- 1);
443 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
446 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
447 unsigned long count
);
449 static long nr_cpus_mask
= -1;
450 static long split_count_mask
= -1;
452 #if defined(HAVE_SYSCONF)
453 static void ht_init_nr_cpus_mask(void)
457 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
463 * round up number of CPUs to next power of two, so we
464 * can use & for modulo.
466 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
467 nr_cpus_mask
= maxcpus
- 1;
469 #else /* #if defined(HAVE_SYSCONF) */
470 static void ht_init_nr_cpus_mask(void)
474 #endif /* #else #if defined(HAVE_SYSCONF) */
477 void alloc_split_items_count(struct cds_lfht
*ht
)
479 struct ht_items_count
*count
;
481 if (nr_cpus_mask
== -1) {
482 ht_init_nr_cpus_mask();
483 if (nr_cpus_mask
< 0)
484 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
486 split_count_mask
= nr_cpus_mask
;
489 assert(split_count_mask
>= 0);
491 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
492 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
493 assert(ht
->split_count
);
495 ht
->split_count
= NULL
;
500 void free_split_items_count(struct cds_lfht
*ht
)
502 poison_free(ht
->split_count
);
505 #if defined(HAVE_SCHED_GETCPU)
507 int ht_get_split_count_index(unsigned long hash
)
511 assert(split_count_mask
>= 0);
512 cpu
= sched_getcpu();
513 if (caa_unlikely(cpu
< 0))
514 return hash
& split_count_mask
;
516 return cpu
& split_count_mask
;
518 #else /* #if defined(HAVE_SCHED_GETCPU) */
520 int ht_get_split_count_index(unsigned long hash
)
522 return hash
& split_count_mask
;
524 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
527 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
529 unsigned long split_count
;
533 if (caa_unlikely(!ht
->split_count
))
535 index
= ht_get_split_count_index(hash
);
536 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
537 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
539 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
541 dbg_printf("add split count %lu\n", split_count
);
542 count
= uatomic_add_return(&ht
->count
,
543 1UL << COUNT_COMMIT_ORDER
);
544 if (caa_likely(count
& (count
- 1)))
546 /* Only if global count is power of 2 */
548 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
550 dbg_printf("add set global %ld\n", count
);
551 cds_lfht_resize_lazy_count(ht
, size
,
552 count
>> (CHAIN_LEN_TARGET
- 1));
556 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
558 unsigned long split_count
;
562 if (caa_unlikely(!ht
->split_count
))
564 index
= ht_get_split_count_index(hash
);
565 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
566 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
568 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
570 dbg_printf("del split count %lu\n", split_count
);
571 count
= uatomic_add_return(&ht
->count
,
572 -(1UL << COUNT_COMMIT_ORDER
));
573 if (caa_likely(count
& (count
- 1)))
575 /* Only if global count is power of 2 */
577 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
579 dbg_printf("del set global %ld\n", count
);
581 * Don't shrink table if the number of nodes is below a
584 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
586 cds_lfht_resize_lazy_count(ht
, size
,
587 count
>> (CHAIN_LEN_TARGET
- 1));
591 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
595 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
597 count
= uatomic_read(&ht
->count
);
599 * Use bucket-local length for small table expand and for
600 * environments lacking per-cpu data support.
602 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
605 dbg_printf("WARNING: large chain length: %u.\n",
607 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
608 cds_lfht_resize_lazy_grow(ht
, size
,
609 cds_lfht_get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
613 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
615 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
619 int is_removed(struct cds_lfht_node
*node
)
621 return ((unsigned long) node
) & REMOVED_FLAG
;
625 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
627 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
631 int is_bucket(struct cds_lfht_node
*node
)
633 return ((unsigned long) node
) & BUCKET_FLAG
;
637 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
639 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
643 int is_removal_owner(struct cds_lfht_node
*node
)
645 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
649 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
651 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
655 struct cds_lfht_node
*get_end(void)
657 return (struct cds_lfht_node
*) END_VALUE
;
661 int is_end(struct cds_lfht_node
*node
)
663 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
667 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
670 unsigned long old1
, old2
;
672 old1
= uatomic_read(ptr
);
677 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
682 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
684 return ht
->mm
->alloc_bucket_table(ht
, order
);
688 * cds_lfht_free_bucket_table() should be called with decreasing order.
689 * When cds_lfht_free_bucket_table(0) is called, it means the whole
693 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
695 return ht
->mm
->free_bucket_table(ht
, order
);
699 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
701 return ht
->bucket_at(ht
, index
);
705 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
709 return bucket_at(ht
, hash
& (size
- 1));
713 * Remove all logically deleted nodes from a bucket up to a certain node key.
716 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
718 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
720 assert(!is_bucket(bucket
));
721 assert(!is_removed(bucket
));
722 assert(!is_bucket(node
));
723 assert(!is_removed(node
));
726 /* We can always skip the bucket node initially */
727 iter
= rcu_dereference(iter_prev
->next
);
728 assert(!is_removed(iter
));
729 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
731 * We should never be called with bucket (start of chain)
732 * and logically removed node (end of path compression
733 * marker) being the actual same node. This would be a
734 * bug in the algorithm implementation.
736 assert(bucket
!= node
);
738 if (caa_unlikely(is_end(iter
)))
740 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
742 next
= rcu_dereference(clear_flag(iter
)->next
);
743 if (caa_likely(is_removed(next
)))
745 iter_prev
= clear_flag(iter
);
748 assert(!is_removed(iter
));
750 new_next
= flag_bucket(clear_flag(next
));
752 new_next
= clear_flag(next
);
753 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
758 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
759 struct cds_lfht_node
*old_node
,
760 struct cds_lfht_node
*old_next
,
761 struct cds_lfht_node
*new_node
)
763 struct cds_lfht_node
*bucket
, *ret_next
;
765 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
768 assert(!is_removed(old_node
));
769 assert(!is_bucket(old_node
));
770 assert(!is_removed(new_node
));
771 assert(!is_bucket(new_node
));
772 assert(new_node
!= old_node
);
774 /* Insert after node to be replaced */
775 if (is_removed(old_next
)) {
777 * Too late, the old node has been removed under us
778 * between lookup and replace. Fail.
782 assert(old_next
== clear_flag(old_next
));
783 assert(new_node
!= old_next
);
784 new_node
->next
= old_next
;
786 * Here is the whole trick for lock-free replace: we add
787 * the replacement node _after_ the node we want to
788 * replace by atomically setting its next pointer at the
789 * same time we set its removal flag. Given that
790 * the lookups/get next use an iterator aware of the
791 * next pointer, they will either skip the old node due
792 * to the removal flag and see the new node, or use
793 * the old node, but will not see the new one.
794 * This is a replacement of a node with another node
795 * that has the same value: we are therefore not
796 * removing a value from the hash table.
798 ret_next
= uatomic_cmpxchg(&old_node
->next
,
799 old_next
, flag_removed(new_node
));
800 if (ret_next
== old_next
)
801 break; /* We performed the replacement. */
806 * Ensure that the old node is not visible to readers anymore:
807 * lookup for the node, and remove it (along with any other
808 * logically removed node) if found.
810 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
811 _cds_lfht_gc_bucket(bucket
, new_node
);
813 assert(is_removed(rcu_dereference(old_node
->next
)));
818 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
819 * mode. A NULL unique_ret allows creation of duplicate keys.
822 void _cds_lfht_add(struct cds_lfht
*ht
,
824 cds_lfht_match_fct match
,
827 struct cds_lfht_node
*node
,
828 struct cds_lfht_iter
*unique_ret
,
831 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
833 struct cds_lfht_node
*bucket
;
835 assert(!is_bucket(node
));
836 assert(!is_removed(node
));
837 bucket
= lookup_bucket(ht
, size
, hash
);
839 uint32_t chain_len
= 0;
842 * iter_prev points to the non-removed node prior to the
846 /* We can always skip the bucket node initially */
847 iter
= rcu_dereference(iter_prev
->next
);
848 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
850 if (caa_unlikely(is_end(iter
)))
852 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
855 /* bucket node is the first node of the identical-hash-value chain */
856 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
859 next
= rcu_dereference(clear_flag(iter
)->next
);
860 if (caa_unlikely(is_removed(next
)))
866 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
867 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
870 * uniquely adding inserts the node as the first
871 * node of the identical-hash-value node chain.
873 * This semantic ensures no duplicated keys
874 * should ever be observable in the table
875 * (including observe one node by one node
876 * by forward iterations)
878 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
882 *unique_ret
= d_iter
;
886 /* Only account for identical reverse hash once */
887 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
889 check_resize(ht
, size
, ++chain_len
);
890 iter_prev
= clear_flag(iter
);
895 assert(node
!= clear_flag(iter
));
896 assert(!is_removed(iter_prev
));
897 assert(!is_removed(iter
));
898 assert(iter_prev
!= node
);
900 node
->next
= clear_flag(iter
);
902 node
->next
= flag_bucket(clear_flag(iter
));
904 new_node
= flag_bucket(node
);
907 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
909 continue; /* retry */
916 assert(!is_removed(iter
));
918 new_next
= flag_bucket(clear_flag(next
));
920 new_next
= clear_flag(next
);
921 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
926 unique_ret
->node
= return_node
;
927 /* unique_ret->next left unset, never used. */
932 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
933 struct cds_lfht_node
*node
)
935 struct cds_lfht_node
*bucket
, *next
;
937 if (!node
) /* Return -ENOENT if asked to delete NULL node */
940 /* logically delete the node */
941 assert(!is_bucket(node
));
942 assert(!is_removed(node
));
943 assert(!is_removal_owner(node
));
946 * We are first checking if the node had previously been
947 * logically removed (this check is not atomic with setting the
948 * logical removal flag). Return -ENOENT if the node had
949 * previously been removed.
951 next
= rcu_dereference(node
->next
);
952 if (caa_unlikely(is_removed(next
)))
954 assert(!is_bucket(next
));
956 * We set the REMOVED_FLAG unconditionally. Note that there may
957 * be more than one concurrent thread setting this flag.
958 * Knowing which wins the race will be known after the garbage
959 * collection phase, stay tuned!
961 uatomic_or(&node
->next
, REMOVED_FLAG
);
962 /* We performed the (logical) deletion. */
965 * Ensure that the node is not visible to readers anymore: lookup for
966 * the node, and remove it (along with any other logically removed node)
969 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
970 _cds_lfht_gc_bucket(bucket
, node
);
972 assert(is_removed(rcu_dereference(node
->next
)));
974 * Last phase: atomically exchange node->next with a version
975 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
976 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
977 * the node and win the removal race.
978 * It is interesting to note that all "add" paths are forbidden
979 * to change the next pointer starting from the point where the
980 * REMOVED_FLAG is set, so here using a read, followed by a
981 * xchg() suffice to guarantee that the xchg() will ever only
982 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
985 if (!is_removal_owner(uatomic_xchg(&node
->next
,
986 flag_removal_owner(node
->next
))))
993 void *partition_resize_thread(void *arg
)
995 struct partition_resize_work
*work
= arg
;
997 work
->ht
->flavor
->register_thread();
998 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
999 work
->ht
->flavor
->unregister_thread();
1004 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1006 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1007 unsigned long start
, unsigned long len
))
1009 unsigned long partition_len
;
1010 struct partition_resize_work
*work
;
1012 unsigned long nr_threads
;
1015 * Note: nr_cpus_mask + 1 is always power of 2.
1016 * We spawn just the number of threads we need to satisfy the minimum
1017 * partition size, up to the number of CPUs in the system.
1019 if (nr_cpus_mask
> 0) {
1020 nr_threads
= min(nr_cpus_mask
+ 1,
1021 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1025 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1026 work
= calloc(nr_threads
, sizeof(*work
));
1028 for (thread
= 0; thread
< nr_threads
; thread
++) {
1029 work
[thread
].ht
= ht
;
1031 work
[thread
].len
= partition_len
;
1032 work
[thread
].start
= thread
* partition_len
;
1033 work
[thread
].fct
= fct
;
1034 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1035 partition_resize_thread
, &work
[thread
]);
1038 for (thread
= 0; thread
< nr_threads
; thread
++) {
1039 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1046 * Holding RCU read lock to protect _cds_lfht_add against memory
1047 * reclaim that could be performed by other call_rcu worker threads (ABA
1050 * When we reach a certain length, we can split this population phase over
1051 * many worker threads, based on the number of CPUs available in the system.
1052 * This should therefore take care of not having the expand lagging behind too
1053 * many concurrent insertion threads by using the scheduler's ability to
1054 * schedule bucket node population fairly with insertions.
1057 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1058 unsigned long start
, unsigned long len
)
1060 unsigned long j
, size
= 1UL << (i
- 1);
1062 assert(i
> MIN_TABLE_ORDER
);
1063 ht
->flavor
->read_lock();
1064 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1065 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1067 assert(j
>= size
&& j
< (size
<< 1));
1068 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1070 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1071 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1073 ht
->flavor
->read_unlock();
1077 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1080 assert(nr_cpus_mask
!= -1);
1081 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1082 ht
->flavor
->thread_online();
1083 init_table_populate_partition(ht
, i
, 0, len
);
1084 ht
->flavor
->thread_offline();
1087 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1091 void init_table(struct cds_lfht
*ht
,
1092 unsigned long first_order
, unsigned long last_order
)
1096 dbg_printf("init table: first_order %lu last_order %lu\n",
1097 first_order
, last_order
);
1098 assert(first_order
> MIN_TABLE_ORDER
);
1099 for (i
= first_order
; i
<= last_order
; i
++) {
1102 len
= 1UL << (i
- 1);
1103 dbg_printf("init order %lu len: %lu\n", i
, len
);
1105 /* Stop expand if the resize target changes under us */
1106 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1109 cds_lfht_alloc_bucket_table(ht
, i
);
1112 * Set all bucket nodes reverse hash values for a level and
1113 * link all bucket nodes into the table.
1115 init_table_populate(ht
, i
, len
);
1118 * Update table size.
1120 cmm_smp_wmb(); /* populate data before RCU size */
1121 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1123 dbg_printf("init new size: %lu\n", 1UL << i
);
1124 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1130 * Holding RCU read lock to protect _cds_lfht_remove against memory
1131 * reclaim that could be performed by other call_rcu worker threads (ABA
1133 * For a single level, we logically remove and garbage collect each node.
1135 * As a design choice, we perform logical removal and garbage collection on a
1136 * node-per-node basis to simplify this algorithm. We also assume keeping good
1137 * cache locality of the operation would overweight possible performance gain
1138 * that could be achieved by batching garbage collection for multiple levels.
1139 * However, this would have to be justified by benchmarks.
1141 * Concurrent removal and add operations are helping us perform garbage
1142 * collection of logically removed nodes. We guarantee that all logically
1143 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1144 * invoked to free a hole level of bucket nodes (after a grace period).
1146 * Logical removal and garbage collection can therefore be done in batch or on a
1147 * node-per-node basis, as long as the guarantee above holds.
1149 * When we reach a certain length, we can split this removal over many worker
1150 * threads, based on the number of CPUs available in the system. This should
1151 * take care of not letting resize process lag behind too many concurrent
1152 * updater threads actively inserting into the hash table.
1155 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1156 unsigned long start
, unsigned long len
)
1158 unsigned long j
, size
= 1UL << (i
- 1);
1160 assert(i
> MIN_TABLE_ORDER
);
1161 ht
->flavor
->read_lock();
1162 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1163 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1164 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1166 assert(j
>= size
&& j
< (size
<< 1));
1167 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1169 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1170 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1171 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1173 ht
->flavor
->read_unlock();
1177 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1180 assert(nr_cpus_mask
!= -1);
1181 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1182 ht
->flavor
->thread_online();
1183 remove_table_partition(ht
, i
, 0, len
);
1184 ht
->flavor
->thread_offline();
1187 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1191 * fini_table() is never called for first_order == 0, which is why
1192 * free_by_rcu_order == 0 can be used as criterion to know if free must
1196 void fini_table(struct cds_lfht
*ht
,
1197 unsigned long first_order
, unsigned long last_order
)
1200 unsigned long free_by_rcu_order
= 0;
1202 dbg_printf("fini table: first_order %lu last_order %lu\n",
1203 first_order
, last_order
);
1204 assert(first_order
> MIN_TABLE_ORDER
);
1205 for (i
= last_order
; i
>= first_order
; i
--) {
1208 len
= 1UL << (i
- 1);
1209 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1211 /* Stop shrink if the resize target changes under us */
1212 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1215 cmm_smp_wmb(); /* populate data before RCU size */
1216 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1219 * We need to wait for all add operations to reach Q.S. (and
1220 * thus use the new table for lookups) before we can start
1221 * releasing the old bucket nodes. Otherwise their lookup will
1222 * return a logically removed node as insert position.
1224 ht
->flavor
->update_synchronize_rcu();
1225 if (free_by_rcu_order
)
1226 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1229 * Set "removed" flag in bucket nodes about to be removed.
1230 * Unlink all now-logically-removed bucket node pointers.
1231 * Concurrent add/remove operation are helping us doing
1234 remove_table(ht
, i
, len
);
1236 free_by_rcu_order
= i
;
1238 dbg_printf("fini new size: %lu\n", 1UL << i
);
1239 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1243 if (free_by_rcu_order
) {
1244 ht
->flavor
->update_synchronize_rcu();
1245 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1250 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1252 struct cds_lfht_node
*prev
, *node
;
1253 unsigned long order
, len
, i
;
1255 cds_lfht_alloc_bucket_table(ht
, 0);
1257 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1258 node
= bucket_at(ht
, 0);
1259 node
->next
= flag_bucket(get_end());
1260 node
->reverse_hash
= 0;
1262 for (order
= 1; order
< cds_lfht_get_count_order_ulong(size
) + 1; order
++) {
1263 len
= 1UL << (order
- 1);
1264 cds_lfht_alloc_bucket_table(ht
, order
);
1266 for (i
= 0; i
< len
; i
++) {
1268 * Now, we are trying to init the node with the
1269 * hash=(len+i) (which is also a bucket with the
1270 * index=(len+i)) and insert it into the hash table,
1271 * so this node has to be inserted after the bucket
1272 * with the index=(len+i)&(len-1)=i. And because there
1273 * is no other non-bucket node nor bucket node with
1274 * larger index/hash inserted, so the bucket node
1275 * being inserted should be inserted directly linked
1276 * after the bucket node with index=i.
1278 prev
= bucket_at(ht
, i
);
1279 node
= bucket_at(ht
, len
+ i
);
1281 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1282 order
, len
+ i
, len
+ i
);
1283 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1285 /* insert after prev */
1286 assert(is_bucket(prev
->next
));
1287 node
->next
= prev
->next
;
1288 prev
->next
= flag_bucket(node
);
1293 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1294 unsigned long min_nr_alloc_buckets
,
1295 unsigned long max_nr_buckets
,
1297 const struct cds_lfht_mm_type
*mm
,
1298 const struct rcu_flavor_struct
*flavor
,
1299 pthread_attr_t
*attr
)
1301 struct cds_lfht
*ht
;
1302 unsigned long order
;
1304 /* min_nr_alloc_buckets must be power of two */
1305 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1308 /* init_size must be power of two */
1309 if (!init_size
|| (init_size
& (init_size
- 1)))
1313 * Memory management plugin default.
1316 if (CAA_BITS_PER_LONG
> 32
1318 && max_nr_buckets
<= (1ULL << 32)) {
1320 * For 64-bit architectures, with max number of
1321 * buckets small enough not to use the entire
1322 * 64-bit memory mapping space (and allowing a
1323 * fair number of hash table instances), use the
1324 * mmap allocator, which is faster than the
1327 mm
= &cds_lfht_mm_mmap
;
1330 * The fallback is to use the order allocator.
1332 mm
= &cds_lfht_mm_order
;
1336 /* max_nr_buckets == 0 for order based mm means infinite */
1337 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1338 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1340 /* max_nr_buckets must be power of two */
1341 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1344 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1345 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1346 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1347 init_size
= min(init_size
, max_nr_buckets
);
1349 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1351 assert(ht
->mm
== mm
);
1352 assert(ht
->bucket_at
== mm
->bucket_at
);
1355 ht
->flavor
= flavor
;
1356 ht
->resize_attr
= attr
;
1357 alloc_split_items_count(ht
);
1358 /* this mutex should not nest in read-side C.S. */
1359 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1360 order
= cds_lfht_get_count_order_ulong(init_size
);
1361 ht
->resize_target
= 1UL << order
;
1362 cds_lfht_create_bucket(ht
, 1UL << order
);
1363 ht
->size
= 1UL << order
;
1367 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1368 cds_lfht_match_fct match
, const void *key
,
1369 struct cds_lfht_iter
*iter
)
1371 struct cds_lfht_node
*node
, *next
, *bucket
;
1372 unsigned long reverse_hash
, size
;
1374 reverse_hash
= bit_reverse_ulong(hash
);
1376 size
= rcu_dereference(ht
->size
);
1377 bucket
= lookup_bucket(ht
, size
, hash
);
1378 /* We can always skip the bucket node initially */
1379 node
= rcu_dereference(bucket
->next
);
1380 node
= clear_flag(node
);
1382 if (caa_unlikely(is_end(node
))) {
1386 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1390 next
= rcu_dereference(node
->next
);
1391 assert(node
== clear_flag(node
));
1392 if (caa_likely(!is_removed(next
))
1394 && node
->reverse_hash
== reverse_hash
1395 && caa_likely(match(node
, key
))) {
1398 node
= clear_flag(next
);
1400 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1405 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1406 const void *key
, struct cds_lfht_iter
*iter
)
1408 struct cds_lfht_node
*node
, *next
;
1409 unsigned long reverse_hash
;
1412 reverse_hash
= node
->reverse_hash
;
1414 node
= clear_flag(next
);
1417 if (caa_unlikely(is_end(node
))) {
1421 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1425 next
= rcu_dereference(node
->next
);
1426 if (caa_likely(!is_removed(next
))
1428 && caa_likely(match(node
, key
))) {
1431 node
= clear_flag(next
);
1433 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1438 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1440 struct cds_lfht_node
*node
, *next
;
1442 node
= clear_flag(iter
->next
);
1444 if (caa_unlikely(is_end(node
))) {
1448 next
= rcu_dereference(node
->next
);
1449 if (caa_likely(!is_removed(next
))
1450 && !is_bucket(next
)) {
1453 node
= clear_flag(next
);
1455 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1460 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1463 * Get next after first bucket node. The first bucket node is the
1464 * first node of the linked list.
1466 iter
->next
= bucket_at(ht
, 0)->next
;
1467 cds_lfht_next(ht
, iter
);
1470 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1471 struct cds_lfht_node
*node
)
1475 node
->reverse_hash
= bit_reverse_ulong(hash
);
1476 size
= rcu_dereference(ht
->size
);
1477 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1478 ht_count_add(ht
, size
, hash
);
1481 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1483 cds_lfht_match_fct match
,
1485 struct cds_lfht_node
*node
)
1488 struct cds_lfht_iter iter
;
1490 node
->reverse_hash
= bit_reverse_ulong(hash
);
1491 size
= rcu_dereference(ht
->size
);
1492 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1493 if (iter
.node
== node
)
1494 ht_count_add(ht
, size
, hash
);
1498 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1500 cds_lfht_match_fct match
,
1502 struct cds_lfht_node
*node
)
1505 struct cds_lfht_iter iter
;
1507 node
->reverse_hash
= bit_reverse_ulong(hash
);
1508 size
= rcu_dereference(ht
->size
);
1510 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1511 if (iter
.node
== node
) {
1512 ht_count_add(ht
, size
, hash
);
1516 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1521 int cds_lfht_replace(struct cds_lfht
*ht
,
1522 struct cds_lfht_iter
*old_iter
,
1524 cds_lfht_match_fct match
,
1526 struct cds_lfht_node
*new_node
)
1530 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1531 if (!old_iter
->node
)
1533 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1535 if (caa_unlikely(!match(old_iter
->node
, key
)))
1537 size
= rcu_dereference(ht
->size
);
1538 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1542 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1544 unsigned long size
, hash
;
1547 size
= rcu_dereference(ht
->size
);
1548 ret
= _cds_lfht_del(ht
, size
, node
);
1550 hash
= bit_reverse_ulong(node
->reverse_hash
);
1551 ht_count_del(ht
, size
, hash
);
1557 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1559 struct cds_lfht_node
*node
;
1560 unsigned long order
, i
, size
;
1562 /* Check that the table is empty */
1563 node
= bucket_at(ht
, 0);
1565 node
= clear_flag(node
)->next
;
1566 if (!is_bucket(node
))
1568 assert(!is_removed(node
));
1569 } while (!is_end(node
));
1571 * size accessed without rcu_dereference because hash table is
1575 /* Internal sanity check: all nodes left should be bucket */
1576 for (i
= 0; i
< size
; i
++) {
1577 node
= bucket_at(ht
, i
);
1578 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1579 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1580 assert(is_bucket(node
->next
));
1583 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1584 cds_lfht_free_bucket_table(ht
, order
);
1590 * Should only be called when no more concurrent readers nor writers can
1591 * possibly access the table.
1593 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1597 /* Wait for in-flight resize operations to complete */
1598 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1599 cmm_smp_mb(); /* Store destroy before load resize */
1600 while (uatomic_read(&ht
->in_progress_resize
))
1601 poll(NULL
, 0, 100); /* wait for 100ms */
1602 ret
= cds_lfht_delete_bucket(ht
);
1605 free_split_items_count(ht
);
1607 *attr
= ht
->resize_attr
;
1612 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1613 long *approx_before
,
1614 unsigned long *count
,
1617 struct cds_lfht_node
*node
, *next
;
1618 unsigned long nr_bucket
= 0, nr_removed
= 0;
1621 if (ht
->split_count
) {
1624 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1625 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1626 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1632 /* Count non-bucket nodes in the table */
1633 node
= bucket_at(ht
, 0);
1635 next
= rcu_dereference(node
->next
);
1636 if (is_removed(next
)) {
1637 if (!is_bucket(next
))
1641 } else if (!is_bucket(next
))
1645 node
= clear_flag(next
);
1646 } while (!is_end(node
));
1647 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
1648 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1650 if (ht
->split_count
) {
1653 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1654 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1655 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1660 /* called with resize mutex held */
1662 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1663 unsigned long old_size
, unsigned long new_size
)
1665 unsigned long old_order
, new_order
;
1667 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1668 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1669 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1670 old_size
, old_order
, new_size
, new_order
);
1671 assert(new_size
> old_size
);
1672 init_table(ht
, old_order
+ 1, new_order
);
1675 /* called with resize mutex held */
1677 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1678 unsigned long old_size
, unsigned long new_size
)
1680 unsigned long old_order
, new_order
;
1682 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1683 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1684 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1685 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1686 old_size
, old_order
, new_size
, new_order
);
1687 assert(new_size
< old_size
);
1689 /* Remove and unlink all bucket nodes to remove. */
1690 fini_table(ht
, new_order
+ 1, old_order
);
1694 /* called with resize mutex held */
1696 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1698 unsigned long new_size
, old_size
;
1701 * Resize table, re-do if the target size has changed under us.
1704 assert(uatomic_read(&ht
->in_progress_resize
));
1705 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1707 ht
->resize_initiated
= 1;
1708 old_size
= ht
->size
;
1709 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1710 if (old_size
< new_size
)
1711 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1712 else if (old_size
> new_size
)
1713 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1714 ht
->resize_initiated
= 0;
1715 /* write resize_initiated before read resize_target */
1717 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1721 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1723 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
1727 void resize_target_update_count(struct cds_lfht
*ht
,
1728 unsigned long count
)
1730 count
= max(count
, MIN_TABLE_SIZE
);
1731 count
= min(count
, ht
->max_nr_buckets
);
1732 uatomic_set(&ht
->resize_target
, count
);
1735 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1737 resize_target_update_count(ht
, new_size
);
1738 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1739 ht
->flavor
->thread_offline();
1740 pthread_mutex_lock(&ht
->resize_mutex
);
1741 _do_cds_lfht_resize(ht
);
1742 pthread_mutex_unlock(&ht
->resize_mutex
);
1743 ht
->flavor
->thread_online();
1747 void do_resize_cb(struct rcu_head
*head
)
1749 struct rcu_resize_work
*work
=
1750 caa_container_of(head
, struct rcu_resize_work
, head
);
1751 struct cds_lfht
*ht
= work
->ht
;
1753 ht
->flavor
->thread_offline();
1754 pthread_mutex_lock(&ht
->resize_mutex
);
1755 _do_cds_lfht_resize(ht
);
1756 pthread_mutex_unlock(&ht
->resize_mutex
);
1757 ht
->flavor
->thread_online();
1759 cmm_smp_mb(); /* finish resize before decrement */
1760 uatomic_dec(&ht
->in_progress_resize
);
1764 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1766 struct rcu_resize_work
*work
;
1768 /* Store resize_target before read resize_initiated */
1770 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
1771 uatomic_inc(&ht
->in_progress_resize
);
1772 cmm_smp_mb(); /* increment resize count before load destroy */
1773 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1774 uatomic_dec(&ht
->in_progress_resize
);
1777 work
= malloc(sizeof(*work
));
1779 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1780 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1785 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1787 unsigned long target_size
= size
<< growth
;
1789 target_size
= min(target_size
, ht
->max_nr_buckets
);
1790 if (resize_target_grow(ht
, target_size
) >= target_size
)
1793 __cds_lfht_resize_lazy_launch(ht
);
1797 * We favor grow operations over shrink. A shrink operation never occurs
1798 * if a grow operation is queued for lazy execution. A grow operation
1799 * cancels any pending shrink lazy execution.
1802 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1803 unsigned long count
)
1805 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1807 count
= max(count
, MIN_TABLE_SIZE
);
1808 count
= min(count
, ht
->max_nr_buckets
);
1810 return; /* Already the right size, no resize needed */
1811 if (count
> size
) { /* lazy grow */
1812 if (resize_target_grow(ht
, count
) >= count
)
1814 } else { /* lazy shrink */
1818 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
1820 break; /* no resize needed */
1822 return; /* growing is/(was just) in progress */
1824 return; /* some other thread do shrink */
1828 __cds_lfht_resize_lazy_launch(ht
);