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>
172 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
174 #define dbg_printf(fmt, args...)
178 * Split-counters lazily update the global counter each 1024
179 * addition/removal. It automatically keeps track of resize required.
180 * We use the bucket length as indicator for need to expand for small
181 * tables and machines lacking per-cpu data suppport.
183 #define COUNT_COMMIT_ORDER 10
184 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
185 #define CHAIN_LEN_TARGET 1
186 #define CHAIN_LEN_RESIZE_THRESHOLD 3
189 * Define the minimum table size.
191 #define MIN_TABLE_ORDER 0
192 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
194 #if (CAA_BITS_PER_LONG == 32)
195 #define MAX_TABLE_ORDER 32
197 #define MAX_TABLE_ORDER 64
201 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
203 #define MIN_PARTITION_PER_THREAD_ORDER 12
204 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
207 #define min(a, b) ((a) < (b) ? (a) : (b))
211 #define max(a, b) ((a) > (b) ? (a) : (b))
215 * The removed flag needs to be updated atomically with the pointer.
216 * It indicates that no node must attach to the node scheduled for
217 * removal, and that node garbage collection must be performed.
218 * The bucket flag does not require to be updated atomically with the
219 * pointer, but it is added as a pointer low bit flag to save space.
221 #define REMOVED_FLAG (1UL << 0)
222 #define BUCKET_FLAG (1UL << 1)
223 #define FLAGS_MASK ((1UL << 2) - 1)
225 /* Value of the end pointer. Should not interact with flags. */
226 #define END_VALUE NULL
229 * ht_items_count: Split-counters counting the number of node addition
230 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
231 * is set at hash table creation.
233 * These are free-running counters, never reset to zero. They count the
234 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
235 * operations to update the global counter. We choose a power-of-2 value
236 * for the trigger to deal with 32 or 64-bit overflow of the counter.
238 struct ht_items_count
{
239 unsigned long add
, del
;
240 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
243 * rcu_table: Contains the size and desired new size if a resize
244 * operation is in progress, as well as the statically-sized array of
245 * bucket table pointers.
248 unsigned long size
; /* always a power of 2, shared (RCU) */
249 unsigned long resize_target
;
250 int resize_initiated
;
253 * Contains the per order-index-level bucket node table. The size
254 * of each bucket node table is half the number of hashes contained
255 * in this order (except for order 0). The minimum allocation size
256 * parameter allows combining the bucket node arrays of the lowermost
257 * levels to improve cache locality for small index orders.
259 struct cds_lfht_node
*tbl
[MAX_TABLE_ORDER
];
263 * cds_lfht: Top-level data structure representing a lock-free hash
264 * table. Defined in the implementation file to make it be an opaque
269 unsigned long min_alloc_buckets_order
;
270 unsigned long min_nr_alloc_buckets
;
271 unsigned long max_nr_buckets
;
274 * We need to put the work threads offline (QSBR) when taking this
275 * mutex, because we use synchronize_rcu within this mutex critical
276 * section, which waits on read-side critical sections, and could
277 * therefore cause grace-period deadlock if we hold off RCU G.P.
280 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
281 unsigned int in_progress_resize
, in_progress_destroy
;
282 const struct rcu_flavor_struct
*flavor
;
283 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
284 long count
; /* global approximate item count */
285 struct ht_items_count
*split_count
; /* split item count */
289 * rcu_resize_work: Contains arguments passed to RCU worker thread
290 * responsible for performing lazy resize.
292 struct rcu_resize_work
{
293 struct rcu_head head
;
298 * partition_resize_work: Contains arguments passed to worker threads
299 * executing the hash table resize on partitions of the hash table
300 * assigned to each processor's worker thread.
302 struct partition_resize_work
{
305 unsigned long i
, start
, len
;
306 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
307 unsigned long start
, unsigned long len
);
311 void _cds_lfht_add(struct cds_lfht
*ht
,
312 cds_lfht_match_fct match
,
315 struct cds_lfht_node
*node
,
316 struct cds_lfht_iter
*unique_ret
,
320 * Algorithm to reverse bits in a word by lookup table, extended to
323 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
324 * Originally from Public Domain.
327 static const uint8_t BitReverseTable256
[256] =
329 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
330 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
331 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
332 R6(0), R6(2), R6(1), R6(3)
339 uint8_t bit_reverse_u8(uint8_t v
)
341 return BitReverseTable256
[v
];
344 static __attribute__((unused
))
345 uint32_t bit_reverse_u32(uint32_t v
)
347 return ((uint32_t) bit_reverse_u8(v
) << 24) |
348 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
349 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
350 ((uint32_t) bit_reverse_u8(v
>> 24));
353 static __attribute__((unused
))
354 uint64_t bit_reverse_u64(uint64_t v
)
356 return ((uint64_t) bit_reverse_u8(v
) << 56) |
357 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
358 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
359 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
360 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
361 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
362 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
363 ((uint64_t) bit_reverse_u8(v
>> 56));
367 unsigned long bit_reverse_ulong(unsigned long v
)
369 #if (CAA_BITS_PER_LONG == 32)
370 return bit_reverse_u32(v
);
372 return bit_reverse_u64(v
);
377 * fls: returns the position of the most significant bit.
378 * Returns 0 if no bit is set, else returns the position of the most
379 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
381 #if defined(__i386) || defined(__x86_64)
383 unsigned int fls_u32(uint32_t x
)
391 : "=r" (r
) : "rm" (x
));
397 #if defined(__x86_64)
399 unsigned int fls_u64(uint64_t x
)
407 : "=r" (r
) : "rm" (x
));
414 static __attribute__((unused
))
415 unsigned int fls_u64(uint64_t x
)
422 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
426 if (!(x
& 0xFFFF000000000000ULL
)) {
430 if (!(x
& 0xFF00000000000000ULL
)) {
434 if (!(x
& 0xF000000000000000ULL
)) {
438 if (!(x
& 0xC000000000000000ULL
)) {
442 if (!(x
& 0x8000000000000000ULL
)) {
451 static __attribute__((unused
))
452 unsigned int fls_u32(uint32_t x
)
458 if (!(x
& 0xFFFF0000U
)) {
462 if (!(x
& 0xFF000000U
)) {
466 if (!(x
& 0xF0000000U
)) {
470 if (!(x
& 0xC0000000U
)) {
474 if (!(x
& 0x80000000U
)) {
482 unsigned int fls_ulong(unsigned long x
)
484 #if (CAA_BITS_PER_LONG == 32)
492 * Return the minimum order for which x <= (1UL << order).
493 * Return -1 if x is 0.
495 int get_count_order_u32(uint32_t x
)
500 return fls_u32(x
- 1);
504 * Return the minimum order for which x <= (1UL << order).
505 * Return -1 if x is 0.
507 int get_count_order_ulong(unsigned long x
)
512 return fls_ulong(x
- 1);
516 #define poison_free(ptr) \
519 memset(ptr, 0x42, sizeof(*(ptr))); \
524 #define poison_free(ptr) free(ptr)
528 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
531 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
532 unsigned long count
);
534 static long nr_cpus_mask
= -1;
535 static long split_count_mask
= -1;
537 #if defined(HAVE_SYSCONF)
538 static void ht_init_nr_cpus_mask(void)
542 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
548 * round up number of CPUs to next power of two, so we
549 * can use & for modulo.
551 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
552 nr_cpus_mask
= maxcpus
- 1;
554 #else /* #if defined(HAVE_SYSCONF) */
555 static void ht_init_nr_cpus_mask(void)
559 #endif /* #else #if defined(HAVE_SYSCONF) */
562 void alloc_split_items_count(struct cds_lfht
*ht
)
564 struct ht_items_count
*count
;
566 if (nr_cpus_mask
== -1) {
567 ht_init_nr_cpus_mask();
568 if (nr_cpus_mask
< 0)
569 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
571 split_count_mask
= nr_cpus_mask
;
574 assert(split_count_mask
>= 0);
576 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
577 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
578 assert(ht
->split_count
);
580 ht
->split_count
= NULL
;
585 void free_split_items_count(struct cds_lfht
*ht
)
587 poison_free(ht
->split_count
);
590 #if defined(HAVE_SCHED_GETCPU)
592 int ht_get_split_count_index(unsigned long hash
)
596 assert(split_count_mask
>= 0);
597 cpu
= sched_getcpu();
598 if (caa_unlikely(cpu
< 0))
599 return hash
& split_count_mask
;
601 return cpu
& split_count_mask
;
603 #else /* #if defined(HAVE_SCHED_GETCPU) */
605 int ht_get_split_count_index(unsigned long hash
)
607 return hash
& split_count_mask
;
609 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
612 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
614 unsigned long split_count
;
617 if (caa_unlikely(!ht
->split_count
))
619 index
= ht_get_split_count_index(hash
);
620 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
621 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
624 dbg_printf("add split count %lu\n", split_count
);
625 count
= uatomic_add_return(&ht
->count
,
626 1UL << COUNT_COMMIT_ORDER
);
628 if (!(count
& (count
- 1))) {
629 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
631 dbg_printf("add set global %ld\n", count
);
632 cds_lfht_resize_lazy_count(ht
, size
,
633 count
>> (CHAIN_LEN_TARGET
- 1));
639 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
641 unsigned long split_count
;
644 if (caa_unlikely(!ht
->split_count
))
646 index
= ht_get_split_count_index(hash
);
647 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
648 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
651 dbg_printf("del split count %lu\n", split_count
);
652 count
= uatomic_add_return(&ht
->count
,
653 -(1UL << COUNT_COMMIT_ORDER
));
655 if (!(count
& (count
- 1))) {
656 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
658 dbg_printf("del set global %ld\n", count
);
660 * Don't shrink table if the number of nodes is below a
663 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
665 cds_lfht_resize_lazy_count(ht
, size
,
666 count
>> (CHAIN_LEN_TARGET
- 1));
672 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
676 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
678 count
= uatomic_read(&ht
->count
);
680 * Use bucket-local length for small table expand and for
681 * environments lacking per-cpu data support.
683 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
686 dbg_printf("WARNING: large chain length: %u.\n",
688 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
689 cds_lfht_resize_lazy_grow(ht
, size
,
690 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
694 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
696 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
700 int is_removed(struct cds_lfht_node
*node
)
702 return ((unsigned long) node
) & REMOVED_FLAG
;
706 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
708 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
712 int is_bucket(struct cds_lfht_node
*node
)
714 return ((unsigned long) node
) & BUCKET_FLAG
;
718 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
720 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
724 struct cds_lfht_node
*get_end(void)
726 return (struct cds_lfht_node
*) END_VALUE
;
730 int is_end(struct cds_lfht_node
*node
)
732 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
736 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
739 unsigned long old1
, old2
;
741 old1
= uatomic_read(ptr
);
746 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
751 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
754 ht
->t
.tbl
[0] = calloc(ht
->min_nr_alloc_buckets
,
755 sizeof(struct cds_lfht_node
));
756 assert(ht
->t
.tbl
[0]);
757 } else if (order
> ht
->min_alloc_buckets_order
) {
758 ht
->t
.tbl
[order
] = calloc(1UL << (order
-1),
759 sizeof(struct cds_lfht_node
));
760 assert(ht
->t
.tbl
[order
]);
762 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
766 * cds_lfht_free_bucket_table() should be called with decreasing order.
767 * When cds_lfht_free_bucket_table(0) is called, it means the whole
771 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
774 poison_free(ht
->t
.tbl
[0]);
775 else if (order
> ht
->min_alloc_buckets_order
)
776 poison_free(ht
->t
.tbl
[order
]);
777 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
781 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
785 if ((__builtin_constant_p(index
) && index
== 0)
786 || index
< ht
->min_nr_alloc_buckets
) {
787 dbg_printf("bucket index %lu order 0 aridx 0\n", index
);
788 return &ht
->t
.tbl
[0][index
];
791 * equivalent to get_count_order_ulong(index + 1), but optimizes
792 * away the non-existing 0 special-case for
793 * get_count_order_ulong.
795 order
= fls_ulong(index
);
796 dbg_printf("bucket index %lu order %lu aridx %lu\n",
797 index
, order
, index
& ((1UL << (order
- 1)) - 1));
798 return &ht
->t
.tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
802 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
806 return bucket_at(ht
, hash
& (size
- 1));
810 * Remove all logically deleted nodes from a bucket up to a certain node key.
813 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
815 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
817 assert(!is_bucket(bucket
));
818 assert(!is_removed(bucket
));
819 assert(!is_bucket(node
));
820 assert(!is_removed(node
));
823 /* We can always skip the bucket node initially */
824 iter
= rcu_dereference(iter_prev
->next
);
825 assert(!is_removed(iter
));
826 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
828 * We should never be called with bucket (start of chain)
829 * and logically removed node (end of path compression
830 * marker) being the actual same node. This would be a
831 * bug in the algorithm implementation.
833 assert(bucket
!= node
);
835 if (caa_unlikely(is_end(iter
)))
837 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
839 next
= rcu_dereference(clear_flag(iter
)->next
);
840 if (caa_likely(is_removed(next
)))
842 iter_prev
= clear_flag(iter
);
845 assert(!is_removed(iter
));
847 new_next
= flag_bucket(clear_flag(next
));
849 new_next
= clear_flag(next
);
850 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
856 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
857 struct cds_lfht_node
*old_node
,
858 struct cds_lfht_node
*old_next
,
859 struct cds_lfht_node
*new_node
)
861 struct cds_lfht_node
*bucket
, *ret_next
;
863 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
866 assert(!is_removed(old_node
));
867 assert(!is_bucket(old_node
));
868 assert(!is_removed(new_node
));
869 assert(!is_bucket(new_node
));
870 assert(new_node
!= old_node
);
872 /* Insert after node to be replaced */
873 if (is_removed(old_next
)) {
875 * Too late, the old node has been removed under us
876 * between lookup and replace. Fail.
880 assert(!is_bucket(old_next
));
881 assert(new_node
!= clear_flag(old_next
));
882 new_node
->next
= clear_flag(old_next
);
884 * Here is the whole trick for lock-free replace: we add
885 * the replacement node _after_ the node we want to
886 * replace by atomically setting its next pointer at the
887 * same time we set its removal flag. Given that
888 * the lookups/get next use an iterator aware of the
889 * next pointer, they will either skip the old node due
890 * to the removal flag and see the new node, or use
891 * the old node, but will not see the new one.
893 ret_next
= uatomic_cmpxchg(&old_node
->next
,
894 old_next
, flag_removed(new_node
));
895 if (ret_next
== old_next
)
896 break; /* We performed the replacement. */
901 * Ensure that the old node is not visible to readers anymore:
902 * lookup for the node, and remove it (along with any other
903 * logically removed node) if found.
905 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
906 _cds_lfht_gc_bucket(bucket
, new_node
);
908 assert(is_removed(rcu_dereference(old_node
->next
)));
913 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
914 * mode. A NULL unique_ret allows creation of duplicate keys.
917 void _cds_lfht_add(struct cds_lfht
*ht
,
918 cds_lfht_match_fct match
,
921 struct cds_lfht_node
*node
,
922 struct cds_lfht_iter
*unique_ret
,
925 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
927 struct cds_lfht_node
*bucket
;
929 assert(!is_bucket(node
));
930 assert(!is_removed(node
));
931 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
933 uint32_t chain_len
= 0;
936 * iter_prev points to the non-removed node prior to the
940 /* We can always skip the bucket node initially */
941 iter
= rcu_dereference(iter_prev
->next
);
942 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
944 if (caa_unlikely(is_end(iter
)))
946 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
949 /* bucket node is the first node of the identical-hash-value chain */
950 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
953 next
= rcu_dereference(clear_flag(iter
)->next
);
954 if (caa_unlikely(is_removed(next
)))
960 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
961 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
964 * uniquely adding inserts the node as the first
965 * node of the identical-hash-value node chain.
967 * This semantic ensures no duplicated keys
968 * should ever be observable in the table
969 * (including observe one node by one node
970 * by forward iterations)
972 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
976 *unique_ret
= d_iter
;
980 /* Only account for identical reverse hash once */
981 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
983 check_resize(ht
, size
, ++chain_len
);
984 iter_prev
= clear_flag(iter
);
989 assert(node
!= clear_flag(iter
));
990 assert(!is_removed(iter_prev
));
991 assert(!is_removed(iter
));
992 assert(iter_prev
!= node
);
994 node
->next
= clear_flag(iter
);
996 node
->next
= flag_bucket(clear_flag(iter
));
998 new_node
= flag_bucket(node
);
1001 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1002 new_node
) != iter
) {
1003 continue; /* retry */
1010 assert(!is_removed(iter
));
1011 if (is_bucket(iter
))
1012 new_next
= flag_bucket(clear_flag(next
));
1014 new_next
= clear_flag(next
);
1015 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1020 unique_ret
->node
= return_node
;
1021 /* unique_ret->next left unset, never used. */
1026 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1027 struct cds_lfht_node
*node
,
1030 struct cds_lfht_node
*bucket
, *next
, *old
;
1032 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1035 /* logically delete the node */
1036 assert(!is_bucket(node
));
1037 assert(!is_removed(node
));
1038 old
= rcu_dereference(node
->next
);
1040 struct cds_lfht_node
*new_next
;
1043 if (caa_unlikely(is_removed(next
)))
1046 assert(is_bucket(next
));
1048 assert(!is_bucket(next
));
1049 new_next
= flag_removed(next
);
1050 old
= uatomic_cmpxchg(&node
->next
, next
, new_next
);
1051 } while (old
!= next
);
1052 /* We performed the (logical) deletion. */
1055 * Ensure that the node is not visible to readers anymore: lookup for
1056 * the node, and remove it (along with any other logically removed node)
1059 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1060 _cds_lfht_gc_bucket(bucket
, node
);
1062 assert(is_removed(rcu_dereference(node
->next
)));
1067 void *partition_resize_thread(void *arg
)
1069 struct partition_resize_work
*work
= arg
;
1071 work
->ht
->flavor
->register_thread();
1072 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1073 work
->ht
->flavor
->unregister_thread();
1078 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1080 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1081 unsigned long start
, unsigned long len
))
1083 unsigned long partition_len
;
1084 struct partition_resize_work
*work
;
1086 unsigned long nr_threads
;
1089 * Note: nr_cpus_mask + 1 is always power of 2.
1090 * We spawn just the number of threads we need to satisfy the minimum
1091 * partition size, up to the number of CPUs in the system.
1093 if (nr_cpus_mask
> 0) {
1094 nr_threads
= min(nr_cpus_mask
+ 1,
1095 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1099 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1100 work
= calloc(nr_threads
, sizeof(*work
));
1102 for (thread
= 0; thread
< nr_threads
; thread
++) {
1103 work
[thread
].ht
= ht
;
1105 work
[thread
].len
= partition_len
;
1106 work
[thread
].start
= thread
* partition_len
;
1107 work
[thread
].fct
= fct
;
1108 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1109 partition_resize_thread
, &work
[thread
]);
1112 for (thread
= 0; thread
< nr_threads
; thread
++) {
1113 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1120 * Holding RCU read lock to protect _cds_lfht_add against memory
1121 * reclaim that could be performed by other call_rcu worker threads (ABA
1124 * When we reach a certain length, we can split this population phase over
1125 * many worker threads, based on the number of CPUs available in the system.
1126 * This should therefore take care of not having the expand lagging behind too
1127 * many concurrent insertion threads by using the scheduler's ability to
1128 * schedule bucket node population fairly with insertions.
1131 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1132 unsigned long start
, unsigned long len
)
1134 unsigned long j
, size
= 1UL << (i
- 1);
1136 assert(i
> MIN_TABLE_ORDER
);
1137 ht
->flavor
->read_lock();
1138 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1139 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1141 assert(j
>= size
&& j
< (size
<< 1));
1142 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1144 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1145 _cds_lfht_add(ht
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1147 ht
->flavor
->read_unlock();
1151 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1154 assert(nr_cpus_mask
!= -1);
1155 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1156 ht
->flavor
->thread_online();
1157 init_table_populate_partition(ht
, i
, 0, len
);
1158 ht
->flavor
->thread_offline();
1161 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1165 void init_table(struct cds_lfht
*ht
,
1166 unsigned long first_order
, unsigned long last_order
)
1170 dbg_printf("init table: first_order %lu last_order %lu\n",
1171 first_order
, last_order
);
1172 assert(first_order
> MIN_TABLE_ORDER
);
1173 for (i
= first_order
; i
<= last_order
; i
++) {
1176 len
= 1UL << (i
- 1);
1177 dbg_printf("init order %lu len: %lu\n", i
, len
);
1179 /* Stop expand if the resize target changes under us */
1180 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (1UL << i
))
1183 cds_lfht_alloc_bucket_table(ht
, i
);
1186 * Set all bucket nodes reverse hash values for a level and
1187 * link all bucket nodes into the table.
1189 init_table_populate(ht
, i
, len
);
1192 * Update table size.
1194 cmm_smp_wmb(); /* populate data before RCU size */
1195 CMM_STORE_SHARED(ht
->t
.size
, 1UL << i
);
1197 dbg_printf("init new size: %lu\n", 1UL << i
);
1198 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1204 * Holding RCU read lock to protect _cds_lfht_remove against memory
1205 * reclaim that could be performed by other call_rcu worker threads (ABA
1207 * For a single level, we logically remove and garbage collect each node.
1209 * As a design choice, we perform logical removal and garbage collection on a
1210 * node-per-node basis to simplify this algorithm. We also assume keeping good
1211 * cache locality of the operation would overweight possible performance gain
1212 * that could be achieved by batching garbage collection for multiple levels.
1213 * However, this would have to be justified by benchmarks.
1215 * Concurrent removal and add operations are helping us perform garbage
1216 * collection of logically removed nodes. We guarantee that all logically
1217 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1218 * invoked to free a hole level of bucket nodes (after a grace period).
1220 * Logical removal and garbage collection can therefore be done in batch or on a
1221 * node-per-node basis, as long as the guarantee above holds.
1223 * When we reach a certain length, we can split this removal over many worker
1224 * threads, based on the number of CPUs available in the system. This should
1225 * take care of not letting resize process lag behind too many concurrent
1226 * updater threads actively inserting into the hash table.
1229 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1230 unsigned long start
, unsigned long len
)
1232 unsigned long j
, size
= 1UL << (i
- 1);
1234 assert(i
> MIN_TABLE_ORDER
);
1235 ht
->flavor
->read_lock();
1236 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1237 struct cds_lfht_node
*fini_node
= bucket_at(ht
, j
);
1239 assert(j
>= size
&& j
< (size
<< 1));
1240 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1242 fini_node
->reverse_hash
= bit_reverse_ulong(j
);
1243 (void) _cds_lfht_del(ht
, size
, fini_node
, 1);
1245 ht
->flavor
->read_unlock();
1249 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1252 assert(nr_cpus_mask
!= -1);
1253 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1254 ht
->flavor
->thread_online();
1255 remove_table_partition(ht
, i
, 0, len
);
1256 ht
->flavor
->thread_offline();
1259 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1263 void fini_table(struct cds_lfht
*ht
,
1264 unsigned long first_order
, unsigned long last_order
)
1267 unsigned long free_by_rcu_order
= 0;
1269 dbg_printf("fini table: first_order %lu last_order %lu\n",
1270 first_order
, last_order
);
1271 assert(first_order
> MIN_TABLE_ORDER
);
1272 for (i
= last_order
; i
>= first_order
; i
--) {
1275 len
= 1UL << (i
- 1);
1276 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1278 /* Stop shrink if the resize target changes under us */
1279 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1282 cmm_smp_wmb(); /* populate data before RCU size */
1283 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1286 * We need to wait for all add operations to reach Q.S. (and
1287 * thus use the new table for lookups) before we can start
1288 * releasing the old bucket nodes. Otherwise their lookup will
1289 * return a logically removed node as insert position.
1291 ht
->flavor
->update_synchronize_rcu();
1292 if (free_by_rcu_order
)
1293 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1296 * Set "removed" flag in bucket nodes about to be removed.
1297 * Unlink all now-logically-removed bucket node pointers.
1298 * Concurrent add/remove operation are helping us doing
1301 remove_table(ht
, i
, len
);
1303 free_by_rcu_order
= i
;
1305 dbg_printf("fini new size: %lu\n", 1UL << i
);
1306 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1310 if (free_by_rcu_order
) {
1311 ht
->flavor
->update_synchronize_rcu();
1312 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1317 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1319 struct cds_lfht_node
*prev
, *node
;
1320 unsigned long order
, len
, i
;
1322 cds_lfht_alloc_bucket_table(ht
, 0);
1324 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1325 node
= bucket_at(ht
, 0);
1326 node
->next
= flag_bucket(get_end());
1327 node
->reverse_hash
= 0;
1329 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1330 len
= 1UL << (order
- 1);
1331 cds_lfht_alloc_bucket_table(ht
, order
);
1333 for (i
= 0; i
< len
; i
++) {
1335 * Now, we are trying to init the node with the
1336 * hash=(len+i) (which is also a bucket with the
1337 * index=(len+i)) and insert it into the hash table,
1338 * so this node has to be inserted after the bucket
1339 * with the index=(len+i)&(len-1)=i. And because there
1340 * is no other non-bucket node nor bucket node with
1341 * larger index/hash inserted, so the bucket node
1342 * being inserted should be inserted directly linked
1343 * after the bucket node with index=i.
1345 prev
= bucket_at(ht
, i
);
1346 node
= bucket_at(ht
, len
+ i
);
1348 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1349 order
, len
+ i
, len
+ i
);
1350 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1352 /* insert after prev */
1353 assert(is_bucket(prev
->next
));
1354 node
->next
= prev
->next
;
1355 prev
->next
= flag_bucket(node
);
1360 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1361 unsigned long min_nr_alloc_buckets
,
1362 unsigned long max_nr_buckets
,
1364 const struct rcu_flavor_struct
*flavor
,
1365 pthread_attr_t
*attr
)
1367 struct cds_lfht
*ht
;
1368 unsigned long order
;
1370 /* min_nr_alloc_buckets must be power of two */
1371 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1374 /* init_size must be power of two */
1375 if (!init_size
|| (init_size
& (init_size
- 1)))
1378 if (!max_nr_buckets
)
1379 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1381 /* max_nr_buckets must be power of two */
1382 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1385 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1386 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1387 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1388 init_size
= min(init_size
, max_nr_buckets
);
1389 ht
= calloc(1, sizeof(struct cds_lfht
));
1392 ht
->flavor
= flavor
;
1393 ht
->resize_attr
= attr
;
1394 alloc_split_items_count(ht
);
1395 /* this mutex should not nest in read-side C.S. */
1396 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1397 order
= get_count_order_ulong(init_size
);
1398 ht
->t
.resize_target
= 1UL << order
;
1399 ht
->min_nr_alloc_buckets
= min_nr_alloc_buckets
;
1400 ht
->min_alloc_buckets_order
= get_count_order_ulong(min_nr_alloc_buckets
);
1401 ht
->max_nr_buckets
= max_nr_buckets
;
1402 cds_lfht_create_bucket(ht
, 1UL << order
);
1403 ht
->t
.size
= 1UL << order
;
1407 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1408 cds_lfht_match_fct match
, const void *key
,
1409 struct cds_lfht_iter
*iter
)
1411 struct cds_lfht_node
*node
, *next
, *bucket
;
1412 unsigned long reverse_hash
, size
;
1414 reverse_hash
= bit_reverse_ulong(hash
);
1416 size
= rcu_dereference(ht
->t
.size
);
1417 bucket
= lookup_bucket(ht
, size
, hash
);
1418 /* We can always skip the bucket node initially */
1419 node
= rcu_dereference(bucket
->next
);
1420 node
= clear_flag(node
);
1422 if (caa_unlikely(is_end(node
))) {
1426 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1430 next
= rcu_dereference(node
->next
);
1431 assert(node
== clear_flag(node
));
1432 if (caa_likely(!is_removed(next
))
1434 && node
->reverse_hash
== reverse_hash
1435 && caa_likely(match(node
, key
))) {
1438 node
= clear_flag(next
);
1440 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1445 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1446 const void *key
, struct cds_lfht_iter
*iter
)
1448 struct cds_lfht_node
*node
, *next
;
1449 unsigned long reverse_hash
;
1452 reverse_hash
= node
->reverse_hash
;
1454 node
= clear_flag(next
);
1457 if (caa_unlikely(is_end(node
))) {
1461 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1465 next
= rcu_dereference(node
->next
);
1466 if (caa_likely(!is_removed(next
))
1468 && caa_likely(match(node
, key
))) {
1471 node
= clear_flag(next
);
1473 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1478 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1480 struct cds_lfht_node
*node
, *next
;
1482 node
= clear_flag(iter
->next
);
1484 if (caa_unlikely(is_end(node
))) {
1488 next
= rcu_dereference(node
->next
);
1489 if (caa_likely(!is_removed(next
))
1490 && !is_bucket(next
)) {
1493 node
= clear_flag(next
);
1495 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1500 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1503 * Get next after first bucket node. The first bucket node is the
1504 * first node of the linked list.
1506 iter
->next
= bucket_at(ht
, 0)->next
;
1507 cds_lfht_next(ht
, iter
);
1510 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1511 struct cds_lfht_node
*node
)
1515 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1516 size
= rcu_dereference(ht
->t
.size
);
1517 _cds_lfht_add(ht
, NULL
, NULL
, size
, node
, NULL
, 0);
1518 ht_count_add(ht
, size
, hash
);
1521 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1523 cds_lfht_match_fct match
,
1525 struct cds_lfht_node
*node
)
1528 struct cds_lfht_iter iter
;
1530 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1531 size
= rcu_dereference(ht
->t
.size
);
1532 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1533 if (iter
.node
== node
)
1534 ht_count_add(ht
, size
, hash
);
1538 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1540 cds_lfht_match_fct match
,
1542 struct cds_lfht_node
*node
)
1545 struct cds_lfht_iter iter
;
1547 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1548 size
= rcu_dereference(ht
->t
.size
);
1550 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1551 if (iter
.node
== node
) {
1552 ht_count_add(ht
, size
, hash
);
1556 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1561 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1562 struct cds_lfht_node
*new_node
)
1566 size
= rcu_dereference(ht
->t
.size
);
1567 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1571 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1573 unsigned long size
, hash
;
1576 size
= rcu_dereference(ht
->t
.size
);
1577 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1579 hash
= bit_reverse_ulong(iter
->node
->reverse_hash
);
1580 ht_count_del(ht
, size
, hash
);
1586 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1588 struct cds_lfht_node
*node
;
1589 unsigned long order
, i
, size
;
1591 /* Check that the table is empty */
1592 node
= bucket_at(ht
, 0);
1594 node
= clear_flag(node
)->next
;
1595 if (!is_bucket(node
))
1597 assert(!is_removed(node
));
1598 } while (!is_end(node
));
1600 * size accessed without rcu_dereference because hash table is
1604 /* Internal sanity check: all nodes left should be bucket */
1605 for (i
= 0; i
< size
; i
++) {
1606 node
= bucket_at(ht
, i
);
1607 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1608 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1609 assert(is_bucket(node
->next
));
1612 for (order
= get_count_order_ulong(size
); (long)order
>= 0; order
--)
1613 cds_lfht_free_bucket_table(ht
, order
);
1619 * Should only be called when no more concurrent readers nor writers can
1620 * possibly access the table.
1622 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1626 /* Wait for in-flight resize operations to complete */
1627 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1628 cmm_smp_mb(); /* Store destroy before load resize */
1629 while (uatomic_read(&ht
->in_progress_resize
))
1630 poll(NULL
, 0, 100); /* wait for 100ms */
1631 ret
= cds_lfht_delete_bucket(ht
);
1634 free_split_items_count(ht
);
1636 *attr
= ht
->resize_attr
;
1641 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1642 long *approx_before
,
1643 unsigned long *count
,
1644 unsigned long *removed
,
1647 struct cds_lfht_node
*node
, *next
;
1648 unsigned long nr_bucket
= 0;
1651 if (ht
->split_count
) {
1654 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1655 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1656 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1663 /* Count non-bucket nodes in the table */
1664 node
= bucket_at(ht
, 0);
1666 next
= rcu_dereference(node
->next
);
1667 if (is_removed(next
)) {
1668 if (!is_bucket(next
))
1672 } else if (!is_bucket(next
))
1676 node
= clear_flag(next
);
1677 } while (!is_end(node
));
1678 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1680 if (ht
->split_count
) {
1683 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1684 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1685 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1690 /* called with resize mutex held */
1692 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1693 unsigned long old_size
, unsigned long new_size
)
1695 unsigned long old_order
, new_order
;
1697 old_order
= get_count_order_ulong(old_size
);
1698 new_order
= get_count_order_ulong(new_size
);
1699 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1700 old_size
, old_order
, new_size
, new_order
);
1701 assert(new_size
> old_size
);
1702 init_table(ht
, old_order
+ 1, new_order
);
1705 /* called with resize mutex held */
1707 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1708 unsigned long old_size
, unsigned long new_size
)
1710 unsigned long old_order
, new_order
;
1712 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1713 old_order
= get_count_order_ulong(old_size
);
1714 new_order
= get_count_order_ulong(new_size
);
1715 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1716 old_size
, old_order
, new_size
, new_order
);
1717 assert(new_size
< old_size
);
1719 /* Remove and unlink all bucket nodes to remove. */
1720 fini_table(ht
, new_order
+ 1, old_order
);
1724 /* called with resize mutex held */
1726 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1728 unsigned long new_size
, old_size
;
1731 * Resize table, re-do if the target size has changed under us.
1734 assert(uatomic_read(&ht
->in_progress_resize
));
1735 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1737 ht
->t
.resize_initiated
= 1;
1738 old_size
= ht
->t
.size
;
1739 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1740 if (old_size
< new_size
)
1741 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1742 else if (old_size
> new_size
)
1743 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1744 ht
->t
.resize_initiated
= 0;
1745 /* write resize_initiated before read resize_target */
1747 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1751 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1753 return _uatomic_xchg_monotonic_increase(&ht
->t
.resize_target
, new_size
);
1757 void resize_target_update_count(struct cds_lfht
*ht
,
1758 unsigned long count
)
1760 count
= max(count
, MIN_TABLE_SIZE
);
1761 count
= min(count
, ht
->max_nr_buckets
);
1762 uatomic_set(&ht
->t
.resize_target
, count
);
1765 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1767 resize_target_update_count(ht
, new_size
);
1768 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1769 ht
->flavor
->thread_offline();
1770 pthread_mutex_lock(&ht
->resize_mutex
);
1771 _do_cds_lfht_resize(ht
);
1772 pthread_mutex_unlock(&ht
->resize_mutex
);
1773 ht
->flavor
->thread_online();
1777 void do_resize_cb(struct rcu_head
*head
)
1779 struct rcu_resize_work
*work
=
1780 caa_container_of(head
, struct rcu_resize_work
, head
);
1781 struct cds_lfht
*ht
= work
->ht
;
1783 ht
->flavor
->thread_offline();
1784 pthread_mutex_lock(&ht
->resize_mutex
);
1785 _do_cds_lfht_resize(ht
);
1786 pthread_mutex_unlock(&ht
->resize_mutex
);
1787 ht
->flavor
->thread_online();
1789 cmm_smp_mb(); /* finish resize before decrement */
1790 uatomic_dec(&ht
->in_progress_resize
);
1794 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1796 struct rcu_resize_work
*work
;
1798 /* Store resize_target before read resize_initiated */
1800 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1801 uatomic_inc(&ht
->in_progress_resize
);
1802 cmm_smp_mb(); /* increment resize count before load destroy */
1803 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1804 uatomic_dec(&ht
->in_progress_resize
);
1807 work
= malloc(sizeof(*work
));
1809 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1810 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1815 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1817 unsigned long target_size
= size
<< growth
;
1819 target_size
= min(target_size
, ht
->max_nr_buckets
);
1820 if (resize_target_grow(ht
, target_size
) >= target_size
)
1823 __cds_lfht_resize_lazy_launch(ht
);
1827 * We favor grow operations over shrink. A shrink operation never occurs
1828 * if a grow operation is queued for lazy execution. A grow operation
1829 * cancels any pending shrink lazy execution.
1832 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1833 unsigned long count
)
1835 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1837 count
= max(count
, MIN_TABLE_SIZE
);
1838 count
= min(count
, ht
->max_nr_buckets
);
1840 return; /* Already the right size, no resize needed */
1841 if (count
> size
) { /* lazy grow */
1842 if (resize_target_grow(ht
, count
) >= count
)
1844 } else { /* lazy shrink */
1848 s
= uatomic_cmpxchg(&ht
->t
.resize_target
, size
, count
);
1850 break; /* no resize needed */
1852 return; /* growing is/(was just) in progress */
1854 return; /* some other thread do shrink */
1858 __cds_lfht_resize_lazy_launch(ht
);