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 * cds_lfht: Top-level data structure representing a lock-free hash
244 * table. Defined in the implementation file to make it be an opaque
248 unsigned long size
; /* always a power of 2, shared (RCU) */
252 * We need to put the work threads offline (QSBR) when taking this
253 * mutex, because we use synchronize_rcu within this mutex critical
254 * section, which waits on read-side critical sections, and could
255 * therefore cause grace-period deadlock if we hold off RCU G.P.
258 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
259 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
260 unsigned int in_progress_resize
, in_progress_destroy
;
261 unsigned long resize_target
;
262 int resize_initiated
;
263 const struct rcu_flavor_struct
*flavor
;
265 long count
; /* global approximate item count */
266 struct ht_items_count
*split_count
; /* split item count */
268 unsigned long min_alloc_buckets_order
;
269 unsigned long min_nr_alloc_buckets
;
270 unsigned long max_nr_buckets
;
272 * Contains the per order-index-level bucket node table. The size
273 * of each bucket node table is half the number of hashes contained
274 * in this order (except for order 0). The minimum allocation size
275 * parameter allows combining the bucket node arrays of the lowermost
276 * levels to improve cache locality for small index orders.
278 struct cds_lfht_node
*tbl
[MAX_TABLE_ORDER
];
282 * rcu_resize_work: Contains arguments passed to RCU worker thread
283 * responsible for performing lazy resize.
285 struct rcu_resize_work
{
286 struct rcu_head head
;
291 * partition_resize_work: Contains arguments passed to worker threads
292 * executing the hash table resize on partitions of the hash table
293 * assigned to each processor's worker thread.
295 struct partition_resize_work
{
298 unsigned long i
, start
, len
;
299 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
300 unsigned long start
, unsigned long len
);
304 void _cds_lfht_add(struct cds_lfht
*ht
,
305 cds_lfht_match_fct match
,
308 struct cds_lfht_node
*node
,
309 struct cds_lfht_iter
*unique_ret
,
313 * Algorithm to reverse bits in a word by lookup table, extended to
316 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
317 * Originally from Public Domain.
320 static const uint8_t BitReverseTable256
[256] =
322 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
323 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
324 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
325 R6(0), R6(2), R6(1), R6(3)
332 uint8_t bit_reverse_u8(uint8_t v
)
334 return BitReverseTable256
[v
];
337 static __attribute__((unused
))
338 uint32_t bit_reverse_u32(uint32_t v
)
340 return ((uint32_t) bit_reverse_u8(v
) << 24) |
341 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
342 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
343 ((uint32_t) bit_reverse_u8(v
>> 24));
346 static __attribute__((unused
))
347 uint64_t bit_reverse_u64(uint64_t v
)
349 return ((uint64_t) bit_reverse_u8(v
) << 56) |
350 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
351 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
352 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
353 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
354 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
355 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
356 ((uint64_t) bit_reverse_u8(v
>> 56));
360 unsigned long bit_reverse_ulong(unsigned long v
)
362 #if (CAA_BITS_PER_LONG == 32)
363 return bit_reverse_u32(v
);
365 return bit_reverse_u64(v
);
370 * fls: returns the position of the most significant bit.
371 * Returns 0 if no bit is set, else returns the position of the most
372 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
374 #if defined(__i386) || defined(__x86_64)
376 unsigned int fls_u32(uint32_t x
)
384 : "=r" (r
) : "rm" (x
));
390 #if defined(__x86_64)
392 unsigned int fls_u64(uint64_t x
)
400 : "=r" (r
) : "rm" (x
));
407 static __attribute__((unused
))
408 unsigned int fls_u64(uint64_t x
)
415 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
419 if (!(x
& 0xFFFF000000000000ULL
)) {
423 if (!(x
& 0xFF00000000000000ULL
)) {
427 if (!(x
& 0xF000000000000000ULL
)) {
431 if (!(x
& 0xC000000000000000ULL
)) {
435 if (!(x
& 0x8000000000000000ULL
)) {
444 static __attribute__((unused
))
445 unsigned int fls_u32(uint32_t x
)
451 if (!(x
& 0xFFFF0000U
)) {
455 if (!(x
& 0xFF000000U
)) {
459 if (!(x
& 0xF0000000U
)) {
463 if (!(x
& 0xC0000000U
)) {
467 if (!(x
& 0x80000000U
)) {
475 unsigned int fls_ulong(unsigned long x
)
477 #if (CAA_BITS_PER_LONG == 32)
485 * Return the minimum order for which x <= (1UL << order).
486 * Return -1 if x is 0.
488 int get_count_order_u32(uint32_t x
)
493 return fls_u32(x
- 1);
497 * Return the minimum order for which x <= (1UL << order).
498 * Return -1 if x is 0.
500 int get_count_order_ulong(unsigned long x
)
505 return fls_ulong(x
- 1);
509 #define poison_free(ptr) \
512 memset(ptr, 0x42, sizeof(*(ptr))); \
517 #define poison_free(ptr) free(ptr)
521 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
524 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
525 unsigned long count
);
527 static long nr_cpus_mask
= -1;
528 static long split_count_mask
= -1;
530 #if defined(HAVE_SYSCONF)
531 static void ht_init_nr_cpus_mask(void)
535 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
541 * round up number of CPUs to next power of two, so we
542 * can use & for modulo.
544 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
545 nr_cpus_mask
= maxcpus
- 1;
547 #else /* #if defined(HAVE_SYSCONF) */
548 static void ht_init_nr_cpus_mask(void)
552 #endif /* #else #if defined(HAVE_SYSCONF) */
555 void alloc_split_items_count(struct cds_lfht
*ht
)
557 struct ht_items_count
*count
;
559 if (nr_cpus_mask
== -1) {
560 ht_init_nr_cpus_mask();
561 if (nr_cpus_mask
< 0)
562 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
564 split_count_mask
= nr_cpus_mask
;
567 assert(split_count_mask
>= 0);
569 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
570 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
571 assert(ht
->split_count
);
573 ht
->split_count
= NULL
;
578 void free_split_items_count(struct cds_lfht
*ht
)
580 poison_free(ht
->split_count
);
583 #if defined(HAVE_SCHED_GETCPU)
585 int ht_get_split_count_index(unsigned long hash
)
589 assert(split_count_mask
>= 0);
590 cpu
= sched_getcpu();
591 if (caa_unlikely(cpu
< 0))
592 return hash
& split_count_mask
;
594 return cpu
& split_count_mask
;
596 #else /* #if defined(HAVE_SCHED_GETCPU) */
598 int ht_get_split_count_index(unsigned long hash
)
600 return hash
& split_count_mask
;
602 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
605 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
607 unsigned long split_count
;
610 if (caa_unlikely(!ht
->split_count
))
612 index
= ht_get_split_count_index(hash
);
613 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
614 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
617 dbg_printf("add split count %lu\n", split_count
);
618 count
= uatomic_add_return(&ht
->count
,
619 1UL << COUNT_COMMIT_ORDER
);
621 if (!(count
& (count
- 1))) {
622 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
624 dbg_printf("add set global %ld\n", count
);
625 cds_lfht_resize_lazy_count(ht
, size
,
626 count
>> (CHAIN_LEN_TARGET
- 1));
632 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
634 unsigned long split_count
;
637 if (caa_unlikely(!ht
->split_count
))
639 index
= ht_get_split_count_index(hash
);
640 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
641 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
644 dbg_printf("del split count %lu\n", split_count
);
645 count
= uatomic_add_return(&ht
->count
,
646 -(1UL << COUNT_COMMIT_ORDER
));
648 if (!(count
& (count
- 1))) {
649 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
651 dbg_printf("del set global %ld\n", count
);
653 * Don't shrink table if the number of nodes is below a
656 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
658 cds_lfht_resize_lazy_count(ht
, size
,
659 count
>> (CHAIN_LEN_TARGET
- 1));
665 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
669 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
671 count
= uatomic_read(&ht
->count
);
673 * Use bucket-local length for small table expand and for
674 * environments lacking per-cpu data support.
676 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
679 dbg_printf("WARNING: large chain length: %u.\n",
681 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
682 cds_lfht_resize_lazy_grow(ht
, size
,
683 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
687 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
689 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
693 int is_removed(struct cds_lfht_node
*node
)
695 return ((unsigned long) node
) & REMOVED_FLAG
;
699 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
701 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
705 int is_bucket(struct cds_lfht_node
*node
)
707 return ((unsigned long) node
) & BUCKET_FLAG
;
711 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
713 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
717 struct cds_lfht_node
*get_end(void)
719 return (struct cds_lfht_node
*) END_VALUE
;
723 int is_end(struct cds_lfht_node
*node
)
725 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
729 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
732 unsigned long old1
, old2
;
734 old1
= uatomic_read(ptr
);
739 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
744 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
747 ht
->tbl
[0] = calloc(ht
->min_nr_alloc_buckets
,
748 sizeof(struct cds_lfht_node
));
750 } else if (order
> ht
->min_alloc_buckets_order
) {
751 ht
->tbl
[order
] = calloc(1UL << (order
-1),
752 sizeof(struct cds_lfht_node
));
753 assert(ht
->tbl
[order
]);
755 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
759 * cds_lfht_free_bucket_table() should be called with decreasing order.
760 * When cds_lfht_free_bucket_table(0) is called, it means the whole
764 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
767 poison_free(ht
->tbl
[0]);
768 else if (order
> ht
->min_alloc_buckets_order
)
769 poison_free(ht
->tbl
[order
]);
770 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
774 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
778 if ((__builtin_constant_p(index
) && index
== 0)
779 || index
< ht
->min_nr_alloc_buckets
) {
780 dbg_printf("bucket index %lu order 0 aridx 0\n", index
);
781 return &ht
->tbl
[0][index
];
784 * equivalent to get_count_order_ulong(index + 1), but optimizes
785 * away the non-existing 0 special-case for
786 * get_count_order_ulong.
788 order
= fls_ulong(index
);
789 dbg_printf("bucket index %lu order %lu aridx %lu\n",
790 index
, order
, index
& ((1UL << (order
- 1)) - 1));
791 return &ht
->tbl
[order
][index
& ((1UL << (order
- 1)) - 1)];
795 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
799 return bucket_at(ht
, hash
& (size
- 1));
803 * Remove all logically deleted nodes from a bucket up to a certain node key.
806 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
808 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
810 assert(!is_bucket(bucket
));
811 assert(!is_removed(bucket
));
812 assert(!is_bucket(node
));
813 assert(!is_removed(node
));
816 /* We can always skip the bucket node initially */
817 iter
= rcu_dereference(iter_prev
->next
);
818 assert(!is_removed(iter
));
819 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
821 * We should never be called with bucket (start of chain)
822 * and logically removed node (end of path compression
823 * marker) being the actual same node. This would be a
824 * bug in the algorithm implementation.
826 assert(bucket
!= node
);
828 if (caa_unlikely(is_end(iter
)))
830 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
832 next
= rcu_dereference(clear_flag(iter
)->next
);
833 if (caa_likely(is_removed(next
)))
835 iter_prev
= clear_flag(iter
);
838 assert(!is_removed(iter
));
840 new_next
= flag_bucket(clear_flag(next
));
842 new_next
= clear_flag(next
);
843 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
849 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
850 struct cds_lfht_node
*old_node
,
851 struct cds_lfht_node
*old_next
,
852 struct cds_lfht_node
*new_node
)
854 struct cds_lfht_node
*bucket
, *ret_next
;
856 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
859 assert(!is_removed(old_node
));
860 assert(!is_bucket(old_node
));
861 assert(!is_removed(new_node
));
862 assert(!is_bucket(new_node
));
863 assert(new_node
!= old_node
);
865 /* Insert after node to be replaced */
866 if (is_removed(old_next
)) {
868 * Too late, the old node has been removed under us
869 * between lookup and replace. Fail.
873 assert(!is_bucket(old_next
));
874 assert(new_node
!= clear_flag(old_next
));
875 new_node
->next
= clear_flag(old_next
);
877 * Here is the whole trick for lock-free replace: we add
878 * the replacement node _after_ the node we want to
879 * replace by atomically setting its next pointer at the
880 * same time we set its removal flag. Given that
881 * the lookups/get next use an iterator aware of the
882 * next pointer, they will either skip the old node due
883 * to the removal flag and see the new node, or use
884 * the old node, but will not see the new one.
886 ret_next
= uatomic_cmpxchg(&old_node
->next
,
887 old_next
, flag_removed(new_node
));
888 if (ret_next
== old_next
)
889 break; /* We performed the replacement. */
894 * Ensure that the old node is not visible to readers anymore:
895 * lookup for the node, and remove it (along with any other
896 * logically removed node) if found.
898 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
899 _cds_lfht_gc_bucket(bucket
, new_node
);
901 assert(is_removed(rcu_dereference(old_node
->next
)));
906 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
907 * mode. A NULL unique_ret allows creation of duplicate keys.
910 void _cds_lfht_add(struct cds_lfht
*ht
,
911 cds_lfht_match_fct match
,
914 struct cds_lfht_node
*node
,
915 struct cds_lfht_iter
*unique_ret
,
918 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
920 struct cds_lfht_node
*bucket
;
922 assert(!is_bucket(node
));
923 assert(!is_removed(node
));
924 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
926 uint32_t chain_len
= 0;
929 * iter_prev points to the non-removed node prior to the
933 /* We can always skip the bucket node initially */
934 iter
= rcu_dereference(iter_prev
->next
);
935 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
937 if (caa_unlikely(is_end(iter
)))
939 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
942 /* bucket node is the first node of the identical-hash-value chain */
943 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
946 next
= rcu_dereference(clear_flag(iter
)->next
);
947 if (caa_unlikely(is_removed(next
)))
953 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
954 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
957 * uniquely adding inserts the node as the first
958 * node of the identical-hash-value node chain.
960 * This semantic ensures no duplicated keys
961 * should ever be observable in the table
962 * (including observe one node by one node
963 * by forward iterations)
965 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
969 *unique_ret
= d_iter
;
973 /* Only account for identical reverse hash once */
974 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
976 check_resize(ht
, size
, ++chain_len
);
977 iter_prev
= clear_flag(iter
);
982 assert(node
!= clear_flag(iter
));
983 assert(!is_removed(iter_prev
));
984 assert(!is_removed(iter
));
985 assert(iter_prev
!= node
);
987 node
->next
= clear_flag(iter
);
989 node
->next
= flag_bucket(clear_flag(iter
));
991 new_node
= flag_bucket(node
);
994 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
996 continue; /* retry */
1003 assert(!is_removed(iter
));
1004 if (is_bucket(iter
))
1005 new_next
= flag_bucket(clear_flag(next
));
1007 new_next
= clear_flag(next
);
1008 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1013 unique_ret
->node
= return_node
;
1014 /* unique_ret->next left unset, never used. */
1019 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1020 struct cds_lfht_node
*node
,
1023 struct cds_lfht_node
*bucket
, *next
, *old
;
1025 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1028 /* logically delete the node */
1029 assert(!is_bucket(node
));
1030 assert(!is_removed(node
));
1031 old
= rcu_dereference(node
->next
);
1033 struct cds_lfht_node
*new_next
;
1036 if (caa_unlikely(is_removed(next
)))
1039 assert(is_bucket(next
));
1041 assert(!is_bucket(next
));
1042 new_next
= flag_removed(next
);
1043 old
= uatomic_cmpxchg(&node
->next
, next
, new_next
);
1044 } while (old
!= next
);
1045 /* We performed the (logical) deletion. */
1048 * Ensure that the node is not visible to readers anymore: lookup for
1049 * the node, and remove it (along with any other logically removed node)
1052 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1053 _cds_lfht_gc_bucket(bucket
, node
);
1055 assert(is_removed(rcu_dereference(node
->next
)));
1060 void *partition_resize_thread(void *arg
)
1062 struct partition_resize_work
*work
= arg
;
1064 work
->ht
->flavor
->register_thread();
1065 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1066 work
->ht
->flavor
->unregister_thread();
1071 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1073 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1074 unsigned long start
, unsigned long len
))
1076 unsigned long partition_len
;
1077 struct partition_resize_work
*work
;
1079 unsigned long nr_threads
;
1082 * Note: nr_cpus_mask + 1 is always power of 2.
1083 * We spawn just the number of threads we need to satisfy the minimum
1084 * partition size, up to the number of CPUs in the system.
1086 if (nr_cpus_mask
> 0) {
1087 nr_threads
= min(nr_cpus_mask
+ 1,
1088 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1092 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1093 work
= calloc(nr_threads
, sizeof(*work
));
1095 for (thread
= 0; thread
< nr_threads
; thread
++) {
1096 work
[thread
].ht
= ht
;
1098 work
[thread
].len
= partition_len
;
1099 work
[thread
].start
= thread
* partition_len
;
1100 work
[thread
].fct
= fct
;
1101 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1102 partition_resize_thread
, &work
[thread
]);
1105 for (thread
= 0; thread
< nr_threads
; thread
++) {
1106 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1113 * Holding RCU read lock to protect _cds_lfht_add against memory
1114 * reclaim that could be performed by other call_rcu worker threads (ABA
1117 * When we reach a certain length, we can split this population phase over
1118 * many worker threads, based on the number of CPUs available in the system.
1119 * This should therefore take care of not having the expand lagging behind too
1120 * many concurrent insertion threads by using the scheduler's ability to
1121 * schedule bucket node population fairly with insertions.
1124 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1125 unsigned long start
, unsigned long len
)
1127 unsigned long j
, size
= 1UL << (i
- 1);
1129 assert(i
> MIN_TABLE_ORDER
);
1130 ht
->flavor
->read_lock();
1131 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1132 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1134 assert(j
>= size
&& j
< (size
<< 1));
1135 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1137 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1138 _cds_lfht_add(ht
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1140 ht
->flavor
->read_unlock();
1144 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1147 assert(nr_cpus_mask
!= -1);
1148 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1149 ht
->flavor
->thread_online();
1150 init_table_populate_partition(ht
, i
, 0, len
);
1151 ht
->flavor
->thread_offline();
1154 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1158 void init_table(struct cds_lfht
*ht
,
1159 unsigned long first_order
, unsigned long last_order
)
1163 dbg_printf("init table: first_order %lu last_order %lu\n",
1164 first_order
, last_order
);
1165 assert(first_order
> MIN_TABLE_ORDER
);
1166 for (i
= first_order
; i
<= last_order
; i
++) {
1169 len
= 1UL << (i
- 1);
1170 dbg_printf("init order %lu len: %lu\n", i
, len
);
1172 /* Stop expand if the resize target changes under us */
1173 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1176 cds_lfht_alloc_bucket_table(ht
, i
);
1179 * Set all bucket nodes reverse hash values for a level and
1180 * link all bucket nodes into the table.
1182 init_table_populate(ht
, i
, len
);
1185 * Update table size.
1187 cmm_smp_wmb(); /* populate data before RCU size */
1188 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1190 dbg_printf("init new size: %lu\n", 1UL << i
);
1191 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1197 * Holding RCU read lock to protect _cds_lfht_remove against memory
1198 * reclaim that could be performed by other call_rcu worker threads (ABA
1200 * For a single level, we logically remove and garbage collect each node.
1202 * As a design choice, we perform logical removal and garbage collection on a
1203 * node-per-node basis to simplify this algorithm. We also assume keeping good
1204 * cache locality of the operation would overweight possible performance gain
1205 * that could be achieved by batching garbage collection for multiple levels.
1206 * However, this would have to be justified by benchmarks.
1208 * Concurrent removal and add operations are helping us perform garbage
1209 * collection of logically removed nodes. We guarantee that all logically
1210 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1211 * invoked to free a hole level of bucket nodes (after a grace period).
1213 * Logical removal and garbage collection can therefore be done in batch or on a
1214 * node-per-node basis, as long as the guarantee above holds.
1216 * When we reach a certain length, we can split this removal over many worker
1217 * threads, based on the number of CPUs available in the system. This should
1218 * take care of not letting resize process lag behind too many concurrent
1219 * updater threads actively inserting into the hash table.
1222 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1223 unsigned long start
, unsigned long len
)
1225 unsigned long j
, size
= 1UL << (i
- 1);
1227 assert(i
> MIN_TABLE_ORDER
);
1228 ht
->flavor
->read_lock();
1229 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1230 struct cds_lfht_node
*fini_node
= bucket_at(ht
, j
);
1232 assert(j
>= size
&& j
< (size
<< 1));
1233 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1235 fini_node
->reverse_hash
= bit_reverse_ulong(j
);
1236 (void) _cds_lfht_del(ht
, size
, fini_node
, 1);
1238 ht
->flavor
->read_unlock();
1242 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1245 assert(nr_cpus_mask
!= -1);
1246 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1247 ht
->flavor
->thread_online();
1248 remove_table_partition(ht
, i
, 0, len
);
1249 ht
->flavor
->thread_offline();
1252 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1256 void fini_table(struct cds_lfht
*ht
,
1257 unsigned long first_order
, unsigned long last_order
)
1260 unsigned long free_by_rcu_order
= 0;
1262 dbg_printf("fini table: first_order %lu last_order %lu\n",
1263 first_order
, last_order
);
1264 assert(first_order
> MIN_TABLE_ORDER
);
1265 for (i
= last_order
; i
>= first_order
; i
--) {
1268 len
= 1UL << (i
- 1);
1269 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1271 /* Stop shrink if the resize target changes under us */
1272 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1275 cmm_smp_wmb(); /* populate data before RCU size */
1276 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1279 * We need to wait for all add operations to reach Q.S. (and
1280 * thus use the new table for lookups) before we can start
1281 * releasing the old bucket nodes. Otherwise their lookup will
1282 * return a logically removed node as insert position.
1284 ht
->flavor
->update_synchronize_rcu();
1285 if (free_by_rcu_order
)
1286 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1289 * Set "removed" flag in bucket nodes about to be removed.
1290 * Unlink all now-logically-removed bucket node pointers.
1291 * Concurrent add/remove operation are helping us doing
1294 remove_table(ht
, i
, len
);
1296 free_by_rcu_order
= i
;
1298 dbg_printf("fini new size: %lu\n", 1UL << i
);
1299 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1303 if (free_by_rcu_order
) {
1304 ht
->flavor
->update_synchronize_rcu();
1305 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1310 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1312 struct cds_lfht_node
*prev
, *node
;
1313 unsigned long order
, len
, i
;
1315 cds_lfht_alloc_bucket_table(ht
, 0);
1317 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1318 node
= bucket_at(ht
, 0);
1319 node
->next
= flag_bucket(get_end());
1320 node
->reverse_hash
= 0;
1322 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1323 len
= 1UL << (order
- 1);
1324 cds_lfht_alloc_bucket_table(ht
, order
);
1326 for (i
= 0; i
< len
; i
++) {
1328 * Now, we are trying to init the node with the
1329 * hash=(len+i) (which is also a bucket with the
1330 * index=(len+i)) and insert it into the hash table,
1331 * so this node has to be inserted after the bucket
1332 * with the index=(len+i)&(len-1)=i. And because there
1333 * is no other non-bucket node nor bucket node with
1334 * larger index/hash inserted, so the bucket node
1335 * being inserted should be inserted directly linked
1336 * after the bucket node with index=i.
1338 prev
= bucket_at(ht
, i
);
1339 node
= bucket_at(ht
, len
+ i
);
1341 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1342 order
, len
+ i
, len
+ i
);
1343 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1345 /* insert after prev */
1346 assert(is_bucket(prev
->next
));
1347 node
->next
= prev
->next
;
1348 prev
->next
= flag_bucket(node
);
1353 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1354 unsigned long min_nr_alloc_buckets
,
1355 unsigned long max_nr_buckets
,
1357 const struct rcu_flavor_struct
*flavor
,
1358 pthread_attr_t
*attr
)
1360 struct cds_lfht
*ht
;
1361 unsigned long order
;
1363 /* min_nr_alloc_buckets must be power of two */
1364 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1367 /* init_size must be power of two */
1368 if (!init_size
|| (init_size
& (init_size
- 1)))
1371 if (!max_nr_buckets
)
1372 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1374 /* max_nr_buckets must be power of two */
1375 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1378 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1379 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1380 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1381 init_size
= min(init_size
, max_nr_buckets
);
1382 ht
= calloc(1, sizeof(struct cds_lfht
));
1385 ht
->flavor
= flavor
;
1386 ht
->resize_attr
= attr
;
1387 alloc_split_items_count(ht
);
1388 /* this mutex should not nest in read-side C.S. */
1389 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1390 order
= get_count_order_ulong(init_size
);
1391 ht
->resize_target
= 1UL << order
;
1392 ht
->min_nr_alloc_buckets
= min_nr_alloc_buckets
;
1393 ht
->min_alloc_buckets_order
= get_count_order_ulong(min_nr_alloc_buckets
);
1394 ht
->max_nr_buckets
= max_nr_buckets
;
1395 cds_lfht_create_bucket(ht
, 1UL << order
);
1396 ht
->size
= 1UL << order
;
1400 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1401 cds_lfht_match_fct match
, const void *key
,
1402 struct cds_lfht_iter
*iter
)
1404 struct cds_lfht_node
*node
, *next
, *bucket
;
1405 unsigned long reverse_hash
, size
;
1407 reverse_hash
= bit_reverse_ulong(hash
);
1409 size
= rcu_dereference(ht
->size
);
1410 bucket
= lookup_bucket(ht
, size
, hash
);
1411 /* We can always skip the bucket node initially */
1412 node
= rcu_dereference(bucket
->next
);
1413 node
= clear_flag(node
);
1415 if (caa_unlikely(is_end(node
))) {
1419 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1423 next
= rcu_dereference(node
->next
);
1424 assert(node
== clear_flag(node
));
1425 if (caa_likely(!is_removed(next
))
1427 && node
->reverse_hash
== reverse_hash
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_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1439 const void *key
, struct cds_lfht_iter
*iter
)
1441 struct cds_lfht_node
*node
, *next
;
1442 unsigned long reverse_hash
;
1445 reverse_hash
= node
->reverse_hash
;
1447 node
= clear_flag(next
);
1450 if (caa_unlikely(is_end(node
))) {
1454 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1458 next
= rcu_dereference(node
->next
);
1459 if (caa_likely(!is_removed(next
))
1461 && caa_likely(match(node
, key
))) {
1464 node
= clear_flag(next
);
1466 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1471 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1473 struct cds_lfht_node
*node
, *next
;
1475 node
= clear_flag(iter
->next
);
1477 if (caa_unlikely(is_end(node
))) {
1481 next
= rcu_dereference(node
->next
);
1482 if (caa_likely(!is_removed(next
))
1483 && !is_bucket(next
)) {
1486 node
= clear_flag(next
);
1488 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1493 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1496 * Get next after first bucket node. The first bucket node is the
1497 * first node of the linked list.
1499 iter
->next
= bucket_at(ht
, 0)->next
;
1500 cds_lfht_next(ht
, iter
);
1503 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1504 struct cds_lfht_node
*node
)
1508 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1509 size
= rcu_dereference(ht
->size
);
1510 _cds_lfht_add(ht
, NULL
, NULL
, size
, node
, NULL
, 0);
1511 ht_count_add(ht
, size
, hash
);
1514 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1516 cds_lfht_match_fct match
,
1518 struct cds_lfht_node
*node
)
1521 struct cds_lfht_iter iter
;
1523 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1524 size
= rcu_dereference(ht
->size
);
1525 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1526 if (iter
.node
== node
)
1527 ht_count_add(ht
, size
, hash
);
1531 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1533 cds_lfht_match_fct match
,
1535 struct cds_lfht_node
*node
)
1538 struct cds_lfht_iter iter
;
1540 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1541 size
= rcu_dereference(ht
->size
);
1543 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1544 if (iter
.node
== node
) {
1545 ht_count_add(ht
, size
, hash
);
1549 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1554 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1555 struct cds_lfht_node
*new_node
)
1559 size
= rcu_dereference(ht
->size
);
1560 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1564 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1566 unsigned long size
, hash
;
1569 size
= rcu_dereference(ht
->size
);
1570 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1572 hash
= bit_reverse_ulong(iter
->node
->reverse_hash
);
1573 ht_count_del(ht
, size
, hash
);
1579 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1581 struct cds_lfht_node
*node
;
1582 unsigned long order
, i
, size
;
1584 /* Check that the table is empty */
1585 node
= bucket_at(ht
, 0);
1587 node
= clear_flag(node
)->next
;
1588 if (!is_bucket(node
))
1590 assert(!is_removed(node
));
1591 } while (!is_end(node
));
1593 * size accessed without rcu_dereference because hash table is
1597 /* Internal sanity check: all nodes left should be bucket */
1598 for (i
= 0; i
< size
; i
++) {
1599 node
= bucket_at(ht
, i
);
1600 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1601 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1602 assert(is_bucket(node
->next
));
1605 for (order
= get_count_order_ulong(size
); (long)order
>= 0; order
--)
1606 cds_lfht_free_bucket_table(ht
, order
);
1612 * Should only be called when no more concurrent readers nor writers can
1613 * possibly access the table.
1615 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1619 /* Wait for in-flight resize operations to complete */
1620 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1621 cmm_smp_mb(); /* Store destroy before load resize */
1622 while (uatomic_read(&ht
->in_progress_resize
))
1623 poll(NULL
, 0, 100); /* wait for 100ms */
1624 ret
= cds_lfht_delete_bucket(ht
);
1627 free_split_items_count(ht
);
1629 *attr
= ht
->resize_attr
;
1634 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1635 long *approx_before
,
1636 unsigned long *count
,
1637 unsigned long *removed
,
1640 struct cds_lfht_node
*node
, *next
;
1641 unsigned long nr_bucket
= 0;
1644 if (ht
->split_count
) {
1647 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1648 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1649 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1656 /* Count non-bucket nodes in the table */
1657 node
= bucket_at(ht
, 0);
1659 next
= rcu_dereference(node
->next
);
1660 if (is_removed(next
)) {
1661 if (!is_bucket(next
))
1665 } else if (!is_bucket(next
))
1669 node
= clear_flag(next
);
1670 } while (!is_end(node
));
1671 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1673 if (ht
->split_count
) {
1676 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1677 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1678 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1683 /* called with resize mutex held */
1685 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1686 unsigned long old_size
, unsigned long new_size
)
1688 unsigned long old_order
, new_order
;
1690 old_order
= get_count_order_ulong(old_size
);
1691 new_order
= get_count_order_ulong(new_size
);
1692 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1693 old_size
, old_order
, new_size
, new_order
);
1694 assert(new_size
> old_size
);
1695 init_table(ht
, old_order
+ 1, new_order
);
1698 /* called with resize mutex held */
1700 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1701 unsigned long old_size
, unsigned long new_size
)
1703 unsigned long old_order
, new_order
;
1705 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1706 old_order
= get_count_order_ulong(old_size
);
1707 new_order
= get_count_order_ulong(new_size
);
1708 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1709 old_size
, old_order
, new_size
, new_order
);
1710 assert(new_size
< old_size
);
1712 /* Remove and unlink all bucket nodes to remove. */
1713 fini_table(ht
, new_order
+ 1, old_order
);
1717 /* called with resize mutex held */
1719 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1721 unsigned long new_size
, old_size
;
1724 * Resize table, re-do if the target size has changed under us.
1727 assert(uatomic_read(&ht
->in_progress_resize
));
1728 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1730 ht
->resize_initiated
= 1;
1731 old_size
= ht
->size
;
1732 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1733 if (old_size
< new_size
)
1734 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1735 else if (old_size
> new_size
)
1736 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1737 ht
->resize_initiated
= 0;
1738 /* write resize_initiated before read resize_target */
1740 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1744 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1746 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
1750 void resize_target_update_count(struct cds_lfht
*ht
,
1751 unsigned long count
)
1753 count
= max(count
, MIN_TABLE_SIZE
);
1754 count
= min(count
, ht
->max_nr_buckets
);
1755 uatomic_set(&ht
->resize_target
, count
);
1758 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1760 resize_target_update_count(ht
, new_size
);
1761 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1762 ht
->flavor
->thread_offline();
1763 pthread_mutex_lock(&ht
->resize_mutex
);
1764 _do_cds_lfht_resize(ht
);
1765 pthread_mutex_unlock(&ht
->resize_mutex
);
1766 ht
->flavor
->thread_online();
1770 void do_resize_cb(struct rcu_head
*head
)
1772 struct rcu_resize_work
*work
=
1773 caa_container_of(head
, struct rcu_resize_work
, head
);
1774 struct cds_lfht
*ht
= work
->ht
;
1776 ht
->flavor
->thread_offline();
1777 pthread_mutex_lock(&ht
->resize_mutex
);
1778 _do_cds_lfht_resize(ht
);
1779 pthread_mutex_unlock(&ht
->resize_mutex
);
1780 ht
->flavor
->thread_online();
1782 cmm_smp_mb(); /* finish resize before decrement */
1783 uatomic_dec(&ht
->in_progress_resize
);
1787 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1789 struct rcu_resize_work
*work
;
1791 /* Store resize_target before read resize_initiated */
1793 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
1794 uatomic_inc(&ht
->in_progress_resize
);
1795 cmm_smp_mb(); /* increment resize count before load destroy */
1796 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1797 uatomic_dec(&ht
->in_progress_resize
);
1800 work
= malloc(sizeof(*work
));
1802 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1803 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1808 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1810 unsigned long target_size
= size
<< growth
;
1812 target_size
= min(target_size
, ht
->max_nr_buckets
);
1813 if (resize_target_grow(ht
, target_size
) >= target_size
)
1816 __cds_lfht_resize_lazy_launch(ht
);
1820 * We favor grow operations over shrink. A shrink operation never occurs
1821 * if a grow operation is queued for lazy execution. A grow operation
1822 * cancels any pending shrink lazy execution.
1825 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1826 unsigned long count
)
1828 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1830 count
= max(count
, MIN_TABLE_SIZE
);
1831 count
= min(count
, ht
->max_nr_buckets
);
1833 return; /* Already the right size, no resize needed */
1834 if (count
> size
) { /* lazy grow */
1835 if (resize_target_grow(ht
, count
) >= count
)
1837 } else { /* lazy shrink */
1841 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
1843 break; /* no resize needed */
1845 return; /* growing is/(was just) in progress */
1847 return; /* some other thread do shrink */
1851 __cds_lfht_resize_lazy_launch(ht
);