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/arch.h>
164 #include <urcu/uatomic.h>
165 #include <urcu/compiler.h>
166 #include <urcu/rculfhash.h>
171 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
173 #define dbg_printf(fmt, args...)
177 * Split-counters lazily update the global counter each 1024
178 * addition/removal. It automatically keeps track of resize required.
179 * We use the bucket length as indicator for need to expand for small
180 * tables and machines lacking per-cpu data suppport.
182 #define COUNT_COMMIT_ORDER 10
183 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
184 #define CHAIN_LEN_TARGET 1
185 #define CHAIN_LEN_RESIZE_THRESHOLD 3
188 * Define the minimum table size.
190 #define MIN_TABLE_ORDER 0
191 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
193 #if (CAA_BITS_PER_LONG == 32)
194 #define MAX_TABLE_ORDER 32
196 #define MAX_TABLE_ORDER 64
200 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
202 #define MIN_PARTITION_PER_THREAD_ORDER 12
203 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
206 #define min(a, b) ((a) < (b) ? (a) : (b))
210 #define max(a, b) ((a) > (b) ? (a) : (b))
214 * The removed flag needs to be updated atomically with the pointer.
215 * It indicates that no node must attach to the node scheduled for
216 * removal, and that node garbage collection must be performed.
217 * The bucket flag does not require to be updated atomically with the
218 * pointer, but it is added as a pointer low bit flag to save space.
220 #define REMOVED_FLAG (1UL << 0)
221 #define BUCKET_FLAG (1UL << 1)
222 #define FLAGS_MASK ((1UL << 2) - 1)
224 /* Value of the end pointer. Should not interact with flags. */
225 #define END_VALUE NULL
228 * ht_items_count: Split-counters counting the number of node addition
229 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
230 * is set at hash table creation.
232 * These are free-running counters, never reset to zero. They count the
233 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
234 * operations to update the global counter. We choose a power-of-2 value
235 * for the trigger to deal with 32 or 64-bit overflow of the counter.
237 struct ht_items_count
{
238 unsigned long add
, del
;
239 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
242 * rcu_level: Contains the per order-index-level bucket node table. The
243 * size of each bucket node table is half the number of hashes contained
244 * in this order (except for order 0). The minimum allocation size
245 * parameter allows combining the bucket node arrays of the lowermost
246 * levels to improve cache locality for small index orders.
249 /* Note: manually update allocation length when adding a field */
250 struct cds_lfht_node nodes
[0];
254 * rcu_table: Contains the size and desired new size if a resize
255 * operation is in progress, as well as the statically-sized array of
256 * rcu_level pointers.
259 unsigned long size
; /* always a power of 2, shared (RCU) */
260 unsigned long resize_target
;
261 int resize_initiated
;
262 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
266 * cds_lfht: Top-level data structure representing a lock-free hash
267 * table. Defined in the implementation file to make it be an opaque
272 unsigned long min_alloc_order
;
273 unsigned long min_alloc_size
;
276 * We need to put the work threads offline (QSBR) when taking this
277 * mutex, because we use synchronize_rcu within this mutex critical
278 * section, which waits on read-side critical sections, and could
279 * therefore cause grace-period deadlock if we hold off RCU G.P.
282 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
283 unsigned int in_progress_resize
, in_progress_destroy
;
284 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
285 void (*func
)(struct rcu_head
*head
));
286 void (*cds_lfht_synchronize_rcu
)(void);
287 void (*cds_lfht_rcu_read_lock
)(void);
288 void (*cds_lfht_rcu_read_unlock
)(void);
289 void (*cds_lfht_rcu_thread_offline
)(void);
290 void (*cds_lfht_rcu_thread_online
)(void);
291 void (*cds_lfht_rcu_register_thread
)(void);
292 void (*cds_lfht_rcu_unregister_thread
)(void);
293 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
294 long count
; /* global approximate item count */
295 struct ht_items_count
*split_count
; /* split item count */
299 * rcu_resize_work: Contains arguments passed to RCU worker thread
300 * responsible for performing lazy resize.
302 struct rcu_resize_work
{
303 struct rcu_head head
;
308 * partition_resize_work: Contains arguments passed to worker threads
309 * executing the hash table resize on partitions of the hash table
310 * assigned to each processor's worker thread.
312 struct partition_resize_work
{
315 unsigned long i
, start
, len
;
316 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
317 unsigned long start
, unsigned long len
);
321 void _cds_lfht_add(struct cds_lfht
*ht
,
322 cds_lfht_match_fct match
,
325 struct cds_lfht_node
*node
,
326 struct cds_lfht_iter
*unique_ret
,
330 * Algorithm to reverse bits in a word by lookup table, extended to
333 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
334 * Originally from Public Domain.
337 static const uint8_t BitReverseTable256
[256] =
339 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
340 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
341 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
342 R6(0), R6(2), R6(1), R6(3)
349 uint8_t bit_reverse_u8(uint8_t v
)
351 return BitReverseTable256
[v
];
354 static __attribute__((unused
))
355 uint32_t bit_reverse_u32(uint32_t v
)
357 return ((uint32_t) bit_reverse_u8(v
) << 24) |
358 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
359 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
360 ((uint32_t) bit_reverse_u8(v
>> 24));
363 static __attribute__((unused
))
364 uint64_t bit_reverse_u64(uint64_t v
)
366 return ((uint64_t) bit_reverse_u8(v
) << 56) |
367 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
368 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
369 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
370 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
371 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
372 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
373 ((uint64_t) bit_reverse_u8(v
>> 56));
377 unsigned long bit_reverse_ulong(unsigned long v
)
379 #if (CAA_BITS_PER_LONG == 32)
380 return bit_reverse_u32(v
);
382 return bit_reverse_u64(v
);
387 * fls: returns the position of the most significant bit.
388 * Returns 0 if no bit is set, else returns the position of the most
389 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
391 #if defined(__i386) || defined(__x86_64)
393 unsigned int fls_u32(uint32_t x
)
401 : "=r" (r
) : "rm" (x
));
407 #if defined(__x86_64)
409 unsigned int fls_u64(uint64_t x
)
417 : "=r" (r
) : "rm" (x
));
424 static __attribute__((unused
))
425 unsigned int fls_u64(uint64_t x
)
432 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
436 if (!(x
& 0xFFFF000000000000ULL
)) {
440 if (!(x
& 0xFF00000000000000ULL
)) {
444 if (!(x
& 0xF000000000000000ULL
)) {
448 if (!(x
& 0xC000000000000000ULL
)) {
452 if (!(x
& 0x8000000000000000ULL
)) {
461 static __attribute__((unused
))
462 unsigned int fls_u32(uint32_t x
)
468 if (!(x
& 0xFFFF0000U
)) {
472 if (!(x
& 0xFF000000U
)) {
476 if (!(x
& 0xF0000000U
)) {
480 if (!(x
& 0xC0000000U
)) {
484 if (!(x
& 0x80000000U
)) {
492 unsigned int fls_ulong(unsigned long x
)
494 #if (CAA_BITS_PER_LONG == 32)
502 * Return the minimum order for which x <= (1UL << order).
503 * Return -1 if x is 0.
505 int get_count_order_u32(uint32_t x
)
510 return fls_u32(x
- 1);
514 * Return the minimum order for which x <= (1UL << order).
515 * Return -1 if x is 0.
517 int get_count_order_ulong(unsigned long x
)
522 return fls_ulong(x
- 1);
526 #define poison_free(ptr) \
529 memset(ptr, 0x42, sizeof(*(ptr))); \
534 #define poison_free(ptr) free(ptr)
538 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
541 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
542 unsigned long count
);
544 static long nr_cpus_mask
= -1;
545 static long split_count_mask
= -1;
547 #if defined(HAVE_SYSCONF)
548 static void ht_init_nr_cpus_mask(void)
552 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
558 * round up number of CPUs to next power of two, so we
559 * can use & for modulo.
561 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
562 nr_cpus_mask
= maxcpus
- 1;
564 #else /* #if defined(HAVE_SYSCONF) */
565 static void ht_init_nr_cpus_mask(void)
569 #endif /* #else #if defined(HAVE_SYSCONF) */
572 void alloc_split_items_count(struct cds_lfht
*ht
)
574 struct ht_items_count
*count
;
576 if (nr_cpus_mask
== -1) {
577 ht_init_nr_cpus_mask();
578 if (nr_cpus_mask
< 0)
579 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
581 split_count_mask
= nr_cpus_mask
;
584 assert(split_count_mask
>= 0);
586 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
587 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
588 assert(ht
->split_count
);
590 ht
->split_count
= NULL
;
595 void free_split_items_count(struct cds_lfht
*ht
)
597 poison_free(ht
->split_count
);
600 #if defined(HAVE_SCHED_GETCPU)
602 int ht_get_split_count_index(unsigned long hash
)
606 assert(split_count_mask
>= 0);
607 cpu
= sched_getcpu();
608 if (caa_unlikely(cpu
< 0))
609 return hash
& split_count_mask
;
611 return cpu
& split_count_mask
;
613 #else /* #if defined(HAVE_SCHED_GETCPU) */
615 int ht_get_split_count_index(unsigned long hash
)
617 return hash
& split_count_mask
;
619 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
622 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
624 unsigned long split_count
;
627 if (caa_unlikely(!ht
->split_count
))
629 index
= ht_get_split_count_index(hash
);
630 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
631 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
634 dbg_printf("add split count %lu\n", split_count
);
635 count
= uatomic_add_return(&ht
->count
,
636 1UL << COUNT_COMMIT_ORDER
);
638 if (!(count
& (count
- 1))) {
639 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
641 dbg_printf("add set global %ld\n", count
);
642 cds_lfht_resize_lazy_count(ht
, size
,
643 count
>> (CHAIN_LEN_TARGET
- 1));
649 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
651 unsigned long split_count
;
654 if (caa_unlikely(!ht
->split_count
))
656 index
= ht_get_split_count_index(hash
);
657 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
658 if (caa_unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
661 dbg_printf("del split count %lu\n", split_count
);
662 count
= uatomic_add_return(&ht
->count
,
663 -(1UL << COUNT_COMMIT_ORDER
));
665 if (!(count
& (count
- 1))) {
666 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
668 dbg_printf("del set global %ld\n", count
);
670 * Don't shrink table if the number of nodes is below a
673 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
675 cds_lfht_resize_lazy_count(ht
, size
,
676 count
>> (CHAIN_LEN_TARGET
- 1));
682 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
686 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
688 count
= uatomic_read(&ht
->count
);
690 * Use bucket-local length for small table expand and for
691 * environments lacking per-cpu data support.
693 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
696 dbg_printf("WARNING: large chain length: %u.\n",
698 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
699 cds_lfht_resize_lazy_grow(ht
, size
,
700 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
704 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
706 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
710 int is_removed(struct cds_lfht_node
*node
)
712 return ((unsigned long) node
) & REMOVED_FLAG
;
716 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
718 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
722 int is_bucket(struct cds_lfht_node
*node
)
724 return ((unsigned long) node
) & BUCKET_FLAG
;
728 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
730 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
734 struct cds_lfht_node
*get_end(void)
736 return (struct cds_lfht_node
*) END_VALUE
;
740 int is_end(struct cds_lfht_node
*node
)
742 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
746 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
749 unsigned long old1
, old2
;
751 old1
= uatomic_read(ptr
);
756 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
761 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
764 ht
->t
.tbl
[0] = calloc(ht
->min_alloc_size
,
765 sizeof(struct cds_lfht_node
));
766 assert(ht
->t
.tbl
[0]);
767 } else if (order
> ht
->min_alloc_order
) {
768 ht
->t
.tbl
[order
] = calloc(1UL << (order
-1),
769 sizeof(struct cds_lfht_node
));
770 assert(ht
->t
.tbl
[order
]);
772 /* Nothing to do for 0 < order && order <= ht->min_alloc_order */
776 * cds_lfht_free_bucket_table() should be called with decreasing order.
777 * When cds_lfht_free_bucket_table(0) is called, it means the whole
781 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
784 poison_free(ht
->t
.tbl
[0]);
785 else if (order
> ht
->min_alloc_order
)
786 poison_free(ht
->t
.tbl
[order
]);
787 /* Nothing to do for 0 < order && order <= ht->min_alloc_order */
791 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
795 if ((__builtin_constant_p(index
) && index
== 0)
796 || index
< ht
->min_alloc_size
) {
797 dbg_printf("bucket index %lu order 0 aridx 0\n", index
);
798 return &ht
->t
.tbl
[0]->nodes
[index
];
801 * equivalent to get_count_order_ulong(index + 1), but optimizes
802 * away the non-existing 0 special-case for
803 * get_count_order_ulong.
805 order
= fls_ulong(index
);
806 dbg_printf("bucket index %lu order %lu aridx %lu\n",
807 index
, order
, index
& ((1UL << (order
- 1)) - 1));
808 return &ht
->t
.tbl
[order
]->nodes
[index
& ((1UL << (order
- 1)) - 1)];
812 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
816 return bucket_at(ht
, hash
& (size
- 1));
820 * Remove all logically deleted nodes from a bucket up to a certain node key.
823 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
825 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
827 assert(!is_bucket(bucket
));
828 assert(!is_removed(bucket
));
829 assert(!is_bucket(node
));
830 assert(!is_removed(node
));
833 /* We can always skip the bucket node initially */
834 iter
= rcu_dereference(iter_prev
->next
);
835 assert(!is_removed(iter
));
836 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
838 * We should never be called with bucket (start of chain)
839 * and logically removed node (end of path compression
840 * marker) being the actual same node. This would be a
841 * bug in the algorithm implementation.
843 assert(bucket
!= node
);
845 if (caa_unlikely(is_end(iter
)))
847 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
849 next
= rcu_dereference(clear_flag(iter
)->next
);
850 if (caa_likely(is_removed(next
)))
852 iter_prev
= clear_flag(iter
);
855 assert(!is_removed(iter
));
857 new_next
= flag_bucket(clear_flag(next
));
859 new_next
= clear_flag(next
);
860 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
866 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
867 struct cds_lfht_node
*old_node
,
868 struct cds_lfht_node
*old_next
,
869 struct cds_lfht_node
*new_node
)
871 struct cds_lfht_node
*bucket
, *ret_next
;
873 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
876 assert(!is_removed(old_node
));
877 assert(!is_bucket(old_node
));
878 assert(!is_removed(new_node
));
879 assert(!is_bucket(new_node
));
880 assert(new_node
!= old_node
);
882 /* Insert after node to be replaced */
883 if (is_removed(old_next
)) {
885 * Too late, the old node has been removed under us
886 * between lookup and replace. Fail.
890 assert(!is_bucket(old_next
));
891 assert(new_node
!= clear_flag(old_next
));
892 new_node
->next
= clear_flag(old_next
);
894 * Here is the whole trick for lock-free replace: we add
895 * the replacement node _after_ the node we want to
896 * replace by atomically setting its next pointer at the
897 * same time we set its removal flag. Given that
898 * the lookups/get next use an iterator aware of the
899 * next pointer, they will either skip the old node due
900 * to the removal flag and see the new node, or use
901 * the old node, but will not see the new one.
903 ret_next
= uatomic_cmpxchg(&old_node
->next
,
904 old_next
, flag_removed(new_node
));
905 if (ret_next
== old_next
)
906 break; /* We performed the replacement. */
911 * Ensure that the old node is not visible to readers anymore:
912 * lookup for the node, and remove it (along with any other
913 * logically removed node) if found.
915 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
916 _cds_lfht_gc_bucket(bucket
, new_node
);
918 assert(is_removed(rcu_dereference(old_node
->next
)));
923 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
924 * mode. A NULL unique_ret allows creation of duplicate keys.
927 void _cds_lfht_add(struct cds_lfht
*ht
,
928 cds_lfht_match_fct match
,
931 struct cds_lfht_node
*node
,
932 struct cds_lfht_iter
*unique_ret
,
935 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
937 struct cds_lfht_node
*bucket
;
939 assert(!is_bucket(node
));
940 assert(!is_removed(node
));
941 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
943 uint32_t chain_len
= 0;
946 * iter_prev points to the non-removed node prior to the
950 /* We can always skip the bucket node initially */
951 iter
= rcu_dereference(iter_prev
->next
);
952 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
954 if (caa_unlikely(is_end(iter
)))
956 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
959 /* bucket node is the first node of the identical-hash-value chain */
960 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
963 next
= rcu_dereference(clear_flag(iter
)->next
);
964 if (caa_unlikely(is_removed(next
)))
970 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
971 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
974 * uniquely adding inserts the node as the first
975 * node of the identical-hash-value node chain.
977 * This semantic ensures no duplicated keys
978 * should ever be observable in the table
979 * (including observe one node by one node
980 * by forward iterations)
982 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
986 *unique_ret
= d_iter
;
990 /* Only account for identical reverse hash once */
991 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
993 check_resize(ht
, size
, ++chain_len
);
994 iter_prev
= clear_flag(iter
);
999 assert(node
!= clear_flag(iter
));
1000 assert(!is_removed(iter_prev
));
1001 assert(!is_removed(iter
));
1002 assert(iter_prev
!= node
);
1004 node
->next
= clear_flag(iter
);
1006 node
->next
= flag_bucket(clear_flag(iter
));
1007 if (is_bucket(iter
))
1008 new_node
= flag_bucket(node
);
1011 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1012 new_node
) != iter
) {
1013 continue; /* retry */
1020 assert(!is_removed(iter
));
1021 if (is_bucket(iter
))
1022 new_next
= flag_bucket(clear_flag(next
));
1024 new_next
= clear_flag(next
);
1025 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1030 unique_ret
->node
= return_node
;
1031 /* unique_ret->next left unset, never used. */
1036 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1037 struct cds_lfht_node
*node
,
1040 struct cds_lfht_node
*bucket
, *next
, *old
;
1042 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1045 /* logically delete the node */
1046 assert(!is_bucket(node
));
1047 assert(!is_removed(node
));
1048 old
= rcu_dereference(node
->next
);
1050 struct cds_lfht_node
*new_next
;
1053 if (caa_unlikely(is_removed(next
)))
1056 assert(is_bucket(next
));
1058 assert(!is_bucket(next
));
1059 new_next
= flag_removed(next
);
1060 old
= uatomic_cmpxchg(&node
->next
, next
, new_next
);
1061 } while (old
!= next
);
1062 /* We performed the (logical) deletion. */
1065 * Ensure that the node is not visible to readers anymore: lookup for
1066 * the node, and remove it (along with any other logically removed node)
1069 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1070 _cds_lfht_gc_bucket(bucket
, node
);
1072 assert(is_removed(rcu_dereference(node
->next
)));
1077 void *partition_resize_thread(void *arg
)
1079 struct partition_resize_work
*work
= arg
;
1081 work
->ht
->cds_lfht_rcu_register_thread();
1082 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1083 work
->ht
->cds_lfht_rcu_unregister_thread();
1088 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1090 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1091 unsigned long start
, unsigned long len
))
1093 unsigned long partition_len
;
1094 struct partition_resize_work
*work
;
1096 unsigned long nr_threads
;
1099 * Note: nr_cpus_mask + 1 is always power of 2.
1100 * We spawn just the number of threads we need to satisfy the minimum
1101 * partition size, up to the number of CPUs in the system.
1103 if (nr_cpus_mask
> 0) {
1104 nr_threads
= min(nr_cpus_mask
+ 1,
1105 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1109 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1110 work
= calloc(nr_threads
, sizeof(*work
));
1112 for (thread
= 0; thread
< nr_threads
; thread
++) {
1113 work
[thread
].ht
= ht
;
1115 work
[thread
].len
= partition_len
;
1116 work
[thread
].start
= thread
* partition_len
;
1117 work
[thread
].fct
= fct
;
1118 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1119 partition_resize_thread
, &work
[thread
]);
1122 for (thread
= 0; thread
< nr_threads
; thread
++) {
1123 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1130 * Holding RCU read lock to protect _cds_lfht_add against memory
1131 * reclaim that could be performed by other call_rcu worker threads (ABA
1134 * When we reach a certain length, we can split this population phase over
1135 * many worker threads, based on the number of CPUs available in the system.
1136 * This should therefore take care of not having the expand lagging behind too
1137 * many concurrent insertion threads by using the scheduler's ability to
1138 * schedule bucket node population fairly with insertions.
1141 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1142 unsigned long start
, unsigned long len
)
1144 unsigned long j
, size
= 1UL << (i
- 1);
1146 assert(i
> MIN_TABLE_ORDER
);
1147 ht
->cds_lfht_rcu_read_lock();
1148 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1149 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1151 assert(j
>= size
&& j
< (size
<< 1));
1152 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1154 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1155 _cds_lfht_add(ht
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1157 ht
->cds_lfht_rcu_read_unlock();
1161 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1164 assert(nr_cpus_mask
!= -1);
1165 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1166 ht
->cds_lfht_rcu_thread_online();
1167 init_table_populate_partition(ht
, i
, 0, len
);
1168 ht
->cds_lfht_rcu_thread_offline();
1171 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1175 void init_table(struct cds_lfht
*ht
,
1176 unsigned long first_order
, unsigned long last_order
)
1180 dbg_printf("init table: first_order %lu last_order %lu\n",
1181 first_order
, last_order
);
1182 assert(first_order
> MIN_TABLE_ORDER
);
1183 for (i
= first_order
; i
<= last_order
; i
++) {
1186 len
= 1UL << (i
- 1);
1187 dbg_printf("init order %lu len: %lu\n", i
, len
);
1189 /* Stop expand if the resize target changes under us */
1190 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (1UL << i
))
1193 cds_lfht_alloc_bucket_table(ht
, i
);
1196 * Set all bucket nodes reverse hash values for a level and
1197 * link all bucket nodes into the table.
1199 init_table_populate(ht
, i
, len
);
1202 * Update table size.
1204 cmm_smp_wmb(); /* populate data before RCU size */
1205 CMM_STORE_SHARED(ht
->t
.size
, 1UL << i
);
1207 dbg_printf("init new size: %lu\n", 1UL << i
);
1208 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1214 * Holding RCU read lock to protect _cds_lfht_remove against memory
1215 * reclaim that could be performed by other call_rcu worker threads (ABA
1217 * For a single level, we logically remove and garbage collect each node.
1219 * As a design choice, we perform logical removal and garbage collection on a
1220 * node-per-node basis to simplify this algorithm. We also assume keeping good
1221 * cache locality of the operation would overweight possible performance gain
1222 * that could be achieved by batching garbage collection for multiple levels.
1223 * However, this would have to be justified by benchmarks.
1225 * Concurrent removal and add operations are helping us perform garbage
1226 * collection of logically removed nodes. We guarantee that all logically
1227 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1228 * invoked to free a hole level of bucket nodes (after a grace period).
1230 * Logical removal and garbage collection can therefore be done in batch or on a
1231 * node-per-node basis, as long as the guarantee above holds.
1233 * When we reach a certain length, we can split this removal over many worker
1234 * threads, based on the number of CPUs available in the system. This should
1235 * take care of not letting resize process lag behind too many concurrent
1236 * updater threads actively inserting into the hash table.
1239 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1240 unsigned long start
, unsigned long len
)
1242 unsigned long j
, size
= 1UL << (i
- 1);
1244 assert(i
> MIN_TABLE_ORDER
);
1245 ht
->cds_lfht_rcu_read_lock();
1246 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1247 struct cds_lfht_node
*fini_node
= bucket_at(ht
, j
);
1249 assert(j
>= size
&& j
< (size
<< 1));
1250 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1252 fini_node
->reverse_hash
= bit_reverse_ulong(j
);
1253 (void) _cds_lfht_del(ht
, size
, fini_node
, 1);
1255 ht
->cds_lfht_rcu_read_unlock();
1259 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1262 assert(nr_cpus_mask
!= -1);
1263 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1264 ht
->cds_lfht_rcu_thread_online();
1265 remove_table_partition(ht
, i
, 0, len
);
1266 ht
->cds_lfht_rcu_thread_offline();
1269 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1273 void fini_table(struct cds_lfht
*ht
,
1274 unsigned long first_order
, unsigned long last_order
)
1277 unsigned long free_by_rcu_order
= 0;
1279 dbg_printf("fini table: first_order %lu last_order %lu\n",
1280 first_order
, last_order
);
1281 assert(first_order
> MIN_TABLE_ORDER
);
1282 for (i
= last_order
; i
>= first_order
; i
--) {
1285 len
= 1UL << (i
- 1);
1286 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1288 /* Stop shrink if the resize target changes under us */
1289 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1292 cmm_smp_wmb(); /* populate data before RCU size */
1293 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1296 * We need to wait for all add operations to reach Q.S. (and
1297 * thus use the new table for lookups) before we can start
1298 * releasing the old bucket nodes. Otherwise their lookup will
1299 * return a logically removed node as insert position.
1301 ht
->cds_lfht_synchronize_rcu();
1302 if (free_by_rcu_order
)
1303 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1306 * Set "removed" flag in bucket nodes about to be removed.
1307 * Unlink all now-logically-removed bucket node pointers.
1308 * Concurrent add/remove operation are helping us doing
1311 remove_table(ht
, i
, len
);
1313 free_by_rcu_order
= i
;
1315 dbg_printf("fini new size: %lu\n", 1UL << i
);
1316 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1320 if (free_by_rcu_order
) {
1321 ht
->cds_lfht_synchronize_rcu();
1322 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1327 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1329 struct cds_lfht_node
*prev
, *node
;
1330 unsigned long order
, len
, i
;
1332 cds_lfht_alloc_bucket_table(ht
, 0);
1334 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1335 node
= bucket_at(ht
, 0);
1336 node
->next
= flag_bucket(get_end());
1337 node
->reverse_hash
= 0;
1339 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1340 len
= 1UL << (order
- 1);
1341 cds_lfht_alloc_bucket_table(ht
, order
);
1343 for (i
= 0; i
< len
; i
++) {
1345 * Now, we are trying to init the node with the
1346 * hash=(len+i) (which is also a bucket with the
1347 * index=(len+i)) and insert it into the hash table,
1348 * so this node has to be inserted after the bucket
1349 * with the index=(len+i)&(len-1)=i. And because there
1350 * is no other non-bucket node nor bucket node with
1351 * larger index/hash inserted, so the bucket node
1352 * being inserted should be inserted directly linked
1353 * after the bucket node with index=i.
1355 prev
= bucket_at(ht
, i
);
1356 node
= bucket_at(ht
, len
+ i
);
1358 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1359 order
, len
+ i
, len
+ i
);
1360 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1362 /* insert after prev */
1363 assert(is_bucket(prev
->next
));
1364 node
->next
= prev
->next
;
1365 prev
->next
= flag_bucket(node
);
1370 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1371 unsigned long min_alloc_size
,
1373 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1374 void (*func
)(struct rcu_head
*head
)),
1375 void (*cds_lfht_synchronize_rcu
)(void),
1376 void (*cds_lfht_rcu_read_lock
)(void),
1377 void (*cds_lfht_rcu_read_unlock
)(void),
1378 void (*cds_lfht_rcu_thread_offline
)(void),
1379 void (*cds_lfht_rcu_thread_online
)(void),
1380 void (*cds_lfht_rcu_register_thread
)(void),
1381 void (*cds_lfht_rcu_unregister_thread
)(void),
1382 pthread_attr_t
*attr
)
1384 struct cds_lfht
*ht
;
1385 unsigned long order
;
1387 /* min_alloc_size must be power of two */
1388 if (!min_alloc_size
|| (min_alloc_size
& (min_alloc_size
- 1)))
1390 /* init_size must be power of two */
1391 if (!init_size
|| (init_size
& (init_size
- 1)))
1393 min_alloc_size
= max(min_alloc_size
, MIN_TABLE_SIZE
);
1394 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1395 ht
= calloc(1, sizeof(struct cds_lfht
));
1398 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1399 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1400 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1401 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1402 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1403 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1404 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1405 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1406 ht
->resize_attr
= attr
;
1407 alloc_split_items_count(ht
);
1408 /* this mutex should not nest in read-side C.S. */
1409 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1410 order
= get_count_order_ulong(init_size
);
1411 ht
->t
.resize_target
= 1UL << order
;
1412 ht
->min_alloc_size
= min_alloc_size
;
1413 ht
->min_alloc_order
= get_count_order_ulong(min_alloc_size
);
1414 cds_lfht_create_bucket(ht
, 1UL << order
);
1415 ht
->t
.size
= 1UL << order
;
1419 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1420 cds_lfht_match_fct match
, const void *key
,
1421 struct cds_lfht_iter
*iter
)
1423 struct cds_lfht_node
*node
, *next
, *bucket
;
1424 unsigned long reverse_hash
, size
;
1426 reverse_hash
= bit_reverse_ulong(hash
);
1428 size
= rcu_dereference(ht
->t
.size
);
1429 bucket
= lookup_bucket(ht
, size
, hash
);
1430 /* We can always skip the bucket node initially */
1431 node
= rcu_dereference(bucket
->next
);
1432 node
= clear_flag(node
);
1434 if (caa_unlikely(is_end(node
))) {
1438 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1442 next
= rcu_dereference(node
->next
);
1443 assert(node
== clear_flag(node
));
1444 if (caa_likely(!is_removed(next
))
1446 && node
->reverse_hash
== reverse_hash
1447 && caa_likely(match(node
, key
))) {
1450 node
= clear_flag(next
);
1452 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1457 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1458 const void *key
, struct cds_lfht_iter
*iter
)
1460 struct cds_lfht_node
*node
, *next
;
1461 unsigned long reverse_hash
;
1464 reverse_hash
= node
->reverse_hash
;
1466 node
= clear_flag(next
);
1469 if (caa_unlikely(is_end(node
))) {
1473 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1477 next
= rcu_dereference(node
->next
);
1478 if (caa_likely(!is_removed(next
))
1480 && caa_likely(match(node
, key
))) {
1483 node
= clear_flag(next
);
1485 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1490 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1492 struct cds_lfht_node
*node
, *next
;
1494 node
= clear_flag(iter
->next
);
1496 if (caa_unlikely(is_end(node
))) {
1500 next
= rcu_dereference(node
->next
);
1501 if (caa_likely(!is_removed(next
))
1502 && !is_bucket(next
)) {
1505 node
= clear_flag(next
);
1507 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1512 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1515 * Get next after first bucket node. The first bucket node is the
1516 * first node of the linked list.
1518 iter
->next
= bucket_at(ht
, 0)->next
;
1519 cds_lfht_next(ht
, iter
);
1522 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1523 struct cds_lfht_node
*node
)
1527 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1528 size
= rcu_dereference(ht
->t
.size
);
1529 _cds_lfht_add(ht
, NULL
, NULL
, size
, node
, NULL
, 0);
1530 ht_count_add(ht
, size
, hash
);
1533 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1535 cds_lfht_match_fct match
,
1537 struct cds_lfht_node
*node
)
1540 struct cds_lfht_iter iter
;
1542 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1543 size
= rcu_dereference(ht
->t
.size
);
1544 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1545 if (iter
.node
== node
)
1546 ht_count_add(ht
, size
, hash
);
1550 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1552 cds_lfht_match_fct match
,
1554 struct cds_lfht_node
*node
)
1557 struct cds_lfht_iter iter
;
1559 node
->reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1560 size
= rcu_dereference(ht
->t
.size
);
1562 _cds_lfht_add(ht
, match
, key
, size
, node
, &iter
, 0);
1563 if (iter
.node
== node
) {
1564 ht_count_add(ht
, size
, hash
);
1568 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1573 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1574 struct cds_lfht_node
*new_node
)
1578 size
= rcu_dereference(ht
->t
.size
);
1579 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1583 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1585 unsigned long size
, hash
;
1588 size
= rcu_dereference(ht
->t
.size
);
1589 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1591 hash
= bit_reverse_ulong(iter
->node
->reverse_hash
);
1592 ht_count_del(ht
, size
, hash
);
1598 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1600 struct cds_lfht_node
*node
;
1601 unsigned long order
, i
, size
;
1603 /* Check that the table is empty */
1604 node
= bucket_at(ht
, 0);
1606 node
= clear_flag(node
)->next
;
1607 if (!is_bucket(node
))
1609 assert(!is_removed(node
));
1610 } while (!is_end(node
));
1612 * size accessed without rcu_dereference because hash table is
1616 /* Internal sanity check: all nodes left should be bucket */
1617 for (i
= 0; i
< size
; i
++) {
1618 node
= bucket_at(ht
, i
);
1619 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1620 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1621 assert(is_bucket(node
->next
));
1624 for (order
= get_count_order_ulong(size
); (long)order
>= 0; order
--)
1625 cds_lfht_free_bucket_table(ht
, order
);
1631 * Should only be called when no more concurrent readers nor writers can
1632 * possibly access the table.
1634 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1638 /* Wait for in-flight resize operations to complete */
1639 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1640 cmm_smp_mb(); /* Store destroy before load resize */
1641 while (uatomic_read(&ht
->in_progress_resize
))
1642 poll(NULL
, 0, 100); /* wait for 100ms */
1643 ret
= cds_lfht_delete_bucket(ht
);
1646 free_split_items_count(ht
);
1648 *attr
= ht
->resize_attr
;
1653 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1654 long *approx_before
,
1655 unsigned long *count
,
1656 unsigned long *removed
,
1659 struct cds_lfht_node
*node
, *next
;
1660 unsigned long nr_bucket
= 0;
1663 if (ht
->split_count
) {
1666 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1667 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1668 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1675 /* Count non-bucket nodes in the table */
1676 node
= bucket_at(ht
, 0);
1678 next
= rcu_dereference(node
->next
);
1679 if (is_removed(next
)) {
1680 if (!is_bucket(next
))
1684 } else if (!is_bucket(next
))
1688 node
= clear_flag(next
);
1689 } while (!is_end(node
));
1690 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1692 if (ht
->split_count
) {
1695 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1696 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1697 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1702 /* called with resize mutex held */
1704 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1705 unsigned long old_size
, unsigned long new_size
)
1707 unsigned long old_order
, new_order
;
1709 old_order
= get_count_order_ulong(old_size
);
1710 new_order
= get_count_order_ulong(new_size
);
1711 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1712 old_size
, old_order
, new_size
, new_order
);
1713 assert(new_size
> old_size
);
1714 init_table(ht
, old_order
+ 1, new_order
);
1717 /* called with resize mutex held */
1719 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1720 unsigned long old_size
, unsigned long new_size
)
1722 unsigned long old_order
, new_order
;
1724 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1725 old_order
= get_count_order_ulong(old_size
);
1726 new_order
= get_count_order_ulong(new_size
);
1727 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1728 old_size
, old_order
, new_size
, new_order
);
1729 assert(new_size
< old_size
);
1731 /* Remove and unlink all bucket nodes to remove. */
1732 fini_table(ht
, new_order
+ 1, old_order
);
1736 /* called with resize mutex held */
1738 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1740 unsigned long new_size
, old_size
;
1743 * Resize table, re-do if the target size has changed under us.
1746 assert(uatomic_read(&ht
->in_progress_resize
));
1747 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1749 ht
->t
.resize_initiated
= 1;
1750 old_size
= ht
->t
.size
;
1751 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1752 if (old_size
< new_size
)
1753 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1754 else if (old_size
> new_size
)
1755 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1756 ht
->t
.resize_initiated
= 0;
1757 /* write resize_initiated before read resize_target */
1759 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1763 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1765 return _uatomic_xchg_monotonic_increase(&ht
->t
.resize_target
, new_size
);
1769 void resize_target_update_count(struct cds_lfht
*ht
,
1770 unsigned long count
)
1772 count
= max(count
, MIN_TABLE_SIZE
);
1773 uatomic_set(&ht
->t
.resize_target
, count
);
1776 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1778 resize_target_update_count(ht
, new_size
);
1779 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1780 ht
->cds_lfht_rcu_thread_offline();
1781 pthread_mutex_lock(&ht
->resize_mutex
);
1782 _do_cds_lfht_resize(ht
);
1783 pthread_mutex_unlock(&ht
->resize_mutex
);
1784 ht
->cds_lfht_rcu_thread_online();
1788 void do_resize_cb(struct rcu_head
*head
)
1790 struct rcu_resize_work
*work
=
1791 caa_container_of(head
, struct rcu_resize_work
, head
);
1792 struct cds_lfht
*ht
= work
->ht
;
1794 ht
->cds_lfht_rcu_thread_offline();
1795 pthread_mutex_lock(&ht
->resize_mutex
);
1796 _do_cds_lfht_resize(ht
);
1797 pthread_mutex_unlock(&ht
->resize_mutex
);
1798 ht
->cds_lfht_rcu_thread_online();
1800 cmm_smp_mb(); /* finish resize before decrement */
1801 uatomic_dec(&ht
->in_progress_resize
);
1805 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1807 struct rcu_resize_work
*work
;
1809 /* Store resize_target before read resize_initiated */
1811 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1812 uatomic_inc(&ht
->in_progress_resize
);
1813 cmm_smp_mb(); /* increment resize count before load destroy */
1814 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1815 uatomic_dec(&ht
->in_progress_resize
);
1818 work
= malloc(sizeof(*work
));
1820 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1821 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1826 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1828 unsigned long target_size
= size
<< growth
;
1830 if (resize_target_grow(ht
, target_size
) >= target_size
)
1833 __cds_lfht_resize_lazy_launch(ht
);
1837 * We favor grow operations over shrink. A shrink operation never occurs
1838 * if a grow operation is queued for lazy execution. A grow operation
1839 * cancels any pending shrink lazy execution.
1842 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1843 unsigned long count
)
1845 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1847 count
= max(count
, MIN_TABLE_SIZE
);
1849 return; /* Already the right size, no resize needed */
1850 if (count
> size
) { /* lazy grow */
1851 if (resize_target_grow(ht
, count
) >= count
)
1853 } else { /* lazy shrink */
1857 s
= uatomic_cmpxchg(&ht
->t
.resize_target
, size
, count
);
1859 break; /* no resize needed */
1861 return; /* growing is/(was just) in progress */
1863 return; /* some other thread do shrink */
1867 __cds_lfht_resize_lazy_launch(ht
);