4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
32 * Some specificities of this Lock-Free Resizable RCU Hash Table
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
94 * this order (except for order 0).
95 * - synchronzie_rcu is used to garbage-collect the old dummy node table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
102 * hash table hash table the last all dummy node tables
103 * order size dummy node 0 1 2 3 4 5 6(index)
110 * 5 32 16 1 1 2 4 8 16
111 * 6 64 32 1 1 2 4 8 16 32
113 * When growing/shrinking, we only focus on the last dummy node table
114 * which size is (!order ? 1 : (1 << (order -1))).
116 * Example for growing/shrinking:
117 * grow hash table from order 5 to 6: init the index=6 dummy node table
118 * shrink hash table from order 6 to 5: fini the index=6 dummy node table
120 * A bit of ascii art explanation:
122 * Order index is the off-by-one compare to the actual power of 2 because
123 * we use index 0 to deal with the 0 special-case.
125 * This shows the nodes for a small table ordered by reversed bits:
137 * This shows the nodes in order of non-reversed bits, linked by
138 * reversed-bit order.
143 * 2 | | 2 010 010 <- |
144 * | | | 3 011 110 | <- |
145 * 3 -> | | | 4 100 001 | |
161 #include <urcu-call-rcu.h>
162 #include <urcu/arch.h>
163 #include <urcu/uatomic.h>
164 #include <urcu/compiler.h>
165 #include <urcu/rculfhash.h>
170 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
172 #define dbg_printf(fmt, args...)
176 * Split-counters lazily update the global counter each 1024
177 * addition/removal. It automatically keeps track of resize required.
178 * We use the bucket length as indicator for need to expand for small
179 * tables and machines lacking per-cpu data suppport.
181 #define COUNT_COMMIT_ORDER 10
182 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
183 #define CHAIN_LEN_TARGET 1
184 #define CHAIN_LEN_RESIZE_THRESHOLD 3
187 * Define the minimum table size.
189 #define MIN_TABLE_SIZE 1
191 #if (CAA_BITS_PER_LONG == 32)
192 #define MAX_TABLE_ORDER 32
194 #define MAX_TABLE_ORDER 64
198 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
200 #define MIN_PARTITION_PER_THREAD_ORDER 12
201 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
204 #define min(a, b) ((a) < (b) ? (a) : (b))
208 #define max(a, b) ((a) > (b) ? (a) : (b))
212 * The removed flag needs to be updated atomically with the pointer.
213 * It indicates that no node must attach to the node scheduled for
214 * removal, and that node garbage collection must be performed.
215 * The dummy flag does not require to be updated atomically with the
216 * pointer, but it is added as a pointer low bit flag to save space.
218 #define REMOVED_FLAG (1UL << 0)
219 #define DUMMY_FLAG (1UL << 1)
220 #define FLAGS_MASK ((1UL << 2) - 1)
222 /* Value of the end pointer. Should not interact with flags. */
223 #define END_VALUE NULL
225 struct ht_items_count
{
226 unsigned long add
, del
;
227 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
230 /* Note: manually update allocation length when adding a field */
231 struct _cds_lfht_node nodes
[0];
235 unsigned long size
; /* always a power of 2, shared (RCU) */
236 unsigned long resize_target
;
237 int resize_initiated
;
238 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
243 cds_lfht_hash_fct hash_fct
;
244 cds_lfht_compare_fct compare_fct
;
245 unsigned long min_alloc_order
;
246 unsigned long min_alloc_size
;
247 unsigned long hash_seed
;
250 * We need to put the work threads offline (QSBR) when taking this
251 * mutex, because we use synchronize_rcu within this mutex critical
252 * section, which waits on read-side critical sections, and could
253 * therefore cause grace-period deadlock if we hold off RCU G.P.
256 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
257 unsigned int in_progress_resize
, in_progress_destroy
;
258 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
259 void (*func
)(struct rcu_head
*head
));
260 void (*cds_lfht_synchronize_rcu
)(void);
261 void (*cds_lfht_rcu_read_lock
)(void);
262 void (*cds_lfht_rcu_read_unlock
)(void);
263 void (*cds_lfht_rcu_thread_offline
)(void);
264 void (*cds_lfht_rcu_thread_online
)(void);
265 void (*cds_lfht_rcu_register_thread
)(void);
266 void (*cds_lfht_rcu_unregister_thread
)(void);
267 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
268 long count
; /* global approximate item count */
269 struct ht_items_count
*split_count
; /* split item count */
272 struct rcu_resize_work
{
273 struct rcu_head head
;
277 struct partition_resize_work
{
280 unsigned long i
, start
, len
;
281 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
282 unsigned long start
, unsigned long len
);
286 void _cds_lfht_add(struct cds_lfht
*ht
,
288 struct cds_lfht_node
*node
,
289 struct cds_lfht_iter
*unique_ret
,
293 * Algorithm to reverse bits in a word by lookup table, extended to
296 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
297 * Originally from Public Domain.
300 static const uint8_t BitReverseTable256
[256] =
302 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
303 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
304 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
305 R6(0), R6(2), R6(1), R6(3)
312 uint8_t bit_reverse_u8(uint8_t v
)
314 return BitReverseTable256
[v
];
317 static __attribute__((unused
))
318 uint32_t bit_reverse_u32(uint32_t v
)
320 return ((uint32_t) bit_reverse_u8(v
) << 24) |
321 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
322 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
323 ((uint32_t) bit_reverse_u8(v
>> 24));
326 static __attribute__((unused
))
327 uint64_t bit_reverse_u64(uint64_t v
)
329 return ((uint64_t) bit_reverse_u8(v
) << 56) |
330 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
331 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
332 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
333 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
334 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
335 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
336 ((uint64_t) bit_reverse_u8(v
>> 56));
340 unsigned long bit_reverse_ulong(unsigned long v
)
342 #if (CAA_BITS_PER_LONG == 32)
343 return bit_reverse_u32(v
);
345 return bit_reverse_u64(v
);
350 * fls: returns the position of the most significant bit.
351 * Returns 0 if no bit is set, else returns the position of the most
352 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
354 #if defined(__i386) || defined(__x86_64)
356 unsigned int fls_u32(uint32_t x
)
364 : "=r" (r
) : "rm" (x
));
370 #if defined(__x86_64)
372 unsigned int fls_u64(uint64_t x
)
380 : "=r" (r
) : "rm" (x
));
387 static __attribute__((unused
))
388 unsigned int fls_u64(uint64_t x
)
395 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
399 if (!(x
& 0xFFFF000000000000ULL
)) {
403 if (!(x
& 0xFF00000000000000ULL
)) {
407 if (!(x
& 0xF000000000000000ULL
)) {
411 if (!(x
& 0xC000000000000000ULL
)) {
415 if (!(x
& 0x8000000000000000ULL
)) {
424 static __attribute__((unused
))
425 unsigned int fls_u32(uint32_t x
)
431 if (!(x
& 0xFFFF0000U
)) {
435 if (!(x
& 0xFF000000U
)) {
439 if (!(x
& 0xF0000000U
)) {
443 if (!(x
& 0xC0000000U
)) {
447 if (!(x
& 0x80000000U
)) {
455 unsigned int fls_ulong(unsigned long x
)
457 #if (CAA_BITS_PER_lONG == 32)
465 * Return the minimum order for which x <= (1UL << order).
466 * Return -1 if x is 0.
468 int get_count_order_u32(uint32_t x
)
473 return fls_u32(x
- 1);
477 * Return the minimum order for which x <= (1UL << order).
478 * Return -1 if x is 0.
480 int get_count_order_ulong(unsigned long x
)
485 return fls_ulong(x
- 1);
489 #define poison_free(ptr) \
491 memset(ptr, 0x42, sizeof(*(ptr))); \
495 #define poison_free(ptr) free(ptr)
499 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
502 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
503 unsigned long count
);
505 static long nr_cpus_mask
= -1;
506 static long split_count_mask
= -1;
508 #if defined(HAVE_SYSCONF)
509 static void ht_init_nr_cpus_mask(void)
513 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
519 * round up number of CPUs to next power of two, so we
520 * can use & for modulo.
522 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
523 nr_cpus_mask
= maxcpus
- 1;
525 #else /* #if defined(HAVE_SYSCONF) */
526 static void ht_init_nr_cpus_mask(void)
530 #endif /* #else #if defined(HAVE_SYSCONF) */
533 struct ht_items_count
*alloc_split_items_count(void)
535 struct ht_items_count
*count
;
537 if (nr_cpus_mask
== -1) {
538 ht_init_nr_cpus_mask();
539 if (nr_cpus_mask
< 0)
540 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
542 split_count_mask
= nr_cpus_mask
;
545 assert(split_count_mask
>= 0);
546 return calloc(split_count_mask
+ 1, sizeof(*count
));
550 void free_split_items_count(struct ht_items_count
*count
)
555 #if defined(HAVE_SCHED_GETCPU)
557 int ht_get_split_count_index(unsigned long hash
)
561 assert(split_count_mask
>= 0);
562 cpu
= sched_getcpu();
563 if (unlikely(cpu
< 0))
564 return hash
& split_count_mask
;
566 return cpu
& split_count_mask
;
568 #else /* #if defined(HAVE_SCHED_GETCPU) */
570 int ht_get_split_count_index(unsigned long hash
)
572 return hash
& split_count_mask
;
574 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
577 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
579 unsigned long split_count
;
582 if (unlikely(!ht
->split_count
))
584 index
= ht_get_split_count_index(hash
);
585 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
586 if (unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
589 dbg_printf("add split count %lu\n", split_count
);
590 count
= uatomic_add_return(&ht
->count
,
591 1UL << COUNT_COMMIT_ORDER
);
593 if (!(count
& (count
- 1))) {
594 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
596 dbg_printf("add set global %ld\n", count
);
597 cds_lfht_resize_lazy_count(ht
, size
,
598 count
>> (CHAIN_LEN_TARGET
- 1));
604 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
606 unsigned long split_count
;
609 if (unlikely(!ht
->split_count
))
611 index
= ht_get_split_count_index(hash
);
612 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
613 if (unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
616 dbg_printf("del split count %lu\n", split_count
);
617 count
= uatomic_add_return(&ht
->count
,
618 -(1UL << COUNT_COMMIT_ORDER
));
620 if (!(count
& (count
- 1))) {
621 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
623 dbg_printf("del set global %ld\n", count
);
625 * Don't shrink table if the number of nodes is below a
628 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
630 cds_lfht_resize_lazy_count(ht
, size
,
631 count
>> (CHAIN_LEN_TARGET
- 1));
637 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
641 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
643 count
= uatomic_read(&ht
->count
);
645 * Use bucket-local length for small table expand and for
646 * environments lacking per-cpu data support.
648 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
651 dbg_printf("WARNING: large chain length: %u.\n",
653 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
654 cds_lfht_resize_lazy(ht
, size
,
655 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
659 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
661 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
665 int is_removed(struct cds_lfht_node
*node
)
667 return ((unsigned long) node
) & REMOVED_FLAG
;
671 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
673 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
677 int is_dummy(struct cds_lfht_node
*node
)
679 return ((unsigned long) node
) & DUMMY_FLAG
;
683 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
685 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
689 struct cds_lfht_node
*get_end(void)
691 return (struct cds_lfht_node
*) END_VALUE
;
695 int is_end(struct cds_lfht_node
*node
)
697 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
701 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
703 unsigned long old1
, old2
;
705 old1
= uatomic_read(ptr
);
710 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
715 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
718 unsigned long index
, order
;
721 index
= hash
& (size
- 1);
723 if (index
< ht
->min_alloc_size
) {
724 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
726 return &ht
->t
.tbl
[0]->nodes
[index
];
729 * equivalent to get_count_order_ulong(index + 1), but optimizes
730 * away the non-existing 0 special-case for
731 * get_count_order_ulong.
733 order
= fls_ulong(index
);
734 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
735 hash
, index
, order
, index
& ((1UL << (order
- 1)) - 1));
736 return &ht
->t
.tbl
[order
]->nodes
[index
& ((1UL << (order
- 1)) - 1)];
740 * Remove all logically deleted nodes from a bucket up to a certain node key.
743 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
745 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
747 assert(!is_dummy(dummy
));
748 assert(!is_removed(dummy
));
749 assert(!is_dummy(node
));
750 assert(!is_removed(node
));
753 /* We can always skip the dummy node initially */
754 iter
= rcu_dereference(iter_prev
->p
.next
);
755 assert(!is_removed(iter
));
756 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
758 * We should never be called with dummy (start of chain)
759 * and logically removed node (end of path compression
760 * marker) being the actual same node. This would be a
761 * bug in the algorithm implementation.
763 assert(dummy
!= node
);
765 if (unlikely(is_end(iter
)))
767 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
769 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
770 if (likely(is_removed(next
)))
772 iter_prev
= clear_flag(iter
);
775 assert(!is_removed(iter
));
777 new_next
= flag_dummy(clear_flag(next
));
779 new_next
= clear_flag(next
);
780 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
786 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
787 struct cds_lfht_node
*old_node
,
788 struct cds_lfht_node
*old_next
,
789 struct cds_lfht_node
*new_node
)
791 struct cds_lfht_node
*dummy
, *ret_next
;
792 struct _cds_lfht_node
*lookup
;
794 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
797 assert(!is_removed(old_node
));
798 assert(!is_dummy(old_node
));
799 assert(!is_removed(new_node
));
800 assert(!is_dummy(new_node
));
801 assert(new_node
!= old_node
);
803 /* Insert after node to be replaced */
804 if (is_removed(old_next
)) {
806 * Too late, the old node has been removed under us
807 * between lookup and replace. Fail.
811 assert(!is_dummy(old_next
));
812 assert(new_node
!= clear_flag(old_next
));
813 new_node
->p
.next
= clear_flag(old_next
);
815 * Here is the whole trick for lock-free replace: we add
816 * the replacement node _after_ the node we want to
817 * replace by atomically setting its next pointer at the
818 * same time we set its removal flag. Given that
819 * the lookups/get next use an iterator aware of the
820 * next pointer, they will either skip the old node due
821 * to the removal flag and see the new node, or use
822 * the old node, but will not see the new one.
824 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
825 old_next
, flag_removed(new_node
));
826 if (ret_next
== old_next
)
827 break; /* We performed the replacement. */
832 * Ensure that the old node is not visible to readers anymore:
833 * lookup for the node, and remove it (along with any other
834 * logically removed node) if found.
836 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
837 dummy
= (struct cds_lfht_node
*) lookup
;
838 _cds_lfht_gc_bucket(dummy
, new_node
);
840 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
845 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
846 * mode. A NULL unique_ret allows creation of duplicate keys.
849 void _cds_lfht_add(struct cds_lfht
*ht
,
851 struct cds_lfht_node
*node
,
852 struct cds_lfht_iter
*unique_ret
,
855 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
857 struct _cds_lfht_node
*lookup
;
859 assert(!is_dummy(node
));
860 assert(!is_removed(node
));
861 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
863 uint32_t chain_len
= 0;
866 * iter_prev points to the non-removed node prior to the
869 iter_prev
= (struct cds_lfht_node
*) lookup
;
870 /* We can always skip the dummy node initially */
871 iter
= rcu_dereference(iter_prev
->p
.next
);
872 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
874 if (unlikely(is_end(iter
)))
876 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
879 /* dummy node is the first node of the identical-hash-value chain */
880 if (dummy
&& clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
)
883 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
884 if (unlikely(is_removed(next
)))
890 && clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
) {
891 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
894 * uniquely adding inserts the node as the first
895 * node of the identical-hash-value node chain.
897 * This semantic ensures no duplicated keys
898 * should ever be observable in the table
899 * (including observe one node by one node
900 * by forward iterations)
902 cds_lfht_next_duplicate(ht
, &d_iter
);
906 *unique_ret
= d_iter
;
910 /* Only account for identical reverse hash once */
911 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
913 check_resize(ht
, size
, ++chain_len
);
914 iter_prev
= clear_flag(iter
);
919 assert(node
!= clear_flag(iter
));
920 assert(!is_removed(iter_prev
));
921 assert(!is_removed(iter
));
922 assert(iter_prev
!= node
);
924 node
->p
.next
= clear_flag(iter
);
926 node
->p
.next
= flag_dummy(clear_flag(iter
));
928 new_node
= flag_dummy(node
);
931 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
933 continue; /* retry */
940 assert(!is_removed(iter
));
942 new_next
= flag_dummy(clear_flag(next
));
944 new_next
= clear_flag(next
);
945 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
950 unique_ret
->node
= return_node
;
951 /* unique_ret->next left unset, never used. */
956 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
957 struct cds_lfht_node
*node
,
960 struct cds_lfht_node
*dummy
, *next
, *old
;
961 struct _cds_lfht_node
*lookup
;
963 if (!node
) /* Return -ENOENT if asked to delete NULL node */
966 /* logically delete the node */
967 assert(!is_dummy(node
));
968 assert(!is_removed(node
));
969 old
= rcu_dereference(node
->p
.next
);
971 struct cds_lfht_node
*new_next
;
974 if (unlikely(is_removed(next
)))
977 assert(is_dummy(next
));
979 assert(!is_dummy(next
));
980 new_next
= flag_removed(next
);
981 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
982 } while (old
!= next
);
983 /* We performed the (logical) deletion. */
986 * Ensure that the node is not visible to readers anymore: lookup for
987 * the node, and remove it (along with any other logically removed node)
990 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
991 dummy
= (struct cds_lfht_node
*) lookup
;
992 _cds_lfht_gc_bucket(dummy
, node
);
994 assert(is_removed(rcu_dereference(node
->p
.next
)));
999 void *partition_resize_thread(void *arg
)
1001 struct partition_resize_work
*work
= arg
;
1003 work
->ht
->cds_lfht_rcu_register_thread();
1004 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1005 work
->ht
->cds_lfht_rcu_unregister_thread();
1010 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1012 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1013 unsigned long start
, unsigned long len
))
1015 unsigned long partition_len
;
1016 struct partition_resize_work
*work
;
1018 unsigned long nr_threads
;
1021 * Note: nr_cpus_mask + 1 is always power of 2.
1022 * We spawn just the number of threads we need to satisfy the minimum
1023 * partition size, up to the number of CPUs in the system.
1025 if (nr_cpus_mask
> 0) {
1026 nr_threads
= min(nr_cpus_mask
+ 1,
1027 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1031 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1032 work
= calloc(nr_threads
, sizeof(*work
));
1034 for (thread
= 0; thread
< nr_threads
; thread
++) {
1035 work
[thread
].ht
= ht
;
1037 work
[thread
].len
= partition_len
;
1038 work
[thread
].start
= thread
* partition_len
;
1039 work
[thread
].fct
= fct
;
1040 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1041 partition_resize_thread
, &work
[thread
]);
1044 for (thread
= 0; thread
< nr_threads
; thread
++) {
1045 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1052 * Holding RCU read lock to protect _cds_lfht_add against memory
1053 * reclaim that could be performed by other call_rcu worker threads (ABA
1056 * When we reach a certain length, we can split this population phase over
1057 * many worker threads, based on the number of CPUs available in the system.
1058 * This should therefore take care of not having the expand lagging behind too
1059 * many concurrent insertion threads by using the scheduler's ability to
1060 * schedule dummy node population fairly with insertions.
1063 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1064 unsigned long start
, unsigned long len
)
1068 assert(i
> ht
->min_alloc_order
);
1069 ht
->cds_lfht_rcu_read_lock();
1070 for (j
= start
; j
< start
+ len
; j
++) {
1071 struct cds_lfht_node
*new_node
=
1072 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1074 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1075 i
, j
, (1UL << (i
- 1)) + j
);
1076 new_node
->p
.reverse_hash
=
1077 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1078 _cds_lfht_add(ht
, 1UL << (i
- 1),
1081 ht
->cds_lfht_rcu_read_unlock();
1085 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1088 assert(nr_cpus_mask
!= -1);
1089 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1090 ht
->cds_lfht_rcu_thread_online();
1091 init_table_populate_partition(ht
, i
, 0, len
);
1092 ht
->cds_lfht_rcu_thread_offline();
1095 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1099 void init_table(struct cds_lfht
*ht
,
1100 unsigned long first_order
, unsigned long last_order
)
1104 dbg_printf("init table: first_order %lu last_order %lu\n",
1105 first_order
, last_order
);
1106 assert(first_order
> ht
->min_alloc_order
);
1107 for (i
= first_order
; i
<= last_order
; i
++) {
1110 len
= 1UL << (i
- 1);
1111 dbg_printf("init order %lu len: %lu\n", i
, len
);
1113 /* Stop expand if the resize target changes under us */
1114 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (1UL << i
))
1117 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1118 assert(ht
->t
.tbl
[i
]);
1121 * Set all dummy nodes reverse hash values for a level and
1122 * link all dummy nodes into the table.
1124 init_table_populate(ht
, i
, len
);
1127 * Update table size.
1129 cmm_smp_wmb(); /* populate data before RCU size */
1130 CMM_STORE_SHARED(ht
->t
.size
, 1UL << i
);
1132 dbg_printf("init new size: %lu\n", 1UL << i
);
1133 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1139 * Holding RCU read lock to protect _cds_lfht_remove against memory
1140 * reclaim that could be performed by other call_rcu worker threads (ABA
1142 * For a single level, we logically remove and garbage collect each node.
1144 * As a design choice, we perform logical removal and garbage collection on a
1145 * node-per-node basis to simplify this algorithm. We also assume keeping good
1146 * cache locality of the operation would overweight possible performance gain
1147 * that could be achieved by batching garbage collection for multiple levels.
1148 * However, this would have to be justified by benchmarks.
1150 * Concurrent removal and add operations are helping us perform garbage
1151 * collection of logically removed nodes. We guarantee that all logically
1152 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1153 * invoked to free a hole level of dummy nodes (after a grace period).
1155 * Logical removal and garbage collection can therefore be done in batch or on a
1156 * node-per-node basis, as long as the guarantee above holds.
1158 * When we reach a certain length, we can split this removal over many worker
1159 * threads, based on the number of CPUs available in the system. This should
1160 * take care of not letting resize process lag behind too many concurrent
1161 * updater threads actively inserting into the hash table.
1164 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1165 unsigned long start
, unsigned long len
)
1169 assert(i
> ht
->min_alloc_order
);
1170 ht
->cds_lfht_rcu_read_lock();
1171 for (j
= start
; j
< start
+ len
; j
++) {
1172 struct cds_lfht_node
*fini_node
=
1173 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1175 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1176 i
, j
, (1UL << (i
- 1)) + j
);
1177 fini_node
->p
.reverse_hash
=
1178 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1179 (void) _cds_lfht_del(ht
, 1UL << (i
- 1), fini_node
, 1);
1181 ht
->cds_lfht_rcu_read_unlock();
1185 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1188 assert(nr_cpus_mask
!= -1);
1189 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1190 ht
->cds_lfht_rcu_thread_online();
1191 remove_table_partition(ht
, i
, 0, len
);
1192 ht
->cds_lfht_rcu_thread_offline();
1195 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1199 void fini_table(struct cds_lfht
*ht
,
1200 unsigned long first_order
, unsigned long last_order
)
1203 void *free_by_rcu
= NULL
;
1205 dbg_printf("fini table: first_order %lu last_order %lu\n",
1206 first_order
, last_order
);
1207 assert(first_order
> ht
->min_alloc_order
);
1208 for (i
= last_order
; i
>= first_order
; i
--) {
1211 len
= 1UL << (i
- 1);
1212 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1214 /* Stop shrink if the resize target changes under us */
1215 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1218 cmm_smp_wmb(); /* populate data before RCU size */
1219 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1222 * We need to wait for all add operations to reach Q.S. (and
1223 * thus use the new table for lookups) before we can start
1224 * releasing the old dummy nodes. Otherwise their lookup will
1225 * return a logically removed node as insert position.
1227 ht
->cds_lfht_synchronize_rcu();
1232 * Set "removed" flag in dummy nodes about to be removed.
1233 * Unlink all now-logically-removed dummy node pointers.
1234 * Concurrent add/remove operation are helping us doing
1237 remove_table(ht
, i
, len
);
1239 free_by_rcu
= ht
->t
.tbl
[i
];
1241 dbg_printf("fini new size: %lu\n", 1UL << i
);
1242 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1247 ht
->cds_lfht_synchronize_rcu();
1253 void cds_lfht_create_dummy(struct cds_lfht
*ht
, unsigned long size
)
1255 struct _cds_lfht_node
*prev
, *node
;
1256 unsigned long order
, len
, i
, j
;
1258 ht
->t
.tbl
[0] = calloc(1, ht
->min_alloc_size
* sizeof(struct _cds_lfht_node
));
1259 assert(ht
->t
.tbl
[0]);
1261 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1262 ht
->t
.tbl
[0]->nodes
[0].next
= flag_dummy(get_end());
1263 ht
->t
.tbl
[0]->nodes
[0].reverse_hash
= 0;
1265 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1266 len
= 1UL << (order
- 1);
1267 if (order
<= ht
->min_alloc_order
) {
1268 ht
->t
.tbl
[order
] = (struct rcu_level
*) (ht
->t
.tbl
[0]->nodes
+ len
);
1270 ht
->t
.tbl
[order
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1271 assert(ht
->t
.tbl
[order
]);
1275 prev
= ht
->t
.tbl
[i
]->nodes
;
1276 for (j
= 0; j
< len
; j
++) {
1277 if (j
& (j
- 1)) { /* Between power of 2 */
1279 } else if (j
) { /* At each power of 2 */
1281 prev
= ht
->t
.tbl
[i
]->nodes
;
1284 node
= &ht
->t
.tbl
[order
]->nodes
[j
];
1285 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1287 node
->next
= prev
->next
;
1288 assert(is_dummy(node
->next
));
1289 node
->reverse_hash
= bit_reverse_ulong(j
+ len
);
1290 prev
->next
= flag_dummy((struct cds_lfht_node
*)node
);
1295 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1296 cds_lfht_compare_fct compare_fct
,
1297 unsigned long hash_seed
,
1298 unsigned long init_size
,
1299 unsigned long min_alloc_size
,
1301 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1302 void (*func
)(struct rcu_head
*head
)),
1303 void (*cds_lfht_synchronize_rcu
)(void),
1304 void (*cds_lfht_rcu_read_lock
)(void),
1305 void (*cds_lfht_rcu_read_unlock
)(void),
1306 void (*cds_lfht_rcu_thread_offline
)(void),
1307 void (*cds_lfht_rcu_thread_online
)(void),
1308 void (*cds_lfht_rcu_register_thread
)(void),
1309 void (*cds_lfht_rcu_unregister_thread
)(void),
1310 pthread_attr_t
*attr
)
1312 struct cds_lfht
*ht
;
1313 unsigned long order
;
1315 /* min_alloc_size must be power of two */
1316 if (!min_alloc_size
|| (min_alloc_size
& (min_alloc_size
- 1)))
1318 /* init_size must be power of two */
1319 if (!init_size
|| (init_size
& (init_size
- 1)))
1321 min_alloc_size
= max(min_alloc_size
, MIN_TABLE_SIZE
);
1322 init_size
= max(init_size
, min_alloc_size
);
1323 ht
= calloc(1, sizeof(struct cds_lfht
));
1325 ht
->hash_fct
= hash_fct
;
1326 ht
->compare_fct
= compare_fct
;
1327 ht
->hash_seed
= hash_seed
;
1328 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1329 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1330 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1331 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1332 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1333 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1334 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1335 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1336 ht
->resize_attr
= attr
;
1337 ht
->split_count
= alloc_split_items_count();
1338 /* this mutex should not nest in read-side C.S. */
1339 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1341 order
= get_count_order_ulong(init_size
);
1342 ht
->t
.resize_target
= 1UL << order
;
1343 cds_lfht_create_dummy(ht
, 1UL << order
);
1344 ht
->t
.size
= 1UL << order
;
1345 ht
->min_alloc_size
= min_alloc_size
;
1346 ht
->min_alloc_order
= get_count_order_ulong(min_alloc_size
);
1350 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1351 struct cds_lfht_iter
*iter
)
1353 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1354 struct _cds_lfht_node
*lookup
;
1355 unsigned long hash
, reverse_hash
, size
;
1357 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1358 reverse_hash
= bit_reverse_ulong(hash
);
1360 size
= rcu_dereference(ht
->t
.size
);
1361 lookup
= lookup_bucket(ht
, size
, hash
);
1362 dummy_node
= (struct cds_lfht_node
*) lookup
;
1363 /* We can always skip the dummy node initially */
1364 node
= rcu_dereference(dummy_node
->p
.next
);
1365 node
= clear_flag(node
);
1367 if (unlikely(is_end(node
))) {
1371 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1375 next
= rcu_dereference(node
->p
.next
);
1376 if (likely(!is_removed(next
))
1378 && clear_flag(node
)->p
.reverse_hash
== reverse_hash
1379 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1382 node
= clear_flag(next
);
1384 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1389 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1391 struct cds_lfht_node
*node
, *next
;
1392 unsigned long reverse_hash
;
1397 reverse_hash
= node
->p
.reverse_hash
;
1399 key_len
= node
->key_len
;
1401 node
= clear_flag(next
);
1404 if (unlikely(is_end(node
))) {
1408 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1412 next
= rcu_dereference(node
->p
.next
);
1413 if (likely(!is_removed(next
))
1415 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1418 node
= clear_flag(next
);
1420 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1425 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1427 struct cds_lfht_node
*node
, *next
;
1429 node
= clear_flag(iter
->next
);
1431 if (unlikely(is_end(node
))) {
1435 next
= rcu_dereference(node
->p
.next
);
1436 if (likely(!is_removed(next
))
1437 && !is_dummy(next
)) {
1440 node
= clear_flag(next
);
1442 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1447 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1449 struct _cds_lfht_node
*lookup
;
1452 * Get next after first dummy node. The first dummy node is the
1453 * first node of the linked list.
1455 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1456 iter
->next
= lookup
->next
;
1457 cds_lfht_next(ht
, iter
);
1460 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1462 unsigned long hash
, size
;
1464 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1465 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1467 size
= rcu_dereference(ht
->t
.size
);
1468 _cds_lfht_add(ht
, size
, node
, NULL
, 0);
1469 ht_count_add(ht
, size
, hash
);
1472 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1473 struct cds_lfht_node
*node
)
1475 unsigned long hash
, size
;
1476 struct cds_lfht_iter iter
;
1478 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1479 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1481 size
= rcu_dereference(ht
->t
.size
);
1482 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1483 if (iter
.node
== node
)
1484 ht_count_add(ht
, size
, hash
);
1488 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1489 struct cds_lfht_node
*node
)
1491 unsigned long hash
, size
;
1492 struct cds_lfht_iter iter
;
1494 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1495 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1497 size
= rcu_dereference(ht
->t
.size
);
1499 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1500 if (iter
.node
== node
) {
1501 ht_count_add(ht
, size
, hash
);
1505 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1510 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1511 struct cds_lfht_node
*new_node
)
1515 size
= rcu_dereference(ht
->t
.size
);
1516 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1520 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1522 unsigned long size
, hash
;
1525 size
= rcu_dereference(ht
->t
.size
);
1526 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1528 hash
= bit_reverse_ulong(iter
->node
->p
.reverse_hash
);
1529 ht_count_del(ht
, size
, hash
);
1535 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1537 struct cds_lfht_node
*node
;
1538 struct _cds_lfht_node
*lookup
;
1539 unsigned long order
, i
, size
;
1541 /* Check that the table is empty */
1542 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1543 node
= (struct cds_lfht_node
*) lookup
;
1545 node
= clear_flag(node
)->p
.next
;
1546 if (!is_dummy(node
))
1548 assert(!is_removed(node
));
1549 } while (!is_end(node
));
1551 * size accessed without rcu_dereference because hash table is
1555 /* Internal sanity check: all nodes left should be dummy */
1556 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1559 len
= !order
? 1 : 1UL << (order
- 1);
1560 for (i
= 0; i
< len
; i
++) {
1561 dbg_printf("delete order %lu i %lu hash %lu\n",
1563 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1564 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1567 if (order
== ht
->min_alloc_order
)
1568 poison_free(ht
->t
.tbl
[0]);
1569 else if (order
> ht
->min_alloc_order
)
1570 poison_free(ht
->t
.tbl
[order
]);
1571 /* Nothing to delete for order < ht->min_alloc_order */
1577 * Should only be called when no more concurrent readers nor writers can
1578 * possibly access the table.
1580 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1584 /* Wait for in-flight resize operations to complete */
1585 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1586 cmm_smp_mb(); /* Store destroy before load resize */
1587 while (uatomic_read(&ht
->in_progress_resize
))
1588 poll(NULL
, 0, 100); /* wait for 100ms */
1589 ret
= cds_lfht_delete_dummy(ht
);
1592 free_split_items_count(ht
->split_count
);
1594 *attr
= ht
->resize_attr
;
1599 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1600 long *approx_before
,
1601 unsigned long *count
,
1602 unsigned long *removed
,
1605 struct cds_lfht_node
*node
, *next
;
1606 struct _cds_lfht_node
*lookup
;
1607 unsigned long nr_dummy
= 0;
1610 if (split_count_mask
>= 0) {
1613 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1614 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1615 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1622 /* Count non-dummy nodes in the table */
1623 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1624 node
= (struct cds_lfht_node
*) lookup
;
1626 next
= rcu_dereference(node
->p
.next
);
1627 if (is_removed(next
)) {
1628 if (!is_dummy(next
))
1632 } else if (!is_dummy(next
))
1636 node
= clear_flag(next
);
1637 } while (!is_end(node
));
1638 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1640 if (split_count_mask
>= 0) {
1643 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1644 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1645 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1650 /* called with resize mutex held */
1652 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1653 unsigned long old_size
, unsigned long new_size
)
1655 unsigned long old_order
, new_order
;
1657 old_order
= get_count_order_ulong(old_size
);
1658 new_order
= get_count_order_ulong(new_size
);
1659 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1660 old_size
, old_order
, new_size
, new_order
);
1661 assert(new_size
> old_size
);
1662 init_table(ht
, old_order
+ 1, new_order
);
1665 /* called with resize mutex held */
1667 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1668 unsigned long old_size
, unsigned long new_size
)
1670 unsigned long old_order
, new_order
;
1672 new_size
= max(new_size
, ht
->min_alloc_size
);
1673 old_order
= get_count_order_ulong(old_size
);
1674 new_order
= get_count_order_ulong(new_size
);
1675 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1676 old_size
, old_order
, new_size
, new_order
);
1677 assert(new_size
< old_size
);
1679 /* Remove and unlink all dummy nodes to remove. */
1680 fini_table(ht
, new_order
+ 1, old_order
);
1684 /* called with resize mutex held */
1686 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1688 unsigned long new_size
, old_size
;
1691 * Resize table, re-do if the target size has changed under us.
1694 assert(uatomic_read(&ht
->in_progress_resize
));
1695 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1697 ht
->t
.resize_initiated
= 1;
1698 old_size
= ht
->t
.size
;
1699 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1700 if (old_size
< new_size
)
1701 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1702 else if (old_size
> new_size
)
1703 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1704 ht
->t
.resize_initiated
= 0;
1705 /* write resize_initiated before read resize_target */
1707 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1711 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1714 return _uatomic_max(&ht
->t
.resize_target
,
1715 size
<< growth_order
);
1719 void resize_target_update_count(struct cds_lfht
*ht
,
1720 unsigned long count
)
1722 count
= max(count
, ht
->min_alloc_size
);
1723 uatomic_set(&ht
->t
.resize_target
, count
);
1726 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1728 resize_target_update_count(ht
, new_size
);
1729 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1730 ht
->cds_lfht_rcu_thread_offline();
1731 pthread_mutex_lock(&ht
->resize_mutex
);
1732 _do_cds_lfht_resize(ht
);
1733 pthread_mutex_unlock(&ht
->resize_mutex
);
1734 ht
->cds_lfht_rcu_thread_online();
1738 void do_resize_cb(struct rcu_head
*head
)
1740 struct rcu_resize_work
*work
=
1741 caa_container_of(head
, struct rcu_resize_work
, head
);
1742 struct cds_lfht
*ht
= work
->ht
;
1744 ht
->cds_lfht_rcu_thread_offline();
1745 pthread_mutex_lock(&ht
->resize_mutex
);
1746 _do_cds_lfht_resize(ht
);
1747 pthread_mutex_unlock(&ht
->resize_mutex
);
1748 ht
->cds_lfht_rcu_thread_online();
1750 cmm_smp_mb(); /* finish resize before decrement */
1751 uatomic_dec(&ht
->in_progress_resize
);
1755 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1757 struct rcu_resize_work
*work
;
1758 unsigned long target_size
;
1760 target_size
= resize_target_update(ht
, size
, growth
);
1761 /* Store resize_target before read resize_initiated */
1763 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1764 uatomic_inc(&ht
->in_progress_resize
);
1765 cmm_smp_mb(); /* increment resize count before load destroy */
1766 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1767 uatomic_dec(&ht
->in_progress_resize
);
1770 work
= malloc(sizeof(*work
));
1772 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1773 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1778 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1779 unsigned long count
)
1781 struct rcu_resize_work
*work
;
1783 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1785 resize_target_update_count(ht
, count
);
1786 /* Store resize_target before read resize_initiated */
1788 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1789 uatomic_inc(&ht
->in_progress_resize
);
1790 cmm_smp_mb(); /* increment resize count before load destroy */
1791 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1792 uatomic_dec(&ht
->in_progress_resize
);
1795 work
= malloc(sizeof(*work
));
1797 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1798 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);