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) \
492 memset(ptr, 0x42, sizeof(*(ptr))); \
497 #define poison_free(ptr) free(ptr)
501 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
504 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
505 unsigned long count
);
507 static long nr_cpus_mask
= -1;
508 static long split_count_mask
= -1;
510 #if defined(HAVE_SYSCONF)
511 static void ht_init_nr_cpus_mask(void)
515 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
521 * round up number of CPUs to next power of two, so we
522 * can use & for modulo.
524 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
525 nr_cpus_mask
= maxcpus
- 1;
527 #else /* #if defined(HAVE_SYSCONF) */
528 static void ht_init_nr_cpus_mask(void)
532 #endif /* #else #if defined(HAVE_SYSCONF) */
535 void alloc_split_items_count(struct cds_lfht
*ht
)
537 struct ht_items_count
*count
;
539 if (nr_cpus_mask
== -1) {
540 ht_init_nr_cpus_mask();
541 if (nr_cpus_mask
< 0)
542 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
544 split_count_mask
= nr_cpus_mask
;
547 assert(split_count_mask
>= 0);
549 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
550 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
551 assert(ht
->split_count
);
553 ht
->split_count
= NULL
;
558 void free_split_items_count(struct cds_lfht
*ht
)
560 poison_free(ht
->split_count
);
563 #if defined(HAVE_SCHED_GETCPU)
565 int ht_get_split_count_index(unsigned long hash
)
569 assert(split_count_mask
>= 0);
570 cpu
= sched_getcpu();
571 if (unlikely(cpu
< 0))
572 return hash
& split_count_mask
;
574 return cpu
& split_count_mask
;
576 #else /* #if defined(HAVE_SCHED_GETCPU) */
578 int ht_get_split_count_index(unsigned long hash
)
580 return hash
& split_count_mask
;
582 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
585 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
587 unsigned long split_count
;
590 if (unlikely(!ht
->split_count
))
592 index
= ht_get_split_count_index(hash
);
593 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
594 if (unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
597 dbg_printf("add split count %lu\n", split_count
);
598 count
= uatomic_add_return(&ht
->count
,
599 1UL << COUNT_COMMIT_ORDER
);
601 if (!(count
& (count
- 1))) {
602 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
604 dbg_printf("add set global %ld\n", count
);
605 cds_lfht_resize_lazy_count(ht
, size
,
606 count
>> (CHAIN_LEN_TARGET
- 1));
612 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
614 unsigned long split_count
;
617 if (unlikely(!ht
->split_count
))
619 index
= ht_get_split_count_index(hash
);
620 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
621 if (unlikely(!(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
624 dbg_printf("del split count %lu\n", split_count
);
625 count
= uatomic_add_return(&ht
->count
,
626 -(1UL << COUNT_COMMIT_ORDER
));
628 if (!(count
& (count
- 1))) {
629 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
631 dbg_printf("del set global %ld\n", count
);
633 * Don't shrink table if the number of nodes is below a
636 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
638 cds_lfht_resize_lazy_count(ht
, size
,
639 count
>> (CHAIN_LEN_TARGET
- 1));
645 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
649 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
651 count
= uatomic_read(&ht
->count
);
653 * Use bucket-local length for small table expand and for
654 * environments lacking per-cpu data support.
656 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
659 dbg_printf("WARNING: large chain length: %u.\n",
661 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
662 cds_lfht_resize_lazy(ht
, size
,
663 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
667 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
669 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
673 int is_removed(struct cds_lfht_node
*node
)
675 return ((unsigned long) node
) & REMOVED_FLAG
;
679 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
681 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
685 int is_dummy(struct cds_lfht_node
*node
)
687 return ((unsigned long) node
) & DUMMY_FLAG
;
691 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
693 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
697 struct cds_lfht_node
*get_end(void)
699 return (struct cds_lfht_node
*) END_VALUE
;
703 int is_end(struct cds_lfht_node
*node
)
705 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
709 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
711 unsigned long old1
, old2
;
713 old1
= uatomic_read(ptr
);
718 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
723 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
726 unsigned long index
, order
;
729 index
= hash
& (size
- 1);
731 if (index
< ht
->min_alloc_size
) {
732 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
734 return &ht
->t
.tbl
[0]->nodes
[index
];
737 * equivalent to get_count_order_ulong(index + 1), but optimizes
738 * away the non-existing 0 special-case for
739 * get_count_order_ulong.
741 order
= fls_ulong(index
);
742 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
743 hash
, index
, order
, index
& ((1UL << (order
- 1)) - 1));
744 return &ht
->t
.tbl
[order
]->nodes
[index
& ((1UL << (order
- 1)) - 1)];
748 * Remove all logically deleted nodes from a bucket up to a certain node key.
751 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
753 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
755 assert(!is_dummy(dummy
));
756 assert(!is_removed(dummy
));
757 assert(!is_dummy(node
));
758 assert(!is_removed(node
));
761 /* We can always skip the dummy node initially */
762 iter
= rcu_dereference(iter_prev
->p
.next
);
763 assert(!is_removed(iter
));
764 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
766 * We should never be called with dummy (start of chain)
767 * and logically removed node (end of path compression
768 * marker) being the actual same node. This would be a
769 * bug in the algorithm implementation.
771 assert(dummy
!= node
);
773 if (unlikely(is_end(iter
)))
775 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
777 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
778 if (likely(is_removed(next
)))
780 iter_prev
= clear_flag(iter
);
783 assert(!is_removed(iter
));
785 new_next
= flag_dummy(clear_flag(next
));
787 new_next
= clear_flag(next
);
788 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
794 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
795 struct cds_lfht_node
*old_node
,
796 struct cds_lfht_node
*old_next
,
797 struct cds_lfht_node
*new_node
)
799 struct cds_lfht_node
*dummy
, *ret_next
;
800 struct _cds_lfht_node
*lookup
;
802 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
805 assert(!is_removed(old_node
));
806 assert(!is_dummy(old_node
));
807 assert(!is_removed(new_node
));
808 assert(!is_dummy(new_node
));
809 assert(new_node
!= old_node
);
811 /* Insert after node to be replaced */
812 if (is_removed(old_next
)) {
814 * Too late, the old node has been removed under us
815 * between lookup and replace. Fail.
819 assert(!is_dummy(old_next
));
820 assert(new_node
!= clear_flag(old_next
));
821 new_node
->p
.next
= clear_flag(old_next
);
823 * Here is the whole trick for lock-free replace: we add
824 * the replacement node _after_ the node we want to
825 * replace by atomically setting its next pointer at the
826 * same time we set its removal flag. Given that
827 * the lookups/get next use an iterator aware of the
828 * next pointer, they will either skip the old node due
829 * to the removal flag and see the new node, or use
830 * the old node, but will not see the new one.
832 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
833 old_next
, flag_removed(new_node
));
834 if (ret_next
== old_next
)
835 break; /* We performed the replacement. */
840 * Ensure that the old node is not visible to readers anymore:
841 * lookup for the node, and remove it (along with any other
842 * logically removed node) if found.
844 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
845 dummy
= (struct cds_lfht_node
*) lookup
;
846 _cds_lfht_gc_bucket(dummy
, new_node
);
848 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
853 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
854 * mode. A NULL unique_ret allows creation of duplicate keys.
857 void _cds_lfht_add(struct cds_lfht
*ht
,
859 struct cds_lfht_node
*node
,
860 struct cds_lfht_iter
*unique_ret
,
863 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
865 struct _cds_lfht_node
*lookup
;
867 assert(!is_dummy(node
));
868 assert(!is_removed(node
));
869 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
871 uint32_t chain_len
= 0;
874 * iter_prev points to the non-removed node prior to the
877 iter_prev
= (struct cds_lfht_node
*) lookup
;
878 /* We can always skip the dummy node initially */
879 iter
= rcu_dereference(iter_prev
->p
.next
);
880 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
882 if (unlikely(is_end(iter
)))
884 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
887 /* dummy node is the first node of the identical-hash-value chain */
888 if (dummy
&& clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
)
891 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
892 if (unlikely(is_removed(next
)))
898 && clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
) {
899 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
902 * uniquely adding inserts the node as the first
903 * node of the identical-hash-value node chain.
905 * This semantic ensures no duplicated keys
906 * should ever be observable in the table
907 * (including observe one node by one node
908 * by forward iterations)
910 cds_lfht_next_duplicate(ht
, &d_iter
);
914 *unique_ret
= d_iter
;
918 /* Only account for identical reverse hash once */
919 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
921 check_resize(ht
, size
, ++chain_len
);
922 iter_prev
= clear_flag(iter
);
927 assert(node
!= clear_flag(iter
));
928 assert(!is_removed(iter_prev
));
929 assert(!is_removed(iter
));
930 assert(iter_prev
!= node
);
932 node
->p
.next
= clear_flag(iter
);
934 node
->p
.next
= flag_dummy(clear_flag(iter
));
936 new_node
= flag_dummy(node
);
939 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
941 continue; /* retry */
948 assert(!is_removed(iter
));
950 new_next
= flag_dummy(clear_flag(next
));
952 new_next
= clear_flag(next
);
953 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
958 unique_ret
->node
= return_node
;
959 /* unique_ret->next left unset, never used. */
964 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
965 struct cds_lfht_node
*node
,
968 struct cds_lfht_node
*dummy
, *next
, *old
;
969 struct _cds_lfht_node
*lookup
;
971 if (!node
) /* Return -ENOENT if asked to delete NULL node */
974 /* logically delete the node */
975 assert(!is_dummy(node
));
976 assert(!is_removed(node
));
977 old
= rcu_dereference(node
->p
.next
);
979 struct cds_lfht_node
*new_next
;
982 if (unlikely(is_removed(next
)))
985 assert(is_dummy(next
));
987 assert(!is_dummy(next
));
988 new_next
= flag_removed(next
);
989 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
990 } while (old
!= next
);
991 /* We performed the (logical) deletion. */
994 * Ensure that the node is not visible to readers anymore: lookup for
995 * the node, and remove it (along with any other logically removed node)
998 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
999 dummy
= (struct cds_lfht_node
*) lookup
;
1000 _cds_lfht_gc_bucket(dummy
, node
);
1002 assert(is_removed(rcu_dereference(node
->p
.next
)));
1007 void *partition_resize_thread(void *arg
)
1009 struct partition_resize_work
*work
= arg
;
1011 work
->ht
->cds_lfht_rcu_register_thread();
1012 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1013 work
->ht
->cds_lfht_rcu_unregister_thread();
1018 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1020 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1021 unsigned long start
, unsigned long len
))
1023 unsigned long partition_len
;
1024 struct partition_resize_work
*work
;
1026 unsigned long nr_threads
;
1029 * Note: nr_cpus_mask + 1 is always power of 2.
1030 * We spawn just the number of threads we need to satisfy the minimum
1031 * partition size, up to the number of CPUs in the system.
1033 if (nr_cpus_mask
> 0) {
1034 nr_threads
= min(nr_cpus_mask
+ 1,
1035 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1039 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1040 work
= calloc(nr_threads
, sizeof(*work
));
1042 for (thread
= 0; thread
< nr_threads
; thread
++) {
1043 work
[thread
].ht
= ht
;
1045 work
[thread
].len
= partition_len
;
1046 work
[thread
].start
= thread
* partition_len
;
1047 work
[thread
].fct
= fct
;
1048 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1049 partition_resize_thread
, &work
[thread
]);
1052 for (thread
= 0; thread
< nr_threads
; thread
++) {
1053 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1060 * Holding RCU read lock to protect _cds_lfht_add against memory
1061 * reclaim that could be performed by other call_rcu worker threads (ABA
1064 * When we reach a certain length, we can split this population phase over
1065 * many worker threads, based on the number of CPUs available in the system.
1066 * This should therefore take care of not having the expand lagging behind too
1067 * many concurrent insertion threads by using the scheduler's ability to
1068 * schedule dummy node population fairly with insertions.
1071 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1072 unsigned long start
, unsigned long len
)
1076 assert(i
> ht
->min_alloc_order
);
1077 ht
->cds_lfht_rcu_read_lock();
1078 for (j
= start
; j
< start
+ len
; j
++) {
1079 struct cds_lfht_node
*new_node
=
1080 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1082 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1083 i
, j
, (1UL << (i
- 1)) + j
);
1084 new_node
->p
.reverse_hash
=
1085 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1086 _cds_lfht_add(ht
, 1UL << (i
- 1),
1089 ht
->cds_lfht_rcu_read_unlock();
1093 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1096 assert(nr_cpus_mask
!= -1);
1097 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1098 ht
->cds_lfht_rcu_thread_online();
1099 init_table_populate_partition(ht
, i
, 0, len
);
1100 ht
->cds_lfht_rcu_thread_offline();
1103 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1107 void init_table(struct cds_lfht
*ht
,
1108 unsigned long first_order
, unsigned long last_order
)
1112 dbg_printf("init table: first_order %lu last_order %lu\n",
1113 first_order
, last_order
);
1114 assert(first_order
> ht
->min_alloc_order
);
1115 for (i
= first_order
; i
<= last_order
; i
++) {
1118 len
= 1UL << (i
- 1);
1119 dbg_printf("init order %lu len: %lu\n", i
, len
);
1121 /* Stop expand if the resize target changes under us */
1122 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (1UL << i
))
1125 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1126 assert(ht
->t
.tbl
[i
]);
1129 * Set all dummy nodes reverse hash values for a level and
1130 * link all dummy nodes into the table.
1132 init_table_populate(ht
, i
, len
);
1135 * Update table size.
1137 cmm_smp_wmb(); /* populate data before RCU size */
1138 CMM_STORE_SHARED(ht
->t
.size
, 1UL << i
);
1140 dbg_printf("init new size: %lu\n", 1UL << i
);
1141 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1147 * Holding RCU read lock to protect _cds_lfht_remove against memory
1148 * reclaim that could be performed by other call_rcu worker threads (ABA
1150 * For a single level, we logically remove and garbage collect each node.
1152 * As a design choice, we perform logical removal and garbage collection on a
1153 * node-per-node basis to simplify this algorithm. We also assume keeping good
1154 * cache locality of the operation would overweight possible performance gain
1155 * that could be achieved by batching garbage collection for multiple levels.
1156 * However, this would have to be justified by benchmarks.
1158 * Concurrent removal and add operations are helping us perform garbage
1159 * collection of logically removed nodes. We guarantee that all logically
1160 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1161 * invoked to free a hole level of dummy nodes (after a grace period).
1163 * Logical removal and garbage collection can therefore be done in batch or on a
1164 * node-per-node basis, as long as the guarantee above holds.
1166 * When we reach a certain length, we can split this removal over many worker
1167 * threads, based on the number of CPUs available in the system. This should
1168 * take care of not letting resize process lag behind too many concurrent
1169 * updater threads actively inserting into the hash table.
1172 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1173 unsigned long start
, unsigned long len
)
1177 assert(i
> ht
->min_alloc_order
);
1178 ht
->cds_lfht_rcu_read_lock();
1179 for (j
= start
; j
< start
+ len
; j
++) {
1180 struct cds_lfht_node
*fini_node
=
1181 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1183 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1184 i
, j
, (1UL << (i
- 1)) + j
);
1185 fini_node
->p
.reverse_hash
=
1186 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1187 (void) _cds_lfht_del(ht
, 1UL << (i
- 1), fini_node
, 1);
1189 ht
->cds_lfht_rcu_read_unlock();
1193 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1196 assert(nr_cpus_mask
!= -1);
1197 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1198 ht
->cds_lfht_rcu_thread_online();
1199 remove_table_partition(ht
, i
, 0, len
);
1200 ht
->cds_lfht_rcu_thread_offline();
1203 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1207 void fini_table(struct cds_lfht
*ht
,
1208 unsigned long first_order
, unsigned long last_order
)
1211 void *free_by_rcu
= NULL
;
1213 dbg_printf("fini table: first_order %lu last_order %lu\n",
1214 first_order
, last_order
);
1215 assert(first_order
> ht
->min_alloc_order
);
1216 for (i
= last_order
; i
>= first_order
; i
--) {
1219 len
= 1UL << (i
- 1);
1220 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1222 /* Stop shrink if the resize target changes under us */
1223 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1226 cmm_smp_wmb(); /* populate data before RCU size */
1227 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1230 * We need to wait for all add operations to reach Q.S. (and
1231 * thus use the new table for lookups) before we can start
1232 * releasing the old dummy nodes. Otherwise their lookup will
1233 * return a logically removed node as insert position.
1235 ht
->cds_lfht_synchronize_rcu();
1240 * Set "removed" flag in dummy nodes about to be removed.
1241 * Unlink all now-logically-removed dummy node pointers.
1242 * Concurrent add/remove operation are helping us doing
1245 remove_table(ht
, i
, len
);
1247 free_by_rcu
= ht
->t
.tbl
[i
];
1249 dbg_printf("fini new size: %lu\n", 1UL << i
);
1250 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1255 ht
->cds_lfht_synchronize_rcu();
1261 void cds_lfht_create_dummy(struct cds_lfht
*ht
, unsigned long size
)
1263 struct _cds_lfht_node
*prev
, *node
;
1264 unsigned long order
, len
, i
, j
;
1266 ht
->t
.tbl
[0] = calloc(1, ht
->min_alloc_size
* sizeof(struct _cds_lfht_node
));
1267 assert(ht
->t
.tbl
[0]);
1269 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1270 ht
->t
.tbl
[0]->nodes
[0].next
= flag_dummy(get_end());
1271 ht
->t
.tbl
[0]->nodes
[0].reverse_hash
= 0;
1273 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1274 len
= 1UL << (order
- 1);
1275 if (order
<= ht
->min_alloc_order
) {
1276 ht
->t
.tbl
[order
] = (struct rcu_level
*) (ht
->t
.tbl
[0]->nodes
+ len
);
1278 ht
->t
.tbl
[order
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1279 assert(ht
->t
.tbl
[order
]);
1283 prev
= ht
->t
.tbl
[i
]->nodes
;
1284 for (j
= 0; j
< len
; j
++) {
1285 if (j
& (j
- 1)) { /* Between power of 2 */
1287 } else if (j
) { /* At each power of 2 */
1289 prev
= ht
->t
.tbl
[i
]->nodes
;
1292 node
= &ht
->t
.tbl
[order
]->nodes
[j
];
1293 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1295 node
->next
= prev
->next
;
1296 assert(is_dummy(node
->next
));
1297 node
->reverse_hash
= bit_reverse_ulong(j
+ len
);
1298 prev
->next
= flag_dummy((struct cds_lfht_node
*)node
);
1303 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1304 cds_lfht_compare_fct compare_fct
,
1305 unsigned long hash_seed
,
1306 unsigned long init_size
,
1307 unsigned long min_alloc_size
,
1309 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1310 void (*func
)(struct rcu_head
*head
)),
1311 void (*cds_lfht_synchronize_rcu
)(void),
1312 void (*cds_lfht_rcu_read_lock
)(void),
1313 void (*cds_lfht_rcu_read_unlock
)(void),
1314 void (*cds_lfht_rcu_thread_offline
)(void),
1315 void (*cds_lfht_rcu_thread_online
)(void),
1316 void (*cds_lfht_rcu_register_thread
)(void),
1317 void (*cds_lfht_rcu_unregister_thread
)(void),
1318 pthread_attr_t
*attr
)
1320 struct cds_lfht
*ht
;
1321 unsigned long order
;
1323 /* min_alloc_size must be power of two */
1324 if (!min_alloc_size
|| (min_alloc_size
& (min_alloc_size
- 1)))
1326 /* init_size must be power of two */
1327 if (!init_size
|| (init_size
& (init_size
- 1)))
1329 min_alloc_size
= max(min_alloc_size
, MIN_TABLE_SIZE
);
1330 init_size
= max(init_size
, min_alloc_size
);
1331 ht
= calloc(1, sizeof(struct cds_lfht
));
1333 ht
->hash_fct
= hash_fct
;
1334 ht
->compare_fct
= compare_fct
;
1335 ht
->hash_seed
= hash_seed
;
1336 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1337 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1338 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1339 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1340 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1341 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1342 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1343 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1344 ht
->resize_attr
= attr
;
1345 alloc_split_items_count(ht
);
1346 /* this mutex should not nest in read-side C.S. */
1347 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1349 order
= get_count_order_ulong(init_size
);
1350 ht
->t
.resize_target
= 1UL << order
;
1351 cds_lfht_create_dummy(ht
, 1UL << order
);
1352 ht
->t
.size
= 1UL << order
;
1353 ht
->min_alloc_size
= min_alloc_size
;
1354 ht
->min_alloc_order
= get_count_order_ulong(min_alloc_size
);
1358 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1359 struct cds_lfht_iter
*iter
)
1361 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1362 struct _cds_lfht_node
*lookup
;
1363 unsigned long hash
, reverse_hash
, size
;
1365 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1366 reverse_hash
= bit_reverse_ulong(hash
);
1368 size
= rcu_dereference(ht
->t
.size
);
1369 lookup
= lookup_bucket(ht
, size
, hash
);
1370 dummy_node
= (struct cds_lfht_node
*) lookup
;
1371 /* We can always skip the dummy node initially */
1372 node
= rcu_dereference(dummy_node
->p
.next
);
1373 node
= clear_flag(node
);
1375 if (unlikely(is_end(node
))) {
1379 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1383 next
= rcu_dereference(node
->p
.next
);
1384 if (likely(!is_removed(next
))
1386 && clear_flag(node
)->p
.reverse_hash
== reverse_hash
1387 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1390 node
= clear_flag(next
);
1392 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1397 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1399 struct cds_lfht_node
*node
, *next
;
1400 unsigned long reverse_hash
;
1405 reverse_hash
= node
->p
.reverse_hash
;
1407 key_len
= node
->key_len
;
1409 node
= clear_flag(next
);
1412 if (unlikely(is_end(node
))) {
1416 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1420 next
= rcu_dereference(node
->p
.next
);
1421 if (likely(!is_removed(next
))
1423 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1426 node
= clear_flag(next
);
1428 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1433 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1435 struct cds_lfht_node
*node
, *next
;
1437 node
= clear_flag(iter
->next
);
1439 if (unlikely(is_end(node
))) {
1443 next
= rcu_dereference(node
->p
.next
);
1444 if (likely(!is_removed(next
))
1445 && !is_dummy(next
)) {
1448 node
= clear_flag(next
);
1450 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1455 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1457 struct _cds_lfht_node
*lookup
;
1460 * Get next after first dummy node. The first dummy node is the
1461 * first node of the linked list.
1463 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1464 iter
->next
= lookup
->next
;
1465 cds_lfht_next(ht
, iter
);
1468 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1470 unsigned long hash
, size
;
1472 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1473 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1475 size
= rcu_dereference(ht
->t
.size
);
1476 _cds_lfht_add(ht
, size
, node
, NULL
, 0);
1477 ht_count_add(ht
, size
, hash
);
1480 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1481 struct cds_lfht_node
*node
)
1483 unsigned long hash
, size
;
1484 struct cds_lfht_iter iter
;
1486 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1487 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1489 size
= rcu_dereference(ht
->t
.size
);
1490 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1491 if (iter
.node
== node
)
1492 ht_count_add(ht
, size
, hash
);
1496 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1497 struct cds_lfht_node
*node
)
1499 unsigned long hash
, size
;
1500 struct cds_lfht_iter iter
;
1502 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1503 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1505 size
= rcu_dereference(ht
->t
.size
);
1507 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1508 if (iter
.node
== node
) {
1509 ht_count_add(ht
, size
, hash
);
1513 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1518 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1519 struct cds_lfht_node
*new_node
)
1523 size
= rcu_dereference(ht
->t
.size
);
1524 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1528 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1530 unsigned long size
, hash
;
1533 size
= rcu_dereference(ht
->t
.size
);
1534 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1536 hash
= bit_reverse_ulong(iter
->node
->p
.reverse_hash
);
1537 ht_count_del(ht
, size
, hash
);
1543 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1545 struct cds_lfht_node
*node
;
1546 struct _cds_lfht_node
*lookup
;
1547 unsigned long order
, i
, size
;
1549 /* Check that the table is empty */
1550 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1551 node
= (struct cds_lfht_node
*) lookup
;
1553 node
= clear_flag(node
)->p
.next
;
1554 if (!is_dummy(node
))
1556 assert(!is_removed(node
));
1557 } while (!is_end(node
));
1559 * size accessed without rcu_dereference because hash table is
1563 /* Internal sanity check: all nodes left should be dummy */
1564 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1567 len
= !order
? 1 : 1UL << (order
- 1);
1568 for (i
= 0; i
< len
; i
++) {
1569 dbg_printf("delete order %lu i %lu hash %lu\n",
1571 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1572 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1575 if (order
== ht
->min_alloc_order
)
1576 poison_free(ht
->t
.tbl
[0]);
1577 else if (order
> ht
->min_alloc_order
)
1578 poison_free(ht
->t
.tbl
[order
]);
1579 /* Nothing to delete for order < ht->min_alloc_order */
1585 * Should only be called when no more concurrent readers nor writers can
1586 * possibly access the table.
1588 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1592 /* Wait for in-flight resize operations to complete */
1593 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1594 cmm_smp_mb(); /* Store destroy before load resize */
1595 while (uatomic_read(&ht
->in_progress_resize
))
1596 poll(NULL
, 0, 100); /* wait for 100ms */
1597 ret
= cds_lfht_delete_dummy(ht
);
1600 free_split_items_count(ht
);
1602 *attr
= ht
->resize_attr
;
1607 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1608 long *approx_before
,
1609 unsigned long *count
,
1610 unsigned long *removed
,
1613 struct cds_lfht_node
*node
, *next
;
1614 struct _cds_lfht_node
*lookup
;
1615 unsigned long nr_dummy
= 0;
1618 if (ht
->split_count
) {
1621 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1622 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1623 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1630 /* Count non-dummy nodes in the table */
1631 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1632 node
= (struct cds_lfht_node
*) lookup
;
1634 next
= rcu_dereference(node
->p
.next
);
1635 if (is_removed(next
)) {
1636 if (!is_dummy(next
))
1640 } else if (!is_dummy(next
))
1644 node
= clear_flag(next
);
1645 } while (!is_end(node
));
1646 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1648 if (ht
->split_count
) {
1651 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1652 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1653 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1658 /* called with resize mutex held */
1660 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1661 unsigned long old_size
, unsigned long new_size
)
1663 unsigned long old_order
, new_order
;
1665 old_order
= get_count_order_ulong(old_size
);
1666 new_order
= get_count_order_ulong(new_size
);
1667 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1668 old_size
, old_order
, new_size
, new_order
);
1669 assert(new_size
> old_size
);
1670 init_table(ht
, old_order
+ 1, new_order
);
1673 /* called with resize mutex held */
1675 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1676 unsigned long old_size
, unsigned long new_size
)
1678 unsigned long old_order
, new_order
;
1680 new_size
= max(new_size
, ht
->min_alloc_size
);
1681 old_order
= get_count_order_ulong(old_size
);
1682 new_order
= get_count_order_ulong(new_size
);
1683 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1684 old_size
, old_order
, new_size
, new_order
);
1685 assert(new_size
< old_size
);
1687 /* Remove and unlink all dummy nodes to remove. */
1688 fini_table(ht
, new_order
+ 1, old_order
);
1692 /* called with resize mutex held */
1694 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1696 unsigned long new_size
, old_size
;
1699 * Resize table, re-do if the target size has changed under us.
1702 assert(uatomic_read(&ht
->in_progress_resize
));
1703 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1705 ht
->t
.resize_initiated
= 1;
1706 old_size
= ht
->t
.size
;
1707 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1708 if (old_size
< new_size
)
1709 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1710 else if (old_size
> new_size
)
1711 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1712 ht
->t
.resize_initiated
= 0;
1713 /* write resize_initiated before read resize_target */
1715 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1719 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1722 return _uatomic_max(&ht
->t
.resize_target
,
1723 size
<< growth_order
);
1727 void resize_target_update_count(struct cds_lfht
*ht
,
1728 unsigned long count
)
1730 count
= max(count
, ht
->min_alloc_size
);
1731 uatomic_set(&ht
->t
.resize_target
, count
);
1734 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1736 resize_target_update_count(ht
, new_size
);
1737 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1738 ht
->cds_lfht_rcu_thread_offline();
1739 pthread_mutex_lock(&ht
->resize_mutex
);
1740 _do_cds_lfht_resize(ht
);
1741 pthread_mutex_unlock(&ht
->resize_mutex
);
1742 ht
->cds_lfht_rcu_thread_online();
1746 void do_resize_cb(struct rcu_head
*head
)
1748 struct rcu_resize_work
*work
=
1749 caa_container_of(head
, struct rcu_resize_work
, head
);
1750 struct cds_lfht
*ht
= work
->ht
;
1752 ht
->cds_lfht_rcu_thread_offline();
1753 pthread_mutex_lock(&ht
->resize_mutex
);
1754 _do_cds_lfht_resize(ht
);
1755 pthread_mutex_unlock(&ht
->resize_mutex
);
1756 ht
->cds_lfht_rcu_thread_online();
1758 cmm_smp_mb(); /* finish resize before decrement */
1759 uatomic_dec(&ht
->in_progress_resize
);
1763 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1765 struct rcu_resize_work
*work
;
1766 unsigned long target_size
;
1768 target_size
= resize_target_update(ht
, size
, growth
);
1769 /* Store resize_target before read resize_initiated */
1771 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1772 uatomic_inc(&ht
->in_progress_resize
);
1773 cmm_smp_mb(); /* increment resize count before load destroy */
1774 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1775 uatomic_dec(&ht
->in_progress_resize
);
1778 work
= malloc(sizeof(*work
));
1780 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1781 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1786 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1787 unsigned long count
)
1789 struct rcu_resize_work
*work
;
1791 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1793 resize_target_update_count(ht
, count
);
1794 /* Store resize_target before read resize_initiated */
1796 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1797 uatomic_inc(&ht
->in_progress_resize
);
1798 cmm_smp_mb(); /* increment resize count before load destroy */
1799 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1800 uatomic_dec(&ht
->in_progress_resize
);
1803 work
= malloc(sizeof(*work
));
1805 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1806 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);