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 * - Per-CPU 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
95 * - call_rcu is used to garbage-collect the old order 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.
100 * A bit of ascii art explanation:
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
105 * This shows the nodes for a small table ordered by reversed bits:
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
123 * 1 | 1 001 100 <- <-
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
128 * 3 -> | | | 4 100 001 | |
144 #include <urcu-call-rcu.h>
145 #include <urcu/arch.h>
146 #include <urcu/uatomic.h>
147 #include <urcu/compiler.h>
148 #include <urcu/rculfhash.h>
153 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
155 #define dbg_printf(fmt, args...)
159 * Per-CPU split-counters lazily update the global counter each 1024
160 * addition/removal. It automatically keeps track of resize required.
161 * We use the bucket length as indicator for need to expand for small
162 * tables and machines lacking per-cpu data suppport.
164 #define COUNT_COMMIT_ORDER 10
165 #define CHAIN_LEN_TARGET 1
166 #define CHAIN_LEN_RESIZE_THRESHOLD 3
169 * Define the minimum table size.
171 #define MIN_TABLE_SIZE 1
173 #if (CAA_BITS_PER_LONG == 32)
174 #define MAX_TABLE_ORDER 32
176 #define MAX_TABLE_ORDER 64
180 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 #define MIN_PARTITION_PER_THREAD_ORDER 12
183 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
186 #define min(a, b) ((a) < (b) ? (a) : (b))
190 #define max(a, b) ((a) > (b) ? (a) : (b))
194 * The removed flag needs to be updated atomically with the pointer.
195 * It indicates that no node must attach to the node scheduled for
196 * removal, and that node garbage collection must be performed.
197 * The dummy flag does not require to be updated atomically with the
198 * pointer, but it is added as a pointer low bit flag to save space.
200 #define REMOVED_FLAG (1UL << 0)
201 #define DUMMY_FLAG (1UL << 1)
202 #define FLAGS_MASK ((1UL << 2) - 1)
204 /* Value of the end pointer. Should not interact with flags. */
205 #define END_VALUE NULL
207 struct ht_items_count
{
208 unsigned long add
, del
;
209 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
212 struct rcu_head head
;
213 struct _cds_lfht_node nodes
[0];
217 unsigned long size
; /* always a power of 2, shared (RCU) */
218 unsigned long resize_target
;
219 int resize_initiated
;
220 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
225 cds_lfht_hash_fct hash_fct
;
226 cds_lfht_compare_fct compare_fct
;
227 unsigned long hash_seed
;
230 * We need to put the work threads offline (QSBR) when taking this
231 * mutex, because we use synchronize_rcu within this mutex critical
232 * section, which waits on read-side critical sections, and could
233 * therefore cause grace-period deadlock if we hold off RCU G.P.
236 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
237 unsigned int in_progress_resize
, in_progress_destroy
;
238 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
239 void (*func
)(struct rcu_head
*head
));
240 void (*cds_lfht_synchronize_rcu
)(void);
241 void (*cds_lfht_rcu_read_lock
)(void);
242 void (*cds_lfht_rcu_read_unlock
)(void);
243 void (*cds_lfht_rcu_thread_offline
)(void);
244 void (*cds_lfht_rcu_thread_online
)(void);
245 void (*cds_lfht_rcu_register_thread
)(void);
246 void (*cds_lfht_rcu_unregister_thread
)(void);
247 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
248 long count
; /* global approximate item count */
249 struct ht_items_count
*percpu_count
; /* per-cpu item count */
252 struct rcu_resize_work
{
253 struct rcu_head head
;
257 struct partition_resize_work
{
259 unsigned long i
, start
, len
;
260 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
261 unsigned long start
, unsigned long len
);
271 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
273 struct cds_lfht_node
*node
,
274 enum add_mode mode
, int dummy
);
277 * Algorithm to reverse bits in a word by lookup table, extended to
280 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
281 * Originally from Public Domain.
284 static const uint8_t BitReverseTable256
[256] =
286 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
287 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
288 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
289 R6(0), R6(2), R6(1), R6(3)
296 uint8_t bit_reverse_u8(uint8_t v
)
298 return BitReverseTable256
[v
];
301 static __attribute__((unused
))
302 uint32_t bit_reverse_u32(uint32_t v
)
304 return ((uint32_t) bit_reverse_u8(v
) << 24) |
305 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
306 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
307 ((uint32_t) bit_reverse_u8(v
>> 24));
310 static __attribute__((unused
))
311 uint64_t bit_reverse_u64(uint64_t v
)
313 return ((uint64_t) bit_reverse_u8(v
) << 56) |
314 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
315 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
316 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
317 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
318 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
319 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
320 ((uint64_t) bit_reverse_u8(v
>> 56));
324 unsigned long bit_reverse_ulong(unsigned long v
)
326 #if (CAA_BITS_PER_LONG == 32)
327 return bit_reverse_u32(v
);
329 return bit_reverse_u64(v
);
334 * fls: returns the position of the most significant bit.
335 * Returns 0 if no bit is set, else returns the position of the most
336 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
338 #if defined(__i386) || defined(__x86_64)
340 unsigned int fls_u32(uint32_t x
)
348 : "=r" (r
) : "rm" (x
));
354 #if defined(__x86_64)
356 unsigned int fls_u64(uint64_t x
)
364 : "=r" (r
) : "rm" (x
));
371 static __attribute__((unused
))
372 unsigned int fls_u64(uint64_t x
)
379 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
383 if (!(x
& 0xFFFF000000000000ULL
)) {
387 if (!(x
& 0xFF00000000000000ULL
)) {
391 if (!(x
& 0xF000000000000000ULL
)) {
395 if (!(x
& 0xC000000000000000ULL
)) {
399 if (!(x
& 0x8000000000000000ULL
)) {
408 static __attribute__((unused
))
409 unsigned int fls_u32(uint32_t x
)
415 if (!(x
& 0xFFFF0000U
)) {
419 if (!(x
& 0xFF000000U
)) {
423 if (!(x
& 0xF0000000U
)) {
427 if (!(x
& 0xC0000000U
)) {
431 if (!(x
& 0x80000000U
)) {
439 unsigned int fls_ulong(unsigned long x
)
441 #if (CAA_BITS_PER_lONG == 32)
448 int get_count_order_u32(uint32_t x
)
452 order
= fls_u32(x
) - 1;
458 int get_count_order_ulong(unsigned long x
)
462 order
= fls_ulong(x
) - 1;
469 #define poison_free(ptr) \
471 memset(ptr, 0x42, sizeof(*(ptr))); \
475 #define poison_free(ptr) free(ptr)
479 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
482 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
483 * available, then we support hash table item accounting.
484 * In the unfortunate event the number of CPUs reported would be
485 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
487 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
490 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
491 unsigned long count
);
493 static long nr_cpus_mask
= -1;
496 struct ht_items_count
*alloc_per_cpu_items_count(void)
498 struct ht_items_count
*count
;
500 switch (nr_cpus_mask
) {
507 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
513 * round up number of CPUs to next power of two, so we
514 * can use & for modulo.
516 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
517 nr_cpus_mask
= maxcpus
- 1;
521 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
526 void free_per_cpu_items_count(struct ht_items_count
*count
)
536 assert(nr_cpus_mask
>= 0);
537 cpu
= sched_getcpu();
538 if (unlikely(cpu
< 0))
541 return cpu
& nr_cpus_mask
;
545 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
547 unsigned long percpu_count
;
550 if (unlikely(!ht
->percpu_count
))
553 if (unlikely(cpu
< 0))
555 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
556 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
559 dbg_printf("add percpu %lu\n", percpu_count
);
560 count
= uatomic_add_return(&ht
->count
,
561 1UL << COUNT_COMMIT_ORDER
);
563 if (!(count
& (count
- 1))) {
564 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
566 dbg_printf("add set global %ld\n", count
);
567 cds_lfht_resize_lazy_count(ht
, size
,
568 count
>> (CHAIN_LEN_TARGET
- 1));
574 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
576 unsigned long percpu_count
;
579 if (unlikely(!ht
->percpu_count
))
582 if (unlikely(cpu
< 0))
584 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
585 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
588 dbg_printf("del percpu %lu\n", percpu_count
);
589 count
= uatomic_add_return(&ht
->count
,
590 -(1UL << COUNT_COMMIT_ORDER
));
592 if (!(count
& (count
- 1))) {
593 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
595 dbg_printf("del set global %ld\n", count
);
597 * Don't shrink table if the number of nodes is below a
600 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (nr_cpus_mask
+ 1))
602 cds_lfht_resize_lazy_count(ht
, size
,
603 count
>> (CHAIN_LEN_TARGET
- 1));
608 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
610 static const long nr_cpus_mask
= -2;
613 struct ht_items_count
*alloc_per_cpu_items_count(void)
619 void free_per_cpu_items_count(struct ht_items_count
*count
)
624 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
629 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
633 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
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 void cds_lfht_free_level(struct rcu_head
*head
)
717 struct rcu_level
*l
=
718 caa_container_of(head
, struct rcu_level
, head
);
723 * Remove all logically deleted nodes from a bucket up to a certain node key.
726 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
728 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
730 assert(!is_dummy(dummy
));
731 assert(!is_removed(dummy
));
732 assert(!is_dummy(node
));
733 assert(!is_removed(node
));
736 /* We can always skip the dummy node initially */
737 iter
= rcu_dereference(iter_prev
->p
.next
);
738 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
740 * We should never be called with dummy (start of chain)
741 * and logically removed node (end of path compression
742 * marker) being the actual same node. This would be a
743 * bug in the algorithm implementation.
745 assert(dummy
!= node
);
747 if (unlikely(is_end(iter
)))
749 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
751 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
752 if (likely(is_removed(next
)))
754 iter_prev
= clear_flag(iter
);
757 assert(!is_removed(iter
));
759 new_next
= flag_dummy(clear_flag(next
));
761 new_next
= clear_flag(next
);
762 if (is_removed(iter
))
763 new_next
= flag_removed(new_next
);
764 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
770 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
771 struct cds_lfht_node
*old_node
,
772 struct cds_lfht_node
*ret_next
,
773 struct cds_lfht_node
*new_node
)
775 struct cds_lfht_node
*dummy
, *old_next
;
776 struct _cds_lfht_node
*lookup
;
778 unsigned long hash
, index
, order
;
780 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
783 assert(!is_removed(old_node
));
784 assert(!is_dummy(old_node
));
785 assert(!is_removed(new_node
));
786 assert(!is_dummy(new_node
));
787 assert(new_node
!= old_node
);
789 /* Insert after node to be replaced */
791 if (is_removed(old_next
)) {
793 * Too late, the old node has been removed under us
794 * between lookup and replace. Fail.
798 assert(!is_dummy(old_next
));
799 assert(new_node
!= clear_flag(old_next
));
800 new_node
->p
.next
= clear_flag(old_next
);
802 * Here is the whole trick for lock-free replace: we add
803 * the replacement node _after_ the node we want to
804 * replace by atomically setting its next pointer at the
805 * same time we set its removal flag. Given that
806 * the lookups/get next use an iterator aware of the
807 * next pointer, they will either skip the old node due
808 * to the removal flag and see the new node, or use
809 * the old node, but will not see the new one.
811 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
812 old_next
, flag_removed(new_node
));
813 } while (ret_next
!= old_next
);
815 /* We performed the replacement. */
819 * Ensure that the old node is not visible to readers anymore:
820 * lookup for the node, and remove it (along with any other
821 * logically removed node) if found.
823 hash
= bit_reverse_ulong(old_node
->p
.reverse_hash
);
825 index
= hash
& (size
- 1);
826 order
= get_count_order_ulong(index
+ 1);
827 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
828 dummy
= (struct cds_lfht_node
*) lookup
;
829 _cds_lfht_gc_bucket(dummy
, new_node
);
832 * Only the flagging action indicated that we (and no other)
833 * replaced the node from the hash table.
836 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
844 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
846 struct cds_lfht_node
*node
,
847 enum add_mode mode
, int dummy
)
849 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
850 *dummy_node
, *return_node
;
851 struct _cds_lfht_node
*lookup
;
852 unsigned long hash
, index
, order
;
854 assert(!is_dummy(node
));
855 assert(!is_removed(node
));
858 node
->p
.next
= flag_dummy(get_end());
859 return node
; /* Initial first add (head) */
861 hash
= 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 index
= hash
& (size
- 1);
870 order
= get_count_order_ulong(index
+ 1);
871 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
872 iter_prev
= (struct cds_lfht_node
*) lookup
;
873 /* We can always skip the dummy node initially */
874 iter
= rcu_dereference(iter_prev
->p
.next
);
875 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
877 if (unlikely(is_end(iter
)))
879 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
881 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
882 if (unlikely(is_removed(next
)))
884 if ((mode
== ADD_UNIQUE
|| mode
== ADD_REPLACE
)
886 && !ht
->compare_fct(node
->key
, node
->key_len
,
887 clear_flag(iter
)->key
,
888 clear_flag(iter
)->key_len
)) {
889 if (mode
== ADD_UNIQUE
)
890 return clear_flag(iter
);
891 else /* mode == ADD_REPLACE */
894 /* Only account for identical reverse hash once */
895 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
897 check_resize(ht
, size
, ++chain_len
);
898 iter_prev
= clear_flag(iter
);
903 assert(node
!= clear_flag(iter
));
904 assert(!is_removed(iter_prev
));
905 assert(!is_removed(iter
));
906 assert(iter_prev
!= node
);
908 node
->p
.next
= clear_flag(iter
);
910 node
->p
.next
= flag_dummy(clear_flag(iter
));
912 new_node
= flag_dummy(node
);
915 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
917 continue; /* retry */
919 if (mode
== ADD_REPLACE
)
921 else /* ADD_DEFAULT and ADD_UNIQUE */
928 if (!_cds_lfht_replace(ht
, size
, clear_flag(iter
), next
,
930 return_node
= clear_flag(iter
);
931 goto end
; /* gc already done */
933 continue; /* retry */
937 assert(!is_removed(iter
));
939 new_next
= flag_dummy(clear_flag(next
));
941 new_next
= clear_flag(next
);
942 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
946 /* Garbage collect logically removed nodes in the bucket */
947 index
= hash
& (size
- 1);
948 order
= get_count_order_ulong(index
+ 1);
949 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
950 dummy_node
= (struct cds_lfht_node
*) lookup
;
951 _cds_lfht_gc_bucket(dummy_node
, node
);
957 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
958 struct cds_lfht_node
*node
,
961 struct cds_lfht_node
*dummy
, *next
, *old
;
962 struct _cds_lfht_node
*lookup
;
964 unsigned long hash
, index
, order
;
966 if (!node
) /* Return -ENOENT if asked to delete NULL node */
969 /* logically delete the node */
970 assert(!is_dummy(node
));
971 assert(!is_removed(node
));
972 old
= rcu_dereference(node
->p
.next
);
974 struct cds_lfht_node
*new_next
;
977 if (unlikely(is_removed(next
)))
980 assert(is_dummy(next
));
982 assert(!is_dummy(next
));
983 new_next
= flag_removed(next
);
984 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
985 } while (old
!= next
);
987 /* We performed the (logical) deletion. */
991 * Ensure that the node is not visible to readers anymore: lookup for
992 * the node, and remove it (along with any other logically removed node)
995 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
997 index
= hash
& (size
- 1);
998 order
= get_count_order_ulong(index
+ 1);
999 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
1000 dummy
= (struct cds_lfht_node
*) lookup
;
1001 _cds_lfht_gc_bucket(dummy
, node
);
1004 * Only the flagging action indicated that we (and no other)
1005 * removed the node from the hash.
1008 assert(is_removed(rcu_dereference(node
->p
.next
)));
1016 void *partition_resize_thread(void *arg
)
1018 struct partition_resize_work
*work
= arg
;
1020 work
->ht
->cds_lfht_rcu_register_thread();
1021 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1022 work
->ht
->cds_lfht_rcu_unregister_thread();
1027 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1029 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1030 unsigned long start
, unsigned long len
))
1032 unsigned long partition_len
;
1033 struct partition_resize_work
*work
;
1035 unsigned long nr_threads
;
1036 pthread_t
*thread_id
;
1039 * Note: nr_cpus_mask + 1 is always power of 2.
1040 * We spawn just the number of threads we need to satisfy the minimum
1041 * partition size, up to the number of CPUs in the system.
1043 if (nr_cpus_mask
> 0) {
1044 nr_threads
= min(nr_cpus_mask
+ 1,
1045 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1049 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1050 work
= calloc(nr_threads
, sizeof(*work
));
1051 thread_id
= calloc(nr_threads
, sizeof(*thread_id
));
1053 for (thread
= 0; thread
< nr_threads
; thread
++) {
1054 work
[thread
].ht
= ht
;
1056 work
[thread
].len
= partition_len
;
1057 work
[thread
].start
= thread
* partition_len
;
1058 work
[thread
].fct
= fct
;
1059 ret
= pthread_create(&thread_id
[thread
], ht
->resize_attr
,
1060 partition_resize_thread
, &work
[thread
]);
1063 for (thread
= 0; thread
< nr_threads
; thread
++) {
1064 ret
= pthread_join(thread_id
[thread
], NULL
);
1072 * Holding RCU read lock to protect _cds_lfht_add against memory
1073 * reclaim that could be performed by other call_rcu worker threads (ABA
1076 * When we reach a certain length, we can split this population phase over
1077 * many worker threads, based on the number of CPUs available in the system.
1078 * This should therefore take care of not having the expand lagging behind too
1079 * many concurrent insertion threads by using the scheduler's ability to
1080 * schedule dummy node population fairly with insertions.
1083 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1084 unsigned long start
, unsigned long len
)
1088 ht
->cds_lfht_rcu_read_lock();
1089 for (j
= start
; j
< start
+ len
; j
++) {
1090 struct cds_lfht_node
*new_node
=
1091 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1093 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1094 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1095 new_node
->p
.reverse_hash
=
1096 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1097 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1098 new_node
, ADD_DEFAULT
, 1);
1100 ht
->cds_lfht_rcu_read_unlock();
1104 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1107 assert(nr_cpus_mask
!= -1);
1108 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1109 ht
->cds_lfht_rcu_thread_online();
1110 init_table_populate_partition(ht
, i
, 0, len
);
1111 ht
->cds_lfht_rcu_thread_offline();
1114 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1118 void init_table(struct cds_lfht
*ht
,
1119 unsigned long first_order
, unsigned long len_order
)
1121 unsigned long i
, end_order
;
1123 dbg_printf("init table: first_order %lu end_order %lu\n",
1124 first_order
, first_order
+ len_order
);
1125 end_order
= first_order
+ len_order
;
1126 for (i
= first_order
; i
< end_order
; i
++) {
1129 len
= !i
? 1 : 1UL << (i
- 1);
1130 dbg_printf("init order %lu len: %lu\n", i
, len
);
1132 /* Stop expand if the resize target changes under us */
1133 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1136 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1137 + (len
* sizeof(struct _cds_lfht_node
)));
1138 assert(ht
->t
.tbl
[i
]);
1141 * Set all dummy nodes reverse hash values for a level and
1142 * link all dummy nodes into the table.
1144 init_table_populate(ht
, i
, len
);
1147 * Update table size.
1149 cmm_smp_wmb(); /* populate data before RCU size */
1150 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1152 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1153 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1159 * Holding RCU read lock to protect _cds_lfht_remove against memory
1160 * reclaim that could be performed by other call_rcu worker threads (ABA
1162 * For a single level, we logically remove and garbage collect each node.
1164 * As a design choice, we perform logical removal and garbage collection on a
1165 * node-per-node basis to simplify this algorithm. We also assume keeping good
1166 * cache locality of the operation would overweight possible performance gain
1167 * that could be achieved by batching garbage collection for multiple levels.
1168 * However, this would have to be justified by benchmarks.
1170 * Concurrent removal and add operations are helping us perform garbage
1171 * collection of logically removed nodes. We guarantee that all logically
1172 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1173 * invoked to free a hole level of dummy nodes (after a grace period).
1175 * Logical removal and garbage collection can therefore be done in batch or on a
1176 * node-per-node basis, as long as the guarantee above holds.
1178 * When we reach a certain length, we can split this removal over many worker
1179 * threads, based on the number of CPUs available in the system. This should
1180 * take care of not letting resize process lag behind too many concurrent
1181 * updater threads actively inserting into the hash table.
1184 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1185 unsigned long start
, unsigned long len
)
1189 ht
->cds_lfht_rcu_read_lock();
1190 for (j
= start
; j
< start
+ len
; j
++) {
1191 struct cds_lfht_node
*fini_node
=
1192 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1194 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1195 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1196 fini_node
->p
.reverse_hash
=
1197 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1198 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1201 ht
->cds_lfht_rcu_read_unlock();
1205 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1208 assert(nr_cpus_mask
!= -1);
1209 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1210 ht
->cds_lfht_rcu_thread_online();
1211 remove_table_partition(ht
, i
, 0, len
);
1212 ht
->cds_lfht_rcu_thread_offline();
1215 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1219 void fini_table(struct cds_lfht
*ht
,
1220 unsigned long first_order
, unsigned long len_order
)
1224 dbg_printf("fini table: first_order %lu end_order %lu\n",
1225 first_order
, first_order
+ len_order
);
1226 end_order
= first_order
+ len_order
;
1227 assert(first_order
> 0);
1228 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1231 len
= !i
? 1 : 1UL << (i
- 1);
1232 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1234 /* Stop shrink if the resize target changes under us */
1235 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1238 cmm_smp_wmb(); /* populate data before RCU size */
1239 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1242 * We need to wait for all add operations to reach Q.S. (and
1243 * thus use the new table for lookups) before we can start
1244 * releasing the old dummy nodes. Otherwise their lookup will
1245 * return a logically removed node as insert position.
1247 ht
->cds_lfht_synchronize_rcu();
1250 * Set "removed" flag in dummy nodes about to be removed.
1251 * Unlink all now-logically-removed dummy node pointers.
1252 * Concurrent add/remove operation are helping us doing
1255 remove_table(ht
, i
, len
);
1257 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1259 dbg_printf("fini new size: %lu\n", 1UL << i
);
1260 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1265 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1266 cds_lfht_compare_fct compare_fct
,
1267 unsigned long hash_seed
,
1268 unsigned long init_size
,
1270 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1271 void (*func
)(struct rcu_head
*head
)),
1272 void (*cds_lfht_synchronize_rcu
)(void),
1273 void (*cds_lfht_rcu_read_lock
)(void),
1274 void (*cds_lfht_rcu_read_unlock
)(void),
1275 void (*cds_lfht_rcu_thread_offline
)(void),
1276 void (*cds_lfht_rcu_thread_online
)(void),
1277 void (*cds_lfht_rcu_register_thread
)(void),
1278 void (*cds_lfht_rcu_unregister_thread
)(void),
1279 pthread_attr_t
*attr
)
1281 struct cds_lfht
*ht
;
1282 unsigned long order
;
1284 /* init_size must be power of two */
1285 if (init_size
&& (init_size
& (init_size
- 1)))
1287 ht
= calloc(1, sizeof(struct cds_lfht
));
1289 ht
->hash_fct
= hash_fct
;
1290 ht
->compare_fct
= compare_fct
;
1291 ht
->hash_seed
= hash_seed
;
1292 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1293 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1294 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1295 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1296 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1297 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1298 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1299 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1300 ht
->resize_attr
= attr
;
1301 ht
->percpu_count
= alloc_per_cpu_items_count();
1302 /* this mutex should not nest in read-side C.S. */
1303 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1304 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1306 ht
->cds_lfht_rcu_thread_offline();
1307 pthread_mutex_lock(&ht
->resize_mutex
);
1308 ht
->t
.resize_target
= 1UL << (order
- 1);
1309 init_table(ht
, 0, order
);
1310 pthread_mutex_unlock(&ht
->resize_mutex
);
1311 ht
->cds_lfht_rcu_thread_online();
1315 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1316 struct cds_lfht_iter
*iter
)
1318 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1319 struct _cds_lfht_node
*lookup
;
1320 unsigned long hash
, reverse_hash
, index
, order
, size
;
1322 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1323 reverse_hash
= bit_reverse_ulong(hash
);
1325 size
= rcu_dereference(ht
->t
.size
);
1326 index
= hash
& (size
- 1);
1327 order
= get_count_order_ulong(index
+ 1);
1328 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1329 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1330 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1331 dummy_node
= (struct cds_lfht_node
*) lookup
;
1332 /* We can always skip the dummy node initially */
1333 node
= rcu_dereference(dummy_node
->p
.next
);
1334 node
= clear_flag(node
);
1336 if (unlikely(is_end(node
))) {
1340 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1344 next
= rcu_dereference(node
->p
.next
);
1345 if (likely(!is_removed(next
))
1347 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1350 node
= clear_flag(next
);
1352 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1357 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1359 struct cds_lfht_node
*node
, *next
;
1360 unsigned long reverse_hash
;
1365 reverse_hash
= node
->p
.reverse_hash
;
1367 key_len
= node
->key_len
;
1369 node
= clear_flag(next
);
1372 if (unlikely(is_end(node
))) {
1376 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1380 next
= rcu_dereference(node
->p
.next
);
1381 if (likely(!is_removed(next
))
1383 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1386 node
= clear_flag(next
);
1388 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1393 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1395 struct cds_lfht_node
*node
, *next
;
1397 node
= clear_flag(iter
->next
);
1399 if (unlikely(is_end(node
))) {
1403 next
= rcu_dereference(node
->p
.next
);
1404 if (likely(!is_removed(next
))
1405 && !is_dummy(next
)) {
1408 node
= clear_flag(next
);
1410 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1415 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1417 struct _cds_lfht_node
*lookup
;
1420 * Get next after first dummy node. The first dummy node is the
1421 * first node of the linked list.
1423 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1424 iter
->next
= lookup
->next
;
1425 cds_lfht_next(ht
, iter
);
1428 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1430 unsigned long hash
, size
;
1432 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1433 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1435 size
= rcu_dereference(ht
->t
.size
);
1436 (void) _cds_lfht_add(ht
, size
, node
, ADD_DEFAULT
, 0);
1437 ht_count_add(ht
, size
);
1440 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1441 struct cds_lfht_node
*node
)
1443 unsigned long hash
, size
;
1444 struct cds_lfht_node
*ret
;
1446 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1447 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1449 size
= rcu_dereference(ht
->t
.size
);
1450 ret
= _cds_lfht_add(ht
, size
, node
, ADD_UNIQUE
, 0);
1452 ht_count_add(ht
, size
);
1456 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1457 struct cds_lfht_node
*node
)
1459 unsigned long hash
, size
;
1460 struct cds_lfht_node
*ret
;
1462 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1463 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1465 size
= rcu_dereference(ht
->t
.size
);
1466 ret
= _cds_lfht_add(ht
, size
, node
, ADD_REPLACE
, 0);
1468 ht_count_add(ht
, size
);
1472 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1473 struct cds_lfht_node
*new_node
)
1477 size
= rcu_dereference(ht
->t
.size
);
1478 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1482 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1487 size
= rcu_dereference(ht
->t
.size
);
1488 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1490 ht_count_del(ht
, size
);
1495 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1497 struct cds_lfht_node
*node
;
1498 struct _cds_lfht_node
*lookup
;
1499 unsigned long order
, i
, size
;
1501 /* Check that the table is empty */
1502 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1503 node
= (struct cds_lfht_node
*) lookup
;
1505 node
= clear_flag(node
)->p
.next
;
1506 if (!is_dummy(node
))
1508 assert(!is_removed(node
));
1509 } while (!is_end(node
));
1511 * size accessed without rcu_dereference because hash table is
1515 /* Internal sanity check: all nodes left should be dummy */
1516 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1519 len
= !order
? 1 : 1UL << (order
- 1);
1520 for (i
= 0; i
< len
; i
++) {
1521 dbg_printf("delete order %lu i %lu hash %lu\n",
1523 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1524 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1526 poison_free(ht
->t
.tbl
[order
]);
1532 * Should only be called when no more concurrent readers nor writers can
1533 * possibly access the table.
1535 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1539 /* Wait for in-flight resize operations to complete */
1540 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1541 cmm_smp_mb(); /* Store destroy before load resize */
1542 while (uatomic_read(&ht
->in_progress_resize
))
1543 poll(NULL
, 0, 100); /* wait for 100ms */
1544 ret
= cds_lfht_delete_dummy(ht
);
1547 free_per_cpu_items_count(ht
->percpu_count
);
1549 *attr
= ht
->resize_attr
;
1554 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1555 long *approx_before
,
1556 unsigned long *count
,
1557 unsigned long *removed
,
1560 struct cds_lfht_node
*node
, *next
;
1561 struct _cds_lfht_node
*lookup
;
1562 unsigned long nr_dummy
= 0;
1565 if (nr_cpus_mask
>= 0) {
1568 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1569 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1570 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1577 /* Count non-dummy nodes in the table */
1578 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1579 node
= (struct cds_lfht_node
*) lookup
;
1581 next
= rcu_dereference(node
->p
.next
);
1582 if (is_removed(next
)) {
1583 if (!is_dummy(next
))
1587 } else if (!is_dummy(next
))
1591 node
= clear_flag(next
);
1592 } while (!is_end(node
));
1593 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1595 if (nr_cpus_mask
>= 0) {
1598 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1599 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1600 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1605 /* called with resize mutex held */
1607 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1608 unsigned long old_size
, unsigned long new_size
)
1610 unsigned long old_order
, new_order
;
1612 old_order
= get_count_order_ulong(old_size
) + 1;
1613 new_order
= get_count_order_ulong(new_size
) + 1;
1614 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1615 old_size
, old_order
, new_size
, new_order
);
1616 assert(new_size
> old_size
);
1617 init_table(ht
, old_order
, new_order
- old_order
);
1620 /* called with resize mutex held */
1622 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1623 unsigned long old_size
, unsigned long new_size
)
1625 unsigned long old_order
, new_order
;
1627 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1628 old_order
= get_count_order_ulong(old_size
) + 1;
1629 new_order
= get_count_order_ulong(new_size
) + 1;
1630 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1631 old_size
, old_order
, new_size
, new_order
);
1632 assert(new_size
< old_size
);
1634 /* Remove and unlink all dummy nodes to remove. */
1635 fini_table(ht
, new_order
, old_order
- new_order
);
1639 /* called with resize mutex held */
1641 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1643 unsigned long new_size
, old_size
;
1646 * Resize table, re-do if the target size has changed under us.
1649 assert(uatomic_read(&ht
->in_progress_resize
));
1650 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1652 ht
->t
.resize_initiated
= 1;
1653 old_size
= ht
->t
.size
;
1654 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1655 if (old_size
< new_size
)
1656 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1657 else if (old_size
> new_size
)
1658 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1659 ht
->t
.resize_initiated
= 0;
1660 /* write resize_initiated before read resize_target */
1662 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1666 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1669 return _uatomic_max(&ht
->t
.resize_target
,
1670 size
<< growth_order
);
1674 void resize_target_update_count(struct cds_lfht
*ht
,
1675 unsigned long count
)
1677 count
= max(count
, MIN_TABLE_SIZE
);
1678 uatomic_set(&ht
->t
.resize_target
, count
);
1681 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1683 resize_target_update_count(ht
, new_size
);
1684 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1685 ht
->cds_lfht_rcu_thread_offline();
1686 pthread_mutex_lock(&ht
->resize_mutex
);
1687 _do_cds_lfht_resize(ht
);
1688 pthread_mutex_unlock(&ht
->resize_mutex
);
1689 ht
->cds_lfht_rcu_thread_online();
1693 void do_resize_cb(struct rcu_head
*head
)
1695 struct rcu_resize_work
*work
=
1696 caa_container_of(head
, struct rcu_resize_work
, head
);
1697 struct cds_lfht
*ht
= work
->ht
;
1699 ht
->cds_lfht_rcu_thread_offline();
1700 pthread_mutex_lock(&ht
->resize_mutex
);
1701 _do_cds_lfht_resize(ht
);
1702 pthread_mutex_unlock(&ht
->resize_mutex
);
1703 ht
->cds_lfht_rcu_thread_online();
1705 cmm_smp_mb(); /* finish resize before decrement */
1706 uatomic_dec(&ht
->in_progress_resize
);
1710 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1712 struct rcu_resize_work
*work
;
1713 unsigned long target_size
;
1715 target_size
= resize_target_update(ht
, size
, growth
);
1716 /* Store resize_target before read resize_initiated */
1718 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1719 uatomic_inc(&ht
->in_progress_resize
);
1720 cmm_smp_mb(); /* increment resize count before load destroy */
1721 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1722 uatomic_dec(&ht
->in_progress_resize
);
1725 work
= malloc(sizeof(*work
));
1727 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1728 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1732 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1735 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1736 unsigned long count
)
1738 struct rcu_resize_work
*work
;
1740 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1742 resize_target_update_count(ht
, count
);
1743 /* Store resize_target before read resize_initiated */
1745 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1746 uatomic_inc(&ht
->in_progress_resize
);
1747 cmm_smp_mb(); /* increment resize count before load destroy */
1748 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1749 uatomic_dec(&ht
->in_progress_resize
);
1752 work
= malloc(sizeof(*work
));
1754 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1755 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);