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 /* Note: manually update allocation length when adding a field */
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
{
260 unsigned long i
, start
, len
;
261 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
262 unsigned long start
, unsigned long len
);
266 void _cds_lfht_add(struct cds_lfht
*ht
,
268 struct cds_lfht_node
*node
,
269 struct cds_lfht_iter
*unique_ret
,
273 * Algorithm to reverse bits in a word by lookup table, extended to
276 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
277 * Originally from Public Domain.
280 static const uint8_t BitReverseTable256
[256] =
282 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
283 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
284 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
285 R6(0), R6(2), R6(1), R6(3)
292 uint8_t bit_reverse_u8(uint8_t v
)
294 return BitReverseTable256
[v
];
297 static __attribute__((unused
))
298 uint32_t bit_reverse_u32(uint32_t v
)
300 return ((uint32_t) bit_reverse_u8(v
) << 24) |
301 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
302 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
303 ((uint32_t) bit_reverse_u8(v
>> 24));
306 static __attribute__((unused
))
307 uint64_t bit_reverse_u64(uint64_t v
)
309 return ((uint64_t) bit_reverse_u8(v
) << 56) |
310 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
311 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
312 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
313 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
314 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
315 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
316 ((uint64_t) bit_reverse_u8(v
>> 56));
320 unsigned long bit_reverse_ulong(unsigned long v
)
322 #if (CAA_BITS_PER_LONG == 32)
323 return bit_reverse_u32(v
);
325 return bit_reverse_u64(v
);
330 * fls: returns the position of the most significant bit.
331 * Returns 0 if no bit is set, else returns the position of the most
332 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
334 #if defined(__i386) || defined(__x86_64)
336 unsigned int fls_u32(uint32_t x
)
344 : "=r" (r
) : "rm" (x
));
350 #if defined(__x86_64)
352 unsigned int fls_u64(uint64_t x
)
360 : "=r" (r
) : "rm" (x
));
367 static __attribute__((unused
))
368 unsigned int fls_u64(uint64_t x
)
375 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
379 if (!(x
& 0xFFFF000000000000ULL
)) {
383 if (!(x
& 0xFF00000000000000ULL
)) {
387 if (!(x
& 0xF000000000000000ULL
)) {
391 if (!(x
& 0xC000000000000000ULL
)) {
395 if (!(x
& 0x8000000000000000ULL
)) {
404 static __attribute__((unused
))
405 unsigned int fls_u32(uint32_t x
)
411 if (!(x
& 0xFFFF0000U
)) {
415 if (!(x
& 0xFF000000U
)) {
419 if (!(x
& 0xF0000000U
)) {
423 if (!(x
& 0xC0000000U
)) {
427 if (!(x
& 0x80000000U
)) {
435 unsigned int fls_ulong(unsigned long x
)
437 #if (CAA_BITS_PER_lONG == 32)
445 * Return the minimum order for which x <= (1UL << order).
446 * Return -1 if x is 0.
448 int get_count_order_u32(uint32_t x
)
453 return fls_u32(x
- 1);
457 * Return the minimum order for which x <= (1UL << order).
458 * Return -1 if x is 0.
460 int get_count_order_ulong(unsigned long x
)
465 return 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 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
718 unsigned long index
, order
;
721 index
= hash
& (size
- 1);
723 * equivalent to get_count_order_ulong(index + 1), but optimizes
724 * away the non-existing 0 special-case for
725 * get_count_order_ulong.
727 order
= fls_ulong(index
);
729 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
730 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
732 return &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
736 * Remove all logically deleted nodes from a bucket up to a certain node key.
739 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
741 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
743 assert(!is_dummy(dummy
));
744 assert(!is_removed(dummy
));
745 assert(!is_dummy(node
));
746 assert(!is_removed(node
));
749 /* We can always skip the dummy node initially */
750 iter
= rcu_dereference(iter_prev
->p
.next
);
751 assert(!is_removed(iter
));
752 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
754 * We should never be called with dummy (start of chain)
755 * and logically removed node (end of path compression
756 * marker) being the actual same node. This would be a
757 * bug in the algorithm implementation.
759 assert(dummy
!= node
);
761 if (unlikely(is_end(iter
)))
763 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
765 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
766 if (likely(is_removed(next
)))
768 iter_prev
= clear_flag(iter
);
771 assert(!is_removed(iter
));
773 new_next
= flag_dummy(clear_flag(next
));
775 new_next
= clear_flag(next
);
776 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
782 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
783 struct cds_lfht_node
*old_node
,
784 struct cds_lfht_node
*old_next
,
785 struct cds_lfht_node
*new_node
)
787 struct cds_lfht_node
*dummy
, *ret_next
;
788 struct _cds_lfht_node
*lookup
;
790 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
793 assert(!is_removed(old_node
));
794 assert(!is_dummy(old_node
));
795 assert(!is_removed(new_node
));
796 assert(!is_dummy(new_node
));
797 assert(new_node
!= old_node
);
799 /* Insert after node to be replaced */
800 if (is_removed(old_next
)) {
802 * Too late, the old node has been removed under us
803 * between lookup and replace. Fail.
807 assert(!is_dummy(old_next
));
808 assert(new_node
!= clear_flag(old_next
));
809 new_node
->p
.next
= clear_flag(old_next
);
811 * Here is the whole trick for lock-free replace: we add
812 * the replacement node _after_ the node we want to
813 * replace by atomically setting its next pointer at the
814 * same time we set its removal flag. Given that
815 * the lookups/get next use an iterator aware of the
816 * next pointer, they will either skip the old node due
817 * to the removal flag and see the new node, or use
818 * the old node, but will not see the new one.
820 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
821 old_next
, flag_removed(new_node
));
822 if (ret_next
== old_next
)
823 break; /* We performed the replacement. */
828 * Ensure that the old node is not visible to readers anymore:
829 * lookup for the node, and remove it (along with any other
830 * logically removed node) if found.
832 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
833 dummy
= (struct cds_lfht_node
*) lookup
;
834 _cds_lfht_gc_bucket(dummy
, new_node
);
836 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
841 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
842 * mode. A NULL unique_ret allows creation of duplicate keys.
845 void _cds_lfht_add(struct cds_lfht
*ht
,
847 struct cds_lfht_node
*node
,
848 struct cds_lfht_iter
*unique_ret
,
851 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
853 struct _cds_lfht_node
*lookup
;
855 assert(!is_dummy(node
));
856 assert(!is_removed(node
));
860 node
->p
.next
= flag_dummy(get_end());
861 return; /* Initial first add (head) */
863 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
865 uint32_t chain_len
= 0;
868 * iter_prev points to the non-removed node prior to the
871 iter_prev
= (struct cds_lfht_node
*) lookup
;
872 /* We can always skip the dummy node initially */
873 iter
= rcu_dereference(iter_prev
->p
.next
);
874 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
876 if (unlikely(is_end(iter
)))
878 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
880 /* dummy node is the first node of the identical-hash-value chain */
881 if (dummy
&& clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
)
883 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
884 if (unlikely(is_removed(next
)))
888 && clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
889 && !ht
->compare_fct(node
->key
, node
->key_len
,
890 clear_flag(iter
)->key
,
891 clear_flag(iter
)->key_len
)) {
892 unique_ret
->node
= clear_flag(iter
);
893 unique_ret
->next
= next
;
896 /* Only account for identical reverse hash once */
897 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
899 check_resize(ht
, size
, ++chain_len
);
900 iter_prev
= clear_flag(iter
);
905 assert(node
!= clear_flag(iter
));
906 assert(!is_removed(iter_prev
));
907 assert(!is_removed(iter
));
908 assert(iter_prev
!= node
);
910 node
->p
.next
= clear_flag(iter
);
912 node
->p
.next
= flag_dummy(clear_flag(iter
));
914 new_node
= flag_dummy(node
);
917 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
919 continue; /* retry */
926 assert(!is_removed(iter
));
928 new_next
= flag_dummy(clear_flag(next
));
930 new_next
= clear_flag(next
);
931 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
936 unique_ret
->node
= return_node
;
937 /* unique_ret->next left unset, never used. */
942 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
943 struct cds_lfht_node
*node
,
946 struct cds_lfht_node
*dummy
, *next
, *old
;
947 struct _cds_lfht_node
*lookup
;
949 if (!node
) /* Return -ENOENT if asked to delete NULL node */
952 /* logically delete the node */
953 assert(!is_dummy(node
));
954 assert(!is_removed(node
));
955 old
= rcu_dereference(node
->p
.next
);
957 struct cds_lfht_node
*new_next
;
960 if (unlikely(is_removed(next
)))
963 assert(is_dummy(next
));
965 assert(!is_dummy(next
));
966 new_next
= flag_removed(next
);
967 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
968 } while (old
!= next
);
969 /* We performed the (logical) deletion. */
972 * Ensure that the node is not visible to readers anymore: lookup for
973 * the node, and remove it (along with any other logically removed node)
976 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
977 dummy
= (struct cds_lfht_node
*) lookup
;
978 _cds_lfht_gc_bucket(dummy
, node
);
980 assert(is_removed(rcu_dereference(node
->p
.next
)));
985 void *partition_resize_thread(void *arg
)
987 struct partition_resize_work
*work
= arg
;
989 work
->ht
->cds_lfht_rcu_register_thread();
990 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
991 work
->ht
->cds_lfht_rcu_unregister_thread();
996 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
998 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
999 unsigned long start
, unsigned long len
))
1001 unsigned long partition_len
;
1002 struct partition_resize_work
*work
;
1004 unsigned long nr_threads
;
1007 * Note: nr_cpus_mask + 1 is always power of 2.
1008 * We spawn just the number of threads we need to satisfy the minimum
1009 * partition size, up to the number of CPUs in the system.
1011 if (nr_cpus_mask
> 0) {
1012 nr_threads
= min(nr_cpus_mask
+ 1,
1013 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1017 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1018 work
= calloc(nr_threads
, sizeof(*work
));
1020 for (thread
= 0; thread
< nr_threads
; thread
++) {
1021 work
[thread
].ht
= ht
;
1023 work
[thread
].len
= partition_len
;
1024 work
[thread
].start
= thread
* partition_len
;
1025 work
[thread
].fct
= fct
;
1026 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1027 partition_resize_thread
, &work
[thread
]);
1030 for (thread
= 0; thread
< nr_threads
; thread
++) {
1031 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1038 * Holding RCU read lock to protect _cds_lfht_add against memory
1039 * reclaim that could be performed by other call_rcu worker threads (ABA
1042 * When we reach a certain length, we can split this population phase over
1043 * many worker threads, based on the number of CPUs available in the system.
1044 * This should therefore take care of not having the expand lagging behind too
1045 * many concurrent insertion threads by using the scheduler's ability to
1046 * schedule dummy node population fairly with insertions.
1049 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1050 unsigned long start
, unsigned long len
)
1054 ht
->cds_lfht_rcu_read_lock();
1055 for (j
= start
; j
< start
+ len
; j
++) {
1056 struct cds_lfht_node
*new_node
=
1057 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1059 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1060 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1061 new_node
->p
.reverse_hash
=
1062 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1063 _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
1066 ht
->cds_lfht_rcu_read_unlock();
1070 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1073 assert(nr_cpus_mask
!= -1);
1074 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1075 ht
->cds_lfht_rcu_thread_online();
1076 init_table_populate_partition(ht
, i
, 0, len
);
1077 ht
->cds_lfht_rcu_thread_offline();
1080 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1084 void init_table(struct cds_lfht
*ht
,
1085 unsigned long first_order
, unsigned long len_order
)
1087 unsigned long i
, end_order
;
1089 dbg_printf("init table: first_order %lu end_order %lu\n",
1090 first_order
, first_order
+ len_order
);
1091 end_order
= first_order
+ len_order
;
1092 for (i
= first_order
; i
< end_order
; i
++) {
1095 len
= !i
? 1 : 1UL << (i
- 1);
1096 dbg_printf("init order %lu len: %lu\n", i
, len
);
1098 /* Stop expand if the resize target changes under us */
1099 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1102 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1103 assert(ht
->t
.tbl
[i
]);
1106 * Set all dummy nodes reverse hash values for a level and
1107 * link all dummy nodes into the table.
1109 init_table_populate(ht
, i
, len
);
1112 * Update table size.
1114 cmm_smp_wmb(); /* populate data before RCU size */
1115 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1117 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1118 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1124 * Holding RCU read lock to protect _cds_lfht_remove against memory
1125 * reclaim that could be performed by other call_rcu worker threads (ABA
1127 * For a single level, we logically remove and garbage collect each node.
1129 * As a design choice, we perform logical removal and garbage collection on a
1130 * node-per-node basis to simplify this algorithm. We also assume keeping good
1131 * cache locality of the operation would overweight possible performance gain
1132 * that could be achieved by batching garbage collection for multiple levels.
1133 * However, this would have to be justified by benchmarks.
1135 * Concurrent removal and add operations are helping us perform garbage
1136 * collection of logically removed nodes. We guarantee that all logically
1137 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1138 * invoked to free a hole level of dummy nodes (after a grace period).
1140 * Logical removal and garbage collection can therefore be done in batch or on a
1141 * node-per-node basis, as long as the guarantee above holds.
1143 * When we reach a certain length, we can split this removal over many worker
1144 * threads, based on the number of CPUs available in the system. This should
1145 * take care of not letting resize process lag behind too many concurrent
1146 * updater threads actively inserting into the hash table.
1149 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1150 unsigned long start
, unsigned long len
)
1154 ht
->cds_lfht_rcu_read_lock();
1155 for (j
= start
; j
< start
+ len
; j
++) {
1156 struct cds_lfht_node
*fini_node
=
1157 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1159 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1160 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1161 fini_node
->p
.reverse_hash
=
1162 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1163 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1166 ht
->cds_lfht_rcu_read_unlock();
1170 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1173 assert(nr_cpus_mask
!= -1);
1174 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1175 ht
->cds_lfht_rcu_thread_online();
1176 remove_table_partition(ht
, i
, 0, len
);
1177 ht
->cds_lfht_rcu_thread_offline();
1180 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1184 void fini_table(struct cds_lfht
*ht
,
1185 unsigned long first_order
, unsigned long len_order
)
1188 void *free_by_rcu
= NULL
;
1190 dbg_printf("fini table: first_order %lu end_order %lu\n",
1191 first_order
, first_order
+ len_order
);
1192 end_order
= first_order
+ len_order
;
1193 assert(first_order
> 0);
1194 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1197 len
= !i
? 1 : 1UL << (i
- 1);
1198 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1200 /* Stop shrink if the resize target changes under us */
1201 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1204 cmm_smp_wmb(); /* populate data before RCU size */
1205 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1208 * We need to wait for all add operations to reach Q.S. (and
1209 * thus use the new table for lookups) before we can start
1210 * releasing the old dummy nodes. Otherwise their lookup will
1211 * return a logically removed node as insert position.
1213 ht
->cds_lfht_synchronize_rcu();
1218 * Set "removed" flag in dummy nodes about to be removed.
1219 * Unlink all now-logically-removed dummy node pointers.
1220 * Concurrent add/remove operation are helping us doing
1223 remove_table(ht
, i
, len
);
1225 free_by_rcu
= ht
->t
.tbl
[i
];
1227 dbg_printf("fini new size: %lu\n", 1UL << i
);
1228 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1233 ht
->cds_lfht_synchronize_rcu();
1238 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1239 cds_lfht_compare_fct compare_fct
,
1240 unsigned long hash_seed
,
1241 unsigned long init_size
,
1243 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1244 void (*func
)(struct rcu_head
*head
)),
1245 void (*cds_lfht_synchronize_rcu
)(void),
1246 void (*cds_lfht_rcu_read_lock
)(void),
1247 void (*cds_lfht_rcu_read_unlock
)(void),
1248 void (*cds_lfht_rcu_thread_offline
)(void),
1249 void (*cds_lfht_rcu_thread_online
)(void),
1250 void (*cds_lfht_rcu_register_thread
)(void),
1251 void (*cds_lfht_rcu_unregister_thread
)(void),
1252 pthread_attr_t
*attr
)
1254 struct cds_lfht
*ht
;
1255 unsigned long order
;
1257 /* init_size must be power of two */
1258 if (init_size
&& (init_size
& (init_size
- 1)))
1260 ht
= calloc(1, sizeof(struct cds_lfht
));
1262 ht
->hash_fct
= hash_fct
;
1263 ht
->compare_fct
= compare_fct
;
1264 ht
->hash_seed
= hash_seed
;
1265 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1266 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1267 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1268 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1269 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1270 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1271 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1272 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1273 ht
->resize_attr
= attr
;
1274 ht
->percpu_count
= alloc_per_cpu_items_count();
1275 /* this mutex should not nest in read-side C.S. */
1276 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1277 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1279 ht
->cds_lfht_rcu_thread_offline();
1280 pthread_mutex_lock(&ht
->resize_mutex
);
1281 ht
->t
.resize_target
= 1UL << (order
- 1);
1282 init_table(ht
, 0, order
);
1283 pthread_mutex_unlock(&ht
->resize_mutex
);
1284 ht
->cds_lfht_rcu_thread_online();
1288 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1289 struct cds_lfht_iter
*iter
)
1291 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1292 struct _cds_lfht_node
*lookup
;
1293 unsigned long hash
, reverse_hash
, size
;
1295 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1296 reverse_hash
= bit_reverse_ulong(hash
);
1298 size
= rcu_dereference(ht
->t
.size
);
1299 lookup
= lookup_bucket(ht
, size
, hash
);
1300 dummy_node
= (struct cds_lfht_node
*) lookup
;
1301 /* We can always skip the dummy node initially */
1302 node
= rcu_dereference(dummy_node
->p
.next
);
1303 node
= clear_flag(node
);
1305 if (unlikely(is_end(node
))) {
1309 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1313 next
= rcu_dereference(node
->p
.next
);
1314 if (likely(!is_removed(next
))
1316 && clear_flag(node
)->p
.reverse_hash
== reverse_hash
1317 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1320 node
= clear_flag(next
);
1322 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1327 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1329 struct cds_lfht_node
*node
, *next
;
1330 unsigned long reverse_hash
;
1335 reverse_hash
= node
->p
.reverse_hash
;
1337 key_len
= node
->key_len
;
1339 node
= clear_flag(next
);
1342 if (unlikely(is_end(node
))) {
1346 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1350 next
= rcu_dereference(node
->p
.next
);
1351 if (likely(!is_removed(next
))
1353 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1356 node
= clear_flag(next
);
1358 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1363 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1365 struct cds_lfht_node
*node
, *next
;
1367 node
= clear_flag(iter
->next
);
1369 if (unlikely(is_end(node
))) {
1373 next
= rcu_dereference(node
->p
.next
);
1374 if (likely(!is_removed(next
))
1375 && !is_dummy(next
)) {
1378 node
= clear_flag(next
);
1380 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1385 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1387 struct _cds_lfht_node
*lookup
;
1390 * Get next after first dummy node. The first dummy node is the
1391 * first node of the linked list.
1393 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1394 iter
->next
= lookup
->next
;
1395 cds_lfht_next(ht
, iter
);
1398 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1400 unsigned long hash
, size
;
1402 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1403 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1405 size
= rcu_dereference(ht
->t
.size
);
1406 _cds_lfht_add(ht
, size
, node
, NULL
, 0);
1407 ht_count_add(ht
, size
);
1410 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1411 struct cds_lfht_node
*node
)
1413 unsigned long hash
, size
;
1414 struct cds_lfht_iter iter
;
1416 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1417 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1419 size
= rcu_dereference(ht
->t
.size
);
1420 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1421 if (iter
.node
== node
)
1422 ht_count_add(ht
, size
);
1426 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1427 struct cds_lfht_node
*node
)
1429 unsigned long hash
, size
;
1430 struct cds_lfht_iter iter
;
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
);
1437 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1438 if (iter
.node
== node
) {
1439 ht_count_add(ht
, size
);
1443 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1448 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1449 struct cds_lfht_node
*new_node
)
1453 size
= rcu_dereference(ht
->t
.size
);
1454 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1458 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1463 size
= rcu_dereference(ht
->t
.size
);
1464 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1466 ht_count_del(ht
, size
);
1471 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1473 struct cds_lfht_node
*node
;
1474 struct _cds_lfht_node
*lookup
;
1475 unsigned long order
, i
, size
;
1477 /* Check that the table is empty */
1478 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1479 node
= (struct cds_lfht_node
*) lookup
;
1481 node
= clear_flag(node
)->p
.next
;
1482 if (!is_dummy(node
))
1484 assert(!is_removed(node
));
1485 } while (!is_end(node
));
1487 * size accessed without rcu_dereference because hash table is
1491 /* Internal sanity check: all nodes left should be dummy */
1492 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1495 len
= !order
? 1 : 1UL << (order
- 1);
1496 for (i
= 0; i
< len
; i
++) {
1497 dbg_printf("delete order %lu i %lu hash %lu\n",
1499 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1500 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1502 poison_free(ht
->t
.tbl
[order
]);
1508 * Should only be called when no more concurrent readers nor writers can
1509 * possibly access the table.
1511 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1515 /* Wait for in-flight resize operations to complete */
1516 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1517 cmm_smp_mb(); /* Store destroy before load resize */
1518 while (uatomic_read(&ht
->in_progress_resize
))
1519 poll(NULL
, 0, 100); /* wait for 100ms */
1520 ret
= cds_lfht_delete_dummy(ht
);
1523 free_per_cpu_items_count(ht
->percpu_count
);
1525 *attr
= ht
->resize_attr
;
1530 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1531 long *approx_before
,
1532 unsigned long *count
,
1533 unsigned long *removed
,
1536 struct cds_lfht_node
*node
, *next
;
1537 struct _cds_lfht_node
*lookup
;
1538 unsigned long nr_dummy
= 0;
1541 if (nr_cpus_mask
>= 0) {
1544 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1545 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1546 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1553 /* Count non-dummy nodes in the table */
1554 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1555 node
= (struct cds_lfht_node
*) lookup
;
1557 next
= rcu_dereference(node
->p
.next
);
1558 if (is_removed(next
)) {
1559 if (!is_dummy(next
))
1563 } else if (!is_dummy(next
))
1567 node
= clear_flag(next
);
1568 } while (!is_end(node
));
1569 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1571 if (nr_cpus_mask
>= 0) {
1574 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1575 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1576 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1581 /* called with resize mutex held */
1583 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1584 unsigned long old_size
, unsigned long new_size
)
1586 unsigned long old_order
, new_order
;
1588 old_order
= get_count_order_ulong(old_size
) + 1;
1589 new_order
= get_count_order_ulong(new_size
) + 1;
1590 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1591 old_size
, old_order
, new_size
, new_order
);
1592 assert(new_size
> old_size
);
1593 init_table(ht
, old_order
, new_order
- old_order
);
1596 /* called with resize mutex held */
1598 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1599 unsigned long old_size
, unsigned long new_size
)
1601 unsigned long old_order
, new_order
;
1603 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1604 old_order
= get_count_order_ulong(old_size
) + 1;
1605 new_order
= get_count_order_ulong(new_size
) + 1;
1606 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1607 old_size
, old_order
, new_size
, new_order
);
1608 assert(new_size
< old_size
);
1610 /* Remove and unlink all dummy nodes to remove. */
1611 fini_table(ht
, new_order
, old_order
- new_order
);
1615 /* called with resize mutex held */
1617 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1619 unsigned long new_size
, old_size
;
1622 * Resize table, re-do if the target size has changed under us.
1625 assert(uatomic_read(&ht
->in_progress_resize
));
1626 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1628 ht
->t
.resize_initiated
= 1;
1629 old_size
= ht
->t
.size
;
1630 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1631 if (old_size
< new_size
)
1632 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1633 else if (old_size
> new_size
)
1634 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1635 ht
->t
.resize_initiated
= 0;
1636 /* write resize_initiated before read resize_target */
1638 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1642 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1645 return _uatomic_max(&ht
->t
.resize_target
,
1646 size
<< growth_order
);
1650 void resize_target_update_count(struct cds_lfht
*ht
,
1651 unsigned long count
)
1653 count
= max(count
, MIN_TABLE_SIZE
);
1654 uatomic_set(&ht
->t
.resize_target
, count
);
1657 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1659 resize_target_update_count(ht
, new_size
);
1660 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1661 ht
->cds_lfht_rcu_thread_offline();
1662 pthread_mutex_lock(&ht
->resize_mutex
);
1663 _do_cds_lfht_resize(ht
);
1664 pthread_mutex_unlock(&ht
->resize_mutex
);
1665 ht
->cds_lfht_rcu_thread_online();
1669 void do_resize_cb(struct rcu_head
*head
)
1671 struct rcu_resize_work
*work
=
1672 caa_container_of(head
, struct rcu_resize_work
, head
);
1673 struct cds_lfht
*ht
= work
->ht
;
1675 ht
->cds_lfht_rcu_thread_offline();
1676 pthread_mutex_lock(&ht
->resize_mutex
);
1677 _do_cds_lfht_resize(ht
);
1678 pthread_mutex_unlock(&ht
->resize_mutex
);
1679 ht
->cds_lfht_rcu_thread_online();
1681 cmm_smp_mb(); /* finish resize before decrement */
1682 uatomic_dec(&ht
->in_progress_resize
);
1686 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1688 struct rcu_resize_work
*work
;
1689 unsigned long target_size
;
1691 target_size
= resize_target_update(ht
, size
, growth
);
1692 /* Store resize_target before read resize_initiated */
1694 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1695 uatomic_inc(&ht
->in_progress_resize
);
1696 cmm_smp_mb(); /* increment resize count before load destroy */
1697 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1698 uatomic_dec(&ht
->in_progress_resize
);
1701 work
= malloc(sizeof(*work
));
1703 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1704 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1708 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1711 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1712 unsigned long count
)
1714 struct rcu_resize_work
*work
;
1716 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1718 resize_target_update_count(ht
, count
);
1719 /* Store resize_target before read resize_initiated */
1721 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1722 uatomic_inc(&ht
->in_progress_resize
);
1723 cmm_smp_mb(); /* increment resize count before load destroy */
1724 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1725 uatomic_dec(&ht
->in_progress_resize
);
1728 work
= malloc(sizeof(*work
));
1730 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1731 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);