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
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 * Per-CPU 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 CHAIN_LEN_TARGET 1
183 #define CHAIN_LEN_RESIZE_THRESHOLD 3
186 * Define the minimum table size.
188 #define MIN_TABLE_SIZE 1
190 #if (CAA_BITS_PER_LONG == 32)
191 #define MAX_TABLE_ORDER 32
193 #define MAX_TABLE_ORDER 64
197 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
199 #define MIN_PARTITION_PER_THREAD_ORDER 12
200 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
203 #define min(a, b) ((a) < (b) ? (a) : (b))
207 #define max(a, b) ((a) > (b) ? (a) : (b))
211 * The removed flag needs to be updated atomically with the pointer.
212 * It indicates that no node must attach to the node scheduled for
213 * removal, and that node garbage collection must be performed.
214 * The dummy flag does not require to be updated atomically with the
215 * pointer, but it is added as a pointer low bit flag to save space.
217 #define REMOVED_FLAG (1UL << 0)
218 #define DUMMY_FLAG (1UL << 1)
219 #define FLAGS_MASK ((1UL << 2) - 1)
221 /* Value of the end pointer. Should not interact with flags. */
222 #define END_VALUE NULL
224 struct ht_items_count
{
225 unsigned long add
, del
;
226 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
229 /* Note: manually update allocation length when adding a field */
230 struct _cds_lfht_node nodes
[0];
234 unsigned long size
; /* always a power of 2, shared (RCU) */
235 unsigned long resize_target
;
236 int resize_initiated
;
237 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
242 cds_lfht_hash_fct hash_fct
;
243 cds_lfht_compare_fct compare_fct
;
244 unsigned long hash_seed
;
247 * We need to put the work threads offline (QSBR) when taking this
248 * mutex, because we use synchronize_rcu within this mutex critical
249 * section, which waits on read-side critical sections, and could
250 * therefore cause grace-period deadlock if we hold off RCU G.P.
253 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
254 unsigned int in_progress_resize
, in_progress_destroy
;
255 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
256 void (*func
)(struct rcu_head
*head
));
257 void (*cds_lfht_synchronize_rcu
)(void);
258 void (*cds_lfht_rcu_read_lock
)(void);
259 void (*cds_lfht_rcu_read_unlock
)(void);
260 void (*cds_lfht_rcu_thread_offline
)(void);
261 void (*cds_lfht_rcu_thread_online
)(void);
262 void (*cds_lfht_rcu_register_thread
)(void);
263 void (*cds_lfht_rcu_unregister_thread
)(void);
264 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
265 long count
; /* global approximate item count */
266 struct ht_items_count
*percpu_count
; /* per-cpu item count */
269 struct rcu_resize_work
{
270 struct rcu_head head
;
274 struct partition_resize_work
{
277 unsigned long i
, start
, len
;
278 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
279 unsigned long start
, unsigned long len
);
283 void _cds_lfht_add(struct cds_lfht
*ht
,
285 struct cds_lfht_node
*node
,
286 struct cds_lfht_iter
*unique_ret
,
290 * Algorithm to reverse bits in a word by lookup table, extended to
293 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
294 * Originally from Public Domain.
297 static const uint8_t BitReverseTable256
[256] =
299 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
300 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
301 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
302 R6(0), R6(2), R6(1), R6(3)
309 uint8_t bit_reverse_u8(uint8_t v
)
311 return BitReverseTable256
[v
];
314 static __attribute__((unused
))
315 uint32_t bit_reverse_u32(uint32_t v
)
317 return ((uint32_t) bit_reverse_u8(v
) << 24) |
318 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
319 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
320 ((uint32_t) bit_reverse_u8(v
>> 24));
323 static __attribute__((unused
))
324 uint64_t bit_reverse_u64(uint64_t v
)
326 return ((uint64_t) bit_reverse_u8(v
) << 56) |
327 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
328 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
329 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
330 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
331 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
332 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
333 ((uint64_t) bit_reverse_u8(v
>> 56));
337 unsigned long bit_reverse_ulong(unsigned long v
)
339 #if (CAA_BITS_PER_LONG == 32)
340 return bit_reverse_u32(v
);
342 return bit_reverse_u64(v
);
347 * fls: returns the position of the most significant bit.
348 * Returns 0 if no bit is set, else returns the position of the most
349 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
351 #if defined(__i386) || defined(__x86_64)
353 unsigned int fls_u32(uint32_t x
)
361 : "=r" (r
) : "rm" (x
));
367 #if defined(__x86_64)
369 unsigned int fls_u64(uint64_t x
)
377 : "=r" (r
) : "rm" (x
));
384 static __attribute__((unused
))
385 unsigned int fls_u64(uint64_t x
)
392 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
396 if (!(x
& 0xFFFF000000000000ULL
)) {
400 if (!(x
& 0xFF00000000000000ULL
)) {
404 if (!(x
& 0xF000000000000000ULL
)) {
408 if (!(x
& 0xC000000000000000ULL
)) {
412 if (!(x
& 0x8000000000000000ULL
)) {
421 static __attribute__((unused
))
422 unsigned int fls_u32(uint32_t x
)
428 if (!(x
& 0xFFFF0000U
)) {
432 if (!(x
& 0xFF000000U
)) {
436 if (!(x
& 0xF0000000U
)) {
440 if (!(x
& 0xC0000000U
)) {
444 if (!(x
& 0x80000000U
)) {
452 unsigned int fls_ulong(unsigned long x
)
454 #if (CAA_BITS_PER_lONG == 32)
462 * Return the minimum order for which x <= (1UL << order).
463 * Return -1 if x is 0.
465 int get_count_order_u32(uint32_t x
)
470 return fls_u32(x
- 1);
474 * Return the minimum order for which x <= (1UL << order).
475 * Return -1 if x is 0.
477 int get_count_order_ulong(unsigned long x
)
482 return fls_ulong(x
- 1);
486 #define poison_free(ptr) \
488 memset(ptr, 0x42, sizeof(*(ptr))); \
492 #define poison_free(ptr) free(ptr)
496 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
499 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
500 * available, then we support hash table item accounting.
501 * In the unfortunate event the number of CPUs reported would be
502 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
504 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
507 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
508 unsigned long count
);
510 static long nr_cpus_mask
= -1;
513 struct ht_items_count
*alloc_per_cpu_items_count(void)
515 struct ht_items_count
*count
;
517 switch (nr_cpus_mask
) {
524 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
530 * round up number of CPUs to next power of two, so we
531 * can use & for modulo.
533 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
534 nr_cpus_mask
= maxcpus
- 1;
538 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
543 void free_per_cpu_items_count(struct ht_items_count
*count
)
553 assert(nr_cpus_mask
>= 0);
554 cpu
= sched_getcpu();
555 if (unlikely(cpu
< 0))
558 return cpu
& nr_cpus_mask
;
562 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
564 unsigned long percpu_count
;
567 if (unlikely(!ht
->percpu_count
))
570 if (unlikely(cpu
< 0))
572 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
573 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
576 dbg_printf("add percpu %lu\n", percpu_count
);
577 count
= uatomic_add_return(&ht
->count
,
578 1UL << COUNT_COMMIT_ORDER
);
580 if (!(count
& (count
- 1))) {
581 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
583 dbg_printf("add set global %ld\n", count
);
584 cds_lfht_resize_lazy_count(ht
, size
,
585 count
>> (CHAIN_LEN_TARGET
- 1));
591 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
593 unsigned long percpu_count
;
596 if (unlikely(!ht
->percpu_count
))
599 if (unlikely(cpu
< 0))
601 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, 1);
602 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
605 dbg_printf("del percpu %lu\n", percpu_count
);
606 count
= uatomic_add_return(&ht
->count
,
607 -(1UL << COUNT_COMMIT_ORDER
));
609 if (!(count
& (count
- 1))) {
610 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
612 dbg_printf("del set global %ld\n", count
);
614 * Don't shrink table if the number of nodes is below a
617 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (nr_cpus_mask
+ 1))
619 cds_lfht_resize_lazy_count(ht
, size
,
620 count
>> (CHAIN_LEN_TARGET
- 1));
625 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
627 static const long nr_cpus_mask
= -2;
630 struct ht_items_count
*alloc_per_cpu_items_count(void)
636 void free_per_cpu_items_count(struct ht_items_count
*count
)
641 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
646 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
650 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
654 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
658 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
660 count
= uatomic_read(&ht
->count
);
662 * Use bucket-local length for small table expand and for
663 * environments lacking per-cpu data support.
665 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
668 dbg_printf("WARNING: large chain length: %u.\n",
670 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
671 cds_lfht_resize_lazy(ht
, size
,
672 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
676 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
678 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
682 int is_removed(struct cds_lfht_node
*node
)
684 return ((unsigned long) node
) & REMOVED_FLAG
;
688 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
690 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
694 int is_dummy(struct cds_lfht_node
*node
)
696 return ((unsigned long) node
) & DUMMY_FLAG
;
700 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
702 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
706 struct cds_lfht_node
*get_end(void)
708 return (struct cds_lfht_node
*) END_VALUE
;
712 int is_end(struct cds_lfht_node
*node
)
714 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
718 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
720 unsigned long old1
, old2
;
722 old1
= uatomic_read(ptr
);
727 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
732 struct _cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
735 unsigned long index
, order
;
738 index
= hash
& (size
- 1);
740 * equivalent to get_count_order_ulong(index + 1), but optimizes
741 * away the non-existing 0 special-case for
742 * get_count_order_ulong.
744 order
= fls_ulong(index
);
746 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
747 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
749 return &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
753 * Remove all logically deleted nodes from a bucket up to a certain node key.
756 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
758 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
760 assert(!is_dummy(dummy
));
761 assert(!is_removed(dummy
));
762 assert(!is_dummy(node
));
763 assert(!is_removed(node
));
766 /* We can always skip the dummy node initially */
767 iter
= rcu_dereference(iter_prev
->p
.next
);
768 assert(!is_removed(iter
));
769 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
771 * We should never be called with dummy (start of chain)
772 * and logically removed node (end of path compression
773 * marker) being the actual same node. This would be a
774 * bug in the algorithm implementation.
776 assert(dummy
!= node
);
778 if (unlikely(is_end(iter
)))
780 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
782 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
783 if (likely(is_removed(next
)))
785 iter_prev
= clear_flag(iter
);
788 assert(!is_removed(iter
));
790 new_next
= flag_dummy(clear_flag(next
));
792 new_next
= clear_flag(next
);
793 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
799 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
800 struct cds_lfht_node
*old_node
,
801 struct cds_lfht_node
*old_next
,
802 struct cds_lfht_node
*new_node
)
804 struct cds_lfht_node
*dummy
, *ret_next
;
805 struct _cds_lfht_node
*lookup
;
807 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
810 assert(!is_removed(old_node
));
811 assert(!is_dummy(old_node
));
812 assert(!is_removed(new_node
));
813 assert(!is_dummy(new_node
));
814 assert(new_node
!= old_node
);
816 /* Insert after node to be replaced */
817 if (is_removed(old_next
)) {
819 * Too late, the old node has been removed under us
820 * between lookup and replace. Fail.
824 assert(!is_dummy(old_next
));
825 assert(new_node
!= clear_flag(old_next
));
826 new_node
->p
.next
= clear_flag(old_next
);
828 * Here is the whole trick for lock-free replace: we add
829 * the replacement node _after_ the node we want to
830 * replace by atomically setting its next pointer at the
831 * same time we set its removal flag. Given that
832 * the lookups/get next use an iterator aware of the
833 * next pointer, they will either skip the old node due
834 * to the removal flag and see the new node, or use
835 * the old node, but will not see the new one.
837 ret_next
= uatomic_cmpxchg(&old_node
->p
.next
,
838 old_next
, flag_removed(new_node
));
839 if (ret_next
== old_next
)
840 break; /* We performed the replacement. */
845 * Ensure that the old node is not visible to readers anymore:
846 * lookup for the node, and remove it (along with any other
847 * logically removed node) if found.
849 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->p
.reverse_hash
));
850 dummy
= (struct cds_lfht_node
*) lookup
;
851 _cds_lfht_gc_bucket(dummy
, new_node
);
853 assert(is_removed(rcu_dereference(old_node
->p
.next
)));
858 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
859 * mode. A NULL unique_ret allows creation of duplicate keys.
862 void _cds_lfht_add(struct cds_lfht
*ht
,
864 struct cds_lfht_node
*node
,
865 struct cds_lfht_iter
*unique_ret
,
868 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
870 struct _cds_lfht_node
*lookup
;
872 assert(!is_dummy(node
));
873 assert(!is_removed(node
));
874 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
876 uint32_t chain_len
= 0;
879 * iter_prev points to the non-removed node prior to the
882 iter_prev
= (struct cds_lfht_node
*) lookup
;
883 /* We can always skip the dummy node initially */
884 iter
= rcu_dereference(iter_prev
->p
.next
);
885 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
887 if (unlikely(is_end(iter
)))
889 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
892 /* dummy node is the first node of the identical-hash-value chain */
893 if (dummy
&& clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
)
896 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
897 if (unlikely(is_removed(next
)))
903 && clear_flag(iter
)->p
.reverse_hash
== node
->p
.reverse_hash
) {
904 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
907 * uniquely adding inserts the node as the first
908 * node of the identical-hash-value node chain.
910 * This semantic ensures no duplicated keys
911 * should ever be observable in the table
912 * (including observe one node by one node
913 * by forward iterations)
915 cds_lfht_next_duplicate(ht
, &d_iter
);
919 *unique_ret
= d_iter
;
923 /* Only account for identical reverse hash once */
924 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
926 check_resize(ht
, size
, ++chain_len
);
927 iter_prev
= clear_flag(iter
);
932 assert(node
!= clear_flag(iter
));
933 assert(!is_removed(iter_prev
));
934 assert(!is_removed(iter
));
935 assert(iter_prev
!= node
);
937 node
->p
.next
= clear_flag(iter
);
939 node
->p
.next
= flag_dummy(clear_flag(iter
));
941 new_node
= flag_dummy(node
);
944 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
946 continue; /* retry */
953 assert(!is_removed(iter
));
955 new_next
= flag_dummy(clear_flag(next
));
957 new_next
= clear_flag(next
);
958 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
963 unique_ret
->node
= return_node
;
964 /* unique_ret->next left unset, never used. */
969 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
970 struct cds_lfht_node
*node
,
973 struct cds_lfht_node
*dummy
, *next
, *old
;
974 struct _cds_lfht_node
*lookup
;
976 if (!node
) /* Return -ENOENT if asked to delete NULL node */
979 /* logically delete the node */
980 assert(!is_dummy(node
));
981 assert(!is_removed(node
));
982 old
= rcu_dereference(node
->p
.next
);
984 struct cds_lfht_node
*new_next
;
987 if (unlikely(is_removed(next
)))
990 assert(is_dummy(next
));
992 assert(!is_dummy(next
));
993 new_next
= flag_removed(next
);
994 old
= uatomic_cmpxchg(&node
->p
.next
, next
, new_next
);
995 } while (old
!= next
);
996 /* We performed the (logical) deletion. */
999 * Ensure that the node is not visible to readers anymore: lookup for
1000 * the node, and remove it (along with any other logically removed node)
1003 lookup
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->p
.reverse_hash
));
1004 dummy
= (struct cds_lfht_node
*) lookup
;
1005 _cds_lfht_gc_bucket(dummy
, node
);
1007 assert(is_removed(rcu_dereference(node
->p
.next
)));
1012 void *partition_resize_thread(void *arg
)
1014 struct partition_resize_work
*work
= arg
;
1016 work
->ht
->cds_lfht_rcu_register_thread();
1017 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1018 work
->ht
->cds_lfht_rcu_unregister_thread();
1023 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1025 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1026 unsigned long start
, unsigned long len
))
1028 unsigned long partition_len
;
1029 struct partition_resize_work
*work
;
1031 unsigned long nr_threads
;
1034 * Note: nr_cpus_mask + 1 is always power of 2.
1035 * We spawn just the number of threads we need to satisfy the minimum
1036 * partition size, up to the number of CPUs in the system.
1038 if (nr_cpus_mask
> 0) {
1039 nr_threads
= min(nr_cpus_mask
+ 1,
1040 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1044 partition_len
= len
>> get_count_order_ulong(nr_threads
);
1045 work
= calloc(nr_threads
, sizeof(*work
));
1047 for (thread
= 0; thread
< nr_threads
; thread
++) {
1048 work
[thread
].ht
= ht
;
1050 work
[thread
].len
= partition_len
;
1051 work
[thread
].start
= thread
* partition_len
;
1052 work
[thread
].fct
= fct
;
1053 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1054 partition_resize_thread
, &work
[thread
]);
1057 for (thread
= 0; thread
< nr_threads
; thread
++) {
1058 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1065 * Holding RCU read lock to protect _cds_lfht_add against memory
1066 * reclaim that could be performed by other call_rcu worker threads (ABA
1069 * When we reach a certain length, we can split this population phase over
1070 * many worker threads, based on the number of CPUs available in the system.
1071 * This should therefore take care of not having the expand lagging behind too
1072 * many concurrent insertion threads by using the scheduler's ability to
1073 * schedule dummy node population fairly with insertions.
1076 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1077 unsigned long start
, unsigned long len
)
1082 ht
->cds_lfht_rcu_read_lock();
1083 for (j
= start
; j
< start
+ len
; j
++) {
1084 struct cds_lfht_node
*new_node
=
1085 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1087 dbg_printf("init populate: i %lu j %lu hash %lu\n",
1088 i
, j
, (1UL << (i
- 1)) + j
);
1089 new_node
->p
.reverse_hash
=
1090 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1091 _cds_lfht_add(ht
, 1UL << (i
- 1),
1094 ht
->cds_lfht_rcu_read_unlock();
1098 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1101 assert(nr_cpus_mask
!= -1);
1102 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1103 ht
->cds_lfht_rcu_thread_online();
1104 init_table_populate_partition(ht
, i
, 0, len
);
1105 ht
->cds_lfht_rcu_thread_offline();
1108 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1112 void init_table(struct cds_lfht
*ht
,
1113 unsigned long first_order
, unsigned long last_order
)
1117 dbg_printf("init table: first_order %lu last_order %lu\n",
1118 first_order
, last_order
);
1119 assert(first_order
> 0);
1120 for (i
= first_order
; i
<= last_order
; i
++) {
1123 len
= 1UL << (i
- 1);
1124 dbg_printf("init order %lu len: %lu\n", i
, len
);
1126 /* Stop expand if the resize target changes under us */
1127 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (1UL << i
))
1130 ht
->t
.tbl
[i
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1131 assert(ht
->t
.tbl
[i
]);
1134 * Set all dummy nodes reverse hash values for a level and
1135 * link all dummy nodes into the table.
1137 init_table_populate(ht
, i
, len
);
1140 * Update table size.
1142 cmm_smp_wmb(); /* populate data before RCU size */
1143 CMM_STORE_SHARED(ht
->t
.size
, 1UL << i
);
1145 dbg_printf("init new size: %lu\n", 1UL << i
);
1146 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1152 * Holding RCU read lock to protect _cds_lfht_remove against memory
1153 * reclaim that could be performed by other call_rcu worker threads (ABA
1155 * For a single level, we logically remove and garbage collect each node.
1157 * As a design choice, we perform logical removal and garbage collection on a
1158 * node-per-node basis to simplify this algorithm. We also assume keeping good
1159 * cache locality of the operation would overweight possible performance gain
1160 * that could be achieved by batching garbage collection for multiple levels.
1161 * However, this would have to be justified by benchmarks.
1163 * Concurrent removal and add operations are helping us perform garbage
1164 * collection of logically removed nodes. We guarantee that all logically
1165 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1166 * invoked to free a hole level of dummy nodes (after a grace period).
1168 * Logical removal and garbage collection can therefore be done in batch or on a
1169 * node-per-node basis, as long as the guarantee above holds.
1171 * When we reach a certain length, we can split this removal over many worker
1172 * threads, based on the number of CPUs available in the system. This should
1173 * take care of not letting resize process lag behind too many concurrent
1174 * updater threads actively inserting into the hash table.
1177 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1178 unsigned long start
, unsigned long len
)
1183 ht
->cds_lfht_rcu_read_lock();
1184 for (j
= start
; j
< start
+ len
; j
++) {
1185 struct cds_lfht_node
*fini_node
=
1186 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1188 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1189 i
, j
, (1UL << (i
- 1)) + j
);
1190 fini_node
->p
.reverse_hash
=
1191 bit_reverse_ulong((1UL << (i
- 1)) + j
);
1192 (void) _cds_lfht_del(ht
, 1UL << (i
- 1), fini_node
, 1);
1194 ht
->cds_lfht_rcu_read_unlock();
1198 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1201 assert(nr_cpus_mask
!= -1);
1202 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1203 ht
->cds_lfht_rcu_thread_online();
1204 remove_table_partition(ht
, i
, 0, len
);
1205 ht
->cds_lfht_rcu_thread_offline();
1208 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1212 void fini_table(struct cds_lfht
*ht
,
1213 unsigned long first_order
, unsigned long last_order
)
1216 void *free_by_rcu
= NULL
;
1218 dbg_printf("fini table: first_order %lu last_order %lu\n",
1219 first_order
, last_order
);
1220 assert(first_order
> 0);
1221 for (i
= last_order
; i
>= first_order
; i
--) {
1224 len
= 1UL << (i
- 1);
1225 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1227 /* Stop shrink if the resize target changes under us */
1228 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1231 cmm_smp_wmb(); /* populate data before RCU size */
1232 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1235 * We need to wait for all add operations to reach Q.S. (and
1236 * thus use the new table for lookups) before we can start
1237 * releasing the old dummy nodes. Otherwise their lookup will
1238 * return a logically removed node as insert position.
1240 ht
->cds_lfht_synchronize_rcu();
1245 * Set "removed" flag in dummy nodes about to be removed.
1246 * Unlink all now-logically-removed dummy node pointers.
1247 * Concurrent add/remove operation are helping us doing
1250 remove_table(ht
, i
, len
);
1252 free_by_rcu
= ht
->t
.tbl
[i
];
1254 dbg_printf("fini new size: %lu\n", 1UL << i
);
1255 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1260 ht
->cds_lfht_synchronize_rcu();
1266 void cds_lfht_create_dummy(struct cds_lfht
*ht
, unsigned long size
)
1268 struct _cds_lfht_node
*prev
, *node
;
1269 unsigned long order
, len
, i
, j
;
1271 ht
->t
.tbl
[0] = calloc(1, sizeof(struct _cds_lfht_node
));
1272 assert(ht
->t
.tbl
[0]);
1274 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1275 ht
->t
.tbl
[0]->nodes
[0].next
= flag_dummy(get_end());
1276 ht
->t
.tbl
[0]->nodes
[0].reverse_hash
= 0;
1278 for (order
= 1; order
< get_count_order_ulong(size
) + 1; order
++) {
1279 len
= 1UL << (order
- 1);
1280 ht
->t
.tbl
[order
] = calloc(1, len
* sizeof(struct _cds_lfht_node
));
1281 assert(ht
->t
.tbl
[order
]);
1284 prev
= ht
->t
.tbl
[i
]->nodes
;
1285 for (j
= 0; j
< len
; j
++) {
1286 if (j
& (j
- 1)) { /* Between power of 2 */
1288 } else if (j
) { /* At each power of 2 */
1290 prev
= ht
->t
.tbl
[i
]->nodes
;
1293 node
= &ht
->t
.tbl
[order
]->nodes
[j
];
1294 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1296 node
->next
= prev
->next
;
1297 assert(is_dummy(node
->next
));
1298 node
->reverse_hash
= bit_reverse_ulong(j
+ len
);
1299 prev
->next
= flag_dummy((struct cds_lfht_node
*)node
);
1304 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1305 cds_lfht_compare_fct compare_fct
,
1306 unsigned long hash_seed
,
1307 unsigned long init_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 /* init_size must be power of two */
1324 if (init_size
&& (init_size
& (init_size
- 1)))
1326 ht
= calloc(1, sizeof(struct cds_lfht
));
1328 ht
->hash_fct
= hash_fct
;
1329 ht
->compare_fct
= compare_fct
;
1330 ht
->hash_seed
= hash_seed
;
1331 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1332 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1333 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1334 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1335 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1336 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1337 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1338 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1339 ht
->resize_attr
= attr
;
1340 ht
->percpu_count
= alloc_per_cpu_items_count();
1341 /* this mutex should not nest in read-side C.S. */
1342 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1344 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
));
1345 ht
->t
.resize_target
= 1UL << order
;
1346 cds_lfht_create_dummy(ht
, 1UL << order
);
1347 ht
->t
.size
= 1UL << order
;
1351 void cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
,
1352 struct cds_lfht_iter
*iter
)
1354 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1355 struct _cds_lfht_node
*lookup
;
1356 unsigned long hash
, reverse_hash
, size
;
1358 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1359 reverse_hash
= bit_reverse_ulong(hash
);
1361 size
= rcu_dereference(ht
->t
.size
);
1362 lookup
= lookup_bucket(ht
, size
, hash
);
1363 dummy_node
= (struct cds_lfht_node
*) lookup
;
1364 /* We can always skip the dummy node initially */
1365 node
= rcu_dereference(dummy_node
->p
.next
);
1366 node
= clear_flag(node
);
1368 if (unlikely(is_end(node
))) {
1372 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1376 next
= rcu_dereference(node
->p
.next
);
1377 if (likely(!is_removed(next
))
1379 && clear_flag(node
)->p
.reverse_hash
== reverse_hash
1380 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1383 node
= clear_flag(next
);
1385 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1390 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1392 struct cds_lfht_node
*node
, *next
;
1393 unsigned long reverse_hash
;
1398 reverse_hash
= node
->p
.reverse_hash
;
1400 key_len
= node
->key_len
;
1402 node
= clear_flag(next
);
1405 if (unlikely(is_end(node
))) {
1409 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1413 next
= rcu_dereference(node
->p
.next
);
1414 if (likely(!is_removed(next
))
1416 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1419 node
= clear_flag(next
);
1421 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1426 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1428 struct cds_lfht_node
*node
, *next
;
1430 node
= clear_flag(iter
->next
);
1432 if (unlikely(is_end(node
))) {
1436 next
= rcu_dereference(node
->p
.next
);
1437 if (likely(!is_removed(next
))
1438 && !is_dummy(next
)) {
1441 node
= clear_flag(next
);
1443 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1448 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1450 struct _cds_lfht_node
*lookup
;
1453 * Get next after first dummy node. The first dummy node is the
1454 * first node of the linked list.
1456 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1457 iter
->next
= lookup
->next
;
1458 cds_lfht_next(ht
, iter
);
1461 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1463 unsigned long hash
, size
;
1465 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1466 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1468 size
= rcu_dereference(ht
->t
.size
);
1469 _cds_lfht_add(ht
, size
, node
, NULL
, 0);
1470 ht_count_add(ht
, size
);
1473 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1474 struct cds_lfht_node
*node
)
1476 unsigned long hash
, size
;
1477 struct cds_lfht_iter iter
;
1479 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1480 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1482 size
= rcu_dereference(ht
->t
.size
);
1483 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1484 if (iter
.node
== node
)
1485 ht_count_add(ht
, size
);
1489 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1490 struct cds_lfht_node
*node
)
1492 unsigned long hash
, size
;
1493 struct cds_lfht_iter iter
;
1495 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1496 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1498 size
= rcu_dereference(ht
->t
.size
);
1500 _cds_lfht_add(ht
, size
, node
, &iter
, 0);
1501 if (iter
.node
== node
) {
1502 ht_count_add(ht
, size
);
1506 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1511 int cds_lfht_replace(struct cds_lfht
*ht
, struct cds_lfht_iter
*old_iter
,
1512 struct cds_lfht_node
*new_node
)
1516 size
= rcu_dereference(ht
->t
.size
);
1517 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1521 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1526 size
= rcu_dereference(ht
->t
.size
);
1527 ret
= _cds_lfht_del(ht
, size
, iter
->node
, 0);
1529 ht_count_del(ht
, size
);
1534 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1536 struct cds_lfht_node
*node
;
1537 struct _cds_lfht_node
*lookup
;
1538 unsigned long order
, i
, size
;
1540 /* Check that the table is empty */
1541 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1542 node
= (struct cds_lfht_node
*) lookup
;
1544 node
= clear_flag(node
)->p
.next
;
1545 if (!is_dummy(node
))
1547 assert(!is_removed(node
));
1548 } while (!is_end(node
));
1550 * size accessed without rcu_dereference because hash table is
1554 /* Internal sanity check: all nodes left should be dummy */
1555 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1558 len
= !order
? 1 : 1UL << (order
- 1);
1559 for (i
= 0; i
< len
; i
++) {
1560 dbg_printf("delete order %lu i %lu hash %lu\n",
1562 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1563 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1565 poison_free(ht
->t
.tbl
[order
]);
1571 * Should only be called when no more concurrent readers nor writers can
1572 * possibly access the table.
1574 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1578 /* Wait for in-flight resize operations to complete */
1579 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1580 cmm_smp_mb(); /* Store destroy before load resize */
1581 while (uatomic_read(&ht
->in_progress_resize
))
1582 poll(NULL
, 0, 100); /* wait for 100ms */
1583 ret
= cds_lfht_delete_dummy(ht
);
1586 free_per_cpu_items_count(ht
->percpu_count
);
1588 *attr
= ht
->resize_attr
;
1593 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1594 long *approx_before
,
1595 unsigned long *count
,
1596 unsigned long *removed
,
1599 struct cds_lfht_node
*node
, *next
;
1600 struct _cds_lfht_node
*lookup
;
1601 unsigned long nr_dummy
= 0;
1604 if (nr_cpus_mask
>= 0) {
1607 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1608 *approx_before
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1609 *approx_before
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1616 /* Count non-dummy nodes in the table */
1617 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1618 node
= (struct cds_lfht_node
*) lookup
;
1620 next
= rcu_dereference(node
->p
.next
);
1621 if (is_removed(next
)) {
1622 if (!is_dummy(next
))
1626 } else if (!is_dummy(next
))
1630 node
= clear_flag(next
);
1631 } while (!is_end(node
));
1632 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1634 if (nr_cpus_mask
>= 0) {
1637 for (i
= 0; i
< nr_cpus_mask
+ 1; i
++) {
1638 *approx_after
+= uatomic_read(&ht
->percpu_count
[i
].add
);
1639 *approx_after
-= uatomic_read(&ht
->percpu_count
[i
].del
);
1644 /* called with resize mutex held */
1646 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1647 unsigned long old_size
, unsigned long new_size
)
1649 unsigned long old_order
, new_order
;
1651 old_order
= get_count_order_ulong(old_size
);
1652 new_order
= get_count_order_ulong(new_size
);
1653 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1654 old_size
, old_order
, new_size
, new_order
);
1655 assert(new_size
> old_size
);
1656 init_table(ht
, old_order
+ 1, new_order
);
1659 /* called with resize mutex held */
1661 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1662 unsigned long old_size
, unsigned long new_size
)
1664 unsigned long old_order
, new_order
;
1666 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1667 old_order
= get_count_order_ulong(old_size
);
1668 new_order
= get_count_order_ulong(new_size
);
1669 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1670 old_size
, old_order
, new_size
, new_order
);
1671 assert(new_size
< old_size
);
1673 /* Remove and unlink all dummy nodes to remove. */
1674 fini_table(ht
, new_order
+ 1, old_order
);
1678 /* called with resize mutex held */
1680 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1682 unsigned long new_size
, old_size
;
1685 * Resize table, re-do if the target size has changed under us.
1688 assert(uatomic_read(&ht
->in_progress_resize
));
1689 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1691 ht
->t
.resize_initiated
= 1;
1692 old_size
= ht
->t
.size
;
1693 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1694 if (old_size
< new_size
)
1695 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1696 else if (old_size
> new_size
)
1697 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1698 ht
->t
.resize_initiated
= 0;
1699 /* write resize_initiated before read resize_target */
1701 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1705 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1708 return _uatomic_max(&ht
->t
.resize_target
,
1709 size
<< growth_order
);
1713 void resize_target_update_count(struct cds_lfht
*ht
,
1714 unsigned long count
)
1716 count
= max(count
, MIN_TABLE_SIZE
);
1717 uatomic_set(&ht
->t
.resize_target
, count
);
1720 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1722 resize_target_update_count(ht
, new_size
);
1723 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1724 ht
->cds_lfht_rcu_thread_offline();
1725 pthread_mutex_lock(&ht
->resize_mutex
);
1726 _do_cds_lfht_resize(ht
);
1727 pthread_mutex_unlock(&ht
->resize_mutex
);
1728 ht
->cds_lfht_rcu_thread_online();
1732 void do_resize_cb(struct rcu_head
*head
)
1734 struct rcu_resize_work
*work
=
1735 caa_container_of(head
, struct rcu_resize_work
, head
);
1736 struct cds_lfht
*ht
= work
->ht
;
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();
1744 cmm_smp_mb(); /* finish resize before decrement */
1745 uatomic_dec(&ht
->in_progress_resize
);
1749 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1751 struct rcu_resize_work
*work
;
1752 unsigned long target_size
;
1754 target_size
= resize_target_update(ht
, size
, growth
);
1755 /* Store resize_target before read resize_initiated */
1757 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1758 uatomic_inc(&ht
->in_progress_resize
);
1759 cmm_smp_mb(); /* increment resize count before load destroy */
1760 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1761 uatomic_dec(&ht
->in_progress_resize
);
1764 work
= malloc(sizeof(*work
));
1766 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1767 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1771 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1774 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1775 unsigned long count
)
1777 struct rcu_resize_work
*work
;
1779 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1781 resize_target_update_count(ht
, count
);
1782 /* Store resize_target before read resize_initiated */
1784 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1785 uatomic_inc(&ht
->in_progress_resize
);
1786 cmm_smp_mb(); /* increment resize count before load destroy */
1787 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1788 uatomic_dec(&ht
->in_progress_resize
);
1791 work
= malloc(sizeof(*work
));
1793 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1794 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);