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.
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/jhash.h>
148 #include <urcu/compiler.h>
149 #include <urcu/rculfhash.h>
154 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
156 #define dbg_printf(fmt, args...)
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
165 #define COUNT_COMMIT_ORDER 10
166 #define CHAIN_LEN_TARGET 1
167 #define CHAIN_LEN_RESIZE_THRESHOLD 3
170 * Define the minimum table size. Protects against hash table resize overload
171 * when too many entries are added quickly before the resize can complete.
172 * This is especially the case if the table could be shrinked to a size of 1.
173 * TODO: we might want to make the add/remove operations help the resize to
174 * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on
177 #define MIN_TABLE_SIZE 128
180 #define max(a, b) ((a) > (b) ? (a) : (b))
184 * The removed flag needs to be updated atomically with the pointer.
185 * The dummy flag does not require to be updated atomically with the
186 * pointer, but it is added as a pointer low bit flag to save space.
188 #define REMOVED_FLAG (1UL << 0)
189 #define DUMMY_FLAG (1UL << 1)
190 #define FLAGS_MASK ((1UL << 2) - 1)
192 struct ht_items_count
{
193 unsigned long add
, remove
;
194 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
197 struct rcu_head head
;
198 struct _cds_lfht_node nodes
[0];
202 unsigned long size
; /* always a power of 2 */
203 unsigned long resize_target
;
204 int resize_initiated
;
205 struct rcu_head head
;
206 struct rcu_level
*tbl
[0];
210 struct rcu_table
*t
; /* shared */
211 cds_lfht_hash_fct hash_fct
;
212 cds_lfht_compare_fct compare_fct
;
213 unsigned long hash_seed
;
214 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
215 unsigned int in_progress_resize
, in_progress_destroy
;
216 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
217 void (*func
)(struct rcu_head
*head
));
218 void (*cds_lfht_synchronize_rcu
)(void);
219 unsigned long count
; /* global approximate item count */
220 struct ht_items_count
*percpu_count
; /* per-cpu item count */
223 struct rcu_resize_work
{
224 struct rcu_head head
;
229 * Algorithm to reverse bits in a word by lookup table, extended to
232 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
233 * Originally from Public Domain.
236 static const uint8_t BitReverseTable256
[256] =
238 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
239 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
240 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
241 R6(0), R6(2), R6(1), R6(3)
248 uint8_t bit_reverse_u8(uint8_t v
)
250 return BitReverseTable256
[v
];
253 static __attribute__((unused
))
254 uint32_t bit_reverse_u32(uint32_t v
)
256 return ((uint32_t) bit_reverse_u8(v
) << 24) |
257 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
258 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
259 ((uint32_t) bit_reverse_u8(v
>> 24));
262 static __attribute__((unused
))
263 uint64_t bit_reverse_u64(uint64_t v
)
265 return ((uint64_t) bit_reverse_u8(v
) << 56) |
266 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
267 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
268 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
269 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
270 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
271 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
272 ((uint64_t) bit_reverse_u8(v
>> 56));
276 unsigned long bit_reverse_ulong(unsigned long v
)
278 #if (CAA_BITS_PER_LONG == 32)
279 return bit_reverse_u32(v
);
281 return bit_reverse_u64(v
);
286 * fls: returns the position of the most significant bit.
287 * Returns 0 if no bit is set, else returns the position of the most
288 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
290 #if defined(__i386) || defined(__x86_64)
292 unsigned int fls_u32(uint32_t x
)
300 : "=r" (r
) : "rm" (x
));
306 #if defined(__x86_64)
308 unsigned int fls_u64(uint64_t x
)
316 : "=r" (r
) : "rm" (x
));
323 static __attribute__((unused
))
324 unsigned int fls_u64(uint64_t x
)
331 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
335 if (!(x
& 0xFFFF000000000000ULL
)) {
339 if (!(x
& 0xFF00000000000000ULL
)) {
343 if (!(x
& 0xF000000000000000ULL
)) {
347 if (!(x
& 0xC000000000000000ULL
)) {
351 if (!(x
& 0x8000000000000000ULL
)) {
360 static __attribute__((unused
))
361 unsigned int fls_u32(uint32_t x
)
367 if (!(x
& 0xFFFF0000U
)) {
371 if (!(x
& 0xFF000000U
)) {
375 if (!(x
& 0xF0000000U
)) {
379 if (!(x
& 0xC0000000U
)) {
383 if (!(x
& 0x80000000U
)) {
391 unsigned int fls_ulong(unsigned long x
)
393 #if (CAA_BITS_PER_lONG == 32)
400 int get_count_order_u32(uint32_t x
)
404 order
= fls_u32(x
) - 1;
410 int get_count_order_ulong(unsigned long x
)
414 order
= fls_ulong(x
) - 1;
421 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
);
424 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
425 * available, then we support hash table item accounting.
426 * In the unfortunate event the number of CPUs reported would be
427 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
429 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
432 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
433 unsigned long count
);
435 static long nr_cpus_mask
= -1;
438 struct ht_items_count
*alloc_per_cpu_items_count(void)
440 struct ht_items_count
*count
;
442 switch (nr_cpus_mask
) {
449 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
455 * round up number of CPUs to next power of two, so we
456 * can use & for modulo.
458 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
459 nr_cpus_mask
= maxcpus
- 1;
463 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
468 void free_per_cpu_items_count(struct ht_items_count
*count
)
478 assert(nr_cpus_mask
>= 0);
479 cpu
= sched_getcpu();
480 if (unlikely(cpu
< 0))
483 return cpu
& nr_cpus_mask
;
487 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
489 unsigned long percpu_count
;
492 if (unlikely(!ht
->percpu_count
))
495 if (unlikely(cpu
< 0))
497 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
498 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
501 dbg_printf("add percpu %lu\n", percpu_count
);
502 count
= uatomic_add_return(&ht
->count
,
503 1UL << COUNT_COMMIT_ORDER
);
505 if (!(count
& (count
- 1))) {
506 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
509 dbg_printf("add set global %lu\n", count
);
510 cds_lfht_resize_lazy_count(ht
, t
,
511 count
>> (CHAIN_LEN_TARGET
- 1));
517 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
519 unsigned long percpu_count
;
522 if (unlikely(!ht
->percpu_count
))
525 if (unlikely(cpu
< 0))
527 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
528 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
531 dbg_printf("remove percpu %lu\n", percpu_count
);
532 count
= uatomic_add_return(&ht
->count
,
533 -(1UL << COUNT_COMMIT_ORDER
));
535 if (!(count
& (count
- 1))) {
536 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
539 dbg_printf("remove set global %lu\n", count
);
540 cds_lfht_resize_lazy_count(ht
, t
,
541 count
>> (CHAIN_LEN_TARGET
- 1));
546 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
548 static const long nr_cpus_mask
= -1;
551 struct ht_items_count
*alloc_per_cpu_items_count(void)
557 void free_per_cpu_items_count(struct ht_items_count
*count
)
562 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
567 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
571 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
575 void check_resize(struct cds_lfht
*ht
, struct rcu_table
*t
,
580 count
= uatomic_read(&ht
->count
);
582 * Use bucket-local length for small table expand and for
583 * environments lacking per-cpu data support.
585 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
588 dbg_printf("WARNING: large chain length: %u.\n",
590 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
591 cds_lfht_resize_lazy(ht
, t
,
592 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
596 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
598 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
602 int is_removed(struct cds_lfht_node
*node
)
604 return ((unsigned long) node
) & REMOVED_FLAG
;
608 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
610 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
614 int is_dummy(struct cds_lfht_node
*node
)
616 return ((unsigned long) node
) & DUMMY_FLAG
;
620 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
622 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
626 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
628 unsigned long old1
, old2
;
630 old1
= uatomic_read(ptr
);
635 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
640 void cds_lfht_free_table_cb(struct rcu_head
*head
)
642 struct rcu_table
*t
=
643 caa_container_of(head
, struct rcu_table
, head
);
648 void cds_lfht_free_level(struct rcu_head
*head
)
650 struct rcu_level
*l
=
651 caa_container_of(head
, struct rcu_level
, head
);
656 * Remove all logically deleted nodes from a bucket up to a certain node key.
659 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
661 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
665 /* We can always skip the dummy node initially */
666 iter
= rcu_dereference(iter_prev
->p
.next
);
667 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
669 if (unlikely(!clear_flag(iter
)))
671 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
673 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
674 if (likely(is_removed(next
)))
676 iter_prev
= clear_flag(iter
);
679 assert(!is_removed(iter
));
681 new_next
= flag_dummy(clear_flag(next
));
683 new_next
= clear_flag(next
);
684 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
689 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
, struct rcu_table
*t
,
690 struct cds_lfht_node
*node
, int unique
, int dummy
)
692 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
694 struct _cds_lfht_node
*lookup
;
695 unsigned long hash
, index
, order
;
699 node
->p
.next
= flag_dummy(NULL
);
700 return node
; /* Initial first add (head) */
702 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
704 uint32_t chain_len
= 0;
707 * iter_prev points to the non-removed node prior to the
710 index
= hash
& (t
->size
- 1);
711 order
= get_count_order_ulong(index
+ 1);
712 lookup
= &t
->tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
713 iter_prev
= (struct cds_lfht_node
*) lookup
;
714 /* We can always skip the dummy node initially */
715 iter
= rcu_dereference(iter_prev
->p
.next
);
716 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
718 if (unlikely(!clear_flag(iter
)))
720 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
722 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
723 if (unlikely(is_removed(next
)))
727 && !ht
->compare_fct(node
->key
, node
->key_len
,
728 clear_flag(iter
)->key
,
729 clear_flag(iter
)->key_len
))
730 return clear_flag(iter
);
731 /* Only account for identical reverse hash once */
732 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
734 check_resize(ht
, t
, ++chain_len
);
735 iter_prev
= clear_flag(iter
);
739 assert(node
!= clear_flag(iter
));
740 assert(!is_removed(iter_prev
));
741 assert(iter_prev
!= node
);
743 node
->p
.next
= clear_flag(iter
);
745 node
->p
.next
= flag_dummy(clear_flag(iter
));
747 new_node
= flag_dummy(node
);
750 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
752 continue; /* retry */
756 assert(!is_removed(iter
));
758 new_next
= flag_dummy(clear_flag(next
));
760 new_next
= clear_flag(next
);
761 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
765 /* Garbage collect logically removed nodes in the bucket */
766 index
= hash
& (t
->size
- 1);
767 order
= get_count_order_ulong(index
+ 1);
768 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
769 dummy_node
= (struct cds_lfht_node
*) lookup
;
770 _cds_lfht_gc_bucket(dummy_node
, node
);
775 int _cds_lfht_remove(struct cds_lfht
*ht
, struct rcu_table
*t
,
776 struct cds_lfht_node
*node
, int dummy_removal
)
778 struct cds_lfht_node
*dummy
, *next
, *old
;
779 struct _cds_lfht_node
*lookup
;
781 unsigned long hash
, index
, order
;
783 /* logically delete the node */
784 old
= rcu_dereference(node
->p
.next
);
787 if (unlikely(is_removed(next
)))
790 assert(is_dummy(next
));
792 assert(!is_dummy(next
));
793 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
795 } while (old
!= next
);
797 /* We performed the (logical) deletion. */
801 node
= clear_flag(node
);
804 * Ensure that the node is not visible to readers anymore: lookup for
805 * the node, and remove it (along with any other logically removed node)
808 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
810 index
= hash
& (t
->size
- 1);
811 order
= get_count_order_ulong(index
+ 1);
812 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
813 dummy
= (struct cds_lfht_node
*) lookup
;
814 _cds_lfht_gc_bucket(dummy
, node
);
817 * Only the flagging action indicated that we (and no other)
818 * removed the node from the hash.
821 assert(is_removed(rcu_dereference(node
->p
.next
)));
828 void init_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
829 unsigned long first_order
, unsigned long len_order
)
831 unsigned long i
, end_order
;
833 dbg_printf("init table: first_order %lu end_order %lu\n",
834 first_order
, first_order
+ len_order
);
835 end_order
= first_order
+ len_order
;
836 t
->size
= !first_order
? 0 : (1UL << (first_order
- 1));
837 for (i
= first_order
; i
< end_order
; i
++) {
838 unsigned long j
, len
;
840 len
= !i
? 1 : 1UL << (i
- 1);
841 dbg_printf("init order %lu len: %lu\n", i
, len
);
842 t
->tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
843 + (len
* sizeof(struct _cds_lfht_node
)));
844 for (j
= 0; j
< len
; j
++) {
845 struct cds_lfht_node
*new_node
=
846 (struct cds_lfht_node
*) &t
->tbl
[i
]->nodes
[j
];
848 dbg_printf("init entry: i %lu j %lu hash %lu\n",
849 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
850 new_node
->p
.reverse_hash
=
851 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
852 (void) _cds_lfht_add(ht
, t
, new_node
, 0, 1);
853 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
856 /* Update table size */
857 t
->size
= !i
? 1 : (1UL << i
);
858 dbg_printf("init new size: %lu\n", t
->size
);
859 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
862 t
->resize_target
= t
->size
;
863 t
->resize_initiated
= 0;
867 void fini_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
868 unsigned long first_order
, unsigned long len_order
)
872 dbg_printf("fini table: first_order %lu end_order %lu\n",
873 first_order
, first_order
+ len_order
);
874 end_order
= first_order
+ len_order
;
875 assert(first_order
> 0);
876 assert(t
->size
== (1UL << (end_order
- 1)));
877 for (i
= end_order
- 1; i
>= first_order
; i
--) {
878 unsigned long j
, len
;
880 len
= !i
? 1 : 1UL << (i
- 1);
881 dbg_printf("fini order %lu len: %lu\n", i
, len
);
882 /* Update table size */
883 t
->size
= 1UL << (i
- 1);
885 for (j
= 0; j
< len
; j
++) {
886 struct cds_lfht_node
*new_node
=
887 (struct cds_lfht_node
*) &t
->tbl
[i
]->nodes
[j
];
889 dbg_printf("fini entry: i %lu j %lu hash %lu\n",
890 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
891 new_node
->p
.reverse_hash
=
892 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
893 (void) _cds_lfht_remove(ht
, t
, new_node
, 1);
894 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
897 ht
->cds_lfht_call_rcu(&t
->tbl
[i
]->head
, cds_lfht_free_level
);
898 dbg_printf("fini new size: %lu\n", t
->size
);
899 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
902 t
->resize_target
= t
->size
;
903 t
->resize_initiated
= 0;
906 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
907 cds_lfht_compare_fct compare_fct
,
908 unsigned long hash_seed
,
909 unsigned long init_size
,
910 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
911 void (*func
)(struct rcu_head
*head
)),
912 void (*cds_lfht_synchronize_rcu
)(void))
917 /* init_size must be power of two */
918 if (init_size
&& (init_size
& (init_size
- 1)))
920 ht
= calloc(1, sizeof(struct cds_lfht
));
921 ht
->hash_fct
= hash_fct
;
922 ht
->compare_fct
= compare_fct
;
923 ht
->hash_seed
= hash_seed
;
924 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
925 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
926 ht
->in_progress_resize
= 0;
927 ht
->percpu_count
= alloc_per_cpu_items_count();
928 /* this mutex should not nest in read-side C.S. */
929 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
930 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
931 ht
->t
= calloc(1, sizeof(struct cds_lfht
)
932 + (order
* sizeof(struct rcu_level
*)));
934 pthread_mutex_lock(&ht
->resize_mutex
);
935 init_table(ht
, ht
->t
, 0, order
);
936 pthread_mutex_unlock(&ht
->resize_mutex
);
940 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
943 struct cds_lfht_node
*node
, *next
;
944 struct _cds_lfht_node
*lookup
;
945 unsigned long hash
, reverse_hash
, index
, order
;
947 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
948 reverse_hash
= bit_reverse_ulong(hash
);
950 t
= rcu_dereference(ht
->t
);
951 index
= hash
& (t
->size
- 1);
952 order
= get_count_order_ulong(index
+ 1);
953 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
954 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
955 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
956 node
= (struct cds_lfht_node
*) lookup
;
960 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
964 next
= rcu_dereference(node
->p
.next
);
965 if (likely(!is_removed(next
))
967 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
970 node
= clear_flag(next
);
972 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
976 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
977 struct cds_lfht_node
*node
)
979 struct cds_lfht_node
*next
;
980 unsigned long reverse_hash
;
984 reverse_hash
= node
->p
.reverse_hash
;
986 key_len
= node
->key_len
;
987 next
= rcu_dereference(node
->p
.next
);
988 node
= clear_flag(next
);
993 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
997 next
= rcu_dereference(node
->p
.next
);
998 if (likely(!is_removed(next
))
1000 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1003 node
= clear_flag(next
);
1005 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1009 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1011 struct rcu_table
*t
;
1014 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1015 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1017 t
= rcu_dereference(ht
->t
);
1018 (void) _cds_lfht_add(ht
, t
, node
, 0, 0);
1019 ht_count_add(ht
, t
);
1022 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1023 struct cds_lfht_node
*node
)
1025 struct rcu_table
*t
;
1027 struct cds_lfht_node
*ret
;
1029 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1030 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1032 t
= rcu_dereference(ht
->t
);
1033 ret
= _cds_lfht_add(ht
, t
, node
, 1, 0);
1035 ht_count_add(ht
, t
);
1039 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1041 struct rcu_table
*t
;
1044 t
= rcu_dereference(ht
->t
);
1045 ret
= _cds_lfht_remove(ht
, t
, node
, 0);
1047 ht_count_remove(ht
, t
);
1052 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1054 struct rcu_table
*t
;
1055 struct cds_lfht_node
*node
;
1056 struct _cds_lfht_node
*lookup
;
1057 unsigned long order
, i
;
1060 /* Check that the table is empty */
1061 lookup
= &t
->tbl
[0]->nodes
[0];
1062 node
= (struct cds_lfht_node
*) lookup
;
1064 node
= clear_flag(node
)->p
.next
;
1065 if (!is_dummy(node
))
1067 assert(!is_removed(node
));
1068 } while (clear_flag(node
));
1069 /* Internal sanity check: all nodes left should be dummy */
1070 for (order
= 0; order
< get_count_order_ulong(t
->size
) + 1; order
++) {
1073 len
= !order
? 1 : 1UL << (order
- 1);
1074 for (i
= 0; i
< len
; i
++) {
1075 dbg_printf("delete order %lu i %lu hash %lu\n",
1077 bit_reverse_ulong(t
->tbl
[order
]->nodes
[i
].reverse_hash
));
1078 assert(is_dummy(t
->tbl
[order
]->nodes
[i
].next
));
1080 free(t
->tbl
[order
]);
1086 * Should only be called when no more concurrent readers nor writers can
1087 * possibly access the table.
1089 int cds_lfht_destroy(struct cds_lfht
*ht
)
1093 /* Wait for in-flight resize operations to complete */
1094 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1095 while (uatomic_read(&ht
->in_progress_resize
))
1096 poll(NULL
, 0, 100); /* wait for 100ms */
1097 ret
= cds_lfht_delete_dummy(ht
);
1101 free_per_cpu_items_count(ht
->percpu_count
);
1106 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1107 unsigned long *count
,
1108 unsigned long *removed
)
1110 struct rcu_table
*t
;
1111 struct cds_lfht_node
*node
, *next
;
1112 struct _cds_lfht_node
*lookup
;
1113 unsigned long nr_dummy
= 0;
1118 t
= rcu_dereference(ht
->t
);
1119 /* Count non-dummy nodes in the table */
1120 lookup
= &t
->tbl
[0]->nodes
[0];
1121 node
= (struct cds_lfht_node
*) lookup
;
1123 next
= rcu_dereference(node
->p
.next
);
1124 if (is_removed(next
)) {
1125 assert(!is_dummy(next
));
1127 } else if (!is_dummy(next
))
1131 node
= clear_flag(next
);
1133 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1136 /* called with resize mutex held */
1138 void _do_cds_lfht_grow(struct cds_lfht
*ht
, struct rcu_table
*old_t
,
1139 unsigned long old_size
, unsigned long new_size
)
1141 unsigned long old_order
, new_order
;
1142 struct rcu_table
*new_t
;
1144 old_order
= get_count_order_ulong(old_size
) + 1;
1145 new_order
= get_count_order_ulong(new_size
) + 1;
1146 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1147 old_size
, old_order
, new_size
, new_order
);
1148 new_t
= malloc(sizeof(struct cds_lfht
)
1149 + (new_order
* sizeof(struct rcu_level
*)));
1150 assert(new_size
> old_size
);
1151 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1152 old_order
* sizeof(struct rcu_level
*));
1153 init_table(ht
, new_t
, old_order
, new_order
- old_order
);
1154 /* Changing table and size atomically wrt lookups */
1155 rcu_assign_pointer(ht
->t
, new_t
);
1156 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1159 /* called with resize mutex held */
1161 void _do_cds_lfht_shrink(struct cds_lfht
*ht
, struct rcu_table
*old_t
,
1162 unsigned long old_size
, unsigned long new_size
)
1164 unsigned long old_order
, new_order
;
1165 struct rcu_table
*new_t
;
1167 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1168 old_order
= get_count_order_ulong(old_size
) + 1;
1169 new_order
= get_count_order_ulong(new_size
) + 1;
1170 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1171 old_size
, old_order
, new_size
, new_order
);
1172 new_t
= malloc(sizeof(struct cds_lfht
)
1173 + (new_order
* sizeof(struct rcu_level
*)));
1174 assert(new_size
< old_size
);
1175 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1176 new_order
* sizeof(struct rcu_level
*));
1177 new_t
->size
= !new_order
? 1 : (1UL << (new_order
- 1));
1178 new_t
->resize_target
= new_t
->size
;
1179 new_t
->resize_initiated
= 0;
1181 /* Changing table and size atomically wrt lookups */
1182 rcu_assign_pointer(ht
->t
, new_t
);
1185 * We need to wait for all reader threads to reach Q.S. (and
1186 * thus use the new table for lookups) before we can start
1187 * releasing the old dummy nodes.
1189 ht
->cds_lfht_synchronize_rcu();
1191 /* Unlink and remove all now-unused dummy node pointers. */
1192 fini_table(ht
, old_t
, new_order
, old_order
- new_order
);
1193 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1197 /* called with resize mutex held */
1199 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1201 unsigned long new_size
, old_size
;
1202 struct rcu_table
*old_t
;
1205 old_size
= old_t
->size
;
1206 new_size
= CMM_LOAD_SHARED(old_t
->resize_target
);
1207 if (old_size
< new_size
)
1208 _do_cds_lfht_grow(ht
, old_t
, old_size
, new_size
);
1209 else if (old_size
> new_size
)
1210 _do_cds_lfht_shrink(ht
, old_t
, old_size
, new_size
);
1212 CMM_STORE_SHARED(old_t
->resize_initiated
, 0);
1216 unsigned long resize_target_update(struct rcu_table
*t
,
1219 return _uatomic_max(&t
->resize_target
,
1220 t
->size
<< growth_order
);
1224 unsigned long resize_target_update_count(struct rcu_table
*t
,
1225 unsigned long count
)
1227 count
= max(count
, MIN_TABLE_SIZE
);
1228 return uatomic_set(&t
->resize_target
, count
);
1231 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1233 struct rcu_table
*t
= rcu_dereference(ht
->t
);
1234 unsigned long target_size
;
1236 target_size
= resize_target_update_count(t
, new_size
);
1237 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1238 pthread_mutex_lock(&ht
->resize_mutex
);
1239 _do_cds_lfht_resize(ht
);
1240 pthread_mutex_unlock(&ht
->resize_mutex
);
1244 void do_resize_cb(struct rcu_head
*head
)
1246 struct rcu_resize_work
*work
=
1247 caa_container_of(head
, struct rcu_resize_work
, head
);
1248 struct cds_lfht
*ht
= work
->ht
;
1250 pthread_mutex_lock(&ht
->resize_mutex
);
1251 _do_cds_lfht_resize(ht
);
1252 pthread_mutex_unlock(&ht
->resize_mutex
);
1254 cmm_smp_mb(); /* finish resize before decrement */
1255 uatomic_dec(&ht
->in_progress_resize
);
1259 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
)
1261 struct rcu_resize_work
*work
;
1262 unsigned long target_size
;
1264 target_size
= resize_target_update(t
, growth
);
1265 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
1266 uatomic_inc(&ht
->in_progress_resize
);
1267 cmm_smp_mb(); /* increment resize count before calling it */
1268 work
= malloc(sizeof(*work
));
1270 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1271 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1275 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1278 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
1279 unsigned long count
)
1281 struct rcu_resize_work
*work
;
1282 unsigned long target_size
;
1284 target_size
= resize_target_update_count(t
, count
);
1285 if (!CMM_LOAD_SHARED(t
->resize_initiated
)) {
1286 uatomic_inc(&ht
->in_progress_resize
);
1287 cmm_smp_mb(); /* increment resize count before calling it */
1288 work
= malloc(sizeof(*work
));
1290 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1291 CMM_STORE_SHARED(t
->resize_initiated
, 1);