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/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
179 #if (CAA_BITS_PER_LONG == 32)
180 #define MAX_TABLE_ORDER 32
182 #define MAX_TABLE_ORDER 64
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 * The dummy flag does not require to be updated atomically with the
196 * pointer, but it is added as a pointer low bit flag to save space.
198 #define REMOVED_FLAG (1UL << 0)
199 #define DUMMY_FLAG (1UL << 1)
200 #define FLAGS_MASK ((1UL << 2) - 1)
202 struct ht_items_count
{
203 unsigned long add
, remove
;
204 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
207 struct rcu_head head
;
208 struct _cds_lfht_node nodes
[0];
212 unsigned long size
; /* always a power of 2, shared (RCU) */
213 unsigned long resize_target
;
214 int resize_initiated
;
215 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
220 cds_lfht_hash_fct hash_fct
;
221 cds_lfht_compare_fct compare_fct
;
222 unsigned long hash_seed
;
225 * We need to put the work threads offline (QSBR) when taking this
226 * mutex, because we use synchronize_rcu within this mutex critical
227 * section, which waits on read-side critical sections, and could
228 * therefore cause grace-period deadlock if we hold off RCU G.P.
231 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
232 unsigned int in_progress_resize
, in_progress_destroy
;
233 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
234 void (*func
)(struct rcu_head
*head
));
235 void (*cds_lfht_synchronize_rcu
)(void);
236 void (*cds_lfht_rcu_read_lock
)(void);
237 void (*cds_lfht_rcu_read_unlock
)(void);
238 void (*cds_lfht_rcu_thread_offline
)(void);
239 void (*cds_lfht_rcu_thread_online
)(void);
240 unsigned long count
; /* global approximate item count */
241 struct ht_items_count
*percpu_count
; /* per-cpu item count */
244 struct rcu_resize_work
{
245 struct rcu_head head
;
250 * Algorithm to reverse bits in a word by lookup table, extended to
253 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
254 * Originally from Public Domain.
257 static const uint8_t BitReverseTable256
[256] =
259 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
260 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
261 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
262 R6(0), R6(2), R6(1), R6(3)
269 uint8_t bit_reverse_u8(uint8_t v
)
271 return BitReverseTable256
[v
];
274 static __attribute__((unused
))
275 uint32_t bit_reverse_u32(uint32_t v
)
277 return ((uint32_t) bit_reverse_u8(v
) << 24) |
278 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
279 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
280 ((uint32_t) bit_reverse_u8(v
>> 24));
283 static __attribute__((unused
))
284 uint64_t bit_reverse_u64(uint64_t v
)
286 return ((uint64_t) bit_reverse_u8(v
) << 56) |
287 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
288 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
289 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
290 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
291 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
292 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
293 ((uint64_t) bit_reverse_u8(v
>> 56));
297 unsigned long bit_reverse_ulong(unsigned long v
)
299 #if (CAA_BITS_PER_LONG == 32)
300 return bit_reverse_u32(v
);
302 return bit_reverse_u64(v
);
307 * fls: returns the position of the most significant bit.
308 * Returns 0 if no bit is set, else returns the position of the most
309 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
311 #if defined(__i386) || defined(__x86_64)
313 unsigned int fls_u32(uint32_t x
)
321 : "=r" (r
) : "rm" (x
));
327 #if defined(__x86_64)
329 unsigned int fls_u64(uint64_t x
)
337 : "=r" (r
) : "rm" (x
));
344 static __attribute__((unused
))
345 unsigned int fls_u64(uint64_t x
)
352 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
356 if (!(x
& 0xFFFF000000000000ULL
)) {
360 if (!(x
& 0xFF00000000000000ULL
)) {
364 if (!(x
& 0xF000000000000000ULL
)) {
368 if (!(x
& 0xC000000000000000ULL
)) {
372 if (!(x
& 0x8000000000000000ULL
)) {
381 static __attribute__((unused
))
382 unsigned int fls_u32(uint32_t x
)
388 if (!(x
& 0xFFFF0000U
)) {
392 if (!(x
& 0xFF000000U
)) {
396 if (!(x
& 0xF0000000U
)) {
400 if (!(x
& 0xC0000000U
)) {
404 if (!(x
& 0x80000000U
)) {
412 unsigned int fls_ulong(unsigned long x
)
414 #if (CAA_BITS_PER_lONG == 32)
421 int get_count_order_u32(uint32_t x
)
425 order
= fls_u32(x
) - 1;
431 int get_count_order_ulong(unsigned long x
)
435 order
= fls_ulong(x
) - 1;
442 #define poison_free(ptr) \
444 memset(ptr, 0x42, sizeof(*(ptr))); \
448 #define poison_free(ptr) free(ptr)
452 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
455 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
456 * available, then we support hash table item accounting.
457 * In the unfortunate event the number of CPUs reported would be
458 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
460 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
463 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
464 unsigned long count
);
466 static long nr_cpus_mask
= -1;
469 struct ht_items_count
*alloc_per_cpu_items_count(void)
471 struct ht_items_count
*count
;
473 switch (nr_cpus_mask
) {
480 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
486 * round up number of CPUs to next power of two, so we
487 * can use & for modulo.
489 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
490 nr_cpus_mask
= maxcpus
- 1;
494 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
499 void free_per_cpu_items_count(struct ht_items_count
*count
)
509 assert(nr_cpus_mask
>= 0);
510 cpu
= sched_getcpu();
511 if (unlikely(cpu
< 0))
514 return cpu
& nr_cpus_mask
;
518 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
520 unsigned long percpu_count
;
523 if (unlikely(!ht
->percpu_count
))
526 if (unlikely(cpu
< 0))
528 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
529 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
532 dbg_printf("add percpu %lu\n", percpu_count
);
533 count
= uatomic_add_return(&ht
->count
,
534 1UL << COUNT_COMMIT_ORDER
);
536 if (!(count
& (count
- 1))) {
537 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
539 dbg_printf("add set global %lu\n", count
);
540 cds_lfht_resize_lazy_count(ht
, size
,
541 count
>> (CHAIN_LEN_TARGET
- 1));
547 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
549 unsigned long percpu_count
;
552 if (unlikely(!ht
->percpu_count
))
555 if (unlikely(cpu
< 0))
557 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
558 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
561 dbg_printf("remove percpu %lu\n", percpu_count
);
562 count
= uatomic_add_return(&ht
->count
,
563 -(1UL << COUNT_COMMIT_ORDER
));
565 if (!(count
& (count
- 1))) {
566 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
568 dbg_printf("remove set global %lu\n", count
);
569 cds_lfht_resize_lazy_count(ht
, size
,
570 count
>> (CHAIN_LEN_TARGET
- 1));
575 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
577 static const long nr_cpus_mask
= -1;
580 struct ht_items_count
*alloc_per_cpu_items_count(void)
586 void free_per_cpu_items_count(struct ht_items_count
*count
)
591 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
596 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
600 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
604 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
608 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
610 count
= uatomic_read(&ht
->count
);
612 * Use bucket-local length for small table expand and for
613 * environments lacking per-cpu data support.
615 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
618 dbg_printf("WARNING: large chain length: %u.\n",
620 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
621 cds_lfht_resize_lazy(ht
, size
,
622 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
626 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
628 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
632 int is_removed(struct cds_lfht_node
*node
)
634 return ((unsigned long) node
) & REMOVED_FLAG
;
638 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
640 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
644 int is_dummy(struct cds_lfht_node
*node
)
646 return ((unsigned long) node
) & DUMMY_FLAG
;
650 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
652 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
656 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
658 unsigned long old1
, old2
;
660 old1
= uatomic_read(ptr
);
665 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
670 void cds_lfht_free_level(struct rcu_head
*head
)
672 struct rcu_level
*l
=
673 caa_container_of(head
, struct rcu_level
, head
);
678 * Remove all logically deleted nodes from a bucket up to a certain node key.
681 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
683 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
685 assert(!is_dummy(dummy
));
686 assert(!is_removed(dummy
));
687 assert(!is_dummy(node
));
688 assert(!is_removed(node
));
691 /* We can always skip the dummy node initially */
692 iter
= rcu_dereference(iter_prev
->p
.next
);
693 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
695 * We should never be called with dummy (start of chain)
696 * and logically removed node (end of path compression
697 * marker) being the actual same node. This would be a
698 * bug in the algorithm implementation.
700 assert(dummy
!= node
);
702 if (unlikely(!clear_flag(iter
)))
704 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
706 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
707 if (likely(is_removed(next
)))
709 iter_prev
= clear_flag(iter
);
712 assert(!is_removed(iter
));
714 new_next
= flag_dummy(clear_flag(next
));
716 new_next
= clear_flag(next
);
717 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
722 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
724 struct cds_lfht_node
*node
,
725 int unique
, int dummy
)
727 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
729 struct _cds_lfht_node
*lookup
;
730 unsigned long hash
, index
, order
;
732 assert(!is_dummy(node
));
733 assert(!is_removed(node
));
736 node
->p
.next
= flag_dummy(NULL
);
737 return node
; /* Initial first add (head) */
739 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
741 uint32_t chain_len
= 0;
744 * iter_prev points to the non-removed node prior to the
747 index
= hash
& (size
- 1);
748 order
= get_count_order_ulong(index
+ 1);
749 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
750 iter_prev
= (struct cds_lfht_node
*) lookup
;
751 /* We can always skip the dummy node initially */
752 iter
= rcu_dereference(iter_prev
->p
.next
);
753 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
755 if (unlikely(!clear_flag(iter
)))
757 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
759 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
760 if (unlikely(is_removed(next
)))
764 && !ht
->compare_fct(node
->key
, node
->key_len
,
765 clear_flag(iter
)->key
,
766 clear_flag(iter
)->key_len
))
767 return clear_flag(iter
);
768 /* Only account for identical reverse hash once */
769 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
771 check_resize(ht
, size
, ++chain_len
);
772 iter_prev
= clear_flag(iter
);
776 assert(node
!= clear_flag(iter
));
777 assert(!is_removed(iter_prev
));
778 assert(!is_removed(iter
));
779 assert(iter_prev
!= node
);
781 node
->p
.next
= clear_flag(iter
);
783 node
->p
.next
= flag_dummy(clear_flag(iter
));
785 new_node
= flag_dummy(node
);
788 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
790 continue; /* retry */
794 assert(!is_removed(iter
));
796 new_next
= flag_dummy(clear_flag(next
));
798 new_next
= clear_flag(next
);
799 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
803 /* Garbage collect logically removed nodes in the bucket */
804 index
= hash
& (size
- 1);
805 order
= get_count_order_ulong(index
+ 1);
806 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
807 dummy_node
= (struct cds_lfht_node
*) lookup
;
808 _cds_lfht_gc_bucket(dummy_node
, node
);
813 int _cds_lfht_remove(struct cds_lfht
*ht
, unsigned long size
,
814 struct cds_lfht_node
*node
,
817 struct cds_lfht_node
*dummy
, *next
, *old
;
818 struct _cds_lfht_node
*lookup
;
820 unsigned long hash
, index
, order
;
822 /* logically delete the node */
823 assert(!is_dummy(node
));
824 assert(!is_removed(node
));
825 old
= rcu_dereference(node
->p
.next
);
828 if (unlikely(is_removed(next
)))
831 assert(is_dummy(next
));
833 assert(!is_dummy(next
));
834 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
836 } while (old
!= next
);
838 /* We performed the (logical) deletion. */
842 * Ensure that the node is not visible to readers anymore: lookup for
843 * the node, and remove it (along with any other logically removed node)
846 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
848 index
= hash
& (size
- 1);
849 order
= get_count_order_ulong(index
+ 1);
850 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
851 dummy
= (struct cds_lfht_node
*) lookup
;
852 _cds_lfht_gc_bucket(dummy
, node
);
855 * Only the flagging action indicated that we (and no other)
856 * removed the node from the hash.
859 assert(is_removed(rcu_dereference(node
->p
.next
)));
866 void init_table_hash(struct cds_lfht
*ht
, unsigned long i
,
871 for (j
= 0; j
< len
; j
++) {
872 struct cds_lfht_node
*new_node
=
873 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
875 dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
876 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
877 new_node
->p
.reverse_hash
=
878 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
879 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
885 void init_table_link(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
889 ht
->cds_lfht_rcu_thread_online();
890 ht
->cds_lfht_rcu_read_lock();
891 for (j
= 0; j
< len
; j
++) {
892 struct cds_lfht_node
*new_node
=
893 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
895 dbg_printf("init link: i %lu j %lu hash %lu\n",
896 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
897 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
899 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
902 ht
->cds_lfht_rcu_read_unlock();
903 ht
->cds_lfht_rcu_thread_offline();
907 * Holding RCU read lock to protect _cds_lfht_add against memory
908 * reclaim that could be performed by other call_rcu worker threads (ABA
912 void init_table(struct cds_lfht
*ht
,
913 unsigned long first_order
, unsigned long len_order
)
915 unsigned long i
, end_order
;
917 dbg_printf("init table: first_order %lu end_order %lu\n",
918 first_order
, first_order
+ len_order
);
919 end_order
= first_order
+ len_order
;
920 for (i
= first_order
; i
< end_order
; i
++) {
923 len
= !i
? 1 : 1UL << (i
- 1);
924 dbg_printf("init order %lu len: %lu\n", i
, len
);
925 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
926 + (len
* sizeof(struct _cds_lfht_node
)));
928 /* Set all dummy nodes reverse hash values for a level */
929 init_table_hash(ht
, i
, len
);
932 * Link all dummy nodes into the table. Concurrent
933 * add/remove are helping us.
935 init_table_link(ht
, i
, len
);
938 * Update table size (after init for now, because no
939 * concurrent updater help (TODO)).
941 cmm_smp_wmb(); /* populate data before RCU size */
942 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
943 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
944 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
950 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
954 ht
->cds_lfht_rcu_thread_online();
955 ht
->cds_lfht_rcu_read_lock();
956 for (j
= 0; j
< len
; j
++) {
957 struct cds_lfht_node
*fini_node
=
958 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
960 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
961 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
962 fini_node
->p
.reverse_hash
=
963 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
964 (void) _cds_lfht_remove(ht
, !i
? 0 : (1UL << (i
- 1)),
966 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
969 ht
->cds_lfht_rcu_read_unlock();
970 ht
->cds_lfht_rcu_thread_offline();
974 * Holding RCU read lock to protect _cds_lfht_remove against memory
975 * reclaim that could be performed by other call_rcu worker threads (ABA
979 void fini_table(struct cds_lfht
*ht
,
980 unsigned long first_order
, unsigned long len_order
)
984 dbg_printf("fini table: first_order %lu end_order %lu\n",
985 first_order
, first_order
+ len_order
);
986 end_order
= first_order
+ len_order
;
987 assert(first_order
> 0);
988 assert(ht
->t
.size
== (1UL << (first_order
- 1)));
989 for (i
= end_order
- 1; i
>= first_order
; i
--) {
992 len
= !i
? 1 : 1UL << (i
- 1);
993 dbg_printf("fini order %lu len: %lu\n", i
, len
);
996 * Set "removed" flag in dummy nodes about to be removed.
997 * Unlink all now-logically-removed dummy node pointers.
998 * Concurrent add/remove operation are helping us doing
1001 remove_table(ht
, i
, len
);
1003 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1005 dbg_printf("fini new size: %lu\n", 1UL << i
);
1006 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1011 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1012 cds_lfht_compare_fct compare_fct
,
1013 unsigned long hash_seed
,
1014 unsigned long init_size
,
1016 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1017 void (*func
)(struct rcu_head
*head
)),
1018 void (*cds_lfht_synchronize_rcu
)(void),
1019 void (*cds_lfht_rcu_read_lock
)(void),
1020 void (*cds_lfht_rcu_read_unlock
)(void),
1021 void (*cds_lfht_rcu_thread_offline
)(void),
1022 void (*cds_lfht_rcu_thread_online
)(void))
1024 struct cds_lfht
*ht
;
1025 unsigned long order
;
1027 /* init_size must be power of two */
1028 if (init_size
&& (init_size
& (init_size
- 1)))
1030 ht
= calloc(1, sizeof(struct cds_lfht
));
1031 ht
->hash_fct
= hash_fct
;
1032 ht
->compare_fct
= compare_fct
;
1033 ht
->hash_seed
= hash_seed
;
1034 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1035 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1036 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1037 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1038 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1039 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1040 ht
->percpu_count
= alloc_per_cpu_items_count();
1041 /* this mutex should not nest in read-side C.S. */
1042 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1043 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1045 ht
->cds_lfht_rcu_thread_offline();
1046 pthread_mutex_lock(&ht
->resize_mutex
);
1047 init_table(ht
, 0, order
);
1048 pthread_mutex_unlock(&ht
->resize_mutex
);
1049 ht
->cds_lfht_rcu_thread_online();
1053 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1055 struct cds_lfht_node
*node
, *next
;
1056 struct _cds_lfht_node
*lookup
;
1057 unsigned long hash
, reverse_hash
, index
, order
, size
;
1059 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1060 reverse_hash
= bit_reverse_ulong(hash
);
1062 size
= rcu_dereference(ht
->t
.size
);
1063 index
= hash
& (size
- 1);
1064 order
= get_count_order_ulong(index
+ 1);
1065 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1066 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1067 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1068 node
= (struct cds_lfht_node
*) lookup
;
1070 if (unlikely(!node
))
1072 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1076 next
= rcu_dereference(node
->p
.next
);
1077 if (likely(!is_removed(next
))
1079 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1082 node
= clear_flag(next
);
1084 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1088 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1089 struct cds_lfht_node
*node
)
1091 struct cds_lfht_node
*next
;
1092 unsigned long reverse_hash
;
1096 reverse_hash
= node
->p
.reverse_hash
;
1098 key_len
= node
->key_len
;
1099 next
= rcu_dereference(node
->p
.next
);
1100 node
= clear_flag(next
);
1103 if (unlikely(!node
))
1105 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1109 next
= rcu_dereference(node
->p
.next
);
1110 if (likely(!is_removed(next
))
1112 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1115 node
= clear_flag(next
);
1117 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1121 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1123 unsigned long hash
, size
;
1125 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1126 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1128 size
= rcu_dereference(ht
->t
.size
);
1129 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1130 ht_count_add(ht
, size
);
1133 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1134 struct cds_lfht_node
*node
)
1136 unsigned long hash
, size
;
1137 struct cds_lfht_node
*ret
;
1139 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1140 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1142 size
= rcu_dereference(ht
->t
.size
);
1143 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1145 ht_count_add(ht
, size
);
1149 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1154 size
= rcu_dereference(ht
->t
.size
);
1155 ret
= _cds_lfht_remove(ht
, size
, node
, 0);
1157 ht_count_remove(ht
, size
);
1162 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1164 struct cds_lfht_node
*node
;
1165 struct _cds_lfht_node
*lookup
;
1166 unsigned long order
, i
, size
;
1168 /* Check that the table is empty */
1169 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1170 node
= (struct cds_lfht_node
*) lookup
;
1172 node
= clear_flag(node
)->p
.next
;
1173 if (!is_dummy(node
))
1175 assert(!is_removed(node
));
1176 } while (clear_flag(node
));
1178 * size accessed without rcu_dereference because hash table is
1182 /* Internal sanity check: all nodes left should be dummy */
1183 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1186 len
= !order
? 1 : 1UL << (order
- 1);
1187 for (i
= 0; i
< len
; i
++) {
1188 dbg_printf("delete order %lu i %lu hash %lu\n",
1190 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1191 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1193 poison_free(ht
->t
.tbl
[order
]);
1199 * Should only be called when no more concurrent readers nor writers can
1200 * possibly access the table.
1202 int cds_lfht_destroy(struct cds_lfht
*ht
)
1206 /* Wait for in-flight resize operations to complete */
1207 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1208 while (uatomic_read(&ht
->in_progress_resize
))
1209 poll(NULL
, 0, 100); /* wait for 100ms */
1210 ret
= cds_lfht_delete_dummy(ht
);
1213 free_per_cpu_items_count(ht
->percpu_count
);
1218 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1219 unsigned long *count
,
1220 unsigned long *removed
)
1222 struct cds_lfht_node
*node
, *next
;
1223 struct _cds_lfht_node
*lookup
;
1224 unsigned long nr_dummy
= 0;
1229 /* Count non-dummy nodes in the table */
1230 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1231 node
= (struct cds_lfht_node
*) lookup
;
1233 next
= rcu_dereference(node
->p
.next
);
1234 if (is_removed(next
)) {
1235 assert(!is_dummy(next
));
1237 } else if (!is_dummy(next
))
1241 node
= clear_flag(next
);
1243 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1246 /* called with resize mutex held */
1248 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1249 unsigned long old_size
, unsigned long new_size
)
1251 unsigned long old_order
, new_order
;
1253 old_order
= get_count_order_ulong(old_size
) + 1;
1254 new_order
= get_count_order_ulong(new_size
) + 1;
1255 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1256 old_size
, old_order
, new_size
, new_order
);
1257 assert(new_size
> old_size
);
1258 init_table(ht
, old_order
, new_order
- old_order
);
1261 /* called with resize mutex held */
1263 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1264 unsigned long old_size
, unsigned long new_size
)
1266 unsigned long old_order
, new_order
;
1268 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1269 old_order
= get_count_order_ulong(old_size
) + 1;
1270 new_order
= get_count_order_ulong(new_size
) + 1;
1271 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1272 old_size
, old_order
, new_size
, new_order
);
1273 assert(new_size
< old_size
);
1275 cmm_smp_wmb(); /* populate data before RCU size */
1276 CMM_STORE_SHARED(ht
->t
.size
, new_size
);
1279 * We need to wait for all add operations to reach Q.S. (and
1280 * thus use the new table for lookups) before we can start
1281 * releasing the old dummy nodes. Otherwise their lookup will
1282 * return a logically removed node as insert position.
1284 ht
->cds_lfht_synchronize_rcu();
1286 /* Remove and unlink all dummy nodes to remove. */
1287 fini_table(ht
, new_order
, old_order
- new_order
);
1291 /* called with resize mutex held */
1293 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1295 unsigned long new_size
, old_size
;
1298 * Resize table, re-do if the target size has changed under us.
1301 ht
->t
.resize_initiated
= 1;
1302 old_size
= ht
->t
.size
;
1303 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1304 if (old_size
< new_size
)
1305 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1306 else if (old_size
> new_size
)
1307 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1308 ht
->t
.resize_initiated
= 0;
1309 /* write resize_initiated before read resize_target */
1311 } while (new_size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1315 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1318 return _uatomic_max(&ht
->t
.resize_target
,
1319 size
<< growth_order
);
1323 void resize_target_update_count(struct cds_lfht
*ht
,
1324 unsigned long count
)
1326 count
= max(count
, MIN_TABLE_SIZE
);
1327 uatomic_set(&ht
->t
.resize_target
, count
);
1330 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1332 resize_target_update_count(ht
, new_size
);
1333 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1334 ht
->cds_lfht_rcu_thread_offline();
1335 pthread_mutex_lock(&ht
->resize_mutex
);
1336 _do_cds_lfht_resize(ht
);
1337 pthread_mutex_unlock(&ht
->resize_mutex
);
1338 ht
->cds_lfht_rcu_thread_online();
1342 void do_resize_cb(struct rcu_head
*head
)
1344 struct rcu_resize_work
*work
=
1345 caa_container_of(head
, struct rcu_resize_work
, head
);
1346 struct cds_lfht
*ht
= work
->ht
;
1348 ht
->cds_lfht_rcu_thread_offline();
1349 pthread_mutex_lock(&ht
->resize_mutex
);
1350 _do_cds_lfht_resize(ht
);
1351 pthread_mutex_unlock(&ht
->resize_mutex
);
1352 ht
->cds_lfht_rcu_thread_online();
1354 cmm_smp_mb(); /* finish resize before decrement */
1355 uatomic_dec(&ht
->in_progress_resize
);
1359 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1361 struct rcu_resize_work
*work
;
1362 unsigned long target_size
;
1364 target_size
= resize_target_update(ht
, size
, growth
);
1365 /* Store resize_target before read resize_initiated */
1367 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1368 uatomic_inc(&ht
->in_progress_resize
);
1369 cmm_smp_mb(); /* increment resize count before calling it */
1370 work
= malloc(sizeof(*work
));
1372 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1373 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1377 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1380 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1381 unsigned long count
)
1383 struct rcu_resize_work
*work
;
1385 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1387 resize_target_update_count(ht
, count
);
1388 /* Store resize_target before read resize_initiated */
1390 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1391 uatomic_inc(&ht
->in_progress_resize
);
1392 cmm_smp_mb(); /* increment resize count before calling it */
1393 work
= malloc(sizeof(*work
));
1395 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1396 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);