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.
172 #define MIN_TABLE_SIZE 128
174 #if (CAA_BITS_PER_LONG == 32)
175 #define MAX_TABLE_ORDER 32
177 #define MAX_TABLE_ORDER 64
181 #define min(a, b) ((a) < (b) ? (a) : (b))
185 #define max(a, b) ((a) > (b) ? (a) : (b))
189 * The removed flag needs to be updated atomically with the pointer.
190 * The dummy flag does not require to be updated atomically with the
191 * pointer, but it is added as a pointer low bit flag to save space.
193 #define REMOVED_FLAG (1UL << 0)
194 #define DUMMY_FLAG (1UL << 1)
195 #define FLAGS_MASK ((1UL << 2) - 1)
197 struct ht_items_count
{
198 unsigned long add
, remove
;
199 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
202 struct rcu_head head
;
203 struct _cds_lfht_node nodes
[0];
207 unsigned long size
; /* always a power of 2, shared (RCU) */
208 unsigned long resize_target
;
209 int resize_initiated
;
210 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
215 cds_lfht_hash_fct hash_fct
;
216 cds_lfht_compare_fct compare_fct
;
217 unsigned long hash_seed
;
220 * We need to put the work threads offline (QSBR) when taking this
221 * mutex, because we use synchronize_rcu within this mutex critical
222 * section, which waits on read-side critical sections, and could
223 * therefore cause grace-period deadlock if we hold off RCU G.P.
226 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
227 unsigned int in_progress_resize
, in_progress_destroy
;
228 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
229 void (*func
)(struct rcu_head
*head
));
230 void (*cds_lfht_synchronize_rcu
)(void);
231 void (*cds_lfht_rcu_read_lock
)(void);
232 void (*cds_lfht_rcu_read_unlock
)(void);
233 void (*cds_lfht_rcu_thread_offline
)(void);
234 void (*cds_lfht_rcu_thread_online
)(void);
235 unsigned long count
; /* global approximate item count */
236 struct ht_items_count
*percpu_count
; /* per-cpu item count */
239 struct rcu_resize_work
{
240 struct rcu_head head
;
245 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
247 struct cds_lfht_node
*node
,
248 int unique
, int dummy
);
251 * Algorithm to reverse bits in a word by lookup table, extended to
254 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
255 * Originally from Public Domain.
258 static const uint8_t BitReverseTable256
[256] =
260 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
261 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
262 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
263 R6(0), R6(2), R6(1), R6(3)
270 uint8_t bit_reverse_u8(uint8_t v
)
272 return BitReverseTable256
[v
];
275 static __attribute__((unused
))
276 uint32_t bit_reverse_u32(uint32_t v
)
278 return ((uint32_t) bit_reverse_u8(v
) << 24) |
279 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
280 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
281 ((uint32_t) bit_reverse_u8(v
>> 24));
284 static __attribute__((unused
))
285 uint64_t bit_reverse_u64(uint64_t v
)
287 return ((uint64_t) bit_reverse_u8(v
) << 56) |
288 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
289 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
290 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
291 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
292 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
293 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
294 ((uint64_t) bit_reverse_u8(v
>> 56));
298 unsigned long bit_reverse_ulong(unsigned long v
)
300 #if (CAA_BITS_PER_LONG == 32)
301 return bit_reverse_u32(v
);
303 return bit_reverse_u64(v
);
308 * fls: returns the position of the most significant bit.
309 * Returns 0 if no bit is set, else returns the position of the most
310 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
312 #if defined(__i386) || defined(__x86_64)
314 unsigned int fls_u32(uint32_t x
)
322 : "=r" (r
) : "rm" (x
));
328 #if defined(__x86_64)
330 unsigned int fls_u64(uint64_t x
)
338 : "=r" (r
) : "rm" (x
));
345 static __attribute__((unused
))
346 unsigned int fls_u64(uint64_t x
)
353 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
357 if (!(x
& 0xFFFF000000000000ULL
)) {
361 if (!(x
& 0xFF00000000000000ULL
)) {
365 if (!(x
& 0xF000000000000000ULL
)) {
369 if (!(x
& 0xC000000000000000ULL
)) {
373 if (!(x
& 0x8000000000000000ULL
)) {
382 static __attribute__((unused
))
383 unsigned int fls_u32(uint32_t x
)
389 if (!(x
& 0xFFFF0000U
)) {
393 if (!(x
& 0xFF000000U
)) {
397 if (!(x
& 0xF0000000U
)) {
401 if (!(x
& 0xC0000000U
)) {
405 if (!(x
& 0x80000000U
)) {
413 unsigned int fls_ulong(unsigned long x
)
415 #if (CAA_BITS_PER_lONG == 32)
422 int get_count_order_u32(uint32_t x
)
426 order
= fls_u32(x
) - 1;
432 int get_count_order_ulong(unsigned long x
)
436 order
= fls_ulong(x
) - 1;
443 #define poison_free(ptr) \
445 memset(ptr, 0x42, sizeof(*(ptr))); \
449 #define poison_free(ptr) free(ptr)
453 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
456 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
457 * available, then we support hash table item accounting.
458 * In the unfortunate event the number of CPUs reported would be
459 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
461 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
464 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
465 unsigned long count
);
467 static long nr_cpus_mask
= -1;
470 struct ht_items_count
*alloc_per_cpu_items_count(void)
472 struct ht_items_count
*count
;
474 switch (nr_cpus_mask
) {
481 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
487 * round up number of CPUs to next power of two, so we
488 * can use & for modulo.
490 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
491 nr_cpus_mask
= maxcpus
- 1;
495 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
500 void free_per_cpu_items_count(struct ht_items_count
*count
)
510 assert(nr_cpus_mask
>= 0);
511 cpu
= sched_getcpu();
512 if (unlikely(cpu
< 0))
515 return cpu
& nr_cpus_mask
;
519 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
521 unsigned long percpu_count
;
524 if (unlikely(!ht
->percpu_count
))
527 if (unlikely(cpu
< 0))
529 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
530 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
533 dbg_printf("add percpu %lu\n", percpu_count
);
534 count
= uatomic_add_return(&ht
->count
,
535 1UL << COUNT_COMMIT_ORDER
);
537 if (!(count
& (count
- 1))) {
538 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
540 dbg_printf("add set global %lu\n", count
);
541 cds_lfht_resize_lazy_count(ht
, size
,
542 count
>> (CHAIN_LEN_TARGET
- 1));
548 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
550 unsigned long percpu_count
;
553 if (unlikely(!ht
->percpu_count
))
556 if (unlikely(cpu
< 0))
558 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
559 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
562 dbg_printf("remove percpu %lu\n", percpu_count
);
563 count
= uatomic_add_return(&ht
->count
,
564 -(1UL << COUNT_COMMIT_ORDER
));
566 if (!(count
& (count
- 1))) {
567 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
569 dbg_printf("remove set global %lu\n", count
);
570 cds_lfht_resize_lazy_count(ht
, size
,
571 count
>> (CHAIN_LEN_TARGET
- 1));
576 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
578 static const long nr_cpus_mask
= -1;
581 struct ht_items_count
*alloc_per_cpu_items_count(void)
587 void free_per_cpu_items_count(struct ht_items_count
*count
)
592 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
597 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
601 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
605 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
609 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
611 count
= uatomic_read(&ht
->count
);
613 * Use bucket-local length for small table expand and for
614 * environments lacking per-cpu data support.
616 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
619 dbg_printf("WARNING: large chain length: %u.\n",
621 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
622 cds_lfht_resize_lazy(ht
, size
,
623 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
627 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
629 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
633 int is_removed(struct cds_lfht_node
*node
)
635 return ((unsigned long) node
) & REMOVED_FLAG
;
639 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
641 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
645 int is_dummy(struct cds_lfht_node
*node
)
647 return ((unsigned long) node
) & DUMMY_FLAG
;
651 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
653 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
657 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
659 unsigned long old1
, old2
;
661 old1
= uatomic_read(ptr
);
666 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
671 void cds_lfht_free_level(struct rcu_head
*head
)
673 struct rcu_level
*l
=
674 caa_container_of(head
, struct rcu_level
, head
);
679 * Remove all logically deleted nodes from a bucket up to a certain node key.
682 int _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
684 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
686 assert(!is_dummy(dummy
));
687 assert(!is_removed(dummy
));
688 assert(!is_dummy(node
));
689 assert(!is_removed(node
));
692 /* We can always skip the dummy node initially */
693 iter
= rcu_dereference(iter_prev
->p
.next
);
694 if (unlikely(iter
== NULL
)) {
696 * We are executing concurrently with a hash table
697 * expand, so we see a dummy node with NULL next value.
698 * Help expand by linking this node into the list and
703 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
705 * We should never be called with dummy (start of chain)
706 * and logically removed node (end of path compression
707 * marker) being the actual same node. This would be a
708 * bug in the algorithm implementation.
710 assert(dummy
!= node
);
712 if (unlikely(!clear_flag(iter
)))
714 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
716 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
717 if (likely(is_removed(next
)))
719 iter_prev
= clear_flag(iter
);
722 assert(!is_removed(iter
));
724 new_next
= flag_dummy(clear_flag(next
));
726 new_next
= clear_flag(next
);
727 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
733 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
735 struct cds_lfht_node
*node
,
736 int unique
, int dummy
)
738 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
740 struct _cds_lfht_node
*lookup
;
741 unsigned long hash
, index
, order
;
743 assert(!is_dummy(node
));
744 assert(!is_removed(node
));
747 node
->p
.next
= flag_dummy(NULL
);
748 return node
; /* Initial first add (head) */
750 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
752 uint32_t chain_len
= 0;
755 * iter_prev points to the non-removed node prior to the
758 index
= hash
& (size
- 1);
759 order
= get_count_order_ulong(index
+ 1);
760 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
761 iter_prev
= (struct cds_lfht_node
*) lookup
;
762 /* We can always skip the dummy node initially */
763 iter
= rcu_dereference(iter_prev
->p
.next
);
764 if (unlikely(iter
== NULL
)) {
766 * We are executing concurrently with a hash table
767 * expand, so we see a dummy node with NULL next value.
768 * Help expand by linking this node into the list and
771 (void) _cds_lfht_add(ht
, size
>> 1, iter_prev
, 0, 1);
772 continue; /* retry */
774 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
777 * When adding a dummy node, we allow concurrent
778 * add/removal to help. If we find the dummy node in
779 * place, skip its insertion.
781 if (unlikely(dummy
&& clear_flag(iter
) == node
))
783 if (unlikely(!clear_flag(iter
)))
785 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
787 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
788 if (unlikely(is_removed(next
)))
792 && !ht
->compare_fct(node
->key
, node
->key_len
,
793 clear_flag(iter
)->key
,
794 clear_flag(iter
)->key_len
))
795 return clear_flag(iter
);
796 /* Only account for identical reverse hash once */
797 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
799 check_resize(ht
, size
, ++chain_len
);
800 iter_prev
= clear_flag(iter
);
804 assert(node
!= clear_flag(iter
));
805 assert(!is_removed(iter_prev
));
806 assert(!is_removed(iter
));
807 assert(iter_prev
!= node
);
809 node
->p
.next
= clear_flag(iter
);
811 node
->p
.next
= flag_dummy(clear_flag(iter
));
813 new_node
= flag_dummy(node
);
816 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
818 continue; /* retry */
822 assert(!is_removed(iter
));
824 new_next
= flag_dummy(clear_flag(next
));
826 new_next
= clear_flag(next
);
827 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
831 /* Garbage collect logically removed nodes in the bucket */
832 index
= hash
& (size
- 1);
833 order
= get_count_order_ulong(index
+ 1);
834 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
835 dummy_node
= (struct cds_lfht_node
*) lookup
;
836 if (_cds_lfht_gc_bucket(dummy_node
, node
)) {
838 (void) _cds_lfht_add(ht
, size
>> 1, dummy_node
, 0, 1);
839 goto gc_end
; /* retry */
845 int _cds_lfht_remove(struct cds_lfht
*ht
, unsigned long size
,
846 struct cds_lfht_node
*node
,
849 struct cds_lfht_node
*dummy
, *next
, *old
;
850 struct _cds_lfht_node
*lookup
;
852 unsigned long hash
, index
, order
;
854 /* logically delete the node */
855 assert(!is_dummy(node
));
856 assert(!is_removed(node
));
857 old
= rcu_dereference(node
->p
.next
);
860 if (unlikely(is_removed(next
)))
863 assert(is_dummy(next
));
865 assert(!is_dummy(next
));
866 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
868 } while (old
!= next
);
870 /* We performed the (logical) deletion. */
874 * Ensure that the node is not visible to readers anymore: lookup for
875 * the node, and remove it (along with any other logically removed node)
879 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
881 index
= hash
& (size
- 1);
882 order
= get_count_order_ulong(index
+ 1);
883 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
884 dummy
= (struct cds_lfht_node
*) lookup
;
885 if (_cds_lfht_gc_bucket(dummy
, node
)) {
887 (void) _cds_lfht_add(ht
, size
>> 1, dummy
, 0, 1);
888 goto gc_retry
; /* retry */
892 * Only the flagging action indicated that we (and no other)
893 * removed the node from the hash.
896 assert(is_removed(rcu_dereference(node
->p
.next
)));
903 void init_table_hash(struct cds_lfht
*ht
, unsigned long i
,
908 for (j
= 0; j
< len
; j
++) {
909 struct cds_lfht_node
*new_node
=
910 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
912 dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
913 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
914 new_node
->p
.reverse_hash
=
915 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
916 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
922 * Holding RCU read lock to protect _cds_lfht_add against memory
923 * reclaim that could be performed by other call_rcu worker threads (ABA
927 void init_table_link(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
931 ht
->cds_lfht_rcu_thread_online();
932 ht
->cds_lfht_rcu_read_lock();
933 for (j
= 0; j
< len
; j
++) {
934 struct cds_lfht_node
*new_node
=
935 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
937 dbg_printf("init link: i %lu j %lu hash %lu\n",
938 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
939 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
941 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
944 ht
->cds_lfht_rcu_read_unlock();
945 ht
->cds_lfht_rcu_thread_offline();
949 void init_table(struct cds_lfht
*ht
,
950 unsigned long first_order
, unsigned long len_order
)
952 unsigned long i
, end_order
;
954 dbg_printf("init table: first_order %lu end_order %lu\n",
955 first_order
, first_order
+ len_order
);
956 end_order
= first_order
+ len_order
;
957 for (i
= first_order
; i
< end_order
; i
++) {
960 len
= !i
? 1 : 1UL << (i
- 1);
961 dbg_printf("init order %lu len: %lu\n", i
, len
);
962 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
963 + (len
* sizeof(struct _cds_lfht_node
)));
965 /* Set all dummy nodes reverse hash values for a level */
966 init_table_hash(ht
, i
, len
);
969 * Update table size. At this point, concurrent add/remove see
970 * dummy nodes with correctly initialized reverse hash value,
971 * but with NULL next pointers. If they do, they can help us
972 * link the dummy nodes into the list and retry.
974 cmm_smp_wmb(); /* populate data before RCU size */
975 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
978 * Link all dummy nodes into the table. Concurrent
979 * add/remove are helping us.
981 init_table_link(ht
, i
, len
);
983 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
984 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
990 * Holding RCU read lock to protect _cds_lfht_remove against memory
991 * reclaim that could be performed by other call_rcu worker threads (ABA
993 * For a single level, we logically remove and garbage collect each node.
995 * As a design choice, we perform logical removal and garbage collection on a
996 * node-per-node basis to simplify this algorithm. We also assume keeping good
997 * cache locality of the operation would overweight possible performance gain
998 * that could be achieved by batching garbage collection for multiple levels.
999 * However, this would have to be justified by benchmarks.
1001 * Concurrent removal and add operations are helping us perform garbage
1002 * collection of logically removed nodes. We guarantee that all logically
1003 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1004 * invoked to free a hole level of dummy nodes (after a grace period).
1006 * Logical removal and garbage collection can therefore be done in batch or on a
1007 * node-per-node basis, as long as the guarantee above holds.
1010 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1014 ht
->cds_lfht_rcu_thread_online();
1015 ht
->cds_lfht_rcu_read_lock();
1016 for (j
= 0; j
< len
; j
++) {
1017 struct cds_lfht_node
*fini_node
=
1018 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1020 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1021 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1022 fini_node
->p
.reverse_hash
=
1023 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1024 (void) _cds_lfht_remove(ht
, !i
? 0 : (1UL << (i
- 1)),
1026 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1029 ht
->cds_lfht_rcu_read_unlock();
1030 ht
->cds_lfht_rcu_thread_offline();
1034 void fini_table(struct cds_lfht
*ht
,
1035 unsigned long first_order
, unsigned long len_order
)
1039 dbg_printf("fini table: first_order %lu end_order %lu\n",
1040 first_order
, first_order
+ len_order
);
1041 end_order
= first_order
+ len_order
;
1042 assert(first_order
> 0);
1043 assert(ht
->t
.size
== (1UL << (first_order
- 1)));
1044 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1047 len
= !i
? 1 : 1UL << (i
- 1);
1048 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1051 * Set "removed" flag in dummy nodes about to be removed.
1052 * Unlink all now-logically-removed dummy node pointers.
1053 * Concurrent add/remove operation are helping us doing
1056 remove_table(ht
, i
, len
);
1058 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1060 dbg_printf("fini new size: %lu\n", 1UL << i
);
1061 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1066 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1067 cds_lfht_compare_fct compare_fct
,
1068 unsigned long hash_seed
,
1069 unsigned long init_size
,
1071 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1072 void (*func
)(struct rcu_head
*head
)),
1073 void (*cds_lfht_synchronize_rcu
)(void),
1074 void (*cds_lfht_rcu_read_lock
)(void),
1075 void (*cds_lfht_rcu_read_unlock
)(void),
1076 void (*cds_lfht_rcu_thread_offline
)(void),
1077 void (*cds_lfht_rcu_thread_online
)(void))
1079 struct cds_lfht
*ht
;
1080 unsigned long order
;
1082 /* init_size must be power of two */
1083 if (init_size
&& (init_size
& (init_size
- 1)))
1085 ht
= calloc(1, sizeof(struct cds_lfht
));
1086 ht
->hash_fct
= hash_fct
;
1087 ht
->compare_fct
= compare_fct
;
1088 ht
->hash_seed
= hash_seed
;
1089 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1090 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1091 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1092 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1093 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1094 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1095 ht
->percpu_count
= alloc_per_cpu_items_count();
1096 /* this mutex should not nest in read-side C.S. */
1097 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1098 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1100 ht
->cds_lfht_rcu_thread_offline();
1101 pthread_mutex_lock(&ht
->resize_mutex
);
1102 init_table(ht
, 0, order
);
1103 pthread_mutex_unlock(&ht
->resize_mutex
);
1104 ht
->cds_lfht_rcu_thread_online();
1108 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1110 struct cds_lfht_node
*node
, *next
;
1111 struct _cds_lfht_node
*lookup
;
1112 unsigned long hash
, reverse_hash
, index
, order
, size
;
1114 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1115 reverse_hash
= bit_reverse_ulong(hash
);
1117 size
= rcu_dereference(ht
->t
.size
);
1118 index
= hash
& (size
- 1);
1119 order
= get_count_order_ulong(index
+ 1);
1120 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1121 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1122 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1123 node
= (struct cds_lfht_node
*) lookup
;
1125 if (unlikely(!node
))
1127 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1131 next
= rcu_dereference(node
->p
.next
);
1132 if (likely(!is_removed(next
))
1134 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1137 node
= clear_flag(next
);
1139 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1143 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1144 struct cds_lfht_node
*node
)
1146 struct cds_lfht_node
*next
;
1147 unsigned long reverse_hash
;
1151 reverse_hash
= node
->p
.reverse_hash
;
1153 key_len
= node
->key_len
;
1154 next
= rcu_dereference(node
->p
.next
);
1155 node
= clear_flag(next
);
1158 if (unlikely(!node
))
1160 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1164 next
= rcu_dereference(node
->p
.next
);
1165 if (likely(!is_removed(next
))
1167 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1170 node
= clear_flag(next
);
1172 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1176 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1178 unsigned long hash
, size
;
1180 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1181 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1183 size
= rcu_dereference(ht
->t
.size
);
1184 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1185 ht_count_add(ht
, size
);
1188 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1189 struct cds_lfht_node
*node
)
1191 unsigned long hash
, size
;
1192 struct cds_lfht_node
*ret
;
1194 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1195 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1197 size
= rcu_dereference(ht
->t
.size
);
1198 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1200 ht_count_add(ht
, size
);
1204 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1209 size
= rcu_dereference(ht
->t
.size
);
1210 ret
= _cds_lfht_remove(ht
, size
, node
, 0);
1212 ht_count_remove(ht
, size
);
1217 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1219 struct cds_lfht_node
*node
;
1220 struct _cds_lfht_node
*lookup
;
1221 unsigned long order
, i
, size
;
1223 /* Check that the table is empty */
1224 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1225 node
= (struct cds_lfht_node
*) lookup
;
1227 node
= clear_flag(node
)->p
.next
;
1228 if (!is_dummy(node
))
1230 assert(!is_removed(node
));
1231 } while (clear_flag(node
));
1233 * size accessed without rcu_dereference because hash table is
1237 /* Internal sanity check: all nodes left should be dummy */
1238 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1241 len
= !order
? 1 : 1UL << (order
- 1);
1242 for (i
= 0; i
< len
; i
++) {
1243 dbg_printf("delete order %lu i %lu hash %lu\n",
1245 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1246 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1248 poison_free(ht
->t
.tbl
[order
]);
1254 * Should only be called when no more concurrent readers nor writers can
1255 * possibly access the table.
1257 int cds_lfht_destroy(struct cds_lfht
*ht
)
1261 /* Wait for in-flight resize operations to complete */
1262 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1263 while (uatomic_read(&ht
->in_progress_resize
))
1264 poll(NULL
, 0, 100); /* wait for 100ms */
1265 ret
= cds_lfht_delete_dummy(ht
);
1268 free_per_cpu_items_count(ht
->percpu_count
);
1273 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1274 unsigned long *count
,
1275 unsigned long *removed
)
1277 struct cds_lfht_node
*node
, *next
;
1278 struct _cds_lfht_node
*lookup
;
1279 unsigned long nr_dummy
= 0;
1284 /* Count non-dummy nodes in the table */
1285 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1286 node
= (struct cds_lfht_node
*) lookup
;
1288 next
= rcu_dereference(node
->p
.next
);
1289 if (is_removed(next
)) {
1290 assert(!is_dummy(next
));
1292 } else if (!is_dummy(next
))
1296 node
= clear_flag(next
);
1298 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1301 /* called with resize mutex held */
1303 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1304 unsigned long old_size
, unsigned long new_size
)
1306 unsigned long old_order
, new_order
;
1308 old_order
= get_count_order_ulong(old_size
) + 1;
1309 new_order
= get_count_order_ulong(new_size
) + 1;
1310 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1311 old_size
, old_order
, new_size
, new_order
);
1312 assert(new_size
> old_size
);
1313 init_table(ht
, old_order
, new_order
- old_order
);
1316 /* called with resize mutex held */
1318 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1319 unsigned long old_size
, unsigned long new_size
)
1321 unsigned long old_order
, new_order
;
1323 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1324 old_order
= get_count_order_ulong(old_size
) + 1;
1325 new_order
= get_count_order_ulong(new_size
) + 1;
1326 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1327 old_size
, old_order
, new_size
, new_order
);
1328 assert(new_size
< old_size
);
1330 cmm_smp_wmb(); /* populate data before RCU size */
1331 CMM_STORE_SHARED(ht
->t
.size
, new_size
);
1334 * We need to wait for all add operations to reach Q.S. (and
1335 * thus use the new table for lookups) before we can start
1336 * releasing the old dummy nodes. Otherwise their lookup will
1337 * return a logically removed node as insert position.
1339 ht
->cds_lfht_synchronize_rcu();
1341 /* Remove and unlink all dummy nodes to remove. */
1342 fini_table(ht
, new_order
, old_order
- new_order
);
1346 /* called with resize mutex held */
1348 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1350 unsigned long new_size
, old_size
;
1353 * Resize table, re-do if the target size has changed under us.
1356 ht
->t
.resize_initiated
= 1;
1357 old_size
= ht
->t
.size
;
1358 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1359 if (old_size
< new_size
)
1360 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1361 else if (old_size
> new_size
)
1362 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1363 ht
->t
.resize_initiated
= 0;
1364 /* write resize_initiated before read resize_target */
1366 } while (new_size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1370 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1373 return _uatomic_max(&ht
->t
.resize_target
,
1374 size
<< growth_order
);
1378 void resize_target_update_count(struct cds_lfht
*ht
,
1379 unsigned long count
)
1381 count
= max(count
, MIN_TABLE_SIZE
);
1382 uatomic_set(&ht
->t
.resize_target
, count
);
1385 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1387 resize_target_update_count(ht
, new_size
);
1388 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1389 ht
->cds_lfht_rcu_thread_offline();
1390 pthread_mutex_lock(&ht
->resize_mutex
);
1391 _do_cds_lfht_resize(ht
);
1392 pthread_mutex_unlock(&ht
->resize_mutex
);
1393 ht
->cds_lfht_rcu_thread_online();
1397 void do_resize_cb(struct rcu_head
*head
)
1399 struct rcu_resize_work
*work
=
1400 caa_container_of(head
, struct rcu_resize_work
, head
);
1401 struct cds_lfht
*ht
= work
->ht
;
1403 ht
->cds_lfht_rcu_thread_offline();
1404 pthread_mutex_lock(&ht
->resize_mutex
);
1405 _do_cds_lfht_resize(ht
);
1406 pthread_mutex_unlock(&ht
->resize_mutex
);
1407 ht
->cds_lfht_rcu_thread_online();
1409 cmm_smp_mb(); /* finish resize before decrement */
1410 uatomic_dec(&ht
->in_progress_resize
);
1414 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1416 struct rcu_resize_work
*work
;
1417 unsigned long target_size
;
1419 target_size
= resize_target_update(ht
, size
, growth
);
1420 /* Store resize_target before read resize_initiated */
1422 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1423 uatomic_inc(&ht
->in_progress_resize
);
1424 cmm_smp_mb(); /* increment resize count before calling it */
1425 work
= malloc(sizeof(*work
));
1427 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1428 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1432 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1435 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1436 unsigned long count
)
1438 struct rcu_resize_work
*work
;
1440 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1442 resize_target_update_count(ht
, count
);
1443 /* Store resize_target before read resize_initiated */
1445 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1446 uatomic_inc(&ht
->in_progress_resize
);
1447 cmm_smp_mb(); /* increment resize count before calling it */
1448 work
= malloc(sizeof(*work
));
1450 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1451 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);