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 1
174 #if (CAA_BITS_PER_LONG == 32)
175 #define MAX_TABLE_ORDER 32
177 #define MAX_TABLE_ORDER 64
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
183 #define MIN_PARTITION_PER_THREAD 4096
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 /* Value of the end pointer. Should not interact with flags. */
203 #define END_VALUE NULL
205 struct ht_items_count
{
206 unsigned long add
, remove
;
207 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
210 struct rcu_head head
;
211 struct _cds_lfht_node nodes
[0];
215 unsigned long size
; /* always a power of 2, shared (RCU) */
216 unsigned long resize_target
;
217 int resize_initiated
;
218 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
223 cds_lfht_hash_fct hash_fct
;
224 cds_lfht_compare_fct compare_fct
;
225 unsigned long hash_seed
;
228 * We need to put the work threads offline (QSBR) when taking this
229 * mutex, because we use synchronize_rcu within this mutex critical
230 * section, which waits on read-side critical sections, and could
231 * therefore cause grace-period deadlock if we hold off RCU G.P.
234 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
235 unsigned int in_progress_resize
, in_progress_destroy
;
236 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
237 void (*func
)(struct rcu_head
*head
));
238 void (*cds_lfht_synchronize_rcu
)(void);
239 void (*cds_lfht_rcu_read_lock
)(void);
240 void (*cds_lfht_rcu_read_unlock
)(void);
241 void (*cds_lfht_rcu_thread_offline
)(void);
242 void (*cds_lfht_rcu_thread_online
)(void);
243 void (*cds_lfht_rcu_register_thread
)(void);
244 void (*cds_lfht_rcu_unregister_thread
)(void);
245 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
246 unsigned long count
; /* global approximate item count */
247 struct ht_items_count
*percpu_count
; /* per-cpu item count */
250 struct rcu_resize_work
{
251 struct rcu_head head
;
255 struct partition_resize_work
{
256 struct rcu_head head
;
258 unsigned long i
, start
, len
;
259 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
260 unsigned long start
, unsigned long len
);
264 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
266 struct cds_lfht_node
*node
,
267 int unique
, int dummy
);
270 * Algorithm to reverse bits in a word by lookup table, extended to
273 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
274 * Originally from Public Domain.
277 static const uint8_t BitReverseTable256
[256] =
279 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
280 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
281 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
282 R6(0), R6(2), R6(1), R6(3)
289 uint8_t bit_reverse_u8(uint8_t v
)
291 return BitReverseTable256
[v
];
294 static __attribute__((unused
))
295 uint32_t bit_reverse_u32(uint32_t v
)
297 return ((uint32_t) bit_reverse_u8(v
) << 24) |
298 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
299 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
300 ((uint32_t) bit_reverse_u8(v
>> 24));
303 static __attribute__((unused
))
304 uint64_t bit_reverse_u64(uint64_t v
)
306 return ((uint64_t) bit_reverse_u8(v
) << 56) |
307 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
308 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
309 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
310 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
311 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
312 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
313 ((uint64_t) bit_reverse_u8(v
>> 56));
317 unsigned long bit_reverse_ulong(unsigned long v
)
319 #if (CAA_BITS_PER_LONG == 32)
320 return bit_reverse_u32(v
);
322 return bit_reverse_u64(v
);
327 * fls: returns the position of the most significant bit.
328 * Returns 0 if no bit is set, else returns the position of the most
329 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
331 #if defined(__i386) || defined(__x86_64)
333 unsigned int fls_u32(uint32_t x
)
341 : "=r" (r
) : "rm" (x
));
347 #if defined(__x86_64)
349 unsigned int fls_u64(uint64_t x
)
357 : "=r" (r
) : "rm" (x
));
364 static __attribute__((unused
))
365 unsigned int fls_u64(uint64_t x
)
372 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
376 if (!(x
& 0xFFFF000000000000ULL
)) {
380 if (!(x
& 0xFF00000000000000ULL
)) {
384 if (!(x
& 0xF000000000000000ULL
)) {
388 if (!(x
& 0xC000000000000000ULL
)) {
392 if (!(x
& 0x8000000000000000ULL
)) {
401 static __attribute__((unused
))
402 unsigned int fls_u32(uint32_t x
)
408 if (!(x
& 0xFFFF0000U
)) {
412 if (!(x
& 0xFF000000U
)) {
416 if (!(x
& 0xF0000000U
)) {
420 if (!(x
& 0xC0000000U
)) {
424 if (!(x
& 0x80000000U
)) {
432 unsigned int fls_ulong(unsigned long x
)
434 #if (CAA_BITS_PER_lONG == 32)
441 int get_count_order_u32(uint32_t x
)
445 order
= fls_u32(x
) - 1;
451 int get_count_order_ulong(unsigned long x
)
455 order
= fls_ulong(x
) - 1;
462 #define poison_free(ptr) \
464 memset(ptr, 0x42, sizeof(*(ptr))); \
468 #define poison_free(ptr) free(ptr)
472 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
475 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
476 * available, then we support hash table item accounting.
477 * In the unfortunate event the number of CPUs reported would be
478 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
480 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
483 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
484 unsigned long count
);
486 static long nr_cpus_mask
= -1;
489 struct ht_items_count
*alloc_per_cpu_items_count(void)
491 struct ht_items_count
*count
;
493 switch (nr_cpus_mask
) {
500 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
506 * round up number of CPUs to next power of two, so we
507 * can use & for modulo.
509 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
510 nr_cpus_mask
= maxcpus
- 1;
514 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
519 void free_per_cpu_items_count(struct ht_items_count
*count
)
529 assert(nr_cpus_mask
>= 0);
530 cpu
= sched_getcpu();
531 if (unlikely(cpu
< 0))
534 return cpu
& nr_cpus_mask
;
538 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
540 unsigned long percpu_count
;
543 if (unlikely(!ht
->percpu_count
))
546 if (unlikely(cpu
< 0))
548 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
549 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
552 dbg_printf("add percpu %lu\n", percpu_count
);
553 count
= uatomic_add_return(&ht
->count
,
554 1UL << COUNT_COMMIT_ORDER
);
556 if (!(count
& (count
- 1))) {
557 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
559 dbg_printf("add set global %lu\n", count
);
560 cds_lfht_resize_lazy_count(ht
, size
,
561 count
>> (CHAIN_LEN_TARGET
- 1));
567 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
569 unsigned long percpu_count
;
572 if (unlikely(!ht
->percpu_count
))
575 if (unlikely(cpu
< 0))
577 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
578 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
581 dbg_printf("remove percpu %lu\n", percpu_count
);
582 count
= uatomic_add_return(&ht
->count
,
583 -(1UL << COUNT_COMMIT_ORDER
));
585 if (!(count
& (count
- 1))) {
586 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
588 dbg_printf("remove set global %lu\n", count
);
589 cds_lfht_resize_lazy_count(ht
, size
,
590 count
>> (CHAIN_LEN_TARGET
- 1));
595 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
597 static const long nr_cpus_mask
= -1;
600 struct ht_items_count
*alloc_per_cpu_items_count(void)
606 void free_per_cpu_items_count(struct ht_items_count
*count
)
611 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
616 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
620 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
624 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
628 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
630 count
= uatomic_read(&ht
->count
);
632 * Use bucket-local length for small table expand and for
633 * environments lacking per-cpu data support.
635 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
638 dbg_printf("WARNING: large chain length: %u.\n",
640 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
641 cds_lfht_resize_lazy(ht
, size
,
642 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
646 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
648 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
652 int is_removed(struct cds_lfht_node
*node
)
654 return ((unsigned long) node
) & REMOVED_FLAG
;
658 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
660 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
664 int is_dummy(struct cds_lfht_node
*node
)
666 return ((unsigned long) node
) & DUMMY_FLAG
;
670 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
672 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
676 struct cds_lfht_node
*get_end(void)
678 return (struct cds_lfht_node
*) END_VALUE
;
682 int is_end(struct cds_lfht_node
*node
)
684 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
688 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
690 unsigned long old1
, old2
;
692 old1
= uatomic_read(ptr
);
697 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
702 void cds_lfht_free_level(struct rcu_head
*head
)
704 struct rcu_level
*l
=
705 caa_container_of(head
, struct rcu_level
, head
);
710 * Remove all logically deleted nodes from a bucket up to a certain node key.
713 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
715 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
717 assert(!is_dummy(dummy
));
718 assert(!is_removed(dummy
));
719 assert(!is_dummy(node
));
720 assert(!is_removed(node
));
723 /* We can always skip the dummy node initially */
724 iter
= rcu_dereference(iter_prev
->p
.next
);
725 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
727 * We should never be called with dummy (start of chain)
728 * and logically removed node (end of path compression
729 * marker) being the actual same node. This would be a
730 * bug in the algorithm implementation.
732 assert(dummy
!= node
);
734 if (unlikely(is_end(iter
)))
736 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
738 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
739 if (likely(is_removed(next
)))
741 iter_prev
= clear_flag(iter
);
744 assert(!is_removed(iter
));
746 new_next
= flag_dummy(clear_flag(next
));
748 new_next
= clear_flag(next
);
749 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
755 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
757 struct cds_lfht_node
*node
,
758 int unique
, int dummy
)
760 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
762 struct _cds_lfht_node
*lookup
;
763 unsigned long hash
, index
, order
;
765 assert(!is_dummy(node
));
766 assert(!is_removed(node
));
769 node
->p
.next
= flag_dummy(get_end());
770 return node
; /* Initial first add (head) */
772 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
774 uint32_t chain_len
= 0;
777 * iter_prev points to the non-removed node prior to the
780 index
= hash
& (size
- 1);
781 order
= get_count_order_ulong(index
+ 1);
782 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
783 iter_prev
= (struct cds_lfht_node
*) lookup
;
784 /* We can always skip the dummy node initially */
785 iter
= rcu_dereference(iter_prev
->p
.next
);
786 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
788 if (unlikely(is_end(iter
)))
790 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
792 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
793 if (unlikely(is_removed(next
)))
797 && !ht
->compare_fct(node
->key
, node
->key_len
,
798 clear_flag(iter
)->key
,
799 clear_flag(iter
)->key_len
))
800 return clear_flag(iter
);
801 /* Only account for identical reverse hash once */
802 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
804 check_resize(ht
, size
, ++chain_len
);
805 iter_prev
= clear_flag(iter
);
809 assert(node
!= clear_flag(iter
));
810 assert(!is_removed(iter_prev
));
811 assert(!is_removed(iter
));
812 assert(iter_prev
!= node
);
814 node
->p
.next
= clear_flag(iter
);
816 node
->p
.next
= flag_dummy(clear_flag(iter
));
818 new_node
= flag_dummy(node
);
821 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
823 continue; /* retry */
827 assert(!is_removed(iter
));
829 new_next
= flag_dummy(clear_flag(next
));
831 new_next
= clear_flag(next
);
832 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
836 /* Garbage collect logically removed nodes in the bucket */
837 index
= hash
& (size
- 1);
838 order
= get_count_order_ulong(index
+ 1);
839 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
840 dummy_node
= (struct cds_lfht_node
*) lookup
;
841 _cds_lfht_gc_bucket(dummy_node
, node
);
846 int _cds_lfht_remove(struct cds_lfht
*ht
, unsigned long size
,
847 struct cds_lfht_node
*node
,
850 struct cds_lfht_node
*dummy
, *next
, *old
;
851 struct _cds_lfht_node
*lookup
;
853 unsigned long hash
, index
, order
;
855 /* logically delete the node */
856 assert(!is_dummy(node
));
857 assert(!is_removed(node
));
858 old
= rcu_dereference(node
->p
.next
);
861 if (unlikely(is_removed(next
)))
864 assert(is_dummy(next
));
866 assert(!is_dummy(next
));
867 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
869 } while (old
!= next
);
871 /* We performed the (logical) deletion. */
875 * Ensure that the node is not visible to readers anymore: lookup for
876 * 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 _cds_lfht_gc_bucket(dummy
, node
);
888 * Only the flagging action indicated that we (and no other)
889 * removed the node from the hash.
892 assert(is_removed(rcu_dereference(node
->p
.next
)));
899 void *partition_resize_thread(void *arg
)
901 struct partition_resize_work
*work
= arg
;
903 work
->ht
->cds_lfht_rcu_register_thread();
904 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
905 work
->ht
->cds_lfht_rcu_unregister_thread();
910 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
912 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
913 unsigned long start
, unsigned long len
))
915 unsigned long partition_len
;
916 struct partition_resize_work
*work
;
918 pthread_t
*thread_id
;
920 /* Note: nr_cpus_mask + 1 is always power of 2 */
921 partition_len
= len
>> get_count_order_ulong(nr_cpus_mask
+ 1);
922 work
= calloc(nr_cpus_mask
+ 1, sizeof(*work
));
923 thread_id
= calloc(nr_cpus_mask
+ 1, sizeof(*thread_id
));
925 for (cpu
= 0; cpu
< nr_cpus_mask
+ 1; cpu
++) {
928 work
[cpu
].len
= partition_len
;
929 work
[cpu
].start
= cpu
* partition_len
;
931 ret
= pthread_create(&thread_id
[cpu
], ht
->resize_attr
,
932 partition_resize_thread
, &work
[cpu
]);
935 for (cpu
= 0; cpu
< nr_cpus_mask
+ 1; cpu
++) {
936 ret
= pthread_join(thread_id
[cpu
], NULL
);
944 * Holding RCU read lock to protect _cds_lfht_add against memory
945 * reclaim that could be performed by other call_rcu worker threads (ABA
948 * When we reach a certain length, we can split this population phase over
949 * many worker threads, based on the number of CPUs available in the system.
950 * This should therefore take care of not having the expand lagging behind too
951 * many concurrent insertion threads by using the scheduler's ability to
952 * schedule dummy node population fairly with insertions.
955 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
956 unsigned long start
, unsigned long len
)
960 ht
->cds_lfht_rcu_read_lock();
961 for (j
= start
; j
< start
+ len
; j
++) {
962 struct cds_lfht_node
*new_node
=
963 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
965 dbg_printf("init populate: i %lu j %lu hash %lu\n",
966 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
967 new_node
->p
.reverse_hash
=
968 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
969 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
971 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
974 ht
->cds_lfht_rcu_read_unlock();
978 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
981 assert(nr_cpus_mask
!= -1);
982 if (nr_cpus_mask
< 0 || len
< (nr_cpus_mask
+ 1) * MIN_PARTITION_PER_THREAD
) {
983 ht
->cds_lfht_rcu_thread_online();
984 init_table_populate_partition(ht
, i
, 0, len
);
985 ht
->cds_lfht_rcu_thread_offline();
988 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
992 void init_table(struct cds_lfht
*ht
,
993 unsigned long first_order
, unsigned long len_order
)
995 unsigned long i
, end_order
;
997 dbg_printf("init table: first_order %lu end_order %lu\n",
998 first_order
, first_order
+ len_order
);
999 end_order
= first_order
+ len_order
;
1000 for (i
= first_order
; i
< end_order
; i
++) {
1003 len
= !i
? 1 : 1UL << (i
- 1);
1004 dbg_printf("init order %lu len: %lu\n", i
, len
);
1006 /* Stop expand if the resize target changes under us */
1007 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1010 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1011 + (len
* sizeof(struct _cds_lfht_node
)));
1012 assert(ht
->t
.tbl
[i
]);
1015 * Set all dummy nodes reverse hash values for a level and
1016 * link all dummy nodes into the table.
1018 init_table_populate(ht
, i
, len
);
1021 * Update table size.
1023 cmm_smp_wmb(); /* populate data before RCU size */
1024 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1026 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1027 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1033 * Holding RCU read lock to protect _cds_lfht_remove against memory
1034 * reclaim that could be performed by other call_rcu worker threads (ABA
1036 * For a single level, we logically remove and garbage collect each node.
1038 * As a design choice, we perform logical removal and garbage collection on a
1039 * node-per-node basis to simplify this algorithm. We also assume keeping good
1040 * cache locality of the operation would overweight possible performance gain
1041 * that could be achieved by batching garbage collection for multiple levels.
1042 * However, this would have to be justified by benchmarks.
1044 * Concurrent removal and add operations are helping us perform garbage
1045 * collection of logically removed nodes. We guarantee that all logically
1046 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1047 * invoked to free a hole level of dummy nodes (after a grace period).
1049 * Logical removal and garbage collection can therefore be done in batch or on a
1050 * node-per-node basis, as long as the guarantee above holds.
1052 * When we reach a certain length, we can split this removal over many worker
1053 * threads, based on the number of CPUs available in the system. This should
1054 * take care of not letting resize process lag behind too many concurrent
1055 * updater threads actively inserting into the hash table.
1058 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1059 unsigned long start
, unsigned long len
)
1063 ht
->cds_lfht_rcu_read_lock();
1064 for (j
= start
; j
< start
+ len
; j
++) {
1065 struct cds_lfht_node
*fini_node
=
1066 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1068 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1069 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1070 fini_node
->p
.reverse_hash
=
1071 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1072 (void) _cds_lfht_remove(ht
, !i
? 0 : (1UL << (i
- 1)),
1074 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1077 ht
->cds_lfht_rcu_read_unlock();
1081 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1084 assert(nr_cpus_mask
!= -1);
1085 if (nr_cpus_mask
< 0 || len
< (nr_cpus_mask
+ 1) * MIN_PARTITION_PER_THREAD
) {
1086 ht
->cds_lfht_rcu_thread_online();
1087 remove_table_partition(ht
, i
, 0, len
);
1088 ht
->cds_lfht_rcu_thread_offline();
1091 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1095 void fini_table(struct cds_lfht
*ht
,
1096 unsigned long first_order
, unsigned long len_order
)
1100 dbg_printf("fini table: first_order %lu end_order %lu\n",
1101 first_order
, first_order
+ len_order
);
1102 end_order
= first_order
+ len_order
;
1103 assert(first_order
> 0);
1104 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1107 len
= !i
? 1 : 1UL << (i
- 1);
1108 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1110 /* Stop shrink if the resize target changes under us */
1111 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1114 cmm_smp_wmb(); /* populate data before RCU size */
1115 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1118 * We need to wait for all add operations to reach Q.S. (and
1119 * thus use the new table for lookups) before we can start
1120 * releasing the old dummy nodes. Otherwise their lookup will
1121 * return a logically removed node as insert position.
1123 ht
->cds_lfht_synchronize_rcu();
1126 * Set "removed" flag in dummy nodes about to be removed.
1127 * Unlink all now-logically-removed dummy node pointers.
1128 * Concurrent add/remove operation are helping us doing
1131 remove_table(ht
, i
, len
);
1133 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1135 dbg_printf("fini new size: %lu\n", 1UL << i
);
1136 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1141 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1142 cds_lfht_compare_fct compare_fct
,
1143 unsigned long hash_seed
,
1144 unsigned long init_size
,
1146 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1147 void (*func
)(struct rcu_head
*head
)),
1148 void (*cds_lfht_synchronize_rcu
)(void),
1149 void (*cds_lfht_rcu_read_lock
)(void),
1150 void (*cds_lfht_rcu_read_unlock
)(void),
1151 void (*cds_lfht_rcu_thread_offline
)(void),
1152 void (*cds_lfht_rcu_thread_online
)(void),
1153 void (*cds_lfht_rcu_register_thread
)(void),
1154 void (*cds_lfht_rcu_unregister_thread
)(void),
1155 pthread_attr_t
*attr
)
1157 struct cds_lfht
*ht
;
1158 unsigned long order
;
1160 /* init_size must be power of two */
1161 if (init_size
&& (init_size
& (init_size
- 1)))
1163 ht
= calloc(1, sizeof(struct cds_lfht
));
1165 ht
->hash_fct
= hash_fct
;
1166 ht
->compare_fct
= compare_fct
;
1167 ht
->hash_seed
= hash_seed
;
1168 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1169 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1170 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1171 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1172 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1173 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1174 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1175 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1176 ht
->resize_attr
= attr
;
1177 ht
->percpu_count
= alloc_per_cpu_items_count();
1178 /* this mutex should not nest in read-side C.S. */
1179 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1180 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1182 ht
->cds_lfht_rcu_thread_offline();
1183 pthread_mutex_lock(&ht
->resize_mutex
);
1184 ht
->t
.resize_target
= 1UL << (order
- 1);
1185 init_table(ht
, 0, order
);
1186 pthread_mutex_unlock(&ht
->resize_mutex
);
1187 ht
->cds_lfht_rcu_thread_online();
1191 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1193 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1194 struct _cds_lfht_node
*lookup
;
1195 unsigned long hash
, reverse_hash
, index
, order
, size
;
1197 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1198 reverse_hash
= bit_reverse_ulong(hash
);
1200 size
= rcu_dereference(ht
->t
.size
);
1201 index
= hash
& (size
- 1);
1202 order
= get_count_order_ulong(index
+ 1);
1203 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1204 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1205 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1206 dummy_node
= (struct cds_lfht_node
*) lookup
;
1207 /* We can always skip the dummy node initially */
1208 node
= rcu_dereference(dummy_node
->p
.next
);
1209 node
= clear_flag(node
);
1211 if (unlikely(is_end(node
))) {
1215 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1219 next
= rcu_dereference(node
->p
.next
);
1220 if (likely(!is_removed(next
))
1222 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1225 node
= clear_flag(next
);
1227 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1231 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1232 struct cds_lfht_node
*node
)
1234 struct cds_lfht_node
*next
;
1235 unsigned long reverse_hash
;
1239 reverse_hash
= node
->p
.reverse_hash
;
1241 key_len
= node
->key_len
;
1242 next
= rcu_dereference(node
->p
.next
);
1243 node
= clear_flag(next
);
1246 if (unlikely(is_end(node
))) {
1250 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1254 next
= rcu_dereference(node
->p
.next
);
1255 if (likely(!is_removed(next
))
1257 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1260 node
= clear_flag(next
);
1262 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1266 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1268 unsigned long hash
, size
;
1270 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1271 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1273 size
= rcu_dereference(ht
->t
.size
);
1274 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1275 ht_count_add(ht
, size
);
1278 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1279 struct cds_lfht_node
*node
)
1281 unsigned long hash
, size
;
1282 struct cds_lfht_node
*ret
;
1284 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1285 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1287 size
= rcu_dereference(ht
->t
.size
);
1288 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1290 ht_count_add(ht
, size
);
1294 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1299 size
= rcu_dereference(ht
->t
.size
);
1300 ret
= _cds_lfht_remove(ht
, size
, node
, 0);
1302 ht_count_remove(ht
, size
);
1307 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1309 struct cds_lfht_node
*node
;
1310 struct _cds_lfht_node
*lookup
;
1311 unsigned long order
, i
, size
;
1313 /* Check that the table is empty */
1314 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1315 node
= (struct cds_lfht_node
*) lookup
;
1317 node
= clear_flag(node
)->p
.next
;
1318 if (!is_dummy(node
))
1320 assert(!is_removed(node
));
1321 } while (!is_end(node
));
1323 * size accessed without rcu_dereference because hash table is
1327 /* Internal sanity check: all nodes left should be dummy */
1328 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1331 len
= !order
? 1 : 1UL << (order
- 1);
1332 for (i
= 0; i
< len
; i
++) {
1333 dbg_printf("delete order %lu i %lu hash %lu\n",
1335 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1336 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1338 poison_free(ht
->t
.tbl
[order
]);
1344 * Should only be called when no more concurrent readers nor writers can
1345 * possibly access the table.
1347 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1351 /* Wait for in-flight resize operations to complete */
1352 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1353 while (uatomic_read(&ht
->in_progress_resize
))
1354 poll(NULL
, 0, 100); /* wait for 100ms */
1355 ret
= cds_lfht_delete_dummy(ht
);
1358 free_per_cpu_items_count(ht
->percpu_count
);
1360 *attr
= ht
->resize_attr
;
1365 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1366 unsigned long *count
,
1367 unsigned long *removed
)
1369 struct cds_lfht_node
*node
, *next
;
1370 struct _cds_lfht_node
*lookup
;
1371 unsigned long nr_dummy
= 0;
1376 /* Count non-dummy nodes in the table */
1377 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1378 node
= (struct cds_lfht_node
*) lookup
;
1380 next
= rcu_dereference(node
->p
.next
);
1381 if (is_removed(next
)) {
1382 assert(!is_dummy(next
));
1384 } else if (!is_dummy(next
))
1388 node
= clear_flag(next
);
1389 } while (!is_end(node
));
1390 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1393 /* called with resize mutex held */
1395 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1396 unsigned long old_size
, unsigned long new_size
)
1398 unsigned long old_order
, new_order
;
1400 old_order
= get_count_order_ulong(old_size
) + 1;
1401 new_order
= get_count_order_ulong(new_size
) + 1;
1402 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1403 old_size
, old_order
, new_size
, new_order
);
1404 assert(new_size
> old_size
);
1405 init_table(ht
, old_order
, new_order
- old_order
);
1408 /* called with resize mutex held */
1410 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1411 unsigned long old_size
, unsigned long new_size
)
1413 unsigned long old_order
, new_order
;
1415 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1416 old_order
= get_count_order_ulong(old_size
) + 1;
1417 new_order
= get_count_order_ulong(new_size
) + 1;
1418 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1419 old_size
, old_order
, new_size
, new_order
);
1420 assert(new_size
< old_size
);
1422 /* Remove and unlink all dummy nodes to remove. */
1423 fini_table(ht
, new_order
, old_order
- new_order
);
1427 /* called with resize mutex held */
1429 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1431 unsigned long new_size
, old_size
;
1434 * Resize table, re-do if the target size has changed under us.
1437 ht
->t
.resize_initiated
= 1;
1438 old_size
= ht
->t
.size
;
1439 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1440 if (old_size
< new_size
)
1441 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1442 else if (old_size
> new_size
)
1443 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1444 ht
->t
.resize_initiated
= 0;
1445 /* write resize_initiated before read resize_target */
1447 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1451 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1454 return _uatomic_max(&ht
->t
.resize_target
,
1455 size
<< growth_order
);
1459 void resize_target_update_count(struct cds_lfht
*ht
,
1460 unsigned long count
)
1462 count
= max(count
, MIN_TABLE_SIZE
);
1463 uatomic_set(&ht
->t
.resize_target
, count
);
1466 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1468 resize_target_update_count(ht
, new_size
);
1469 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1470 ht
->cds_lfht_rcu_thread_offline();
1471 pthread_mutex_lock(&ht
->resize_mutex
);
1472 _do_cds_lfht_resize(ht
);
1473 pthread_mutex_unlock(&ht
->resize_mutex
);
1474 ht
->cds_lfht_rcu_thread_online();
1478 void do_resize_cb(struct rcu_head
*head
)
1480 struct rcu_resize_work
*work
=
1481 caa_container_of(head
, struct rcu_resize_work
, head
);
1482 struct cds_lfht
*ht
= work
->ht
;
1484 ht
->cds_lfht_rcu_thread_offline();
1485 pthread_mutex_lock(&ht
->resize_mutex
);
1486 _do_cds_lfht_resize(ht
);
1487 pthread_mutex_unlock(&ht
->resize_mutex
);
1488 ht
->cds_lfht_rcu_thread_online();
1490 cmm_smp_mb(); /* finish resize before decrement */
1491 uatomic_dec(&ht
->in_progress_resize
);
1495 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1497 struct rcu_resize_work
*work
;
1498 unsigned long target_size
;
1500 target_size
= resize_target_update(ht
, size
, growth
);
1501 /* Store resize_target before read resize_initiated */
1503 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1504 uatomic_inc(&ht
->in_progress_resize
);
1505 cmm_smp_mb(); /* increment resize count before calling it */
1506 work
= malloc(sizeof(*work
));
1508 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1509 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1513 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1516 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1517 unsigned long count
)
1519 struct rcu_resize_work
*work
;
1521 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1523 resize_target_update_count(ht
, count
);
1524 /* Store resize_target before read resize_initiated */
1526 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1527 uatomic_inc(&ht
->in_progress_resize
);
1528 cmm_smp_mb(); /* increment resize count before calling it */
1529 work
= malloc(sizeof(*work
));
1531 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1532 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);