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 #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 /* Value of the end pointer. Should not interact with flags. */
198 #define END_VALUE NULL
200 struct ht_items_count
{
201 unsigned long add
, remove
;
202 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
205 struct rcu_head head
;
206 struct _cds_lfht_node nodes
[0];
210 unsigned long size
; /* always a power of 2, shared (RCU) */
211 unsigned long resize_target
;
212 int resize_initiated
;
213 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
218 cds_lfht_hash_fct hash_fct
;
219 cds_lfht_compare_fct compare_fct
;
220 unsigned long hash_seed
;
223 * We need to put the work threads offline (QSBR) when taking this
224 * mutex, because we use synchronize_rcu within this mutex critical
225 * section, which waits on read-side critical sections, and could
226 * therefore cause grace-period deadlock if we hold off RCU G.P.
229 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
230 unsigned int in_progress_resize
, in_progress_destroy
;
231 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
232 void (*func
)(struct rcu_head
*head
));
233 void (*cds_lfht_synchronize_rcu
)(void);
234 void (*cds_lfht_rcu_read_lock
)(void);
235 void (*cds_lfht_rcu_read_unlock
)(void);
236 void (*cds_lfht_rcu_thread_offline
)(void);
237 void (*cds_lfht_rcu_thread_online
)(void);
238 unsigned long count
; /* global approximate item count */
239 struct ht_items_count
*percpu_count
; /* per-cpu item count */
242 struct rcu_resize_work
{
243 struct rcu_head head
;
248 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
250 struct cds_lfht_node
*node
,
251 int unique
, int dummy
);
254 * Algorithm to reverse bits in a word by lookup table, extended to
257 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
258 * Originally from Public Domain.
261 static const uint8_t BitReverseTable256
[256] =
263 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
264 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
265 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
266 R6(0), R6(2), R6(1), R6(3)
273 uint8_t bit_reverse_u8(uint8_t v
)
275 return BitReverseTable256
[v
];
278 static __attribute__((unused
))
279 uint32_t bit_reverse_u32(uint32_t v
)
281 return ((uint32_t) bit_reverse_u8(v
) << 24) |
282 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
283 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
284 ((uint32_t) bit_reverse_u8(v
>> 24));
287 static __attribute__((unused
))
288 uint64_t bit_reverse_u64(uint64_t v
)
290 return ((uint64_t) bit_reverse_u8(v
) << 56) |
291 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
292 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
293 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
294 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
295 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
296 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
297 ((uint64_t) bit_reverse_u8(v
>> 56));
301 unsigned long bit_reverse_ulong(unsigned long v
)
303 #if (CAA_BITS_PER_LONG == 32)
304 return bit_reverse_u32(v
);
306 return bit_reverse_u64(v
);
311 * fls: returns the position of the most significant bit.
312 * Returns 0 if no bit is set, else returns the position of the most
313 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
315 #if defined(__i386) || defined(__x86_64)
317 unsigned int fls_u32(uint32_t x
)
325 : "=r" (r
) : "rm" (x
));
331 #if defined(__x86_64)
333 unsigned int fls_u64(uint64_t x
)
341 : "=r" (r
) : "rm" (x
));
348 static __attribute__((unused
))
349 unsigned int fls_u64(uint64_t x
)
356 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
360 if (!(x
& 0xFFFF000000000000ULL
)) {
364 if (!(x
& 0xFF00000000000000ULL
)) {
368 if (!(x
& 0xF000000000000000ULL
)) {
372 if (!(x
& 0xC000000000000000ULL
)) {
376 if (!(x
& 0x8000000000000000ULL
)) {
385 static __attribute__((unused
))
386 unsigned int fls_u32(uint32_t x
)
392 if (!(x
& 0xFFFF0000U
)) {
396 if (!(x
& 0xFF000000U
)) {
400 if (!(x
& 0xF0000000U
)) {
404 if (!(x
& 0xC0000000U
)) {
408 if (!(x
& 0x80000000U
)) {
416 unsigned int fls_ulong(unsigned long x
)
418 #if (CAA_BITS_PER_lONG == 32)
425 int get_count_order_u32(uint32_t x
)
429 order
= fls_u32(x
) - 1;
435 int get_count_order_ulong(unsigned long x
)
439 order
= fls_ulong(x
) - 1;
446 #define poison_free(ptr) \
448 memset(ptr, 0x42, sizeof(*(ptr))); \
452 #define poison_free(ptr) free(ptr)
456 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
459 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
460 * available, then we support hash table item accounting.
461 * In the unfortunate event the number of CPUs reported would be
462 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
464 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
467 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
468 unsigned long count
);
470 static long nr_cpus_mask
= -1;
473 struct ht_items_count
*alloc_per_cpu_items_count(void)
475 struct ht_items_count
*count
;
477 switch (nr_cpus_mask
) {
484 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
490 * round up number of CPUs to next power of two, so we
491 * can use & for modulo.
493 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
494 nr_cpus_mask
= maxcpus
- 1;
498 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
503 void free_per_cpu_items_count(struct ht_items_count
*count
)
513 assert(nr_cpus_mask
>= 0);
514 cpu
= sched_getcpu();
515 if (unlikely(cpu
< 0))
518 return cpu
& nr_cpus_mask
;
522 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
524 unsigned long percpu_count
;
527 if (unlikely(!ht
->percpu_count
))
530 if (unlikely(cpu
< 0))
532 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
533 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
536 dbg_printf("add percpu %lu\n", percpu_count
);
537 count
= uatomic_add_return(&ht
->count
,
538 1UL << COUNT_COMMIT_ORDER
);
540 if (!(count
& (count
- 1))) {
541 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
543 dbg_printf("add set global %lu\n", count
);
544 cds_lfht_resize_lazy_count(ht
, size
,
545 count
>> (CHAIN_LEN_TARGET
- 1));
551 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
553 unsigned long percpu_count
;
556 if (unlikely(!ht
->percpu_count
))
559 if (unlikely(cpu
< 0))
561 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
562 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
565 dbg_printf("remove percpu %lu\n", percpu_count
);
566 count
= uatomic_add_return(&ht
->count
,
567 -(1UL << COUNT_COMMIT_ORDER
));
569 if (!(count
& (count
- 1))) {
570 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
572 dbg_printf("remove set global %lu\n", count
);
573 cds_lfht_resize_lazy_count(ht
, size
,
574 count
>> (CHAIN_LEN_TARGET
- 1));
579 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
581 static const long nr_cpus_mask
= -1;
584 struct ht_items_count
*alloc_per_cpu_items_count(void)
590 void free_per_cpu_items_count(struct ht_items_count
*count
)
595 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
600 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
604 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
608 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
612 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
614 count
= uatomic_read(&ht
->count
);
616 * Use bucket-local length for small table expand and for
617 * environments lacking per-cpu data support.
619 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
622 dbg_printf("WARNING: large chain length: %u.\n",
624 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
625 cds_lfht_resize_lazy(ht
, size
,
626 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
630 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
632 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
636 int is_removed(struct cds_lfht_node
*node
)
638 return ((unsigned long) node
) & REMOVED_FLAG
;
642 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
644 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
648 int is_dummy(struct cds_lfht_node
*node
)
650 return ((unsigned long) node
) & DUMMY_FLAG
;
654 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
656 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
660 struct cds_lfht_node
*get_end(void)
662 return (struct cds_lfht_node
*) END_VALUE
;
666 int is_end(struct cds_lfht_node
*node
)
668 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
672 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
674 unsigned long old1
, old2
;
676 old1
= uatomic_read(ptr
);
681 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
686 void cds_lfht_free_level(struct rcu_head
*head
)
688 struct rcu_level
*l
=
689 caa_container_of(head
, struct rcu_level
, head
);
694 * Remove all logically deleted nodes from a bucket up to a certain node key.
697 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
699 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
701 assert(!is_dummy(dummy
));
702 assert(!is_removed(dummy
));
703 assert(!is_dummy(node
));
704 assert(!is_removed(node
));
707 /* We can always skip the dummy node initially */
708 iter
= rcu_dereference(iter_prev
->p
.next
);
709 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
711 * We should never be called with dummy (start of chain)
712 * and logically removed node (end of path compression
713 * marker) being the actual same node. This would be a
714 * bug in the algorithm implementation.
716 assert(dummy
!= node
);
718 if (unlikely(is_end(iter
)))
720 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
722 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
723 if (likely(is_removed(next
)))
725 iter_prev
= clear_flag(iter
);
728 assert(!is_removed(iter
));
730 new_next
= flag_dummy(clear_flag(next
));
732 new_next
= clear_flag(next
);
733 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
739 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
741 struct cds_lfht_node
*node
,
742 int unique
, int dummy
)
744 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
746 struct _cds_lfht_node
*lookup
;
747 unsigned long hash
, index
, order
;
749 assert(!is_dummy(node
));
750 assert(!is_removed(node
));
753 node
->p
.next
= flag_dummy(get_end());
754 return node
; /* Initial first add (head) */
756 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
758 uint32_t chain_len
= 0;
761 * iter_prev points to the non-removed node prior to the
764 index
= hash
& (size
- 1);
765 order
= get_count_order_ulong(index
+ 1);
766 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
767 iter_prev
= (struct cds_lfht_node
*) lookup
;
768 /* We can always skip the dummy node initially */
769 iter
= rcu_dereference(iter_prev
->p
.next
);
770 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
772 if (unlikely(is_end(iter
)))
774 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
776 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
777 if (unlikely(is_removed(next
)))
781 && !ht
->compare_fct(node
->key
, node
->key_len
,
782 clear_flag(iter
)->key
,
783 clear_flag(iter
)->key_len
))
784 return clear_flag(iter
);
785 /* Only account for identical reverse hash once */
786 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
788 check_resize(ht
, size
, ++chain_len
);
789 iter_prev
= clear_flag(iter
);
793 assert(node
!= clear_flag(iter
));
794 assert(!is_removed(iter_prev
));
795 assert(!is_removed(iter
));
796 assert(iter_prev
!= node
);
798 node
->p
.next
= clear_flag(iter
);
800 node
->p
.next
= flag_dummy(clear_flag(iter
));
802 new_node
= flag_dummy(node
);
805 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
807 continue; /* retry */
811 assert(!is_removed(iter
));
813 new_next
= flag_dummy(clear_flag(next
));
815 new_next
= clear_flag(next
);
816 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
820 /* Garbage collect logically removed nodes in the bucket */
821 index
= hash
& (size
- 1);
822 order
= get_count_order_ulong(index
+ 1);
823 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
824 dummy_node
= (struct cds_lfht_node
*) lookup
;
825 _cds_lfht_gc_bucket(dummy_node
, node
);
830 int _cds_lfht_remove(struct cds_lfht
*ht
, unsigned long size
,
831 struct cds_lfht_node
*node
,
834 struct cds_lfht_node
*dummy
, *next
, *old
;
835 struct _cds_lfht_node
*lookup
;
837 unsigned long hash
, index
, order
;
839 /* logically delete the node */
840 assert(!is_dummy(node
));
841 assert(!is_removed(node
));
842 old
= rcu_dereference(node
->p
.next
);
845 if (unlikely(is_removed(next
)))
848 assert(is_dummy(next
));
850 assert(!is_dummy(next
));
851 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
853 } while (old
!= next
);
855 /* We performed the (logical) deletion. */
859 * Ensure that the node is not visible to readers anymore: lookup for
860 * the node, and remove it (along with any other logically removed node)
863 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
865 index
= hash
& (size
- 1);
866 order
= get_count_order_ulong(index
+ 1);
867 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
868 dummy
= (struct cds_lfht_node
*) lookup
;
869 _cds_lfht_gc_bucket(dummy
, node
);
872 * Only the flagging action indicated that we (and no other)
873 * removed the node from the hash.
876 assert(is_removed(rcu_dereference(node
->p
.next
)));
883 * Holding RCU read lock to protect _cds_lfht_add against memory
884 * reclaim that could be performed by other call_rcu worker threads (ABA
888 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
892 ht
->cds_lfht_rcu_thread_online();
893 ht
->cds_lfht_rcu_read_lock();
894 for (j
= 0; j
< len
; j
++) {
895 struct cds_lfht_node
*new_node
=
896 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
898 dbg_printf("init populate: i %lu j %lu hash %lu\n",
899 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
900 new_node
->p
.reverse_hash
=
901 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
902 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
904 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
907 ht
->cds_lfht_rcu_read_unlock();
908 ht
->cds_lfht_rcu_thread_offline();
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
);
926 /* Stop expand if the resize target changes under us */
927 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
930 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
931 + (len
* sizeof(struct _cds_lfht_node
)));
934 * Set all dummy nodes reverse hash values for a level and
935 * link all dummy nodes into the table.
937 init_table_populate(ht
, i
, len
);
942 cmm_smp_wmb(); /* populate data before RCU size */
943 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
945 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
946 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
952 * Holding RCU read lock to protect _cds_lfht_remove against memory
953 * reclaim that could be performed by other call_rcu worker threads (ABA
955 * For a single level, we logically remove and garbage collect each node.
957 * As a design choice, we perform logical removal and garbage collection on a
958 * node-per-node basis to simplify this algorithm. We also assume keeping good
959 * cache locality of the operation would overweight possible performance gain
960 * that could be achieved by batching garbage collection for multiple levels.
961 * However, this would have to be justified by benchmarks.
963 * Concurrent removal and add operations are helping us perform garbage
964 * collection of logically removed nodes. We guarantee that all logically
965 * removed nodes have been garbage-collected (unlinked) before call_rcu is
966 * invoked to free a hole level of dummy nodes (after a grace period).
968 * Logical removal and garbage collection can therefore be done in batch or on a
969 * node-per-node basis, as long as the guarantee above holds.
972 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
976 ht
->cds_lfht_rcu_thread_online();
977 ht
->cds_lfht_rcu_read_lock();
978 for (j
= 0; j
< len
; j
++) {
979 struct cds_lfht_node
*fini_node
=
980 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
982 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
983 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
984 fini_node
->p
.reverse_hash
=
985 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
986 (void) _cds_lfht_remove(ht
, !i
? 0 : (1UL << (i
- 1)),
988 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
991 ht
->cds_lfht_rcu_read_unlock();
992 ht
->cds_lfht_rcu_thread_offline();
996 void fini_table(struct cds_lfht
*ht
,
997 unsigned long first_order
, unsigned long len_order
)
1001 dbg_printf("fini table: first_order %lu end_order %lu\n",
1002 first_order
, first_order
+ len_order
);
1003 end_order
= first_order
+ len_order
;
1004 assert(first_order
> 0);
1005 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1008 len
= !i
? 1 : 1UL << (i
- 1);
1009 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1011 /* Stop shrink if the resize target changes under us */
1012 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1015 cmm_smp_wmb(); /* populate data before RCU size */
1016 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1019 * We need to wait for all add operations to reach Q.S. (and
1020 * thus use the new table for lookups) before we can start
1021 * releasing the old dummy nodes. Otherwise their lookup will
1022 * return a logically removed node as insert position.
1024 ht
->cds_lfht_synchronize_rcu();
1027 * Set "removed" flag in dummy nodes about to be removed.
1028 * Unlink all now-logically-removed dummy node pointers.
1029 * Concurrent add/remove operation are helping us doing
1032 remove_table(ht
, i
, len
);
1034 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1036 dbg_printf("fini new size: %lu\n", 1UL << i
);
1037 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1042 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1043 cds_lfht_compare_fct compare_fct
,
1044 unsigned long hash_seed
,
1045 unsigned long init_size
,
1047 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1048 void (*func
)(struct rcu_head
*head
)),
1049 void (*cds_lfht_synchronize_rcu
)(void),
1050 void (*cds_lfht_rcu_read_lock
)(void),
1051 void (*cds_lfht_rcu_read_unlock
)(void),
1052 void (*cds_lfht_rcu_thread_offline
)(void),
1053 void (*cds_lfht_rcu_thread_online
)(void))
1055 struct cds_lfht
*ht
;
1056 unsigned long order
;
1058 /* init_size must be power of two */
1059 if (init_size
&& (init_size
& (init_size
- 1)))
1061 ht
= calloc(1, sizeof(struct cds_lfht
));
1062 ht
->hash_fct
= hash_fct
;
1063 ht
->compare_fct
= compare_fct
;
1064 ht
->hash_seed
= hash_seed
;
1065 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1066 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1067 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1068 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1069 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1070 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1071 ht
->percpu_count
= alloc_per_cpu_items_count();
1072 /* this mutex should not nest in read-side C.S. */
1073 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1074 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1076 ht
->cds_lfht_rcu_thread_offline();
1077 pthread_mutex_lock(&ht
->resize_mutex
);
1078 ht
->t
.resize_target
= 1UL << (order
- 1);
1079 init_table(ht
, 0, order
);
1080 pthread_mutex_unlock(&ht
->resize_mutex
);
1081 ht
->cds_lfht_rcu_thread_online();
1085 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1087 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1088 struct _cds_lfht_node
*lookup
;
1089 unsigned long hash
, reverse_hash
, index
, order
, size
;
1091 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1092 reverse_hash
= bit_reverse_ulong(hash
);
1094 size
= rcu_dereference(ht
->t
.size
);
1095 index
= hash
& (size
- 1);
1096 order
= get_count_order_ulong(index
+ 1);
1097 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1098 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1099 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1100 dummy_node
= (struct cds_lfht_node
*) lookup
;
1101 /* We can always skip the dummy node initially */
1102 node
= rcu_dereference(dummy_node
->p
.next
);
1103 node
= clear_flag(node
);
1105 if (unlikely(is_end(node
))) {
1109 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1113 next
= rcu_dereference(node
->p
.next
);
1114 if (likely(!is_removed(next
))
1116 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1119 node
= clear_flag(next
);
1121 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1125 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1126 struct cds_lfht_node
*node
)
1128 struct cds_lfht_node
*next
;
1129 unsigned long reverse_hash
;
1133 reverse_hash
= node
->p
.reverse_hash
;
1135 key_len
= node
->key_len
;
1136 next
= rcu_dereference(node
->p
.next
);
1137 node
= clear_flag(next
);
1140 if (unlikely(is_end(node
))) {
1144 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1148 next
= rcu_dereference(node
->p
.next
);
1149 if (likely(!is_removed(next
))
1151 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1154 node
= clear_flag(next
);
1156 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1160 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1162 unsigned long hash
, size
;
1164 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1165 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1167 size
= rcu_dereference(ht
->t
.size
);
1168 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1169 ht_count_add(ht
, size
);
1172 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1173 struct cds_lfht_node
*node
)
1175 unsigned long hash
, size
;
1176 struct cds_lfht_node
*ret
;
1178 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1179 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1181 size
= rcu_dereference(ht
->t
.size
);
1182 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1184 ht_count_add(ht
, size
);
1188 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1193 size
= rcu_dereference(ht
->t
.size
);
1194 ret
= _cds_lfht_remove(ht
, size
, node
, 0);
1196 ht_count_remove(ht
, size
);
1201 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1203 struct cds_lfht_node
*node
;
1204 struct _cds_lfht_node
*lookup
;
1205 unsigned long order
, i
, size
;
1207 /* Check that the table is empty */
1208 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1209 node
= (struct cds_lfht_node
*) lookup
;
1211 node
= clear_flag(node
)->p
.next
;
1212 if (!is_dummy(node
))
1214 assert(!is_removed(node
));
1215 } while (!is_end(node
));
1217 * size accessed without rcu_dereference because hash table is
1221 /* Internal sanity check: all nodes left should be dummy */
1222 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1225 len
= !order
? 1 : 1UL << (order
- 1);
1226 for (i
= 0; i
< len
; i
++) {
1227 dbg_printf("delete order %lu i %lu hash %lu\n",
1229 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1230 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1232 poison_free(ht
->t
.tbl
[order
]);
1238 * Should only be called when no more concurrent readers nor writers can
1239 * possibly access the table.
1241 int cds_lfht_destroy(struct cds_lfht
*ht
)
1245 /* Wait for in-flight resize operations to complete */
1246 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1247 while (uatomic_read(&ht
->in_progress_resize
))
1248 poll(NULL
, 0, 100); /* wait for 100ms */
1249 ret
= cds_lfht_delete_dummy(ht
);
1252 free_per_cpu_items_count(ht
->percpu_count
);
1257 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1258 unsigned long *count
,
1259 unsigned long *removed
)
1261 struct cds_lfht_node
*node
, *next
;
1262 struct _cds_lfht_node
*lookup
;
1263 unsigned long nr_dummy
= 0;
1268 /* Count non-dummy nodes in the table */
1269 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1270 node
= (struct cds_lfht_node
*) lookup
;
1272 next
= rcu_dereference(node
->p
.next
);
1273 if (is_removed(next
)) {
1274 assert(!is_dummy(next
));
1276 } else if (!is_dummy(next
))
1280 node
= clear_flag(next
);
1281 } while (!is_end(node
));
1282 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1285 /* called with resize mutex held */
1287 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1288 unsigned long old_size
, unsigned long new_size
)
1290 unsigned long old_order
, new_order
;
1292 old_order
= get_count_order_ulong(old_size
) + 1;
1293 new_order
= get_count_order_ulong(new_size
) + 1;
1294 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1295 old_size
, old_order
, new_size
, new_order
);
1296 assert(new_size
> old_size
);
1297 init_table(ht
, old_order
, new_order
- old_order
);
1300 /* called with resize mutex held */
1302 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1303 unsigned long old_size
, unsigned long new_size
)
1305 unsigned long old_order
, new_order
;
1307 new_size
= max(new_size
, MIN_TABLE_SIZE
);
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
);
1314 /* Remove and unlink all dummy nodes to remove. */
1315 fini_table(ht
, new_order
, old_order
- new_order
);
1319 /* called with resize mutex held */
1321 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1323 unsigned long new_size
, old_size
;
1326 * Resize table, re-do if the target size has changed under us.
1329 ht
->t
.resize_initiated
= 1;
1330 old_size
= ht
->t
.size
;
1331 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1332 if (old_size
< new_size
)
1333 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1334 else if (old_size
> new_size
)
1335 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1336 ht
->t
.resize_initiated
= 0;
1337 /* write resize_initiated before read resize_target */
1339 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1343 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1346 return _uatomic_max(&ht
->t
.resize_target
,
1347 size
<< growth_order
);
1351 void resize_target_update_count(struct cds_lfht
*ht
,
1352 unsigned long count
)
1354 count
= max(count
, MIN_TABLE_SIZE
);
1355 uatomic_set(&ht
->t
.resize_target
, count
);
1358 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1360 resize_target_update_count(ht
, new_size
);
1361 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1362 ht
->cds_lfht_rcu_thread_offline();
1363 pthread_mutex_lock(&ht
->resize_mutex
);
1364 _do_cds_lfht_resize(ht
);
1365 pthread_mutex_unlock(&ht
->resize_mutex
);
1366 ht
->cds_lfht_rcu_thread_online();
1370 void do_resize_cb(struct rcu_head
*head
)
1372 struct rcu_resize_work
*work
=
1373 caa_container_of(head
, struct rcu_resize_work
, head
);
1374 struct cds_lfht
*ht
= work
->ht
;
1376 ht
->cds_lfht_rcu_thread_offline();
1377 pthread_mutex_lock(&ht
->resize_mutex
);
1378 _do_cds_lfht_resize(ht
);
1379 pthread_mutex_unlock(&ht
->resize_mutex
);
1380 ht
->cds_lfht_rcu_thread_online();
1382 cmm_smp_mb(); /* finish resize before decrement */
1383 uatomic_dec(&ht
->in_progress_resize
);
1387 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1389 struct rcu_resize_work
*work
;
1390 unsigned long target_size
;
1392 target_size
= resize_target_update(ht
, size
, growth
);
1393 /* Store resize_target before read resize_initiated */
1395 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1396 uatomic_inc(&ht
->in_progress_resize
);
1397 cmm_smp_mb(); /* increment resize count before calling it */
1398 work
= malloc(sizeof(*work
));
1400 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1401 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1405 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1408 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1409 unsigned long count
)
1411 struct rcu_resize_work
*work
;
1413 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1415 resize_target_update_count(ht
, count
);
1416 /* Store resize_target before read resize_initiated */
1418 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1419 uatomic_inc(&ht
->in_progress_resize
);
1420 cmm_smp_mb(); /* increment resize count before calling it */
1421 work
= malloc(sizeof(*work
));
1423 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1424 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);