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_ORDER 12
184 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
187 #define min(a, b) ((a) < (b) ? (a) : (b))
191 #define max(a, b) ((a) > (b) ? (a) : (b))
195 * The removed flag needs to be updated atomically with the pointer.
196 * The dummy flag does not require to be updated atomically with the
197 * pointer, but it is added as a pointer low bit flag to save space.
199 #define REMOVED_FLAG (1UL << 0)
200 #define DUMMY_FLAG (1UL << 1)
201 #define FLAGS_MASK ((1UL << 2) - 1)
203 /* Value of the end pointer. Should not interact with flags. */
204 #define END_VALUE NULL
206 struct ht_items_count
{
207 unsigned long add
, del
;
208 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
211 struct rcu_head head
;
212 struct _cds_lfht_node nodes
[0];
216 unsigned long size
; /* always a power of 2, shared (RCU) */
217 unsigned long resize_target
;
218 int resize_initiated
;
219 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
224 cds_lfht_hash_fct hash_fct
;
225 cds_lfht_compare_fct compare_fct
;
226 unsigned long hash_seed
;
229 * We need to put the work threads offline (QSBR) when taking this
230 * mutex, because we use synchronize_rcu within this mutex critical
231 * section, which waits on read-side critical sections, and could
232 * therefore cause grace-period deadlock if we hold off RCU G.P.
235 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
236 unsigned int in_progress_resize
, in_progress_destroy
;
237 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
238 void (*func
)(struct rcu_head
*head
));
239 void (*cds_lfht_synchronize_rcu
)(void);
240 void (*cds_lfht_rcu_read_lock
)(void);
241 void (*cds_lfht_rcu_read_unlock
)(void);
242 void (*cds_lfht_rcu_thread_offline
)(void);
243 void (*cds_lfht_rcu_thread_online
)(void);
244 void (*cds_lfht_rcu_register_thread
)(void);
245 void (*cds_lfht_rcu_unregister_thread
)(void);
246 pthread_attr_t
*resize_attr
; /* Resize threads attributes */
247 unsigned long count
; /* global approximate item count */
248 struct ht_items_count
*percpu_count
; /* per-cpu item count */
251 struct rcu_resize_work
{
252 struct rcu_head head
;
256 struct partition_resize_work
{
257 struct rcu_head head
;
259 unsigned long i
, start
, len
;
260 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
261 unsigned long start
, unsigned long len
);
265 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
267 struct cds_lfht_node
*node
,
268 int unique
, int dummy
);
271 * Algorithm to reverse bits in a word by lookup table, extended to
274 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
275 * Originally from Public Domain.
278 static const uint8_t BitReverseTable256
[256] =
280 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
281 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
282 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
283 R6(0), R6(2), R6(1), R6(3)
290 uint8_t bit_reverse_u8(uint8_t v
)
292 return BitReverseTable256
[v
];
295 static __attribute__((unused
))
296 uint32_t bit_reverse_u32(uint32_t v
)
298 return ((uint32_t) bit_reverse_u8(v
) << 24) |
299 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
300 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
301 ((uint32_t) bit_reverse_u8(v
>> 24));
304 static __attribute__((unused
))
305 uint64_t bit_reverse_u64(uint64_t v
)
307 return ((uint64_t) bit_reverse_u8(v
) << 56) |
308 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
309 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
310 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
311 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
312 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
313 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
314 ((uint64_t) bit_reverse_u8(v
>> 56));
318 unsigned long bit_reverse_ulong(unsigned long v
)
320 #if (CAA_BITS_PER_LONG == 32)
321 return bit_reverse_u32(v
);
323 return bit_reverse_u64(v
);
328 * fls: returns the position of the most significant bit.
329 * Returns 0 if no bit is set, else returns the position of the most
330 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
332 #if defined(__i386) || defined(__x86_64)
334 unsigned int fls_u32(uint32_t x
)
342 : "=r" (r
) : "rm" (x
));
348 #if defined(__x86_64)
350 unsigned int fls_u64(uint64_t x
)
358 : "=r" (r
) : "rm" (x
));
365 static __attribute__((unused
))
366 unsigned int fls_u64(uint64_t x
)
373 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
377 if (!(x
& 0xFFFF000000000000ULL
)) {
381 if (!(x
& 0xFF00000000000000ULL
)) {
385 if (!(x
& 0xF000000000000000ULL
)) {
389 if (!(x
& 0xC000000000000000ULL
)) {
393 if (!(x
& 0x8000000000000000ULL
)) {
402 static __attribute__((unused
))
403 unsigned int fls_u32(uint32_t x
)
409 if (!(x
& 0xFFFF0000U
)) {
413 if (!(x
& 0xFF000000U
)) {
417 if (!(x
& 0xF0000000U
)) {
421 if (!(x
& 0xC0000000U
)) {
425 if (!(x
& 0x80000000U
)) {
433 unsigned int fls_ulong(unsigned long x
)
435 #if (CAA_BITS_PER_lONG == 32)
442 int get_count_order_u32(uint32_t x
)
446 order
= fls_u32(x
) - 1;
452 int get_count_order_ulong(unsigned long x
)
456 order
= fls_ulong(x
) - 1;
463 #define poison_free(ptr) \
465 memset(ptr, 0x42, sizeof(*(ptr))); \
469 #define poison_free(ptr) free(ptr)
473 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
476 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
477 * available, then we support hash table item accounting.
478 * In the unfortunate event the number of CPUs reported would be
479 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
481 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
484 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
485 unsigned long count
);
487 static long nr_cpus_mask
= -1;
490 struct ht_items_count
*alloc_per_cpu_items_count(void)
492 struct ht_items_count
*count
;
494 switch (nr_cpus_mask
) {
501 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
507 * round up number of CPUs to next power of two, so we
508 * can use & for modulo.
510 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
511 nr_cpus_mask
= maxcpus
- 1;
515 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
520 void free_per_cpu_items_count(struct ht_items_count
*count
)
530 assert(nr_cpus_mask
>= 0);
531 cpu
= sched_getcpu();
532 if (unlikely(cpu
< 0))
535 return cpu
& nr_cpus_mask
;
539 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
541 unsigned long percpu_count
;
544 if (unlikely(!ht
->percpu_count
))
547 if (unlikely(cpu
< 0))
549 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
550 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
553 dbg_printf("add percpu %lu\n", percpu_count
);
554 count
= uatomic_add_return(&ht
->count
,
555 1UL << COUNT_COMMIT_ORDER
);
557 if (!(count
& (count
- 1))) {
558 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
560 dbg_printf("add set global %lu\n", count
);
561 cds_lfht_resize_lazy_count(ht
, size
,
562 count
>> (CHAIN_LEN_TARGET
- 1));
568 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
570 unsigned long percpu_count
;
573 if (unlikely(!ht
->percpu_count
))
576 if (unlikely(cpu
< 0))
578 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].del
, -1);
579 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
582 dbg_printf("del percpu %lu\n", percpu_count
);
583 count
= uatomic_add_return(&ht
->count
,
584 -(1UL << COUNT_COMMIT_ORDER
));
586 if (!(count
& (count
- 1))) {
587 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
589 dbg_printf("del set global %lu\n", count
);
590 cds_lfht_resize_lazy_count(ht
, size
,
591 count
>> (CHAIN_LEN_TARGET
- 1));
596 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
598 static const long nr_cpus_mask
= -1;
601 struct ht_items_count
*alloc_per_cpu_items_count(void)
607 void free_per_cpu_items_count(struct ht_items_count
*count
)
612 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
617 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
)
621 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
625 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
629 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
631 count
= uatomic_read(&ht
->count
);
633 * Use bucket-local length for small table expand and for
634 * environments lacking per-cpu data support.
636 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
639 dbg_printf("WARNING: large chain length: %u.\n",
641 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
642 cds_lfht_resize_lazy(ht
, size
,
643 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
647 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
649 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
653 int is_removed(struct cds_lfht_node
*node
)
655 return ((unsigned long) node
) & REMOVED_FLAG
;
659 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
661 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
665 int is_dummy(struct cds_lfht_node
*node
)
667 return ((unsigned long) node
) & DUMMY_FLAG
;
671 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
673 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
677 struct cds_lfht_node
*get_end(void)
679 return (struct cds_lfht_node
*) END_VALUE
;
683 int is_end(struct cds_lfht_node
*node
)
685 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
689 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
691 unsigned long old1
, old2
;
693 old1
= uatomic_read(ptr
);
698 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
703 void cds_lfht_free_level(struct rcu_head
*head
)
705 struct rcu_level
*l
=
706 caa_container_of(head
, struct rcu_level
, head
);
711 * Remove all logically deleted nodes from a bucket up to a certain node key.
714 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
716 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
718 assert(!is_dummy(dummy
));
719 assert(!is_removed(dummy
));
720 assert(!is_dummy(node
));
721 assert(!is_removed(node
));
724 /* We can always skip the dummy node initially */
725 iter
= rcu_dereference(iter_prev
->p
.next
);
726 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
728 * We should never be called with dummy (start of chain)
729 * and logically removed node (end of path compression
730 * marker) being the actual same node. This would be a
731 * bug in the algorithm implementation.
733 assert(dummy
!= node
);
735 if (unlikely(is_end(iter
)))
737 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
739 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
740 if (likely(is_removed(next
)))
742 iter_prev
= clear_flag(iter
);
745 assert(!is_removed(iter
));
747 new_next
= flag_dummy(clear_flag(next
));
749 new_next
= clear_flag(next
);
750 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
756 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
758 struct cds_lfht_node
*node
,
759 int unique
, int dummy
)
761 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
763 struct _cds_lfht_node
*lookup
;
764 unsigned long hash
, index
, order
;
766 assert(!is_dummy(node
));
767 assert(!is_removed(node
));
770 node
->p
.next
= flag_dummy(get_end());
771 return node
; /* Initial first add (head) */
773 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
775 uint32_t chain_len
= 0;
778 * iter_prev points to the non-removed node prior to the
781 index
= hash
& (size
- 1);
782 order
= get_count_order_ulong(index
+ 1);
783 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
784 iter_prev
= (struct cds_lfht_node
*) lookup
;
785 /* We can always skip the dummy node initially */
786 iter
= rcu_dereference(iter_prev
->p
.next
);
787 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
789 if (unlikely(is_end(iter
)))
791 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
793 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
794 if (unlikely(is_removed(next
)))
798 && !ht
->compare_fct(node
->key
, node
->key_len
,
799 clear_flag(iter
)->key
,
800 clear_flag(iter
)->key_len
))
801 return clear_flag(iter
);
802 /* Only account for identical reverse hash once */
803 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
805 check_resize(ht
, size
, ++chain_len
);
806 iter_prev
= clear_flag(iter
);
810 assert(node
!= clear_flag(iter
));
811 assert(!is_removed(iter_prev
));
812 assert(!is_removed(iter
));
813 assert(iter_prev
!= node
);
815 node
->p
.next
= clear_flag(iter
);
817 node
->p
.next
= flag_dummy(clear_flag(iter
));
819 new_node
= flag_dummy(node
);
822 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
824 continue; /* retry */
828 assert(!is_removed(iter
));
830 new_next
= flag_dummy(clear_flag(next
));
832 new_next
= clear_flag(next
);
833 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
837 /* Garbage collect logically removed nodes in the bucket */
838 index
= hash
& (size
- 1);
839 order
= get_count_order_ulong(index
+ 1);
840 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
841 dummy_node
= (struct cds_lfht_node
*) lookup
;
842 _cds_lfht_gc_bucket(dummy_node
, node
);
847 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
848 struct cds_lfht_node
*node
,
851 struct cds_lfht_node
*dummy
, *next
, *old
;
852 struct _cds_lfht_node
*lookup
;
854 unsigned long hash
, index
, order
;
856 /* logically delete the node */
857 assert(!is_dummy(node
));
858 assert(!is_removed(node
));
859 old
= rcu_dereference(node
->p
.next
);
862 if (unlikely(is_removed(next
)))
865 assert(is_dummy(next
));
867 assert(!is_dummy(next
));
868 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
870 } while (old
!= next
);
872 /* We performed the (logical) deletion. */
876 * Ensure that the node is not visible to readers anymore: lookup for
877 * the node, and remove it (along with any other logically removed node)
880 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
882 index
= hash
& (size
- 1);
883 order
= get_count_order_ulong(index
+ 1);
884 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
885 dummy
= (struct cds_lfht_node
*) lookup
;
886 _cds_lfht_gc_bucket(dummy
, node
);
889 * Only the flagging action indicated that we (and no other)
890 * removed the node from the hash.
893 assert(is_removed(rcu_dereference(node
->p
.next
)));
900 void *partition_resize_thread(void *arg
)
902 struct partition_resize_work
*work
= arg
;
904 work
->ht
->cds_lfht_rcu_register_thread();
905 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
906 work
->ht
->cds_lfht_rcu_unregister_thread();
911 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
913 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
914 unsigned long start
, unsigned long len
))
916 unsigned long partition_len
;
917 struct partition_resize_work
*work
;
919 unsigned long nr_threads
;
920 pthread_t
*thread_id
;
923 * Note: nr_cpus_mask + 1 is always power of 2.
924 * We spawn just the number of threads we need to satisfy the minimum
925 * partition size, up to the number of CPUs in the system.
927 nr_threads
= min(nr_cpus_mask
+ 1,
928 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
929 partition_len
= len
>> get_count_order_ulong(nr_threads
);
930 work
= calloc(nr_threads
, sizeof(*work
));
931 thread_id
= calloc(nr_threads
, sizeof(*thread_id
));
933 for (thread
= 0; thread
< nr_threads
; thread
++) {
934 work
[thread
].ht
= ht
;
936 work
[thread
].len
= partition_len
;
937 work
[thread
].start
= thread
* partition_len
;
938 work
[thread
].fct
= fct
;
939 ret
= pthread_create(&thread_id
[thread
], ht
->resize_attr
,
940 partition_resize_thread
, &work
[thread
]);
943 for (thread
= 0; thread
< nr_threads
; thread
++) {
944 ret
= pthread_join(thread_id
[thread
], NULL
);
952 * Holding RCU read lock to protect _cds_lfht_add against memory
953 * reclaim that could be performed by other call_rcu worker threads (ABA
956 * When we reach a certain length, we can split this population phase over
957 * many worker threads, based on the number of CPUs available in the system.
958 * This should therefore take care of not having the expand lagging behind too
959 * many concurrent insertion threads by using the scheduler's ability to
960 * schedule dummy node population fairly with insertions.
963 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
964 unsigned long start
, unsigned long len
)
968 ht
->cds_lfht_rcu_read_lock();
969 for (j
= start
; j
< start
+ len
; j
++) {
970 struct cds_lfht_node
*new_node
=
971 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
973 dbg_printf("init populate: i %lu j %lu hash %lu\n",
974 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
975 new_node
->p
.reverse_hash
=
976 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
977 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
979 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
982 ht
->cds_lfht_rcu_read_unlock();
986 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
989 assert(nr_cpus_mask
!= -1);
990 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
991 ht
->cds_lfht_rcu_thread_online();
992 init_table_populate_partition(ht
, i
, 0, len
);
993 ht
->cds_lfht_rcu_thread_offline();
996 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1000 void init_table(struct cds_lfht
*ht
,
1001 unsigned long first_order
, unsigned long len_order
)
1003 unsigned long i
, end_order
;
1005 dbg_printf("init table: first_order %lu end_order %lu\n",
1006 first_order
, first_order
+ len_order
);
1007 end_order
= first_order
+ len_order
;
1008 for (i
= first_order
; i
< end_order
; i
++) {
1011 len
= !i
? 1 : 1UL << (i
- 1);
1012 dbg_printf("init order %lu len: %lu\n", i
, len
);
1014 /* Stop expand if the resize target changes under us */
1015 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1018 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1019 + (len
* sizeof(struct _cds_lfht_node
)));
1020 assert(ht
->t
.tbl
[i
]);
1023 * Set all dummy nodes reverse hash values for a level and
1024 * link all dummy nodes into the table.
1026 init_table_populate(ht
, i
, len
);
1029 * Update table size.
1031 cmm_smp_wmb(); /* populate data before RCU size */
1032 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1034 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1035 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1041 * Holding RCU read lock to protect _cds_lfht_remove against memory
1042 * reclaim that could be performed by other call_rcu worker threads (ABA
1044 * For a single level, we logically remove and garbage collect each node.
1046 * As a design choice, we perform logical removal and garbage collection on a
1047 * node-per-node basis to simplify this algorithm. We also assume keeping good
1048 * cache locality of the operation would overweight possible performance gain
1049 * that could be achieved by batching garbage collection for multiple levels.
1050 * However, this would have to be justified by benchmarks.
1052 * Concurrent removal and add operations are helping us perform garbage
1053 * collection of logically removed nodes. We guarantee that all logically
1054 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1055 * invoked to free a hole level of dummy nodes (after a grace period).
1057 * Logical removal and garbage collection can therefore be done in batch or on a
1058 * node-per-node basis, as long as the guarantee above holds.
1060 * When we reach a certain length, we can split this removal over many worker
1061 * threads, based on the number of CPUs available in the system. This should
1062 * take care of not letting resize process lag behind too many concurrent
1063 * updater threads actively inserting into the hash table.
1066 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1067 unsigned long start
, unsigned long len
)
1071 ht
->cds_lfht_rcu_read_lock();
1072 for (j
= start
; j
< start
+ len
; j
++) {
1073 struct cds_lfht_node
*fini_node
=
1074 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1076 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1077 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1078 fini_node
->p
.reverse_hash
=
1079 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1080 (void) _cds_lfht_del(ht
, !i
? 0 : (1UL << (i
- 1)),
1082 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1085 ht
->cds_lfht_rcu_read_unlock();
1089 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1092 assert(nr_cpus_mask
!= -1);
1093 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1094 ht
->cds_lfht_rcu_thread_online();
1095 remove_table_partition(ht
, i
, 0, len
);
1096 ht
->cds_lfht_rcu_thread_offline();
1099 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1103 void fini_table(struct cds_lfht
*ht
,
1104 unsigned long first_order
, unsigned long len_order
)
1108 dbg_printf("fini table: first_order %lu end_order %lu\n",
1109 first_order
, first_order
+ len_order
);
1110 end_order
= first_order
+ len_order
;
1111 assert(first_order
> 0);
1112 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1115 len
= !i
? 1 : 1UL << (i
- 1);
1116 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1118 /* Stop shrink if the resize target changes under us */
1119 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1122 cmm_smp_wmb(); /* populate data before RCU size */
1123 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1126 * We need to wait for all add operations to reach Q.S. (and
1127 * thus use the new table for lookups) before we can start
1128 * releasing the old dummy nodes. Otherwise their lookup will
1129 * return a logically removed node as insert position.
1131 ht
->cds_lfht_synchronize_rcu();
1134 * Set "removed" flag in dummy nodes about to be removed.
1135 * Unlink all now-logically-removed dummy node pointers.
1136 * Concurrent add/remove operation are helping us doing
1139 remove_table(ht
, i
, len
);
1141 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1143 dbg_printf("fini new size: %lu\n", 1UL << i
);
1144 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1149 struct cds_lfht
*_cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1150 cds_lfht_compare_fct compare_fct
,
1151 unsigned long hash_seed
,
1152 unsigned long init_size
,
1154 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1155 void (*func
)(struct rcu_head
*head
)),
1156 void (*cds_lfht_synchronize_rcu
)(void),
1157 void (*cds_lfht_rcu_read_lock
)(void),
1158 void (*cds_lfht_rcu_read_unlock
)(void),
1159 void (*cds_lfht_rcu_thread_offline
)(void),
1160 void (*cds_lfht_rcu_thread_online
)(void),
1161 void (*cds_lfht_rcu_register_thread
)(void),
1162 void (*cds_lfht_rcu_unregister_thread
)(void),
1163 pthread_attr_t
*attr
)
1165 struct cds_lfht
*ht
;
1166 unsigned long order
;
1168 /* init_size must be power of two */
1169 if (init_size
&& (init_size
& (init_size
- 1)))
1171 ht
= calloc(1, sizeof(struct cds_lfht
));
1173 ht
->hash_fct
= hash_fct
;
1174 ht
->compare_fct
= compare_fct
;
1175 ht
->hash_seed
= hash_seed
;
1176 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1177 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1178 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1179 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1180 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1181 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1182 ht
->cds_lfht_rcu_register_thread
= cds_lfht_rcu_register_thread
;
1183 ht
->cds_lfht_rcu_unregister_thread
= cds_lfht_rcu_unregister_thread
;
1184 ht
->resize_attr
= attr
;
1185 ht
->percpu_count
= alloc_per_cpu_items_count();
1186 /* this mutex should not nest in read-side C.S. */
1187 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1188 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1190 ht
->cds_lfht_rcu_thread_offline();
1191 pthread_mutex_lock(&ht
->resize_mutex
);
1192 ht
->t
.resize_target
= 1UL << (order
- 1);
1193 init_table(ht
, 0, order
);
1194 pthread_mutex_unlock(&ht
->resize_mutex
);
1195 ht
->cds_lfht_rcu_thread_online();
1199 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1201 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1202 struct _cds_lfht_node
*lookup
;
1203 unsigned long hash
, reverse_hash
, index
, order
, size
;
1205 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1206 reverse_hash
= bit_reverse_ulong(hash
);
1208 size
= rcu_dereference(ht
->t
.size
);
1209 index
= hash
& (size
- 1);
1210 order
= get_count_order_ulong(index
+ 1);
1211 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1212 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1213 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1214 dummy_node
= (struct cds_lfht_node
*) lookup
;
1215 /* We can always skip the dummy node initially */
1216 node
= rcu_dereference(dummy_node
->p
.next
);
1217 node
= clear_flag(node
);
1219 if (unlikely(is_end(node
))) {
1223 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1227 next
= rcu_dereference(node
->p
.next
);
1228 if (likely(!is_removed(next
))
1230 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1233 node
= clear_flag(next
);
1235 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1239 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1240 struct cds_lfht_node
*node
)
1242 struct cds_lfht_node
*next
;
1243 unsigned long reverse_hash
;
1247 reverse_hash
= node
->p
.reverse_hash
;
1249 key_len
= node
->key_len
;
1250 next
= rcu_dereference(node
->p
.next
);
1251 node
= clear_flag(next
);
1254 if (unlikely(is_end(node
))) {
1258 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1262 next
= rcu_dereference(node
->p
.next
);
1263 if (likely(!is_removed(next
))
1265 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1268 node
= clear_flag(next
);
1270 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1274 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1276 unsigned long hash
, size
;
1278 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1279 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1281 size
= rcu_dereference(ht
->t
.size
);
1282 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1283 ht_count_add(ht
, size
);
1286 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1287 struct cds_lfht_node
*node
)
1289 unsigned long hash
, size
;
1290 struct cds_lfht_node
*ret
;
1292 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1293 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1295 size
= rcu_dereference(ht
->t
.size
);
1296 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1298 ht_count_add(ht
, size
);
1302 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1307 size
= rcu_dereference(ht
->t
.size
);
1308 ret
= _cds_lfht_del(ht
, size
, node
, 0);
1310 ht_count_del(ht
, size
);
1315 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1317 struct cds_lfht_node
*node
;
1318 struct _cds_lfht_node
*lookup
;
1319 unsigned long order
, i
, size
;
1321 /* Check that the table is empty */
1322 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1323 node
= (struct cds_lfht_node
*) lookup
;
1325 node
= clear_flag(node
)->p
.next
;
1326 if (!is_dummy(node
))
1328 assert(!is_removed(node
));
1329 } while (!is_end(node
));
1331 * size accessed without rcu_dereference because hash table is
1335 /* Internal sanity check: all nodes left should be dummy */
1336 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1339 len
= !order
? 1 : 1UL << (order
- 1);
1340 for (i
= 0; i
< len
; i
++) {
1341 dbg_printf("delete order %lu i %lu hash %lu\n",
1343 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1344 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1346 poison_free(ht
->t
.tbl
[order
]);
1352 * Should only be called when no more concurrent readers nor writers can
1353 * possibly access the table.
1355 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1359 /* Wait for in-flight resize operations to complete */
1360 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1361 while (uatomic_read(&ht
->in_progress_resize
))
1362 poll(NULL
, 0, 100); /* wait for 100ms */
1363 ret
= cds_lfht_delete_dummy(ht
);
1366 free_per_cpu_items_count(ht
->percpu_count
);
1368 *attr
= ht
->resize_attr
;
1373 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1374 unsigned long *count
,
1375 unsigned long *removed
)
1377 struct cds_lfht_node
*node
, *next
;
1378 struct _cds_lfht_node
*lookup
;
1379 unsigned long nr_dummy
= 0;
1384 /* Count non-dummy nodes in the table */
1385 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1386 node
= (struct cds_lfht_node
*) lookup
;
1388 next
= rcu_dereference(node
->p
.next
);
1389 if (is_removed(next
)) {
1390 assert(!is_dummy(next
));
1392 } else if (!is_dummy(next
))
1396 node
= clear_flag(next
);
1397 } while (!is_end(node
));
1398 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1401 /* called with resize mutex held */
1403 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1404 unsigned long old_size
, unsigned long new_size
)
1406 unsigned long old_order
, new_order
;
1408 old_order
= get_count_order_ulong(old_size
) + 1;
1409 new_order
= get_count_order_ulong(new_size
) + 1;
1410 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1411 old_size
, old_order
, new_size
, new_order
);
1412 assert(new_size
> old_size
);
1413 init_table(ht
, old_order
, new_order
- old_order
);
1416 /* called with resize mutex held */
1418 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1419 unsigned long old_size
, unsigned long new_size
)
1421 unsigned long old_order
, new_order
;
1423 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1424 old_order
= get_count_order_ulong(old_size
) + 1;
1425 new_order
= get_count_order_ulong(new_size
) + 1;
1426 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1427 old_size
, old_order
, new_size
, new_order
);
1428 assert(new_size
< old_size
);
1430 /* Remove and unlink all dummy nodes to remove. */
1431 fini_table(ht
, new_order
, old_order
- new_order
);
1435 /* called with resize mutex held */
1437 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1439 unsigned long new_size
, old_size
;
1442 * Resize table, re-do if the target size has changed under us.
1445 ht
->t
.resize_initiated
= 1;
1446 old_size
= ht
->t
.size
;
1447 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1448 if (old_size
< new_size
)
1449 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1450 else if (old_size
> new_size
)
1451 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1452 ht
->t
.resize_initiated
= 0;
1453 /* write resize_initiated before read resize_target */
1455 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1459 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1462 return _uatomic_max(&ht
->t
.resize_target
,
1463 size
<< growth_order
);
1467 void resize_target_update_count(struct cds_lfht
*ht
,
1468 unsigned long count
)
1470 count
= max(count
, MIN_TABLE_SIZE
);
1471 uatomic_set(&ht
->t
.resize_target
, count
);
1474 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1476 resize_target_update_count(ht
, new_size
);
1477 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1478 ht
->cds_lfht_rcu_thread_offline();
1479 pthread_mutex_lock(&ht
->resize_mutex
);
1480 _do_cds_lfht_resize(ht
);
1481 pthread_mutex_unlock(&ht
->resize_mutex
);
1482 ht
->cds_lfht_rcu_thread_online();
1486 void do_resize_cb(struct rcu_head
*head
)
1488 struct rcu_resize_work
*work
=
1489 caa_container_of(head
, struct rcu_resize_work
, head
);
1490 struct cds_lfht
*ht
= work
->ht
;
1492 ht
->cds_lfht_rcu_thread_offline();
1493 pthread_mutex_lock(&ht
->resize_mutex
);
1494 _do_cds_lfht_resize(ht
);
1495 pthread_mutex_unlock(&ht
->resize_mutex
);
1496 ht
->cds_lfht_rcu_thread_online();
1498 cmm_smp_mb(); /* finish resize before decrement */
1499 uatomic_dec(&ht
->in_progress_resize
);
1503 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1505 struct rcu_resize_work
*work
;
1506 unsigned long target_size
;
1508 target_size
= resize_target_update(ht
, size
, growth
);
1509 /* Store resize_target before read resize_initiated */
1511 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1512 uatomic_inc(&ht
->in_progress_resize
);
1513 cmm_smp_mb(); /* increment resize count before calling it */
1514 work
= malloc(sizeof(*work
));
1516 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1517 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1521 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1524 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1525 unsigned long count
)
1527 struct rcu_resize_work
*work
;
1529 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1531 resize_target_update_count(ht
, count
);
1532 /* Store resize_target before read resize_initiated */
1534 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
1535 uatomic_inc(&ht
->in_progress_resize
);
1536 cmm_smp_mb(); /* increment resize count before calling it */
1537 work
= malloc(sizeof(*work
));
1539 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1540 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);