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...)
163 * Per-CPU split-counters lazily update the global counter each 1024
164 * addition/removal. It automatically keeps track of resize required.
165 * We use the bucket length as indicator for need to expand for small
166 * tables and machines lacking per-cpu data suppport.
168 #define COUNT_COMMIT_ORDER 10
169 #define CHAIN_LEN_TARGET 1
170 #define CHAIN_LEN_RESIZE_THRESHOLD 3
173 * Define the minimum table size. Protects against hash table resize overload
174 * when too many entries are added quickly before the resize can complete.
175 * This is especially the case if the table could be shrinked to a size of 1.
176 * TODO: we might want to make the add/remove operations help the resize to
177 * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on
180 #define MIN_TABLE_SIZE 128
183 #define max(a, b) ((a) > (b) ? (a) : (b))
187 * The removed flag needs to be updated atomically with the pointer.
188 * The dummy flag does not require to be updated atomically with the
189 * pointer, but it is added as a pointer low bit flag to save space.
191 #define REMOVED_FLAG (1UL << 0)
192 #define DUMMY_FLAG (1UL << 1)
193 #define FLAGS_MASK ((1UL << 2) - 1)
195 struct ht_items_count
{
196 unsigned long add
, remove
;
197 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
200 struct rcu_head head
;
201 struct _cds_lfht_node nodes
[0];
205 unsigned long size
; /* always a power of 2 */
206 unsigned long resize_target
;
207 int resize_initiated
;
208 struct rcu_head head
;
209 struct rcu_level
*tbl
[0];
213 struct rcu_table
*t
; /* shared */
214 cds_lfht_hash_fct hash_fct
;
215 cds_lfht_compare_fct compare_fct
;
216 unsigned long hash_seed
;
218 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
219 unsigned int in_progress_resize
, in_progress_destroy
;
220 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
221 void (*func
)(struct rcu_head
*head
));
222 void (*cds_lfht_synchronize_rcu
)(void);
223 unsigned long count
; /* global approximate item count */
224 struct ht_items_count
*percpu_count
; /* per-cpu item count */
227 struct rcu_resize_work
{
228 struct rcu_head head
;
233 * Algorithm to reverse bits in a word by lookup table, extended to
236 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
237 * Originally from Public Domain.
240 static const uint8_t BitReverseTable256
[256] =
242 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
243 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
244 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
245 R6(0), R6(2), R6(1), R6(3)
252 uint8_t bit_reverse_u8(uint8_t v
)
254 return BitReverseTable256
[v
];
257 static __attribute__((unused
))
258 uint32_t bit_reverse_u32(uint32_t v
)
260 return ((uint32_t) bit_reverse_u8(v
) << 24) |
261 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
262 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
263 ((uint32_t) bit_reverse_u8(v
>> 24));
266 static __attribute__((unused
))
267 uint64_t bit_reverse_u64(uint64_t v
)
269 return ((uint64_t) bit_reverse_u8(v
) << 56) |
270 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
271 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
272 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
273 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
274 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
275 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
276 ((uint64_t) bit_reverse_u8(v
>> 56));
280 unsigned long bit_reverse_ulong(unsigned long v
)
282 #if (CAA_BITS_PER_LONG == 32)
283 return bit_reverse_u32(v
);
285 return bit_reverse_u64(v
);
290 * fls: returns the position of the most significant bit.
291 * Returns 0 if no bit is set, else returns the position of the most
292 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
294 #if defined(__i386) || defined(__x86_64)
296 unsigned int fls_u32(uint32_t x
)
304 : "=r" (r
) : "rm" (x
));
310 #if defined(__x86_64)
312 unsigned int fls_u64(uint64_t x
)
320 : "=r" (r
) : "rm" (x
));
327 static __attribute__((unused
))
328 unsigned int fls_u64(uint64_t x
)
335 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
339 if (!(x
& 0xFFFF000000000000ULL
)) {
343 if (!(x
& 0xFF00000000000000ULL
)) {
347 if (!(x
& 0xF000000000000000ULL
)) {
351 if (!(x
& 0xC000000000000000ULL
)) {
355 if (!(x
& 0x8000000000000000ULL
)) {
364 static __attribute__((unused
))
365 unsigned int fls_u32(uint32_t x
)
371 if (!(x
& 0xFFFF0000U
)) {
375 if (!(x
& 0xFF000000U
)) {
379 if (!(x
& 0xF0000000U
)) {
383 if (!(x
& 0xC0000000U
)) {
387 if (!(x
& 0x80000000U
)) {
395 unsigned int fls_ulong(unsigned long x
)
397 #if (CAA_BITS_PER_lONG == 32)
404 int get_count_order_u32(uint32_t x
)
408 order
= fls_u32(x
) - 1;
414 int get_count_order_ulong(unsigned long x
)
418 order
= fls_ulong(x
) - 1;
425 #define poison_free(ptr) \
427 memset(ptr, 0x42, sizeof(*(ptr))); \
431 #define poison_free(ptr) free(ptr)
435 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
);
438 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
439 * available, then we support hash table item accounting.
440 * In the unfortunate event the number of CPUs reported would be
441 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
443 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
446 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
447 unsigned long count
);
449 static long nr_cpus_mask
= -1;
452 struct ht_items_count
*alloc_per_cpu_items_count(void)
454 struct ht_items_count
*count
;
456 switch (nr_cpus_mask
) {
463 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
469 * round up number of CPUs to next power of two, so we
470 * can use & for modulo.
472 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
473 nr_cpus_mask
= maxcpus
- 1;
477 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
482 void free_per_cpu_items_count(struct ht_items_count
*count
)
492 assert(nr_cpus_mask
>= 0);
493 cpu
= sched_getcpu();
494 if (unlikely(cpu
< 0))
497 return cpu
& nr_cpus_mask
;
501 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
503 unsigned long percpu_count
;
506 if (unlikely(!ht
->percpu_count
))
509 if (unlikely(cpu
< 0))
511 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
512 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
515 dbg_printf("add percpu %lu\n", percpu_count
);
516 count
= uatomic_add_return(&ht
->count
,
517 1UL << COUNT_COMMIT_ORDER
);
519 if (!(count
& (count
- 1))) {
520 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
523 dbg_printf("add set global %lu\n", count
);
524 cds_lfht_resize_lazy_count(ht
, t
,
525 count
>> (CHAIN_LEN_TARGET
- 1));
531 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
533 unsigned long percpu_count
;
536 if (unlikely(!ht
->percpu_count
))
539 if (unlikely(cpu
< 0))
541 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
542 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
545 dbg_printf("remove percpu %lu\n", percpu_count
);
546 count
= uatomic_add_return(&ht
->count
,
547 -(1UL << COUNT_COMMIT_ORDER
));
549 if (!(count
& (count
- 1))) {
550 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
)
553 dbg_printf("remove set global %lu\n", count
);
554 cds_lfht_resize_lazy_count(ht
, t
,
555 count
>> (CHAIN_LEN_TARGET
- 1));
560 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
562 static const long nr_cpus_mask
= -1;
565 struct ht_items_count
*alloc_per_cpu_items_count(void)
571 void free_per_cpu_items_count(struct ht_items_count
*count
)
576 void ht_count_add(struct cds_lfht
*ht
, struct rcu_table
*t
)
581 void ht_count_remove(struct cds_lfht
*ht
, struct rcu_table
*t
)
585 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
589 void check_resize(struct cds_lfht
*ht
, struct rcu_table
*t
,
594 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
596 count
= uatomic_read(&ht
->count
);
598 * Use bucket-local length for small table expand and for
599 * environments lacking per-cpu data support.
601 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
604 dbg_printf("WARNING: large chain length: %u.\n",
606 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
607 cds_lfht_resize_lazy(ht
, t
,
608 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
612 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
614 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
618 int is_removed(struct cds_lfht_node
*node
)
620 return ((unsigned long) node
) & REMOVED_FLAG
;
624 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
626 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
630 int is_dummy(struct cds_lfht_node
*node
)
632 return ((unsigned long) node
) & DUMMY_FLAG
;
636 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
638 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
642 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
644 unsigned long old1
, old2
;
646 old1
= uatomic_read(ptr
);
651 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
656 void cds_lfht_free_table_cb(struct rcu_head
*head
)
658 struct rcu_table
*t
=
659 caa_container_of(head
, struct rcu_table
, head
);
664 void cds_lfht_free_level(struct rcu_head
*head
)
666 struct rcu_level
*l
=
667 caa_container_of(head
, struct rcu_level
, head
);
672 * Remove all logically deleted nodes from a bucket up to a certain node key.
675 void _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
677 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
678 struct cds_lfht_node
*iter_trace
[64];
679 unsigned long trace_idx
= 0;
681 memset(iter_trace
, 0, sizeof(iter_trace
));
682 assert(!is_dummy(dummy
));
683 assert(!is_removed(dummy
));
684 assert(!is_dummy(node
));
685 assert(!is_removed(node
));
687 iter_trace
[trace_idx
++ & (64 - 1)] = (void *) 0x1;
689 /* We can always skip the dummy node initially */
690 iter
= rcu_dereference(iter_prev
->p
.next
);
691 iter_trace
[trace_idx
++ & (64 - 1)] = iter
;
692 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
694 * We should never be called with dummy (start of chain)
695 * and logically removed node (end of path compression
696 * marker) being the actual same node. This would be a
697 * bug in the algorithm implementation.
699 assert(dummy
!= node
);
701 if (unlikely(!clear_flag(iter
)))
703 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
705 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
706 if (likely(is_removed(next
)))
708 iter_prev
= clear_flag(iter
);
710 iter_trace
[trace_idx
++ & (64 - 1)] = iter
;
712 assert(!is_removed(iter
));
714 new_next
= flag_dummy(clear_flag(next
));
716 new_next
= clear_flag(next
);
717 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
718 iter_trace
[trace_idx
++ & (64 - 1)] = (void *) 0x2;
723 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
, struct rcu_table
*t
,
724 struct cds_lfht_node
*node
, int unique
, int dummy
)
726 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
728 struct _cds_lfht_node
*lookup
;
729 unsigned long hash
, index
, order
;
731 assert(!is_dummy(node
));
732 assert(!is_removed(node
));
735 node
->p
.next
= flag_dummy(NULL
);
736 return node
; /* Initial first add (head) */
738 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
740 uint32_t chain_len
= 0;
743 * iter_prev points to the non-removed node prior to the
746 index
= hash
& (t
->size
- 1);
747 order
= get_count_order_ulong(index
+ 1);
748 lookup
= &t
->tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
749 iter_prev
= (struct cds_lfht_node
*) lookup
;
750 /* We can always skip the dummy node initially */
751 iter
= rcu_dereference(iter_prev
->p
.next
);
752 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
754 /* TODO: check if removed */
755 if (unlikely(!clear_flag(iter
)))
757 /* TODO: check if removed */
758 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
760 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
761 if (unlikely(is_removed(next
)))
765 && !ht
->compare_fct(node
->key
, node
->key_len
,
766 clear_flag(iter
)->key
,
767 clear_flag(iter
)->key_len
))
768 return clear_flag(iter
);
769 /* Only account for identical reverse hash once */
770 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
772 check_resize(ht
, t
, ++chain_len
);
773 iter_prev
= clear_flag(iter
);
777 assert(node
!= clear_flag(iter
));
778 assert(!is_removed(iter_prev
));
779 assert(!is_removed(iter
));
780 assert(iter_prev
!= node
);
782 node
->p
.next
= clear_flag(iter
);
784 node
->p
.next
= flag_dummy(clear_flag(iter
));
786 new_node
= flag_dummy(node
);
789 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
791 continue; /* retry */
795 assert(!is_removed(iter
));
797 new_next
= flag_dummy(clear_flag(next
));
799 new_next
= clear_flag(next
);
800 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
804 /* Garbage collect logically removed nodes in the bucket */
805 index
= hash
& (t
->size
- 1);
806 order
= get_count_order_ulong(index
+ 1);
807 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
808 dummy_node
= (struct cds_lfht_node
*) lookup
;
809 _cds_lfht_gc_bucket(dummy_node
, node
);
814 int _cds_lfht_remove(struct cds_lfht
*ht
, struct rcu_table
*t
,
815 struct cds_lfht_node
*node
, int dummy_removal
)
817 struct cds_lfht_node
*dummy
, *next
, *old
;
818 struct _cds_lfht_node
*lookup
;
820 unsigned long hash
, index
, order
;
822 /* logically delete the node */
823 assert(!is_dummy(node
));
824 assert(!is_removed(node
));
825 old
= rcu_dereference(node
->p
.next
);
828 if (unlikely(is_removed(next
)))
831 assert(is_dummy(next
));
833 assert(!is_dummy(next
));
834 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
836 } while (old
!= next
);
838 /* We performed the (logical) deletion. */
842 * Ensure that the node is not visible to readers anymore: lookup for
843 * the node, and remove it (along with any other logically removed node)
846 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
848 index
= hash
& (t
->size
- 1);
849 order
= get_count_order_ulong(index
+ 1);
850 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
851 dummy
= (struct cds_lfht_node
*) lookup
;
852 _cds_lfht_gc_bucket(dummy
, node
);
855 * Only the flagging action indicated that we (and no other)
856 * removed the node from the hash.
859 assert(is_removed(rcu_dereference(node
->p
.next
)));
866 void init_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
867 unsigned long first_order
, unsigned long len_order
)
869 unsigned long i
, end_order
;
871 dbg_printf("init table: first_order %lu end_order %lu\n",
872 first_order
, first_order
+ len_order
);
873 end_order
= first_order
+ len_order
;
874 t
->size
= !first_order
? 0 : (1UL << (first_order
- 1));
875 for (i
= first_order
; i
< end_order
; i
++) {
876 unsigned long j
, len
;
878 len
= !i
? 1 : 1UL << (i
- 1);
879 dbg_printf("init order %lu len: %lu\n", i
, len
);
880 t
->tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
881 + (len
* sizeof(struct _cds_lfht_node
)));
882 for (j
= 0; j
< len
; j
++) {
883 struct cds_lfht_node
*new_node
=
884 (struct cds_lfht_node
*) &t
->tbl
[i
]->nodes
[j
];
886 dbg_printf("init entry: i %lu j %lu hash %lu\n",
887 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
888 new_node
->p
.reverse_hash
=
889 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
890 (void) _cds_lfht_add(ht
, t
, new_node
, 0, 1);
891 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
894 /* Update table size */
895 t
->size
= !i
? 1 : (1UL << i
);
896 dbg_printf("init new size: %lu\n", t
->size
);
897 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
900 t
->resize_target
= t
->size
;
901 t
->resize_initiated
= 0;
905 void fini_table(struct cds_lfht
*ht
, struct rcu_table
*t
,
906 unsigned long first_order
, unsigned long len_order
)
910 dbg_printf("fini table: first_order %lu end_order %lu\n",
911 first_order
, first_order
+ len_order
);
912 end_order
= first_order
+ len_order
;
913 assert(first_order
> 0);
914 assert(t
->size
== (1UL << (end_order
- 1)));
915 for (i
= end_order
- 1; i
>= first_order
; i
--) {
916 unsigned long j
, len
;
918 len
= !i
? 1 : 1UL << (i
- 1);
919 dbg_printf("fini order %lu len: %lu\n", i
, len
);
921 * Update table size. Need to shrink this table prior to
922 * removal so gc lookups use non-logically-removed dummy
925 t
->size
= 1UL << (i
- 1);
927 for (j
= 0; j
< len
; j
++) {
928 struct cds_lfht_node
*fini_node
=
929 (struct cds_lfht_node
*) &t
->tbl
[i
]->nodes
[j
];
931 dbg_printf("fini entry: i %lu j %lu hash %lu\n",
932 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
933 fini_node
->p
.reverse_hash
=
934 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
935 (void) _cds_lfht_remove(ht
, t
, fini_node
, 1);
936 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
939 ht
->cds_lfht_call_rcu(&t
->tbl
[i
]->head
, cds_lfht_free_level
);
940 dbg_printf("fini new size: %lu\n", t
->size
);
941 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
944 t
->resize_target
= t
->size
;
945 t
->resize_initiated
= 0;
948 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
949 cds_lfht_compare_fct compare_fct
,
950 unsigned long hash_seed
,
951 unsigned long init_size
,
953 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
954 void (*func
)(struct rcu_head
*head
)),
955 void (*cds_lfht_synchronize_rcu
)(void))
960 /* init_size must be power of two */
961 if (init_size
&& (init_size
& (init_size
- 1)))
963 ht
= calloc(1, sizeof(struct cds_lfht
));
964 ht
->hash_fct
= hash_fct
;
965 ht
->compare_fct
= compare_fct
;
966 ht
->hash_seed
= hash_seed
;
967 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
968 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
969 ht
->in_progress_resize
= 0;
970 ht
->percpu_count
= alloc_per_cpu_items_count();
971 /* this mutex should not nest in read-side C.S. */
972 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
973 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
974 ht
->t
= calloc(1, sizeof(struct cds_lfht
)
975 + (order
* sizeof(struct rcu_level
*)));
978 pthread_mutex_lock(&ht
->resize_mutex
);
979 init_table(ht
, ht
->t
, 0, order
);
980 pthread_mutex_unlock(&ht
->resize_mutex
);
984 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
987 struct cds_lfht_node
*node
, *next
;
988 struct _cds_lfht_node
*lookup
;
989 unsigned long hash
, reverse_hash
, index
, order
;
991 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
992 reverse_hash
= bit_reverse_ulong(hash
);
994 t
= rcu_dereference(ht
->t
);
995 index
= hash
& (t
->size
- 1);
996 order
= get_count_order_ulong(index
+ 1);
997 lookup
= &t
->tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
998 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
999 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1000 node
= (struct cds_lfht_node
*) lookup
;
1002 if (unlikely(!node
))
1004 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1008 next
= rcu_dereference(node
->p
.next
);
1009 if (likely(!is_removed(next
))
1011 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1014 node
= clear_flag(next
);
1016 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1020 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1021 struct cds_lfht_node
*node
)
1023 struct cds_lfht_node
*next
;
1024 unsigned long reverse_hash
;
1028 reverse_hash
= node
->p
.reverse_hash
;
1030 key_len
= node
->key_len
;
1031 next
= rcu_dereference(node
->p
.next
);
1032 node
= clear_flag(next
);
1035 if (unlikely(!node
))
1037 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1041 next
= rcu_dereference(node
->p
.next
);
1042 if (likely(!is_removed(next
))
1044 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1047 node
= clear_flag(next
);
1049 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1053 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1055 struct rcu_table
*t
;
1058 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1059 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1061 t
= rcu_dereference(ht
->t
);
1062 (void) _cds_lfht_add(ht
, t
, node
, 0, 0);
1063 ht_count_add(ht
, t
);
1066 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1067 struct cds_lfht_node
*node
)
1069 struct rcu_table
*t
;
1071 struct cds_lfht_node
*ret
;
1073 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1074 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1076 t
= rcu_dereference(ht
->t
);
1077 ret
= _cds_lfht_add(ht
, t
, node
, 1, 0);
1079 ht_count_add(ht
, t
);
1083 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1085 struct rcu_table
*t
;
1088 t
= rcu_dereference(ht
->t
);
1089 ret
= _cds_lfht_remove(ht
, t
, node
, 0);
1091 ht_count_remove(ht
, t
);
1096 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1098 struct rcu_table
*t
;
1099 struct cds_lfht_node
*node
;
1100 struct _cds_lfht_node
*lookup
;
1101 unsigned long order
, i
;
1104 /* Check that the table is empty */
1105 lookup
= &t
->tbl
[0]->nodes
[0];
1106 node
= (struct cds_lfht_node
*) lookup
;
1108 node
= clear_flag(node
)->p
.next
;
1109 if (!is_dummy(node
))
1111 assert(!is_removed(node
));
1112 } while (clear_flag(node
));
1113 /* Internal sanity check: all nodes left should be dummy */
1114 for (order
= 0; order
< get_count_order_ulong(t
->size
) + 1; order
++) {
1117 len
= !order
? 1 : 1UL << (order
- 1);
1118 for (i
= 0; i
< len
; i
++) {
1119 dbg_printf("delete order %lu i %lu hash %lu\n",
1121 bit_reverse_ulong(t
->tbl
[order
]->nodes
[i
].reverse_hash
));
1122 assert(is_dummy(t
->tbl
[order
]->nodes
[i
].next
));
1124 poison_free(t
->tbl
[order
]);
1130 * Should only be called when no more concurrent readers nor writers can
1131 * possibly access the table.
1133 int cds_lfht_destroy(struct cds_lfht
*ht
)
1137 /* Wait for in-flight resize operations to complete */
1138 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1139 while (uatomic_read(&ht
->in_progress_resize
))
1140 poll(NULL
, 0, 100); /* wait for 100ms */
1141 ret
= cds_lfht_delete_dummy(ht
);
1145 free_per_cpu_items_count(ht
->percpu_count
);
1150 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1151 unsigned long *count
,
1152 unsigned long *removed
)
1154 struct rcu_table
*t
;
1155 struct cds_lfht_node
*node
, *next
;
1156 struct _cds_lfht_node
*lookup
;
1157 unsigned long nr_dummy
= 0;
1162 t
= rcu_dereference(ht
->t
);
1163 /* Count non-dummy nodes in the table */
1164 lookup
= &t
->tbl
[0]->nodes
[0];
1165 node
= (struct cds_lfht_node
*) lookup
;
1167 next
= rcu_dereference(node
->p
.next
);
1168 if (is_removed(next
)) {
1169 assert(!is_dummy(next
));
1171 } else if (!is_dummy(next
))
1175 node
= clear_flag(next
);
1177 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1180 /* called with resize mutex held */
1182 void _do_cds_lfht_grow(struct cds_lfht
*ht
, struct rcu_table
*old_t
,
1183 unsigned long old_size
, unsigned long new_size
)
1185 unsigned long old_order
, new_order
;
1186 struct rcu_table
*new_t
;
1188 old_order
= get_count_order_ulong(old_size
) + 1;
1189 new_order
= get_count_order_ulong(new_size
) + 1;
1190 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1191 old_size
, old_order
, new_size
, new_order
);
1192 new_t
= malloc(sizeof(struct cds_lfht
)
1193 + (new_order
* sizeof(struct rcu_level
*)));
1194 assert(new_size
> old_size
);
1195 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1196 old_order
* sizeof(struct rcu_level
*));
1197 init_table(ht
, new_t
, old_order
, new_order
- old_order
);
1198 /* Changing table and size atomically wrt lookups */
1199 rcu_assign_pointer(ht
->t
, new_t
);
1200 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1203 /* called with resize mutex held */
1205 void _do_cds_lfht_shrink(struct cds_lfht
*ht
, struct rcu_table
*old_t
,
1206 unsigned long old_size
, unsigned long new_size
)
1208 unsigned long old_order
, new_order
;
1209 struct rcu_table
*new_t
;
1211 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1212 old_order
= get_count_order_ulong(old_size
) + 1;
1213 new_order
= get_count_order_ulong(new_size
) + 1;
1214 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1215 old_size
, old_order
, new_size
, new_order
);
1216 new_t
= malloc(sizeof(struct cds_lfht
)
1217 + (new_order
* sizeof(struct rcu_level
*)));
1218 assert(new_size
< old_size
);
1219 memcpy(&new_t
->tbl
, &old_t
->tbl
,
1220 new_order
* sizeof(struct rcu_level
*));
1221 new_t
->size
= !new_order
? 1 : (1UL << (new_order
- 1));
1222 assert(new_t
->size
== new_size
);
1223 new_t
->resize_target
= new_t
->size
;
1224 new_t
->resize_initiated
= 0;
1226 /* Changing table and size atomically wrt lookups */
1227 rcu_assign_pointer(ht
->t
, new_t
);
1230 * We need to wait for all add operations to reach Q.S. (and
1231 * thus use the new table for lookups) before we can start
1232 * releasing the old dummy nodes. Otherwise their lookup will
1233 * return a logically removed node as insert position.
1235 ht
->cds_lfht_synchronize_rcu();
1237 /* Unlink and remove all now-unused dummy node pointers. */
1238 fini_table(ht
, old_t
, new_order
, old_order
- new_order
);
1239 ht
->cds_lfht_call_rcu(&old_t
->head
, cds_lfht_free_table_cb
);
1243 /* called with resize mutex held */
1245 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1247 unsigned long new_size
, old_size
;
1248 struct rcu_table
*old_t
;
1251 old_size
= old_t
->size
;
1252 new_size
= CMM_LOAD_SHARED(old_t
->resize_target
);
1253 if (old_size
< new_size
)
1254 _do_cds_lfht_grow(ht
, old_t
, old_size
, new_size
);
1255 else if (old_size
> new_size
)
1256 _do_cds_lfht_shrink(ht
, old_t
, old_size
, new_size
);
1258 CMM_STORE_SHARED(old_t
->resize_initiated
, 0);
1262 unsigned long resize_target_update(struct rcu_table
*t
,
1265 return _uatomic_max(&t
->resize_target
,
1266 t
->size
<< growth_order
);
1270 void resize_target_update_count(struct rcu_table
*t
,
1271 unsigned long count
)
1273 count
= max(count
, MIN_TABLE_SIZE
);
1274 uatomic_set(&t
->resize_target
, count
);
1277 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1279 struct rcu_table
*t
= rcu_dereference(ht
->t
);
1281 resize_target_update_count(t
, new_size
);
1282 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1283 pthread_mutex_lock(&ht
->resize_mutex
);
1284 _do_cds_lfht_resize(ht
);
1285 pthread_mutex_unlock(&ht
->resize_mutex
);
1289 void do_resize_cb(struct rcu_head
*head
)
1291 struct rcu_resize_work
*work
=
1292 caa_container_of(head
, struct rcu_resize_work
, head
);
1293 struct cds_lfht
*ht
= work
->ht
;
1295 pthread_mutex_lock(&ht
->resize_mutex
);
1296 _do_cds_lfht_resize(ht
);
1297 pthread_mutex_unlock(&ht
->resize_mutex
);
1299 cmm_smp_mb(); /* finish resize before decrement */
1300 uatomic_dec(&ht
->in_progress_resize
);
1304 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, struct rcu_table
*t
, int growth
)
1306 struct rcu_resize_work
*work
;
1307 unsigned long target_size
;
1309 target_size
= resize_target_update(t
, growth
);
1310 if (!CMM_LOAD_SHARED(t
->resize_initiated
) && t
->size
< target_size
) {
1311 uatomic_inc(&ht
->in_progress_resize
);
1312 cmm_smp_mb(); /* increment resize count before calling it */
1313 work
= malloc(sizeof(*work
));
1315 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1316 CMM_STORE_SHARED(t
->resize_initiated
, 1);
1320 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1323 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, struct rcu_table
*t
,
1324 unsigned long count
)
1326 struct rcu_resize_work
*work
;
1328 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1330 resize_target_update_count(t
, count
);
1331 if (!CMM_LOAD_SHARED(t
->resize_initiated
)) {
1332 uatomic_inc(&ht
->in_progress_resize
);
1333 cmm_smp_mb(); /* increment resize count before calling it */
1334 work
= malloc(sizeof(*work
));
1336 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1337 CMM_STORE_SHARED(t
->resize_initiated
, 1);