4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
32 * Some specificities of this Lock-Free Resizable RCU Hash Table
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
100 * A bit of ascii art explanation:
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
105 * This shows the nodes for a small table ordered by reversed bits:
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
123 * 1 | 1 001 100 <- <-
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
128 * 3 -> | | | 4 100 001 | |
144 #include <urcu-call-rcu.h>
145 #include <urcu/arch.h>
146 #include <urcu/uatomic.h>
147 #include <urcu/jhash.h>
148 #include <urcu/compiler.h>
149 #include <urcu/rculfhash.h>
154 #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
156 #define dbg_printf(fmt, args...)
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
165 #define COUNT_COMMIT_ORDER 10
166 #define CHAIN_LEN_TARGET 1
167 #define CHAIN_LEN_RESIZE_THRESHOLD 3
170 * Define the minimum table size.
172 //#define MIN_TABLE_SIZE 128
173 #define MIN_TABLE_SIZE 1
175 #if (CAA_BITS_PER_LONG == 32)
176 #define MAX_TABLE_ORDER 32
178 #define MAX_TABLE_ORDER 64
182 #define min(a, b) ((a) < (b) ? (a) : (b))
186 #define max(a, b) ((a) > (b) ? (a) : (b))
190 * The removed flag needs to be updated atomically with the pointer.
191 * The dummy flag does not require to be updated atomically with the
192 * pointer, but it is added as a pointer low bit flag to save space.
194 #define REMOVED_FLAG (1UL << 0)
195 #define DUMMY_FLAG (1UL << 1)
196 #define FLAGS_MASK ((1UL << 2) - 1)
198 /* Value of the end pointer. Should not interact with flags. */
199 #define END_VALUE 0x4
201 struct ht_items_count
{
202 unsigned long add
, remove
;
203 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
206 struct rcu_head head
;
207 struct _cds_lfht_node nodes
[0];
211 unsigned long size
; /* always a power of 2, shared (RCU) */
212 unsigned long resize_target
;
213 int resize_initiated
;
214 struct rcu_level
*tbl
[MAX_TABLE_ORDER
];
219 cds_lfht_hash_fct hash_fct
;
220 cds_lfht_compare_fct compare_fct
;
221 unsigned long hash_seed
;
224 * We need to put the work threads offline (QSBR) when taking this
225 * mutex, because we use synchronize_rcu within this mutex critical
226 * section, which waits on read-side critical sections, and could
227 * therefore cause grace-period deadlock if we hold off RCU G.P.
230 pthread_mutex_t resize_mutex
; /* resize mutex: add/del mutex */
231 unsigned int in_progress_resize
, in_progress_destroy
;
232 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
233 void (*func
)(struct rcu_head
*head
));
234 void (*cds_lfht_synchronize_rcu
)(void);
235 void (*cds_lfht_rcu_read_lock
)(void);
236 void (*cds_lfht_rcu_read_unlock
)(void);
237 void (*cds_lfht_rcu_thread_offline
)(void);
238 void (*cds_lfht_rcu_thread_online
)(void);
239 unsigned long count
; /* global approximate item count */
240 struct ht_items_count
*percpu_count
; /* per-cpu item count */
243 struct rcu_resize_work
{
244 struct rcu_head head
;
249 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
251 struct cds_lfht_node
*node
,
252 int unique
, int dummy
);
255 * Algorithm to reverse bits in a word by lookup table, extended to
258 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
259 * Originally from Public Domain.
262 static const uint8_t BitReverseTable256
[256] =
264 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
265 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
266 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
267 R6(0), R6(2), R6(1), R6(3)
274 uint8_t bit_reverse_u8(uint8_t v
)
276 return BitReverseTable256
[v
];
279 static __attribute__((unused
))
280 uint32_t bit_reverse_u32(uint32_t v
)
282 return ((uint32_t) bit_reverse_u8(v
) << 24) |
283 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
284 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
285 ((uint32_t) bit_reverse_u8(v
>> 24));
288 static __attribute__((unused
))
289 uint64_t bit_reverse_u64(uint64_t v
)
291 return ((uint64_t) bit_reverse_u8(v
) << 56) |
292 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
293 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
294 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
295 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
296 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
297 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
298 ((uint64_t) bit_reverse_u8(v
>> 56));
302 unsigned long bit_reverse_ulong(unsigned long v
)
304 #if (CAA_BITS_PER_LONG == 32)
305 return bit_reverse_u32(v
);
307 return bit_reverse_u64(v
);
312 * fls: returns the position of the most significant bit.
313 * Returns 0 if no bit is set, else returns the position of the most
314 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
316 #if defined(__i386) || defined(__x86_64)
318 unsigned int fls_u32(uint32_t x
)
326 : "=r" (r
) : "rm" (x
));
332 #if defined(__x86_64)
334 unsigned int fls_u64(uint64_t x
)
342 : "=r" (r
) : "rm" (x
));
349 static __attribute__((unused
))
350 unsigned int fls_u64(uint64_t x
)
357 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
361 if (!(x
& 0xFFFF000000000000ULL
)) {
365 if (!(x
& 0xFF00000000000000ULL
)) {
369 if (!(x
& 0xF000000000000000ULL
)) {
373 if (!(x
& 0xC000000000000000ULL
)) {
377 if (!(x
& 0x8000000000000000ULL
)) {
386 static __attribute__((unused
))
387 unsigned int fls_u32(uint32_t x
)
393 if (!(x
& 0xFFFF0000U
)) {
397 if (!(x
& 0xFF000000U
)) {
401 if (!(x
& 0xF0000000U
)) {
405 if (!(x
& 0xC0000000U
)) {
409 if (!(x
& 0x80000000U
)) {
417 unsigned int fls_ulong(unsigned long x
)
419 #if (CAA_BITS_PER_lONG == 32)
426 int get_count_order_u32(uint32_t x
)
430 order
= fls_u32(x
) - 1;
436 int get_count_order_ulong(unsigned long x
)
440 order
= fls_ulong(x
) - 1;
447 #define poison_free(ptr) \
449 memset(ptr, 0x42, sizeof(*(ptr))); \
453 #define poison_free(ptr) free(ptr)
457 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
);
460 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
461 * available, then we support hash table item accounting.
462 * In the unfortunate event the number of CPUs reported would be
463 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
465 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
468 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
469 unsigned long count
);
471 static long nr_cpus_mask
= -1;
474 struct ht_items_count
*alloc_per_cpu_items_count(void)
476 struct ht_items_count
*count
;
478 switch (nr_cpus_mask
) {
485 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
491 * round up number of CPUs to next power of two, so we
492 * can use & for modulo.
494 maxcpus
= 1UL << get_count_order_ulong(maxcpus
);
495 nr_cpus_mask
= maxcpus
- 1;
499 return calloc(nr_cpus_mask
+ 1, sizeof(*count
));
504 void free_per_cpu_items_count(struct ht_items_count
*count
)
514 assert(nr_cpus_mask
>= 0);
515 cpu
= sched_getcpu();
516 if (unlikely(cpu
< 0))
519 return cpu
& nr_cpus_mask
;
523 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
525 unsigned long percpu_count
;
528 if (unlikely(!ht
->percpu_count
))
531 if (unlikely(cpu
< 0))
533 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].add
, 1);
534 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
537 dbg_printf("add percpu %lu\n", percpu_count
);
538 count
= uatomic_add_return(&ht
->count
,
539 1UL << COUNT_COMMIT_ORDER
);
541 if (!(count
& (count
- 1))) {
542 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
544 dbg_printf("add set global %lu\n", count
);
545 cds_lfht_resize_lazy_count(ht
, size
,
546 count
>> (CHAIN_LEN_TARGET
- 1));
552 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
554 unsigned long percpu_count
;
557 if (unlikely(!ht
->percpu_count
))
560 if (unlikely(cpu
< 0))
562 percpu_count
= uatomic_add_return(&ht
->percpu_count
[cpu
].remove
, -1);
563 if (unlikely(!(percpu_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))) {
566 dbg_printf("remove percpu %lu\n", percpu_count
);
567 count
= uatomic_add_return(&ht
->count
,
568 -(1UL << COUNT_COMMIT_ORDER
));
570 if (!(count
& (count
- 1))) {
571 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
573 dbg_printf("remove set global %lu\n", count
);
574 cds_lfht_resize_lazy_count(ht
, size
,
575 count
>> (CHAIN_LEN_TARGET
- 1));
580 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
582 static const long nr_cpus_mask
= -1;
585 struct ht_items_count
*alloc_per_cpu_items_count(void)
591 void free_per_cpu_items_count(struct ht_items_count
*count
)
596 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
)
601 void ht_count_remove(struct cds_lfht
*ht
, unsigned long size
)
605 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
609 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
613 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
615 count
= uatomic_read(&ht
->count
);
617 * Use bucket-local length for small table expand and for
618 * environments lacking per-cpu data support.
620 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
623 dbg_printf("WARNING: large chain length: %u.\n",
625 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
626 cds_lfht_resize_lazy(ht
, size
,
627 get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
631 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
633 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
637 int is_removed(struct cds_lfht_node
*node
)
639 return ((unsigned long) node
) & REMOVED_FLAG
;
643 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
645 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
649 int is_dummy(struct cds_lfht_node
*node
)
651 return ((unsigned long) node
) & DUMMY_FLAG
;
655 struct cds_lfht_node
*flag_dummy(struct cds_lfht_node
*node
)
657 return (struct cds_lfht_node
*) (((unsigned long) node
) | DUMMY_FLAG
);
661 struct cds_lfht_node
*get_end(void)
663 return (struct cds_lfht_node
*) END_VALUE
;
667 int is_end(struct cds_lfht_node
*node
)
669 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
673 unsigned long _uatomic_max(unsigned long *ptr
, unsigned long v
)
675 unsigned long old1
, old2
;
677 old1
= uatomic_read(ptr
);
682 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
687 void cds_lfht_free_level(struct rcu_head
*head
)
689 struct rcu_level
*l
=
690 caa_container_of(head
, struct rcu_level
, head
);
695 * Remove all logically deleted nodes from a bucket up to a certain node key.
698 int _cds_lfht_gc_bucket(struct cds_lfht_node
*dummy
, struct cds_lfht_node
*node
)
700 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
702 assert(!is_dummy(dummy
));
703 assert(!is_removed(dummy
));
704 assert(!is_dummy(node
));
705 assert(!is_removed(node
));
708 /* We can always skip the dummy node initially */
709 iter
= rcu_dereference(iter_prev
->p
.next
);
710 if (unlikely(iter
== NULL
)) {
712 * We are executing concurrently with a hash table
713 * expand, so we see a dummy node with NULL next value.
714 * Help expand by linking this node into the list and
719 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
721 * We should never be called with dummy (start of chain)
722 * and logically removed node (end of path compression
723 * marker) being the actual same node. This would be a
724 * bug in the algorithm implementation.
726 assert(dummy
!= node
);
728 assert(iter
!= NULL
);
729 if (unlikely(is_end(iter
)))
731 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
733 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
734 if (likely(is_removed(next
)))
736 iter_prev
= clear_flag(iter
);
739 assert(!is_removed(iter
));
741 new_next
= flag_dummy(clear_flag(next
));
743 new_next
= clear_flag(next
);
744 assert(new_next
!= NULL
);
745 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
751 struct cds_lfht_node
*_cds_lfht_add(struct cds_lfht
*ht
,
753 struct cds_lfht_node
*node
,
754 int unique
, int dummy
)
756 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
758 struct _cds_lfht_node
*lookup
;
759 unsigned long hash
, index
, order
;
762 assert(!is_dummy(node
));
763 assert(!is_removed(node
));
766 node
->p
.next
= flag_dummy(get_end());
767 return node
; /* Initial first add (head) */
769 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
771 uint32_t chain_len
= 0;
774 * iter_prev points to the non-removed node prior to the
777 index
= hash
& (size
- 1);
778 order
= get_count_order_ulong(index
+ 1);
779 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& ((!order
? 0 : (1UL << (order
- 1))) - 1)];
780 iter_prev
= (struct cds_lfht_node
*) lookup
;
781 /* We can always skip the dummy node initially */
782 iter
= rcu_dereference(iter_prev
->p
.next
);
783 if (unlikely(iter
== NULL
)) {
785 * We are executing concurrently with a hash table
786 * expand, so we see a dummy node with NULL next value.
787 * Help expand by linking this node into the list and
790 (void) _cds_lfht_add(ht
, size
>> 1, iter_prev
, 0, 1);
791 continue; /* retry */
793 assert(iter_prev
->p
.reverse_hash
<= node
->p
.reverse_hash
);
795 assert(iter
!= NULL
);
797 * When adding a dummy node, we allow concurrent
798 * add/removal to help. If we find the dummy node in
799 * place, skip its insertion.
801 if (unlikely(dummy
&& clear_flag(iter
) == node
))
803 if (unlikely(is_end(iter
)))
805 if (likely(clear_flag(iter
)->p
.reverse_hash
> node
->p
.reverse_hash
))
807 next
= rcu_dereference(clear_flag(iter
)->p
.next
);
808 if (unlikely(is_removed(next
)))
812 && !ht
->compare_fct(node
->key
, node
->key_len
,
813 clear_flag(iter
)->key
,
814 clear_flag(iter
)->key_len
))
815 return clear_flag(iter
);
816 /* Only account for identical reverse hash once */
817 if (iter_prev
->p
.reverse_hash
!= clear_flag(iter
)->p
.reverse_hash
819 check_resize(ht
, size
, ++chain_len
);
820 iter_prev
= clear_flag(iter
);
824 assert(node
!= clear_flag(iter
));
825 assert(!is_removed(iter_prev
));
826 assert(!is_removed(iter
));
827 assert(iter_prev
!= node
);
829 node
->p
.next
= clear_flag(iter
);
832 * Dummy node insertion is performed concurrently (help
833 * scheme). We try to link its next node, and if this
834 * succeeds, it _means_ it's us who link this dummy node
835 * into the table. force_dummy is set as soon as we
836 * succeed this cmpxchg within this function.
839 if (uatomic_cmpxchg(&node
->p
.next
, NULL
,
840 flag_dummy(clear_flag(iter
))) != NULL
) {
845 node
->p
.next
= flag_dummy(clear_flag(iter
));
849 new_node
= flag_dummy(node
);
852 assert(new_node
!= NULL
);
853 if (uatomic_cmpxchg(&iter_prev
->p
.next
, iter
,
855 continue; /* retry */
859 assert(!is_removed(iter
));
861 new_next
= flag_dummy(clear_flag(next
));
863 new_next
= clear_flag(next
);
864 assert(new_next
!= NULL
);
865 (void) uatomic_cmpxchg(&iter_prev
->p
.next
, iter
, new_next
);
869 /* Garbage collect logically removed nodes in the bucket */
870 index
= hash
& (size
- 1);
871 order
= get_count_order_ulong(index
+ 1);
872 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
873 dummy_node
= (struct cds_lfht_node
*) lookup
;
874 if (_cds_lfht_gc_bucket(dummy_node
, node
)) {
876 (void) _cds_lfht_add(ht
, size
>> 1, dummy_node
, 0, 1);
877 goto gc_end
; /* retry */
883 int _cds_lfht_remove(struct cds_lfht
*ht
, unsigned long size
,
884 struct cds_lfht_node
*node
,
887 struct cds_lfht_node
*dummy
, *next
, *old
;
888 struct _cds_lfht_node
*lookup
;
890 unsigned long hash
, index
, order
;
892 /* logically delete the node */
893 assert(!is_dummy(node
));
894 assert(!is_removed(node
));
895 old
= rcu_dereference(node
->p
.next
);
898 if (unlikely(is_removed(next
)))
901 assert(is_dummy(next
));
903 assert(!is_dummy(next
));
904 assert(next
!= NULL
);
905 old
= uatomic_cmpxchg(&node
->p
.next
, next
,
907 } while (old
!= next
);
909 /* We performed the (logical) deletion. */
913 * Ensure that the node is not visible to readers anymore: lookup for
914 * the node, and remove it (along with any other logically removed node)
918 hash
= bit_reverse_ulong(node
->p
.reverse_hash
);
920 index
= hash
& (size
- 1);
921 order
= get_count_order_ulong(index
+ 1);
922 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1)) - 1))];
923 dummy
= (struct cds_lfht_node
*) lookup
;
924 if (_cds_lfht_gc_bucket(dummy
, node
)) {
926 (void) _cds_lfht_add(ht
, size
>> 1, dummy
, 0, 1);
927 goto gc_retry
; /* retry */
931 * Only the flagging action indicated that we (and no other)
932 * removed the node from the hash.
935 assert(is_removed(rcu_dereference(node
->p
.next
)));
942 void init_table_hash(struct cds_lfht
*ht
, unsigned long i
,
947 for (j
= 0; j
< len
; j
++) {
948 struct cds_lfht_node
*new_node
=
949 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
951 dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
952 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
953 new_node
->p
.reverse_hash
=
954 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
955 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
961 * Holding RCU read lock to protect _cds_lfht_add against memory
962 * reclaim that could be performed by other call_rcu worker threads (ABA
966 void init_table_link(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
970 ht
->cds_lfht_rcu_thread_online();
971 ht
->cds_lfht_rcu_read_lock();
972 for (j
= 0; j
< len
; j
++) {
973 struct cds_lfht_node
*new_node
=
974 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
976 dbg_printf("init link: i %lu j %lu hash %lu\n",
977 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
978 (void) _cds_lfht_add(ht
, !i
? 0 : (1UL << (i
- 1)),
980 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
983 ht
->cds_lfht_rcu_read_unlock();
984 ht
->cds_lfht_rcu_thread_offline();
988 void init_table(struct cds_lfht
*ht
,
989 unsigned long first_order
, unsigned long len_order
)
991 unsigned long i
, end_order
;
993 dbg_printf("init table: first_order %lu end_order %lu\n",
994 first_order
, first_order
+ len_order
);
995 end_order
= first_order
+ len_order
;
996 for (i
= first_order
; i
< end_order
; i
++) {
999 len
= !i
? 1 : 1UL << (i
- 1);
1000 dbg_printf("init order %lu len: %lu\n", i
, len
);
1002 /* Stop expand if the resize target changes under us */
1003 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) < (!i
? 1 : (1UL << i
)))
1006 ht
->t
.tbl
[i
] = calloc(1, sizeof(struct rcu_level
)
1007 + (len
* sizeof(struct _cds_lfht_node
)));
1009 /* Set all dummy nodes reverse hash values for a level */
1010 init_table_hash(ht
, i
, len
);
1013 * Update table size. At this point, concurrent add/remove see
1014 * dummy nodes with correctly initialized reverse hash value,
1015 * but with NULL next pointers. If they do, they can help us
1016 * link the dummy nodes into the list and retry.
1018 cmm_smp_wmb(); /* populate data before RCU size */
1019 CMM_STORE_SHARED(ht
->t
.size
, !i
? 1 : (1UL << i
));
1022 * Link all dummy nodes into the table. Concurrent
1023 * add/remove are helping us.
1025 init_table_link(ht
, i
, len
);
1027 dbg_printf("init new size: %lu\n", !i
? 1 : (1UL << i
));
1028 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1034 * Holding RCU read lock to protect _cds_lfht_remove against memory
1035 * reclaim that could be performed by other call_rcu worker threads (ABA
1037 * For a single level, we logically remove and garbage collect each node.
1039 * As a design choice, we perform logical removal and garbage collection on a
1040 * node-per-node basis to simplify this algorithm. We also assume keeping good
1041 * cache locality of the operation would overweight possible performance gain
1042 * that could be achieved by batching garbage collection for multiple levels.
1043 * However, this would have to be justified by benchmarks.
1045 * Concurrent removal and add operations are helping us perform garbage
1046 * collection of logically removed nodes. We guarantee that all logically
1047 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1048 * invoked to free a hole level of dummy nodes (after a grace period).
1050 * Logical removal and garbage collection can therefore be done in batch or on a
1051 * node-per-node basis, as long as the guarantee above holds.
1054 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1058 ht
->cds_lfht_rcu_thread_online();
1059 ht
->cds_lfht_rcu_read_lock();
1060 for (j
= 0; j
< len
; j
++) {
1061 struct cds_lfht_node
*fini_node
=
1062 (struct cds_lfht_node
*) &ht
->t
.tbl
[i
]->nodes
[j
];
1064 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1065 i
, j
, !i
? 0 : (1UL << (i
- 1)) + j
);
1066 fini_node
->p
.reverse_hash
=
1067 bit_reverse_ulong(!i
? 0 : (1UL << (i
- 1)) + j
);
1068 (void) _cds_lfht_remove(ht
, !i
? 0 : (1UL << (i
- 1)),
1070 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1073 ht
->cds_lfht_rcu_read_unlock();
1074 ht
->cds_lfht_rcu_thread_offline();
1078 void fini_table(struct cds_lfht
*ht
,
1079 unsigned long first_order
, unsigned long len_order
)
1083 dbg_printf("fini table: first_order %lu end_order %lu\n",
1084 first_order
, first_order
+ len_order
);
1085 end_order
= first_order
+ len_order
;
1086 assert(first_order
> 0);
1087 for (i
= end_order
- 1; i
>= first_order
; i
--) {
1090 len
= !i
? 1 : 1UL << (i
- 1);
1091 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1093 /* Stop shrink if the resize target changes under us */
1094 if (CMM_LOAD_SHARED(ht
->t
.resize_target
) > (1UL << (i
- 1)))
1097 cmm_smp_wmb(); /* populate data before RCU size */
1098 CMM_STORE_SHARED(ht
->t
.size
, 1UL << (i
- 1));
1101 * We need to wait for all add operations to reach Q.S. (and
1102 * thus use the new table for lookups) before we can start
1103 * releasing the old dummy nodes. Otherwise their lookup will
1104 * return a logically removed node as insert position.
1106 ht
->cds_lfht_synchronize_rcu();
1109 * Set "removed" flag in dummy nodes about to be removed.
1110 * Unlink all now-logically-removed dummy node pointers.
1111 * Concurrent add/remove operation are helping us doing
1114 remove_table(ht
, i
, len
);
1116 ht
->cds_lfht_call_rcu(&ht
->t
.tbl
[i
]->head
, cds_lfht_free_level
);
1118 dbg_printf("fini new size: %lu\n", 1UL << i
);
1119 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1124 struct cds_lfht
*cds_lfht_new(cds_lfht_hash_fct hash_fct
,
1125 cds_lfht_compare_fct compare_fct
,
1126 unsigned long hash_seed
,
1127 unsigned long init_size
,
1129 void (*cds_lfht_call_rcu
)(struct rcu_head
*head
,
1130 void (*func
)(struct rcu_head
*head
)),
1131 void (*cds_lfht_synchronize_rcu
)(void),
1132 void (*cds_lfht_rcu_read_lock
)(void),
1133 void (*cds_lfht_rcu_read_unlock
)(void),
1134 void (*cds_lfht_rcu_thread_offline
)(void),
1135 void (*cds_lfht_rcu_thread_online
)(void))
1137 struct cds_lfht
*ht
;
1138 unsigned long order
;
1140 /* init_size must be power of two */
1141 if (init_size
&& (init_size
& (init_size
- 1)))
1143 ht
= calloc(1, sizeof(struct cds_lfht
));
1144 ht
->hash_fct
= hash_fct
;
1145 ht
->compare_fct
= compare_fct
;
1146 ht
->hash_seed
= hash_seed
;
1147 ht
->cds_lfht_call_rcu
= cds_lfht_call_rcu
;
1148 ht
->cds_lfht_synchronize_rcu
= cds_lfht_synchronize_rcu
;
1149 ht
->cds_lfht_rcu_read_lock
= cds_lfht_rcu_read_lock
;
1150 ht
->cds_lfht_rcu_read_unlock
= cds_lfht_rcu_read_unlock
;
1151 ht
->cds_lfht_rcu_thread_offline
= cds_lfht_rcu_thread_offline
;
1152 ht
->cds_lfht_rcu_thread_online
= cds_lfht_rcu_thread_online
;
1153 ht
->percpu_count
= alloc_per_cpu_items_count();
1154 /* this mutex should not nest in read-side C.S. */
1155 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1156 order
= get_count_order_ulong(max(init_size
, MIN_TABLE_SIZE
)) + 1;
1158 ht
->cds_lfht_rcu_thread_offline();
1159 pthread_mutex_lock(&ht
->resize_mutex
);
1160 ht
->t
.resize_target
= 1UL << (order
- 1);
1161 init_table(ht
, 0, order
);
1162 pthread_mutex_unlock(&ht
->resize_mutex
);
1163 ht
->cds_lfht_rcu_thread_online();
1167 struct cds_lfht_node
*cds_lfht_lookup(struct cds_lfht
*ht
, void *key
, size_t key_len
)
1169 struct cds_lfht_node
*node
, *next
, *dummy_node
;
1170 struct _cds_lfht_node
*lookup
;
1171 unsigned long hash
, reverse_hash
, index
, order
, size
;
1173 hash
= ht
->hash_fct(key
, key_len
, ht
->hash_seed
);
1174 reverse_hash
= bit_reverse_ulong(hash
);
1177 size
= rcu_dereference(ht
->t
.size
);
1178 index
= hash
& (size
- 1);
1179 order
= get_count_order_ulong(index
+ 1);
1180 lookup
= &ht
->t
.tbl
[order
]->nodes
[index
& (!order
? 0 : ((1UL << (order
- 1))) - 1)];
1181 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
1182 hash
, index
, order
, index
& (!order
? 0 : ((1UL << (order
- 1)) - 1)));
1183 dummy_node
= (struct cds_lfht_node
*) lookup
;
1184 /* We can always skip the dummy node initially */
1185 node
= rcu_dereference(dummy_node
->p
.next
);
1186 if (unlikely(node
== NULL
)) {
1188 * We are executing concurrently with a hash table
1189 * expand, so we see a dummy node with NULL next value.
1190 * Help expand by linking this node into the list and
1193 (void) _cds_lfht_add(ht
, size
>> 1, dummy_node
, 0, 1);
1194 goto restart
; /* retry */
1196 node
= clear_flag(node
);
1198 if (unlikely(is_end(node
))) {
1202 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1206 next
= rcu_dereference(node
->p
.next
);
1207 if (likely(!is_removed(next
))
1209 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1212 node
= clear_flag(next
);
1214 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1218 struct cds_lfht_node
*cds_lfht_next(struct cds_lfht
*ht
,
1219 struct cds_lfht_node
*node
)
1221 struct cds_lfht_node
*next
;
1222 unsigned long reverse_hash
;
1226 reverse_hash
= node
->p
.reverse_hash
;
1228 key_len
= node
->key_len
;
1229 next
= rcu_dereference(node
->p
.next
);
1230 node
= clear_flag(next
);
1233 if (unlikely(is_end(node
))) {
1237 if (unlikely(node
->p
.reverse_hash
> reverse_hash
)) {
1241 next
= rcu_dereference(node
->p
.next
);
1242 if (likely(!is_removed(next
))
1244 && likely(!ht
->compare_fct(node
->key
, node
->key_len
, key
, key_len
))) {
1247 node
= clear_flag(next
);
1249 assert(!node
|| !is_dummy(rcu_dereference(node
->p
.next
)));
1253 void cds_lfht_add(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1255 unsigned long hash
, size
;
1257 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1258 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1260 size
= rcu_dereference(ht
->t
.size
);
1261 (void) _cds_lfht_add(ht
, size
, node
, 0, 0);
1262 ht_count_add(ht
, size
);
1265 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1266 struct cds_lfht_node
*node
)
1268 unsigned long hash
, size
;
1269 struct cds_lfht_node
*ret
;
1271 hash
= ht
->hash_fct(node
->key
, node
->key_len
, ht
->hash_seed
);
1272 node
->p
.reverse_hash
= bit_reverse_ulong((unsigned long) hash
);
1274 size
= rcu_dereference(ht
->t
.size
);
1275 ret
= _cds_lfht_add(ht
, size
, node
, 1, 0);
1277 ht_count_add(ht
, size
);
1281 int cds_lfht_remove(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1286 size
= rcu_dereference(ht
->t
.size
);
1287 ret
= _cds_lfht_remove(ht
, size
, node
, 0);
1289 ht_count_remove(ht
, size
);
1294 int cds_lfht_delete_dummy(struct cds_lfht
*ht
)
1296 struct cds_lfht_node
*node
;
1297 struct _cds_lfht_node
*lookup
;
1298 unsigned long order
, i
, size
;
1300 /* Check that the table is empty */
1301 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1302 node
= (struct cds_lfht_node
*) lookup
;
1304 node
= clear_flag(node
)->p
.next
;
1305 if (!is_dummy(node
))
1307 assert(!is_removed(node
));
1308 } while (!is_end(node
));
1310 * size accessed without rcu_dereference because hash table is
1314 /* Internal sanity check: all nodes left should be dummy */
1315 for (order
= 0; order
< get_count_order_ulong(size
) + 1; order
++) {
1318 len
= !order
? 1 : 1UL << (order
- 1);
1319 for (i
= 0; i
< len
; i
++) {
1320 dbg_printf("delete order %lu i %lu hash %lu\n",
1322 bit_reverse_ulong(ht
->t
.tbl
[order
]->nodes
[i
].reverse_hash
));
1323 assert(is_dummy(ht
->t
.tbl
[order
]->nodes
[i
].next
));
1325 poison_free(ht
->t
.tbl
[order
]);
1331 * Should only be called when no more concurrent readers nor writers can
1332 * possibly access the table.
1334 int cds_lfht_destroy(struct cds_lfht
*ht
)
1338 /* Wait for in-flight resize operations to complete */
1339 CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1340 while (uatomic_read(&ht
->in_progress_resize
))
1341 poll(NULL
, 0, 100); /* wait for 100ms */
1342 ret
= cds_lfht_delete_dummy(ht
);
1345 free_per_cpu_items_count(ht
->percpu_count
);
1350 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1351 unsigned long *count
,
1352 unsigned long *removed
)
1354 struct cds_lfht_node
*node
, *next
;
1355 struct _cds_lfht_node
*lookup
;
1356 unsigned long nr_dummy
= 0;
1361 /* Count non-dummy nodes in the table */
1362 lookup
= &ht
->t
.tbl
[0]->nodes
[0];
1363 node
= (struct cds_lfht_node
*) lookup
;
1365 next
= rcu_dereference(node
->p
.next
);
1366 if (is_removed(next
)) {
1367 assert(!is_dummy(next
));
1369 } else if (!is_dummy(next
))
1373 node
= clear_flag(next
);
1374 } while (!is_end(node
));
1375 dbg_printf("number of dummy nodes: %lu\n", nr_dummy
);
1378 /* called with resize mutex held */
1380 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1381 unsigned long old_size
, unsigned long new_size
)
1383 unsigned long old_order
, new_order
;
1385 old_order
= get_count_order_ulong(old_size
) + 1;
1386 new_order
= get_count_order_ulong(new_size
) + 1;
1387 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1388 old_size
, old_order
, new_size
, new_order
);
1389 assert(new_size
> old_size
);
1390 init_table(ht
, old_order
, new_order
- old_order
);
1393 /* called with resize mutex held */
1395 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1396 unsigned long old_size
, unsigned long new_size
)
1398 unsigned long old_order
, new_order
;
1400 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1401 old_order
= get_count_order_ulong(old_size
) + 1;
1402 new_order
= get_count_order_ulong(new_size
) + 1;
1403 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1404 old_size
, old_order
, new_size
, new_order
);
1405 assert(new_size
< old_size
);
1407 /* Remove and unlink all dummy nodes to remove. */
1408 fini_table(ht
, new_order
, old_order
- new_order
);
1412 /* called with resize mutex held */
1414 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1416 unsigned long new_size
, old_size
;
1419 * Resize table, re-do if the target size has changed under us.
1422 ht
->t
.resize_initiated
= 1;
1423 old_size
= ht
->t
.size
;
1424 new_size
= CMM_LOAD_SHARED(ht
->t
.resize_target
);
1425 if (old_size
< new_size
)
1426 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1427 else if (old_size
> new_size
)
1428 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1429 ht
->t
.resize_initiated
= 0;
1430 /* write resize_initiated before read resize_target */
1432 } while (ht
->t
.size
!= CMM_LOAD_SHARED(ht
->t
.resize_target
));
1436 unsigned long resize_target_update(struct cds_lfht
*ht
, unsigned long size
,
1439 return _uatomic_max(&ht
->t
.resize_target
,
1440 size
<< growth_order
);
1444 void resize_target_update_count(struct cds_lfht
*ht
,
1445 unsigned long count
)
1447 count
= max(count
, MIN_TABLE_SIZE
);
1448 uatomic_set(&ht
->t
.resize_target
, count
);
1451 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1453 resize_target_update_count(ht
, new_size
);
1454 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1455 ht
->cds_lfht_rcu_thread_offline();
1456 pthread_mutex_lock(&ht
->resize_mutex
);
1457 _do_cds_lfht_resize(ht
);
1458 pthread_mutex_unlock(&ht
->resize_mutex
);
1459 ht
->cds_lfht_rcu_thread_online();
1463 void do_resize_cb(struct rcu_head
*head
)
1465 struct rcu_resize_work
*work
=
1466 caa_container_of(head
, struct rcu_resize_work
, head
);
1467 struct cds_lfht
*ht
= work
->ht
;
1469 ht
->cds_lfht_rcu_thread_offline();
1470 pthread_mutex_lock(&ht
->resize_mutex
);
1471 _do_cds_lfht_resize(ht
);
1472 pthread_mutex_unlock(&ht
->resize_mutex
);
1473 ht
->cds_lfht_rcu_thread_online();
1475 cmm_smp_mb(); /* finish resize before decrement */
1476 uatomic_dec(&ht
->in_progress_resize
);
1480 void cds_lfht_resize_lazy(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1482 struct rcu_resize_work
*work
;
1483 unsigned long target_size
;
1485 target_size
= resize_target_update(ht
, size
, growth
);
1486 /* Store resize_target before read resize_initiated */
1488 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
) && size
< target_size
) {
1489 uatomic_inc(&ht
->in_progress_resize
);
1490 cmm_smp_mb(); /* increment resize count before calling it */
1491 work
= malloc(sizeof(*work
));
1493 ht
->cds_lfht_call_rcu(&work
->head
, do_resize_cb
);
1494 CMM_STORE_SHARED(ht
->t
.resize_initiated
, 1);
1498 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1501 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1502 unsigned long count
)
1504 struct rcu_resize_work
*work
;
1506 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1508 resize_target_update_count(ht
, count
);
1509 /* Store resize_target before read resize_initiated */
1511 if (!CMM_LOAD_SHARED(ht
->t
.resize_initiated
)) {
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);