4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups and use the returned objects safely by delaying
38 * memory reclaim of a grace period.
39 * - Add and remove operations are lock-free, and do not need to
40 * allocate memory. They need to be executed within RCU read-side
41 * critical section to ensure the objects they read are valid and to
42 * deal with the cmpxchg ABA problem.
43 * - add and add_unique operations are supported. add_unique checks if
44 * the node key already exists in the hash table. It ensures no key
46 * - The resize operation executes concurrently with add/remove/lookup.
47 * - Hash table nodes are contained within a split-ordered list. This
48 * list is ordered by incrementing reversed-bits-hash value.
49 * - An index of bucket nodes is kept. These bucket nodes are the hash
50 * table "buckets", and they are also chained together in the
51 * split-ordered list, which allows recursive expansion.
52 * - The resize operation for small tables only allows expanding the hash table.
53 * It is triggered automatically by detecting long chains in the add
55 * - The resize operation for larger tables (and available through an
56 * API) allows both expanding and shrinking the hash table.
57 * - Split-counters are used to keep track of the number of
58 * nodes within the hash table for automatic resize triggering.
59 * - Resize operation initiated by long chain detection is executed by a
60 * call_rcu thread, which keeps lock-freedom of add and remove.
61 * - Resize operations are protected by a mutex.
62 * - The removal operation is split in two parts: first, a "removed"
63 * flag is set in the next pointer within the node to remove. Then,
64 * a "garbage collection" is performed in the bucket containing the
65 * removed node (from the start of the bucket up to the removed node).
66 * All encountered nodes with "removed" flag set in their next
67 * pointers are removed from the linked-list. If the cmpxchg used for
68 * removal fails (due to concurrent garbage-collection or concurrent
69 * add), we retry from the beginning of the bucket. This ensures that
70 * the node with "removed" flag set is removed from the hash table
71 * (not visible to lookups anymore) before the RCU read-side critical
72 * section held across removal ends. Furthermore, this ensures that
73 * the node with "removed" flag set is removed from the linked-list
74 * before its memory is reclaimed. Only the thread which removal
75 * successfully set the "removed" flag (with a cmpxchg) into a node's
76 * next pointer is considered to have succeeded its removal (and thus
77 * owns the node to reclaim). Because we garbage-collect starting from
78 * an invariant node (the start-of-bucket bucket node) up to the
79 * "removed" node (or find a reverse-hash that is higher), we are sure
80 * that a successful traversal of the chain leads to a chain that is
81 * present in the linked-list (the start node is never removed) and
82 * that is does not contain the "removed" node anymore, even if
83 * concurrent delete/add operations are changing the structure of the
85 * - The add operation performs gargage collection of buckets if it
86 * encounters nodes with removed flag set in the bucket where it wants
87 * to add its new node. This ensures lock-freedom of add operation by
88 * helping the remover unlink nodes from the list rather than to wait
90 * - A RCU "order table" indexed by log2(hash index) is copied and
91 * expanded by the resize operation. This order table allows finding
92 * the "bucket node" tables.
93 * - There is one bucket node table per hash index order. The size of
94 * each bucket node table is half the number of hashes contained in
95 * this order (except for order 0).
96 * - synchronzie_rcu is used to garbage-collect the old bucket node table.
97 * - The per-order bucket node tables contain a compact version of the
98 * hash table nodes. These tables are invariant after they are
99 * populated into the hash table.
101 * Bucket node tables:
103 * hash table hash table the last all bucket node tables
104 * order size bucket node 0 1 2 3 4 5 6(index)
111 * 5 32 16 1 1 2 4 8 16
112 * 6 64 32 1 1 2 4 8 16 32
114 * When growing/shrinking, we only focus on the last bucket node table
115 * which size is (!order ? 1 : (1 << (order -1))).
117 * Example for growing/shrinking:
118 * grow hash table from order 5 to 6: init the index=6 bucket node table
119 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
121 * A bit of ascii art explanation:
123 * Order index is the off-by-one compare to the actual power of 2 because
124 * we use index 0 to deal with the 0 special-case.
126 * This shows the nodes for a small table ordered by reversed bits:
138 * This shows the nodes in order of non-reversed bits, linked by
139 * reversed-bit order.
144 * 2 | | 2 010 010 <- |
145 * | | | 3 011 110 | <- |
146 * 3 -> | | | 4 100 001 | |
162 #include <urcu-call-rcu.h>
163 #include <urcu/arch.h>
164 #include <urcu/uatomic.h>
165 #include <urcu/compiler.h>
169 #include "rculfhash.h"
170 #include "rculfhash-internal.h"
171 #include "urcu-flavor.h"
174 * We need to lock pthread exit, which deadlocks __nptl_setxid in the
176 * This work-around will be allowed to be removed when runas.c gets
177 * changed to do an exec() before issuing seteuid/setegid.
178 * See http://sourceware.org/bugzilla/show_bug.cgi?id=10184 for details.
180 pthread_mutex_t lttng_libc_state_lock
= PTHREAD_MUTEX_INITIALIZER
;
183 * Split-counters lazily update the global counter each 1024
184 * addition/removal. It automatically keeps track of resize required.
185 * We use the bucket length as indicator for need to expand for small
186 * tables and machines lacking per-cpu data suppport.
188 #define COUNT_COMMIT_ORDER 10
189 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
190 #define CHAIN_LEN_TARGET 1
191 #define CHAIN_LEN_RESIZE_THRESHOLD 3
194 * Define the minimum table size.
196 #define MIN_TABLE_ORDER 0
197 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
200 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
202 #define MIN_PARTITION_PER_THREAD_ORDER 12
203 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
206 * The removed flag needs to be updated atomically with the pointer.
207 * It indicates that no node must attach to the node scheduled for
208 * removal, and that node garbage collection must be performed.
209 * The bucket flag does not require to be updated atomically with the
210 * pointer, but it is added as a pointer low bit flag to save space.
212 #define REMOVED_FLAG (1UL << 0)
213 #define BUCKET_FLAG (1UL << 1)
214 #define REMOVAL_OWNER_FLAG (1UL << 2)
215 #define FLAGS_MASK ((1UL << 3) - 1)
217 /* Value of the end pointer. Should not interact with flags. */
218 #define END_VALUE NULL
220 DEFINE_RCU_FLAVOR(rcu_flavor
);
223 * ht_items_count: Split-counters counting the number of node addition
224 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
225 * is set at hash table creation.
227 * These are free-running counters, never reset to zero. They count the
228 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
229 * operations to update the global counter. We choose a power-of-2 value
230 * for the trigger to deal with 32 or 64-bit overflow of the counter.
232 struct ht_items_count
{
233 unsigned long add
, del
;
234 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
237 * rcu_resize_work: Contains arguments passed to RCU worker thread
238 * responsible for performing lazy resize.
240 struct rcu_resize_work
{
241 struct rcu_head head
;
246 * partition_resize_work: Contains arguments passed to worker threads
247 * executing the hash table resize on partitions of the hash table
248 * assigned to each processor's worker thread.
250 struct partition_resize_work
{
253 unsigned long i
, start
, len
;
254 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
255 unsigned long start
, unsigned long len
);
259 * Algorithm to reverse bits in a word by lookup table, extended to
262 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
263 * Originally from Public Domain.
266 static const uint8_t BitReverseTable256
[256] =
268 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
269 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
270 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
271 R6(0), R6(2), R6(1), R6(3)
278 uint8_t bit_reverse_u8(uint8_t v
)
280 return BitReverseTable256
[v
];
283 static __attribute__((unused
))
284 uint32_t bit_reverse_u32(uint32_t v
)
286 return ((uint32_t) bit_reverse_u8(v
) << 24) |
287 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
288 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
289 ((uint32_t) bit_reverse_u8(v
>> 24));
292 static __attribute__((unused
))
293 uint64_t bit_reverse_u64(uint64_t v
)
295 return ((uint64_t) bit_reverse_u8(v
) << 56) |
296 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
297 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
298 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
299 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
300 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
301 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
302 ((uint64_t) bit_reverse_u8(v
>> 56));
306 unsigned long bit_reverse_ulong(unsigned long v
)
308 #if (CAA_BITS_PER_LONG == 32)
309 return bit_reverse_u32(v
);
311 return bit_reverse_u64(v
);
316 * fls: returns the position of the most significant bit.
317 * Returns 0 if no bit is set, else returns the position of the most
318 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
320 #if defined(__i386) || defined(__x86_64)
322 unsigned int fls_u32(uint32_t x
)
330 : "=r" (r
) : "rm" (x
));
336 #if defined(__x86_64)
338 unsigned int fls_u64(uint64_t x
)
346 : "=r" (r
) : "rm" (x
));
353 static __attribute__((unused
))
354 unsigned int fls_u64(uint64_t x
)
361 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
365 if (!(x
& 0xFFFF000000000000ULL
)) {
369 if (!(x
& 0xFF00000000000000ULL
)) {
373 if (!(x
& 0xF000000000000000ULL
)) {
377 if (!(x
& 0xC000000000000000ULL
)) {
381 if (!(x
& 0x8000000000000000ULL
)) {
390 static __attribute__((unused
))
391 unsigned int fls_u32(uint32_t x
)
397 if (!(x
& 0xFFFF0000U
)) {
401 if (!(x
& 0xFF000000U
)) {
405 if (!(x
& 0xF0000000U
)) {
409 if (!(x
& 0xC0000000U
)) {
413 if (!(x
& 0x80000000U
)) {
421 unsigned int cds_lfht_fls_ulong(unsigned long x
)
423 #if (CAA_BITS_PER_LONG == 32)
431 * Return the minimum order for which x <= (1UL << order).
432 * Return -1 if x is 0.
434 int cds_lfht_get_count_order_u32(uint32_t x
)
439 return fls_u32(x
- 1);
443 * Return the minimum order for which x <= (1UL << order).
444 * Return -1 if x is 0.
446 int cds_lfht_get_count_order_ulong(unsigned long x
)
451 return cds_lfht_fls_ulong(x
- 1);
455 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
458 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
459 unsigned long count
);
461 static long nr_cpus_mask
= -1;
462 static long split_count_mask
= -1;
464 #if defined(HAVE_SYSCONF)
465 static void ht_init_nr_cpus_mask(void)
469 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
475 * round up number of CPUs to next power of two, so we
476 * can use & for modulo.
478 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
479 nr_cpus_mask
= maxcpus
- 1;
481 #else /* #if defined(HAVE_SYSCONF) */
482 static void ht_init_nr_cpus_mask(void)
486 #endif /* #else #if defined(HAVE_SYSCONF) */
489 void alloc_split_items_count(struct cds_lfht
*ht
)
491 struct ht_items_count
*count
;
493 if (nr_cpus_mask
== -1) {
494 ht_init_nr_cpus_mask();
495 if (nr_cpus_mask
< 0)
496 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
498 split_count_mask
= nr_cpus_mask
;
501 assert(split_count_mask
>= 0);
503 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
504 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
505 assert(ht
->split_count
);
507 ht
->split_count
= NULL
;
512 void free_split_items_count(struct cds_lfht
*ht
)
514 poison_free(ht
->split_count
);
517 #if defined(HAVE_SCHED_GETCPU)
519 int ht_get_split_count_index(unsigned long hash
)
523 assert(split_count_mask
>= 0);
524 cpu
= sched_getcpu();
525 if (caa_unlikely(cpu
< 0))
526 return hash
& split_count_mask
;
528 return cpu
& split_count_mask
;
530 #else /* #if defined(HAVE_SCHED_GETCPU) */
532 int ht_get_split_count_index(unsigned long hash
)
534 return hash
& split_count_mask
;
536 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
539 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
541 unsigned long split_count
;
545 if (caa_unlikely(!ht
->split_count
))
547 index
= ht_get_split_count_index(hash
);
548 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
549 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
551 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
553 dbg_printf("add split count %lu\n", split_count
);
554 count
= uatomic_add_return(&ht
->count
,
555 1UL << COUNT_COMMIT_ORDER
);
556 if (caa_likely(count
& (count
- 1)))
558 /* Only if global count is power of 2 */
560 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
562 dbg_printf("add set global %ld\n", count
);
563 cds_lfht_resize_lazy_count(ht
, size
,
564 count
>> (CHAIN_LEN_TARGET
- 1));
568 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
570 unsigned long split_count
;
574 if (caa_unlikely(!ht
->split_count
))
576 index
= ht_get_split_count_index(hash
);
577 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
578 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
580 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
582 dbg_printf("del split count %lu\n", split_count
);
583 count
= uatomic_add_return(&ht
->count
,
584 -(1UL << COUNT_COMMIT_ORDER
));
585 if (caa_likely(count
& (count
- 1)))
587 /* Only if global count is power of 2 */
589 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
591 dbg_printf("del set global %ld\n", count
);
593 * Don't shrink table if the number of nodes is below a
596 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
598 cds_lfht_resize_lazy_count(ht
, size
,
599 count
>> (CHAIN_LEN_TARGET
- 1));
603 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
607 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
609 count
= uatomic_read(&ht
->count
);
611 * Use bucket-local length for small table expand and for
612 * environments lacking per-cpu data support.
614 if (count
>= (1UL << COUNT_COMMIT_ORDER
))
617 dbg_printf("WARNING: large chain length: %u.\n",
619 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
)
620 cds_lfht_resize_lazy_grow(ht
, size
,
621 cds_lfht_get_count_order_u32(chain_len
- (CHAIN_LEN_TARGET
- 1)));
625 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
627 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
631 int is_removed(struct cds_lfht_node
*node
)
633 return ((unsigned long) node
) & REMOVED_FLAG
;
637 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
639 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
643 int is_bucket(struct cds_lfht_node
*node
)
645 return ((unsigned long) node
) & BUCKET_FLAG
;
649 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
651 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
655 int is_removal_owner(struct cds_lfht_node
*node
)
657 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
661 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
663 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
667 struct cds_lfht_node
*get_end(void)
669 return (struct cds_lfht_node
*) END_VALUE
;
673 int is_end(struct cds_lfht_node
*node
)
675 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
679 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
682 unsigned long old1
, old2
;
684 old1
= uatomic_read(ptr
);
689 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
694 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
696 return ht
->mm
->alloc_bucket_table(ht
, order
);
700 * cds_lfht_free_bucket_table() should be called with decreasing order.
701 * When cds_lfht_free_bucket_table(0) is called, it means the whole
705 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
707 return ht
->mm
->free_bucket_table(ht
, order
);
711 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
713 return ht
->bucket_at(ht
, index
);
717 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
721 return bucket_at(ht
, hash
& (size
- 1));
725 * Remove all logically deleted nodes from a bucket up to a certain node key.
728 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
730 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
732 assert(!is_bucket(bucket
));
733 assert(!is_removed(bucket
));
734 assert(!is_bucket(node
));
735 assert(!is_removed(node
));
738 /* We can always skip the bucket node initially */
739 iter
= rcu_dereference(iter_prev
->next
);
740 assert(!is_removed(iter
));
741 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
743 * We should never be called with bucket (start of chain)
744 * and logically removed node (end of path compression
745 * marker) being the actual same node. This would be a
746 * bug in the algorithm implementation.
748 assert(bucket
!= node
);
750 if (caa_unlikely(is_end(iter
)))
752 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
754 next
= rcu_dereference(clear_flag(iter
)->next
);
755 if (caa_likely(is_removed(next
)))
757 iter_prev
= clear_flag(iter
);
760 assert(!is_removed(iter
));
762 new_next
= flag_bucket(clear_flag(next
));
764 new_next
= clear_flag(next
);
765 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
770 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
771 struct cds_lfht_node
*old_node
,
772 struct cds_lfht_node
*old_next
,
773 struct cds_lfht_node
*new_node
)
775 struct cds_lfht_node
*bucket
, *ret_next
;
777 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
780 assert(!is_removed(old_node
));
781 assert(!is_bucket(old_node
));
782 assert(!is_removed(new_node
));
783 assert(!is_bucket(new_node
));
784 assert(new_node
!= old_node
);
786 /* Insert after node to be replaced */
787 if (is_removed(old_next
)) {
789 * Too late, the old node has been removed under us
790 * between lookup and replace. Fail.
794 assert(old_next
== clear_flag(old_next
));
795 assert(new_node
!= old_next
);
796 new_node
->next
= old_next
;
798 * Here is the whole trick for lock-free replace: we add
799 * the replacement node _after_ the node we want to
800 * replace by atomically setting its next pointer at the
801 * same time we set its removal flag. Given that
802 * the lookups/get next use an iterator aware of the
803 * next pointer, they will either skip the old node due
804 * to the removal flag and see the new node, or use
805 * the old node, but will not see the new one.
806 * This is a replacement of a node with another node
807 * that has the same value: we are therefore not
808 * removing a value from the hash table.
810 ret_next
= uatomic_cmpxchg(&old_node
->next
,
811 old_next
, flag_removed(new_node
));
812 if (ret_next
== old_next
)
813 break; /* We performed the replacement. */
818 * Ensure that the old node is not visible to readers anymore:
819 * lookup for the node, and remove it (along with any other
820 * logically removed node) if found.
822 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
823 _cds_lfht_gc_bucket(bucket
, new_node
);
825 assert(is_removed(rcu_dereference(old_node
->next
)));
830 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
831 * mode. A NULL unique_ret allows creation of duplicate keys.
834 void _cds_lfht_add(struct cds_lfht
*ht
,
836 cds_lfht_match_fct match
,
839 struct cds_lfht_node
*node
,
840 struct cds_lfht_iter
*unique_ret
,
843 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
845 struct cds_lfht_node
*bucket
;
847 assert(!is_bucket(node
));
848 assert(!is_removed(node
));
849 bucket
= lookup_bucket(ht
, size
, hash
);
851 uint32_t chain_len
= 0;
854 * iter_prev points to the non-removed node prior to the
858 /* We can always skip the bucket node initially */
859 iter
= rcu_dereference(iter_prev
->next
);
860 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
862 if (caa_unlikely(is_end(iter
)))
864 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
867 /* bucket node is the first node of the identical-hash-value chain */
868 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
871 next
= rcu_dereference(clear_flag(iter
)->next
);
872 if (caa_unlikely(is_removed(next
)))
878 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
879 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
882 * uniquely adding inserts the node as the first
883 * node of the identical-hash-value node chain.
885 * This semantic ensures no duplicated keys
886 * should ever be observable in the table
887 * (including observe one node by one node
888 * by forward iterations)
890 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
894 *unique_ret
= d_iter
;
898 /* Only account for identical reverse hash once */
899 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
901 check_resize(ht
, size
, ++chain_len
);
902 iter_prev
= clear_flag(iter
);
907 assert(node
!= clear_flag(iter
));
908 assert(!is_removed(iter_prev
));
909 assert(!is_removed(iter
));
910 assert(iter_prev
!= node
);
912 node
->next
= clear_flag(iter
);
914 node
->next
= flag_bucket(clear_flag(iter
));
916 new_node
= flag_bucket(node
);
919 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
921 continue; /* retry */
928 assert(!is_removed(iter
));
930 new_next
= flag_bucket(clear_flag(next
));
932 new_next
= clear_flag(next
);
933 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
938 unique_ret
->node
= return_node
;
939 /* unique_ret->next left unset, never used. */
944 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
945 struct cds_lfht_node
*node
)
947 struct cds_lfht_node
*bucket
, *next
;
949 if (!node
) /* Return -ENOENT if asked to delete NULL node */
952 /* logically delete the node */
953 assert(!is_bucket(node
));
954 assert(!is_removed(node
));
955 assert(!is_removal_owner(node
));
958 * We are first checking if the node had previously been
959 * logically removed (this check is not atomic with setting the
960 * logical removal flag). Return -ENOENT if the node had
961 * previously been removed.
963 next
= rcu_dereference(node
->next
);
964 if (caa_unlikely(is_removed(next
)))
966 assert(!is_bucket(next
));
968 * We set the REMOVED_FLAG unconditionally. Note that there may
969 * be more than one concurrent thread setting this flag.
970 * Knowing which wins the race will be known after the garbage
971 * collection phase, stay tuned!
973 uatomic_or(&node
->next
, REMOVED_FLAG
);
974 /* We performed the (logical) deletion. */
977 * Ensure that the node is not visible to readers anymore: lookup for
978 * the node, and remove it (along with any other logically removed node)
981 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
982 _cds_lfht_gc_bucket(bucket
, node
);
984 assert(is_removed(rcu_dereference(node
->next
)));
986 * Last phase: atomically exchange node->next with a version
987 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
988 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
989 * the node and win the removal race.
990 * It is interesting to note that all "add" paths are forbidden
991 * to change the next pointer starting from the point where the
992 * REMOVED_FLAG is set, so here using a read, followed by a
993 * xchg() suffice to guarantee that the xchg() will ever only
994 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
997 if (!is_removal_owner(uatomic_xchg(&node
->next
,
998 flag_removal_owner(node
->next
))))
1005 void *partition_resize_thread(void *arg
)
1007 struct partition_resize_work
*work
= arg
;
1009 work
->ht
->flavor
->register_thread();
1010 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1011 work
->ht
->flavor
->unregister_thread();
1016 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1018 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1019 unsigned long start
, unsigned long len
))
1021 unsigned long partition_len
;
1022 struct partition_resize_work
*work
;
1024 unsigned long nr_threads
;
1027 * Note: nr_cpus_mask + 1 is always power of 2.
1028 * We spawn just the number of threads we need to satisfy the minimum
1029 * partition size, up to the number of CPUs in the system.
1031 if (nr_cpus_mask
> 0) {
1032 nr_threads
= min(nr_cpus_mask
+ 1,
1033 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1037 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1038 work
= calloc(nr_threads
, sizeof(*work
));
1040 pthread_mutex_lock(<tng_libc_state_lock
);
1041 for (thread
= 0; thread
< nr_threads
; thread
++) {
1042 work
[thread
].ht
= ht
;
1044 work
[thread
].len
= partition_len
;
1045 work
[thread
].start
= thread
* partition_len
;
1046 work
[thread
].fct
= fct
;
1047 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1048 partition_resize_thread
, &work
[thread
]);
1051 for (thread
= 0; thread
< nr_threads
; thread
++) {
1052 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1055 pthread_mutex_unlock(<tng_libc_state_lock
);
1060 * Holding RCU read lock to protect _cds_lfht_add against memory
1061 * reclaim that could be performed by other call_rcu worker threads (ABA
1064 * When we reach a certain length, we can split this population phase over
1065 * many worker threads, based on the number of CPUs available in the system.
1066 * This should therefore take care of not having the expand lagging behind too
1067 * many concurrent insertion threads by using the scheduler's ability to
1068 * schedule bucket node population fairly with insertions.
1071 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1072 unsigned long start
, unsigned long len
)
1074 unsigned long j
, size
= 1UL << (i
- 1);
1076 assert(i
> MIN_TABLE_ORDER
);
1077 ht
->flavor
->read_lock();
1078 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1079 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1081 assert(j
>= size
&& j
< (size
<< 1));
1082 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1084 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1085 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1087 ht
->flavor
->read_unlock();
1091 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1094 assert(nr_cpus_mask
!= -1);
1095 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1096 ht
->flavor
->thread_online();
1097 init_table_populate_partition(ht
, i
, 0, len
);
1098 ht
->flavor
->thread_offline();
1101 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1105 void init_table(struct cds_lfht
*ht
,
1106 unsigned long first_order
, unsigned long last_order
)
1110 dbg_printf("init table: first_order %lu last_order %lu\n",
1111 first_order
, last_order
);
1112 assert(first_order
> MIN_TABLE_ORDER
);
1113 for (i
= first_order
; i
<= last_order
; i
++) {
1116 len
= 1UL << (i
- 1);
1117 dbg_printf("init order %lu len: %lu\n", i
, len
);
1119 /* Stop expand if the resize target changes under us */
1120 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1123 cds_lfht_alloc_bucket_table(ht
, i
);
1126 * Set all bucket nodes reverse hash values for a level and
1127 * link all bucket nodes into the table.
1129 init_table_populate(ht
, i
, len
);
1132 * Update table size.
1134 cmm_smp_wmb(); /* populate data before RCU size */
1135 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1137 dbg_printf("init new size: %lu\n", 1UL << i
);
1138 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1144 * Holding RCU read lock to protect _cds_lfht_remove against memory
1145 * reclaim that could be performed by other call_rcu worker threads (ABA
1147 * For a single level, we logically remove and garbage collect each node.
1149 * As a design choice, we perform logical removal and garbage collection on a
1150 * node-per-node basis to simplify this algorithm. We also assume keeping good
1151 * cache locality of the operation would overweight possible performance gain
1152 * that could be achieved by batching garbage collection for multiple levels.
1153 * However, this would have to be justified by benchmarks.
1155 * Concurrent removal and add operations are helping us perform garbage
1156 * collection of logically removed nodes. We guarantee that all logically
1157 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1158 * invoked to free a hole level of bucket nodes (after a grace period).
1160 * Logical removal and garbage collection can therefore be done in batch or on a
1161 * node-per-node basis, as long as the guarantee above holds.
1163 * When we reach a certain length, we can split this removal over many worker
1164 * threads, based on the number of CPUs available in the system. This should
1165 * take care of not letting resize process lag behind too many concurrent
1166 * updater threads actively inserting into the hash table.
1169 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1170 unsigned long start
, unsigned long len
)
1172 unsigned long j
, size
= 1UL << (i
- 1);
1174 assert(i
> MIN_TABLE_ORDER
);
1175 ht
->flavor
->read_lock();
1176 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1177 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1178 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1180 assert(j
>= size
&& j
< (size
<< 1));
1181 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1183 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1184 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1185 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1187 ht
->flavor
->read_unlock();
1191 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1194 assert(nr_cpus_mask
!= -1);
1195 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1196 ht
->flavor
->thread_online();
1197 remove_table_partition(ht
, i
, 0, len
);
1198 ht
->flavor
->thread_offline();
1201 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1205 * fini_table() is never called for first_order == 0, which is why
1206 * free_by_rcu_order == 0 can be used as criterion to know if free must
1210 void fini_table(struct cds_lfht
*ht
,
1211 unsigned long first_order
, unsigned long last_order
)
1214 unsigned long free_by_rcu_order
= 0;
1216 dbg_printf("fini table: first_order %lu last_order %lu\n",
1217 first_order
, last_order
);
1218 assert(first_order
> MIN_TABLE_ORDER
);
1219 for (i
= last_order
; i
>= first_order
; i
--) {
1222 len
= 1UL << (i
- 1);
1223 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1225 /* Stop shrink if the resize target changes under us */
1226 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1229 cmm_smp_wmb(); /* populate data before RCU size */
1230 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1233 * We need to wait for all add operations to reach Q.S. (and
1234 * thus use the new table for lookups) before we can start
1235 * releasing the old bucket nodes. Otherwise their lookup will
1236 * return a logically removed node as insert position.
1238 ht
->flavor
->update_synchronize_rcu();
1239 if (free_by_rcu_order
)
1240 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1243 * Set "removed" flag in bucket nodes about to be removed.
1244 * Unlink all now-logically-removed bucket node pointers.
1245 * Concurrent add/remove operation are helping us doing
1248 remove_table(ht
, i
, len
);
1250 free_by_rcu_order
= i
;
1252 dbg_printf("fini new size: %lu\n", 1UL << i
);
1253 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1257 if (free_by_rcu_order
) {
1258 ht
->flavor
->update_synchronize_rcu();
1259 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1264 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1266 struct cds_lfht_node
*prev
, *node
;
1267 unsigned long order
, len
, i
;
1269 cds_lfht_alloc_bucket_table(ht
, 0);
1271 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1272 node
= bucket_at(ht
, 0);
1273 node
->next
= flag_bucket(get_end());
1274 node
->reverse_hash
= 0;
1276 for (order
= 1; order
< cds_lfht_get_count_order_ulong(size
) + 1; order
++) {
1277 len
= 1UL << (order
- 1);
1278 cds_lfht_alloc_bucket_table(ht
, order
);
1280 for (i
= 0; i
< len
; i
++) {
1282 * Now, we are trying to init the node with the
1283 * hash=(len+i) (which is also a bucket with the
1284 * index=(len+i)) and insert it into the hash table,
1285 * so this node has to be inserted after the bucket
1286 * with the index=(len+i)&(len-1)=i. And because there
1287 * is no other non-bucket node nor bucket node with
1288 * larger index/hash inserted, so the bucket node
1289 * being inserted should be inserted directly linked
1290 * after the bucket node with index=i.
1292 prev
= bucket_at(ht
, i
);
1293 node
= bucket_at(ht
, len
+ i
);
1295 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1296 order
, len
+ i
, len
+ i
);
1297 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1299 /* insert after prev */
1300 assert(is_bucket(prev
->next
));
1301 node
->next
= prev
->next
;
1302 prev
->next
= flag_bucket(node
);
1307 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1308 unsigned long min_nr_alloc_buckets
,
1309 unsigned long max_nr_buckets
,
1311 const struct cds_lfht_mm_type
*mm
,
1312 const struct rcu_flavor_struct
*flavor
,
1313 pthread_attr_t
*attr
)
1315 struct cds_lfht
*ht
;
1316 unsigned long order
;
1318 /* min_nr_alloc_buckets must be power of two */
1319 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1322 /* init_size must be power of two */
1323 if (!init_size
|| (init_size
& (init_size
- 1)))
1327 * Memory management plugin default.
1330 if (CAA_BITS_PER_LONG
> 32
1332 && max_nr_buckets
<= (1ULL << 32)) {
1334 * For 64-bit architectures, with max number of
1335 * buckets small enough not to use the entire
1336 * 64-bit memory mapping space (and allowing a
1337 * fair number of hash table instances), use the
1338 * mmap allocator, which is faster than the
1341 mm
= &cds_lfht_mm_mmap
;
1344 * The fallback is to use the order allocator.
1346 mm
= &cds_lfht_mm_order
;
1350 /* max_nr_buckets == 0 for order based mm means infinite */
1351 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1352 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1354 /* max_nr_buckets must be power of two */
1355 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1358 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1359 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1360 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1361 init_size
= min(init_size
, max_nr_buckets
);
1363 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1365 assert(ht
->mm
== mm
);
1366 assert(ht
->bucket_at
== mm
->bucket_at
);
1369 ht
->flavor
= flavor
;
1370 ht
->resize_attr
= attr
;
1371 alloc_split_items_count(ht
);
1372 /* this mutex should not nest in read-side C.S. */
1373 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1374 order
= cds_lfht_get_count_order_ulong(init_size
);
1375 ht
->resize_target
= 1UL << order
;
1376 cds_lfht_create_bucket(ht
, 1UL << order
);
1377 ht
->size
= 1UL << order
;
1381 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1382 cds_lfht_match_fct match
, const void *key
,
1383 struct cds_lfht_iter
*iter
)
1385 struct cds_lfht_node
*node
, *next
, *bucket
;
1386 unsigned long reverse_hash
, size
;
1388 reverse_hash
= bit_reverse_ulong(hash
);
1390 size
= rcu_dereference(ht
->size
);
1391 bucket
= lookup_bucket(ht
, size
, hash
);
1392 /* We can always skip the bucket node initially */
1393 node
= rcu_dereference(bucket
->next
);
1394 node
= clear_flag(node
);
1396 if (caa_unlikely(is_end(node
))) {
1400 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1404 next
= rcu_dereference(node
->next
);
1405 assert(node
== clear_flag(node
));
1406 if (caa_likely(!is_removed(next
))
1408 && node
->reverse_hash
== reverse_hash
1409 && caa_likely(match(node
, key
))) {
1412 node
= clear_flag(next
);
1414 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1419 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1420 const void *key
, struct cds_lfht_iter
*iter
)
1422 struct cds_lfht_node
*node
, *next
;
1423 unsigned long reverse_hash
;
1426 reverse_hash
= node
->reverse_hash
;
1428 node
= clear_flag(next
);
1431 if (caa_unlikely(is_end(node
))) {
1435 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1439 next
= rcu_dereference(node
->next
);
1440 if (caa_likely(!is_removed(next
))
1442 && caa_likely(match(node
, key
))) {
1445 node
= clear_flag(next
);
1447 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1452 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1454 struct cds_lfht_node
*node
, *next
;
1456 node
= clear_flag(iter
->next
);
1458 if (caa_unlikely(is_end(node
))) {
1462 next
= rcu_dereference(node
->next
);
1463 if (caa_likely(!is_removed(next
))
1464 && !is_bucket(next
)) {
1467 node
= clear_flag(next
);
1469 assert(!node
|| !is_bucket(rcu_dereference(node
->next
)));
1474 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1477 * Get next after first bucket node. The first bucket node is the
1478 * first node of the linked list.
1480 iter
->next
= bucket_at(ht
, 0)->next
;
1481 cds_lfht_next(ht
, iter
);
1484 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1485 struct cds_lfht_node
*node
)
1489 node
->reverse_hash
= bit_reverse_ulong(hash
);
1490 size
= rcu_dereference(ht
->size
);
1491 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1492 ht_count_add(ht
, size
, hash
);
1495 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1497 cds_lfht_match_fct match
,
1499 struct cds_lfht_node
*node
)
1502 struct cds_lfht_iter iter
;
1504 node
->reverse_hash
= bit_reverse_ulong(hash
);
1505 size
= rcu_dereference(ht
->size
);
1506 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1507 if (iter
.node
== node
)
1508 ht_count_add(ht
, size
, hash
);
1512 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1514 cds_lfht_match_fct match
,
1516 struct cds_lfht_node
*node
)
1519 struct cds_lfht_iter iter
;
1521 node
->reverse_hash
= bit_reverse_ulong(hash
);
1522 size
= rcu_dereference(ht
->size
);
1524 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1525 if (iter
.node
== node
) {
1526 ht_count_add(ht
, size
, hash
);
1530 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1535 int cds_lfht_replace(struct cds_lfht
*ht
,
1536 struct cds_lfht_iter
*old_iter
,
1538 cds_lfht_match_fct match
,
1540 struct cds_lfht_node
*new_node
)
1544 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1545 if (!old_iter
->node
)
1547 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1549 if (caa_unlikely(!match(old_iter
->node
, key
)))
1551 size
= rcu_dereference(ht
->size
);
1552 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1556 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1558 unsigned long size
, hash
;
1561 size
= rcu_dereference(ht
->size
);
1562 ret
= _cds_lfht_del(ht
, size
, node
);
1564 hash
= bit_reverse_ulong(node
->reverse_hash
);
1565 ht_count_del(ht
, size
, hash
);
1571 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1573 struct cds_lfht_node
*node
;
1574 unsigned long order
, i
, size
;
1576 /* Check that the table is empty */
1577 node
= bucket_at(ht
, 0);
1579 node
= clear_flag(node
)->next
;
1580 if (!is_bucket(node
))
1582 assert(!is_removed(node
));
1583 } while (!is_end(node
));
1585 * size accessed without rcu_dereference because hash table is
1589 /* Internal sanity check: all nodes left should be bucket */
1590 for (i
= 0; i
< size
; i
++) {
1591 node
= bucket_at(ht
, i
);
1592 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1593 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1594 assert(is_bucket(node
->next
));
1597 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1598 cds_lfht_free_bucket_table(ht
, order
);
1604 * Should only be called when no more concurrent readers nor writers can
1605 * possibly access the table.
1607 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1611 /* Wait for in-flight resize operations to complete */
1612 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1613 cmm_smp_mb(); /* Store destroy before load resize */
1614 while (uatomic_read(&ht
->in_progress_resize
))
1615 poll(NULL
, 0, 100); /* wait for 100ms */
1616 ret
= cds_lfht_delete_bucket(ht
);
1619 free_split_items_count(ht
);
1621 *attr
= ht
->resize_attr
;
1626 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1627 long *approx_before
,
1628 unsigned long *count
,
1631 struct cds_lfht_node
*node
, *next
;
1632 unsigned long nr_bucket
= 0, nr_removed
= 0;
1635 if (ht
->split_count
) {
1638 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1639 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1640 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1646 /* Count non-bucket nodes in the table */
1647 node
= bucket_at(ht
, 0);
1649 next
= rcu_dereference(node
->next
);
1650 if (is_removed(next
)) {
1651 if (!is_bucket(next
))
1655 } else if (!is_bucket(next
))
1659 node
= clear_flag(next
);
1660 } while (!is_end(node
));
1661 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
1662 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1664 if (ht
->split_count
) {
1667 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1668 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1669 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1674 /* called with resize mutex held */
1676 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1677 unsigned long old_size
, unsigned long new_size
)
1679 unsigned long old_order
, new_order
;
1681 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1682 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1683 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1684 old_size
, old_order
, new_size
, new_order
);
1685 assert(new_size
> old_size
);
1686 init_table(ht
, old_order
+ 1, new_order
);
1689 /* called with resize mutex held */
1691 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1692 unsigned long old_size
, unsigned long new_size
)
1694 unsigned long old_order
, new_order
;
1696 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1697 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1698 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1699 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1700 old_size
, old_order
, new_size
, new_order
);
1701 assert(new_size
< old_size
);
1703 /* Remove and unlink all bucket nodes to remove. */
1704 fini_table(ht
, new_order
+ 1, old_order
);
1708 /* called with resize mutex held */
1710 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1712 unsigned long new_size
, old_size
;
1715 * Resize table, re-do if the target size has changed under us.
1718 assert(uatomic_read(&ht
->in_progress_resize
));
1719 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1721 ht
->resize_initiated
= 1;
1722 old_size
= ht
->size
;
1723 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1724 if (old_size
< new_size
)
1725 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1726 else if (old_size
> new_size
)
1727 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1728 ht
->resize_initiated
= 0;
1729 /* write resize_initiated before read resize_target */
1731 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1735 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1737 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
1741 void resize_target_update_count(struct cds_lfht
*ht
,
1742 unsigned long count
)
1744 count
= max(count
, MIN_TABLE_SIZE
);
1745 count
= min(count
, ht
->max_nr_buckets
);
1746 uatomic_set(&ht
->resize_target
, count
);
1749 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1751 resize_target_update_count(ht
, new_size
);
1752 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1753 ht
->flavor
->thread_offline();
1754 pthread_mutex_lock(&ht
->resize_mutex
);
1755 _do_cds_lfht_resize(ht
);
1756 pthread_mutex_unlock(&ht
->resize_mutex
);
1757 ht
->flavor
->thread_online();
1761 void do_resize_cb(struct rcu_head
*head
)
1763 struct rcu_resize_work
*work
=
1764 caa_container_of(head
, struct rcu_resize_work
, head
);
1765 struct cds_lfht
*ht
= work
->ht
;
1767 ht
->flavor
->thread_offline();
1768 pthread_mutex_lock(&ht
->resize_mutex
);
1769 _do_cds_lfht_resize(ht
);
1770 pthread_mutex_unlock(&ht
->resize_mutex
);
1771 ht
->flavor
->thread_online();
1773 cmm_smp_mb(); /* finish resize before decrement */
1774 uatomic_dec(&ht
->in_progress_resize
);
1778 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1780 struct rcu_resize_work
*work
;
1782 /* Store resize_target before read resize_initiated */
1784 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
1785 uatomic_inc(&ht
->in_progress_resize
);
1786 cmm_smp_mb(); /* increment resize count before load destroy */
1787 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1788 uatomic_dec(&ht
->in_progress_resize
);
1791 work
= malloc(sizeof(*work
));
1793 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1794 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1799 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
1801 unsigned long target_size
= size
<< growth
;
1803 target_size
= min(target_size
, ht
->max_nr_buckets
);
1804 if (resize_target_grow(ht
, target_size
) >= target_size
)
1807 __cds_lfht_resize_lazy_launch(ht
);
1811 * We favor grow operations over shrink. A shrink operation never occurs
1812 * if a grow operation is queued for lazy execution. A grow operation
1813 * cancels any pending shrink lazy execution.
1816 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
1817 unsigned long count
)
1819 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
1821 count
= max(count
, MIN_TABLE_SIZE
);
1822 count
= min(count
, ht
->max_nr_buckets
);
1824 return; /* Already the right size, no resize needed */
1825 if (count
> size
) { /* lazy grow */
1826 if (resize_target_grow(ht
, count
) >= count
)
1828 } else { /* lazy shrink */
1832 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
1834 break; /* no resize needed */
1836 return; /* growing is/(was just) in progress */
1838 return; /* some other thread do shrink */
1842 __cds_lfht_resize_lazy_launch(ht
);