1 // SPDX-FileCopyrightText: 2010-2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
2 // SPDX-FileCopyrightText: 2011 Lai Jiangshan <laijs@cn.fujitsu.com>
4 // SPDX-License-Identifier: LGPL-2.1-or-later
7 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
11 * Based on the following articles:
12 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
13 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
14 * - Michael, M. M. High performance dynamic lock-free hash tables
15 * and list-based sets. In Proceedings of the fourteenth annual ACM
16 * symposium on Parallel algorithms and architectures, ACM Press,
19 * Some specificities of this Lock-Free Resizable RCU Hash Table
22 * - RCU read-side critical section allows readers to perform hash
23 * table lookups, as well as traversals, and use the returned objects
24 * safely by allowing memory reclaim to take place only after a grace
26 * - Add and remove operations are lock-free, and do not need to
27 * allocate memory. They need to be executed within RCU read-side
28 * critical section to ensure the objects they read are valid and to
29 * deal with the cmpxchg ABA problem.
30 * - add and add_unique operations are supported. add_unique checks if
31 * the node key already exists in the hash table. It ensures not to
32 * populate a duplicate key if the node key already exists in the hash
34 * - The resize operation executes concurrently with
35 * add/add_unique/add_replace/remove/lookup/traversal.
36 * - Hash table nodes are contained within a split-ordered list. This
37 * list is ordered by incrementing reversed-bits-hash value.
38 * - An index of bucket nodes is kept. These bucket nodes are the hash
39 * table "buckets". These buckets are internal nodes that allow to
40 * perform a fast hash lookup, similarly to a skip list. These
41 * buckets are chained together in the split-ordered list, which
42 * allows recursive expansion by inserting new buckets between the
43 * existing buckets. The split-ordered list allows adding new buckets
44 * between existing buckets as the table needs to grow.
45 * - The resize operation for small tables only allows expanding the
46 * hash table. It is triggered automatically by detecting long chains
47 * in the add operation.
48 * - The resize operation for larger tables (and available through an
49 * API) allows both expanding and shrinking the hash table.
50 * - Split-counters are used to keep track of the number of
51 * nodes within the hash table for automatic resize triggering.
52 * - Resize operation initiated by long chain detection is executed by a
53 * worker thread, which keeps lock-freedom of add and remove.
54 * - Resize operations are protected by a mutex.
55 * - The removal operation is split in two parts: first, a "removed"
56 * flag is set in the next pointer within the node to remove. Then,
57 * a "garbage collection" is performed in the bucket containing the
58 * removed node (from the start of the bucket up to the removed node).
59 * All encountered nodes with "removed" flag set in their next
60 * pointers are removed from the linked-list. If the cmpxchg used for
61 * removal fails (due to concurrent garbage-collection or concurrent
62 * add), we retry from the beginning of the bucket. This ensures that
63 * the node with "removed" flag set is removed from the hash table
64 * (not visible to lookups anymore) before the RCU read-side critical
65 * section held across removal ends. Furthermore, this ensures that
66 * the node with "removed" flag set is removed from the linked-list
67 * before its memory is reclaimed. After setting the "removal" flag,
68 * only the thread which removal is the first to set the "removal
69 * owner" flag (with an xchg) into a node's next pointer is considered
70 * to have succeeded its removal (and thus owns the node to reclaim).
71 * Because we garbage-collect starting from an invariant node (the
72 * start-of-bucket bucket node) up to the "removed" node (or find a
73 * reverse-hash that is higher), we are sure that a successful
74 * traversal of the chain leads to a chain that is present in the
75 * linked-list (the start node is never removed) and that it does not
76 * contain the "removed" node anymore, even if concurrent delete/add
77 * operations are changing the structure of the list concurrently.
78 * - The add operations perform garbage collection of buckets if they
79 * encounter nodes with removed flag set in the bucket where they want
80 * to add their new node. This ensures lock-freedom of add operation by
81 * helping the remover unlink nodes from the list rather than to wait
83 * - There are three memory backends for the hash table buckets: the
84 * "order table", the "chunks", and the "mmap".
85 * - These bucket containers contain a compact version of the hash table
87 * - The RCU "order table":
88 * - has a first level table indexed by log2(hash index) which is
89 * copied and expanded by the resize operation. This order table
90 * allows finding the "bucket node" tables.
91 * - There is one bucket node table per hash index order. The size of
92 * each bucket node table is half the number of hashes contained in
93 * this order (except for order 0).
94 * - The RCU "chunks" is best suited for close interaction with a page
95 * allocator. It uses a linear array as index to "chunks" containing
96 * each the same number of buckets.
97 * - The RCU "mmap" memory backend uses a single memory map to hold
99 * - synchronize_rcu is used to garbage-collect the old bucket node table.
101 * Ordering Guarantees:
103 * To discuss these guarantees, we first define "read" operation as any
104 * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
105 * cds_lfht_first, cds_lfht_next operation, as well as
106 * cds_lfht_add_unique (failure).
108 * We define "read traversal" operation as any of the following
109 * group of operations
110 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
111 * (and/or cds_lfht_next, although less common).
112 * - cds_lfht_add_unique (failure) followed by iteration with
113 * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
115 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
116 * cds_lfht_next_duplicate, although less common).
118 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
119 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
121 * When cds_lfht_add_unique succeeds (returns the node passed as
122 * parameter), it acts as a "write" operation. When cds_lfht_add_unique
123 * fails (returns a node different from the one passed as parameter), it
124 * acts as a "read" operation. A cds_lfht_add_unique failure is a
125 * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
126 * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
129 * We define "prior" and "later" node as nodes observable by reads and
130 * read traversals respectively before and after a write or sequence of
133 * Hash-table operations are often cascaded, for example, the pointer
134 * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
135 * whose return value might in turn be passed to another hash-table
136 * operation. This entire cascaded series of operations must be enclosed
137 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
140 * The following ordering guarantees are offered by this hash table:
142 * A.1) "read" after "write": if there is ordering between a write and a
143 * later read, then the read is guaranteed to see the write or some
145 * A.2) "read traversal" after "write": given that there is dependency
146 * ordering between reads in a "read traversal", if there is
147 * ordering between a write and the first read of the traversal,
148 * then the "read traversal" is guaranteed to see the write or
150 * B.1) "write" after "read": if there is ordering between a read and a
151 * later write, then the read will never see the write.
152 * B.2) "write" after "read traversal": given that there is dependency
153 * ordering between reads in a "read traversal", if there is
154 * ordering between the last read of the traversal and a later
155 * write, then the "read traversal" will never see the write.
156 * C) "write" while "read traversal": if a write occurs during a "read
157 * traversal", the traversal may, or may not, see the write.
158 * D.1) "write" after "write": if there is ordering between a write and
159 * a later write, then the later write is guaranteed to see the
160 * effects of the first write.
161 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
162 * order to any pair of concurrent conflicting writes.
163 * Non-conflicting writes (for example, to different keys) are
165 * E) If a grace period separates a "del" or "replace" operation
166 * and a subsequent operation, then that subsequent operation is
167 * guaranteed not to see the removed item.
168 * F) Uniqueness guarantee: given a hash table that does not contain
169 * duplicate items for a given key, there will only be one item in
170 * the hash table after an arbitrary sequence of add_unique and/or
171 * add_replace operations. Note, however, that a pair of
172 * concurrent read operations might well access two different items
174 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
175 * memory barrier), then the second lookup will return the same
176 * node as the previous lookup, or some later node.
177 * G.2) A "read traversal" that starts after the end of a prior "read
178 * traversal" (ordered by memory barriers) is guaranteed to see the
179 * same nodes as the previous traversal, or some later nodes.
180 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
181 * example, if a pair of reads to the same key run concurrently
182 * with an insertion of that same key, the reads remain unordered
183 * regardless of their return values. In other words, you cannot
184 * rely on the values returned by the reads to deduce ordering.
186 * Progress guarantees:
188 * * Reads are wait-free. These operations always move forward in the
189 * hash table linked list, and this list has no loop.
190 * * Writes are lock-free. Any retry loop performed by a write operation
191 * is triggered by progress made within another update operation.
193 * Bucket node tables:
195 * hash table hash table the last all bucket node tables
196 * order size bucket node 0 1 2 3 4 5 6(index)
203 * 5 32 16 1 1 2 4 8 16
204 * 6 64 32 1 1 2 4 8 16 32
206 * When growing/shrinking, we only focus on the last bucket node table
207 * which size is (!order ? 1 : (1 << (order -1))).
209 * Example for growing/shrinking:
210 * grow hash table from order 5 to 6: init the index=6 bucket node table
211 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
213 * A bit of ascii art explanation:
215 * The order index is the off-by-one compared to the actual power of 2
216 * because we use index 0 to deal with the 0 special-case.
218 * This shows the nodes for a small table ordered by reversed bits:
230 * This shows the nodes in order of non-reversed bits, linked by
231 * reversed-bit order.
236 * 2 | | 2 010 010 <- |
237 * | | | 3 011 110 | <- |
238 * 3 -> | | | 4 100 001 | |
253 #include "compat-getcpu.h"
254 #include <urcu/assert.h>
255 #include <urcu/pointer.h>
256 #include <urcu/call-rcu.h>
257 #include <urcu/flavor.h>
258 #include <urcu/arch.h>
259 #include <urcu/uatomic.h>
260 #include <urcu/compiler.h>
261 #include <urcu/rculfhash.h>
265 #include "rculfhash-internal.h"
266 #include "workqueue.h"
267 #include "urcu-die.h"
268 #include "urcu-utils.h"
269 #include "compat-smp.h"
272 * Split-counters lazily update the global counter each 1024
273 * addition/removal. It automatically keeps track of resize required.
274 * We use the bucket length as indicator for need to expand for small
275 * tables and machines lacking per-cpu data support.
277 #define COUNT_COMMIT_ORDER 10
278 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
279 #define CHAIN_LEN_TARGET 1
280 #define CHAIN_LEN_RESIZE_THRESHOLD 3
283 * Define the minimum table size.
285 #define MIN_TABLE_ORDER 0
286 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
289 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
291 #define MIN_PARTITION_PER_THREAD_ORDER 12
292 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
295 * The removed flag needs to be updated atomically with the pointer.
296 * It indicates that no node must attach to the node scheduled for
297 * removal, and that node garbage collection must be performed.
298 * The bucket flag does not require to be updated atomically with the
299 * pointer, but it is added as a pointer low bit flag to save space.
300 * The "removal owner" flag is used to detect which of the "del"
301 * operation that has set the "removed flag" gets to return the removed
302 * node to its caller. Note that the replace operation does not need to
303 * iteract with the "removal owner" flag, because it validates that
304 * the "removed" flag is not set before performing its cmpxchg.
306 #define REMOVED_FLAG (1UL << 0)
307 #define BUCKET_FLAG (1UL << 1)
308 #define REMOVAL_OWNER_FLAG (1UL << 2)
309 #define FLAGS_MASK ((1UL << 3) - 1)
311 /* Value of the end pointer. Should not interact with flags. */
312 #define END_VALUE NULL
315 * ht_items_count: Split-counters counting the number of node addition
316 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
317 * is set at hash table creation.
319 * These are free-running counters, never reset to zero. They count the
320 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
321 * operations to update the global counter. We choose a power-of-2 value
322 * for the trigger to deal with 32 or 64-bit overflow of the counter.
324 struct ht_items_count
{
325 unsigned long add
, del
;
326 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
329 * resize_work: Contains arguments passed to worker thread
330 * responsible for performing lazy resize.
333 struct urcu_work work
;
338 * partition_resize_work: Contains arguments passed to worker threads
339 * executing the hash table resize on partitions of the hash table
340 * assigned to each processor's worker thread.
342 struct partition_resize_work
{
345 unsigned long i
, start
, len
;
346 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
347 unsigned long start
, unsigned long len
);
350 enum nr_cpus_mask_state
{
351 NR_CPUS_MASK_INIT_FAILED
= -2,
352 NR_CPUS_MASK_UNINITIALIZED
= -1,
355 static struct urcu_workqueue
*cds_lfht_workqueue
;
358 * Mutex ensuring mutual exclusion between workqueue initialization and
359 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
361 static pthread_mutex_t cds_lfht_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
363 static struct urcu_atfork cds_lfht_atfork
;
366 * atfork handler nesting counters. Handle being registered to many urcu
367 * flavors, thus being possibly invoked more than once in the
368 * pthread_atfork list of callbacks.
370 static int cds_lfht_workqueue_atfork_nesting
;
372 static void __attribute__((destructor
)) cds_lfht_exit(void);
373 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
);
375 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
378 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
383 #define cds_lfht_iter_debug_assert(...) urcu_posix_assert(__VA_ARGS__)
388 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
__attribute__((unused
)),
389 struct cds_lfht_iter
*iter
__attribute__((unused
)))
393 #define cds_lfht_iter_debug_assert(...)
398 * Algorithm to reverse bits in a word by lookup table, extended to
401 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
402 * Originally from Public Domain.
405 static const uint8_t BitReverseTable256
[256] =
407 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
408 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
409 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
410 R6(0), R6(2), R6(1), R6(3)
417 uint8_t bit_reverse_u8(uint8_t v
)
419 return BitReverseTable256
[v
];
422 #if (CAA_BITS_PER_LONG == 32)
424 uint32_t bit_reverse_u32(uint32_t v
)
426 return ((uint32_t) bit_reverse_u8(v
) << 24) |
427 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
428 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
429 ((uint32_t) bit_reverse_u8(v
>> 24));
433 uint64_t bit_reverse_u64(uint64_t v
)
435 return ((uint64_t) bit_reverse_u8(v
) << 56) |
436 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
437 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
438 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
439 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
440 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
441 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
442 ((uint64_t) bit_reverse_u8(v
>> 56));
447 unsigned long bit_reverse_ulong(unsigned long v
)
449 #if (CAA_BITS_PER_LONG == 32)
450 return bit_reverse_u32(v
);
452 return bit_reverse_u64(v
);
457 * fls: returns the position of the most significant bit.
458 * Returns 0 if no bit is set, else returns the position of the most
459 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
461 #if defined(URCU_ARCH_X86)
463 unsigned int fls_u32(uint32_t x
)
467 __asm__ ("bsrl %1,%0\n\t"
471 : "=r" (r
) : "rm" (x
));
477 #if defined(URCU_ARCH_AMD64)
479 unsigned int fls_u64(uint64_t x
)
483 __asm__ ("bsrq %1,%0\n\t"
487 : "=r" (r
) : "rm" (x
));
494 static __attribute__((unused
))
495 unsigned int fls_u64(uint64_t x
)
502 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
506 if (!(x
& 0xFFFF000000000000ULL
)) {
510 if (!(x
& 0xFF00000000000000ULL
)) {
514 if (!(x
& 0xF000000000000000ULL
)) {
518 if (!(x
& 0xC000000000000000ULL
)) {
522 if (!(x
& 0x8000000000000000ULL
)) {
531 static __attribute__((unused
))
532 unsigned int fls_u32(uint32_t x
)
538 if (!(x
& 0xFFFF0000U
)) {
542 if (!(x
& 0xFF000000U
)) {
546 if (!(x
& 0xF0000000U
)) {
550 if (!(x
& 0xC0000000U
)) {
554 if (!(x
& 0x80000000U
)) {
562 unsigned int cds_lfht_fls_ulong(unsigned long x
)
564 #if (CAA_BITS_PER_LONG == 32)
572 * Return the minimum order for which x <= (1UL << order).
573 * Return -1 if x is 0.
576 int cds_lfht_get_count_order_u32(uint32_t x
)
581 return fls_u32(x
- 1);
585 * Return the minimum order for which x <= (1UL << order).
586 * Return -1 if x is 0.
588 int cds_lfht_get_count_order_ulong(unsigned long x
)
593 return cds_lfht_fls_ulong(x
- 1);
597 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
600 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
601 unsigned long count
);
603 static void mutex_lock(pthread_mutex_t
*mutex
)
607 #ifndef DISTRUST_SIGNALS_EXTREME
608 ret
= pthread_mutex_lock(mutex
);
611 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
612 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
613 if (ret
!= EBUSY
&& ret
!= EINTR
)
615 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
616 uatomic_store(&URCU_TLS(rcu_reader
).need_mb
, 0, CMM_SEQ_CST
);
618 (void) poll(NULL
, 0, 10);
620 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
623 static void mutex_unlock(pthread_mutex_t
*mutex
)
627 ret
= pthread_mutex_unlock(mutex
);
632 static long nr_cpus_mask
= NR_CPUS_MASK_UNINITIALIZED
;
633 static long split_count_mask
= -1;
634 static int split_count_order
= -1;
636 static void ht_init_nr_cpus_mask(void)
640 maxcpus
= get_possible_cpus_array_len();
642 nr_cpus_mask
= NR_CPUS_MASK_INIT_FAILED
;
646 * round up number of CPUs to next power of two, so we
647 * can use & for modulo.
649 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
650 nr_cpus_mask
= maxcpus
- 1;
654 void alloc_split_items_count(struct cds_lfht
*ht
)
656 if (nr_cpus_mask
== NR_CPUS_MASK_UNINITIALIZED
) {
657 ht_init_nr_cpus_mask();
658 if (nr_cpus_mask
< 0)
659 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
661 split_count_mask
= nr_cpus_mask
;
663 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
666 urcu_posix_assert(split_count_mask
>= 0);
668 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
669 ht
->split_count
= calloc(split_count_mask
+ 1,
670 sizeof(struct ht_items_count
));
671 urcu_posix_assert(ht
->split_count
);
673 ht
->split_count
= NULL
;
678 void free_split_items_count(struct cds_lfht
*ht
)
680 poison_free(ht
->split_count
);
684 int ht_get_split_count_index(unsigned long hash
)
688 urcu_posix_assert(split_count_mask
>= 0);
689 cpu
= urcu_sched_getcpu();
690 if (caa_unlikely(cpu
< 0))
691 return hash
& split_count_mask
;
693 return cpu
& split_count_mask
;
697 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
699 unsigned long split_count
, count
;
702 if (caa_unlikely(!ht
->split_count
))
704 index
= ht_get_split_count_index(hash
);
705 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
706 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
708 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
710 dbg_printf("add split count %lu\n", split_count
);
711 count
= uatomic_add_return(&ht
->count
,
712 1UL << COUNT_COMMIT_ORDER
);
713 if (caa_likely(count
& (count
- 1)))
715 /* Only if global count is power of 2 */
717 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
719 dbg_printf("add set global %lu\n", count
);
720 cds_lfht_resize_lazy_count(ht
, size
,
721 count
>> (CHAIN_LEN_TARGET
- 1));
725 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
727 unsigned long split_count
, count
;
730 if (caa_unlikely(!ht
->split_count
))
732 index
= ht_get_split_count_index(hash
);
733 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
734 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
736 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
738 dbg_printf("del split count %lu\n", split_count
);
739 count
= uatomic_add_return(&ht
->count
,
740 -(1UL << COUNT_COMMIT_ORDER
));
741 if (caa_likely(count
& (count
- 1)))
743 /* Only if global count is power of 2 */
745 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
747 dbg_printf("del set global %lu\n", count
);
749 * Don't shrink table if the number of nodes is below a
752 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
754 cds_lfht_resize_lazy_count(ht
, size
,
755 count
>> (CHAIN_LEN_TARGET
- 1));
759 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
763 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
765 count
= uatomic_read(&ht
->count
);
767 * Use bucket-local length for small table expand and for
768 * environments lacking per-cpu data support.
770 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
773 dbg_printf("WARNING: large chain length: %u.\n",
775 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
779 * Ideal growth calculated based on chain length.
781 growth
= cds_lfht_get_count_order_u32(chain_len
782 - (CHAIN_LEN_TARGET
- 1));
783 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
785 >= (1UL << (COUNT_COMMIT_ORDER
786 + split_count_order
))) {
788 * If ideal growth expands the hash table size
789 * beyond the "small hash table" sizes, use the
790 * maximum small hash table size to attempt
791 * expanding the hash table. This only applies
792 * when node accounting is available, otherwise
793 * the chain length is used to expand the hash
794 * table in every case.
796 growth
= COUNT_COMMIT_ORDER
+ split_count_order
797 - cds_lfht_get_count_order_ulong(size
);
801 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
806 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
808 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
812 int is_removed(const struct cds_lfht_node
*node
)
814 return ((unsigned long) node
) & REMOVED_FLAG
;
818 int is_bucket(struct cds_lfht_node
*node
)
820 return ((unsigned long) node
) & BUCKET_FLAG
;
824 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
826 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
830 int is_removal_owner(struct cds_lfht_node
*node
)
832 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
836 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
838 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
842 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
844 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
848 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
850 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
854 struct cds_lfht_node
*get_end(void)
856 return (struct cds_lfht_node
*) END_VALUE
;
860 int is_end(struct cds_lfht_node
*node
)
862 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
866 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
869 unsigned long old1
, old2
;
871 old1
= uatomic_read(ptr
);
878 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
883 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
885 return ht
->mm
->alloc_bucket_table(ht
, order
);
889 * cds_lfht_free_bucket_table() should be called with decreasing order.
890 * When cds_lfht_free_bucket_table(0) is called, it means the whole
894 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
896 return ht
->mm
->free_bucket_table(ht
, order
);
900 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
902 return ht
->bucket_at(ht
, index
);
906 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
909 urcu_posix_assert(size
> 0);
910 return bucket_at(ht
, hash
& (size
- 1));
914 * Remove all logically deleted nodes from a bucket up to a certain node key.
917 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
919 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
921 urcu_posix_assert(!is_bucket(bucket
));
922 urcu_posix_assert(!is_removed(bucket
));
923 urcu_posix_assert(!is_removal_owner(bucket
));
924 urcu_posix_assert(!is_bucket(node
));
925 urcu_posix_assert(!is_removed(node
));
926 urcu_posix_assert(!is_removal_owner(node
));
929 /* We can always skip the bucket node initially */
930 iter
= rcu_dereference(iter_prev
->next
);
931 urcu_posix_assert(!is_removed(iter
));
932 urcu_posix_assert(!is_removal_owner(iter
));
933 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
935 * We should never be called with bucket (start of chain)
936 * and logically removed node (end of path compression
937 * marker) being the actual same node. This would be a
938 * bug in the algorithm implementation.
940 urcu_posix_assert(bucket
!= node
);
942 if (caa_unlikely(is_end(iter
)))
944 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
946 next
= rcu_dereference(clear_flag(iter
)->next
);
947 if (caa_likely(is_removed(next
)))
949 iter_prev
= clear_flag(iter
);
952 urcu_posix_assert(!is_removed(iter
));
953 urcu_posix_assert(!is_removal_owner(iter
));
955 new_next
= flag_bucket(clear_flag(next
));
957 new_next
= clear_flag(next
);
958 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
963 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
964 struct cds_lfht_node
*old_node
,
965 struct cds_lfht_node
*old_next
,
966 struct cds_lfht_node
*new_node
)
968 struct cds_lfht_node
*bucket
, *ret_next
;
970 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
973 urcu_posix_assert(!is_removed(old_node
));
974 urcu_posix_assert(!is_removal_owner(old_node
));
975 urcu_posix_assert(!is_bucket(old_node
));
976 urcu_posix_assert(!is_removed(new_node
));
977 urcu_posix_assert(!is_removal_owner(new_node
));
978 urcu_posix_assert(!is_bucket(new_node
));
979 urcu_posix_assert(new_node
!= old_node
);
981 /* Insert after node to be replaced */
982 if (is_removed(old_next
)) {
984 * Too late, the old node has been removed under us
985 * between lookup and replace. Fail.
989 urcu_posix_assert(old_next
== clear_flag(old_next
));
990 urcu_posix_assert(new_node
!= old_next
);
992 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
993 * flag. It is either set atomically at the same time
994 * (replace) or after (del).
996 urcu_posix_assert(!is_removal_owner(old_next
));
997 new_node
->next
= old_next
;
999 * Here is the whole trick for lock-free replace: we add
1000 * the replacement node _after_ the node we want to
1001 * replace by atomically setting its next pointer at the
1002 * same time we set its removal flag. Given that
1003 * the lookups/get next use an iterator aware of the
1004 * next pointer, they will either skip the old node due
1005 * to the removal flag and see the new node, or use
1006 * the old node, but will not see the new one.
1007 * This is a replacement of a node with another node
1008 * that has the same value: we are therefore not
1009 * removing a value from the hash table. We set both the
1010 * REMOVED and REMOVAL_OWNER flags atomically so we own
1011 * the node after successful cmpxchg.
1013 ret_next
= uatomic_cmpxchg(&old_node
->next
,
1014 old_next
, flag_removed_or_removal_owner(new_node
));
1015 if (ret_next
== old_next
)
1016 break; /* We performed the replacement. */
1017 old_next
= ret_next
;
1021 * Ensure that the old node is not visible to readers anymore:
1022 * lookup for the node, and remove it (along with any other
1023 * logically removed node) if found.
1025 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
1026 _cds_lfht_gc_bucket(bucket
, new_node
);
1028 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
1033 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1034 * mode. A NULL unique_ret allows creation of duplicate keys.
1037 void _cds_lfht_add(struct cds_lfht
*ht
,
1039 cds_lfht_match_fct match
,
1042 struct cds_lfht_node
*node
,
1043 struct cds_lfht_iter
*unique_ret
,
1046 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
1048 struct cds_lfht_node
*bucket
;
1050 urcu_posix_assert(!is_bucket(node
));
1051 urcu_posix_assert(!is_removed(node
));
1052 urcu_posix_assert(!is_removal_owner(node
));
1053 bucket
= lookup_bucket(ht
, size
, hash
);
1055 uint32_t chain_len
= 0;
1058 * iter_prev points to the non-removed node prior to the
1062 /* We can always skip the bucket node initially */
1063 iter
= rcu_dereference(iter_prev
->next
);
1064 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1066 if (caa_unlikely(is_end(iter
)))
1068 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1071 /* bucket node is the first node of the identical-hash-value chain */
1072 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1075 next
= rcu_dereference(clear_flag(iter
)->next
);
1076 if (caa_unlikely(is_removed(next
)))
1082 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1083 struct cds_lfht_iter d_iter
= {
1086 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1092 * uniquely adding inserts the node as the first
1093 * node of the identical-hash-value node chain.
1095 * This semantic ensures no duplicated keys
1096 * should ever be observable in the table
1097 * (including traversing the table node by
1098 * node by forward iterations)
1100 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1104 *unique_ret
= d_iter
;
1108 /* Only account for identical reverse hash once */
1109 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1110 && !is_bucket(next
))
1111 check_resize(ht
, size
, ++chain_len
);
1112 iter_prev
= clear_flag(iter
);
1117 urcu_posix_assert(node
!= clear_flag(iter
));
1118 urcu_posix_assert(!is_removed(iter_prev
));
1119 urcu_posix_assert(!is_removal_owner(iter_prev
));
1120 urcu_posix_assert(!is_removed(iter
));
1121 urcu_posix_assert(!is_removal_owner(iter
));
1122 urcu_posix_assert(iter_prev
!= node
);
1124 node
->next
= clear_flag(iter
);
1126 node
->next
= flag_bucket(clear_flag(iter
));
1127 if (is_bucket(iter
))
1128 new_node
= flag_bucket(node
);
1131 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1132 new_node
) != iter
) {
1133 continue; /* retry */
1140 urcu_posix_assert(!is_removed(iter
));
1141 urcu_posix_assert(!is_removal_owner(iter
));
1142 if (is_bucket(iter
))
1143 new_next
= flag_bucket(clear_flag(next
));
1145 new_next
= clear_flag(next
);
1146 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1151 unique_ret
->node
= return_node
;
1152 /* unique_ret->next left unset, never used. */
1157 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1158 struct cds_lfht_node
*node
)
1160 struct cds_lfht_node
*bucket
, *next
;
1161 uintptr_t *node_next
;
1163 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1166 /* logically delete the node */
1167 urcu_posix_assert(!is_bucket(node
));
1168 urcu_posix_assert(!is_removed(node
));
1169 urcu_posix_assert(!is_removal_owner(node
));
1172 * We are first checking if the node had previously been
1173 * logically removed (this check is not atomic with setting the
1174 * logical removal flag). Return -ENOENT if the node had
1175 * previously been removed.
1177 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1178 if (caa_unlikely(is_removed(next
)))
1180 urcu_posix_assert(!is_bucket(next
));
1182 * The del operation semantic guarantees a full memory barrier
1183 * before the uatomic_or atomic commit of the deletion flag.
1185 * We set the REMOVED_FLAG unconditionally. Note that there may
1186 * be more than one concurrent thread setting this flag.
1187 * Knowing which wins the race will be known after the garbage
1188 * collection phase, stay tuned!
1190 * NOTE: The node_next variable is present to avoid breaking
1191 * strict-aliasing rules.
1193 node_next
= (uintptr_t*)&node
->next
;
1194 uatomic_or_mo(node_next
, REMOVED_FLAG
, CMM_RELEASE
);
1196 /* We performed the (logical) deletion. */
1199 * Ensure that the node is not visible to readers anymore: lookup for
1200 * the node, and remove it (along with any other logically removed node)
1203 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1204 _cds_lfht_gc_bucket(bucket
, node
);
1206 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1208 * Last phase: atomically exchange node->next with a version
1209 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1210 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1211 * the node and win the removal race.
1212 * It is interesting to note that all "add" paths are forbidden
1213 * to change the next pointer starting from the point where the
1214 * REMOVED_FLAG is set, so here using a read, followed by a
1215 * xchg() suffice to guarantee that the xchg() will ever only
1216 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1219 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1220 flag_removal_owner(uatomic_load(&node
->next
, CMM_RELAXED
)))))
1227 void *partition_resize_thread(void *arg
)
1229 struct partition_resize_work
*work
= arg
;
1231 work
->ht
->flavor
->register_thread();
1232 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1233 work
->ht
->flavor
->unregister_thread();
1238 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1240 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1241 unsigned long start
, unsigned long len
))
1243 unsigned long partition_len
, start
= 0;
1244 struct partition_resize_work
*work
;
1246 unsigned long thread
, nr_threads
;
1247 sigset_t newmask
, oldmask
;
1249 urcu_posix_assert(nr_cpus_mask
!= NR_CPUS_MASK_UNINITIALIZED
);
1250 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
)
1254 * Note: nr_cpus_mask + 1 is always power of 2.
1255 * We spawn just the number of threads we need to satisfy the minimum
1256 * partition size, up to the number of CPUs in the system.
1258 if (nr_cpus_mask
> 0) {
1259 nr_threads
= min_t(unsigned long, nr_cpus_mask
+ 1,
1260 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1264 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1265 work
= calloc(nr_threads
, sizeof(*work
));
1267 dbg_printf("error allocating for resize, single-threading\n");
1271 ret
= sigfillset(&newmask
);
1272 urcu_posix_assert(!ret
);
1273 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
1274 urcu_posix_assert(!ret
);
1276 for (thread
= 0; thread
< nr_threads
; thread
++) {
1277 work
[thread
].ht
= ht
;
1279 work
[thread
].len
= partition_len
;
1280 work
[thread
].start
= thread
* partition_len
;
1281 work
[thread
].fct
= fct
;
1282 ret
= pthread_create(&(work
[thread
].thread_id
),
1283 ht
->caller_resize_attr
? &ht
->resize_attr
: NULL
,
1284 partition_resize_thread
, &work
[thread
]);
1285 if (ret
== EAGAIN
) {
1287 * Out of resources: wait and join the threads
1288 * we've created, then handle leftovers.
1290 dbg_printf("error spawning for resize, single-threading\n");
1291 start
= work
[thread
].start
;
1293 nr_threads
= thread
;
1296 urcu_posix_assert(!ret
);
1299 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
1300 urcu_posix_assert(!ret
);
1302 for (thread
= 0; thread
< nr_threads
; thread
++) {
1303 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1304 urcu_posix_assert(!ret
);
1309 * A pthread_create failure above will either lead in us having
1310 * no threads to join or starting at a non-zero offset,
1311 * fallback to single thread processing of leftovers.
1313 if (start
== 0 && nr_threads
> 0)
1316 fct(ht
, i
, start
, len
);
1320 * Holding RCU read lock to protect _cds_lfht_add against memory
1321 * reclaim that could be performed by other worker threads (ABA
1324 * When we reach a certain length, we can split this population phase over
1325 * many worker threads, based on the number of CPUs available in the system.
1326 * This should therefore take care of not having the expand lagging behind too
1327 * many concurrent insertion threads by using the scheduler's ability to
1328 * schedule bucket node population fairly with insertions.
1331 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1332 unsigned long start
, unsigned long len
)
1334 unsigned long j
, size
= 1UL << (i
- 1);
1336 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1337 ht
->flavor
->read_lock();
1338 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1339 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1341 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1342 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1344 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1345 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1347 ht
->flavor
->read_unlock();
1351 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1354 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1358 void init_table(struct cds_lfht
*ht
,
1359 unsigned long first_order
, unsigned long last_order
)
1363 dbg_printf("init table: first_order %lu last_order %lu\n",
1364 first_order
, last_order
);
1365 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1366 for (i
= first_order
; i
<= last_order
; i
++) {
1369 len
= 1UL << (i
- 1);
1370 dbg_printf("init order %lu len: %lu\n", i
, len
);
1372 /* Stop expand if the resize target changes under us */
1373 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1376 cds_lfht_alloc_bucket_table(ht
, i
);
1379 * Set all bucket nodes reverse hash values for a level and
1380 * link all bucket nodes into the table.
1382 init_table_populate(ht
, i
, len
);
1385 * Update table size.
1387 * Populate data before RCU size.
1389 uatomic_store(&ht
->size
, 1UL << i
, CMM_RELEASE
);
1391 dbg_printf("init new size: %lu\n", 1UL << i
);
1392 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1398 * Holding RCU read lock to protect _cds_lfht_remove against memory
1399 * reclaim that could be performed by other worker threads (ABA
1401 * For a single level, we logically remove and garbage collect each node.
1403 * As a design choice, we perform logical removal and garbage collection on a
1404 * node-per-node basis to simplify this algorithm. We also assume keeping good
1405 * cache locality of the operation would overweight possible performance gain
1406 * that could be achieved by batching garbage collection for multiple levels.
1407 * However, this would have to be justified by benchmarks.
1409 * Concurrent removal and add operations are helping us perform garbage
1410 * collection of logically removed nodes. We guarantee that all logically
1411 * removed nodes have been garbage-collected (unlinked) before work
1412 * enqueue is invoked to free a hole level of bucket nodes (after a
1415 * Logical removal and garbage collection can therefore be done in batch
1416 * or on a node-per-node basis, as long as the guarantee above holds.
1418 * When we reach a certain length, we can split this removal over many worker
1419 * threads, based on the number of CPUs available in the system. This should
1420 * take care of not letting resize process lag behind too many concurrent
1421 * updater threads actively inserting into the hash table.
1424 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1425 unsigned long start
, unsigned long len
)
1427 unsigned long j
, size
= 1UL << (i
- 1);
1429 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1430 ht
->flavor
->read_lock();
1431 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1432 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1433 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1434 uintptr_t *fini_bucket_next
;
1436 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1437 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1439 /* Set the REMOVED_FLAG to freeze the ->next for gc.
1441 * NOTE: The fini_bucket_next variable is present to
1442 * avoid breaking strict-aliasing rules.
1444 fini_bucket_next
= (uintptr_t*)&fini_bucket
->next
;
1445 uatomic_or(fini_bucket_next
, REMOVED_FLAG
);
1446 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1448 ht
->flavor
->read_unlock();
1452 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1454 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1458 * fini_table() is never called for first_order == 0, which is why
1459 * free_by_rcu_order == 0 can be used as criterion to know if free must
1463 void fini_table(struct cds_lfht
*ht
,
1464 unsigned long first_order
, unsigned long last_order
)
1466 unsigned long free_by_rcu_order
= 0, i
;
1468 dbg_printf("fini table: first_order %lu last_order %lu\n",
1469 first_order
, last_order
);
1470 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1471 for (i
= last_order
; i
>= first_order
; i
--) {
1474 len
= 1UL << (i
- 1);
1475 dbg_printf("fini order %ld len: %lu\n", i
, len
);
1477 /* Stop shrink if the resize target changes under us */
1478 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1481 cmm_smp_wmb(); /* populate data before RCU size */
1482 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1485 * We need to wait for all add operations to reach Q.S. (and
1486 * thus use the new table for lookups) before we can start
1487 * releasing the old bucket nodes. Otherwise their lookup will
1488 * return a logically removed node as insert position.
1490 ht
->flavor
->update_synchronize_rcu();
1491 if (free_by_rcu_order
)
1492 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1495 * Set "removed" flag in bucket nodes about to be removed.
1496 * Unlink all now-logically-removed bucket node pointers.
1497 * Concurrent add/remove operation are helping us doing
1500 remove_table(ht
, i
, len
);
1502 free_by_rcu_order
= i
;
1504 dbg_printf("fini new size: %lu\n", 1UL << i
);
1505 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1509 if (free_by_rcu_order
) {
1510 ht
->flavor
->update_synchronize_rcu();
1511 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1516 * Never called with size < 1.
1519 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1521 struct cds_lfht_node
*prev
, *node
;
1522 unsigned long order
, len
, i
;
1525 cds_lfht_alloc_bucket_table(ht
, 0);
1527 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1528 node
= bucket_at(ht
, 0);
1529 node
->next
= flag_bucket(get_end());
1530 node
->reverse_hash
= 0;
1532 bucket_order
= cds_lfht_get_count_order_ulong(size
);
1533 urcu_posix_assert(bucket_order
>= 0);
1535 for (order
= 1; order
< (unsigned long) bucket_order
+ 1; order
++) {
1536 len
= 1UL << (order
- 1);
1537 cds_lfht_alloc_bucket_table(ht
, order
);
1539 for (i
= 0; i
< len
; i
++) {
1541 * Now, we are trying to init the node with the
1542 * hash=(len+i) (which is also a bucket with the
1543 * index=(len+i)) and insert it into the hash table,
1544 * so this node has to be inserted after the bucket
1545 * with the index=(len+i)&(len-1)=i. And because there
1546 * is no other non-bucket node nor bucket node with
1547 * larger index/hash inserted, so the bucket node
1548 * being inserted should be inserted directly linked
1549 * after the bucket node with index=i.
1551 prev
= bucket_at(ht
, i
);
1552 node
= bucket_at(ht
, len
+ i
);
1554 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1555 order
, len
+ i
, len
+ i
);
1556 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1558 /* insert after prev */
1559 urcu_posix_assert(is_bucket(prev
->next
));
1560 node
->next
= prev
->next
;
1561 prev
->next
= flag_bucket(node
);
1566 #if (CAA_BITS_PER_LONG > 32)
1568 * For 64-bit architectures, with max number of buckets small enough not to
1569 * use the entire 64-bit memory mapping space (and allowing a fair number of
1570 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1571 * fallback to the order allocator.
1574 const struct cds_lfht_mm_type
*get_mm_type(unsigned long max_nr_buckets
)
1576 if (max_nr_buckets
&& max_nr_buckets
<= (1ULL << 32))
1577 return &cds_lfht_mm_mmap
;
1579 return &cds_lfht_mm_order
;
1583 * For 32-bit architectures, use the order allocator.
1586 const struct cds_lfht_mm_type
*get_mm_type(
1587 unsigned long max_nr_buckets
__attribute__((unused
)))
1589 return &cds_lfht_mm_order
;
1593 void cds_lfht_node_init_deleted(struct cds_lfht_node
*node
)
1595 cds_lfht_node_init(node
);
1596 node
->next
= flag_removed(NULL
);
1599 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1600 unsigned long min_nr_alloc_buckets
,
1601 unsigned long max_nr_buckets
,
1603 const struct cds_lfht_mm_type
*mm
,
1604 const struct rcu_flavor_struct
*flavor
,
1605 pthread_attr_t
*attr
)
1607 struct cds_lfht
*ht
;
1608 unsigned long order
;
1610 /* min_nr_alloc_buckets must be power of two */
1611 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1614 /* init_size must be power of two */
1615 if (!init_size
|| (init_size
& (init_size
- 1)))
1619 * Memory management plugin default.
1622 mm
= get_mm_type(max_nr_buckets
);
1624 /* max_nr_buckets == 0 for order based mm means infinite */
1625 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1626 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1628 /* max_nr_buckets must be power of two */
1629 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1632 if (flags
& CDS_LFHT_AUTO_RESIZE
)
1633 cds_lfht_init_worker(flavor
);
1635 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1636 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1637 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1638 init_size
= min(init_size
, max_nr_buckets
);
1640 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1641 urcu_posix_assert(ht
);
1642 urcu_posix_assert(ht
->mm
== mm
);
1643 urcu_posix_assert(ht
->bucket_at
== mm
->bucket_at
);
1646 ht
->flavor
= flavor
;
1647 ht
->caller_resize_attr
= attr
;
1649 ht
->resize_attr
= *attr
;
1650 alloc_split_items_count(ht
);
1651 /* this mutex should not nest in read-side C.S. */
1652 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1653 order
= cds_lfht_get_count_order_ulong(init_size
);
1654 ht
->resize_target
= 1UL << order
;
1655 cds_lfht_create_bucket(ht
, 1UL << order
);
1656 ht
->size
= 1UL << order
;
1660 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1661 cds_lfht_match_fct match
, const void *key
,
1662 struct cds_lfht_iter
*iter
)
1664 struct cds_lfht_node
*node
, *next
, *bucket
;
1665 unsigned long reverse_hash
, size
;
1667 cds_lfht_iter_debug_set_ht(ht
, iter
);
1669 reverse_hash
= bit_reverse_ulong(hash
);
1672 * Use load acquire instead of rcu_dereference because there is no
1673 * dependency between the table size and the dereference of the bucket
1676 * This acquire is paired with the store release in init_table().
1678 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1679 bucket
= lookup_bucket(ht
, size
, hash
);
1680 /* We can always skip the bucket node initially */
1681 node
= rcu_dereference(bucket
->next
);
1682 node
= clear_flag(node
);
1684 if (caa_unlikely(is_end(node
))) {
1688 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1692 next
= rcu_dereference(node
->next
);
1693 urcu_posix_assert(node
== clear_flag(node
));
1694 if (caa_likely(!is_removed(next
))
1696 && node
->reverse_hash
== reverse_hash
1697 && caa_likely(match(node
, key
))) {
1700 node
= clear_flag(next
);
1702 urcu_posix_assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1707 void cds_lfht_next_duplicate(struct cds_lfht
*ht
__attribute__((unused
)),
1708 cds_lfht_match_fct match
,
1709 const void *key
, struct cds_lfht_iter
*iter
)
1711 struct cds_lfht_node
*node
, *next
;
1712 unsigned long reverse_hash
;
1714 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1716 reverse_hash
= node
->reverse_hash
;
1718 node
= clear_flag(next
);
1721 if (caa_unlikely(is_end(node
))) {
1725 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1729 next
= rcu_dereference(node
->next
);
1730 if (caa_likely(!is_removed(next
))
1732 && caa_likely(match(node
, key
))) {
1735 node
= clear_flag(next
);
1737 urcu_posix_assert(!node
|| !is_bucket(uatomic_load(&node
->next
, CMM_RELAXED
)));
1742 void cds_lfht_next(struct cds_lfht
*ht
__attribute__((unused
)),
1743 struct cds_lfht_iter
*iter
)
1745 struct cds_lfht_node
*node
, *next
;
1747 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1748 node
= clear_flag(iter
->next
);
1750 if (caa_unlikely(is_end(node
))) {
1754 next
= rcu_dereference(node
->next
);
1755 if (caa_likely(!is_removed(next
))
1756 && !is_bucket(next
)) {
1759 node
= clear_flag(next
);
1761 urcu_posix_assert(!node
|| !is_bucket(uatomic_load(&node
->next
, CMM_RELAXED
)));
1766 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1768 cds_lfht_iter_debug_set_ht(ht
, iter
);
1770 * Get next after first bucket node. The first bucket node is the
1771 * first node of the linked list.
1773 iter
->next
= uatomic_load(&bucket_at(ht
, 0)->next
, CMM_CONSUME
);
1774 cds_lfht_next(ht
, iter
);
1777 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1778 struct cds_lfht_node
*node
)
1782 node
->reverse_hash
= bit_reverse_ulong(hash
);
1783 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1784 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1785 ht_count_add(ht
, size
, hash
);
1788 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1790 cds_lfht_match_fct match
,
1792 struct cds_lfht_node
*node
)
1795 struct cds_lfht_iter iter
;
1797 node
->reverse_hash
= bit_reverse_ulong(hash
);
1798 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1799 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1800 if (iter
.node
== node
)
1801 ht_count_add(ht
, size
, hash
);
1805 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1807 cds_lfht_match_fct match
,
1809 struct cds_lfht_node
*node
)
1812 struct cds_lfht_iter iter
;
1814 node
->reverse_hash
= bit_reverse_ulong(hash
);
1815 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1817 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1818 if (iter
.node
== node
) {
1819 ht_count_add(ht
, size
, hash
);
1823 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1828 int cds_lfht_replace(struct cds_lfht
*ht
,
1829 struct cds_lfht_iter
*old_iter
,
1831 cds_lfht_match_fct match
,
1833 struct cds_lfht_node
*new_node
)
1837 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1838 if (!old_iter
->node
)
1840 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1842 if (caa_unlikely(!match(old_iter
->node
, key
)))
1844 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1845 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1849 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1854 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1855 ret
= _cds_lfht_del(ht
, size
, node
);
1859 hash
= bit_reverse_ulong(node
->reverse_hash
);
1860 ht_count_del(ht
, size
, hash
);
1865 int cds_lfht_is_node_deleted(const struct cds_lfht_node
*node
)
1867 return is_removed(CMM_LOAD_SHARED(node
->next
));
1871 bool cds_lfht_is_empty(struct cds_lfht
*ht
)
1873 struct cds_lfht_node
*node
, *next
;
1877 was_online
= ht
->flavor
->read_ongoing();
1879 ht
->flavor
->thread_online();
1880 ht
->flavor
->read_lock();
1882 /* Check that the table is empty */
1883 node
= bucket_at(ht
, 0);
1885 next
= rcu_dereference(node
->next
);
1886 if (!is_bucket(next
)) {
1890 node
= clear_flag(next
);
1891 } while (!is_end(node
));
1893 ht
->flavor
->read_unlock();
1894 ht
->flavor
->thread_offline();
1900 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1902 struct cds_lfht_node
*node
;
1903 unsigned long order
, i
, size
;
1905 /* Check that the table is empty */
1906 node
= bucket_at(ht
, 0);
1908 node
= clear_flag(node
)->next
;
1909 if (!is_bucket(node
))
1911 urcu_posix_assert(!is_removed(node
));
1912 urcu_posix_assert(!is_removal_owner(node
));
1913 } while (!is_end(node
));
1915 * size accessed without rcu_dereference because hash table is
1919 /* Internal sanity check: all nodes left should be buckets */
1920 for (i
= 0; i
< size
; i
++) {
1921 node
= bucket_at(ht
, i
);
1922 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1923 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1924 urcu_posix_assert(is_bucket(node
->next
));
1927 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1928 cds_lfht_free_bucket_table(ht
, order
);
1934 void do_auto_resize_destroy_cb(struct urcu_work
*work
)
1936 struct cds_lfht
*ht
= caa_container_of(work
, struct cds_lfht
, destroy_work
);
1939 ht
->flavor
->register_thread();
1940 ret
= cds_lfht_delete_bucket(ht
);
1943 free_split_items_count(ht
);
1944 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1947 ht
->flavor
->unregister_thread();
1952 * Should only be called when no more concurrent readers nor writers can
1953 * possibly access the table.
1955 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1959 if (ht
->flags
& CDS_LFHT_AUTO_RESIZE
) {
1961 * Perform error-checking for emptiness before queuing
1962 * work, so we can return error to the caller. This runs
1963 * concurrently with ongoing resize.
1965 if (!cds_lfht_is_empty(ht
))
1967 /* Cancel ongoing resize operations. */
1968 uatomic_store(&ht
->in_progress_destroy
, 1, CMM_RELAXED
);
1970 *attr
= ht
->caller_resize_attr
;
1971 ht
->caller_resize_attr
= NULL
;
1974 * Queue destroy work after prior queued resize
1975 * operations. Given there are no concurrent writers
1976 * accessing the hash table at this point, no resize
1977 * operations can be queued after this destroy work.
1979 urcu_workqueue_queue_work(cds_lfht_workqueue
,
1980 &ht
->destroy_work
, do_auto_resize_destroy_cb
);
1983 ret
= cds_lfht_delete_bucket(ht
);
1986 free_split_items_count(ht
);
1988 *attr
= ht
->caller_resize_attr
;
1989 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1996 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1997 long *approx_before
,
1998 unsigned long *count
,
2001 struct cds_lfht_node
*node
, *next
;
2002 unsigned long nr_bucket
= 0, nr_removed
= 0;
2005 if (ht
->split_count
) {
2008 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
2009 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
2010 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
2016 /* Count non-bucket nodes in the table */
2017 node
= bucket_at(ht
, 0);
2019 next
= rcu_dereference(node
->next
);
2020 if (is_removed(next
)) {
2021 if (!is_bucket(next
))
2025 } else if (!is_bucket(next
))
2029 node
= clear_flag(next
);
2030 } while (!is_end(node
));
2031 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
2032 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
2034 if (ht
->split_count
) {
2037 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
2038 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
2039 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
2044 /* called with resize mutex held */
2046 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
2047 unsigned long old_size
, unsigned long new_size
)
2049 unsigned long old_order
, new_order
;
2051 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2052 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2053 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2054 old_size
, old_order
, new_size
, new_order
);
2055 urcu_posix_assert(new_size
> old_size
);
2056 init_table(ht
, old_order
+ 1, new_order
);
2059 /* called with resize mutex held */
2061 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
2062 unsigned long old_size
, unsigned long new_size
)
2064 unsigned long old_order
, new_order
;
2066 new_size
= max(new_size
, MIN_TABLE_SIZE
);
2067 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2068 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2069 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2070 old_size
, old_order
, new_size
, new_order
);
2071 urcu_posix_assert(new_size
< old_size
);
2073 /* Remove and unlink all bucket nodes to remove. */
2074 fini_table(ht
, new_order
+ 1, old_order
);
2078 /* called with resize mutex held */
2080 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
2082 unsigned long new_size
, old_size
;
2085 * Resize table, re-do if the target size has changed under us.
2088 if (uatomic_load(&ht
->in_progress_destroy
, CMM_RELAXED
))
2091 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2093 old_size
= ht
->size
;
2094 new_size
= uatomic_load(&ht
->resize_target
, CMM_RELAXED
);
2095 if (old_size
< new_size
)
2096 _do_cds_lfht_grow(ht
, old_size
, new_size
);
2097 else if (old_size
> new_size
)
2098 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
2100 uatomic_store(&ht
->resize_initiated
, 0, CMM_RELAXED
);
2101 /* write resize_initiated before read resize_target */
2103 } while (ht
->size
!= uatomic_load(&ht
->resize_target
, CMM_RELAXED
));
2107 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
2109 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
2113 void resize_target_update_count(struct cds_lfht
*ht
,
2114 unsigned long count
)
2116 count
= max(count
, MIN_TABLE_SIZE
);
2117 count
= min(count
, ht
->max_nr_buckets
);
2118 uatomic_set(&ht
->resize_target
, count
);
2121 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
2123 resize_target_update_count(ht
, new_size
);
2126 * Set flags has early as possible even in contention case.
2128 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2130 mutex_lock(&ht
->resize_mutex
);
2131 _do_cds_lfht_resize(ht
);
2132 mutex_unlock(&ht
->resize_mutex
);
2136 void do_resize_cb(struct urcu_work
*work
)
2138 struct resize_work
*resize_work
=
2139 caa_container_of(work
, struct resize_work
, work
);
2140 struct cds_lfht
*ht
= resize_work
->ht
;
2142 ht
->flavor
->register_thread();
2143 mutex_lock(&ht
->resize_mutex
);
2144 _do_cds_lfht_resize(ht
);
2145 mutex_unlock(&ht
->resize_mutex
);
2146 ht
->flavor
->unregister_thread();
2151 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
2153 struct resize_work
*work
;
2156 * Store to resize_target is before read resize_initiated as guaranteed
2157 * by either cmpxchg or _uatomic_xchg_monotonic_increase.
2159 if (!uatomic_load(&ht
->resize_initiated
, CMM_RELAXED
)) {
2160 if (uatomic_load(&ht
->in_progress_destroy
, CMM_RELAXED
)) {
2163 work
= malloc(sizeof(*work
));
2165 dbg_printf("error allocating resize work, bailing out\n");
2169 urcu_workqueue_queue_work(cds_lfht_workqueue
,
2170 &work
->work
, do_resize_cb
);
2171 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2176 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2178 unsigned long target_size
= size
<< growth
;
2180 target_size
= min(target_size
, ht
->max_nr_buckets
);
2181 if (resize_target_grow(ht
, target_size
) >= target_size
)
2184 __cds_lfht_resize_lazy_launch(ht
);
2188 * We favor grow operations over shrink. A shrink operation never occurs
2189 * if a grow operation is queued for lazy execution. A grow operation
2190 * cancels any pending shrink lazy execution.
2193 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2194 unsigned long count
)
2196 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2198 count
= max(count
, MIN_TABLE_SIZE
);
2199 count
= min(count
, ht
->max_nr_buckets
);
2201 return; /* Already the right size, no resize needed */
2202 if (count
> size
) { /* lazy grow */
2203 if (resize_target_grow(ht
, count
) >= count
)
2205 } else { /* lazy shrink */
2209 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2211 break; /* no resize needed */
2213 return; /* growing is/(was just) in progress */
2215 return; /* some other thread do shrink */
2219 __cds_lfht_resize_lazy_launch(ht
);
2222 static void cds_lfht_before_fork(void *priv
__attribute__((unused
)))
2224 if (cds_lfht_workqueue_atfork_nesting
++)
2226 mutex_lock(&cds_lfht_fork_mutex
);
2227 if (!cds_lfht_workqueue
)
2229 urcu_workqueue_pause_worker(cds_lfht_workqueue
);
2232 static void cds_lfht_after_fork_parent(void *priv
__attribute__((unused
)))
2234 if (--cds_lfht_workqueue_atfork_nesting
)
2236 if (!cds_lfht_workqueue
)
2238 urcu_workqueue_resume_worker(cds_lfht_workqueue
);
2240 mutex_unlock(&cds_lfht_fork_mutex
);
2243 static void cds_lfht_after_fork_child(void *priv
__attribute__((unused
)))
2245 if (--cds_lfht_workqueue_atfork_nesting
)
2247 if (!cds_lfht_workqueue
)
2249 urcu_workqueue_create_worker(cds_lfht_workqueue
);
2251 mutex_unlock(&cds_lfht_fork_mutex
);
2254 static struct urcu_atfork cds_lfht_atfork
= {
2255 .before_fork
= cds_lfht_before_fork
,
2256 .after_fork_parent
= cds_lfht_after_fork_parent
,
2257 .after_fork_child
= cds_lfht_after_fork_child
,
2260 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
)
2262 flavor
->register_rculfhash_atfork(&cds_lfht_atfork
);
2264 mutex_lock(&cds_lfht_fork_mutex
);
2265 if (!cds_lfht_workqueue
)
2266 cds_lfht_workqueue
= urcu_workqueue_create(0, -1, NULL
,
2267 NULL
, NULL
, NULL
, NULL
, NULL
, NULL
, NULL
);
2268 mutex_unlock(&cds_lfht_fork_mutex
);
2271 static void cds_lfht_exit(void)
2273 mutex_lock(&cds_lfht_fork_mutex
);
2274 if (cds_lfht_workqueue
) {
2275 urcu_workqueue_flush_queued_work(cds_lfht_workqueue
);
2276 urcu_workqueue_destroy(cds_lfht_workqueue
);
2277 cds_lfht_workqueue
= NULL
;
2279 mutex_unlock(&cds_lfht_fork_mutex
);