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 static struct urcu_workqueue
*cds_lfht_workqueue
;
353 * Mutex ensuring mutual exclusion between workqueue initialization and
354 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
356 static pthread_mutex_t cds_lfht_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
358 static struct urcu_atfork cds_lfht_atfork
;
361 * atfork handler nesting counters. Handle being registered to many urcu
362 * flavors, thus being possibly invoked more than once in the
363 * pthread_atfork list of callbacks.
365 static int cds_lfht_workqueue_atfork_nesting
;
367 static void __attribute__((destructor
)) cds_lfht_exit(void);
368 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
);
370 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
373 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
378 #define cds_lfht_iter_debug_assert(...) urcu_posix_assert(__VA_ARGS__)
383 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
__attribute__((unused
)),
384 struct cds_lfht_iter
*iter
__attribute__((unused
)))
388 #define cds_lfht_iter_debug_assert(...)
393 * Algorithm to reverse bits in a word by lookup table, extended to
396 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
397 * Originally from Public Domain.
400 static const uint8_t BitReverseTable256
[256] =
402 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
403 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
404 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
405 R6(0), R6(2), R6(1), R6(3)
412 uint8_t bit_reverse_u8(uint8_t v
)
414 return BitReverseTable256
[v
];
417 #if (CAA_BITS_PER_LONG == 32)
419 uint32_t bit_reverse_u32(uint32_t v
)
421 return ((uint32_t) bit_reverse_u8(v
) << 24) |
422 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
423 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
424 ((uint32_t) bit_reverse_u8(v
>> 24));
428 uint64_t bit_reverse_u64(uint64_t v
)
430 return ((uint64_t) bit_reverse_u8(v
) << 56) |
431 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
432 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
433 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
434 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
435 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
436 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
437 ((uint64_t) bit_reverse_u8(v
>> 56));
442 unsigned long bit_reverse_ulong(unsigned long v
)
444 #if (CAA_BITS_PER_LONG == 32)
445 return bit_reverse_u32(v
);
447 return bit_reverse_u64(v
);
452 * fls: returns the position of the most significant bit.
453 * Returns 0 if no bit is set, else returns the position of the most
454 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
456 #if defined(URCU_ARCH_X86)
458 unsigned int fls_u32(uint32_t x
)
462 __asm__ ("bsrl %1,%0\n\t"
466 : "=r" (r
) : "rm" (x
));
472 #if defined(URCU_ARCH_AMD64)
474 unsigned int fls_u64(uint64_t x
)
478 __asm__ ("bsrq %1,%0\n\t"
482 : "=r" (r
) : "rm" (x
));
489 static __attribute__((unused
))
490 unsigned int fls_u64(uint64_t x
)
497 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
501 if (!(x
& 0xFFFF000000000000ULL
)) {
505 if (!(x
& 0xFF00000000000000ULL
)) {
509 if (!(x
& 0xF000000000000000ULL
)) {
513 if (!(x
& 0xC000000000000000ULL
)) {
517 if (!(x
& 0x8000000000000000ULL
)) {
526 static __attribute__((unused
))
527 unsigned int fls_u32(uint32_t x
)
533 if (!(x
& 0xFFFF0000U
)) {
537 if (!(x
& 0xFF000000U
)) {
541 if (!(x
& 0xF0000000U
)) {
545 if (!(x
& 0xC0000000U
)) {
549 if (!(x
& 0x80000000U
)) {
557 unsigned int cds_lfht_fls_ulong(unsigned long x
)
559 #if (CAA_BITS_PER_LONG == 32)
567 * Return the minimum order for which x <= (1UL << order).
568 * Return -1 if x is 0.
571 int cds_lfht_get_count_order_u32(uint32_t x
)
576 return fls_u32(x
- 1);
580 * Return the minimum order for which x <= (1UL << order).
581 * Return -1 if x is 0.
583 int cds_lfht_get_count_order_ulong(unsigned long x
)
588 return cds_lfht_fls_ulong(x
- 1);
592 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
595 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
596 unsigned long count
);
598 static void mutex_lock(pthread_mutex_t
*mutex
)
602 #ifndef DISTRUST_SIGNALS_EXTREME
603 ret
= pthread_mutex_lock(mutex
);
606 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
607 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
608 if (ret
!= EBUSY
&& ret
!= EINTR
)
610 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
611 uatomic_store(&URCU_TLS(rcu_reader
).need_mb
, 0, CMM_SEQ_CST
);
613 (void) poll(NULL
, 0, 10);
615 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
618 static void mutex_unlock(pthread_mutex_t
*mutex
)
622 ret
= pthread_mutex_unlock(mutex
);
627 static long nr_cpus_mask
= -1;
628 static long split_count_mask
= -1;
629 static int split_count_order
= -1;
631 static void ht_init_nr_cpus_mask(void)
635 maxcpus
= get_possible_cpus_array_len();
641 * round up number of CPUs to next power of two, so we
642 * can use & for modulo.
644 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
645 nr_cpus_mask
= maxcpus
- 1;
649 void alloc_split_items_count(struct cds_lfht
*ht
)
651 if (nr_cpus_mask
== -1) {
652 ht_init_nr_cpus_mask();
653 if (nr_cpus_mask
< 0)
654 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
656 split_count_mask
= nr_cpus_mask
;
658 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
661 urcu_posix_assert(split_count_mask
>= 0);
663 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
664 ht
->split_count
= calloc(split_count_mask
+ 1,
665 sizeof(struct ht_items_count
));
666 urcu_posix_assert(ht
->split_count
);
668 ht
->split_count
= NULL
;
673 void free_split_items_count(struct cds_lfht
*ht
)
675 poison_free(ht
->split_count
);
679 int ht_get_split_count_index(unsigned long hash
)
683 urcu_posix_assert(split_count_mask
>= 0);
684 cpu
= urcu_sched_getcpu();
685 if (caa_unlikely(cpu
< 0))
686 return hash
& split_count_mask
;
688 return cpu
& split_count_mask
;
692 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
694 unsigned long split_count
, count
;
697 if (caa_unlikely(!ht
->split_count
))
699 index
= ht_get_split_count_index(hash
);
700 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
701 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
703 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
705 dbg_printf("add split count %lu\n", split_count
);
706 count
= uatomic_add_return(&ht
->count
,
707 1UL << COUNT_COMMIT_ORDER
);
708 if (caa_likely(count
& (count
- 1)))
710 /* Only if global count is power of 2 */
712 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
714 dbg_printf("add set global %lu\n", count
);
715 cds_lfht_resize_lazy_count(ht
, size
,
716 count
>> (CHAIN_LEN_TARGET
- 1));
720 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
722 unsigned long split_count
, count
;
725 if (caa_unlikely(!ht
->split_count
))
727 index
= ht_get_split_count_index(hash
);
728 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
729 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
731 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
733 dbg_printf("del split count %lu\n", split_count
);
734 count
= uatomic_add_return(&ht
->count
,
735 -(1UL << COUNT_COMMIT_ORDER
));
736 if (caa_likely(count
& (count
- 1)))
738 /* Only if global count is power of 2 */
740 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
742 dbg_printf("del set global %lu\n", count
);
744 * Don't shrink table if the number of nodes is below a
747 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
749 cds_lfht_resize_lazy_count(ht
, size
,
750 count
>> (CHAIN_LEN_TARGET
- 1));
754 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
758 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
760 count
= uatomic_read(&ht
->count
);
762 * Use bucket-local length for small table expand and for
763 * environments lacking per-cpu data support.
765 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
768 dbg_printf("WARNING: large chain length: %u.\n",
770 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
774 * Ideal growth calculated based on chain length.
776 growth
= cds_lfht_get_count_order_u32(chain_len
777 - (CHAIN_LEN_TARGET
- 1));
778 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
780 >= (1UL << (COUNT_COMMIT_ORDER
781 + split_count_order
))) {
783 * If ideal growth expands the hash table size
784 * beyond the "small hash table" sizes, use the
785 * maximum small hash table size to attempt
786 * expanding the hash table. This only applies
787 * when node accounting is available, otherwise
788 * the chain length is used to expand the hash
789 * table in every case.
791 growth
= COUNT_COMMIT_ORDER
+ split_count_order
792 - cds_lfht_get_count_order_ulong(size
);
796 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
801 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
803 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
807 int is_removed(const struct cds_lfht_node
*node
)
809 return ((unsigned long) node
) & REMOVED_FLAG
;
813 int is_bucket(struct cds_lfht_node
*node
)
815 return ((unsigned long) node
) & BUCKET_FLAG
;
819 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
821 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
825 int is_removal_owner(struct cds_lfht_node
*node
)
827 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
831 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
833 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
837 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
839 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
843 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
845 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
849 struct cds_lfht_node
*get_end(void)
851 return (struct cds_lfht_node
*) END_VALUE
;
855 int is_end(struct cds_lfht_node
*node
)
857 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
861 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
864 unsigned long old1
, old2
;
866 old1
= uatomic_read(ptr
);
873 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
878 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
880 return ht
->mm
->alloc_bucket_table(ht
, order
);
884 * cds_lfht_free_bucket_table() should be called with decreasing order.
885 * When cds_lfht_free_bucket_table(0) is called, it means the whole
889 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
891 return ht
->mm
->free_bucket_table(ht
, order
);
895 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
897 return ht
->bucket_at(ht
, index
);
901 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
904 urcu_posix_assert(size
> 0);
905 return bucket_at(ht
, hash
& (size
- 1));
909 * Remove all logically deleted nodes from a bucket up to a certain node key.
912 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
914 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
916 urcu_posix_assert(!is_bucket(bucket
));
917 urcu_posix_assert(!is_removed(bucket
));
918 urcu_posix_assert(!is_removal_owner(bucket
));
919 urcu_posix_assert(!is_bucket(node
));
920 urcu_posix_assert(!is_removed(node
));
921 urcu_posix_assert(!is_removal_owner(node
));
924 /* We can always skip the bucket node initially */
925 iter
= rcu_dereference(iter_prev
->next
);
926 urcu_posix_assert(!is_removed(iter
));
927 urcu_posix_assert(!is_removal_owner(iter
));
928 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
930 * We should never be called with bucket (start of chain)
931 * and logically removed node (end of path compression
932 * marker) being the actual same node. This would be a
933 * bug in the algorithm implementation.
935 urcu_posix_assert(bucket
!= node
);
937 if (caa_unlikely(is_end(iter
)))
939 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
941 next
= rcu_dereference(clear_flag(iter
)->next
);
942 if (caa_likely(is_removed(next
)))
944 iter_prev
= clear_flag(iter
);
947 urcu_posix_assert(!is_removed(iter
));
948 urcu_posix_assert(!is_removal_owner(iter
));
950 new_next
= flag_bucket(clear_flag(next
));
952 new_next
= clear_flag(next
);
953 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
958 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
959 struct cds_lfht_node
*old_node
,
960 struct cds_lfht_node
*old_next
,
961 struct cds_lfht_node
*new_node
)
963 struct cds_lfht_node
*bucket
, *ret_next
;
965 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
968 urcu_posix_assert(!is_removed(old_node
));
969 urcu_posix_assert(!is_removal_owner(old_node
));
970 urcu_posix_assert(!is_bucket(old_node
));
971 urcu_posix_assert(!is_removed(new_node
));
972 urcu_posix_assert(!is_removal_owner(new_node
));
973 urcu_posix_assert(!is_bucket(new_node
));
974 urcu_posix_assert(new_node
!= old_node
);
976 /* Insert after node to be replaced */
977 if (is_removed(old_next
)) {
979 * Too late, the old node has been removed under us
980 * between lookup and replace. Fail.
984 urcu_posix_assert(old_next
== clear_flag(old_next
));
985 urcu_posix_assert(new_node
!= old_next
);
987 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
988 * flag. It is either set atomically at the same time
989 * (replace) or after (del).
991 urcu_posix_assert(!is_removal_owner(old_next
));
992 new_node
->next
= old_next
;
994 * Here is the whole trick for lock-free replace: we add
995 * the replacement node _after_ the node we want to
996 * replace by atomically setting its next pointer at the
997 * same time we set its removal flag. Given that
998 * the lookups/get next use an iterator aware of the
999 * next pointer, they will either skip the old node due
1000 * to the removal flag and see the new node, or use
1001 * the old node, but will not see the new one.
1002 * This is a replacement of a node with another node
1003 * that has the same value: we are therefore not
1004 * removing a value from the hash table. We set both the
1005 * REMOVED and REMOVAL_OWNER flags atomically so we own
1006 * the node after successful cmpxchg.
1008 ret_next
= uatomic_cmpxchg(&old_node
->next
,
1009 old_next
, flag_removed_or_removal_owner(new_node
));
1010 if (ret_next
== old_next
)
1011 break; /* We performed the replacement. */
1012 old_next
= ret_next
;
1016 * Ensure that the old node is not visible to readers anymore:
1017 * lookup for the node, and remove it (along with any other
1018 * logically removed node) if found.
1020 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
1021 _cds_lfht_gc_bucket(bucket
, new_node
);
1023 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
1028 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1029 * mode. A NULL unique_ret allows creation of duplicate keys.
1032 void _cds_lfht_add(struct cds_lfht
*ht
,
1034 cds_lfht_match_fct match
,
1037 struct cds_lfht_node
*node
,
1038 struct cds_lfht_iter
*unique_ret
,
1041 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
1043 struct cds_lfht_node
*bucket
;
1045 urcu_posix_assert(!is_bucket(node
));
1046 urcu_posix_assert(!is_removed(node
));
1047 urcu_posix_assert(!is_removal_owner(node
));
1048 bucket
= lookup_bucket(ht
, size
, hash
);
1050 uint32_t chain_len
= 0;
1053 * iter_prev points to the non-removed node prior to the
1057 /* We can always skip the bucket node initially */
1058 iter
= rcu_dereference(iter_prev
->next
);
1059 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1061 if (caa_unlikely(is_end(iter
)))
1063 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1066 /* bucket node is the first node of the identical-hash-value chain */
1067 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1070 next
= rcu_dereference(clear_flag(iter
)->next
);
1071 if (caa_unlikely(is_removed(next
)))
1077 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1078 struct cds_lfht_iter d_iter
= {
1081 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1087 * uniquely adding inserts the node as the first
1088 * node of the identical-hash-value node chain.
1090 * This semantic ensures no duplicated keys
1091 * should ever be observable in the table
1092 * (including traversing the table node by
1093 * node by forward iterations)
1095 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1099 *unique_ret
= d_iter
;
1103 /* Only account for identical reverse hash once */
1104 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1105 && !is_bucket(next
))
1106 check_resize(ht
, size
, ++chain_len
);
1107 iter_prev
= clear_flag(iter
);
1112 urcu_posix_assert(node
!= clear_flag(iter
));
1113 urcu_posix_assert(!is_removed(iter_prev
));
1114 urcu_posix_assert(!is_removal_owner(iter_prev
));
1115 urcu_posix_assert(!is_removed(iter
));
1116 urcu_posix_assert(!is_removal_owner(iter
));
1117 urcu_posix_assert(iter_prev
!= node
);
1119 node
->next
= clear_flag(iter
);
1121 node
->next
= flag_bucket(clear_flag(iter
));
1122 if (is_bucket(iter
))
1123 new_node
= flag_bucket(node
);
1126 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1127 new_node
) != iter
) {
1128 continue; /* retry */
1135 urcu_posix_assert(!is_removed(iter
));
1136 urcu_posix_assert(!is_removal_owner(iter
));
1137 if (is_bucket(iter
))
1138 new_next
= flag_bucket(clear_flag(next
));
1140 new_next
= clear_flag(next
);
1141 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1146 unique_ret
->node
= return_node
;
1147 /* unique_ret->next left unset, never used. */
1152 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1153 struct cds_lfht_node
*node
)
1155 struct cds_lfht_node
*bucket
, *next
;
1156 uintptr_t *node_next
;
1158 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1161 /* logically delete the node */
1162 urcu_posix_assert(!is_bucket(node
));
1163 urcu_posix_assert(!is_removed(node
));
1164 urcu_posix_assert(!is_removal_owner(node
));
1167 * We are first checking if the node had previously been
1168 * logically removed (this check is not atomic with setting the
1169 * logical removal flag). Return -ENOENT if the node had
1170 * previously been removed.
1172 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1173 if (caa_unlikely(is_removed(next
)))
1175 urcu_posix_assert(!is_bucket(next
));
1177 * The del operation semantic guarantees a full memory barrier
1178 * before the uatomic_or atomic commit of the deletion flag.
1180 * We set the REMOVED_FLAG unconditionally. Note that there may
1181 * be more than one concurrent thread setting this flag.
1182 * Knowing which wins the race will be known after the garbage
1183 * collection phase, stay tuned!
1185 * NOTE: The node_next variable is present to avoid breaking
1186 * strict-aliasing rules.
1188 node_next
= (uintptr_t*)&node
->next
;
1189 uatomic_or_mo(node_next
, REMOVED_FLAG
, CMM_RELEASE
);
1191 /* We performed the (logical) deletion. */
1194 * Ensure that the node is not visible to readers anymore: lookup for
1195 * the node, and remove it (along with any other logically removed node)
1198 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1199 _cds_lfht_gc_bucket(bucket
, node
);
1201 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1203 * Last phase: atomically exchange node->next with a version
1204 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1205 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1206 * the node and win the removal race.
1207 * It is interesting to note that all "add" paths are forbidden
1208 * to change the next pointer starting from the point where the
1209 * REMOVED_FLAG is set, so here using a read, followed by a
1210 * xchg() suffice to guarantee that the xchg() will ever only
1211 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1214 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1215 flag_removal_owner(uatomic_load(&node
->next
, CMM_RELAXED
)))))
1222 void *partition_resize_thread(void *arg
)
1224 struct partition_resize_work
*work
= arg
;
1226 work
->ht
->flavor
->register_thread();
1227 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1228 work
->ht
->flavor
->unregister_thread();
1233 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1235 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1236 unsigned long start
, unsigned long len
))
1238 unsigned long partition_len
, start
= 0;
1239 struct partition_resize_work
*work
;
1241 unsigned long thread
, nr_threads
;
1242 sigset_t newmask
, oldmask
;
1244 urcu_posix_assert(nr_cpus_mask
!= -1);
1245 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
)
1249 * Note: nr_cpus_mask + 1 is always power of 2.
1250 * We spawn just the number of threads we need to satisfy the minimum
1251 * partition size, up to the number of CPUs in the system.
1253 if (nr_cpus_mask
> 0) {
1254 nr_threads
= min_t(unsigned long, nr_cpus_mask
+ 1,
1255 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1259 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1260 work
= calloc(nr_threads
, sizeof(*work
));
1262 dbg_printf("error allocating for resize, single-threading\n");
1266 ret
= sigfillset(&newmask
);
1267 urcu_posix_assert(!ret
);
1268 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
1269 urcu_posix_assert(!ret
);
1271 for (thread
= 0; thread
< nr_threads
; thread
++) {
1272 work
[thread
].ht
= ht
;
1274 work
[thread
].len
= partition_len
;
1275 work
[thread
].start
= thread
* partition_len
;
1276 work
[thread
].fct
= fct
;
1277 ret
= pthread_create(&(work
[thread
].thread_id
),
1278 ht
->caller_resize_attr
? &ht
->resize_attr
: NULL
,
1279 partition_resize_thread
, &work
[thread
]);
1280 if (ret
== EAGAIN
) {
1282 * Out of resources: wait and join the threads
1283 * we've created, then handle leftovers.
1285 dbg_printf("error spawning for resize, single-threading\n");
1286 start
= work
[thread
].start
;
1288 nr_threads
= thread
;
1291 urcu_posix_assert(!ret
);
1294 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
1295 urcu_posix_assert(!ret
);
1297 for (thread
= 0; thread
< nr_threads
; thread
++) {
1298 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1299 urcu_posix_assert(!ret
);
1304 * A pthread_create failure above will either lead in us having
1305 * no threads to join or starting at a non-zero offset,
1306 * fallback to single thread processing of leftovers.
1308 if (start
== 0 && nr_threads
> 0)
1311 fct(ht
, i
, start
, len
);
1315 * Holding RCU read lock to protect _cds_lfht_add against memory
1316 * reclaim that could be performed by other worker threads (ABA
1319 * When we reach a certain length, we can split this population phase over
1320 * many worker threads, based on the number of CPUs available in the system.
1321 * This should therefore take care of not having the expand lagging behind too
1322 * many concurrent insertion threads by using the scheduler's ability to
1323 * schedule bucket node population fairly with insertions.
1326 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1327 unsigned long start
, unsigned long len
)
1329 unsigned long j
, size
= 1UL << (i
- 1);
1331 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1332 ht
->flavor
->read_lock();
1333 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1334 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1336 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1337 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1339 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1340 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1342 ht
->flavor
->read_unlock();
1346 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1349 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1353 void init_table(struct cds_lfht
*ht
,
1354 unsigned long first_order
, unsigned long last_order
)
1358 dbg_printf("init table: first_order %lu last_order %lu\n",
1359 first_order
, last_order
);
1360 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1361 for (i
= first_order
; i
<= last_order
; i
++) {
1364 len
= 1UL << (i
- 1);
1365 dbg_printf("init order %lu len: %lu\n", i
, len
);
1367 /* Stop expand if the resize target changes under us */
1368 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1371 cds_lfht_alloc_bucket_table(ht
, i
);
1374 * Set all bucket nodes reverse hash values for a level and
1375 * link all bucket nodes into the table.
1377 init_table_populate(ht
, i
, len
);
1380 * Update table size.
1382 * Populate data before RCU size.
1384 uatomic_store(&ht
->size
, 1UL << i
, CMM_RELEASE
);
1386 dbg_printf("init new size: %lu\n", 1UL << i
);
1387 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1393 * Holding RCU read lock to protect _cds_lfht_remove against memory
1394 * reclaim that could be performed by other worker threads (ABA
1396 * For a single level, we logically remove and garbage collect each node.
1398 * As a design choice, we perform logical removal and garbage collection on a
1399 * node-per-node basis to simplify this algorithm. We also assume keeping good
1400 * cache locality of the operation would overweight possible performance gain
1401 * that could be achieved by batching garbage collection for multiple levels.
1402 * However, this would have to be justified by benchmarks.
1404 * Concurrent removal and add operations are helping us perform garbage
1405 * collection of logically removed nodes. We guarantee that all logically
1406 * removed nodes have been garbage-collected (unlinked) before work
1407 * enqueue is invoked to free a hole level of bucket nodes (after a
1410 * Logical removal and garbage collection can therefore be done in batch
1411 * or on a node-per-node basis, as long as the guarantee above holds.
1413 * When we reach a certain length, we can split this removal over many worker
1414 * threads, based on the number of CPUs available in the system. This should
1415 * take care of not letting resize process lag behind too many concurrent
1416 * updater threads actively inserting into the hash table.
1419 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1420 unsigned long start
, unsigned long len
)
1422 unsigned long j
, size
= 1UL << (i
- 1);
1424 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1425 ht
->flavor
->read_lock();
1426 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1427 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1428 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1429 uintptr_t *fini_bucket_next
;
1431 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1432 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1434 /* Set the REMOVED_FLAG to freeze the ->next for gc.
1436 * NOTE: The fini_bucket_next variable is present to
1437 * avoid breaking strict-aliasing rules.
1439 fini_bucket_next
= (uintptr_t*)&fini_bucket
->next
;
1440 uatomic_or(fini_bucket_next
, REMOVED_FLAG
);
1441 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1443 ht
->flavor
->read_unlock();
1447 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1449 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1453 * fini_table() is never called for first_order == 0, which is why
1454 * free_by_rcu_order == 0 can be used as criterion to know if free must
1458 void fini_table(struct cds_lfht
*ht
,
1459 unsigned long first_order
, unsigned long last_order
)
1461 unsigned long free_by_rcu_order
= 0, i
;
1463 dbg_printf("fini table: first_order %lu last_order %lu\n",
1464 first_order
, last_order
);
1465 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1466 for (i
= last_order
; i
>= first_order
; i
--) {
1469 len
= 1UL << (i
- 1);
1470 dbg_printf("fini order %ld len: %lu\n", i
, len
);
1472 /* Stop shrink if the resize target changes under us */
1473 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1476 cmm_smp_wmb(); /* populate data before RCU size */
1477 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1480 * We need to wait for all add operations to reach Q.S. (and
1481 * thus use the new table for lookups) before we can start
1482 * releasing the old bucket nodes. Otherwise their lookup will
1483 * return a logically removed node as insert position.
1485 ht
->flavor
->update_synchronize_rcu();
1486 if (free_by_rcu_order
)
1487 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1490 * Set "removed" flag in bucket nodes about to be removed.
1491 * Unlink all now-logically-removed bucket node pointers.
1492 * Concurrent add/remove operation are helping us doing
1495 remove_table(ht
, i
, len
);
1497 free_by_rcu_order
= i
;
1499 dbg_printf("fini new size: %lu\n", 1UL << i
);
1500 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1504 if (free_by_rcu_order
) {
1505 ht
->flavor
->update_synchronize_rcu();
1506 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1511 * Never called with size < 1.
1514 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1516 struct cds_lfht_node
*prev
, *node
;
1517 unsigned long order
, len
, i
;
1520 cds_lfht_alloc_bucket_table(ht
, 0);
1522 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1523 node
= bucket_at(ht
, 0);
1524 node
->next
= flag_bucket(get_end());
1525 node
->reverse_hash
= 0;
1527 bucket_order
= cds_lfht_get_count_order_ulong(size
);
1528 urcu_posix_assert(bucket_order
>= 0);
1530 for (order
= 1; order
< (unsigned long) bucket_order
+ 1; order
++) {
1531 len
= 1UL << (order
- 1);
1532 cds_lfht_alloc_bucket_table(ht
, order
);
1534 for (i
= 0; i
< len
; i
++) {
1536 * Now, we are trying to init the node with the
1537 * hash=(len+i) (which is also a bucket with the
1538 * index=(len+i)) and insert it into the hash table,
1539 * so this node has to be inserted after the bucket
1540 * with the index=(len+i)&(len-1)=i. And because there
1541 * is no other non-bucket node nor bucket node with
1542 * larger index/hash inserted, so the bucket node
1543 * being inserted should be inserted directly linked
1544 * after the bucket node with index=i.
1546 prev
= bucket_at(ht
, i
);
1547 node
= bucket_at(ht
, len
+ i
);
1549 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1550 order
, len
+ i
, len
+ i
);
1551 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1553 /* insert after prev */
1554 urcu_posix_assert(is_bucket(prev
->next
));
1555 node
->next
= prev
->next
;
1556 prev
->next
= flag_bucket(node
);
1561 #if (CAA_BITS_PER_LONG > 32)
1563 * For 64-bit architectures, with max number of buckets small enough not to
1564 * use the entire 64-bit memory mapping space (and allowing a fair number of
1565 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1566 * fallback to the order allocator.
1569 const struct cds_lfht_mm_type
*get_mm_type(unsigned long max_nr_buckets
)
1571 if (max_nr_buckets
&& max_nr_buckets
<= (1ULL << 32))
1572 return &cds_lfht_mm_mmap
;
1574 return &cds_lfht_mm_order
;
1578 * For 32-bit architectures, use the order allocator.
1581 const struct cds_lfht_mm_type
*get_mm_type(
1582 unsigned long max_nr_buckets
__attribute__((unused
)))
1584 return &cds_lfht_mm_order
;
1588 void cds_lfht_node_init_deleted(struct cds_lfht_node
*node
)
1590 cds_lfht_node_init(node
);
1591 node
->next
= flag_removed(NULL
);
1594 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1595 unsigned long min_nr_alloc_buckets
,
1596 unsigned long max_nr_buckets
,
1598 const struct cds_lfht_mm_type
*mm
,
1599 const struct rcu_flavor_struct
*flavor
,
1600 pthread_attr_t
*attr
)
1602 struct cds_lfht
*ht
;
1603 unsigned long order
;
1605 /* min_nr_alloc_buckets must be power of two */
1606 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1609 /* init_size must be power of two */
1610 if (!init_size
|| (init_size
& (init_size
- 1)))
1614 * Memory management plugin default.
1617 mm
= get_mm_type(max_nr_buckets
);
1619 /* max_nr_buckets == 0 for order based mm means infinite */
1620 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1621 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1623 /* max_nr_buckets must be power of two */
1624 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1627 if (flags
& CDS_LFHT_AUTO_RESIZE
)
1628 cds_lfht_init_worker(flavor
);
1630 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1631 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1632 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1633 init_size
= min(init_size
, max_nr_buckets
);
1635 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1636 urcu_posix_assert(ht
);
1637 urcu_posix_assert(ht
->mm
== mm
);
1638 urcu_posix_assert(ht
->bucket_at
== mm
->bucket_at
);
1641 ht
->flavor
= flavor
;
1642 ht
->caller_resize_attr
= attr
;
1644 ht
->resize_attr
= *attr
;
1645 alloc_split_items_count(ht
);
1646 /* this mutex should not nest in read-side C.S. */
1647 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1648 order
= cds_lfht_get_count_order_ulong(init_size
);
1649 ht
->resize_target
= 1UL << order
;
1650 cds_lfht_create_bucket(ht
, 1UL << order
);
1651 ht
->size
= 1UL << order
;
1655 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1656 cds_lfht_match_fct match
, const void *key
,
1657 struct cds_lfht_iter
*iter
)
1659 struct cds_lfht_node
*node
, *next
, *bucket
;
1660 unsigned long reverse_hash
, size
;
1662 cds_lfht_iter_debug_set_ht(ht
, iter
);
1664 reverse_hash
= bit_reverse_ulong(hash
);
1667 * Use load acquire instead of rcu_dereference because there is no
1668 * dependency between the table size and the dereference of the bucket
1671 * This acquire is paired with the store release in init_table().
1673 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1674 bucket
= lookup_bucket(ht
, size
, hash
);
1675 /* We can always skip the bucket node initially */
1676 node
= rcu_dereference(bucket
->next
);
1677 node
= clear_flag(node
);
1679 if (caa_unlikely(is_end(node
))) {
1683 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1687 next
= rcu_dereference(node
->next
);
1688 urcu_posix_assert(node
== clear_flag(node
));
1689 if (caa_likely(!is_removed(next
))
1691 && node
->reverse_hash
== reverse_hash
1692 && caa_likely(match(node
, key
))) {
1695 node
= clear_flag(next
);
1697 urcu_posix_assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1702 void cds_lfht_next_duplicate(struct cds_lfht
*ht
__attribute__((unused
)),
1703 cds_lfht_match_fct match
,
1704 const void *key
, struct cds_lfht_iter
*iter
)
1706 struct cds_lfht_node
*node
, *next
;
1707 unsigned long reverse_hash
;
1709 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1711 reverse_hash
= node
->reverse_hash
;
1713 node
= clear_flag(next
);
1716 if (caa_unlikely(is_end(node
))) {
1720 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1724 next
= rcu_dereference(node
->next
);
1725 if (caa_likely(!is_removed(next
))
1727 && caa_likely(match(node
, key
))) {
1730 node
= clear_flag(next
);
1732 urcu_posix_assert(!node
|| !is_bucket(uatomic_load(&node
->next
, CMM_RELAXED
)));
1737 void cds_lfht_next(struct cds_lfht
*ht
__attribute__((unused
)),
1738 struct cds_lfht_iter
*iter
)
1740 struct cds_lfht_node
*node
, *next
;
1742 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1743 node
= clear_flag(iter
->next
);
1745 if (caa_unlikely(is_end(node
))) {
1749 next
= rcu_dereference(node
->next
);
1750 if (caa_likely(!is_removed(next
))
1751 && !is_bucket(next
)) {
1754 node
= clear_flag(next
);
1756 urcu_posix_assert(!node
|| !is_bucket(uatomic_load(&node
->next
, CMM_RELAXED
)));
1761 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1763 cds_lfht_iter_debug_set_ht(ht
, iter
);
1765 * Get next after first bucket node. The first bucket node is the
1766 * first node of the linked list.
1768 iter
->next
= uatomic_load(&bucket_at(ht
, 0)->next
, CMM_CONSUME
);
1769 cds_lfht_next(ht
, iter
);
1772 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1773 struct cds_lfht_node
*node
)
1777 node
->reverse_hash
= bit_reverse_ulong(hash
);
1778 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1779 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1780 ht_count_add(ht
, size
, hash
);
1783 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1785 cds_lfht_match_fct match
,
1787 struct cds_lfht_node
*node
)
1790 struct cds_lfht_iter iter
;
1792 node
->reverse_hash
= bit_reverse_ulong(hash
);
1793 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1794 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1795 if (iter
.node
== node
)
1796 ht_count_add(ht
, size
, hash
);
1800 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1802 cds_lfht_match_fct match
,
1804 struct cds_lfht_node
*node
)
1807 struct cds_lfht_iter iter
;
1809 node
->reverse_hash
= bit_reverse_ulong(hash
);
1810 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1812 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1813 if (iter
.node
== node
) {
1814 ht_count_add(ht
, size
, hash
);
1818 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1823 int cds_lfht_replace(struct cds_lfht
*ht
,
1824 struct cds_lfht_iter
*old_iter
,
1826 cds_lfht_match_fct match
,
1828 struct cds_lfht_node
*new_node
)
1832 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1833 if (!old_iter
->node
)
1835 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1837 if (caa_unlikely(!match(old_iter
->node
, key
)))
1839 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1840 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1844 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1849 size
= uatomic_load(&ht
->size
, CMM_ACQUIRE
);
1850 ret
= _cds_lfht_del(ht
, size
, node
);
1854 hash
= bit_reverse_ulong(node
->reverse_hash
);
1855 ht_count_del(ht
, size
, hash
);
1860 int cds_lfht_is_node_deleted(const struct cds_lfht_node
*node
)
1862 return is_removed(CMM_LOAD_SHARED(node
->next
));
1866 bool cds_lfht_is_empty(struct cds_lfht
*ht
)
1868 struct cds_lfht_node
*node
, *next
;
1872 was_online
= ht
->flavor
->read_ongoing();
1874 ht
->flavor
->thread_online();
1875 ht
->flavor
->read_lock();
1877 /* Check that the table is empty */
1878 node
= bucket_at(ht
, 0);
1880 next
= rcu_dereference(node
->next
);
1881 if (!is_bucket(next
)) {
1885 node
= clear_flag(next
);
1886 } while (!is_end(node
));
1888 ht
->flavor
->read_unlock();
1889 ht
->flavor
->thread_offline();
1895 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1897 struct cds_lfht_node
*node
;
1898 unsigned long order
, i
, size
;
1900 /* Check that the table is empty */
1901 node
= bucket_at(ht
, 0);
1903 node
= clear_flag(node
)->next
;
1904 if (!is_bucket(node
))
1906 urcu_posix_assert(!is_removed(node
));
1907 urcu_posix_assert(!is_removal_owner(node
));
1908 } while (!is_end(node
));
1910 * size accessed without rcu_dereference because hash table is
1914 /* Internal sanity check: all nodes left should be buckets */
1915 for (i
= 0; i
< size
; i
++) {
1916 node
= bucket_at(ht
, i
);
1917 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1918 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1919 urcu_posix_assert(is_bucket(node
->next
));
1922 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1923 cds_lfht_free_bucket_table(ht
, order
);
1929 void do_auto_resize_destroy_cb(struct urcu_work
*work
)
1931 struct cds_lfht
*ht
= caa_container_of(work
, struct cds_lfht
, destroy_work
);
1934 ht
->flavor
->register_thread();
1935 ret
= cds_lfht_delete_bucket(ht
);
1938 free_split_items_count(ht
);
1939 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1942 ht
->flavor
->unregister_thread();
1947 * Should only be called when no more concurrent readers nor writers can
1948 * possibly access the table.
1950 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1954 if (ht
->flags
& CDS_LFHT_AUTO_RESIZE
) {
1956 * Perform error-checking for emptiness before queuing
1957 * work, so we can return error to the caller. This runs
1958 * concurrently with ongoing resize.
1960 if (!cds_lfht_is_empty(ht
))
1962 /* Cancel ongoing resize operations. */
1963 uatomic_store(&ht
->in_progress_destroy
, 1, CMM_RELAXED
);
1965 *attr
= ht
->caller_resize_attr
;
1966 ht
->caller_resize_attr
= NULL
;
1969 * Queue destroy work after prior queued resize
1970 * operations. Given there are no concurrent writers
1971 * accessing the hash table at this point, no resize
1972 * operations can be queued after this destroy work.
1974 urcu_workqueue_queue_work(cds_lfht_workqueue
,
1975 &ht
->destroy_work
, do_auto_resize_destroy_cb
);
1978 ret
= cds_lfht_delete_bucket(ht
);
1981 free_split_items_count(ht
);
1983 *attr
= ht
->caller_resize_attr
;
1984 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1991 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1992 long *approx_before
,
1993 unsigned long *count
,
1996 struct cds_lfht_node
*node
, *next
;
1997 unsigned long nr_bucket
= 0, nr_removed
= 0;
2000 if (ht
->split_count
) {
2003 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
2004 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
2005 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
2011 /* Count non-bucket nodes in the table */
2012 node
= bucket_at(ht
, 0);
2014 next
= rcu_dereference(node
->next
);
2015 if (is_removed(next
)) {
2016 if (!is_bucket(next
))
2020 } else if (!is_bucket(next
))
2024 node
= clear_flag(next
);
2025 } while (!is_end(node
));
2026 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
2027 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
2029 if (ht
->split_count
) {
2032 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
2033 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
2034 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
2039 /* called with resize mutex held */
2041 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
2042 unsigned long old_size
, unsigned long new_size
)
2044 unsigned long old_order
, new_order
;
2046 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2047 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2048 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2049 old_size
, old_order
, new_size
, new_order
);
2050 urcu_posix_assert(new_size
> old_size
);
2051 init_table(ht
, old_order
+ 1, new_order
);
2054 /* called with resize mutex held */
2056 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
2057 unsigned long old_size
, unsigned long new_size
)
2059 unsigned long old_order
, new_order
;
2061 new_size
= max(new_size
, MIN_TABLE_SIZE
);
2062 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2063 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2064 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2065 old_size
, old_order
, new_size
, new_order
);
2066 urcu_posix_assert(new_size
< old_size
);
2068 /* Remove and unlink all bucket nodes to remove. */
2069 fini_table(ht
, new_order
+ 1, old_order
);
2073 /* called with resize mutex held */
2075 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
2077 unsigned long new_size
, old_size
;
2080 * Resize table, re-do if the target size has changed under us.
2083 if (uatomic_load(&ht
->in_progress_destroy
, CMM_RELAXED
))
2086 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2088 old_size
= ht
->size
;
2089 new_size
= uatomic_load(&ht
->resize_target
, CMM_RELAXED
);
2090 if (old_size
< new_size
)
2091 _do_cds_lfht_grow(ht
, old_size
, new_size
);
2092 else if (old_size
> new_size
)
2093 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
2095 uatomic_store(&ht
->resize_initiated
, 0, CMM_RELAXED
);
2096 /* write resize_initiated before read resize_target */
2098 } while (ht
->size
!= uatomic_load(&ht
->resize_target
, CMM_RELAXED
));
2102 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
2104 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
2108 void resize_target_update_count(struct cds_lfht
*ht
,
2109 unsigned long count
)
2111 count
= max(count
, MIN_TABLE_SIZE
);
2112 count
= min(count
, ht
->max_nr_buckets
);
2113 uatomic_set(&ht
->resize_target
, count
);
2116 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
2118 resize_target_update_count(ht
, new_size
);
2121 * Set flags has early as possible even in contention case.
2123 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2125 mutex_lock(&ht
->resize_mutex
);
2126 _do_cds_lfht_resize(ht
);
2127 mutex_unlock(&ht
->resize_mutex
);
2131 void do_resize_cb(struct urcu_work
*work
)
2133 struct resize_work
*resize_work
=
2134 caa_container_of(work
, struct resize_work
, work
);
2135 struct cds_lfht
*ht
= resize_work
->ht
;
2137 ht
->flavor
->register_thread();
2138 mutex_lock(&ht
->resize_mutex
);
2139 _do_cds_lfht_resize(ht
);
2140 mutex_unlock(&ht
->resize_mutex
);
2141 ht
->flavor
->unregister_thread();
2146 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
2148 struct resize_work
*work
;
2151 * Store to resize_target is before read resize_initiated as guaranteed
2152 * by either cmpxchg or _uatomic_xchg_monotonic_increase.
2154 if (!uatomic_load(&ht
->resize_initiated
, CMM_RELAXED
)) {
2155 if (uatomic_load(&ht
->in_progress_destroy
, CMM_RELAXED
)) {
2158 work
= malloc(sizeof(*work
));
2160 dbg_printf("error allocating resize work, bailing out\n");
2164 urcu_workqueue_queue_work(cds_lfht_workqueue
,
2165 &work
->work
, do_resize_cb
);
2166 uatomic_store(&ht
->resize_initiated
, 1, CMM_RELAXED
);
2171 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2173 unsigned long target_size
= size
<< growth
;
2175 target_size
= min(target_size
, ht
->max_nr_buckets
);
2176 if (resize_target_grow(ht
, target_size
) >= target_size
)
2179 __cds_lfht_resize_lazy_launch(ht
);
2183 * We favor grow operations over shrink. A shrink operation never occurs
2184 * if a grow operation is queued for lazy execution. A grow operation
2185 * cancels any pending shrink lazy execution.
2188 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2189 unsigned long count
)
2191 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2193 count
= max(count
, MIN_TABLE_SIZE
);
2194 count
= min(count
, ht
->max_nr_buckets
);
2196 return; /* Already the right size, no resize needed */
2197 if (count
> size
) { /* lazy grow */
2198 if (resize_target_grow(ht
, count
) >= count
)
2200 } else { /* lazy shrink */
2204 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2206 break; /* no resize needed */
2208 return; /* growing is/(was just) in progress */
2210 return; /* some other thread do shrink */
2214 __cds_lfht_resize_lazy_launch(ht
);
2217 static void cds_lfht_before_fork(void *priv
__attribute__((unused
)))
2219 if (cds_lfht_workqueue_atfork_nesting
++)
2221 mutex_lock(&cds_lfht_fork_mutex
);
2222 if (!cds_lfht_workqueue
)
2224 urcu_workqueue_pause_worker(cds_lfht_workqueue
);
2227 static void cds_lfht_after_fork_parent(void *priv
__attribute__((unused
)))
2229 if (--cds_lfht_workqueue_atfork_nesting
)
2231 if (!cds_lfht_workqueue
)
2233 urcu_workqueue_resume_worker(cds_lfht_workqueue
);
2235 mutex_unlock(&cds_lfht_fork_mutex
);
2238 static void cds_lfht_after_fork_child(void *priv
__attribute__((unused
)))
2240 if (--cds_lfht_workqueue_atfork_nesting
)
2242 if (!cds_lfht_workqueue
)
2244 urcu_workqueue_create_worker(cds_lfht_workqueue
);
2246 mutex_unlock(&cds_lfht_fork_mutex
);
2249 static struct urcu_atfork cds_lfht_atfork
= {
2250 .before_fork
= cds_lfht_before_fork
,
2251 .after_fork_parent
= cds_lfht_after_fork_parent
,
2252 .after_fork_child
= cds_lfht_after_fork_child
,
2255 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
)
2257 flavor
->register_rculfhash_atfork(&cds_lfht_atfork
);
2259 mutex_lock(&cds_lfht_fork_mutex
);
2260 if (!cds_lfht_workqueue
)
2261 cds_lfht_workqueue
= urcu_workqueue_create(0, -1, NULL
,
2262 NULL
, NULL
, NULL
, NULL
, NULL
, NULL
, NULL
);
2263 mutex_unlock(&cds_lfht_fork_mutex
);
2266 static void cds_lfht_exit(void)
2268 mutex_lock(&cds_lfht_fork_mutex
);
2269 if (cds_lfht_workqueue
) {
2270 urcu_workqueue_flush_queued_work(cds_lfht_workqueue
);
2271 urcu_workqueue_destroy(cds_lfht_workqueue
);
2272 cds_lfht_workqueue
= NULL
;
2274 mutex_unlock(&cds_lfht_fork_mutex
);