4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
33 * Some specificities of this Lock-Free Resizable RCU Hash Table
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups, as well as traversals, and use the returned objects
38 * safely by allowing memory reclaim to take place only after a grace
40 * - Add and remove operations are lock-free, and do not need to
41 * allocate memory. They need to be executed within RCU read-side
42 * critical section to ensure the objects they read are valid and to
43 * deal with the cmpxchg ABA problem.
44 * - add and add_unique operations are supported. add_unique checks if
45 * the node key already exists in the hash table. It ensures not to
46 * populate a duplicate key if the node key already exists in the hash
48 * - The resize operation executes concurrently with
49 * add/add_unique/add_replace/remove/lookup/traversal.
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
52 * - An index of bucket nodes is kept. These bucket nodes are the hash
53 * table "buckets". These buckets are internal nodes that allow to
54 * perform a fast hash lookup, similarly to a skip list. These
55 * buckets are chained together in the split-ordered list, which
56 * allows recursive expansion by inserting new buckets between the
57 * existing buckets. The split-ordered list allows adding new buckets
58 * between existing buckets as the table needs to grow.
59 * - The resize operation for small tables only allows expanding the
60 * hash table. It is triggered automatically by detecting long chains
61 * in the add operation.
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
64 * - Split-counters are used to keep track of the number of
65 * nodes within the hash table for automatic resize triggering.
66 * - Resize operation initiated by long chain detection is executed by a
67 * call_rcu thread, which keeps lock-freedom of add and remove.
68 * - Resize operations are protected by a mutex.
69 * - The removal operation is split in two parts: first, a "removed"
70 * flag is set in the next pointer within the node to remove. Then,
71 * a "garbage collection" is performed in the bucket containing the
72 * removed node (from the start of the bucket up to the removed node).
73 * All encountered nodes with "removed" flag set in their next
74 * pointers are removed from the linked-list. If the cmpxchg used for
75 * removal fails (due to concurrent garbage-collection or concurrent
76 * add), we retry from the beginning of the bucket. This ensures that
77 * the node with "removed" flag set is removed from the hash table
78 * (not visible to lookups anymore) before the RCU read-side critical
79 * section held across removal ends. Furthermore, this ensures that
80 * the node with "removed" flag set is removed from the linked-list
81 * before its memory is reclaimed. After setting the "removal" flag,
82 * only the thread which removal is the first to set the "removal
83 * owner" flag (with an xchg) into a node's next pointer is considered
84 * to have succeeded its removal (and thus owns the node to reclaim).
85 * Because we garbage-collect starting from an invariant node (the
86 * start-of-bucket bucket node) up to the "removed" node (or find a
87 * reverse-hash that is higher), we are sure that a successful
88 * traversal of the chain leads to a chain that is present in the
89 * linked-list (the start node is never removed) and that it does not
90 * contain the "removed" node anymore, even if concurrent delete/add
91 * operations are changing the structure of the list concurrently.
92 * - The add operations perform garbage collection of buckets if they
93 * encounter nodes with removed flag set in the bucket where they want
94 * to add their new node. This ensures lock-freedom of add operation by
95 * helping the remover unlink nodes from the list rather than to wait
97 * - There are three memory backends for the hash table buckets: the
98 * "order table", the "chunks", and the "mmap".
99 * - These bucket containers contain a compact version of the hash table
101 * - The RCU "order table":
102 * - has a first level table indexed by log2(hash index) which is
103 * copied and expanded by the resize operation. This order table
104 * allows finding the "bucket node" tables.
105 * - There is one bucket node table per hash index order. The size of
106 * each bucket node table is half the number of hashes contained in
107 * this order (except for order 0).
108 * - The RCU "chunks" is best suited for close interaction with a page
109 * allocator. It uses a linear array as index to "chunks" containing
110 * each the same number of buckets.
111 * - The RCU "mmap" memory backend uses a single memory map to hold
113 * - synchronize_rcu is used to garbage-collect the old bucket node table.
115 * Ordering Guarantees:
117 * To discuss these guarantees, we first define "read" operation as any
118 * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
119 * cds_lfht_first, cds_lfht_next operation, as well as
120 * cds_lfht_add_unique (failure).
122 * We define "read traversal" operation as any of the following
123 * group of operations
124 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
125 * (and/or cds_lfht_next, although less common).
126 * - cds_lfht_add_unique (failure) followed by iteration with
127 * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
129 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
130 * cds_lfht_next_duplicate, although less common).
132 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
133 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
135 * When cds_lfht_add_unique succeeds (returns the node passed as
136 * parameter), it acts as a "write" operation. When cds_lfht_add_unique
137 * fails (returns a node different from the one passed as parameter), it
138 * acts as a "read" operation. A cds_lfht_add_unique failure is a
139 * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
140 * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
143 * We define "prior" and "later" node as nodes observable by reads and
144 * read traversals respectively before and after a write or sequence of
147 * Hash-table operations are often cascaded, for example, the pointer
148 * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
149 * whose return value might in turn be passed to another hash-table
150 * operation. This entire cascaded series of operations must be enclosed
151 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
154 * The following ordering guarantees are offered by this hash table:
156 * A.1) "read" after "write": if there is ordering between a write and a
157 * later read, then the read is guaranteed to see the write or some
159 * A.2) "read traversal" after "write": given that there is dependency
160 * ordering between reads in a "read traversal", if there is
161 * ordering between a write and the first read of the traversal,
162 * then the "read traversal" is guaranteed to see the write or
164 * B.1) "write" after "read": if there is ordering between a read and a
165 * later write, then the read will never see the write.
166 * B.2) "write" after "read traversal": given that there is dependency
167 * ordering between reads in a "read traversal", if there is
168 * ordering between the last read of the traversal and a later
169 * write, then the "read traversal" will never see the write.
170 * C) "write" while "read traversal": if a write occurs during a "read
171 * traversal", the traversal may, or may not, see the write.
172 * D.1) "write" after "write": if there is ordering between a write and
173 * a later write, then the later write is guaranteed to see the
174 * effects of the first write.
175 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
176 * order to any pair of concurrent conflicting writes.
177 * Non-conflicting writes (for example, to different keys) are
179 * E) If a grace period separates a "del" or "replace" operation
180 * and a subsequent operation, then that subsequent operation is
181 * guaranteed not to see the removed item.
182 * F) Uniqueness guarantee: given a hash table that does not contain
183 * duplicate items for a given key, there will only be one item in
184 * the hash table after an arbitrary sequence of add_unique and/or
185 * add_replace operations. Note, however, that a pair of
186 * concurrent read operations might well access two different items
188 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
189 * memory barrier), then the second lookup will return the same
190 * node as the previous lookup, or some later node.
191 * G.2) A "read traversal" that starts after the end of a prior "read
192 * traversal" (ordered by memory barriers) is guaranteed to see the
193 * same nodes as the previous traversal, or some later nodes.
194 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
195 * example, if a pair of reads to the same key run concurrently
196 * with an insertion of that same key, the reads remain unordered
197 * regardless of their return values. In other words, you cannot
198 * rely on the values returned by the reads to deduce ordering.
200 * Progress guarantees:
202 * * Reads are wait-free. These operations always move forward in the
203 * hash table linked list, and this list has no loop.
204 * * Writes are lock-free. Any retry loop performed by a write operation
205 * is triggered by progress made within another update operation.
207 * Bucket node tables:
209 * hash table hash table the last all bucket node tables
210 * order size bucket node 0 1 2 3 4 5 6(index)
217 * 5 32 16 1 1 2 4 8 16
218 * 6 64 32 1 1 2 4 8 16 32
220 * When growing/shrinking, we only focus on the last bucket node table
221 * which size is (!order ? 1 : (1 << (order -1))).
223 * Example for growing/shrinking:
224 * grow hash table from order 5 to 6: init the index=6 bucket node table
225 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
227 * A bit of ascii art explanation:
229 * The order index is the off-by-one compared to the actual power of 2
230 * because we use index 0 to deal with the 0 special-case.
232 * This shows the nodes for a small table ordered by reversed bits:
244 * This shows the nodes in order of non-reversed bits, linked by
245 * reversed-bit order.
250 * 2 | | 2 010 010 <- |
251 * | | | 3 011 110 | <- |
252 * 3 -> | | | 4 100 001 | |
269 #include "compat-getcpu.h"
271 #include <urcu-call-rcu.h>
272 #include <urcu-flavor.h>
273 #include <urcu/arch.h>
274 #include <urcu/uatomic.h>
275 #include <urcu/compiler.h>
276 #include <urcu/rculfhash.h>
277 #include <rculfhash-internal.h>
282 * Split-counters lazily update the global counter each 1024
283 * addition/removal. It automatically keeps track of resize required.
284 * We use the bucket length as indicator for need to expand for small
285 * tables and machines lacking per-cpu data support.
287 #define COUNT_COMMIT_ORDER 10
288 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
289 #define CHAIN_LEN_TARGET 1
290 #define CHAIN_LEN_RESIZE_THRESHOLD 3
293 * Define the minimum table size.
295 #define MIN_TABLE_ORDER 0
296 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
299 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
301 #define MIN_PARTITION_PER_THREAD_ORDER 12
302 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
305 * The removed flag needs to be updated atomically with the pointer.
306 * It indicates that no node must attach to the node scheduled for
307 * removal, and that node garbage collection must be performed.
308 * The bucket flag does not require to be updated atomically with the
309 * pointer, but it is added as a pointer low bit flag to save space.
310 * The "removal owner" flag is used to detect which of the "del"
311 * operation that has set the "removed flag" gets to return the removed
312 * node to its caller. Note that the replace operation does not need to
313 * iteract with the "removal owner" flag, because it validates that
314 * the "removed" flag is not set before performing its cmpxchg.
316 #define REMOVED_FLAG (1UL << 0)
317 #define BUCKET_FLAG (1UL << 1)
318 #define REMOVAL_OWNER_FLAG (1UL << 2)
319 #define FLAGS_MASK ((1UL << 3) - 1)
321 /* Value of the end pointer. Should not interact with flags. */
322 #define END_VALUE NULL
325 * ht_items_count: Split-counters counting the number of node addition
326 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
327 * is set at hash table creation.
329 * These are free-running counters, never reset to zero. They count the
330 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
331 * operations to update the global counter. We choose a power-of-2 value
332 * for the trigger to deal with 32 or 64-bit overflow of the counter.
334 struct ht_items_count
{
335 unsigned long add
, del
;
336 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
339 * rcu_resize_work: Contains arguments passed to RCU worker thread
340 * responsible for performing lazy resize.
342 struct rcu_resize_work
{
343 struct rcu_head head
;
348 * partition_resize_work: Contains arguments passed to worker threads
349 * executing the hash table resize on partitions of the hash table
350 * assigned to each processor's worker thread.
352 struct partition_resize_work
{
355 unsigned long i
, start
, len
;
356 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
357 unsigned long start
, unsigned long len
);
361 * Algorithm to reverse bits in a word by lookup table, extended to
364 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
365 * Originally from Public Domain.
368 static const uint8_t BitReverseTable256
[256] =
370 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
371 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
372 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
373 R6(0), R6(2), R6(1), R6(3)
380 uint8_t bit_reverse_u8(uint8_t v
)
382 return BitReverseTable256
[v
];
385 #if (CAA_BITS_PER_LONG == 32)
387 uint32_t bit_reverse_u32(uint32_t v
)
389 return ((uint32_t) bit_reverse_u8(v
) << 24) |
390 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
391 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
392 ((uint32_t) bit_reverse_u8(v
>> 24));
396 uint64_t bit_reverse_u64(uint64_t v
)
398 return ((uint64_t) bit_reverse_u8(v
) << 56) |
399 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
400 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
401 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
402 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
403 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
404 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
405 ((uint64_t) bit_reverse_u8(v
>> 56));
410 unsigned long bit_reverse_ulong(unsigned long v
)
412 #if (CAA_BITS_PER_LONG == 32)
413 return bit_reverse_u32(v
);
415 return bit_reverse_u64(v
);
420 * fls: returns the position of the most significant bit.
421 * Returns 0 if no bit is set, else returns the position of the most
422 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
424 #if defined(__i386) || defined(__x86_64)
426 unsigned int fls_u32(uint32_t x
)
430 __asm__ ("bsrl %1,%0\n\t"
434 : "=r" (r
) : "rm" (x
));
440 #if defined(__x86_64)
442 unsigned int fls_u64(uint64_t x
)
446 __asm__ ("bsrq %1,%0\n\t"
450 : "=r" (r
) : "rm" (x
));
457 static __attribute__((unused
))
458 unsigned int fls_u64(uint64_t x
)
465 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
469 if (!(x
& 0xFFFF000000000000ULL
)) {
473 if (!(x
& 0xFF00000000000000ULL
)) {
477 if (!(x
& 0xF000000000000000ULL
)) {
481 if (!(x
& 0xC000000000000000ULL
)) {
485 if (!(x
& 0x8000000000000000ULL
)) {
494 static __attribute__((unused
))
495 unsigned int fls_u32(uint32_t x
)
501 if (!(x
& 0xFFFF0000U
)) {
505 if (!(x
& 0xFF000000U
)) {
509 if (!(x
& 0xF0000000U
)) {
513 if (!(x
& 0xC0000000U
)) {
517 if (!(x
& 0x80000000U
)) {
525 unsigned int cds_lfht_fls_ulong(unsigned long x
)
527 #if (CAA_BITS_PER_LONG == 32)
535 * Return the minimum order for which x <= (1UL << order).
536 * Return -1 if x is 0.
538 int cds_lfht_get_count_order_u32(uint32_t x
)
543 return fls_u32(x
- 1);
547 * Return the minimum order for which x <= (1UL << order).
548 * Return -1 if x is 0.
550 int cds_lfht_get_count_order_ulong(unsigned long x
)
555 return cds_lfht_fls_ulong(x
- 1);
559 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
562 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
563 unsigned long count
);
565 static long nr_cpus_mask
= -1;
566 static long split_count_mask
= -1;
567 static int split_count_order
= -1;
569 #if defined(HAVE_SYSCONF)
570 static void ht_init_nr_cpus_mask(void)
574 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
580 * round up number of CPUs to next power of two, so we
581 * can use & for modulo.
583 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
584 nr_cpus_mask
= maxcpus
- 1;
586 #else /* #if defined(HAVE_SYSCONF) */
587 static void ht_init_nr_cpus_mask(void)
591 #endif /* #else #if defined(HAVE_SYSCONF) */
594 void alloc_split_items_count(struct cds_lfht
*ht
)
596 if (nr_cpus_mask
== -1) {
597 ht_init_nr_cpus_mask();
598 if (nr_cpus_mask
< 0)
599 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
601 split_count_mask
= nr_cpus_mask
;
603 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
606 assert(split_count_mask
>= 0);
608 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
609 ht
->split_count
= calloc(split_count_mask
+ 1,
610 sizeof(struct ht_items_count
));
611 assert(ht
->split_count
);
613 ht
->split_count
= NULL
;
618 void free_split_items_count(struct cds_lfht
*ht
)
620 poison_free(ht
->split_count
);
624 int ht_get_split_count_index(unsigned long hash
)
628 assert(split_count_mask
>= 0);
629 cpu
= urcu_sched_getcpu();
630 if (caa_unlikely(cpu
< 0))
631 return hash
& split_count_mask
;
633 return cpu
& split_count_mask
;
637 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
639 unsigned long split_count
;
643 if (caa_unlikely(!ht
->split_count
))
645 index
= ht_get_split_count_index(hash
);
646 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
647 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
649 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
651 dbg_printf("add split count %lu\n", split_count
);
652 count
= uatomic_add_return(&ht
->count
,
653 1UL << COUNT_COMMIT_ORDER
);
654 if (caa_likely(count
& (count
- 1)))
656 /* Only if global count is power of 2 */
658 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
660 dbg_printf("add set global %ld\n", count
);
661 cds_lfht_resize_lazy_count(ht
, size
,
662 count
>> (CHAIN_LEN_TARGET
- 1));
666 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
668 unsigned long split_count
;
672 if (caa_unlikely(!ht
->split_count
))
674 index
= ht_get_split_count_index(hash
);
675 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
676 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
678 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
680 dbg_printf("del split count %lu\n", split_count
);
681 count
= uatomic_add_return(&ht
->count
,
682 -(1UL << COUNT_COMMIT_ORDER
));
683 if (caa_likely(count
& (count
- 1)))
685 /* Only if global count is power of 2 */
687 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
689 dbg_printf("del set global %ld\n", count
);
691 * Don't shrink table if the number of nodes is below a
694 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
696 cds_lfht_resize_lazy_count(ht
, size
,
697 count
>> (CHAIN_LEN_TARGET
- 1));
701 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
705 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
707 count
= uatomic_read(&ht
->count
);
709 * Use bucket-local length for small table expand and for
710 * environments lacking per-cpu data support.
712 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
715 dbg_printf("WARNING: large chain length: %u.\n",
717 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
721 * Ideal growth calculated based on chain length.
723 growth
= cds_lfht_get_count_order_u32(chain_len
724 - (CHAIN_LEN_TARGET
- 1));
725 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
727 >= (1UL << (COUNT_COMMIT_ORDER
728 + split_count_order
))) {
730 * If ideal growth expands the hash table size
731 * beyond the "small hash table" sizes, use the
732 * maximum small hash table size to attempt
733 * expanding the hash table. This only applies
734 * when node accounting is available, otherwise
735 * the chain length is used to expand the hash
736 * table in every case.
738 growth
= COUNT_COMMIT_ORDER
+ split_count_order
739 - cds_lfht_get_count_order_ulong(size
);
743 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
748 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
750 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
754 int is_removed(struct cds_lfht_node
*node
)
756 return ((unsigned long) node
) & REMOVED_FLAG
;
760 int is_bucket(struct cds_lfht_node
*node
)
762 return ((unsigned long) node
) & BUCKET_FLAG
;
766 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
768 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
772 int is_removal_owner(struct cds_lfht_node
*node
)
774 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
778 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
780 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
784 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
786 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
790 struct cds_lfht_node
*get_end(void)
792 return (struct cds_lfht_node
*) END_VALUE
;
796 int is_end(struct cds_lfht_node
*node
)
798 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
802 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
805 unsigned long old1
, old2
;
807 old1
= uatomic_read(ptr
);
812 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
817 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
819 return ht
->mm
->alloc_bucket_table(ht
, order
);
823 * cds_lfht_free_bucket_table() should be called with decreasing order.
824 * When cds_lfht_free_bucket_table(0) is called, it means the whole
828 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
830 return ht
->mm
->free_bucket_table(ht
, order
);
834 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
836 return ht
->bucket_at(ht
, index
);
840 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
844 return bucket_at(ht
, hash
& (size
- 1));
848 * Remove all logically deleted nodes from a bucket up to a certain node key.
851 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
853 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
855 assert(!is_bucket(bucket
));
856 assert(!is_removed(bucket
));
857 assert(!is_removal_owner(bucket
));
858 assert(!is_bucket(node
));
859 assert(!is_removed(node
));
860 assert(!is_removal_owner(node
));
863 /* We can always skip the bucket node initially */
864 iter
= rcu_dereference(iter_prev
->next
);
865 assert(!is_removed(iter
));
866 assert(!is_removal_owner(iter
));
867 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
869 * We should never be called with bucket (start of chain)
870 * and logically removed node (end of path compression
871 * marker) being the actual same node. This would be a
872 * bug in the algorithm implementation.
874 assert(bucket
!= node
);
876 if (caa_unlikely(is_end(iter
)))
878 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
880 next
= rcu_dereference(clear_flag(iter
)->next
);
881 if (caa_likely(is_removed(next
)))
883 iter_prev
= clear_flag(iter
);
886 assert(!is_removed(iter
));
887 assert(!is_removal_owner(iter
));
889 new_next
= flag_bucket(clear_flag(next
));
891 new_next
= clear_flag(next
);
892 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
897 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
898 struct cds_lfht_node
*old_node
,
899 struct cds_lfht_node
*old_next
,
900 struct cds_lfht_node
*new_node
)
902 struct cds_lfht_node
*bucket
, *ret_next
;
904 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
907 assert(!is_removed(old_node
));
908 assert(!is_removal_owner(old_node
));
909 assert(!is_bucket(old_node
));
910 assert(!is_removed(new_node
));
911 assert(!is_removal_owner(new_node
));
912 assert(!is_bucket(new_node
));
913 assert(new_node
!= old_node
);
915 /* Insert after node to be replaced */
916 if (is_removed(old_next
)) {
918 * Too late, the old node has been removed under us
919 * between lookup and replace. Fail.
923 assert(old_next
== clear_flag(old_next
));
924 assert(new_node
!= old_next
);
926 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
927 * flag. It is either set atomically at the same time
928 * (replace) or after (del).
930 assert(!is_removal_owner(old_next
));
931 new_node
->next
= old_next
;
933 * Here is the whole trick for lock-free replace: we add
934 * the replacement node _after_ the node we want to
935 * replace by atomically setting its next pointer at the
936 * same time we set its removal flag. Given that
937 * the lookups/get next use an iterator aware of the
938 * next pointer, they will either skip the old node due
939 * to the removal flag and see the new node, or use
940 * the old node, but will not see the new one.
941 * This is a replacement of a node with another node
942 * that has the same value: we are therefore not
943 * removing a value from the hash table. We set both the
944 * REMOVED and REMOVAL_OWNER flags atomically so we own
945 * the node after successful cmpxchg.
947 ret_next
= uatomic_cmpxchg(&old_node
->next
,
948 old_next
, flag_removed_or_removal_owner(new_node
));
949 if (ret_next
== old_next
)
950 break; /* We performed the replacement. */
955 * Ensure that the old node is not visible to readers anymore:
956 * lookup for the node, and remove it (along with any other
957 * logically removed node) if found.
959 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
960 _cds_lfht_gc_bucket(bucket
, new_node
);
962 assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
967 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
968 * mode. A NULL unique_ret allows creation of duplicate keys.
971 void _cds_lfht_add(struct cds_lfht
*ht
,
973 cds_lfht_match_fct match
,
976 struct cds_lfht_node
*node
,
977 struct cds_lfht_iter
*unique_ret
,
980 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
982 struct cds_lfht_node
*bucket
;
984 assert(!is_bucket(node
));
985 assert(!is_removed(node
));
986 assert(!is_removal_owner(node
));
987 bucket
= lookup_bucket(ht
, size
, hash
);
989 uint32_t chain_len
= 0;
992 * iter_prev points to the non-removed node prior to the
996 /* We can always skip the bucket node initially */
997 iter
= rcu_dereference(iter_prev
->next
);
998 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1000 if (caa_unlikely(is_end(iter
)))
1002 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1005 /* bucket node is the first node of the identical-hash-value chain */
1006 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1009 next
= rcu_dereference(clear_flag(iter
)->next
);
1010 if (caa_unlikely(is_removed(next
)))
1016 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1017 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
1020 * uniquely adding inserts the node as the first
1021 * node of the identical-hash-value node chain.
1023 * This semantic ensures no duplicated keys
1024 * should ever be observable in the table
1025 * (including traversing the table node by
1026 * node by forward iterations)
1028 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1032 *unique_ret
= d_iter
;
1036 /* Only account for identical reverse hash once */
1037 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1038 && !is_bucket(next
))
1039 check_resize(ht
, size
, ++chain_len
);
1040 iter_prev
= clear_flag(iter
);
1045 assert(node
!= clear_flag(iter
));
1046 assert(!is_removed(iter_prev
));
1047 assert(!is_removal_owner(iter_prev
));
1048 assert(!is_removed(iter
));
1049 assert(!is_removal_owner(iter
));
1050 assert(iter_prev
!= node
);
1052 node
->next
= clear_flag(iter
);
1054 node
->next
= flag_bucket(clear_flag(iter
));
1055 if (is_bucket(iter
))
1056 new_node
= flag_bucket(node
);
1059 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1060 new_node
) != iter
) {
1061 continue; /* retry */
1068 assert(!is_removed(iter
));
1069 assert(!is_removal_owner(iter
));
1070 if (is_bucket(iter
))
1071 new_next
= flag_bucket(clear_flag(next
));
1073 new_next
= clear_flag(next
);
1074 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1079 unique_ret
->node
= return_node
;
1080 /* unique_ret->next left unset, never used. */
1085 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1086 struct cds_lfht_node
*node
)
1088 struct cds_lfht_node
*bucket
, *next
;
1090 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1093 /* logically delete the node */
1094 assert(!is_bucket(node
));
1095 assert(!is_removed(node
));
1096 assert(!is_removal_owner(node
));
1099 * We are first checking if the node had previously been
1100 * logically removed (this check is not atomic with setting the
1101 * logical removal flag). Return -ENOENT if the node had
1102 * previously been removed.
1104 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1105 if (caa_unlikely(is_removed(next
)))
1107 assert(!is_bucket(next
));
1109 * The del operation semantic guarantees a full memory barrier
1110 * before the uatomic_or atomic commit of the deletion flag.
1112 cmm_smp_mb__before_uatomic_or();
1114 * We set the REMOVED_FLAG unconditionally. Note that there may
1115 * be more than one concurrent thread setting this flag.
1116 * Knowing which wins the race will be known after the garbage
1117 * collection phase, stay tuned!
1119 uatomic_or(&node
->next
, REMOVED_FLAG
);
1120 /* We performed the (logical) deletion. */
1123 * Ensure that the node is not visible to readers anymore: lookup for
1124 * the node, and remove it (along with any other logically removed node)
1127 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1128 _cds_lfht_gc_bucket(bucket
, node
);
1130 assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1132 * Last phase: atomically exchange node->next with a version
1133 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1134 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1135 * the node and win the removal race.
1136 * It is interesting to note that all "add" paths are forbidden
1137 * to change the next pointer starting from the point where the
1138 * REMOVED_FLAG is set, so here using a read, followed by a
1139 * xchg() suffice to guarantee that the xchg() will ever only
1140 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1143 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1144 flag_removal_owner(node
->next
))))
1151 void *partition_resize_thread(void *arg
)
1153 struct partition_resize_work
*work
= arg
;
1155 work
->ht
->flavor
->register_thread();
1156 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1157 work
->ht
->flavor
->unregister_thread();
1162 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1164 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1165 unsigned long start
, unsigned long len
))
1167 unsigned long partition_len
, start
= 0;
1168 struct partition_resize_work
*work
;
1170 unsigned long nr_threads
;
1172 assert(nr_cpus_mask
!= -1);
1173 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
)
1177 * Note: nr_cpus_mask + 1 is always power of 2.
1178 * We spawn just the number of threads we need to satisfy the minimum
1179 * partition size, up to the number of CPUs in the system.
1181 if (nr_cpus_mask
> 0) {
1182 nr_threads
= min(nr_cpus_mask
+ 1,
1183 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1187 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1188 work
= calloc(nr_threads
, sizeof(*work
));
1190 dbg_printf("error allocating for resize, single-threading\n");
1193 for (thread
= 0; thread
< nr_threads
; thread
++) {
1194 work
[thread
].ht
= ht
;
1196 work
[thread
].len
= partition_len
;
1197 work
[thread
].start
= thread
* partition_len
;
1198 work
[thread
].fct
= fct
;
1199 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1200 partition_resize_thread
, &work
[thread
]);
1201 if (ret
== EAGAIN
) {
1203 * Out of resources: wait and join the threads
1204 * we've created, then handle leftovers.
1206 dbg_printf("error spawning for resize, single-threading\n");
1207 start
= work
[thread
].start
;
1209 nr_threads
= thread
;
1214 for (thread
= 0; thread
< nr_threads
; thread
++) {
1215 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1221 * A pthread_create failure above will either lead in us having
1222 * no threads to join or starting at a non-zero offset,
1223 * fallback to single thread processing of leftovers.
1225 if (start
== 0 && nr_threads
> 0)
1228 ht
->flavor
->thread_online();
1229 fct(ht
, i
, start
, len
);
1230 ht
->flavor
->thread_offline();
1234 * Holding RCU read lock to protect _cds_lfht_add against memory
1235 * reclaim that could be performed by other call_rcu worker threads (ABA
1238 * When we reach a certain length, we can split this population phase over
1239 * many worker threads, based on the number of CPUs available in the system.
1240 * This should therefore take care of not having the expand lagging behind too
1241 * many concurrent insertion threads by using the scheduler's ability to
1242 * schedule bucket node population fairly with insertions.
1245 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1246 unsigned long start
, unsigned long len
)
1248 unsigned long j
, size
= 1UL << (i
- 1);
1250 assert(i
> MIN_TABLE_ORDER
);
1251 ht
->flavor
->read_lock();
1252 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1253 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1255 assert(j
>= size
&& j
< (size
<< 1));
1256 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1258 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1259 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1261 ht
->flavor
->read_unlock();
1265 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1268 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1272 void init_table(struct cds_lfht
*ht
,
1273 unsigned long first_order
, unsigned long last_order
)
1277 dbg_printf("init table: first_order %lu last_order %lu\n",
1278 first_order
, last_order
);
1279 assert(first_order
> MIN_TABLE_ORDER
);
1280 for (i
= first_order
; i
<= last_order
; i
++) {
1283 len
= 1UL << (i
- 1);
1284 dbg_printf("init order %lu len: %lu\n", i
, len
);
1286 /* Stop expand if the resize target changes under us */
1287 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1290 cds_lfht_alloc_bucket_table(ht
, i
);
1293 * Set all bucket nodes reverse hash values for a level and
1294 * link all bucket nodes into the table.
1296 init_table_populate(ht
, i
, len
);
1299 * Update table size.
1301 cmm_smp_wmb(); /* populate data before RCU size */
1302 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1304 dbg_printf("init new size: %lu\n", 1UL << i
);
1305 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1311 * Holding RCU read lock to protect _cds_lfht_remove against memory
1312 * reclaim that could be performed by other call_rcu worker threads (ABA
1314 * For a single level, we logically remove and garbage collect each node.
1316 * As a design choice, we perform logical removal and garbage collection on a
1317 * node-per-node basis to simplify this algorithm. We also assume keeping good
1318 * cache locality of the operation would overweight possible performance gain
1319 * that could be achieved by batching garbage collection for multiple levels.
1320 * However, this would have to be justified by benchmarks.
1322 * Concurrent removal and add operations are helping us perform garbage
1323 * collection of logically removed nodes. We guarantee that all logically
1324 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1325 * invoked to free a hole level of bucket nodes (after a grace period).
1327 * Logical removal and garbage collection can therefore be done in batch
1328 * or on a node-per-node basis, as long as the guarantee above holds.
1330 * When we reach a certain length, we can split this removal over many worker
1331 * threads, based on the number of CPUs available in the system. This should
1332 * take care of not letting resize process lag behind too many concurrent
1333 * updater threads actively inserting into the hash table.
1336 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1337 unsigned long start
, unsigned long len
)
1339 unsigned long j
, size
= 1UL << (i
- 1);
1341 assert(i
> MIN_TABLE_ORDER
);
1342 ht
->flavor
->read_lock();
1343 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1344 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1345 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1347 assert(j
>= size
&& j
< (size
<< 1));
1348 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1350 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1351 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1352 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1354 ht
->flavor
->read_unlock();
1358 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1360 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1364 * fini_table() is never called for first_order == 0, which is why
1365 * free_by_rcu_order == 0 can be used as criterion to know if free must
1369 void fini_table(struct cds_lfht
*ht
,
1370 unsigned long first_order
, unsigned long last_order
)
1373 unsigned long free_by_rcu_order
= 0;
1375 dbg_printf("fini table: first_order %lu last_order %lu\n",
1376 first_order
, last_order
);
1377 assert(first_order
> MIN_TABLE_ORDER
);
1378 for (i
= last_order
; i
>= first_order
; i
--) {
1381 len
= 1UL << (i
- 1);
1382 dbg_printf("fini order %ld len: %lu\n", i
, len
);
1384 /* Stop shrink if the resize target changes under us */
1385 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1388 cmm_smp_wmb(); /* populate data before RCU size */
1389 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1392 * We need to wait for all add operations to reach Q.S. (and
1393 * thus use the new table for lookups) before we can start
1394 * releasing the old bucket nodes. Otherwise their lookup will
1395 * return a logically removed node as insert position.
1397 ht
->flavor
->update_synchronize_rcu();
1398 if (free_by_rcu_order
)
1399 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1402 * Set "removed" flag in bucket nodes about to be removed.
1403 * Unlink all now-logically-removed bucket node pointers.
1404 * Concurrent add/remove operation are helping us doing
1407 remove_table(ht
, i
, len
);
1409 free_by_rcu_order
= i
;
1411 dbg_printf("fini new size: %lu\n", 1UL << i
);
1412 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1416 if (free_by_rcu_order
) {
1417 ht
->flavor
->update_synchronize_rcu();
1418 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1423 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1425 struct cds_lfht_node
*prev
, *node
;
1426 unsigned long order
, len
, i
;
1428 cds_lfht_alloc_bucket_table(ht
, 0);
1430 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1431 node
= bucket_at(ht
, 0);
1432 node
->next
= flag_bucket(get_end());
1433 node
->reverse_hash
= 0;
1435 for (order
= 1; order
< cds_lfht_get_count_order_ulong(size
) + 1; order
++) {
1436 len
= 1UL << (order
- 1);
1437 cds_lfht_alloc_bucket_table(ht
, order
);
1439 for (i
= 0; i
< len
; i
++) {
1441 * Now, we are trying to init the node with the
1442 * hash=(len+i) (which is also a bucket with the
1443 * index=(len+i)) and insert it into the hash table,
1444 * so this node has to be inserted after the bucket
1445 * with the index=(len+i)&(len-1)=i. And because there
1446 * is no other non-bucket node nor bucket node with
1447 * larger index/hash inserted, so the bucket node
1448 * being inserted should be inserted directly linked
1449 * after the bucket node with index=i.
1451 prev
= bucket_at(ht
, i
);
1452 node
= bucket_at(ht
, len
+ i
);
1454 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1455 order
, len
+ i
, len
+ i
);
1456 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1458 /* insert after prev */
1459 assert(is_bucket(prev
->next
));
1460 node
->next
= prev
->next
;
1461 prev
->next
= flag_bucket(node
);
1466 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1467 unsigned long min_nr_alloc_buckets
,
1468 unsigned long max_nr_buckets
,
1470 const struct cds_lfht_mm_type
*mm
,
1471 const struct rcu_flavor_struct
*flavor
,
1472 pthread_attr_t
*attr
)
1474 struct cds_lfht
*ht
;
1475 unsigned long order
;
1477 /* min_nr_alloc_buckets must be power of two */
1478 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1481 /* init_size must be power of two */
1482 if (!init_size
|| (init_size
& (init_size
- 1)))
1486 * Memory management plugin default.
1489 if (CAA_BITS_PER_LONG
> 32
1491 && max_nr_buckets
<= (1ULL << 32)) {
1493 * For 64-bit architectures, with max number of
1494 * buckets small enough not to use the entire
1495 * 64-bit memory mapping space (and allowing a
1496 * fair number of hash table instances), use the
1497 * mmap allocator, which is faster than the
1500 mm
= &cds_lfht_mm_mmap
;
1503 * The fallback is to use the order allocator.
1505 mm
= &cds_lfht_mm_order
;
1509 /* max_nr_buckets == 0 for order based mm means infinite */
1510 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1511 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1513 /* max_nr_buckets must be power of two */
1514 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1517 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1518 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1519 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1520 init_size
= min(init_size
, max_nr_buckets
);
1522 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1524 assert(ht
->mm
== mm
);
1525 assert(ht
->bucket_at
== mm
->bucket_at
);
1528 ht
->flavor
= flavor
;
1529 ht
->resize_attr
= attr
;
1530 alloc_split_items_count(ht
);
1531 /* this mutex should not nest in read-side C.S. */
1532 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1533 order
= cds_lfht_get_count_order_ulong(init_size
);
1534 ht
->resize_target
= 1UL << order
;
1535 cds_lfht_create_bucket(ht
, 1UL << order
);
1536 ht
->size
= 1UL << order
;
1540 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1541 cds_lfht_match_fct match
, const void *key
,
1542 struct cds_lfht_iter
*iter
)
1544 struct cds_lfht_node
*node
, *next
, *bucket
;
1545 unsigned long reverse_hash
, size
;
1547 reverse_hash
= bit_reverse_ulong(hash
);
1549 size
= rcu_dereference(ht
->size
);
1550 bucket
= lookup_bucket(ht
, size
, hash
);
1551 /* We can always skip the bucket node initially */
1552 node
= rcu_dereference(bucket
->next
);
1553 node
= clear_flag(node
);
1555 if (caa_unlikely(is_end(node
))) {
1559 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1563 next
= rcu_dereference(node
->next
);
1564 assert(node
== clear_flag(node
));
1565 if (caa_likely(!is_removed(next
))
1567 && node
->reverse_hash
== reverse_hash
1568 && caa_likely(match(node
, key
))) {
1571 node
= clear_flag(next
);
1573 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1578 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1579 const void *key
, struct cds_lfht_iter
*iter
)
1581 struct cds_lfht_node
*node
, *next
;
1582 unsigned long reverse_hash
;
1585 reverse_hash
= node
->reverse_hash
;
1587 node
= clear_flag(next
);
1590 if (caa_unlikely(is_end(node
))) {
1594 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1598 next
= rcu_dereference(node
->next
);
1599 if (caa_likely(!is_removed(next
))
1601 && caa_likely(match(node
, key
))) {
1604 node
= clear_flag(next
);
1606 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1611 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1613 struct cds_lfht_node
*node
, *next
;
1615 node
= clear_flag(iter
->next
);
1617 if (caa_unlikely(is_end(node
))) {
1621 next
= rcu_dereference(node
->next
);
1622 if (caa_likely(!is_removed(next
))
1623 && !is_bucket(next
)) {
1626 node
= clear_flag(next
);
1628 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1633 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1636 * Get next after first bucket node. The first bucket node is the
1637 * first node of the linked list.
1639 iter
->next
= bucket_at(ht
, 0)->next
;
1640 cds_lfht_next(ht
, iter
);
1643 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1644 struct cds_lfht_node
*node
)
1648 node
->reverse_hash
= bit_reverse_ulong(hash
);
1649 size
= rcu_dereference(ht
->size
);
1650 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1651 ht_count_add(ht
, size
, hash
);
1654 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1656 cds_lfht_match_fct match
,
1658 struct cds_lfht_node
*node
)
1661 struct cds_lfht_iter iter
;
1663 node
->reverse_hash
= bit_reverse_ulong(hash
);
1664 size
= rcu_dereference(ht
->size
);
1665 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1666 if (iter
.node
== node
)
1667 ht_count_add(ht
, size
, hash
);
1671 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1673 cds_lfht_match_fct match
,
1675 struct cds_lfht_node
*node
)
1678 struct cds_lfht_iter iter
;
1680 node
->reverse_hash
= bit_reverse_ulong(hash
);
1681 size
= rcu_dereference(ht
->size
);
1683 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1684 if (iter
.node
== node
) {
1685 ht_count_add(ht
, size
, hash
);
1689 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1694 int cds_lfht_replace(struct cds_lfht
*ht
,
1695 struct cds_lfht_iter
*old_iter
,
1697 cds_lfht_match_fct match
,
1699 struct cds_lfht_node
*new_node
)
1703 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1704 if (!old_iter
->node
)
1706 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1708 if (caa_unlikely(!match(old_iter
->node
, key
)))
1710 size
= rcu_dereference(ht
->size
);
1711 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1715 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1720 size
= rcu_dereference(ht
->size
);
1721 ret
= _cds_lfht_del(ht
, size
, node
);
1725 hash
= bit_reverse_ulong(node
->reverse_hash
);
1726 ht_count_del(ht
, size
, hash
);
1731 int cds_lfht_is_node_deleted(struct cds_lfht_node
*node
)
1733 return is_removed(CMM_LOAD_SHARED(node
->next
));
1737 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1739 struct cds_lfht_node
*node
;
1740 unsigned long order
, i
, size
;
1742 /* Check that the table is empty */
1743 node
= bucket_at(ht
, 0);
1745 node
= clear_flag(node
)->next
;
1746 if (!is_bucket(node
))
1748 assert(!is_removed(node
));
1749 assert(!is_removal_owner(node
));
1750 } while (!is_end(node
));
1752 * size accessed without rcu_dereference because hash table is
1756 /* Internal sanity check: all nodes left should be buckets */
1757 for (i
= 0; i
< size
; i
++) {
1758 node
= bucket_at(ht
, i
);
1759 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1760 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1761 assert(is_bucket(node
->next
));
1764 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1765 cds_lfht_free_bucket_table(ht
, order
);
1771 * Should only be called when no more concurrent readers nor writers can
1772 * possibly access the table.
1774 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1776 int ret
, was_online
;
1778 /* Wait for in-flight resize operations to complete */
1779 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1780 cmm_smp_mb(); /* Store destroy before load resize */
1781 was_online
= ht
->flavor
->read_ongoing();
1783 ht
->flavor
->thread_offline();
1784 /* Calling with RCU read-side held is an error. */
1785 if (ht
->flavor
->read_ongoing()) {
1788 ht
->flavor
->thread_online();
1791 while (uatomic_read(&ht
->in_progress_resize
))
1792 poll(NULL
, 0, 100); /* wait for 100ms */
1794 ht
->flavor
->thread_online();
1795 ret
= cds_lfht_delete_bucket(ht
);
1798 free_split_items_count(ht
);
1800 *attr
= ht
->resize_attr
;
1806 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1807 long *approx_before
,
1808 unsigned long *count
,
1811 struct cds_lfht_node
*node
, *next
;
1812 unsigned long nr_bucket
= 0, nr_removed
= 0;
1815 if (ht
->split_count
) {
1818 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1819 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1820 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1826 /* Count non-bucket nodes in the table */
1827 node
= bucket_at(ht
, 0);
1829 next
= rcu_dereference(node
->next
);
1830 if (is_removed(next
)) {
1831 if (!is_bucket(next
))
1835 } else if (!is_bucket(next
))
1839 node
= clear_flag(next
);
1840 } while (!is_end(node
));
1841 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
1842 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1844 if (ht
->split_count
) {
1847 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1848 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1849 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1854 /* called with resize mutex held */
1856 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1857 unsigned long old_size
, unsigned long new_size
)
1859 unsigned long old_order
, new_order
;
1861 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1862 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1863 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1864 old_size
, old_order
, new_size
, new_order
);
1865 assert(new_size
> old_size
);
1866 init_table(ht
, old_order
+ 1, new_order
);
1869 /* called with resize mutex held */
1871 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1872 unsigned long old_size
, unsigned long new_size
)
1874 unsigned long old_order
, new_order
;
1876 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1877 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1878 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1879 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1880 old_size
, old_order
, new_size
, new_order
);
1881 assert(new_size
< old_size
);
1883 /* Remove and unlink all bucket nodes to remove. */
1884 fini_table(ht
, new_order
+ 1, old_order
);
1888 /* called with resize mutex held */
1890 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1892 unsigned long new_size
, old_size
;
1895 * Resize table, re-do if the target size has changed under us.
1898 assert(uatomic_read(&ht
->in_progress_resize
));
1899 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1901 ht
->resize_initiated
= 1;
1902 old_size
= ht
->size
;
1903 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1904 if (old_size
< new_size
)
1905 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1906 else if (old_size
> new_size
)
1907 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1908 ht
->resize_initiated
= 0;
1909 /* write resize_initiated before read resize_target */
1911 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1915 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1917 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
1921 void resize_target_update_count(struct cds_lfht
*ht
,
1922 unsigned long count
)
1924 count
= max(count
, MIN_TABLE_SIZE
);
1925 count
= min(count
, ht
->max_nr_buckets
);
1926 uatomic_set(&ht
->resize_target
, count
);
1929 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1933 was_online
= ht
->flavor
->read_ongoing();
1935 ht
->flavor
->thread_offline();
1936 /* Calling with RCU read-side held is an error. */
1937 if (ht
->flavor
->read_ongoing()) {
1938 static int print_once
;
1940 if (!CMM_LOAD_SHARED(print_once
))
1941 fprintf(stderr
, "[error] rculfhash: cds_lfht_resize "
1942 "called with RCU read-side lock held.\n");
1943 CMM_STORE_SHARED(print_once
, 1);
1947 resize_target_update_count(ht
, new_size
);
1948 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1949 pthread_mutex_lock(&ht
->resize_mutex
);
1950 _do_cds_lfht_resize(ht
);
1951 pthread_mutex_unlock(&ht
->resize_mutex
);
1954 ht
->flavor
->thread_online();
1958 void do_resize_cb(struct rcu_head
*head
)
1960 struct rcu_resize_work
*work
=
1961 caa_container_of(head
, struct rcu_resize_work
, head
);
1962 struct cds_lfht
*ht
= work
->ht
;
1964 ht
->flavor
->thread_offline();
1965 pthread_mutex_lock(&ht
->resize_mutex
);
1966 _do_cds_lfht_resize(ht
);
1967 pthread_mutex_unlock(&ht
->resize_mutex
);
1968 ht
->flavor
->thread_online();
1970 cmm_smp_mb(); /* finish resize before decrement */
1971 uatomic_dec(&ht
->in_progress_resize
);
1975 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1977 struct rcu_resize_work
*work
;
1979 /* Store resize_target before read resize_initiated */
1981 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
1982 uatomic_inc(&ht
->in_progress_resize
);
1983 cmm_smp_mb(); /* increment resize count before load destroy */
1984 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1985 uatomic_dec(&ht
->in_progress_resize
);
1988 work
= malloc(sizeof(*work
));
1990 dbg_printf("error allocating resize work, bailing out\n");
1991 uatomic_dec(&ht
->in_progress_resize
);
1995 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1996 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
2001 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2003 unsigned long target_size
= size
<< growth
;
2005 target_size
= min(target_size
, ht
->max_nr_buckets
);
2006 if (resize_target_grow(ht
, target_size
) >= target_size
)
2009 __cds_lfht_resize_lazy_launch(ht
);
2013 * We favor grow operations over shrink. A shrink operation never occurs
2014 * if a grow operation is queued for lazy execution. A grow operation
2015 * cancels any pending shrink lazy execution.
2018 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2019 unsigned long count
)
2021 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2023 count
= max(count
, MIN_TABLE_SIZE
);
2024 count
= min(count
, ht
->max_nr_buckets
);
2026 return; /* Already the right size, no resize needed */
2027 if (count
> size
) { /* lazy grow */
2028 if (resize_target_grow(ht
, count
) >= count
)
2030 } else { /* lazy shrink */
2034 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2036 break; /* no resize needed */
2038 return; /* growing is/(was just) in progress */
2040 return; /* some other thread do shrink */
2044 __cds_lfht_resize_lazy_launch(ht
);