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 * worker 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 | |
268 #include "compat-getcpu.h"
269 #include <urcu/pointer.h>
270 #include <urcu/call-rcu.h>
271 #include <urcu/flavor.h>
272 #include <urcu/arch.h>
273 #include <urcu/uatomic.h>
274 #include <urcu/compiler.h>
275 #include <urcu/rculfhash.h>
276 #include <rculfhash-internal.h>
280 #include "workqueue.h"
281 #include "urcu-die.h"
284 * Split-counters lazily update the global counter each 1024
285 * addition/removal. It automatically keeps track of resize required.
286 * We use the bucket length as indicator for need to expand for small
287 * tables and machines lacking per-cpu data support.
289 #define COUNT_COMMIT_ORDER 10
290 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
291 #define CHAIN_LEN_TARGET 1
292 #define CHAIN_LEN_RESIZE_THRESHOLD 3
295 * Define the minimum table size.
297 #define MIN_TABLE_ORDER 0
298 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
301 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
303 #define MIN_PARTITION_PER_THREAD_ORDER 12
304 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
307 * The removed flag needs to be updated atomically with the pointer.
308 * It indicates that no node must attach to the node scheduled for
309 * removal, and that node garbage collection must be performed.
310 * The bucket flag does not require to be updated atomically with the
311 * pointer, but it is added as a pointer low bit flag to save space.
312 * The "removal owner" flag is used to detect which of the "del"
313 * operation that has set the "removed flag" gets to return the removed
314 * node to its caller. Note that the replace operation does not need to
315 * iteract with the "removal owner" flag, because it validates that
316 * the "removed" flag is not set before performing its cmpxchg.
318 #define REMOVED_FLAG (1UL << 0)
319 #define BUCKET_FLAG (1UL << 1)
320 #define REMOVAL_OWNER_FLAG (1UL << 2)
321 #define FLAGS_MASK ((1UL << 3) - 1)
323 /* Value of the end pointer. Should not interact with flags. */
324 #define END_VALUE NULL
327 * ht_items_count: Split-counters counting the number of node addition
328 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
329 * is set at hash table creation.
331 * These are free-running counters, never reset to zero. They count the
332 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
333 * operations to update the global counter. We choose a power-of-2 value
334 * for the trigger to deal with 32 or 64-bit overflow of the counter.
336 struct ht_items_count
{
337 unsigned long add
, del
;
338 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
341 * resize_work: Contains arguments passed to worker thread
342 * responsible for performing lazy resize.
345 struct urcu_work work
;
350 * partition_resize_work: Contains arguments passed to worker threads
351 * executing the hash table resize on partitions of the hash table
352 * assigned to each processor's worker thread.
354 struct partition_resize_work
{
357 unsigned long i
, start
, len
;
358 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
359 unsigned long start
, unsigned long len
);
362 static struct urcu_workqueue
*cds_lfht_workqueue
;
363 static unsigned long cds_lfht_workqueue_user_count
;
366 * Mutex ensuring mutual exclusion between workqueue initialization and
367 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
369 static pthread_mutex_t cds_lfht_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
371 static struct urcu_atfork cds_lfht_atfork
;
374 * atfork handler nesting counters. Handle being registered to many urcu
375 * flavors, thus being possibly invoked more than once in the
376 * pthread_atfork list of callbacks.
378 static int cds_lfht_workqueue_atfork_nesting
;
380 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
);
381 static void cds_lfht_fini_worker(const struct rcu_flavor_struct
*flavor
);
383 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
386 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
391 #define cds_lfht_iter_debug_assert(...) assert(__VA_ARGS__)
396 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
400 #define cds_lfht_iter_debug_assert(...)
405 * Algorithm to reverse bits in a word by lookup table, extended to
408 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
409 * Originally from Public Domain.
412 static const uint8_t BitReverseTable256
[256] =
414 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
415 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
416 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
417 R6(0), R6(2), R6(1), R6(3)
424 uint8_t bit_reverse_u8(uint8_t v
)
426 return BitReverseTable256
[v
];
429 #if (CAA_BITS_PER_LONG == 32)
431 uint32_t bit_reverse_u32(uint32_t v
)
433 return ((uint32_t) bit_reverse_u8(v
) << 24) |
434 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
435 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
436 ((uint32_t) bit_reverse_u8(v
>> 24));
440 uint64_t bit_reverse_u64(uint64_t v
)
442 return ((uint64_t) bit_reverse_u8(v
) << 56) |
443 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
444 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
445 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
446 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
447 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
448 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
449 ((uint64_t) bit_reverse_u8(v
>> 56));
454 unsigned long bit_reverse_ulong(unsigned long v
)
456 #if (CAA_BITS_PER_LONG == 32)
457 return bit_reverse_u32(v
);
459 return bit_reverse_u64(v
);
464 * fls: returns the position of the most significant bit.
465 * Returns 0 if no bit is set, else returns the position of the most
466 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
468 #if defined(__i386) || defined(__x86_64)
470 unsigned int fls_u32(uint32_t x
)
474 __asm__ ("bsrl %1,%0\n\t"
478 : "=r" (r
) : "rm" (x
));
484 #if defined(__x86_64)
486 unsigned int fls_u64(uint64_t x
)
490 __asm__ ("bsrq %1,%0\n\t"
494 : "=r" (r
) : "rm" (x
));
501 static __attribute__((unused
))
502 unsigned int fls_u64(uint64_t x
)
509 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
513 if (!(x
& 0xFFFF000000000000ULL
)) {
517 if (!(x
& 0xFF00000000000000ULL
)) {
521 if (!(x
& 0xF000000000000000ULL
)) {
525 if (!(x
& 0xC000000000000000ULL
)) {
529 if (!(x
& 0x8000000000000000ULL
)) {
538 static __attribute__((unused
))
539 unsigned int fls_u32(uint32_t x
)
545 if (!(x
& 0xFFFF0000U
)) {
549 if (!(x
& 0xFF000000U
)) {
553 if (!(x
& 0xF0000000U
)) {
557 if (!(x
& 0xC0000000U
)) {
561 if (!(x
& 0x80000000U
)) {
569 unsigned int cds_lfht_fls_ulong(unsigned long x
)
571 #if (CAA_BITS_PER_LONG == 32)
579 * Return the minimum order for which x <= (1UL << order).
580 * Return -1 if x is 0.
582 int cds_lfht_get_count_order_u32(uint32_t x
)
587 return fls_u32(x
- 1);
591 * Return the minimum order for which x <= (1UL << order).
592 * Return -1 if x is 0.
594 int cds_lfht_get_count_order_ulong(unsigned long x
)
599 return cds_lfht_fls_ulong(x
- 1);
603 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
606 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
607 unsigned long count
);
609 static void mutex_lock(pthread_mutex_t
*mutex
)
613 #ifndef DISTRUST_SIGNALS_EXTREME
614 ret
= pthread_mutex_lock(mutex
);
617 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
618 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
619 if (ret
!= EBUSY
&& ret
!= EINTR
)
621 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
623 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
626 (void) poll(NULL
, 0, 10);
628 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
631 static void mutex_unlock(pthread_mutex_t
*mutex
)
635 ret
= pthread_mutex_unlock(mutex
);
640 static long nr_cpus_mask
= -1;
641 static long split_count_mask
= -1;
642 static int split_count_order
= -1;
644 #if defined(HAVE_SYSCONF)
645 static void ht_init_nr_cpus_mask(void)
649 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
655 * round up number of CPUs to next power of two, so we
656 * can use & for modulo.
658 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
659 nr_cpus_mask
= maxcpus
- 1;
661 #else /* #if defined(HAVE_SYSCONF) */
662 static void ht_init_nr_cpus_mask(void)
666 #endif /* #else #if defined(HAVE_SYSCONF) */
669 void alloc_split_items_count(struct cds_lfht
*ht
)
671 if (nr_cpus_mask
== -1) {
672 ht_init_nr_cpus_mask();
673 if (nr_cpus_mask
< 0)
674 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
676 split_count_mask
= nr_cpus_mask
;
678 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
681 assert(split_count_mask
>= 0);
683 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
684 ht
->split_count
= calloc(split_count_mask
+ 1,
685 sizeof(struct ht_items_count
));
686 assert(ht
->split_count
);
688 ht
->split_count
= NULL
;
693 void free_split_items_count(struct cds_lfht
*ht
)
695 poison_free(ht
->split_count
);
699 int ht_get_split_count_index(unsigned long hash
)
703 assert(split_count_mask
>= 0);
704 cpu
= urcu_sched_getcpu();
705 if (caa_unlikely(cpu
< 0))
706 return hash
& split_count_mask
;
708 return cpu
& split_count_mask
;
712 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
714 unsigned long split_count
;
718 if (caa_unlikely(!ht
->split_count
))
720 index
= ht_get_split_count_index(hash
);
721 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
722 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
724 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
726 dbg_printf("add split count %lu\n", split_count
);
727 count
= uatomic_add_return(&ht
->count
,
728 1UL << COUNT_COMMIT_ORDER
);
729 if (caa_likely(count
& (count
- 1)))
731 /* Only if global count is power of 2 */
733 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
735 dbg_printf("add set global %ld\n", count
);
736 cds_lfht_resize_lazy_count(ht
, size
,
737 count
>> (CHAIN_LEN_TARGET
- 1));
741 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
743 unsigned long split_count
;
747 if (caa_unlikely(!ht
->split_count
))
749 index
= ht_get_split_count_index(hash
);
750 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
751 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
753 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
755 dbg_printf("del split count %lu\n", split_count
);
756 count
= uatomic_add_return(&ht
->count
,
757 -(1UL << COUNT_COMMIT_ORDER
));
758 if (caa_likely(count
& (count
- 1)))
760 /* Only if global count is power of 2 */
762 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
764 dbg_printf("del set global %ld\n", count
);
766 * Don't shrink table if the number of nodes is below a
769 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
771 cds_lfht_resize_lazy_count(ht
, size
,
772 count
>> (CHAIN_LEN_TARGET
- 1));
776 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
780 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
782 count
= uatomic_read(&ht
->count
);
784 * Use bucket-local length for small table expand and for
785 * environments lacking per-cpu data support.
787 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
790 dbg_printf("WARNING: large chain length: %u.\n",
792 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
796 * Ideal growth calculated based on chain length.
798 growth
= cds_lfht_get_count_order_u32(chain_len
799 - (CHAIN_LEN_TARGET
- 1));
800 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
802 >= (1UL << (COUNT_COMMIT_ORDER
803 + split_count_order
))) {
805 * If ideal growth expands the hash table size
806 * beyond the "small hash table" sizes, use the
807 * maximum small hash table size to attempt
808 * expanding the hash table. This only applies
809 * when node accounting is available, otherwise
810 * the chain length is used to expand the hash
811 * table in every case.
813 growth
= COUNT_COMMIT_ORDER
+ split_count_order
814 - cds_lfht_get_count_order_ulong(size
);
818 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
823 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
825 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
829 int is_removed(struct cds_lfht_node
*node
)
831 return ((unsigned long) node
) & REMOVED_FLAG
;
835 int is_bucket(struct cds_lfht_node
*node
)
837 return ((unsigned long) node
) & BUCKET_FLAG
;
841 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
843 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
847 int is_removal_owner(struct cds_lfht_node
*node
)
849 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
853 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
855 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
859 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
861 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
865 struct cds_lfht_node
*get_end(void)
867 return (struct cds_lfht_node
*) END_VALUE
;
871 int is_end(struct cds_lfht_node
*node
)
873 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
877 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
880 unsigned long old1
, old2
;
882 old1
= uatomic_read(ptr
);
887 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
892 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
894 return ht
->mm
->alloc_bucket_table(ht
, order
);
898 * cds_lfht_free_bucket_table() should be called with decreasing order.
899 * When cds_lfht_free_bucket_table(0) is called, it means the whole
903 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
905 return ht
->mm
->free_bucket_table(ht
, order
);
909 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
911 return ht
->bucket_at(ht
, index
);
915 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
919 return bucket_at(ht
, hash
& (size
- 1));
923 * Remove all logically deleted nodes from a bucket up to a certain node key.
926 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
928 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
930 assert(!is_bucket(bucket
));
931 assert(!is_removed(bucket
));
932 assert(!is_removal_owner(bucket
));
933 assert(!is_bucket(node
));
934 assert(!is_removed(node
));
935 assert(!is_removal_owner(node
));
938 /* We can always skip the bucket node initially */
939 iter
= rcu_dereference(iter_prev
->next
);
940 assert(!is_removed(iter
));
941 assert(!is_removal_owner(iter
));
942 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
944 * We should never be called with bucket (start of chain)
945 * and logically removed node (end of path compression
946 * marker) being the actual same node. This would be a
947 * bug in the algorithm implementation.
949 assert(bucket
!= node
);
951 if (caa_unlikely(is_end(iter
)))
953 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
955 next
= rcu_dereference(clear_flag(iter
)->next
);
956 if (caa_likely(is_removed(next
)))
958 iter_prev
= clear_flag(iter
);
961 assert(!is_removed(iter
));
962 assert(!is_removal_owner(iter
));
964 new_next
= flag_bucket(clear_flag(next
));
966 new_next
= clear_flag(next
);
967 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
972 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
973 struct cds_lfht_node
*old_node
,
974 struct cds_lfht_node
*old_next
,
975 struct cds_lfht_node
*new_node
)
977 struct cds_lfht_node
*bucket
, *ret_next
;
979 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
982 assert(!is_removed(old_node
));
983 assert(!is_removal_owner(old_node
));
984 assert(!is_bucket(old_node
));
985 assert(!is_removed(new_node
));
986 assert(!is_removal_owner(new_node
));
987 assert(!is_bucket(new_node
));
988 assert(new_node
!= old_node
);
990 /* Insert after node to be replaced */
991 if (is_removed(old_next
)) {
993 * Too late, the old node has been removed under us
994 * between lookup and replace. Fail.
998 assert(old_next
== clear_flag(old_next
));
999 assert(new_node
!= old_next
);
1001 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
1002 * flag. It is either set atomically at the same time
1003 * (replace) or after (del).
1005 assert(!is_removal_owner(old_next
));
1006 new_node
->next
= old_next
;
1008 * Here is the whole trick for lock-free replace: we add
1009 * the replacement node _after_ the node we want to
1010 * replace by atomically setting its next pointer at the
1011 * same time we set its removal flag. Given that
1012 * the lookups/get next use an iterator aware of the
1013 * next pointer, they will either skip the old node due
1014 * to the removal flag and see the new node, or use
1015 * the old node, but will not see the new one.
1016 * This is a replacement of a node with another node
1017 * that has the same value: we are therefore not
1018 * removing a value from the hash table. We set both the
1019 * REMOVED and REMOVAL_OWNER flags atomically so we own
1020 * the node after successful cmpxchg.
1022 ret_next
= uatomic_cmpxchg(&old_node
->next
,
1023 old_next
, flag_removed_or_removal_owner(new_node
));
1024 if (ret_next
== old_next
)
1025 break; /* We performed the replacement. */
1026 old_next
= ret_next
;
1030 * Ensure that the old node is not visible to readers anymore:
1031 * lookup for the node, and remove it (along with any other
1032 * logically removed node) if found.
1034 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
1035 _cds_lfht_gc_bucket(bucket
, new_node
);
1037 assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
1042 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1043 * mode. A NULL unique_ret allows creation of duplicate keys.
1046 void _cds_lfht_add(struct cds_lfht
*ht
,
1048 cds_lfht_match_fct match
,
1051 struct cds_lfht_node
*node
,
1052 struct cds_lfht_iter
*unique_ret
,
1055 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
1057 struct cds_lfht_node
*bucket
;
1059 assert(!is_bucket(node
));
1060 assert(!is_removed(node
));
1061 assert(!is_removal_owner(node
));
1062 bucket
= lookup_bucket(ht
, size
, hash
);
1064 uint32_t chain_len
= 0;
1067 * iter_prev points to the non-removed node prior to the
1071 /* We can always skip the bucket node initially */
1072 iter
= rcu_dereference(iter_prev
->next
);
1073 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1075 if (caa_unlikely(is_end(iter
)))
1077 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1080 /* bucket node is the first node of the identical-hash-value chain */
1081 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1084 next
= rcu_dereference(clear_flag(iter
)->next
);
1085 if (caa_unlikely(is_removed(next
)))
1091 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1092 struct cds_lfht_iter d_iter
= {
1095 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1101 * uniquely adding inserts the node as the first
1102 * node of the identical-hash-value node chain.
1104 * This semantic ensures no duplicated keys
1105 * should ever be observable in the table
1106 * (including traversing the table node by
1107 * node by forward iterations)
1109 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1113 *unique_ret
= d_iter
;
1117 /* Only account for identical reverse hash once */
1118 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1119 && !is_bucket(next
))
1120 check_resize(ht
, size
, ++chain_len
);
1121 iter_prev
= clear_flag(iter
);
1126 assert(node
!= clear_flag(iter
));
1127 assert(!is_removed(iter_prev
));
1128 assert(!is_removal_owner(iter_prev
));
1129 assert(!is_removed(iter
));
1130 assert(!is_removal_owner(iter
));
1131 assert(iter_prev
!= node
);
1133 node
->next
= clear_flag(iter
);
1135 node
->next
= flag_bucket(clear_flag(iter
));
1136 if (is_bucket(iter
))
1137 new_node
= flag_bucket(node
);
1140 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1141 new_node
) != iter
) {
1142 continue; /* retry */
1149 assert(!is_removed(iter
));
1150 assert(!is_removal_owner(iter
));
1151 if (is_bucket(iter
))
1152 new_next
= flag_bucket(clear_flag(next
));
1154 new_next
= clear_flag(next
);
1155 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1160 unique_ret
->node
= return_node
;
1161 /* unique_ret->next left unset, never used. */
1166 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1167 struct cds_lfht_node
*node
)
1169 struct cds_lfht_node
*bucket
, *next
;
1171 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1174 /* logically delete the node */
1175 assert(!is_bucket(node
));
1176 assert(!is_removed(node
));
1177 assert(!is_removal_owner(node
));
1180 * We are first checking if the node had previously been
1181 * logically removed (this check is not atomic with setting the
1182 * logical removal flag). Return -ENOENT if the node had
1183 * previously been removed.
1185 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1186 if (caa_unlikely(is_removed(next
)))
1188 assert(!is_bucket(next
));
1190 * The del operation semantic guarantees a full memory barrier
1191 * before the uatomic_or atomic commit of the deletion flag.
1193 cmm_smp_mb__before_uatomic_or();
1195 * We set the REMOVED_FLAG unconditionally. Note that there may
1196 * be more than one concurrent thread setting this flag.
1197 * Knowing which wins the race will be known after the garbage
1198 * collection phase, stay tuned!
1200 uatomic_or(&node
->next
, REMOVED_FLAG
);
1201 /* We performed the (logical) deletion. */
1204 * Ensure that the node is not visible to readers anymore: lookup for
1205 * the node, and remove it (along with any other logically removed node)
1208 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1209 _cds_lfht_gc_bucket(bucket
, node
);
1211 assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1213 * Last phase: atomically exchange node->next with a version
1214 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1215 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1216 * the node and win the removal race.
1217 * It is interesting to note that all "add" paths are forbidden
1218 * to change the next pointer starting from the point where the
1219 * REMOVED_FLAG is set, so here using a read, followed by a
1220 * xchg() suffice to guarantee that the xchg() will ever only
1221 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1224 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1225 flag_removal_owner(node
->next
))))
1232 void *partition_resize_thread(void *arg
)
1234 struct partition_resize_work
*work
= arg
;
1236 work
->ht
->flavor
->register_thread();
1237 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1238 work
->ht
->flavor
->unregister_thread();
1243 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1245 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1246 unsigned long start
, unsigned long len
))
1248 unsigned long partition_len
, start
= 0;
1249 struct partition_resize_work
*work
;
1251 unsigned long nr_threads
;
1253 assert(nr_cpus_mask
!= -1);
1254 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
)
1258 * Note: nr_cpus_mask + 1 is always power of 2.
1259 * We spawn just the number of threads we need to satisfy the minimum
1260 * partition size, up to the number of CPUs in the system.
1262 if (nr_cpus_mask
> 0) {
1263 nr_threads
= min(nr_cpus_mask
+ 1,
1264 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1268 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1269 work
= calloc(nr_threads
, sizeof(*work
));
1271 dbg_printf("error allocating for resize, single-threading\n");
1274 for (thread
= 0; thread
< nr_threads
; thread
++) {
1275 work
[thread
].ht
= ht
;
1277 work
[thread
].len
= partition_len
;
1278 work
[thread
].start
= thread
* partition_len
;
1279 work
[thread
].fct
= fct
;
1280 ret
= pthread_create(&(work
[thread
].thread_id
), ht
->resize_attr
,
1281 partition_resize_thread
, &work
[thread
]);
1282 if (ret
== EAGAIN
) {
1284 * Out of resources: wait and join the threads
1285 * we've created, then handle leftovers.
1287 dbg_printf("error spawning for resize, single-threading\n");
1288 start
= work
[thread
].start
;
1290 nr_threads
= thread
;
1295 for (thread
= 0; thread
< nr_threads
; thread
++) {
1296 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1302 * A pthread_create failure above will either lead in us having
1303 * no threads to join or starting at a non-zero offset,
1304 * fallback to single thread processing of leftovers.
1306 if (start
== 0 && nr_threads
> 0)
1309 fct(ht
, i
, start
, len
);
1313 * Holding RCU read lock to protect _cds_lfht_add against memory
1314 * reclaim that could be performed by other worker threads (ABA
1317 * When we reach a certain length, we can split this population phase over
1318 * many worker threads, based on the number of CPUs available in the system.
1319 * This should therefore take care of not having the expand lagging behind too
1320 * many concurrent insertion threads by using the scheduler's ability to
1321 * schedule bucket node population fairly with insertions.
1324 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1325 unsigned long start
, unsigned long len
)
1327 unsigned long j
, size
= 1UL << (i
- 1);
1329 assert(i
> MIN_TABLE_ORDER
);
1330 ht
->flavor
->read_lock();
1331 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1332 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1334 assert(j
>= size
&& j
< (size
<< 1));
1335 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1337 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1338 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1340 ht
->flavor
->read_unlock();
1344 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1347 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1351 void init_table(struct cds_lfht
*ht
,
1352 unsigned long first_order
, unsigned long last_order
)
1356 dbg_printf("init table: first_order %lu last_order %lu\n",
1357 first_order
, last_order
);
1358 assert(first_order
> MIN_TABLE_ORDER
);
1359 for (i
= first_order
; i
<= last_order
; i
++) {
1362 len
= 1UL << (i
- 1);
1363 dbg_printf("init order %lu len: %lu\n", i
, len
);
1365 /* Stop expand if the resize target changes under us */
1366 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1369 cds_lfht_alloc_bucket_table(ht
, i
);
1372 * Set all bucket nodes reverse hash values for a level and
1373 * link all bucket nodes into the table.
1375 init_table_populate(ht
, i
, len
);
1378 * Update table size.
1380 cmm_smp_wmb(); /* populate data before RCU size */
1381 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1383 dbg_printf("init new size: %lu\n", 1UL << i
);
1384 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1390 * Holding RCU read lock to protect _cds_lfht_remove against memory
1391 * reclaim that could be performed by other worker threads (ABA
1393 * For a single level, we logically remove and garbage collect each node.
1395 * As a design choice, we perform logical removal and garbage collection on a
1396 * node-per-node basis to simplify this algorithm. We also assume keeping good
1397 * cache locality of the operation would overweight possible performance gain
1398 * that could be achieved by batching garbage collection for multiple levels.
1399 * However, this would have to be justified by benchmarks.
1401 * Concurrent removal and add operations are helping us perform garbage
1402 * collection of logically removed nodes. We guarantee that all logically
1403 * removed nodes have been garbage-collected (unlinked) before work
1404 * enqueue is invoked to free a hole level of bucket nodes (after a
1407 * Logical removal and garbage collection can therefore be done in batch
1408 * or on a node-per-node basis, as long as the guarantee above holds.
1410 * When we reach a certain length, we can split this removal over many worker
1411 * threads, based on the number of CPUs available in the system. This should
1412 * take care of not letting resize process lag behind too many concurrent
1413 * updater threads actively inserting into the hash table.
1416 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1417 unsigned long start
, unsigned long len
)
1419 unsigned long j
, size
= 1UL << (i
- 1);
1421 assert(i
> MIN_TABLE_ORDER
);
1422 ht
->flavor
->read_lock();
1423 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1424 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1425 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1427 assert(j
>= size
&& j
< (size
<< 1));
1428 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1430 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1431 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1432 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1434 ht
->flavor
->read_unlock();
1438 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1440 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1444 * fini_table() is never called for first_order == 0, which is why
1445 * free_by_rcu_order == 0 can be used as criterion to know if free must
1449 void fini_table(struct cds_lfht
*ht
,
1450 unsigned long first_order
, unsigned long last_order
)
1453 unsigned long free_by_rcu_order
= 0;
1455 dbg_printf("fini table: first_order %lu last_order %lu\n",
1456 first_order
, last_order
);
1457 assert(first_order
> MIN_TABLE_ORDER
);
1458 for (i
= last_order
; i
>= first_order
; i
--) {
1461 len
= 1UL << (i
- 1);
1462 dbg_printf("fini order %ld len: %lu\n", i
, len
);
1464 /* Stop shrink if the resize target changes under us */
1465 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1468 cmm_smp_wmb(); /* populate data before RCU size */
1469 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1472 * We need to wait for all add operations to reach Q.S. (and
1473 * thus use the new table for lookups) before we can start
1474 * releasing the old bucket nodes. Otherwise their lookup will
1475 * return a logically removed node as insert position.
1477 ht
->flavor
->update_synchronize_rcu();
1478 if (free_by_rcu_order
)
1479 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1482 * Set "removed" flag in bucket nodes about to be removed.
1483 * Unlink all now-logically-removed bucket node pointers.
1484 * Concurrent add/remove operation are helping us doing
1487 remove_table(ht
, i
, len
);
1489 free_by_rcu_order
= i
;
1491 dbg_printf("fini new size: %lu\n", 1UL << i
);
1492 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1496 if (free_by_rcu_order
) {
1497 ht
->flavor
->update_synchronize_rcu();
1498 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1503 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1505 struct cds_lfht_node
*prev
, *node
;
1506 unsigned long order
, len
, i
;
1508 cds_lfht_alloc_bucket_table(ht
, 0);
1510 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1511 node
= bucket_at(ht
, 0);
1512 node
->next
= flag_bucket(get_end());
1513 node
->reverse_hash
= 0;
1515 for (order
= 1; order
< cds_lfht_get_count_order_ulong(size
) + 1; order
++) {
1516 len
= 1UL << (order
- 1);
1517 cds_lfht_alloc_bucket_table(ht
, order
);
1519 for (i
= 0; i
< len
; i
++) {
1521 * Now, we are trying to init the node with the
1522 * hash=(len+i) (which is also a bucket with the
1523 * index=(len+i)) and insert it into the hash table,
1524 * so this node has to be inserted after the bucket
1525 * with the index=(len+i)&(len-1)=i. And because there
1526 * is no other non-bucket node nor bucket node with
1527 * larger index/hash inserted, so the bucket node
1528 * being inserted should be inserted directly linked
1529 * after the bucket node with index=i.
1531 prev
= bucket_at(ht
, i
);
1532 node
= bucket_at(ht
, len
+ i
);
1534 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1535 order
, len
+ i
, len
+ i
);
1536 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1538 /* insert after prev */
1539 assert(is_bucket(prev
->next
));
1540 node
->next
= prev
->next
;
1541 prev
->next
= flag_bucket(node
);
1546 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1547 unsigned long min_nr_alloc_buckets
,
1548 unsigned long max_nr_buckets
,
1550 const struct cds_lfht_mm_type
*mm
,
1551 const struct rcu_flavor_struct
*flavor
,
1552 pthread_attr_t
*attr
)
1554 struct cds_lfht
*ht
;
1555 unsigned long order
;
1557 /* min_nr_alloc_buckets must be power of two */
1558 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1561 /* init_size must be power of two */
1562 if (!init_size
|| (init_size
& (init_size
- 1)))
1566 * Memory management plugin default.
1569 if (CAA_BITS_PER_LONG
> 32
1571 && max_nr_buckets
<= (1ULL << 32)) {
1573 * For 64-bit architectures, with max number of
1574 * buckets small enough not to use the entire
1575 * 64-bit memory mapping space (and allowing a
1576 * fair number of hash table instances), use the
1577 * mmap allocator, which is faster than the
1580 mm
= &cds_lfht_mm_mmap
;
1583 * The fallback is to use the order allocator.
1585 mm
= &cds_lfht_mm_order
;
1589 /* max_nr_buckets == 0 for order based mm means infinite */
1590 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1591 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1593 /* max_nr_buckets must be power of two */
1594 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1597 if (flags
& CDS_LFHT_AUTO_RESIZE
)
1598 cds_lfht_init_worker(flavor
);
1600 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1601 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1602 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1603 init_size
= min(init_size
, max_nr_buckets
);
1605 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1607 assert(ht
->mm
== mm
);
1608 assert(ht
->bucket_at
== mm
->bucket_at
);
1611 ht
->flavor
= flavor
;
1612 ht
->resize_attr
= attr
;
1613 alloc_split_items_count(ht
);
1614 /* this mutex should not nest in read-side C.S. */
1615 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1616 order
= cds_lfht_get_count_order_ulong(init_size
);
1617 ht
->resize_target
= 1UL << order
;
1618 cds_lfht_create_bucket(ht
, 1UL << order
);
1619 ht
->size
= 1UL << order
;
1623 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1624 cds_lfht_match_fct match
, const void *key
,
1625 struct cds_lfht_iter
*iter
)
1627 struct cds_lfht_node
*node
, *next
, *bucket
;
1628 unsigned long reverse_hash
, size
;
1630 cds_lfht_iter_debug_set_ht(ht
, iter
);
1632 reverse_hash
= bit_reverse_ulong(hash
);
1634 size
= rcu_dereference(ht
->size
);
1635 bucket
= lookup_bucket(ht
, size
, hash
);
1636 /* We can always skip the bucket node initially */
1637 node
= rcu_dereference(bucket
->next
);
1638 node
= clear_flag(node
);
1640 if (caa_unlikely(is_end(node
))) {
1644 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1648 next
= rcu_dereference(node
->next
);
1649 assert(node
== clear_flag(node
));
1650 if (caa_likely(!is_removed(next
))
1652 && node
->reverse_hash
== reverse_hash
1653 && caa_likely(match(node
, key
))) {
1656 node
= clear_flag(next
);
1658 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1663 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1664 const void *key
, struct cds_lfht_iter
*iter
)
1666 struct cds_lfht_node
*node
, *next
;
1667 unsigned long reverse_hash
;
1669 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1671 reverse_hash
= node
->reverse_hash
;
1673 node
= clear_flag(next
);
1676 if (caa_unlikely(is_end(node
))) {
1680 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1684 next
= rcu_dereference(node
->next
);
1685 if (caa_likely(!is_removed(next
))
1687 && caa_likely(match(node
, key
))) {
1690 node
= clear_flag(next
);
1692 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1697 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1699 struct cds_lfht_node
*node
, *next
;
1701 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1702 node
= clear_flag(iter
->next
);
1704 if (caa_unlikely(is_end(node
))) {
1708 next
= rcu_dereference(node
->next
);
1709 if (caa_likely(!is_removed(next
))
1710 && !is_bucket(next
)) {
1713 node
= clear_flag(next
);
1715 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1720 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1722 cds_lfht_iter_debug_set_ht(ht
, iter
);
1724 * Get next after first bucket node. The first bucket node is the
1725 * first node of the linked list.
1727 iter
->next
= bucket_at(ht
, 0)->next
;
1728 cds_lfht_next(ht
, iter
);
1731 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1732 struct cds_lfht_node
*node
)
1736 node
->reverse_hash
= bit_reverse_ulong(hash
);
1737 size
= rcu_dereference(ht
->size
);
1738 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1739 ht_count_add(ht
, size
, hash
);
1742 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1744 cds_lfht_match_fct match
,
1746 struct cds_lfht_node
*node
)
1749 struct cds_lfht_iter iter
;
1751 node
->reverse_hash
= bit_reverse_ulong(hash
);
1752 size
= rcu_dereference(ht
->size
);
1753 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1754 if (iter
.node
== node
)
1755 ht_count_add(ht
, size
, hash
);
1759 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1761 cds_lfht_match_fct match
,
1763 struct cds_lfht_node
*node
)
1766 struct cds_lfht_iter iter
;
1768 node
->reverse_hash
= bit_reverse_ulong(hash
);
1769 size
= rcu_dereference(ht
->size
);
1771 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1772 if (iter
.node
== node
) {
1773 ht_count_add(ht
, size
, hash
);
1777 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1782 int cds_lfht_replace(struct cds_lfht
*ht
,
1783 struct cds_lfht_iter
*old_iter
,
1785 cds_lfht_match_fct match
,
1787 struct cds_lfht_node
*new_node
)
1791 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1792 if (!old_iter
->node
)
1794 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1796 if (caa_unlikely(!match(old_iter
->node
, key
)))
1798 size
= rcu_dereference(ht
->size
);
1799 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1803 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1808 size
= rcu_dereference(ht
->size
);
1809 ret
= _cds_lfht_del(ht
, size
, node
);
1813 hash
= bit_reverse_ulong(node
->reverse_hash
);
1814 ht_count_del(ht
, size
, hash
);
1819 int cds_lfht_is_node_deleted(struct cds_lfht_node
*node
)
1821 return is_removed(CMM_LOAD_SHARED(node
->next
));
1825 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1827 struct cds_lfht_node
*node
;
1828 unsigned long order
, i
, size
;
1830 /* Check that the table is empty */
1831 node
= bucket_at(ht
, 0);
1833 node
= clear_flag(node
)->next
;
1834 if (!is_bucket(node
))
1836 assert(!is_removed(node
));
1837 assert(!is_removal_owner(node
));
1838 } while (!is_end(node
));
1840 * size accessed without rcu_dereference because hash table is
1844 /* Internal sanity check: all nodes left should be buckets */
1845 for (i
= 0; i
< size
; i
++) {
1846 node
= bucket_at(ht
, i
);
1847 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1848 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1849 assert(is_bucket(node
->next
));
1852 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1853 cds_lfht_free_bucket_table(ht
, order
);
1859 * Should only be called when no more concurrent readers nor writers can
1860 * possibly access the table.
1862 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1866 if (ht
->flags
& CDS_LFHT_AUTO_RESIZE
) {
1867 /* Cancel ongoing resize operations. */
1868 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1869 /* Wait for in-flight resize operations to complete */
1870 urcu_workqueue_flush_queued_work(cds_lfht_workqueue
);
1872 ret
= cds_lfht_delete_bucket(ht
);
1875 free_split_items_count(ht
);
1877 *attr
= ht
->resize_attr
;
1878 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1881 if (ht
->flags
& CDS_LFHT_AUTO_RESIZE
)
1882 cds_lfht_fini_worker(ht
->flavor
);
1887 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1888 long *approx_before
,
1889 unsigned long *count
,
1892 struct cds_lfht_node
*node
, *next
;
1893 unsigned long nr_bucket
= 0, nr_removed
= 0;
1896 if (ht
->split_count
) {
1899 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1900 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1901 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1907 /* Count non-bucket nodes in the table */
1908 node
= bucket_at(ht
, 0);
1910 next
= rcu_dereference(node
->next
);
1911 if (is_removed(next
)) {
1912 if (!is_bucket(next
))
1916 } else if (!is_bucket(next
))
1920 node
= clear_flag(next
);
1921 } while (!is_end(node
));
1922 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
1923 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1925 if (ht
->split_count
) {
1928 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1929 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1930 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1935 /* called with resize mutex held */
1937 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1938 unsigned long old_size
, unsigned long new_size
)
1940 unsigned long old_order
, new_order
;
1942 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1943 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1944 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1945 old_size
, old_order
, new_size
, new_order
);
1946 assert(new_size
> old_size
);
1947 init_table(ht
, old_order
+ 1, new_order
);
1950 /* called with resize mutex held */
1952 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1953 unsigned long old_size
, unsigned long new_size
)
1955 unsigned long old_order
, new_order
;
1957 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1958 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1959 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1960 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1961 old_size
, old_order
, new_size
, new_order
);
1962 assert(new_size
< old_size
);
1964 /* Remove and unlink all bucket nodes to remove. */
1965 fini_table(ht
, new_order
+ 1, old_order
);
1969 /* called with resize mutex held */
1971 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1973 unsigned long new_size
, old_size
;
1976 * Resize table, re-do if the target size has changed under us.
1979 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1981 ht
->resize_initiated
= 1;
1982 old_size
= ht
->size
;
1983 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1984 if (old_size
< new_size
)
1985 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1986 else if (old_size
> new_size
)
1987 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1988 ht
->resize_initiated
= 0;
1989 /* write resize_initiated before read resize_target */
1991 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1995 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1997 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
2001 void resize_target_update_count(struct cds_lfht
*ht
,
2002 unsigned long count
)
2004 count
= max(count
, MIN_TABLE_SIZE
);
2005 count
= min(count
, ht
->max_nr_buckets
);
2006 uatomic_set(&ht
->resize_target
, count
);
2009 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
2011 resize_target_update_count(ht
, new_size
);
2012 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
2013 mutex_lock(&ht
->resize_mutex
);
2014 _do_cds_lfht_resize(ht
);
2015 mutex_unlock(&ht
->resize_mutex
);
2019 void do_resize_cb(struct urcu_work
*work
)
2021 struct resize_work
*resize_work
=
2022 caa_container_of(work
, struct resize_work
, work
);
2023 struct cds_lfht
*ht
= resize_work
->ht
;
2025 ht
->flavor
->register_thread();
2026 mutex_lock(&ht
->resize_mutex
);
2027 _do_cds_lfht_resize(ht
);
2028 mutex_unlock(&ht
->resize_mutex
);
2029 ht
->flavor
->unregister_thread();
2034 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
2036 struct resize_work
*work
;
2038 /* Store resize_target before read resize_initiated */
2040 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
2041 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
2044 work
= malloc(sizeof(*work
));
2046 dbg_printf("error allocating resize work, bailing out\n");
2050 urcu_workqueue_queue_work(cds_lfht_workqueue
,
2051 &work
->work
, do_resize_cb
);
2052 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
2057 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2059 unsigned long target_size
= size
<< growth
;
2061 target_size
= min(target_size
, ht
->max_nr_buckets
);
2062 if (resize_target_grow(ht
, target_size
) >= target_size
)
2065 __cds_lfht_resize_lazy_launch(ht
);
2069 * We favor grow operations over shrink. A shrink operation never occurs
2070 * if a grow operation is queued for lazy execution. A grow operation
2071 * cancels any pending shrink lazy execution.
2074 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2075 unsigned long count
)
2077 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2079 count
= max(count
, MIN_TABLE_SIZE
);
2080 count
= min(count
, ht
->max_nr_buckets
);
2082 return; /* Already the right size, no resize needed */
2083 if (count
> size
) { /* lazy grow */
2084 if (resize_target_grow(ht
, count
) >= count
)
2086 } else { /* lazy shrink */
2090 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2092 break; /* no resize needed */
2094 return; /* growing is/(was just) in progress */
2096 return; /* some other thread do shrink */
2100 __cds_lfht_resize_lazy_launch(ht
);
2103 static void cds_lfht_before_fork(void *priv
)
2105 if (cds_lfht_workqueue_atfork_nesting
++)
2107 mutex_lock(&cds_lfht_fork_mutex
);
2108 if (!cds_lfht_workqueue
)
2110 urcu_workqueue_pause_worker(cds_lfht_workqueue
);
2113 static void cds_lfht_after_fork_parent(void *priv
)
2115 if (--cds_lfht_workqueue_atfork_nesting
)
2117 if (!cds_lfht_workqueue
)
2119 urcu_workqueue_resume_worker(cds_lfht_workqueue
);
2121 mutex_unlock(&cds_lfht_fork_mutex
);
2124 static void cds_lfht_after_fork_child(void *priv
)
2126 if (--cds_lfht_workqueue_atfork_nesting
)
2128 if (!cds_lfht_workqueue
)
2130 urcu_workqueue_create_worker(cds_lfht_workqueue
);
2132 mutex_unlock(&cds_lfht_fork_mutex
);
2135 static struct urcu_atfork cds_lfht_atfork
= {
2136 .before_fork
= cds_lfht_before_fork
,
2137 .after_fork_parent
= cds_lfht_after_fork_parent
,
2138 .after_fork_child
= cds_lfht_after_fork_child
,
2141 /* Block all signals to ensure we don't disturb the application. */
2142 static void cds_lfht_worker_init(struct urcu_workqueue
*workqueue
,
2148 /* Block signal for entire process, so only our thread processes it. */
2149 ret
= sigfillset(&mask
);
2152 ret
= pthread_sigmask(SIG_BLOCK
, &mask
, NULL
);
2157 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
)
2159 flavor
->register_rculfhash_atfork(&cds_lfht_atfork
);
2161 mutex_lock(&cds_lfht_fork_mutex
);
2162 if (cds_lfht_workqueue_user_count
++)
2164 cds_lfht_workqueue
= urcu_workqueue_create(0, -1, NULL
,
2165 NULL
, cds_lfht_worker_init
, NULL
, NULL
, NULL
, NULL
, NULL
);
2167 mutex_unlock(&cds_lfht_fork_mutex
);
2170 static void cds_lfht_fini_worker(const struct rcu_flavor_struct
*flavor
)
2172 mutex_lock(&cds_lfht_fork_mutex
);
2173 if (--cds_lfht_workqueue_user_count
)
2175 urcu_workqueue_destroy(cds_lfht_workqueue
);
2176 cds_lfht_workqueue
= NULL
;
2178 mutex_unlock(&cds_lfht_fork_mutex
);
2180 flavor
->unregister_rculfhash_atfork(&cds_lfht_atfork
);