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,
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 | |
270 #include <urcu-call-rcu.h>
271 #include <urcu/arch.h>
272 #include <urcu/uatomic.h>
273 #include <urcu/compiler.h>
277 #include "rculfhash.h"
278 #include "rculfhash-internal.h"
279 #include "urcu-flavor.h"
281 #include <common/common.h>
284 * We need to lock pthread exit, which deadlocks __nptl_setxid in the runas
285 * clone. This work-around will be allowed to be removed when runas.c gets
286 * changed to do an exec() before issuing seteuid/setegid. See
287 * http://sourceware.org/bugzilla/show_bug.cgi?id=10184 for details.
289 pthread_mutex_t lttng_libc_state_lock
= PTHREAD_MUTEX_INITIALIZER
;
292 * Split-counters lazily update the global counter each 1024
293 * addition/removal. It automatically keeps track of resize required.
294 * We use the bucket length as indicator for need to expand for small
295 * tables and machines lacking per-cpu data suppport.
297 #define COUNT_COMMIT_ORDER 10
298 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
299 #define CHAIN_LEN_TARGET 1
300 #define CHAIN_LEN_RESIZE_THRESHOLD 3
303 * Define the minimum table size.
305 #define MIN_TABLE_ORDER 0
306 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
309 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
311 #define MIN_PARTITION_PER_THREAD_ORDER 12
312 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
315 * The removed flag needs to be updated atomically with the pointer.
316 * It indicates that no node must attach to the node scheduled for
317 * removal, and that node garbage collection must be performed.
318 * The bucket flag does not require to be updated atomically with the
319 * pointer, but it is added as a pointer low bit flag to save space.
320 * The "removal owner" flag is used to detect which of the "del"
321 * operation that has set the "removed flag" gets to return the removed
322 * node to its caller. Note that the replace operation does not need to
323 * iteract with the "removal owner" flag, because it validates that
324 * the "removed" flag is not set before performing its cmpxchg.
326 #define REMOVED_FLAG (1UL << 0)
327 #define BUCKET_FLAG (1UL << 1)
328 #define REMOVAL_OWNER_FLAG (1UL << 2)
329 #define FLAGS_MASK ((1UL << 3) - 1)
331 /* Value of the end pointer. Should not interact with flags. */
332 #define END_VALUE NULL
335 * ht_items_count: Split-counters counting the number of node addition
336 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
337 * is set at hash table creation.
339 * These are free-running counters, never reset to zero. They count the
340 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
341 * operations to update the global counter. We choose a power-of-2 value
342 * for the trigger to deal with 32 or 64-bit overflow of the counter.
344 struct ht_items_count
{
345 unsigned long add
, del
;
346 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
349 * rcu_resize_work: Contains arguments passed to RCU worker thread
350 * responsible for performing lazy resize.
352 struct rcu_resize_work
{
353 struct rcu_head head
;
358 * partition_resize_work: Contains arguments passed to worker threads
359 * executing the hash table resize on partitions of the hash table
360 * assigned to each processor's worker thread.
362 struct partition_resize_work
{
365 unsigned long i
, start
, len
;
366 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
367 unsigned long start
, unsigned long len
);
371 * Algorithm to reverse bits in a word by lookup table, extended to
374 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
375 * Originally from Public Domain.
378 static const uint8_t BitReverseTable256
[256] =
380 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
381 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
382 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
383 R6(0), R6(2), R6(1), R6(3)
390 uint8_t bit_reverse_u8(uint8_t v
)
392 return BitReverseTable256
[v
];
395 static __attribute__((unused
))
396 uint32_t bit_reverse_u32(uint32_t v
)
398 return ((uint32_t) bit_reverse_u8(v
) << 24) |
399 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
400 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
401 ((uint32_t) bit_reverse_u8(v
>> 24));
404 static __attribute__((unused
))
405 uint64_t bit_reverse_u64(uint64_t v
)
407 return ((uint64_t) bit_reverse_u8(v
) << 56) |
408 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
409 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
410 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
411 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
412 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
413 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
414 ((uint64_t) bit_reverse_u8(v
>> 56));
418 unsigned long bit_reverse_ulong(unsigned long v
)
420 #if (CAA_BITS_PER_LONG == 32)
421 return bit_reverse_u32(v
);
423 return bit_reverse_u64(v
);
428 * fls: returns the position of the most significant bit.
429 * Returns 0 if no bit is set, else returns the position of the most
430 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
432 #if defined(__i386) || defined(__x86_64)
434 unsigned int fls_u32(uint32_t x
)
442 : "=r" (r
) : "rm" (x
));
448 #if defined(__x86_64)
450 unsigned int fls_u64(uint64_t x
)
458 : "=r" (r
) : "rm" (x
));
465 static __attribute__((unused
))
466 unsigned int fls_u64(uint64_t x
)
473 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
477 if (!(x
& 0xFFFF000000000000ULL
)) {
481 if (!(x
& 0xFF00000000000000ULL
)) {
485 if (!(x
& 0xF000000000000000ULL
)) {
489 if (!(x
& 0xC000000000000000ULL
)) {
493 if (!(x
& 0x8000000000000000ULL
)) {
502 static __attribute__((unused
))
503 unsigned int fls_u32(uint32_t x
)
509 if (!(x
& 0xFFFF0000U
)) {
513 if (!(x
& 0xFF000000U
)) {
517 if (!(x
& 0xF0000000U
)) {
521 if (!(x
& 0xC0000000U
)) {
525 if (!(x
& 0x80000000U
)) {
533 unsigned int cds_lfht_fls_ulong(unsigned long x
)
535 #if (CAA_BITS_PER_LONG == 32)
543 * Return the minimum order for which x <= (1UL << order).
544 * Return -1 if x is 0.
546 int cds_lfht_get_count_order_u32(uint32_t x
)
551 return fls_u32(x
- 1);
555 * Return the minimum order for which x <= (1UL << order).
556 * Return -1 if x is 0.
558 int cds_lfht_get_count_order_ulong(unsigned long x
)
563 return cds_lfht_fls_ulong(x
- 1);
567 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
570 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
571 unsigned long count
);
573 static long nr_cpus_mask
= -1;
574 static long split_count_mask
= -1;
575 static int split_count_order
= -1;
577 #if defined(HAVE_SYSCONF)
578 static void ht_init_nr_cpus_mask(void)
582 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
588 * round up number of CPUs to next power of two, so we
589 * can use & for modulo.
591 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
592 nr_cpus_mask
= maxcpus
- 1;
594 #else /* #if defined(HAVE_SYSCONF) */
595 static void ht_init_nr_cpus_mask(void)
599 #endif /* #else #if defined(HAVE_SYSCONF) */
602 void alloc_split_items_count(struct cds_lfht
*ht
)
604 struct ht_items_count
*count
;
606 if (nr_cpus_mask
== -1) {
607 ht_init_nr_cpus_mask();
608 if (nr_cpus_mask
< 0)
609 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
611 split_count_mask
= nr_cpus_mask
;
613 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
616 assert(split_count_mask
>= 0);
618 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
619 ht
->split_count
= calloc(split_count_mask
+ 1, sizeof(*count
));
620 assert(ht
->split_count
);
622 ht
->split_count
= NULL
;
627 void free_split_items_count(struct cds_lfht
*ht
)
629 poison_free(ht
->split_count
);
632 #if defined(HAVE_SCHED_GETCPU) && !defined(VALGRIND)
634 int ht_get_split_count_index(unsigned long hash
)
638 assert(split_count_mask
>= 0);
639 cpu
= sched_getcpu();
640 if (caa_unlikely(cpu
< 0))
641 return hash
& split_count_mask
;
643 return cpu
& split_count_mask
;
645 #else /* #if defined(HAVE_SCHED_GETCPU) */
647 int ht_get_split_count_index(unsigned long hash
)
649 return hash
& split_count_mask
;
651 #endif /* #else #if defined(HAVE_SCHED_GETCPU) */
654 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
656 unsigned long split_count
;
660 if (caa_unlikely(!ht
->split_count
))
662 index
= ht_get_split_count_index(hash
);
663 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
664 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
666 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
668 dbg_printf("add split count %lu\n", split_count
);
669 count
= uatomic_add_return(&ht
->count
,
670 1UL << COUNT_COMMIT_ORDER
);
671 if (caa_likely(count
& (count
- 1)))
673 /* Only if global count is power of 2 */
675 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
677 dbg_printf("add set global %ld\n", count
);
678 cds_lfht_resize_lazy_count(ht
, size
,
679 count
>> (CHAIN_LEN_TARGET
- 1));
683 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
685 unsigned long split_count
;
689 if (caa_unlikely(!ht
->split_count
))
691 index
= ht_get_split_count_index(hash
);
692 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
693 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
695 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
697 dbg_printf("del split count %lu\n", split_count
);
698 count
= uatomic_add_return(&ht
->count
,
699 -(1UL << COUNT_COMMIT_ORDER
));
700 if (caa_likely(count
& (count
- 1)))
702 /* Only if global count is power of 2 */
704 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
706 dbg_printf("del set global %ld\n", count
);
708 * Don't shrink table if the number of nodes is below a
711 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
713 cds_lfht_resize_lazy_count(ht
, size
,
714 count
>> (CHAIN_LEN_TARGET
- 1));
718 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
722 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
724 count
= uatomic_read(&ht
->count
);
726 * Use bucket-local length for small table expand and for
727 * environments lacking per-cpu data support.
729 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
732 dbg_printf("WARNING: large chain length: %u.\n",
734 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
738 * Ideal growth calculated based on chain length.
740 growth
= cds_lfht_get_count_order_u32(chain_len
741 - (CHAIN_LEN_TARGET
- 1));
742 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
744 >= (1UL << (COUNT_COMMIT_ORDER
745 + split_count_order
))) {
747 * If ideal growth expands the hash table size
748 * beyond the "small hash table" sizes, use the
749 * maximum small hash table size to attempt
750 * expanding the hash table. This only applies
751 * when node accounting is available, otherwise
752 * the chain length is used to expand the hash
753 * table in every case.
755 growth
= COUNT_COMMIT_ORDER
+ split_count_order
756 - cds_lfht_get_count_order_ulong(size
);
760 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
765 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
767 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
771 int is_removed(struct cds_lfht_node
*node
)
773 return ((unsigned long) node
) & REMOVED_FLAG
;
777 int is_bucket(struct cds_lfht_node
*node
)
779 return ((unsigned long) node
) & BUCKET_FLAG
;
783 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
785 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
789 int is_removal_owner(struct cds_lfht_node
*node
)
791 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
795 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
797 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
801 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
803 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
807 struct cds_lfht_node
*get_end(void)
809 return (struct cds_lfht_node
*) END_VALUE
;
813 int is_end(struct cds_lfht_node
*node
)
815 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
819 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
822 unsigned long old1
, old2
;
824 old1
= uatomic_read(ptr
);
829 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
834 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
836 return ht
->mm
->alloc_bucket_table(ht
, order
);
840 * cds_lfht_free_bucket_table() should be called with decreasing order.
841 * When cds_lfht_free_bucket_table(0) is called, it means the whole
845 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
847 return ht
->mm
->free_bucket_table(ht
, order
);
851 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
853 return ht
->bucket_at(ht
, index
);
857 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
861 return bucket_at(ht
, hash
& (size
- 1));
865 * Remove all logically deleted nodes from a bucket up to a certain node key.
868 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
870 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
872 assert(!is_bucket(bucket
));
873 assert(!is_removed(bucket
));
874 assert(!is_bucket(node
));
875 assert(!is_removed(node
));
878 /* We can always skip the bucket node initially */
879 iter
= rcu_dereference(iter_prev
->next
);
880 assert(!is_removed(iter
));
881 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
883 * We should never be called with bucket (start of chain)
884 * and logically removed node (end of path compression
885 * marker) being the actual same node. This would be a
886 * bug in the algorithm implementation.
888 assert(bucket
!= node
);
890 if (caa_unlikely(is_end(iter
)))
892 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
894 next
= rcu_dereference(clear_flag(iter
)->next
);
895 if (caa_likely(is_removed(next
)))
897 iter_prev
= clear_flag(iter
);
900 assert(!is_removed(iter
));
902 new_next
= flag_bucket(clear_flag(next
));
904 new_next
= clear_flag(next
);
905 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
910 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
911 struct cds_lfht_node
*old_node
,
912 struct cds_lfht_node
*old_next
,
913 struct cds_lfht_node
*new_node
)
915 struct cds_lfht_node
*bucket
, *ret_next
;
917 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
920 assert(!is_removed(old_node
));
921 assert(!is_bucket(old_node
));
922 assert(!is_removed(new_node
));
923 assert(!is_bucket(new_node
));
924 assert(new_node
!= old_node
);
926 /* Insert after node to be replaced */
927 if (is_removed(old_next
)) {
929 * Too late, the old node has been removed under us
930 * between lookup and replace. Fail.
934 assert(old_next
== clear_flag(old_next
));
935 assert(new_node
!= old_next
);
937 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
938 * flag. It is either set atomically at the same time
939 * (replace) or after (del).
941 assert(!is_removal_owner(old_next
));
942 new_node
->next
= old_next
;
944 * Here is the whole trick for lock-free replace: we add
945 * the replacement node _after_ the node we want to
946 * replace by atomically setting its next pointer at the
947 * same time we set its removal flag. Given that
948 * the lookups/get next use an iterator aware of the
949 * next pointer, they will either skip the old node due
950 * to the removal flag and see the new node, or use
951 * the old node, but will not see the new one.
952 * This is a replacement of a node with another node
953 * that has the same value: we are therefore not
954 * removing a value from the hash table. We set both the
955 * REMOVED and REMOVAL_OWNER flags atomically so we own
956 * the node after successful cmpxchg.
958 ret_next
= uatomic_cmpxchg(&old_node
->next
,
959 old_next
, flag_removed_or_removal_owner(new_node
));
960 if (ret_next
== old_next
)
961 break; /* We performed the replacement. */
966 * Ensure that the old node is not visible to readers anymore:
967 * lookup for the node, and remove it (along with any other
968 * logically removed node) if found.
970 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
971 _cds_lfht_gc_bucket(bucket
, new_node
);
973 assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
978 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
979 * mode. A NULL unique_ret allows creation of duplicate keys.
982 void _cds_lfht_add(struct cds_lfht
*ht
,
984 cds_lfht_match_fct match
,
987 struct cds_lfht_node
*node
,
988 struct cds_lfht_iter
*unique_ret
,
991 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
993 struct cds_lfht_node
*bucket
;
995 assert(!is_bucket(node
));
996 assert(!is_removed(node
));
997 bucket
= lookup_bucket(ht
, size
, hash
);
999 uint32_t chain_len
= 0;
1002 * iter_prev points to the non-removed node prior to the
1006 /* We can always skip the bucket node initially */
1007 iter
= rcu_dereference(iter_prev
->next
);
1008 assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1010 if (caa_unlikely(is_end(iter
)))
1012 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1015 /* bucket node is the first node of the identical-hash-value chain */
1016 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1019 next
= rcu_dereference(clear_flag(iter
)->next
);
1020 if (caa_unlikely(is_removed(next
)))
1026 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1027 struct cds_lfht_iter d_iter
= { .node
= node
, .next
= iter
, };
1030 * uniquely adding inserts the node as the first
1031 * node of the identical-hash-value node chain.
1033 * This semantic ensures no duplicated keys
1034 * should ever be observable in the table
1035 * (including traversing the table node by
1036 * node by forward iterations)
1038 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1042 *unique_ret
= d_iter
;
1046 /* Only account for identical reverse hash once */
1047 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1048 && !is_bucket(next
))
1049 check_resize(ht
, size
, ++chain_len
);
1050 iter_prev
= clear_flag(iter
);
1055 assert(node
!= clear_flag(iter
));
1056 assert(!is_removed(iter_prev
));
1057 assert(!is_removed(iter
));
1058 assert(iter_prev
!= node
);
1060 node
->next
= clear_flag(iter
);
1062 node
->next
= flag_bucket(clear_flag(iter
));
1063 if (is_bucket(iter
))
1064 new_node
= flag_bucket(node
);
1067 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1068 new_node
) != iter
) {
1069 continue; /* retry */
1076 assert(!is_removed(iter
));
1077 if (is_bucket(iter
))
1078 new_next
= flag_bucket(clear_flag(next
));
1080 new_next
= clear_flag(next
);
1081 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1086 unique_ret
->node
= return_node
;
1087 /* unique_ret->next left unset, never used. */
1092 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1093 struct cds_lfht_node
*node
)
1095 struct cds_lfht_node
*bucket
, *next
;
1097 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1100 /* logically delete the node */
1101 assert(!is_bucket(node
));
1102 assert(!is_removed(node
));
1103 assert(!is_removal_owner(node
));
1106 * We are first checking if the node had previously been
1107 * logically removed (this check is not atomic with setting the
1108 * logical removal flag). Return -ENOENT if the node had
1109 * previously been removed.
1111 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1112 if (caa_unlikely(is_removed(next
)))
1114 assert(!is_bucket(next
));
1116 * The del operation semantic guarantees a full memory barrier
1117 * before the uatomic_or atomic commit of the deletion flag.
1119 cmm_smp_mb__before_uatomic_or();
1121 * We set the REMOVED_FLAG unconditionally. Note that there may
1122 * be more than one concurrent thread setting this flag.
1123 * Knowing which wins the race will be known after the garbage
1124 * collection phase, stay tuned!
1126 uatomic_or(&node
->next
, REMOVED_FLAG
);
1127 /* We performed the (logical) deletion. */
1130 * Ensure that the node is not visible to readers anymore: lookup for
1131 * the node, and remove it (along with any other logically removed node)
1134 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1135 _cds_lfht_gc_bucket(bucket
, node
);
1137 assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1139 * Last phase: atomically exchange node->next with a version
1140 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1141 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1142 * the node and win the removal race.
1143 * It is interesting to note that all "add" paths are forbidden
1144 * to change the next pointer starting from the point where the
1145 * REMOVED_FLAG is set, so here using a read, followed by a
1146 * xchg() suffice to guarantee that the xchg() will ever only
1147 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1150 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1151 flag_removal_owner(node
->next
))))
1158 void *partition_resize_thread(void *arg
)
1160 struct partition_resize_work
*work
= arg
;
1162 work
->ht
->flavor
->register_thread();
1163 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1164 work
->ht
->flavor
->unregister_thread();
1169 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1171 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1172 unsigned long start
, unsigned long len
))
1174 unsigned long partition_len
;
1175 struct partition_resize_work
*work
;
1177 unsigned long nr_threads
;
1180 * Note: nr_cpus_mask + 1 is always power of 2.
1181 * We spawn just the number of threads we need to satisfy the minimum
1182 * partition size, up to the number of CPUs in the system.
1184 if (nr_cpus_mask
> 0) {
1185 nr_threads
= min(nr_cpus_mask
+ 1,
1186 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1190 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1191 work
= calloc(nr_threads
, sizeof(*work
));
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
]);
1203 for (thread
= 0; thread
< nr_threads
; thread
++) {
1204 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1211 * Holding RCU read lock to protect _cds_lfht_add against memory
1212 * reclaim that could be performed by other call_rcu worker threads (ABA
1215 * When we reach a certain length, we can split this population phase over
1216 * many worker threads, based on the number of CPUs available in the system.
1217 * This should therefore take care of not having the expand lagging behind too
1218 * many concurrent insertion threads by using the scheduler's ability to
1219 * schedule bucket node population fairly with insertions.
1222 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1223 unsigned long start
, unsigned long len
)
1225 unsigned long j
, size
= 1UL << (i
- 1);
1227 assert(i
> MIN_TABLE_ORDER
);
1228 ht
->flavor
->read_lock();
1229 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1230 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1232 assert(j
>= size
&& j
< (size
<< 1));
1233 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1235 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1236 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1238 ht
->flavor
->read_unlock();
1242 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1245 assert(nr_cpus_mask
!= -1);
1246 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1247 ht
->flavor
->thread_online();
1248 init_table_populate_partition(ht
, i
, 0, len
);
1249 ht
->flavor
->thread_offline();
1252 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1256 void init_table(struct cds_lfht
*ht
,
1257 unsigned long first_order
, unsigned long last_order
)
1261 dbg_printf("init table: first_order %lu last_order %lu\n",
1262 first_order
, last_order
);
1263 assert(first_order
> MIN_TABLE_ORDER
);
1264 for (i
= first_order
; i
<= last_order
; i
++) {
1267 len
= 1UL << (i
- 1);
1268 dbg_printf("init order %lu len: %lu\n", i
, len
);
1270 /* Stop expand if the resize target changes under us */
1271 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1274 cds_lfht_alloc_bucket_table(ht
, i
);
1277 * Set all bucket nodes reverse hash values for a level and
1278 * link all bucket nodes into the table.
1280 init_table_populate(ht
, i
, len
);
1283 * Update table size.
1285 cmm_smp_wmb(); /* populate data before RCU size */
1286 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1288 dbg_printf("init new size: %lu\n", 1UL << i
);
1289 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1295 * Holding RCU read lock to protect _cds_lfht_remove against memory
1296 * reclaim that could be performed by other call_rcu worker threads (ABA
1298 * For a single level, we logically remove and garbage collect each node.
1300 * As a design choice, we perform logical removal and garbage collection on a
1301 * node-per-node basis to simplify this algorithm. We also assume keeping good
1302 * cache locality of the operation would overweight possible performance gain
1303 * that could be achieved by batching garbage collection for multiple levels.
1304 * However, this would have to be justified by benchmarks.
1306 * Concurrent removal and add operations are helping us perform garbage
1307 * collection of logically removed nodes. We guarantee that all logically
1308 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1309 * invoked to free a hole level of bucket nodes (after a grace period).
1311 * Logical removal and garbage collection can therefore be done in batch
1312 * or on a node-per-node basis, as long as the guarantee above holds.
1314 * When we reach a certain length, we can split this removal over many worker
1315 * threads, based on the number of CPUs available in the system. This should
1316 * take care of not letting resize process lag behind too many concurrent
1317 * updater threads actively inserting into the hash table.
1320 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1321 unsigned long start
, unsigned long len
)
1323 unsigned long j
, size
= 1UL << (i
- 1);
1325 assert(i
> MIN_TABLE_ORDER
);
1326 ht
->flavor
->read_lock();
1327 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1328 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1329 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1331 assert(j
>= size
&& j
< (size
<< 1));
1332 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1334 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1335 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1336 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1338 ht
->flavor
->read_unlock();
1342 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1345 assert(nr_cpus_mask
!= -1);
1346 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
) {
1347 ht
->flavor
->thread_online();
1348 remove_table_partition(ht
, i
, 0, len
);
1349 ht
->flavor
->thread_offline();
1352 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1356 * fini_table() is never called for first_order == 0, which is why
1357 * free_by_rcu_order == 0 can be used as criterion to know if free must
1361 void fini_table(struct cds_lfht
*ht
,
1362 unsigned long first_order
, unsigned long last_order
)
1365 unsigned long free_by_rcu_order
= 0;
1367 dbg_printf("fini table: first_order %lu last_order %lu\n",
1368 first_order
, last_order
);
1369 assert(first_order
> MIN_TABLE_ORDER
);
1370 for (i
= last_order
; i
>= first_order
; i
--) {
1373 len
= 1UL << (i
- 1);
1374 dbg_printf("fini order %lu len: %lu\n", i
, len
);
1376 /* Stop shrink if the resize target changes under us */
1377 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1380 cmm_smp_wmb(); /* populate data before RCU size */
1381 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1384 * We need to wait for all add operations to reach Q.S. (and
1385 * thus use the new table for lookups) before we can start
1386 * releasing the old bucket nodes. Otherwise their lookup will
1387 * return a logically removed node as insert position.
1389 ht
->flavor
->update_synchronize_rcu();
1390 if (free_by_rcu_order
)
1391 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1394 * Set "removed" flag in bucket nodes about to be removed.
1395 * Unlink all now-logically-removed bucket node pointers.
1396 * Concurrent add/remove operation are helping us doing
1399 remove_table(ht
, i
, len
);
1401 free_by_rcu_order
= i
;
1403 dbg_printf("fini new size: %lu\n", 1UL << i
);
1404 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1408 if (free_by_rcu_order
) {
1409 ht
->flavor
->update_synchronize_rcu();
1410 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1415 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1417 struct cds_lfht_node
*prev
, *node
;
1418 unsigned long order
, len
, i
;
1420 cds_lfht_alloc_bucket_table(ht
, 0);
1422 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1423 node
= bucket_at(ht
, 0);
1424 node
->next
= flag_bucket(get_end());
1425 node
->reverse_hash
= 0;
1427 for (order
= 1; order
< cds_lfht_get_count_order_ulong(size
) + 1; order
++) {
1428 len
= 1UL << (order
- 1);
1429 cds_lfht_alloc_bucket_table(ht
, order
);
1431 for (i
= 0; i
< len
; i
++) {
1433 * Now, we are trying to init the node with the
1434 * hash=(len+i) (which is also a bucket with the
1435 * index=(len+i)) and insert it into the hash table,
1436 * so this node has to be inserted after the bucket
1437 * with the index=(len+i)&(len-1)=i. And because there
1438 * is no other non-bucket node nor bucket node with
1439 * larger index/hash inserted, so the bucket node
1440 * being inserted should be inserted directly linked
1441 * after the bucket node with index=i.
1443 prev
= bucket_at(ht
, i
);
1444 node
= bucket_at(ht
, len
+ i
);
1446 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1447 order
, len
+ i
, len
+ i
);
1448 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1450 /* insert after prev */
1451 assert(is_bucket(prev
->next
));
1452 node
->next
= prev
->next
;
1453 prev
->next
= flag_bucket(node
);
1458 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1459 unsigned long min_nr_alloc_buckets
,
1460 unsigned long max_nr_buckets
,
1462 const struct cds_lfht_mm_type
*mm
,
1463 const struct rcu_flavor_struct
*flavor
,
1464 pthread_attr_t
*attr
)
1466 struct cds_lfht
*ht
;
1467 unsigned long order
;
1469 /* min_nr_alloc_buckets must be power of two */
1470 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1473 /* init_size must be power of two */
1474 if (!init_size
|| (init_size
& (init_size
- 1)))
1478 * Memory management plugin default.
1481 if (CAA_BITS_PER_LONG
> 32
1483 && max_nr_buckets
<= (1ULL << 32)) {
1485 * For 64-bit architectures, with max number of
1486 * buckets small enough not to use the entire
1487 * 64-bit memory mapping space (and allowing a
1488 * fair number of hash table instances), use the
1489 * mmap allocator, which is faster than the
1492 mm
= &cds_lfht_mm_mmap
;
1495 * The fallback is to use the order allocator.
1497 mm
= &cds_lfht_mm_order
;
1501 /* max_nr_buckets == 0 for order based mm means infinite */
1502 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1503 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1505 /* max_nr_buckets must be power of two */
1506 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1509 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1510 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1511 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1512 init_size
= min(init_size
, max_nr_buckets
);
1514 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1516 assert(ht
->mm
== mm
);
1517 assert(ht
->bucket_at
== mm
->bucket_at
);
1520 ht
->flavor
= flavor
;
1521 ht
->resize_attr
= attr
;
1522 alloc_split_items_count(ht
);
1523 /* this mutex should not nest in read-side C.S. */
1524 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1525 order
= cds_lfht_get_count_order_ulong(init_size
);
1526 ht
->resize_target
= 1UL << order
;
1527 cds_lfht_create_bucket(ht
, 1UL << order
);
1528 ht
->size
= 1UL << order
;
1532 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1533 cds_lfht_match_fct match
, const void *key
,
1534 struct cds_lfht_iter
*iter
)
1536 struct cds_lfht_node
*node
, *next
, *bucket
;
1537 unsigned long reverse_hash
, size
;
1539 reverse_hash
= bit_reverse_ulong(hash
);
1541 size
= rcu_dereference(ht
->size
);
1542 bucket
= lookup_bucket(ht
, size
, hash
);
1543 /* We can always skip the bucket node initially */
1544 node
= rcu_dereference(bucket
->next
);
1545 node
= clear_flag(node
);
1547 if (caa_unlikely(is_end(node
))) {
1551 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1555 next
= rcu_dereference(node
->next
);
1556 assert(node
== clear_flag(node
));
1557 if (caa_likely(!is_removed(next
))
1559 && node
->reverse_hash
== reverse_hash
1560 && caa_likely(match(node
, key
))) {
1563 node
= clear_flag(next
);
1565 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1570 void cds_lfht_next_duplicate(struct cds_lfht
*ht
, cds_lfht_match_fct match
,
1571 const void *key
, struct cds_lfht_iter
*iter
)
1573 struct cds_lfht_node
*node
, *next
;
1574 unsigned long reverse_hash
;
1577 reverse_hash
= node
->reverse_hash
;
1579 node
= clear_flag(next
);
1582 if (caa_unlikely(is_end(node
))) {
1586 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1590 next
= rcu_dereference(node
->next
);
1591 if (caa_likely(!is_removed(next
))
1593 && caa_likely(match(node
, key
))) {
1596 node
= clear_flag(next
);
1598 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1603 void cds_lfht_next(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1605 struct cds_lfht_node
*node
, *next
;
1607 node
= clear_flag(iter
->next
);
1609 if (caa_unlikely(is_end(node
))) {
1613 next
= rcu_dereference(node
->next
);
1614 if (caa_likely(!is_removed(next
))
1615 && !is_bucket(next
)) {
1618 node
= clear_flag(next
);
1620 assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1625 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1628 * Get next after first bucket node. The first bucket node is the
1629 * first node of the linked list.
1631 iter
->next
= bucket_at(ht
, 0)->next
;
1632 cds_lfht_next(ht
, iter
);
1635 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1636 struct cds_lfht_node
*node
)
1640 node
->reverse_hash
= bit_reverse_ulong(hash
);
1641 size
= rcu_dereference(ht
->size
);
1642 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1643 ht_count_add(ht
, size
, hash
);
1646 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1648 cds_lfht_match_fct match
,
1650 struct cds_lfht_node
*node
)
1653 struct cds_lfht_iter iter
;
1655 node
->reverse_hash
= bit_reverse_ulong(hash
);
1656 size
= rcu_dereference(ht
->size
);
1657 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1658 if (iter
.node
== node
)
1659 ht_count_add(ht
, size
, hash
);
1663 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1665 cds_lfht_match_fct match
,
1667 struct cds_lfht_node
*node
)
1670 struct cds_lfht_iter iter
;
1672 node
->reverse_hash
= bit_reverse_ulong(hash
);
1673 size
= rcu_dereference(ht
->size
);
1675 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1676 if (iter
.node
== node
) {
1677 ht_count_add(ht
, size
, hash
);
1681 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1686 int cds_lfht_replace(struct cds_lfht
*ht
,
1687 struct cds_lfht_iter
*old_iter
,
1689 cds_lfht_match_fct match
,
1691 struct cds_lfht_node
*new_node
)
1695 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1696 if (!old_iter
->node
)
1698 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1700 if (caa_unlikely(!match(old_iter
->node
, key
)))
1702 size
= rcu_dereference(ht
->size
);
1703 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1707 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1709 unsigned long size
, hash
;
1712 size
= rcu_dereference(ht
->size
);
1713 ret
= _cds_lfht_del(ht
, size
, node
);
1715 hash
= bit_reverse_ulong(node
->reverse_hash
);
1716 ht_count_del(ht
, size
, hash
);
1721 int cds_lfht_is_node_deleted(struct cds_lfht_node
*node
)
1723 return is_removed(CMM_LOAD_SHARED(node
->next
));
1727 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1729 struct cds_lfht_node
*node
;
1730 unsigned long order
, i
, size
;
1732 /* Check that the table is empty */
1733 node
= bucket_at(ht
, 0);
1735 node
= clear_flag(node
)->next
;
1736 if (!is_bucket(node
))
1738 assert(!is_removed(node
));
1739 } while (!is_end(node
));
1741 * size accessed without rcu_dereference because hash table is
1745 /* Internal sanity check: all nodes left should be buckets */
1746 for (i
= 0; i
< size
; i
++) {
1747 node
= bucket_at(ht
, i
);
1748 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1749 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1750 assert(is_bucket(node
->next
));
1753 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1754 cds_lfht_free_bucket_table(ht
, order
);
1760 * Should only be called when no more concurrent readers nor writers can
1761 * possibly access the table.
1763 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1766 #ifdef rcu_read_ongoing_mb
1770 /* Wait for in-flight resize operations to complete */
1771 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1772 cmm_smp_mb(); /* Store destroy before load resize */
1773 #ifdef rcu_read_ongoing_mb
1774 was_online
= ht
->flavor
->read_ongoing();
1776 ht
->flavor
->thread_offline();
1777 /* Calling with RCU read-side held is an error. */
1778 if (ht
->flavor
->read_ongoing()) {
1781 ht
->flavor
->thread_online();
1785 while (uatomic_read(&ht
->in_progress_resize
))
1786 (void) poll(NULL
, 0, 100); /* wait for 100ms */
1787 ret
= cds_lfht_delete_bucket(ht
);
1790 free_split_items_count(ht
);
1792 *attr
= ht
->resize_attr
;
1794 #ifdef rcu_read_ongoing_mb
1800 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1801 long *approx_before
,
1802 unsigned long *count
,
1805 struct cds_lfht_node
*node
, *next
;
1806 unsigned long nr_bucket
= 0, nr_removed
= 0;
1809 if (ht
->split_count
) {
1812 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1813 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1814 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1820 /* Count non-bucket nodes in the table */
1821 node
= bucket_at(ht
, 0);
1823 next
= rcu_dereference(node
->next
);
1824 if (is_removed(next
)) {
1825 if (!is_bucket(next
))
1829 } else if (!is_bucket(next
))
1833 node
= clear_flag(next
);
1834 } while (!is_end(node
));
1835 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
1836 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
1838 if (ht
->split_count
) {
1841 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1842 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
1843 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
1848 /* called with resize mutex held */
1850 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
1851 unsigned long old_size
, unsigned long new_size
)
1853 unsigned long old_order
, new_order
;
1855 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1856 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1857 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1858 old_size
, old_order
, new_size
, new_order
);
1859 assert(new_size
> old_size
);
1860 init_table(ht
, old_order
+ 1, new_order
);
1863 /* called with resize mutex held */
1865 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
1866 unsigned long old_size
, unsigned long new_size
)
1868 unsigned long old_order
, new_order
;
1870 new_size
= max(new_size
, MIN_TABLE_SIZE
);
1871 old_order
= cds_lfht_get_count_order_ulong(old_size
);
1872 new_order
= cds_lfht_get_count_order_ulong(new_size
);
1873 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1874 old_size
, old_order
, new_size
, new_order
);
1875 assert(new_size
< old_size
);
1877 /* Remove and unlink all bucket nodes to remove. */
1878 fini_table(ht
, new_order
+ 1, old_order
);
1882 /* called with resize mutex held */
1884 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
1886 unsigned long new_size
, old_size
;
1889 * Resize table, re-do if the target size has changed under us.
1892 assert(uatomic_read(&ht
->in_progress_resize
));
1893 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1895 ht
->resize_initiated
= 1;
1896 old_size
= ht
->size
;
1897 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
1898 if (old_size
< new_size
)
1899 _do_cds_lfht_grow(ht
, old_size
, new_size
);
1900 else if (old_size
> new_size
)
1901 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
1902 ht
->resize_initiated
= 0;
1903 /* write resize_initiated before read resize_target */
1905 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
1909 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
1911 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
1915 void resize_target_update_count(struct cds_lfht
*ht
,
1916 unsigned long count
)
1918 count
= max(count
, MIN_TABLE_SIZE
);
1919 count
= min(count
, ht
->max_nr_buckets
);
1920 uatomic_set(&ht
->resize_target
, count
);
1923 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
1925 #ifdef rcu_read_ongoing_mb
1928 was_online
= ht
->flavor
->read_ongoing();
1930 ht
->flavor
->thread_offline();
1931 /* Calling with RCU read-side held is an error. */
1932 if (ht
->flavor
->read_ongoing()) {
1933 static int print_once
;
1935 if (!CMM_LOAD_SHARED(print_once
))
1936 fprintf(stderr
, "[error] rculfhash: cds_lfht_resize "
1937 "called with RCU read-side lock held.\n");
1938 CMM_STORE_SHARED(print_once
, 1);
1943 resize_target_update_count(ht
, new_size
);
1944 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1945 pthread_mutex_lock(&ht
->resize_mutex
);
1946 _do_cds_lfht_resize(ht
);
1947 pthread_mutex_unlock(&ht
->resize_mutex
);
1948 #ifdef rcu_read_ongoing_mb
1951 ht
->flavor
->thread_online();
1956 void do_resize_cb(struct rcu_head
*head
)
1958 struct rcu_resize_work
*work
=
1959 caa_container_of(head
, struct rcu_resize_work
, head
);
1960 struct cds_lfht
*ht
= work
->ht
;
1962 ht
->flavor
->thread_offline();
1963 pthread_mutex_lock(&ht
->resize_mutex
);
1964 _do_cds_lfht_resize(ht
);
1965 pthread_mutex_unlock(&ht
->resize_mutex
);
1966 ht
->flavor
->thread_online();
1968 cmm_smp_mb(); /* finish resize before decrement */
1969 uatomic_dec(&ht
->in_progress_resize
);
1973 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
1975 struct rcu_resize_work
*work
;
1977 /* Store resize_target before read resize_initiated */
1979 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
1980 uatomic_inc(&ht
->in_progress_resize
);
1981 cmm_smp_mb(); /* increment resize count before load destroy */
1982 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
1983 uatomic_dec(&ht
->in_progress_resize
);
1986 work
= zmalloc(sizeof(*work
));
1988 dbg_printf("error allocating resize work, bailing out\n");
1989 uatomic_dec(&ht
->in_progress_resize
);
1993 ht
->flavor
->update_call_rcu(&work
->head
, do_resize_cb
);
1994 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
1999 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2001 unsigned long target_size
= size
<< growth
;
2003 target_size
= min(target_size
, ht
->max_nr_buckets
);
2004 if (resize_target_grow(ht
, target_size
) >= target_size
)
2007 __cds_lfht_resize_lazy_launch(ht
);
2011 * We favor grow operations over shrink. A shrink operation never occurs
2012 * if a grow operation is queued for lazy execution. A grow operation
2013 * cancels any pending shrink lazy execution.
2016 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2017 unsigned long count
)
2019 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2021 count
= max(count
, MIN_TABLE_SIZE
);
2022 count
= min(count
, ht
->max_nr_buckets
);
2024 return; /* Already the right size, no resize needed */
2025 if (count
> size
) { /* lazy grow */
2026 if (resize_target_grow(ht
, count
) >= count
)
2028 } else { /* lazy shrink */
2032 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2034 break; /* no resize needed */
2036 return; /* growing is/(was just) in progress */
2038 return; /* some other thread do shrink */
2042 __cds_lfht_resize_lazy_launch(ht
);