1 // SPDX-FileCopyrightText: 2010-2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
2 // SPDX-FileCopyrightText: 2011 Lai Jiangshan <laijs@cn.fujitsu.com>
4 // SPDX-License-Identifier: LGPL-2.1-or-later
7 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
11 * Based on the following articles:
12 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
13 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
14 * - Michael, M. M. High performance dynamic lock-free hash tables
15 * and list-based sets. In Proceedings of the fourteenth annual ACM
16 * symposium on Parallel algorithms and architectures, ACM Press,
19 * Some specificities of this Lock-Free Resizable RCU Hash Table
22 * - RCU read-side critical section allows readers to perform hash
23 * table lookups, as well as traversals, and use the returned objects
24 * safely by allowing memory reclaim to take place only after a grace
26 * - Add and remove operations are lock-free, and do not need to
27 * allocate memory. They need to be executed within RCU read-side
28 * critical section to ensure the objects they read are valid and to
29 * deal with the cmpxchg ABA problem.
30 * - add and add_unique operations are supported. add_unique checks if
31 * the node key already exists in the hash table. It ensures not to
32 * populate a duplicate key if the node key already exists in the hash
34 * - The resize operation executes concurrently with
35 * add/add_unique/add_replace/remove/lookup/traversal.
36 * - Hash table nodes are contained within a split-ordered list. This
37 * list is ordered by incrementing reversed-bits-hash value.
38 * - An index of bucket nodes is kept. These bucket nodes are the hash
39 * table "buckets". These buckets are internal nodes that allow to
40 * perform a fast hash lookup, similarly to a skip list. These
41 * buckets are chained together in the split-ordered list, which
42 * allows recursive expansion by inserting new buckets between the
43 * existing buckets. The split-ordered list allows adding new buckets
44 * between existing buckets as the table needs to grow.
45 * - The resize operation for small tables only allows expanding the
46 * hash table. It is triggered automatically by detecting long chains
47 * in the add operation.
48 * - The resize operation for larger tables (and available through an
49 * API) allows both expanding and shrinking the hash table.
50 * - Split-counters are used to keep track of the number of
51 * nodes within the hash table for automatic resize triggering.
52 * - Resize operation initiated by long chain detection is executed by a
53 * worker thread, which keeps lock-freedom of add and remove.
54 * - Resize operations are protected by a mutex.
55 * - The removal operation is split in two parts: first, a "removed"
56 * flag is set in the next pointer within the node to remove. Then,
57 * a "garbage collection" is performed in the bucket containing the
58 * removed node (from the start of the bucket up to the removed node).
59 * All encountered nodes with "removed" flag set in their next
60 * pointers are removed from the linked-list. If the cmpxchg used for
61 * removal fails (due to concurrent garbage-collection or concurrent
62 * add), we retry from the beginning of the bucket. This ensures that
63 * the node with "removed" flag set is removed from the hash table
64 * (not visible to lookups anymore) before the RCU read-side critical
65 * section held across removal ends. Furthermore, this ensures that
66 * the node with "removed" flag set is removed from the linked-list
67 * before its memory is reclaimed. After setting the "removal" flag,
68 * only the thread which removal is the first to set the "removal
69 * owner" flag (with an xchg) into a node's next pointer is considered
70 * to have succeeded its removal (and thus owns the node to reclaim).
71 * Because we garbage-collect starting from an invariant node (the
72 * start-of-bucket bucket node) up to the "removed" node (or find a
73 * reverse-hash that is higher), we are sure that a successful
74 * traversal of the chain leads to a chain that is present in the
75 * linked-list (the start node is never removed) and that it does not
76 * contain the "removed" node anymore, even if concurrent delete/add
77 * operations are changing the structure of the list concurrently.
78 * - The add operations perform garbage collection of buckets if they
79 * encounter nodes with removed flag set in the bucket where they want
80 * to add their new node. This ensures lock-freedom of add operation by
81 * helping the remover unlink nodes from the list rather than to wait
83 * - There are three memory backends for the hash table buckets: the
84 * "order table", the "chunks", and the "mmap".
85 * - These bucket containers contain a compact version of the hash table
87 * - The RCU "order table":
88 * - has a first level table indexed by log2(hash index) which is
89 * copied and expanded by the resize operation. This order table
90 * allows finding the "bucket node" tables.
91 * - There is one bucket node table per hash index order. The size of
92 * each bucket node table is half the number of hashes contained in
93 * this order (except for order 0).
94 * - The RCU "chunks" is best suited for close interaction with a page
95 * allocator. It uses a linear array as index to "chunks" containing
96 * each the same number of buckets.
97 * - The RCU "mmap" memory backend uses a single memory map to hold
99 * - synchronize_rcu is used to garbage-collect the old bucket node table.
101 * Ordering Guarantees:
103 * To discuss these guarantees, we first define "read" operation as any
104 * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
105 * cds_lfht_first, cds_lfht_next operation, as well as
106 * cds_lfht_add_unique (failure).
108 * We define "read traversal" operation as any of the following
109 * group of operations
110 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
111 * (and/or cds_lfht_next, although less common).
112 * - cds_lfht_add_unique (failure) followed by iteration with
113 * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
115 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
116 * cds_lfht_next_duplicate, although less common).
118 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
119 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
121 * When cds_lfht_add_unique succeeds (returns the node passed as
122 * parameter), it acts as a "write" operation. When cds_lfht_add_unique
123 * fails (returns a node different from the one passed as parameter), it
124 * acts as a "read" operation. A cds_lfht_add_unique failure is a
125 * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
126 * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
129 * We define "prior" and "later" node as nodes observable by reads and
130 * read traversals respectively before and after a write or sequence of
133 * Hash-table operations are often cascaded, for example, the pointer
134 * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
135 * whose return value might in turn be passed to another hash-table
136 * operation. This entire cascaded series of operations must be enclosed
137 * by a pair of matching rcu_read_lock() and rcu_read_unlock()
140 * The following ordering guarantees are offered by this hash table:
142 * A.1) "read" after "write": if there is ordering between a write and a
143 * later read, then the read is guaranteed to see the write or some
145 * A.2) "read traversal" after "write": given that there is dependency
146 * ordering between reads in a "read traversal", if there is
147 * ordering between a write and the first read of the traversal,
148 * then the "read traversal" is guaranteed to see the write or
150 * B.1) "write" after "read": if there is ordering between a read and a
151 * later write, then the read will never see the write.
152 * B.2) "write" after "read traversal": given that there is dependency
153 * ordering between reads in a "read traversal", if there is
154 * ordering between the last read of the traversal and a later
155 * write, then the "read traversal" will never see the write.
156 * C) "write" while "read traversal": if a write occurs during a "read
157 * traversal", the traversal may, or may not, see the write.
158 * D.1) "write" after "write": if there is ordering between a write and
159 * a later write, then the later write is guaranteed to see the
160 * effects of the first write.
161 * D.2) Concurrent "write" pairs: The system will assign an arbitrary
162 * order to any pair of concurrent conflicting writes.
163 * Non-conflicting writes (for example, to different keys) are
165 * E) If a grace period separates a "del" or "replace" operation
166 * and a subsequent operation, then that subsequent operation is
167 * guaranteed not to see the removed item.
168 * F) Uniqueness guarantee: given a hash table that does not contain
169 * duplicate items for a given key, there will only be one item in
170 * the hash table after an arbitrary sequence of add_unique and/or
171 * add_replace operations. Note, however, that a pair of
172 * concurrent read operations might well access two different items
174 * G.1) If a pair of lookups for a given key are ordered (e.g. by a
175 * memory barrier), then the second lookup will return the same
176 * node as the previous lookup, or some later node.
177 * G.2) A "read traversal" that starts after the end of a prior "read
178 * traversal" (ordered by memory barriers) is guaranteed to see the
179 * same nodes as the previous traversal, or some later nodes.
180 * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
181 * example, if a pair of reads to the same key run concurrently
182 * with an insertion of that same key, the reads remain unordered
183 * regardless of their return values. In other words, you cannot
184 * rely on the values returned by the reads to deduce ordering.
186 * Progress guarantees:
188 * * Reads are wait-free. These operations always move forward in the
189 * hash table linked list, and this list has no loop.
190 * * Writes are lock-free. Any retry loop performed by a write operation
191 * is triggered by progress made within another update operation.
193 * Bucket node tables:
195 * hash table hash table the last all bucket node tables
196 * order size bucket node 0 1 2 3 4 5 6(index)
203 * 5 32 16 1 1 2 4 8 16
204 * 6 64 32 1 1 2 4 8 16 32
206 * When growing/shrinking, we only focus on the last bucket node table
207 * which size is (!order ? 1 : (1 << (order -1))).
209 * Example for growing/shrinking:
210 * grow hash table from order 5 to 6: init the index=6 bucket node table
211 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
213 * A bit of ascii art explanation:
215 * The order index is the off-by-one compared to the actual power of 2
216 * because we use index 0 to deal with the 0 special-case.
218 * This shows the nodes for a small table ordered by reversed bits:
230 * This shows the nodes in order of non-reversed bits, linked by
231 * reversed-bit order.
236 * 2 | | 2 010 010 <- |
237 * | | | 3 011 110 | <- |
238 * 3 -> | | | 4 100 001 | |
253 #include "compat-getcpu.h"
254 #include <urcu/assert.h>
255 #include <urcu/pointer.h>
256 #include <urcu/call-rcu.h>
257 #include <urcu/flavor.h>
258 #include <urcu/arch.h>
259 #include <urcu/uatomic.h>
260 #include <urcu/compiler.h>
261 #include <urcu/rculfhash.h>
262 #include <urcu/static/urcu-signal-nr.h>
266 #include "rculfhash-internal.h"
267 #include "workqueue.h"
268 #include "urcu-die.h"
269 #include "urcu-utils.h"
270 #include "compat-smp.h"
273 * Split-counters lazily update the global counter each 1024
274 * addition/removal. It automatically keeps track of resize required.
275 * We use the bucket length as indicator for need to expand for small
276 * tables and machines lacking per-cpu data support.
278 #define COUNT_COMMIT_ORDER 10
279 #define DEFAULT_SPLIT_COUNT_MASK 0xFUL
280 #define CHAIN_LEN_TARGET 1
281 #define CHAIN_LEN_RESIZE_THRESHOLD 3
284 * Define the minimum table size.
286 #define MIN_TABLE_ORDER 0
287 #define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
290 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
292 #define MIN_PARTITION_PER_THREAD_ORDER 12
293 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
296 * The removed flag needs to be updated atomically with the pointer.
297 * It indicates that no node must attach to the node scheduled for
298 * removal, and that node garbage collection must be performed.
299 * The bucket flag does not require to be updated atomically with the
300 * pointer, but it is added as a pointer low bit flag to save space.
301 * The "removal owner" flag is used to detect which of the "del"
302 * operation that has set the "removed flag" gets to return the removed
303 * node to its caller. Note that the replace operation does not need to
304 * iteract with the "removal owner" flag, because it validates that
305 * the "removed" flag is not set before performing its cmpxchg.
307 #define REMOVED_FLAG (1UL << 0)
308 #define BUCKET_FLAG (1UL << 1)
309 #define REMOVAL_OWNER_FLAG (1UL << 2)
310 #define FLAGS_MASK ((1UL << 3) - 1)
312 /* Value of the end pointer. Should not interact with flags. */
313 #define END_VALUE NULL
316 * ht_items_count: Split-counters counting the number of node addition
317 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
318 * is set at hash table creation.
320 * These are free-running counters, never reset to zero. They count the
321 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
322 * operations to update the global counter. We choose a power-of-2 value
323 * for the trigger to deal with 32 or 64-bit overflow of the counter.
325 struct ht_items_count
{
326 unsigned long add
, del
;
327 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
330 * resize_work: Contains arguments passed to worker thread
331 * responsible for performing lazy resize.
334 struct urcu_work work
;
339 * partition_resize_work: Contains arguments passed to worker threads
340 * executing the hash table resize on partitions of the hash table
341 * assigned to each processor's worker thread.
343 struct partition_resize_work
{
346 unsigned long i
, start
, len
;
347 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
348 unsigned long start
, unsigned long len
);
351 static struct urcu_workqueue
*cds_lfht_workqueue
;
354 * Mutex ensuring mutual exclusion between workqueue initialization and
355 * fork handlers. cds_lfht_fork_mutex nests inside call_rcu_mutex.
357 static pthread_mutex_t cds_lfht_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
359 static struct urcu_atfork cds_lfht_atfork
;
362 * atfork handler nesting counters. Handle being registered to many urcu
363 * flavors, thus being possibly invoked more than once in the
364 * pthread_atfork list of callbacks.
366 static int cds_lfht_workqueue_atfork_nesting
;
368 static void __attribute__((destructor
)) cds_lfht_exit(void);
369 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
);
371 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
374 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
379 #define cds_lfht_iter_debug_assert(...) urcu_posix_assert(__VA_ARGS__)
384 void cds_lfht_iter_debug_set_ht(struct cds_lfht
*ht
__attribute__((unused
)),
385 struct cds_lfht_iter
*iter
__attribute__((unused
)))
389 #define cds_lfht_iter_debug_assert(...)
394 * Algorithm to reverse bits in a word by lookup table, extended to
397 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
398 * Originally from Public Domain.
401 static const uint8_t BitReverseTable256
[256] =
403 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
404 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
405 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
406 R6(0), R6(2), R6(1), R6(3)
413 uint8_t bit_reverse_u8(uint8_t v
)
415 return BitReverseTable256
[v
];
418 #if (CAA_BITS_PER_LONG == 32)
420 uint32_t bit_reverse_u32(uint32_t v
)
422 return ((uint32_t) bit_reverse_u8(v
) << 24) |
423 ((uint32_t) bit_reverse_u8(v
>> 8) << 16) |
424 ((uint32_t) bit_reverse_u8(v
>> 16) << 8) |
425 ((uint32_t) bit_reverse_u8(v
>> 24));
429 uint64_t bit_reverse_u64(uint64_t v
)
431 return ((uint64_t) bit_reverse_u8(v
) << 56) |
432 ((uint64_t) bit_reverse_u8(v
>> 8) << 48) |
433 ((uint64_t) bit_reverse_u8(v
>> 16) << 40) |
434 ((uint64_t) bit_reverse_u8(v
>> 24) << 32) |
435 ((uint64_t) bit_reverse_u8(v
>> 32) << 24) |
436 ((uint64_t) bit_reverse_u8(v
>> 40) << 16) |
437 ((uint64_t) bit_reverse_u8(v
>> 48) << 8) |
438 ((uint64_t) bit_reverse_u8(v
>> 56));
443 unsigned long bit_reverse_ulong(unsigned long v
)
445 #if (CAA_BITS_PER_LONG == 32)
446 return bit_reverse_u32(v
);
448 return bit_reverse_u64(v
);
453 * fls: returns the position of the most significant bit.
454 * Returns 0 if no bit is set, else returns the position of the most
455 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
457 #if defined(URCU_ARCH_X86)
459 unsigned int fls_u32(uint32_t x
)
463 __asm__ ("bsrl %1,%0\n\t"
467 : "=r" (r
) : "rm" (x
));
473 #if defined(URCU_ARCH_AMD64)
475 unsigned int fls_u64(uint64_t x
)
479 __asm__ ("bsrq %1,%0\n\t"
483 : "=r" (r
) : "rm" (x
));
490 static __attribute__((unused
))
491 unsigned int fls_u64(uint64_t x
)
498 if (!(x
& 0xFFFFFFFF00000000ULL
)) {
502 if (!(x
& 0xFFFF000000000000ULL
)) {
506 if (!(x
& 0xFF00000000000000ULL
)) {
510 if (!(x
& 0xF000000000000000ULL
)) {
514 if (!(x
& 0xC000000000000000ULL
)) {
518 if (!(x
& 0x8000000000000000ULL
)) {
527 static __attribute__((unused
))
528 unsigned int fls_u32(uint32_t x
)
534 if (!(x
& 0xFFFF0000U
)) {
538 if (!(x
& 0xFF000000U
)) {
542 if (!(x
& 0xF0000000U
)) {
546 if (!(x
& 0xC0000000U
)) {
550 if (!(x
& 0x80000000U
)) {
558 unsigned int cds_lfht_fls_ulong(unsigned long x
)
560 #if (CAA_BITS_PER_LONG == 32)
568 * Return the minimum order for which x <= (1UL << order).
569 * Return -1 if x is 0.
572 int cds_lfht_get_count_order_u32(uint32_t x
)
577 return fls_u32(x
- 1);
581 * Return the minimum order for which x <= (1UL << order).
582 * Return -1 if x is 0.
584 int cds_lfht_get_count_order_ulong(unsigned long x
)
589 return cds_lfht_fls_ulong(x
- 1);
593 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
);
596 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
597 unsigned long count
);
599 static void mutex_lock(pthread_mutex_t
*mutex
)
603 #ifndef DISTRUST_SIGNALS_EXTREME
604 ret
= pthread_mutex_lock(mutex
);
607 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
608 while ((ret
= pthread_mutex_trylock(mutex
)) != 0) {
609 if (ret
!= EBUSY
&& ret
!= EINTR
)
611 if (CMM_LOAD_SHARED(URCU_TLS(rcu_reader
).need_mb
)) {
613 _CMM_STORE_SHARED(URCU_TLS(rcu_reader
).need_mb
, 0);
616 (void) poll(NULL
, 0, 10);
618 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
621 static void mutex_unlock(pthread_mutex_t
*mutex
)
625 ret
= pthread_mutex_unlock(mutex
);
630 static long nr_cpus_mask
= -1;
631 static long split_count_mask
= -1;
632 static int split_count_order
= -1;
634 static void ht_init_nr_cpus_mask(void)
638 maxcpus
= get_possible_cpus_array_len();
644 * round up number of CPUs to next power of two, so we
645 * can use & for modulo.
647 maxcpus
= 1UL << cds_lfht_get_count_order_ulong(maxcpus
);
648 nr_cpus_mask
= maxcpus
- 1;
652 void alloc_split_items_count(struct cds_lfht
*ht
)
654 if (nr_cpus_mask
== -1) {
655 ht_init_nr_cpus_mask();
656 if (nr_cpus_mask
< 0)
657 split_count_mask
= DEFAULT_SPLIT_COUNT_MASK
;
659 split_count_mask
= nr_cpus_mask
;
661 cds_lfht_get_count_order_ulong(split_count_mask
+ 1);
664 urcu_posix_assert(split_count_mask
>= 0);
666 if (ht
->flags
& CDS_LFHT_ACCOUNTING
) {
667 ht
->split_count
= calloc(split_count_mask
+ 1,
668 sizeof(struct ht_items_count
));
669 urcu_posix_assert(ht
->split_count
);
671 ht
->split_count
= NULL
;
676 void free_split_items_count(struct cds_lfht
*ht
)
678 poison_free(ht
->split_count
);
682 int ht_get_split_count_index(unsigned long hash
)
686 urcu_posix_assert(split_count_mask
>= 0);
687 cpu
= urcu_sched_getcpu();
688 if (caa_unlikely(cpu
< 0))
689 return hash
& split_count_mask
;
691 return cpu
& split_count_mask
;
695 void ht_count_add(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
697 unsigned long split_count
, count
;
700 if (caa_unlikely(!ht
->split_count
))
702 index
= ht_get_split_count_index(hash
);
703 split_count
= uatomic_add_return(&ht
->split_count
[index
].add
, 1);
704 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
706 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
708 dbg_printf("add split count %lu\n", split_count
);
709 count
= uatomic_add_return(&ht
->count
,
710 1UL << COUNT_COMMIT_ORDER
);
711 if (caa_likely(count
& (count
- 1)))
713 /* Only if global count is power of 2 */
715 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) < size
)
717 dbg_printf("add set global %lu\n", count
);
718 cds_lfht_resize_lazy_count(ht
, size
,
719 count
>> (CHAIN_LEN_TARGET
- 1));
723 void ht_count_del(struct cds_lfht
*ht
, unsigned long size
, unsigned long hash
)
725 unsigned long split_count
, count
;
728 if (caa_unlikely(!ht
->split_count
))
730 index
= ht_get_split_count_index(hash
);
731 split_count
= uatomic_add_return(&ht
->split_count
[index
].del
, 1);
732 if (caa_likely(split_count
& ((1UL << COUNT_COMMIT_ORDER
) - 1)))
734 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
736 dbg_printf("del split count %lu\n", split_count
);
737 count
= uatomic_add_return(&ht
->count
,
738 -(1UL << COUNT_COMMIT_ORDER
));
739 if (caa_likely(count
& (count
- 1)))
741 /* Only if global count is power of 2 */
743 if ((count
>> CHAIN_LEN_RESIZE_THRESHOLD
) >= size
)
745 dbg_printf("del set global %lu\n", count
);
747 * Don't shrink table if the number of nodes is below a
750 if (count
< (1UL << COUNT_COMMIT_ORDER
) * (split_count_mask
+ 1))
752 cds_lfht_resize_lazy_count(ht
, size
,
753 count
>> (CHAIN_LEN_TARGET
- 1));
757 void check_resize(struct cds_lfht
*ht
, unsigned long size
, uint32_t chain_len
)
761 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
763 count
= uatomic_read(&ht
->count
);
765 * Use bucket-local length for small table expand and for
766 * environments lacking per-cpu data support.
768 if (count
>= (1UL << (COUNT_COMMIT_ORDER
+ split_count_order
)))
771 dbg_printf("WARNING: large chain length: %u.\n",
773 if (chain_len
>= CHAIN_LEN_RESIZE_THRESHOLD
) {
777 * Ideal growth calculated based on chain length.
779 growth
= cds_lfht_get_count_order_u32(chain_len
780 - (CHAIN_LEN_TARGET
- 1));
781 if ((ht
->flags
& CDS_LFHT_ACCOUNTING
)
783 >= (1UL << (COUNT_COMMIT_ORDER
784 + split_count_order
))) {
786 * If ideal growth expands the hash table size
787 * beyond the "small hash table" sizes, use the
788 * maximum small hash table size to attempt
789 * expanding the hash table. This only applies
790 * when node accounting is available, otherwise
791 * the chain length is used to expand the hash
792 * table in every case.
794 growth
= COUNT_COMMIT_ORDER
+ split_count_order
795 - cds_lfht_get_count_order_ulong(size
);
799 cds_lfht_resize_lazy_grow(ht
, size
, growth
);
804 struct cds_lfht_node
*clear_flag(struct cds_lfht_node
*node
)
806 return (struct cds_lfht_node
*) (((unsigned long) node
) & ~FLAGS_MASK
);
810 int is_removed(const struct cds_lfht_node
*node
)
812 return ((unsigned long) node
) & REMOVED_FLAG
;
816 int is_bucket(struct cds_lfht_node
*node
)
818 return ((unsigned long) node
) & BUCKET_FLAG
;
822 struct cds_lfht_node
*flag_bucket(struct cds_lfht_node
*node
)
824 return (struct cds_lfht_node
*) (((unsigned long) node
) | BUCKET_FLAG
);
828 int is_removal_owner(struct cds_lfht_node
*node
)
830 return ((unsigned long) node
) & REMOVAL_OWNER_FLAG
;
834 struct cds_lfht_node
*flag_removed(struct cds_lfht_node
*node
)
836 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
);
840 struct cds_lfht_node
*flag_removal_owner(struct cds_lfht_node
*node
)
842 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVAL_OWNER_FLAG
);
846 struct cds_lfht_node
*flag_removed_or_removal_owner(struct cds_lfht_node
*node
)
848 return (struct cds_lfht_node
*) (((unsigned long) node
) | REMOVED_FLAG
| REMOVAL_OWNER_FLAG
);
852 struct cds_lfht_node
*get_end(void)
854 return (struct cds_lfht_node
*) END_VALUE
;
858 int is_end(struct cds_lfht_node
*node
)
860 return clear_flag(node
) == (struct cds_lfht_node
*) END_VALUE
;
864 unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr
,
867 unsigned long old1
, old2
;
869 old1
= uatomic_read(ptr
);
874 } while ((old1
= uatomic_cmpxchg(ptr
, old2
, v
)) != old2
);
879 void cds_lfht_alloc_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
881 return ht
->mm
->alloc_bucket_table(ht
, order
);
885 * cds_lfht_free_bucket_table() should be called with decreasing order.
886 * When cds_lfht_free_bucket_table(0) is called, it means the whole
890 void cds_lfht_free_bucket_table(struct cds_lfht
*ht
, unsigned long order
)
892 return ht
->mm
->free_bucket_table(ht
, order
);
896 struct cds_lfht_node
*bucket_at(struct cds_lfht
*ht
, unsigned long index
)
898 return ht
->bucket_at(ht
, index
);
902 struct cds_lfht_node
*lookup_bucket(struct cds_lfht
*ht
, unsigned long size
,
905 urcu_posix_assert(size
> 0);
906 return bucket_at(ht
, hash
& (size
- 1));
910 * Remove all logically deleted nodes from a bucket up to a certain node key.
913 void _cds_lfht_gc_bucket(struct cds_lfht_node
*bucket
, struct cds_lfht_node
*node
)
915 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_next
;
917 urcu_posix_assert(!is_bucket(bucket
));
918 urcu_posix_assert(!is_removed(bucket
));
919 urcu_posix_assert(!is_removal_owner(bucket
));
920 urcu_posix_assert(!is_bucket(node
));
921 urcu_posix_assert(!is_removed(node
));
922 urcu_posix_assert(!is_removal_owner(node
));
925 /* We can always skip the bucket node initially */
926 iter
= rcu_dereference(iter_prev
->next
);
927 urcu_posix_assert(!is_removed(iter
));
928 urcu_posix_assert(!is_removal_owner(iter
));
929 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
931 * We should never be called with bucket (start of chain)
932 * and logically removed node (end of path compression
933 * marker) being the actual same node. This would be a
934 * bug in the algorithm implementation.
936 urcu_posix_assert(bucket
!= node
);
938 if (caa_unlikely(is_end(iter
)))
940 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
942 next
= rcu_dereference(clear_flag(iter
)->next
);
943 if (caa_likely(is_removed(next
)))
945 iter_prev
= clear_flag(iter
);
948 urcu_posix_assert(!is_removed(iter
));
949 urcu_posix_assert(!is_removal_owner(iter
));
951 new_next
= flag_bucket(clear_flag(next
));
953 new_next
= clear_flag(next
);
954 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
959 int _cds_lfht_replace(struct cds_lfht
*ht
, unsigned long size
,
960 struct cds_lfht_node
*old_node
,
961 struct cds_lfht_node
*old_next
,
962 struct cds_lfht_node
*new_node
)
964 struct cds_lfht_node
*bucket
, *ret_next
;
966 if (!old_node
) /* Return -ENOENT if asked to replace NULL node */
969 urcu_posix_assert(!is_removed(old_node
));
970 urcu_posix_assert(!is_removal_owner(old_node
));
971 urcu_posix_assert(!is_bucket(old_node
));
972 urcu_posix_assert(!is_removed(new_node
));
973 urcu_posix_assert(!is_removal_owner(new_node
));
974 urcu_posix_assert(!is_bucket(new_node
));
975 urcu_posix_assert(new_node
!= old_node
);
977 /* Insert after node to be replaced */
978 if (is_removed(old_next
)) {
980 * Too late, the old node has been removed under us
981 * between lookup and replace. Fail.
985 urcu_posix_assert(old_next
== clear_flag(old_next
));
986 urcu_posix_assert(new_node
!= old_next
);
988 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
989 * flag. It is either set atomically at the same time
990 * (replace) or after (del).
992 urcu_posix_assert(!is_removal_owner(old_next
));
993 new_node
->next
= old_next
;
995 * Here is the whole trick for lock-free replace: we add
996 * the replacement node _after_ the node we want to
997 * replace by atomically setting its next pointer at the
998 * same time we set its removal flag. Given that
999 * the lookups/get next use an iterator aware of the
1000 * next pointer, they will either skip the old node due
1001 * to the removal flag and see the new node, or use
1002 * the old node, but will not see the new one.
1003 * This is a replacement of a node with another node
1004 * that has the same value: we are therefore not
1005 * removing a value from the hash table. We set both the
1006 * REMOVED and REMOVAL_OWNER flags atomically so we own
1007 * the node after successful cmpxchg.
1009 ret_next
= uatomic_cmpxchg(&old_node
->next
,
1010 old_next
, flag_removed_or_removal_owner(new_node
));
1011 if (ret_next
== old_next
)
1012 break; /* We performed the replacement. */
1013 old_next
= ret_next
;
1017 * Ensure that the old node is not visible to readers anymore:
1018 * lookup for the node, and remove it (along with any other
1019 * logically removed node) if found.
1021 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(old_node
->reverse_hash
));
1022 _cds_lfht_gc_bucket(bucket
, new_node
);
1024 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(old_node
->next
)));
1029 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
1030 * mode. A NULL unique_ret allows creation of duplicate keys.
1033 void _cds_lfht_add(struct cds_lfht
*ht
,
1035 cds_lfht_match_fct match
,
1038 struct cds_lfht_node
*node
,
1039 struct cds_lfht_iter
*unique_ret
,
1042 struct cds_lfht_node
*iter_prev
, *iter
, *next
, *new_node
, *new_next
,
1044 struct cds_lfht_node
*bucket
;
1046 urcu_posix_assert(!is_bucket(node
));
1047 urcu_posix_assert(!is_removed(node
));
1048 urcu_posix_assert(!is_removal_owner(node
));
1049 bucket
= lookup_bucket(ht
, size
, hash
);
1051 uint32_t chain_len
= 0;
1054 * iter_prev points to the non-removed node prior to the
1058 /* We can always skip the bucket node initially */
1059 iter
= rcu_dereference(iter_prev
->next
);
1060 urcu_posix_assert(iter_prev
->reverse_hash
<= node
->reverse_hash
);
1062 if (caa_unlikely(is_end(iter
)))
1064 if (caa_likely(clear_flag(iter
)->reverse_hash
> node
->reverse_hash
))
1067 /* bucket node is the first node of the identical-hash-value chain */
1068 if (bucket_flag
&& clear_flag(iter
)->reverse_hash
== node
->reverse_hash
)
1071 next
= rcu_dereference(clear_flag(iter
)->next
);
1072 if (caa_unlikely(is_removed(next
)))
1078 && clear_flag(iter
)->reverse_hash
== node
->reverse_hash
) {
1079 struct cds_lfht_iter d_iter
= {
1082 #ifdef CONFIG_CDS_LFHT_ITER_DEBUG
1088 * uniquely adding inserts the node as the first
1089 * node of the identical-hash-value node chain.
1091 * This semantic ensures no duplicated keys
1092 * should ever be observable in the table
1093 * (including traversing the table node by
1094 * node by forward iterations)
1096 cds_lfht_next_duplicate(ht
, match
, key
, &d_iter
);
1100 *unique_ret
= d_iter
;
1104 /* Only account for identical reverse hash once */
1105 if (iter_prev
->reverse_hash
!= clear_flag(iter
)->reverse_hash
1106 && !is_bucket(next
))
1107 check_resize(ht
, size
, ++chain_len
);
1108 iter_prev
= clear_flag(iter
);
1113 urcu_posix_assert(node
!= clear_flag(iter
));
1114 urcu_posix_assert(!is_removed(iter_prev
));
1115 urcu_posix_assert(!is_removal_owner(iter_prev
));
1116 urcu_posix_assert(!is_removed(iter
));
1117 urcu_posix_assert(!is_removal_owner(iter
));
1118 urcu_posix_assert(iter_prev
!= node
);
1120 node
->next
= clear_flag(iter
);
1122 node
->next
= flag_bucket(clear_flag(iter
));
1123 if (is_bucket(iter
))
1124 new_node
= flag_bucket(node
);
1127 if (uatomic_cmpxchg(&iter_prev
->next
, iter
,
1128 new_node
) != iter
) {
1129 continue; /* retry */
1136 urcu_posix_assert(!is_removed(iter
));
1137 urcu_posix_assert(!is_removal_owner(iter
));
1138 if (is_bucket(iter
))
1139 new_next
= flag_bucket(clear_flag(next
));
1141 new_next
= clear_flag(next
);
1142 (void) uatomic_cmpxchg(&iter_prev
->next
, iter
, new_next
);
1147 unique_ret
->node
= return_node
;
1148 /* unique_ret->next left unset, never used. */
1153 int _cds_lfht_del(struct cds_lfht
*ht
, unsigned long size
,
1154 struct cds_lfht_node
*node
)
1156 struct cds_lfht_node
*bucket
, *next
;
1158 if (!node
) /* Return -ENOENT if asked to delete NULL node */
1161 /* logically delete the node */
1162 urcu_posix_assert(!is_bucket(node
));
1163 urcu_posix_assert(!is_removed(node
));
1164 urcu_posix_assert(!is_removal_owner(node
));
1167 * We are first checking if the node had previously been
1168 * logically removed (this check is not atomic with setting the
1169 * logical removal flag). Return -ENOENT if the node had
1170 * previously been removed.
1172 next
= CMM_LOAD_SHARED(node
->next
); /* next is not dereferenced */
1173 if (caa_unlikely(is_removed(next
)))
1175 urcu_posix_assert(!is_bucket(next
));
1177 * The del operation semantic guarantees a full memory barrier
1178 * before the uatomic_or atomic commit of the deletion flag.
1180 cmm_smp_mb__before_uatomic_or();
1182 * We set the REMOVED_FLAG unconditionally. Note that there may
1183 * be more than one concurrent thread setting this flag.
1184 * Knowing which wins the race will be known after the garbage
1185 * collection phase, stay tuned!
1187 uatomic_or(&node
->next
, REMOVED_FLAG
);
1188 /* We performed the (logical) deletion. */
1191 * Ensure that the node is not visible to readers anymore: lookup for
1192 * the node, and remove it (along with any other logically removed node)
1195 bucket
= lookup_bucket(ht
, size
, bit_reverse_ulong(node
->reverse_hash
));
1196 _cds_lfht_gc_bucket(bucket
, node
);
1198 urcu_posix_assert(is_removed(CMM_LOAD_SHARED(node
->next
)));
1200 * Last phase: atomically exchange node->next with a version
1201 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1202 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1203 * the node and win the removal race.
1204 * It is interesting to note that all "add" paths are forbidden
1205 * to change the next pointer starting from the point where the
1206 * REMOVED_FLAG is set, so here using a read, followed by a
1207 * xchg() suffice to guarantee that the xchg() will ever only
1208 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1211 if (!is_removal_owner(uatomic_xchg(&node
->next
,
1212 flag_removal_owner(node
->next
))))
1219 void *partition_resize_thread(void *arg
)
1221 struct partition_resize_work
*work
= arg
;
1223 work
->ht
->flavor
->register_thread();
1224 work
->fct(work
->ht
, work
->i
, work
->start
, work
->len
);
1225 work
->ht
->flavor
->unregister_thread();
1230 void partition_resize_helper(struct cds_lfht
*ht
, unsigned long i
,
1232 void (*fct
)(struct cds_lfht
*ht
, unsigned long i
,
1233 unsigned long start
, unsigned long len
))
1235 unsigned long partition_len
, start
= 0;
1236 struct partition_resize_work
*work
;
1238 unsigned long thread
, nr_threads
;
1239 sigset_t newmask
, oldmask
;
1241 urcu_posix_assert(nr_cpus_mask
!= -1);
1242 if (nr_cpus_mask
< 0 || len
< 2 * MIN_PARTITION_PER_THREAD
)
1246 * Note: nr_cpus_mask + 1 is always power of 2.
1247 * We spawn just the number of threads we need to satisfy the minimum
1248 * partition size, up to the number of CPUs in the system.
1250 if (nr_cpus_mask
> 0) {
1251 nr_threads
= min_t(unsigned long, nr_cpus_mask
+ 1,
1252 len
>> MIN_PARTITION_PER_THREAD_ORDER
);
1256 partition_len
= len
>> cds_lfht_get_count_order_ulong(nr_threads
);
1257 work
= calloc(nr_threads
, sizeof(*work
));
1259 dbg_printf("error allocating for resize, single-threading\n");
1263 ret
= sigfillset(&newmask
);
1264 urcu_posix_assert(!ret
);
1265 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
1266 urcu_posix_assert(!ret
);
1268 for (thread
= 0; thread
< nr_threads
; thread
++) {
1269 work
[thread
].ht
= ht
;
1271 work
[thread
].len
= partition_len
;
1272 work
[thread
].start
= thread
* partition_len
;
1273 work
[thread
].fct
= fct
;
1274 ret
= pthread_create(&(work
[thread
].thread_id
),
1275 ht
->caller_resize_attr
? &ht
->resize_attr
: NULL
,
1276 partition_resize_thread
, &work
[thread
]);
1277 if (ret
== EAGAIN
) {
1279 * Out of resources: wait and join the threads
1280 * we've created, then handle leftovers.
1282 dbg_printf("error spawning for resize, single-threading\n");
1283 start
= work
[thread
].start
;
1285 nr_threads
= thread
;
1288 urcu_posix_assert(!ret
);
1291 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
1292 urcu_posix_assert(!ret
);
1294 for (thread
= 0; thread
< nr_threads
; thread
++) {
1295 ret
= pthread_join(work
[thread
].thread_id
, NULL
);
1296 urcu_posix_assert(!ret
);
1301 * A pthread_create failure above will either lead in us having
1302 * no threads to join or starting at a non-zero offset,
1303 * fallback to single thread processing of leftovers.
1305 if (start
== 0 && nr_threads
> 0)
1308 fct(ht
, i
, start
, len
);
1312 * Holding RCU read lock to protect _cds_lfht_add against memory
1313 * reclaim that could be performed by other worker threads (ABA
1316 * When we reach a certain length, we can split this population phase over
1317 * many worker threads, based on the number of CPUs available in the system.
1318 * This should therefore take care of not having the expand lagging behind too
1319 * many concurrent insertion threads by using the scheduler's ability to
1320 * schedule bucket node population fairly with insertions.
1323 void init_table_populate_partition(struct cds_lfht
*ht
, unsigned long i
,
1324 unsigned long start
, unsigned long len
)
1326 unsigned long j
, size
= 1UL << (i
- 1);
1328 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1329 ht
->flavor
->read_lock();
1330 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1331 struct cds_lfht_node
*new_node
= bucket_at(ht
, j
);
1333 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1334 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1336 new_node
->reverse_hash
= bit_reverse_ulong(j
);
1337 _cds_lfht_add(ht
, j
, NULL
, NULL
, size
, new_node
, NULL
, 1);
1339 ht
->flavor
->read_unlock();
1343 void init_table_populate(struct cds_lfht
*ht
, unsigned long i
,
1346 partition_resize_helper(ht
, i
, len
, init_table_populate_partition
);
1350 void init_table(struct cds_lfht
*ht
,
1351 unsigned long first_order
, unsigned long last_order
)
1355 dbg_printf("init table: first_order %lu last_order %lu\n",
1356 first_order
, last_order
);
1357 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1358 for (i
= first_order
; i
<= last_order
; i
++) {
1361 len
= 1UL << (i
- 1);
1362 dbg_printf("init order %lu len: %lu\n", i
, len
);
1364 /* Stop expand if the resize target changes under us */
1365 if (CMM_LOAD_SHARED(ht
->resize_target
) < (1UL << i
))
1368 cds_lfht_alloc_bucket_table(ht
, i
);
1371 * Set all bucket nodes reverse hash values for a level and
1372 * link all bucket nodes into the table.
1374 init_table_populate(ht
, i
, len
);
1377 * Update table size.
1379 cmm_smp_wmb(); /* populate data before RCU size */
1380 CMM_STORE_SHARED(ht
->size
, 1UL << i
);
1382 dbg_printf("init new size: %lu\n", 1UL << i
);
1383 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1389 * Holding RCU read lock to protect _cds_lfht_remove against memory
1390 * reclaim that could be performed by other worker threads (ABA
1392 * For a single level, we logically remove and garbage collect each node.
1394 * As a design choice, we perform logical removal and garbage collection on a
1395 * node-per-node basis to simplify this algorithm. We also assume keeping good
1396 * cache locality of the operation would overweight possible performance gain
1397 * that could be achieved by batching garbage collection for multiple levels.
1398 * However, this would have to be justified by benchmarks.
1400 * Concurrent removal and add operations are helping us perform garbage
1401 * collection of logically removed nodes. We guarantee that all logically
1402 * removed nodes have been garbage-collected (unlinked) before work
1403 * enqueue is invoked to free a hole level of bucket nodes (after a
1406 * Logical removal and garbage collection can therefore be done in batch
1407 * or on a node-per-node basis, as long as the guarantee above holds.
1409 * When we reach a certain length, we can split this removal over many worker
1410 * threads, based on the number of CPUs available in the system. This should
1411 * take care of not letting resize process lag behind too many concurrent
1412 * updater threads actively inserting into the hash table.
1415 void remove_table_partition(struct cds_lfht
*ht
, unsigned long i
,
1416 unsigned long start
, unsigned long len
)
1418 unsigned long j
, size
= 1UL << (i
- 1);
1420 urcu_posix_assert(i
> MIN_TABLE_ORDER
);
1421 ht
->flavor
->read_lock();
1422 for (j
= size
+ start
; j
< size
+ start
+ len
; j
++) {
1423 struct cds_lfht_node
*fini_bucket
= bucket_at(ht
, j
);
1424 struct cds_lfht_node
*parent_bucket
= bucket_at(ht
, j
- size
);
1426 urcu_posix_assert(j
>= size
&& j
< (size
<< 1));
1427 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1429 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1430 uatomic_or(&fini_bucket
->next
, REMOVED_FLAG
);
1431 _cds_lfht_gc_bucket(parent_bucket
, fini_bucket
);
1433 ht
->flavor
->read_unlock();
1437 void remove_table(struct cds_lfht
*ht
, unsigned long i
, unsigned long len
)
1439 partition_resize_helper(ht
, i
, len
, remove_table_partition
);
1443 * fini_table() is never called for first_order == 0, which is why
1444 * free_by_rcu_order == 0 can be used as criterion to know if free must
1448 void fini_table(struct cds_lfht
*ht
,
1449 unsigned long first_order
, unsigned long last_order
)
1451 unsigned long free_by_rcu_order
= 0, i
;
1453 dbg_printf("fini table: first_order %lu last_order %lu\n",
1454 first_order
, last_order
);
1455 urcu_posix_assert(first_order
> MIN_TABLE_ORDER
);
1456 for (i
= last_order
; i
>= first_order
; i
--) {
1459 len
= 1UL << (i
- 1);
1460 dbg_printf("fini order %ld len: %lu\n", i
, len
);
1462 /* Stop shrink if the resize target changes under us */
1463 if (CMM_LOAD_SHARED(ht
->resize_target
) > (1UL << (i
- 1)))
1466 cmm_smp_wmb(); /* populate data before RCU size */
1467 CMM_STORE_SHARED(ht
->size
, 1UL << (i
- 1));
1470 * We need to wait for all add operations to reach Q.S. (and
1471 * thus use the new table for lookups) before we can start
1472 * releasing the old bucket nodes. Otherwise their lookup will
1473 * return a logically removed node as insert position.
1475 ht
->flavor
->update_synchronize_rcu();
1476 if (free_by_rcu_order
)
1477 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1480 * Set "removed" flag in bucket nodes about to be removed.
1481 * Unlink all now-logically-removed bucket node pointers.
1482 * Concurrent add/remove operation are helping us doing
1485 remove_table(ht
, i
, len
);
1487 free_by_rcu_order
= i
;
1489 dbg_printf("fini new size: %lu\n", 1UL << i
);
1490 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
1494 if (free_by_rcu_order
) {
1495 ht
->flavor
->update_synchronize_rcu();
1496 cds_lfht_free_bucket_table(ht
, free_by_rcu_order
);
1501 * Never called with size < 1.
1504 void cds_lfht_create_bucket(struct cds_lfht
*ht
, unsigned long size
)
1506 struct cds_lfht_node
*prev
, *node
;
1507 unsigned long order
, len
, i
;
1510 cds_lfht_alloc_bucket_table(ht
, 0);
1512 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1513 node
= bucket_at(ht
, 0);
1514 node
->next
= flag_bucket(get_end());
1515 node
->reverse_hash
= 0;
1517 bucket_order
= cds_lfht_get_count_order_ulong(size
);
1518 urcu_posix_assert(bucket_order
>= 0);
1520 for (order
= 1; order
< (unsigned long) bucket_order
+ 1; order
++) {
1521 len
= 1UL << (order
- 1);
1522 cds_lfht_alloc_bucket_table(ht
, order
);
1524 for (i
= 0; i
< len
; i
++) {
1526 * Now, we are trying to init the node with the
1527 * hash=(len+i) (which is also a bucket with the
1528 * index=(len+i)) and insert it into the hash table,
1529 * so this node has to be inserted after the bucket
1530 * with the index=(len+i)&(len-1)=i. And because there
1531 * is no other non-bucket node nor bucket node with
1532 * larger index/hash inserted, so the bucket node
1533 * being inserted should be inserted directly linked
1534 * after the bucket node with index=i.
1536 prev
= bucket_at(ht
, i
);
1537 node
= bucket_at(ht
, len
+ i
);
1539 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
1540 order
, len
+ i
, len
+ i
);
1541 node
->reverse_hash
= bit_reverse_ulong(len
+ i
);
1543 /* insert after prev */
1544 urcu_posix_assert(is_bucket(prev
->next
));
1545 node
->next
= prev
->next
;
1546 prev
->next
= flag_bucket(node
);
1551 #if (CAA_BITS_PER_LONG > 32)
1553 * For 64-bit architectures, with max number of buckets small enough not to
1554 * use the entire 64-bit memory mapping space (and allowing a fair number of
1555 * hash table instances), use the mmap allocator, which is faster. Otherwise,
1556 * fallback to the order allocator.
1559 const struct cds_lfht_mm_type
*get_mm_type(unsigned long max_nr_buckets
)
1561 if (max_nr_buckets
&& max_nr_buckets
<= (1ULL << 32))
1562 return &cds_lfht_mm_mmap
;
1564 return &cds_lfht_mm_order
;
1568 * For 32-bit architectures, use the order allocator.
1571 const struct cds_lfht_mm_type
*get_mm_type(
1572 unsigned long max_nr_buckets
__attribute__((unused
)))
1574 return &cds_lfht_mm_order
;
1578 void cds_lfht_node_init_deleted(struct cds_lfht_node
*node
)
1580 cds_lfht_node_init(node
);
1581 node
->next
= flag_removed(NULL
);
1584 struct cds_lfht
*_cds_lfht_new(unsigned long init_size
,
1585 unsigned long min_nr_alloc_buckets
,
1586 unsigned long max_nr_buckets
,
1588 const struct cds_lfht_mm_type
*mm
,
1589 const struct rcu_flavor_struct
*flavor
,
1590 pthread_attr_t
*attr
)
1592 struct cds_lfht
*ht
;
1593 unsigned long order
;
1595 /* min_nr_alloc_buckets must be power of two */
1596 if (!min_nr_alloc_buckets
|| (min_nr_alloc_buckets
& (min_nr_alloc_buckets
- 1)))
1599 /* init_size must be power of two */
1600 if (!init_size
|| (init_size
& (init_size
- 1)))
1604 * Memory management plugin default.
1607 mm
= get_mm_type(max_nr_buckets
);
1609 /* max_nr_buckets == 0 for order based mm means infinite */
1610 if (mm
== &cds_lfht_mm_order
&& !max_nr_buckets
)
1611 max_nr_buckets
= 1UL << (MAX_TABLE_ORDER
- 1);
1613 /* max_nr_buckets must be power of two */
1614 if (!max_nr_buckets
|| (max_nr_buckets
& (max_nr_buckets
- 1)))
1617 if (flags
& CDS_LFHT_AUTO_RESIZE
)
1618 cds_lfht_init_worker(flavor
);
1620 min_nr_alloc_buckets
= max(min_nr_alloc_buckets
, MIN_TABLE_SIZE
);
1621 init_size
= max(init_size
, MIN_TABLE_SIZE
);
1622 max_nr_buckets
= max(max_nr_buckets
, min_nr_alloc_buckets
);
1623 init_size
= min(init_size
, max_nr_buckets
);
1625 ht
= mm
->alloc_cds_lfht(min_nr_alloc_buckets
, max_nr_buckets
);
1626 urcu_posix_assert(ht
);
1627 urcu_posix_assert(ht
->mm
== mm
);
1628 urcu_posix_assert(ht
->bucket_at
== mm
->bucket_at
);
1631 ht
->flavor
= flavor
;
1632 ht
->caller_resize_attr
= attr
;
1634 ht
->resize_attr
= *attr
;
1635 alloc_split_items_count(ht
);
1636 /* this mutex should not nest in read-side C.S. */
1637 pthread_mutex_init(&ht
->resize_mutex
, NULL
);
1638 order
= cds_lfht_get_count_order_ulong(init_size
);
1639 ht
->resize_target
= 1UL << order
;
1640 cds_lfht_create_bucket(ht
, 1UL << order
);
1641 ht
->size
= 1UL << order
;
1645 void cds_lfht_lookup(struct cds_lfht
*ht
, unsigned long hash
,
1646 cds_lfht_match_fct match
, const void *key
,
1647 struct cds_lfht_iter
*iter
)
1649 struct cds_lfht_node
*node
, *next
, *bucket
;
1650 unsigned long reverse_hash
, size
;
1652 cds_lfht_iter_debug_set_ht(ht
, iter
);
1654 reverse_hash
= bit_reverse_ulong(hash
);
1656 size
= rcu_dereference(ht
->size
);
1657 bucket
= lookup_bucket(ht
, size
, hash
);
1658 /* We can always skip the bucket node initially */
1659 node
= rcu_dereference(bucket
->next
);
1660 node
= clear_flag(node
);
1662 if (caa_unlikely(is_end(node
))) {
1666 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1670 next
= rcu_dereference(node
->next
);
1671 urcu_posix_assert(node
== clear_flag(node
));
1672 if (caa_likely(!is_removed(next
))
1674 && node
->reverse_hash
== reverse_hash
1675 && caa_likely(match(node
, key
))) {
1678 node
= clear_flag(next
);
1680 urcu_posix_assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1685 void cds_lfht_next_duplicate(struct cds_lfht
*ht
__attribute__((unused
)),
1686 cds_lfht_match_fct match
,
1687 const void *key
, struct cds_lfht_iter
*iter
)
1689 struct cds_lfht_node
*node
, *next
;
1690 unsigned long reverse_hash
;
1692 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1694 reverse_hash
= node
->reverse_hash
;
1696 node
= clear_flag(next
);
1699 if (caa_unlikely(is_end(node
))) {
1703 if (caa_unlikely(node
->reverse_hash
> reverse_hash
)) {
1707 next
= rcu_dereference(node
->next
);
1708 if (caa_likely(!is_removed(next
))
1710 && caa_likely(match(node
, key
))) {
1713 node
= clear_flag(next
);
1715 urcu_posix_assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1720 void cds_lfht_next(struct cds_lfht
*ht
__attribute__((unused
)),
1721 struct cds_lfht_iter
*iter
)
1723 struct cds_lfht_node
*node
, *next
;
1725 cds_lfht_iter_debug_assert(ht
== iter
->lfht
);
1726 node
= clear_flag(iter
->next
);
1728 if (caa_unlikely(is_end(node
))) {
1732 next
= rcu_dereference(node
->next
);
1733 if (caa_likely(!is_removed(next
))
1734 && !is_bucket(next
)) {
1737 node
= clear_flag(next
);
1739 urcu_posix_assert(!node
|| !is_bucket(CMM_LOAD_SHARED(node
->next
)));
1744 void cds_lfht_first(struct cds_lfht
*ht
, struct cds_lfht_iter
*iter
)
1746 cds_lfht_iter_debug_set_ht(ht
, iter
);
1748 * Get next after first bucket node. The first bucket node is the
1749 * first node of the linked list.
1751 iter
->next
= bucket_at(ht
, 0)->next
;
1752 cds_lfht_next(ht
, iter
);
1755 void cds_lfht_add(struct cds_lfht
*ht
, unsigned long hash
,
1756 struct cds_lfht_node
*node
)
1760 node
->reverse_hash
= bit_reverse_ulong(hash
);
1761 size
= rcu_dereference(ht
->size
);
1762 _cds_lfht_add(ht
, hash
, NULL
, NULL
, size
, node
, NULL
, 0);
1763 ht_count_add(ht
, size
, hash
);
1766 struct cds_lfht_node
*cds_lfht_add_unique(struct cds_lfht
*ht
,
1768 cds_lfht_match_fct match
,
1770 struct cds_lfht_node
*node
)
1773 struct cds_lfht_iter iter
;
1775 node
->reverse_hash
= bit_reverse_ulong(hash
);
1776 size
= rcu_dereference(ht
->size
);
1777 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1778 if (iter
.node
== node
)
1779 ht_count_add(ht
, size
, hash
);
1783 struct cds_lfht_node
*cds_lfht_add_replace(struct cds_lfht
*ht
,
1785 cds_lfht_match_fct match
,
1787 struct cds_lfht_node
*node
)
1790 struct cds_lfht_iter iter
;
1792 node
->reverse_hash
= bit_reverse_ulong(hash
);
1793 size
= rcu_dereference(ht
->size
);
1795 _cds_lfht_add(ht
, hash
, match
, key
, size
, node
, &iter
, 0);
1796 if (iter
.node
== node
) {
1797 ht_count_add(ht
, size
, hash
);
1801 if (!_cds_lfht_replace(ht
, size
, iter
.node
, iter
.next
, node
))
1806 int cds_lfht_replace(struct cds_lfht
*ht
,
1807 struct cds_lfht_iter
*old_iter
,
1809 cds_lfht_match_fct match
,
1811 struct cds_lfht_node
*new_node
)
1815 new_node
->reverse_hash
= bit_reverse_ulong(hash
);
1816 if (!old_iter
->node
)
1818 if (caa_unlikely(old_iter
->node
->reverse_hash
!= new_node
->reverse_hash
))
1820 if (caa_unlikely(!match(old_iter
->node
, key
)))
1822 size
= rcu_dereference(ht
->size
);
1823 return _cds_lfht_replace(ht
, size
, old_iter
->node
, old_iter
->next
,
1827 int cds_lfht_del(struct cds_lfht
*ht
, struct cds_lfht_node
*node
)
1832 size
= rcu_dereference(ht
->size
);
1833 ret
= _cds_lfht_del(ht
, size
, node
);
1837 hash
= bit_reverse_ulong(node
->reverse_hash
);
1838 ht_count_del(ht
, size
, hash
);
1843 int cds_lfht_is_node_deleted(const struct cds_lfht_node
*node
)
1845 return is_removed(CMM_LOAD_SHARED(node
->next
));
1849 bool cds_lfht_is_empty(struct cds_lfht
*ht
)
1851 struct cds_lfht_node
*node
, *next
;
1855 was_online
= ht
->flavor
->read_ongoing();
1857 ht
->flavor
->thread_online();
1858 ht
->flavor
->read_lock();
1860 /* Check that the table is empty */
1861 node
= bucket_at(ht
, 0);
1863 next
= rcu_dereference(node
->next
);
1864 if (!is_bucket(next
)) {
1868 node
= clear_flag(next
);
1869 } while (!is_end(node
));
1871 ht
->flavor
->read_unlock();
1872 ht
->flavor
->thread_offline();
1878 int cds_lfht_delete_bucket(struct cds_lfht
*ht
)
1880 struct cds_lfht_node
*node
;
1881 unsigned long order
, i
, size
;
1883 /* Check that the table is empty */
1884 node
= bucket_at(ht
, 0);
1886 node
= clear_flag(node
)->next
;
1887 if (!is_bucket(node
))
1889 urcu_posix_assert(!is_removed(node
));
1890 urcu_posix_assert(!is_removal_owner(node
));
1891 } while (!is_end(node
));
1893 * size accessed without rcu_dereference because hash table is
1897 /* Internal sanity check: all nodes left should be buckets */
1898 for (i
= 0; i
< size
; i
++) {
1899 node
= bucket_at(ht
, i
);
1900 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1901 i
, i
, bit_reverse_ulong(node
->reverse_hash
));
1902 urcu_posix_assert(is_bucket(node
->next
));
1905 for (order
= cds_lfht_get_count_order_ulong(size
); (long)order
>= 0; order
--)
1906 cds_lfht_free_bucket_table(ht
, order
);
1912 void do_auto_resize_destroy_cb(struct urcu_work
*work
)
1914 struct cds_lfht
*ht
= caa_container_of(work
, struct cds_lfht
, destroy_work
);
1917 ht
->flavor
->register_thread();
1918 ret
= cds_lfht_delete_bucket(ht
);
1921 free_split_items_count(ht
);
1922 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1925 ht
->flavor
->unregister_thread();
1930 * Should only be called when no more concurrent readers nor writers can
1931 * possibly access the table.
1933 int cds_lfht_destroy(struct cds_lfht
*ht
, pthread_attr_t
**attr
)
1937 if (ht
->flags
& CDS_LFHT_AUTO_RESIZE
) {
1939 * Perform error-checking for emptiness before queuing
1940 * work, so we can return error to the caller. This runs
1941 * concurrently with ongoing resize.
1943 if (!cds_lfht_is_empty(ht
))
1945 /* Cancel ongoing resize operations. */
1946 _CMM_STORE_SHARED(ht
->in_progress_destroy
, 1);
1948 *attr
= ht
->caller_resize_attr
;
1949 ht
->caller_resize_attr
= NULL
;
1952 * Queue destroy work after prior queued resize
1953 * operations. Given there are no concurrent writers
1954 * accessing the hash table at this point, no resize
1955 * operations can be queued after this destroy work.
1957 urcu_workqueue_queue_work(cds_lfht_workqueue
,
1958 &ht
->destroy_work
, do_auto_resize_destroy_cb
);
1961 ret
= cds_lfht_delete_bucket(ht
);
1964 free_split_items_count(ht
);
1966 *attr
= ht
->caller_resize_attr
;
1967 ret
= pthread_mutex_destroy(&ht
->resize_mutex
);
1974 void cds_lfht_count_nodes(struct cds_lfht
*ht
,
1975 long *approx_before
,
1976 unsigned long *count
,
1979 struct cds_lfht_node
*node
, *next
;
1980 unsigned long nr_bucket
= 0, nr_removed
= 0;
1983 if (ht
->split_count
) {
1986 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
1987 *approx_before
+= uatomic_read(&ht
->split_count
[i
].add
);
1988 *approx_before
-= uatomic_read(&ht
->split_count
[i
].del
);
1994 /* Count non-bucket nodes in the table */
1995 node
= bucket_at(ht
, 0);
1997 next
= rcu_dereference(node
->next
);
1998 if (is_removed(next
)) {
1999 if (!is_bucket(next
))
2003 } else if (!is_bucket(next
))
2007 node
= clear_flag(next
);
2008 } while (!is_end(node
));
2009 dbg_printf("number of logically removed nodes: %lu\n", nr_removed
);
2010 dbg_printf("number of bucket nodes: %lu\n", nr_bucket
);
2012 if (ht
->split_count
) {
2015 for (i
= 0; i
< split_count_mask
+ 1; i
++) {
2016 *approx_after
+= uatomic_read(&ht
->split_count
[i
].add
);
2017 *approx_after
-= uatomic_read(&ht
->split_count
[i
].del
);
2022 /* called with resize mutex held */
2024 void _do_cds_lfht_grow(struct cds_lfht
*ht
,
2025 unsigned long old_size
, unsigned long new_size
)
2027 unsigned long old_order
, new_order
;
2029 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2030 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2031 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2032 old_size
, old_order
, new_size
, new_order
);
2033 urcu_posix_assert(new_size
> old_size
);
2034 init_table(ht
, old_order
+ 1, new_order
);
2037 /* called with resize mutex held */
2039 void _do_cds_lfht_shrink(struct cds_lfht
*ht
,
2040 unsigned long old_size
, unsigned long new_size
)
2042 unsigned long old_order
, new_order
;
2044 new_size
= max(new_size
, MIN_TABLE_SIZE
);
2045 old_order
= cds_lfht_get_count_order_ulong(old_size
);
2046 new_order
= cds_lfht_get_count_order_ulong(new_size
);
2047 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
2048 old_size
, old_order
, new_size
, new_order
);
2049 urcu_posix_assert(new_size
< old_size
);
2051 /* Remove and unlink all bucket nodes to remove. */
2052 fini_table(ht
, new_order
+ 1, old_order
);
2056 /* called with resize mutex held */
2058 void _do_cds_lfht_resize(struct cds_lfht
*ht
)
2060 unsigned long new_size
, old_size
;
2063 * Resize table, re-do if the target size has changed under us.
2066 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
))
2068 ht
->resize_initiated
= 1;
2069 old_size
= ht
->size
;
2070 new_size
= CMM_LOAD_SHARED(ht
->resize_target
);
2071 if (old_size
< new_size
)
2072 _do_cds_lfht_grow(ht
, old_size
, new_size
);
2073 else if (old_size
> new_size
)
2074 _do_cds_lfht_shrink(ht
, old_size
, new_size
);
2075 ht
->resize_initiated
= 0;
2076 /* write resize_initiated before read resize_target */
2078 } while (ht
->size
!= CMM_LOAD_SHARED(ht
->resize_target
));
2082 unsigned long resize_target_grow(struct cds_lfht
*ht
, unsigned long new_size
)
2084 return _uatomic_xchg_monotonic_increase(&ht
->resize_target
, new_size
);
2088 void resize_target_update_count(struct cds_lfht
*ht
,
2089 unsigned long count
)
2091 count
= max(count
, MIN_TABLE_SIZE
);
2092 count
= min(count
, ht
->max_nr_buckets
);
2093 uatomic_set(&ht
->resize_target
, count
);
2096 void cds_lfht_resize(struct cds_lfht
*ht
, unsigned long new_size
)
2098 resize_target_update_count(ht
, new_size
);
2099 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
2100 mutex_lock(&ht
->resize_mutex
);
2101 _do_cds_lfht_resize(ht
);
2102 mutex_unlock(&ht
->resize_mutex
);
2106 void do_resize_cb(struct urcu_work
*work
)
2108 struct resize_work
*resize_work
=
2109 caa_container_of(work
, struct resize_work
, work
);
2110 struct cds_lfht
*ht
= resize_work
->ht
;
2112 ht
->flavor
->register_thread();
2113 mutex_lock(&ht
->resize_mutex
);
2114 _do_cds_lfht_resize(ht
);
2115 mutex_unlock(&ht
->resize_mutex
);
2116 ht
->flavor
->unregister_thread();
2121 void __cds_lfht_resize_lazy_launch(struct cds_lfht
*ht
)
2123 struct resize_work
*work
;
2125 /* Store resize_target before read resize_initiated */
2127 if (!CMM_LOAD_SHARED(ht
->resize_initiated
)) {
2128 if (CMM_LOAD_SHARED(ht
->in_progress_destroy
)) {
2131 work
= malloc(sizeof(*work
));
2133 dbg_printf("error allocating resize work, bailing out\n");
2137 urcu_workqueue_queue_work(cds_lfht_workqueue
,
2138 &work
->work
, do_resize_cb
);
2139 CMM_STORE_SHARED(ht
->resize_initiated
, 1);
2144 void cds_lfht_resize_lazy_grow(struct cds_lfht
*ht
, unsigned long size
, int growth
)
2146 unsigned long target_size
= size
<< growth
;
2148 target_size
= min(target_size
, ht
->max_nr_buckets
);
2149 if (resize_target_grow(ht
, target_size
) >= target_size
)
2152 __cds_lfht_resize_lazy_launch(ht
);
2156 * We favor grow operations over shrink. A shrink operation never occurs
2157 * if a grow operation is queued for lazy execution. A grow operation
2158 * cancels any pending shrink lazy execution.
2161 void cds_lfht_resize_lazy_count(struct cds_lfht
*ht
, unsigned long size
,
2162 unsigned long count
)
2164 if (!(ht
->flags
& CDS_LFHT_AUTO_RESIZE
))
2166 count
= max(count
, MIN_TABLE_SIZE
);
2167 count
= min(count
, ht
->max_nr_buckets
);
2169 return; /* Already the right size, no resize needed */
2170 if (count
> size
) { /* lazy grow */
2171 if (resize_target_grow(ht
, count
) >= count
)
2173 } else { /* lazy shrink */
2177 s
= uatomic_cmpxchg(&ht
->resize_target
, size
, count
);
2179 break; /* no resize needed */
2181 return; /* growing is/(was just) in progress */
2183 return; /* some other thread do shrink */
2187 __cds_lfht_resize_lazy_launch(ht
);
2190 static void cds_lfht_before_fork(void *priv
__attribute__((unused
)))
2192 if (cds_lfht_workqueue_atfork_nesting
++)
2194 mutex_lock(&cds_lfht_fork_mutex
);
2195 if (!cds_lfht_workqueue
)
2197 urcu_workqueue_pause_worker(cds_lfht_workqueue
);
2200 static void cds_lfht_after_fork_parent(void *priv
__attribute__((unused
)))
2202 if (--cds_lfht_workqueue_atfork_nesting
)
2204 if (!cds_lfht_workqueue
)
2206 urcu_workqueue_resume_worker(cds_lfht_workqueue
);
2208 mutex_unlock(&cds_lfht_fork_mutex
);
2211 static void cds_lfht_after_fork_child(void *priv
__attribute__((unused
)))
2213 if (--cds_lfht_workqueue_atfork_nesting
)
2215 if (!cds_lfht_workqueue
)
2217 urcu_workqueue_create_worker(cds_lfht_workqueue
);
2219 mutex_unlock(&cds_lfht_fork_mutex
);
2222 static struct urcu_atfork cds_lfht_atfork
= {
2223 .before_fork
= cds_lfht_before_fork
,
2224 .after_fork_parent
= cds_lfht_after_fork_parent
,
2225 .after_fork_child
= cds_lfht_after_fork_child
,
2228 static void cds_lfht_init_worker(const struct rcu_flavor_struct
*flavor
)
2230 flavor
->register_rculfhash_atfork(&cds_lfht_atfork
);
2232 mutex_lock(&cds_lfht_fork_mutex
);
2233 if (!cds_lfht_workqueue
)
2234 cds_lfht_workqueue
= urcu_workqueue_create(0, -1, NULL
,
2235 NULL
, NULL
, NULL
, NULL
, NULL
, NULL
, NULL
);
2236 mutex_unlock(&cds_lfht_fork_mutex
);
2239 static void cds_lfht_exit(void)
2241 mutex_lock(&cds_lfht_fork_mutex
);
2242 if (cds_lfht_workqueue
) {
2243 urcu_workqueue_flush_queued_work(cds_lfht_workqueue
);
2244 urcu_workqueue_destroy(cds_lfht_workqueue
);
2245 cds_lfht_workqueue
= NULL
;
2247 mutex_unlock(&cds_lfht_fork_mutex
);