Fix: syscall-compat.h MacOSX support
[urcu.git] / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
1475579c 4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
abc490a1
MD
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
0dcf4847 7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
abc490a1
MD
8 *
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.
13 *
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.
18 *
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
5e28c532
MD
22 */
23
e753ff5a
MD
24/*
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,
31 * (2002), 73-82.
32 *
1475579c 33 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
1f67ba50
MD
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
39 * period.
e753ff5a
MD
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
1f67ba50
MD
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
47 * table.
48 * - The resize operation executes concurrently with
49 * add/add_unique/add_replace/remove/lookup/traversal.
e753ff5a
MD
50 * - Hash table nodes are contained within a split-ordered list. This
51 * list is ordered by incrementing reversed-bits-hash value.
1ee8f000 52 * - An index of bucket nodes is kept. These bucket nodes are the hash
1f67ba50
MD
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.
1475579c
MD
62 * - The resize operation for larger tables (and available through an
63 * API) allows both expanding and shrinking the hash table.
4c42f1b8 64 * - Split-counters are used to keep track of the number of
1475579c 65 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
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
5c4ca589
MD
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
1f67ba50 89 * linked-list (the start node is never removed) and that it does not
5c4ca589
MD
90 * contain the "removed" node anymore, even if concurrent delete/add
91 * operations are changing the structure of the list concurrently.
1f67ba50
MD
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
29e669f6
MD
95 * helping the remover unlink nodes from the list rather than to wait
96 * for it do to so.
1f67ba50
MD
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
100 * nodes.
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
112 * all buckets.
5f177b1c 113 * - synchronize_rcu is used to garbage-collect the old bucket node table.
93d46c39 114 *
7f949215 115 * Ordering Guarantees:
0f5543cb 116 *
7f949215
MD
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
67ecffc0 120 * cds_lfht_add_unique (failure).
7f949215
MD
121 *
122 * We define "read traversal" operation as any of the following
123 * group of operations
0f5543cb 124 * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
7f949215
MD
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
128 * common).
129 * - cds_lfht_first followed iteration with cds_lfht_next (and/or
130 * cds_lfht_next_duplicate, although less common).
0f5543cb 131 *
bf09adc7 132 * We define "write" operations as any of cds_lfht_add, cds_lfht_replace,
7f949215
MD
133 * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
134 *
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
141 * (failure).
142 *
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
145 * write operations.
146 *
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()
152 * operations.
153 *
154 * The following ordering guarantees are offered by this hash table:
155 *
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
158 * later write.
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
163 * some later write.
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
178 * unordered.
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
187 * with that key.
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.
199 *
200 * Progress guarantees:
201 *
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.
0f5543cb 206 *
1ee8f000 207 * Bucket node tables:
93d46c39 208 *
1ee8f000
LJ
209 * hash table hash table the last all bucket node tables
210 * order size bucket node 0 1 2 3 4 5 6(index)
93d46c39
LJ
211 * table size
212 * 0 1 1 1
213 * 1 2 1 1 1
214 * 2 4 2 1 1 2
215 * 3 8 4 1 1 2 4
216 * 4 16 8 1 1 2 4 8
217 * 5 32 16 1 1 2 4 8 16
218 * 6 64 32 1 1 2 4 8 16 32
219 *
1ee8f000 220 * When growing/shrinking, we only focus on the last bucket node table
93d46c39
LJ
221 * which size is (!order ? 1 : (1 << (order -1))).
222 *
223 * Example for growing/shrinking:
1ee8f000
LJ
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
93d46c39 226 *
1475579c 227 * A bit of ascii art explanation:
67ecffc0 228 *
1f67ba50
MD
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.
67ecffc0 231 *
1475579c 232 * This shows the nodes for a small table ordered by reversed bits:
67ecffc0 233 *
1475579c
MD
234 * bits reverse
235 * 0 000 000
236 * 4 100 001
237 * 2 010 010
238 * 6 110 011
239 * 1 001 100
240 * 5 101 101
241 * 3 011 110
242 * 7 111 111
67ecffc0
MD
243 *
244 * This shows the nodes in order of non-reversed bits, linked by
1475579c 245 * reversed-bit order.
67ecffc0 246 *
1475579c
MD
247 * order bits reverse
248 * 0 0 000 000
0adc36a8
LJ
249 * 1 | 1 001 100 <-
250 * 2 | | 2 010 010 <- |
f6fdd688 251 * | | | 3 011 110 | <- |
1475579c
MD
252 * 3 -> | | | 4 100 001 | |
253 * -> | | 5 101 101 |
254 * -> | 6 110 011
255 * -> 7 111 111
e753ff5a
MD
256 */
257
2ed95849 258#define _LGPL_SOURCE
125f41db 259#define _GNU_SOURCE
2ed95849 260#include <stdlib.h>
e0ba718a
MD
261#include <errno.h>
262#include <assert.h>
263#include <stdio.h>
abc490a1 264#include <stdint.h>
f000907d 265#include <string.h>
125f41db 266#include <sched.h>
95747f9e 267#include <unistd.h>
e0ba718a 268
15cfbec7 269#include "config.h"
a47dd11c 270#include "compat-getcpu.h"
95747f9e 271#include <urcu-pointer.h>
abc490a1 272#include <urcu-call-rcu.h>
7b17c13e 273#include <urcu-flavor.h>
a42cc659
MD
274#include <urcu/arch.h>
275#include <urcu/uatomic.h>
a42cc659 276#include <urcu/compiler.h>
abc490a1 277#include <urcu/rculfhash.h>
0b6aa001 278#include <rculfhash-internal.h>
5e28c532 279#include <stdio.h>
464a1ec9 280#include <pthread.h>
44395fb7 281
f8994aee 282/*
4c42f1b8 283 * Split-counters lazily update the global counter each 1024
f8994aee
MD
284 * addition/removal. It automatically keeps track of resize required.
285 * We use the bucket length as indicator for need to expand for small
ffa11a18 286 * tables and machines lacking per-cpu data support.
f8994aee
MD
287 */
288#define COUNT_COMMIT_ORDER 10
4ddbb355 289#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
290#define CHAIN_LEN_TARGET 1
291#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 292
cd95516d 293/*
76a73da8 294 * Define the minimum table size.
cd95516d 295 */
d0d8f9aa
LJ
296#define MIN_TABLE_ORDER 0
297#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 298
b7d619b0 299/*
1ee8f000 300 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 301 */
6083a889
MD
302#define MIN_PARTITION_PER_THREAD_ORDER 12
303#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 304
d95bd160
MD
305/*
306 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 307 * It indicates that no node must attach to the node scheduled for
b198f0fd 308 * removal, and that node garbage collection must be performed.
1ee8f000 309 * The bucket flag does not require to be updated atomically with the
d95bd160 310 * pointer, but it is added as a pointer low bit flag to save space.
1f67ba50
MD
311 * The "removal owner" flag is used to detect which of the "del"
312 * operation that has set the "removed flag" gets to return the removed
313 * node to its caller. Note that the replace operation does not need to
314 * iteract with the "removal owner" flag, because it validates that
315 * the "removed" flag is not set before performing its cmpxchg.
d95bd160 316 */
d37166c6 317#define REMOVED_FLAG (1UL << 0)
1ee8f000 318#define BUCKET_FLAG (1UL << 1)
db00ccc3
MD
319#define REMOVAL_OWNER_FLAG (1UL << 2)
320#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 321
bb7b2f26 322/* Value of the end pointer. Should not interact with flags. */
f9c80341 323#define END_VALUE NULL
bb7b2f26 324
7f52427b
MD
325/*
326 * ht_items_count: Split-counters counting the number of node addition
327 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
328 * is set at hash table creation.
329 *
330 * These are free-running counters, never reset to zero. They count the
331 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
332 * operations to update the global counter. We choose a power-of-2 value
333 * for the trigger to deal with 32 or 64-bit overflow of the counter.
334 */
df44348d 335struct ht_items_count {
860d07e8 336 unsigned long add, del;
df44348d
MD
337} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
338
7f52427b
MD
339/*
340 * rcu_resize_work: Contains arguments passed to RCU worker thread
341 * responsible for performing lazy resize.
342 */
abc490a1
MD
343struct rcu_resize_work {
344 struct rcu_head head;
14044b37 345 struct cds_lfht *ht;
abc490a1 346};
2ed95849 347
7f52427b
MD
348/*
349 * partition_resize_work: Contains arguments passed to worker threads
350 * executing the hash table resize on partitions of the hash table
351 * assigned to each processor's worker thread.
352 */
b7d619b0 353struct partition_resize_work {
1af6e26e 354 pthread_t thread_id;
b7d619b0
MD
355 struct cds_lfht *ht;
356 unsigned long i, start, len;
357 void (*fct)(struct cds_lfht *ht, unsigned long i,
358 unsigned long start, unsigned long len);
359};
360
abc490a1
MD
361/*
362 * Algorithm to reverse bits in a word by lookup table, extended to
363 * 64-bit words.
f9830efd 364 * Source:
abc490a1 365 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 366 * Originally from Public Domain.
abc490a1
MD
367 */
368
67ecffc0 369static const uint8_t BitReverseTable256[256] =
2ed95849 370{
abc490a1
MD
371#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
372#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
373#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
374 R6(0), R6(2), R6(1), R6(3)
375};
376#undef R2
377#undef R4
378#undef R6
2ed95849 379
abc490a1
MD
380static
381uint8_t bit_reverse_u8(uint8_t v)
382{
383 return BitReverseTable256[v];
384}
ab7d5fc6 385
95bc7fb9
MD
386#if (CAA_BITS_PER_LONG == 32)
387static
abc490a1
MD
388uint32_t bit_reverse_u32(uint32_t v)
389{
67ecffc0
MD
390 return ((uint32_t) bit_reverse_u8(v) << 24) |
391 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
392 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
abc490a1 393 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849 394}
95bc7fb9
MD
395#else
396static
abc490a1 397uint64_t bit_reverse_u64(uint64_t v)
2ed95849 398{
67ecffc0
MD
399 return ((uint64_t) bit_reverse_u8(v) << 56) |
400 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
abc490a1
MD
401 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
402 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
67ecffc0
MD
403 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
404 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
abc490a1
MD
405 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
406 ((uint64_t) bit_reverse_u8(v >> 56));
407}
95bc7fb9 408#endif
abc490a1
MD
409
410static
411unsigned long bit_reverse_ulong(unsigned long v)
412{
413#if (CAA_BITS_PER_LONG == 32)
414 return bit_reverse_u32(v);
415#else
416 return bit_reverse_u64(v);
417#endif
418}
419
f9830efd 420/*
24365af7
MD
421 * fls: returns the position of the most significant bit.
422 * Returns 0 if no bit is set, else returns the position of the most
423 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 424 */
24365af7
MD
425#if defined(__i386) || defined(__x86_64)
426static inline
427unsigned int fls_u32(uint32_t x)
f9830efd 428{
24365af7
MD
429 int r;
430
e1789ce2 431 __asm__ ("bsrl %1,%0\n\t"
24365af7
MD
432 "jnz 1f\n\t"
433 "movl $-1,%0\n\t"
434 "1:\n\t"
435 : "=r" (r) : "rm" (x));
436 return r + 1;
437}
438#define HAS_FLS_U32
439#endif
440
441#if defined(__x86_64)
442static inline
443unsigned int fls_u64(uint64_t x)
444{
445 long r;
446
e1789ce2 447 __asm__ ("bsrq %1,%0\n\t"
24365af7
MD
448 "jnz 1f\n\t"
449 "movq $-1,%0\n\t"
450 "1:\n\t"
451 : "=r" (r) : "rm" (x));
452 return r + 1;
453}
454#define HAS_FLS_U64
455#endif
456
457#ifndef HAS_FLS_U64
458static __attribute__((unused))
459unsigned int fls_u64(uint64_t x)
460{
461 unsigned int r = 64;
462
463 if (!x)
464 return 0;
465
466 if (!(x & 0xFFFFFFFF00000000ULL)) {
467 x <<= 32;
468 r -= 32;
469 }
470 if (!(x & 0xFFFF000000000000ULL)) {
471 x <<= 16;
472 r -= 16;
473 }
474 if (!(x & 0xFF00000000000000ULL)) {
475 x <<= 8;
476 r -= 8;
477 }
478 if (!(x & 0xF000000000000000ULL)) {
479 x <<= 4;
480 r -= 4;
481 }
482 if (!(x & 0xC000000000000000ULL)) {
483 x <<= 2;
484 r -= 2;
485 }
486 if (!(x & 0x8000000000000000ULL)) {
487 x <<= 1;
488 r -= 1;
489 }
490 return r;
491}
492#endif
493
494#ifndef HAS_FLS_U32
495static __attribute__((unused))
496unsigned int fls_u32(uint32_t x)
497{
498 unsigned int r = 32;
f9830efd 499
24365af7
MD
500 if (!x)
501 return 0;
502 if (!(x & 0xFFFF0000U)) {
503 x <<= 16;
504 r -= 16;
505 }
506 if (!(x & 0xFF000000U)) {
507 x <<= 8;
508 r -= 8;
509 }
510 if (!(x & 0xF0000000U)) {
511 x <<= 4;
512 r -= 4;
513 }
514 if (!(x & 0xC0000000U)) {
515 x <<= 2;
516 r -= 2;
517 }
518 if (!(x & 0x80000000U)) {
519 x <<= 1;
520 r -= 1;
521 }
522 return r;
523}
524#endif
525
5bc6b66f 526unsigned int cds_lfht_fls_ulong(unsigned long x)
f9830efd 527{
6887cc5e 528#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
529 return fls_u32(x);
530#else
531 return fls_u64(x);
532#endif
533}
f9830efd 534
920f8ef6
LJ
535/*
536 * Return the minimum order for which x <= (1UL << order).
537 * Return -1 if x is 0.
538 */
5bc6b66f 539int cds_lfht_get_count_order_u32(uint32_t x)
24365af7 540{
920f8ef6
LJ
541 if (!x)
542 return -1;
24365af7 543
920f8ef6 544 return fls_u32(x - 1);
24365af7
MD
545}
546
920f8ef6
LJ
547/*
548 * Return the minimum order for which x <= (1UL << order).
549 * Return -1 if x is 0.
550 */
5bc6b66f 551int cds_lfht_get_count_order_ulong(unsigned long x)
24365af7 552{
920f8ef6
LJ
553 if (!x)
554 return -1;
24365af7 555
5bc6b66f 556 return cds_lfht_fls_ulong(x - 1);
f9830efd
MD
557}
558
559static
ab65b890 560void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 561
f8994aee 562static
4105056a 563void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
564 unsigned long count);
565
df44348d 566static long nr_cpus_mask = -1;
4c42f1b8 567static long split_count_mask = -1;
e53ab1eb 568static int split_count_order = -1;
4c42f1b8 569
4ddbb355 570#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
571static void ht_init_nr_cpus_mask(void)
572{
573 long maxcpus;
574
575 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
576 if (maxcpus <= 0) {
577 nr_cpus_mask = -2;
578 return;
579 }
580 /*
581 * round up number of CPUs to next power of two, so we
582 * can use & for modulo.
583 */
5bc6b66f 584 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
585 nr_cpus_mask = maxcpus - 1;
586}
4ddbb355
LJ
587#else /* #if defined(HAVE_SYSCONF) */
588static void ht_init_nr_cpus_mask(void)
589{
590 nr_cpus_mask = -2;
591}
592#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
593
594static
5afadd12 595void alloc_split_items_count(struct cds_lfht *ht)
df44348d 596{
4c42f1b8
LJ
597 if (nr_cpus_mask == -1) {
598 ht_init_nr_cpus_mask();
4ddbb355
LJ
599 if (nr_cpus_mask < 0)
600 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
601 else
602 split_count_mask = nr_cpus_mask;
e53ab1eb
MD
603 split_count_order =
604 cds_lfht_get_count_order_ulong(split_count_mask + 1);
df44348d 605 }
4c42f1b8 606
4ddbb355 607 assert(split_count_mask >= 0);
5afadd12
LJ
608
609 if (ht->flags & CDS_LFHT_ACCOUNTING) {
95bc7fb9
MD
610 ht->split_count = calloc(split_count_mask + 1,
611 sizeof(struct ht_items_count));
5afadd12
LJ
612 assert(ht->split_count);
613 } else {
614 ht->split_count = NULL;
615 }
df44348d
MD
616}
617
618static
5afadd12 619void free_split_items_count(struct cds_lfht *ht)
df44348d 620{
5afadd12 621 poison_free(ht->split_count);
df44348d
MD
622}
623
624static
14360f1c 625int ht_get_split_count_index(unsigned long hash)
df44348d
MD
626{
627 int cpu;
628
4c42f1b8 629 assert(split_count_mask >= 0);
a47dd11c 630 cpu = urcu_sched_getcpu();
8ed51e04 631 if (caa_unlikely(cpu < 0))
14360f1c 632 return hash & split_count_mask;
df44348d 633 else
4c42f1b8 634 return cpu & split_count_mask;
df44348d
MD
635}
636
637static
14360f1c 638void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 639{
4c42f1b8
LJ
640 unsigned long split_count;
641 int index;
314558bf 642 long count;
df44348d 643
8ed51e04 644 if (caa_unlikely(!ht->split_count))
3171717f 645 return;
14360f1c 646 index = ht_get_split_count_index(hash);
4c42f1b8 647 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
648 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
649 return;
650 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
651
652 dbg_printf("add split count %lu\n", split_count);
653 count = uatomic_add_return(&ht->count,
654 1UL << COUNT_COMMIT_ORDER);
4c299dcb 655 if (caa_likely(count & (count - 1)))
314558bf
MD
656 return;
657 /* Only if global count is power of 2 */
658
659 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
660 return;
661 dbg_printf("add set global %ld\n", count);
662 cds_lfht_resize_lazy_count(ht, size,
663 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
664}
665
666static
14360f1c 667void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 668{
4c42f1b8
LJ
669 unsigned long split_count;
670 int index;
314558bf 671 long count;
df44348d 672
8ed51e04 673 if (caa_unlikely(!ht->split_count))
3171717f 674 return;
14360f1c 675 index = ht_get_split_count_index(hash);
4c42f1b8 676 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
677 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
678 return;
679 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
680
681 dbg_printf("del split count %lu\n", split_count);
682 count = uatomic_add_return(&ht->count,
683 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 684 if (caa_likely(count & (count - 1)))
314558bf
MD
685 return;
686 /* Only if global count is power of 2 */
687
688 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
689 return;
690 dbg_printf("del set global %ld\n", count);
691 /*
692 * Don't shrink table if the number of nodes is below a
693 * certain threshold.
694 */
695 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
696 return;
697 cds_lfht_resize_lazy_count(ht, size,
698 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
699}
700
f9830efd 701static
4105056a 702void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 703{
f8994aee
MD
704 unsigned long count;
705
b8af5011
MD
706 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
707 return;
f8994aee
MD
708 count = uatomic_read(&ht->count);
709 /*
710 * Use bucket-local length for small table expand and for
711 * environments lacking per-cpu data support.
712 */
e53ab1eb 713 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
f8994aee 714 return;
24365af7 715 if (chain_len > 100)
f0c29ed7 716 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 717 chain_len);
e53ab1eb
MD
718 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
719 int growth;
720
721 /*
722 * Ideal growth calculated based on chain length.
723 */
724 growth = cds_lfht_get_count_order_u32(chain_len
725 - (CHAIN_LEN_TARGET - 1));
726 if ((ht->flags & CDS_LFHT_ACCOUNTING)
727 && (size << growth)
728 >= (1UL << (COUNT_COMMIT_ORDER
729 + split_count_order))) {
730 /*
731 * If ideal growth expands the hash table size
732 * beyond the "small hash table" sizes, use the
733 * maximum small hash table size to attempt
734 * expanding the hash table. This only applies
735 * when node accounting is available, otherwise
736 * the chain length is used to expand the hash
737 * table in every case.
738 */
739 growth = COUNT_COMMIT_ORDER + split_count_order
740 - cds_lfht_get_count_order_ulong(size);
741 if (growth <= 0)
742 return;
743 }
744 cds_lfht_resize_lazy_grow(ht, size, growth);
745 }
f9830efd
MD
746}
747
abc490a1 748static
14044b37 749struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 750{
14044b37 751 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
752}
753
754static
14044b37 755int is_removed(struct cds_lfht_node *node)
abc490a1 756{
d37166c6 757 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
758}
759
f5596c94 760static
1ee8f000 761int is_bucket(struct cds_lfht_node *node)
f5596c94 762{
1ee8f000 763 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
764}
765
766static
1ee8f000 767struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 768{
1ee8f000 769 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 770}
bb7b2f26 771
db00ccc3
MD
772static
773int is_removal_owner(struct cds_lfht_node *node)
774{
775 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
776}
777
778static
779struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
780{
781 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
782}
783
71bb3aca
MD
784static
785struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
786{
787 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
788}
789
bb7b2f26
MD
790static
791struct cds_lfht_node *get_end(void)
792{
793 return (struct cds_lfht_node *) END_VALUE;
794}
795
796static
797int is_end(struct cds_lfht_node *node)
798{
799 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
800}
801
abc490a1 802static
ab65b890
LJ
803unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
804 unsigned long v)
abc490a1
MD
805{
806 unsigned long old1, old2;
807
808 old1 = uatomic_read(ptr);
809 do {
810 old2 = old1;
811 if (old2 >= v)
f9830efd 812 return old2;
abc490a1 813 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 814 return old2;
abc490a1
MD
815}
816
48f1b16d
LJ
817static
818void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
819{
0b6aa001 820 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
821}
822
823/*
824 * cds_lfht_free_bucket_table() should be called with decreasing order.
825 * When cds_lfht_free_bucket_table(0) is called, it means the whole
826 * lfht is destroyed.
827 */
828static
829void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
830{
0b6aa001 831 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
832}
833
9d72a73f
LJ
834static inline
835struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 836{
0b6aa001 837 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
838}
839
9d72a73f
LJ
840static inline
841struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
842 unsigned long hash)
843{
844 assert(size > 0);
845 return bucket_at(ht, hash & (size - 1));
846}
847
273399de
MD
848/*
849 * Remove all logically deleted nodes from a bucket up to a certain node key.
850 */
851static
1ee8f000 852void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 853{
14044b37 854 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 855
1ee8f000
LJ
856 assert(!is_bucket(bucket));
857 assert(!is_removed(bucket));
2f943cd7 858 assert(!is_removal_owner(bucket));
1ee8f000 859 assert(!is_bucket(node));
c90201ac 860 assert(!is_removed(node));
2f943cd7 861 assert(!is_removal_owner(node));
273399de 862 for (;;) {
1ee8f000
LJ
863 iter_prev = bucket;
864 /* We can always skip the bucket node initially */
04db56f8 865 iter = rcu_dereference(iter_prev->next);
b4cb483f 866 assert(!is_removed(iter));
2f943cd7 867 assert(!is_removal_owner(iter));
04db56f8 868 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 869 /*
1ee8f000 870 * We should never be called with bucket (start of chain)
bd4db153
MD
871 * and logically removed node (end of path compression
872 * marker) being the actual same node. This would be a
873 * bug in the algorithm implementation.
874 */
1ee8f000 875 assert(bucket != node);
273399de 876 for (;;) {
8ed51e04 877 if (caa_unlikely(is_end(iter)))
f9c80341 878 return;
04db56f8 879 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 880 return;
04db56f8 881 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 882 if (caa_likely(is_removed(next)))
273399de 883 break;
b453eae1 884 iter_prev = clear_flag(iter);
273399de
MD
885 iter = next;
886 }
b198f0fd 887 assert(!is_removed(iter));
2f943cd7 888 assert(!is_removal_owner(iter));
1ee8f000
LJ
889 if (is_bucket(iter))
890 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
891 else
892 new_next = clear_flag(next);
04db56f8 893 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
894 }
895}
896
9357c415
MD
897static
898int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
899 struct cds_lfht_node *old_node,
3fb86f26 900 struct cds_lfht_node *old_next,
9357c415
MD
901 struct cds_lfht_node *new_node)
902{
04db56f8 903 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
904
905 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 906 return -ENOENT;
9357c415
MD
907
908 assert(!is_removed(old_node));
2f943cd7 909 assert(!is_removal_owner(old_node));
1ee8f000 910 assert(!is_bucket(old_node));
9357c415 911 assert(!is_removed(new_node));
2f943cd7 912 assert(!is_removal_owner(new_node));
1ee8f000 913 assert(!is_bucket(new_node));
9357c415 914 assert(new_node != old_node);
3fb86f26 915 for (;;) {
9357c415 916 /* Insert after node to be replaced */
9357c415
MD
917 if (is_removed(old_next)) {
918 /*
919 * Too late, the old node has been removed under us
920 * between lookup and replace. Fail.
921 */
7801dadd 922 return -ENOENT;
9357c415 923 }
feda2722
LJ
924 assert(old_next == clear_flag(old_next));
925 assert(new_node != old_next);
71bb3aca
MD
926 /*
927 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
928 * flag. It is either set atomically at the same time
929 * (replace) or after (del).
930 */
931 assert(!is_removal_owner(old_next));
feda2722 932 new_node->next = old_next;
9357c415
MD
933 /*
934 * Here is the whole trick for lock-free replace: we add
935 * the replacement node _after_ the node we want to
936 * replace by atomically setting its next pointer at the
937 * same time we set its removal flag. Given that
938 * the lookups/get next use an iterator aware of the
939 * next pointer, they will either skip the old node due
940 * to the removal flag and see the new node, or use
941 * the old node, but will not see the new one.
db00ccc3
MD
942 * This is a replacement of a node with another node
943 * that has the same value: we are therefore not
71bb3aca
MD
944 * removing a value from the hash table. We set both the
945 * REMOVED and REMOVAL_OWNER flags atomically so we own
946 * the node after successful cmpxchg.
9357c415 947 */
04db56f8 948 ret_next = uatomic_cmpxchg(&old_node->next,
71bb3aca 949 old_next, flag_removed_or_removal_owner(new_node));
3fb86f26 950 if (ret_next == old_next)
7801dadd 951 break; /* We performed the replacement. */
3fb86f26
LJ
952 old_next = ret_next;
953 }
9357c415 954
9357c415
MD
955 /*
956 * Ensure that the old node is not visible to readers anymore:
957 * lookup for the node, and remove it (along with any other
958 * logically removed node) if found.
959 */
04db56f8
LJ
960 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
961 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 962
a85eff52 963 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
7801dadd 964 return 0;
9357c415
MD
965}
966
83beee94
MD
967/*
968 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
969 * mode. A NULL unique_ret allows creation of duplicate keys.
970 */
abc490a1 971static
83beee94 972void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 973 unsigned long hash,
0422d92c 974 cds_lfht_match_fct match,
996ff57c 975 const void *key,
83beee94
MD
976 unsigned long size,
977 struct cds_lfht_node *node,
978 struct cds_lfht_iter *unique_ret,
1ee8f000 979 int bucket_flag)
abc490a1 980{
14044b37 981 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 982 *return_node;
04db56f8 983 struct cds_lfht_node *bucket;
abc490a1 984
1ee8f000 985 assert(!is_bucket(node));
c90201ac 986 assert(!is_removed(node));
2f943cd7 987 assert(!is_removal_owner(node));
91a75cc5 988 bucket = lookup_bucket(ht, size, hash);
abc490a1 989 for (;;) {
adc0de68 990 uint32_t chain_len = 0;
abc490a1 991
11519af6
MD
992 /*
993 * iter_prev points to the non-removed node prior to the
994 * insert location.
11519af6 995 */
04db56f8 996 iter_prev = bucket;
1ee8f000 997 /* We can always skip the bucket node initially */
04db56f8
LJ
998 iter = rcu_dereference(iter_prev->next);
999 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 1000 for (;;) {
8ed51e04 1001 if (caa_unlikely(is_end(iter)))
273399de 1002 goto insert;
04db56f8 1003 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 1004 goto insert;
238cc06e 1005
1ee8f000
LJ
1006 /* bucket node is the first node of the identical-hash-value chain */
1007 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 1008 goto insert;
238cc06e 1009
04db56f8 1010 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 1011 if (caa_unlikely(is_removed(next)))
9dba85be 1012 goto gc_node;
238cc06e
LJ
1013
1014 /* uniquely add */
83beee94 1015 if (unique_ret
1ee8f000 1016 && !is_bucket(next)
04db56f8 1017 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
1018 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1019
1020 /*
1021 * uniquely adding inserts the node as the first
1022 * node of the identical-hash-value node chain.
1023 *
1024 * This semantic ensures no duplicated keys
1025 * should ever be observable in the table
1f67ba50
MD
1026 * (including traversing the table node by
1027 * node by forward iterations)
238cc06e 1028 */
04db56f8 1029 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
1030 if (!d_iter.node)
1031 goto insert;
1032
1033 *unique_ret = d_iter;
83beee94 1034 return;
48ed1c18 1035 }
238cc06e 1036
11519af6 1037 /* Only account for identical reverse hash once */
04db56f8 1038 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1039 && !is_bucket(next))
4105056a 1040 check_resize(ht, size, ++chain_len);
11519af6 1041 iter_prev = clear_flag(iter);
273399de 1042 iter = next;
abc490a1 1043 }
48ed1c18 1044
273399de 1045 insert:
7ec59d3b 1046 assert(node != clear_flag(iter));
11519af6 1047 assert(!is_removed(iter_prev));
2f943cd7 1048 assert(!is_removal_owner(iter_prev));
c90201ac 1049 assert(!is_removed(iter));
2f943cd7 1050 assert(!is_removal_owner(iter));
f000907d 1051 assert(iter_prev != node);
1ee8f000 1052 if (!bucket_flag)
04db56f8 1053 node->next = clear_flag(iter);
f9c80341 1054 else
1ee8f000
LJ
1055 node->next = flag_bucket(clear_flag(iter));
1056 if (is_bucket(iter))
1057 new_node = flag_bucket(node);
f5596c94
MD
1058 else
1059 new_node = node;
04db56f8 1060 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1061 new_node) != iter) {
273399de 1062 continue; /* retry */
48ed1c18 1063 } else {
83beee94 1064 return_node = node;
960c9e4f 1065 goto end;
48ed1c18
MD
1066 }
1067
9dba85be
MD
1068 gc_node:
1069 assert(!is_removed(iter));
2f943cd7 1070 assert(!is_removal_owner(iter));
1ee8f000
LJ
1071 if (is_bucket(iter))
1072 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1073 else
1074 new_next = clear_flag(next);
04db56f8 1075 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1076 /* retry */
464a1ec9 1077 }
9357c415 1078end:
83beee94
MD
1079 if (unique_ret) {
1080 unique_ret->node = return_node;
1081 /* unique_ret->next left unset, never used. */
1082 }
abc490a1 1083}
464a1ec9 1084
abc490a1 1085static
860d07e8 1086int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1087 struct cds_lfht_node *node)
abc490a1 1088{
db00ccc3 1089 struct cds_lfht_node *bucket, *next;
5e28c532 1090
9357c415 1091 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1092 return -ENOENT;
9357c415 1093
7ec59d3b 1094 /* logically delete the node */
1ee8f000 1095 assert(!is_bucket(node));
c90201ac 1096 assert(!is_removed(node));
db00ccc3 1097 assert(!is_removal_owner(node));
48ed1c18 1098
db00ccc3
MD
1099 /*
1100 * We are first checking if the node had previously been
1101 * logically removed (this check is not atomic with setting the
1102 * logical removal flag). Return -ENOENT if the node had
1103 * previously been removed.
1104 */
a85eff52 1105 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
db00ccc3
MD
1106 if (caa_unlikely(is_removed(next)))
1107 return -ENOENT;
b65ec430 1108 assert(!is_bucket(next));
196f4fab
MD
1109 /*
1110 * The del operation semantic guarantees a full memory barrier
1111 * before the uatomic_or atomic commit of the deletion flag.
1112 */
1113 cmm_smp_mb__before_uatomic_or();
db00ccc3
MD
1114 /*
1115 * We set the REMOVED_FLAG unconditionally. Note that there may
1116 * be more than one concurrent thread setting this flag.
1117 * Knowing which wins the race will be known after the garbage
1118 * collection phase, stay tuned!
1119 */
1120 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1121 /* We performed the (logical) deletion. */
7ec59d3b
MD
1122
1123 /*
1124 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1125 * the node, and remove it (along with any other logically removed node)
1126 * if found.
11519af6 1127 */
04db56f8
LJ
1128 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1129 _cds_lfht_gc_bucket(bucket, node);
743f9143 1130
a85eff52 1131 assert(is_removed(CMM_LOAD_SHARED(node->next)));
db00ccc3
MD
1132 /*
1133 * Last phase: atomically exchange node->next with a version
1134 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1135 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1136 * the node and win the removal race.
1137 * It is interesting to note that all "add" paths are forbidden
1138 * to change the next pointer starting from the point where the
1139 * REMOVED_FLAG is set, so here using a read, followed by a
1140 * xchg() suffice to guarantee that the xchg() will ever only
1141 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1142 * was already set).
1143 */
1144 if (!is_removal_owner(uatomic_xchg(&node->next,
1145 flag_removal_owner(node->next))))
1146 return 0;
1147 else
1148 return -ENOENT;
abc490a1 1149}
2ed95849 1150
b7d619b0
MD
1151static
1152void *partition_resize_thread(void *arg)
1153{
1154 struct partition_resize_work *work = arg;
1155
7b17c13e 1156 work->ht->flavor->register_thread();
b7d619b0 1157 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1158 work->ht->flavor->unregister_thread();
b7d619b0
MD
1159 return NULL;
1160}
1161
1162static
1163void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1164 unsigned long len,
1165 void (*fct)(struct cds_lfht *ht, unsigned long i,
1166 unsigned long start, unsigned long len))
1167{
e54ec2f5 1168 unsigned long partition_len, start = 0;
b7d619b0 1169 struct partition_resize_work *work;
6083a889
MD
1170 int thread, ret;
1171 unsigned long nr_threads;
b7d619b0 1172
d7f3ba4c
EW
1173 assert(nr_cpus_mask != -1);
1174 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1175 goto fallback;
1176
6083a889
MD
1177 /*
1178 * Note: nr_cpus_mask + 1 is always power of 2.
1179 * We spawn just the number of threads we need to satisfy the minimum
1180 * partition size, up to the number of CPUs in the system.
1181 */
91452a6a
MD
1182 if (nr_cpus_mask > 0) {
1183 nr_threads = min(nr_cpus_mask + 1,
1184 len >> MIN_PARTITION_PER_THREAD_ORDER);
1185 } else {
1186 nr_threads = 1;
1187 }
5bc6b66f 1188 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1189 work = calloc(nr_threads, sizeof(*work));
7c75d498
EW
1190 if (!work) {
1191 dbg_printf("error allocating for resize, single-threading\n");
1192 goto fallback;
1193 }
6083a889
MD
1194 for (thread = 0; thread < nr_threads; thread++) {
1195 work[thread].ht = ht;
1196 work[thread].i = i;
1197 work[thread].len = partition_len;
1198 work[thread].start = thread * partition_len;
1199 work[thread].fct = fct;
1af6e26e 1200 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1201 partition_resize_thread, &work[thread]);
e54ec2f5
EW
1202 if (ret == EAGAIN) {
1203 /*
1204 * Out of resources: wait and join the threads
1205 * we've created, then handle leftovers.
1206 */
1207 dbg_printf("error spawning for resize, single-threading\n");
1208 start = work[thread].start;
1209 len -= start;
1210 nr_threads = thread;
1211 break;
1212 }
b7d619b0
MD
1213 assert(!ret);
1214 }
6083a889 1215 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1216 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1217 assert(!ret);
1218 }
1219 free(work);
e54ec2f5
EW
1220
1221 /*
1222 * A pthread_create failure above will either lead in us having
1223 * no threads to join or starting at a non-zero offset,
1224 * fallback to single thread processing of leftovers.
1225 */
1226 if (start == 0 && nr_threads > 0)
1227 return;
7c75d498
EW
1228fallback:
1229 ht->flavor->thread_online();
e54ec2f5 1230 fct(ht, i, start, len);
7c75d498 1231 ht->flavor->thread_offline();
b7d619b0
MD
1232}
1233
e8de508e
MD
1234/*
1235 * Holding RCU read lock to protect _cds_lfht_add against memory
1236 * reclaim that could be performed by other call_rcu worker threads (ABA
1237 * problem).
9ee0fc9a 1238 *
b7d619b0 1239 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1240 * many worker threads, based on the number of CPUs available in the system.
1241 * This should therefore take care of not having the expand lagging behind too
1242 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1243 * schedule bucket node population fairly with insertions.
e8de508e 1244 */
4105056a 1245static
b7d619b0
MD
1246void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1247 unsigned long start, unsigned long len)
4105056a 1248{
9d72a73f 1249 unsigned long j, size = 1UL << (i - 1);
4105056a 1250
d0d8f9aa 1251 assert(i > MIN_TABLE_ORDER);
7b17c13e 1252 ht->flavor->read_lock();
9d72a73f
LJ
1253 for (j = size + start; j < size + start + len; j++) {
1254 struct cds_lfht_node *new_node = bucket_at(ht, j);
1255
1256 assert(j >= size && j < (size << 1));
1257 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1258 i, j, j);
1259 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1260 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1261 }
7b17c13e 1262 ht->flavor->read_unlock();
b7d619b0
MD
1263}
1264
1265static
1266void init_table_populate(struct cds_lfht *ht, unsigned long i,
1267 unsigned long len)
1268{
b7d619b0 1269 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1270}
1271
abc490a1 1272static
4105056a 1273void init_table(struct cds_lfht *ht,
93d46c39 1274 unsigned long first_order, unsigned long last_order)
24365af7 1275{
93d46c39 1276 unsigned long i;
24365af7 1277
93d46c39
LJ
1278 dbg_printf("init table: first_order %lu last_order %lu\n",
1279 first_order, last_order);
d0d8f9aa 1280 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1281 for (i = first_order; i <= last_order; i++) {
4105056a 1282 unsigned long len;
24365af7 1283
4f6e90b7 1284 len = 1UL << (i - 1);
f0c29ed7 1285 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1286
1287 /* Stop expand if the resize target changes under us */
7b3893e4 1288 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1289 break;
1290
48f1b16d 1291 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1292
4105056a 1293 /*
1ee8f000
LJ
1294 * Set all bucket nodes reverse hash values for a level and
1295 * link all bucket nodes into the table.
4105056a 1296 */
dc1da8f6 1297 init_table_populate(ht, i, len);
4105056a 1298
f9c80341
MD
1299 /*
1300 * Update table size.
1301 */
1302 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1303 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1304
4f6e90b7 1305 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1306 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1307 break;
1308 }
1309}
1310
e8de508e
MD
1311/*
1312 * Holding RCU read lock to protect _cds_lfht_remove against memory
1313 * reclaim that could be performed by other call_rcu worker threads (ABA
1314 * problem).
1315 * For a single level, we logically remove and garbage collect each node.
1316 *
1317 * As a design choice, we perform logical removal and garbage collection on a
1318 * node-per-node basis to simplify this algorithm. We also assume keeping good
1319 * cache locality of the operation would overweight possible performance gain
1320 * that could be achieved by batching garbage collection for multiple levels.
1321 * However, this would have to be justified by benchmarks.
1322 *
1323 * Concurrent removal and add operations are helping us perform garbage
1324 * collection of logically removed nodes. We guarantee that all logically
1325 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1326 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e 1327 *
1f67ba50
MD
1328 * Logical removal and garbage collection can therefore be done in batch
1329 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1330 *
b7d619b0
MD
1331 * When we reach a certain length, we can split this removal over many worker
1332 * threads, based on the number of CPUs available in the system. This should
1333 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1334 * updater threads actively inserting into the hash table.
e8de508e 1335 */
4105056a 1336static
b7d619b0
MD
1337void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1338 unsigned long start, unsigned long len)
4105056a 1339{
9d72a73f 1340 unsigned long j, size = 1UL << (i - 1);
4105056a 1341
d0d8f9aa 1342 assert(i > MIN_TABLE_ORDER);
7b17c13e 1343 ht->flavor->read_lock();
9d72a73f 1344 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1345 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1346 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1347
1348 assert(j >= size && j < (size << 1));
1349 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1350 i, j, j);
2e2ce1e9
LJ
1351 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1352 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1353 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1354 }
7b17c13e 1355 ht->flavor->read_unlock();
b7d619b0
MD
1356}
1357
1358static
1359void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1360{
b7d619b0 1361 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1362}
1363
61adb337
MD
1364/*
1365 * fini_table() is never called for first_order == 0, which is why
1366 * free_by_rcu_order == 0 can be used as criterion to know if free must
1367 * be called.
1368 */
1475579c 1369static
4105056a 1370void fini_table(struct cds_lfht *ht,
93d46c39 1371 unsigned long first_order, unsigned long last_order)
1475579c 1372{
93d46c39 1373 long i;
48f1b16d 1374 unsigned long free_by_rcu_order = 0;
1475579c 1375
93d46c39
LJ
1376 dbg_printf("fini table: first_order %lu last_order %lu\n",
1377 first_order, last_order);
d0d8f9aa 1378 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1379 for (i = last_order; i >= first_order; i--) {
4105056a 1380 unsigned long len;
1475579c 1381
4f6e90b7 1382 len = 1UL << (i - 1);
e15df1cc 1383 dbg_printf("fini order %ld len: %lu\n", i, len);
4105056a 1384
4d676753 1385 /* Stop shrink if the resize target changes under us */
7b3893e4 1386 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1387 break;
1388
1389 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1390 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1391
1392 /*
1393 * We need to wait for all add operations to reach Q.S. (and
1394 * thus use the new table for lookups) before we can start
1ee8f000 1395 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1396 * return a logically removed node as insert position.
1397 */
7b17c13e 1398 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1399 if (free_by_rcu_order)
1400 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1401
21263e21 1402 /*
1ee8f000
LJ
1403 * Set "removed" flag in bucket nodes about to be removed.
1404 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1405 * Concurrent add/remove operation are helping us doing
1406 * the gc.
21263e21 1407 */
4105056a
MD
1408 remove_table(ht, i, len);
1409
48f1b16d 1410 free_by_rcu_order = i;
4105056a
MD
1411
1412 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1413 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1414 break;
1415 }
0d14ceb2 1416
48f1b16d 1417 if (free_by_rcu_order) {
7b17c13e 1418 ht->flavor->update_synchronize_rcu();
48f1b16d 1419 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1420 }
1475579c
MD
1421}
1422
ff0d69de 1423static
1ee8f000 1424void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1425{
04db56f8 1426 struct cds_lfht_node *prev, *node;
9d72a73f 1427 unsigned long order, len, i;
ff0d69de 1428
48f1b16d 1429 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1430
9d72a73f
LJ
1431 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1432 node = bucket_at(ht, 0);
1433 node->next = flag_bucket(get_end());
1434 node->reverse_hash = 0;
ff0d69de 1435
5bc6b66f 1436 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
ff0d69de 1437 len = 1UL << (order - 1);
48f1b16d 1438 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1439
9d72a73f
LJ
1440 for (i = 0; i < len; i++) {
1441 /*
1442 * Now, we are trying to init the node with the
1443 * hash=(len+i) (which is also a bucket with the
1444 * index=(len+i)) and insert it into the hash table,
1445 * so this node has to be inserted after the bucket
1446 * with the index=(len+i)&(len-1)=i. And because there
1447 * is no other non-bucket node nor bucket node with
1448 * larger index/hash inserted, so the bucket node
1449 * being inserted should be inserted directly linked
1450 * after the bucket node with index=i.
1451 */
1452 prev = bucket_at(ht, i);
1453 node = bucket_at(ht, len + i);
ff0d69de 1454
1ee8f000 1455 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1456 order, len + i, len + i);
1457 node->reverse_hash = bit_reverse_ulong(len + i);
1458
1459 /* insert after prev */
1460 assert(is_bucket(prev->next));
ff0d69de 1461 node->next = prev->next;
1ee8f000 1462 prev->next = flag_bucket(node);
ff0d69de
LJ
1463 }
1464 }
1465}
1466
0422d92c 1467struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1468 unsigned long min_nr_alloc_buckets,
747d725c 1469 unsigned long max_nr_buckets,
b8af5011 1470 int flags,
0b6aa001 1471 const struct cds_lfht_mm_type *mm,
7b17c13e 1472 const struct rcu_flavor_struct *flavor,
b7d619b0 1473 pthread_attr_t *attr)
abc490a1 1474{
14044b37 1475 struct cds_lfht *ht;
24365af7 1476 unsigned long order;
abc490a1 1477
0722081a
LJ
1478 /* min_nr_alloc_buckets must be power of two */
1479 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1480 return NULL;
747d725c 1481
8129be4e 1482 /* init_size must be power of two */
5488222b 1483 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1484 return NULL;
747d725c 1485
c1888f3a
MD
1486 /*
1487 * Memory management plugin default.
1488 */
1489 if (!mm) {
5a2141a7
MD
1490 if (CAA_BITS_PER_LONG > 32
1491 && max_nr_buckets
c1888f3a
MD
1492 && max_nr_buckets <= (1ULL << 32)) {
1493 /*
1494 * For 64-bit architectures, with max number of
1495 * buckets small enough not to use the entire
1496 * 64-bit memory mapping space (and allowing a
1497 * fair number of hash table instances), use the
1498 * mmap allocator, which is faster than the
1499 * order allocator.
1500 */
1501 mm = &cds_lfht_mm_mmap;
1502 } else {
1503 /*
1504 * The fallback is to use the order allocator.
1505 */
1506 mm = &cds_lfht_mm_order;
1507 }
1508 }
1509
0b6aa001
LJ
1510 /* max_nr_buckets == 0 for order based mm means infinite */
1511 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1512 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1513
1514 /* max_nr_buckets must be power of two */
1515 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1516 return NULL;
1517
0722081a 1518 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1519 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1520 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1521 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1522
1523 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1524 assert(ht);
0b6aa001
LJ
1525 assert(ht->mm == mm);
1526 assert(ht->bucket_at == mm->bucket_at);
1527
b5d6b20f 1528 ht->flags = flags;
7b17c13e 1529 ht->flavor = flavor;
b7d619b0 1530 ht->resize_attr = attr;
5afadd12 1531 alloc_split_items_count(ht);
abc490a1
MD
1532 /* this mutex should not nest in read-side C.S. */
1533 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1534 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1535 ht->resize_target = 1UL << order;
1ee8f000 1536 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1537 ht->size = 1UL << order;
abc490a1
MD
1538 return ht;
1539}
1540
6f554439 1541void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1542 cds_lfht_match_fct match, const void *key,
6f554439 1543 struct cds_lfht_iter *iter)
2ed95849 1544{
04db56f8 1545 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1546 unsigned long reverse_hash, size;
2ed95849 1547
abc490a1 1548 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1549
7b3893e4 1550 size = rcu_dereference(ht->size);
04db56f8 1551 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1552 /* We can always skip the bucket node initially */
04db56f8 1553 node = rcu_dereference(bucket->next);
bb7b2f26 1554 node = clear_flag(node);
2ed95849 1555 for (;;) {
8ed51e04 1556 if (caa_unlikely(is_end(node))) {
96ad1112 1557 node = next = NULL;
abc490a1 1558 break;
bb7b2f26 1559 }
04db56f8 1560 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1561 node = next = NULL;
abc490a1 1562 break;
2ed95849 1563 }
04db56f8 1564 next = rcu_dereference(node->next);
7f52427b 1565 assert(node == clear_flag(node));
8ed51e04 1566 if (caa_likely(!is_removed(next))
1ee8f000 1567 && !is_bucket(next)
04db56f8 1568 && node->reverse_hash == reverse_hash
0422d92c 1569 && caa_likely(match(node, key))) {
273399de 1570 break;
2ed95849 1571 }
1b81fe1a 1572 node = clear_flag(next);
2ed95849 1573 }
a85eff52 1574 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1575 iter->node = node;
1576 iter->next = next;
abc490a1 1577}
e0ba718a 1578
0422d92c 1579void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1580 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1581{
adc0de68 1582 struct cds_lfht_node *node, *next;
a481e5ff 1583 unsigned long reverse_hash;
a481e5ff 1584
adc0de68 1585 node = iter->node;
04db56f8 1586 reverse_hash = node->reverse_hash;
adc0de68 1587 next = iter->next;
a481e5ff
MD
1588 node = clear_flag(next);
1589
1590 for (;;) {
8ed51e04 1591 if (caa_unlikely(is_end(node))) {
96ad1112 1592 node = next = NULL;
a481e5ff 1593 break;
bb7b2f26 1594 }
04db56f8 1595 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1596 node = next = NULL;
a481e5ff
MD
1597 break;
1598 }
04db56f8 1599 next = rcu_dereference(node->next);
8ed51e04 1600 if (caa_likely(!is_removed(next))
1ee8f000 1601 && !is_bucket(next)
04db56f8 1602 && caa_likely(match(node, key))) {
a481e5ff
MD
1603 break;
1604 }
1605 node = clear_flag(next);
1606 }
a85eff52 1607 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1608 iter->node = node;
1609 iter->next = next;
a481e5ff
MD
1610}
1611
4e9b9fbf
MD
1612void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1613{
1614 struct cds_lfht_node *node, *next;
1615
853395e1 1616 node = clear_flag(iter->next);
4e9b9fbf 1617 for (;;) {
8ed51e04 1618 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1619 node = next = NULL;
1620 break;
1621 }
04db56f8 1622 next = rcu_dereference(node->next);
8ed51e04 1623 if (caa_likely(!is_removed(next))
1ee8f000 1624 && !is_bucket(next)) {
4e9b9fbf
MD
1625 break;
1626 }
1627 node = clear_flag(next);
1628 }
a85eff52 1629 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
4e9b9fbf
MD
1630 iter->node = node;
1631 iter->next = next;
1632}
1633
1634void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1635{
4e9b9fbf 1636 /*
1ee8f000 1637 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1638 * first node of the linked list.
1639 */
9d72a73f 1640 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1641 cds_lfht_next(ht, iter);
1642}
1643
0422d92c
MD
1644void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1645 struct cds_lfht_node *node)
abc490a1 1646{
0422d92c 1647 unsigned long size;
ab7d5fc6 1648
709bacf9 1649 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1650 size = rcu_dereference(ht->size);
91a75cc5 1651 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1652 ht_count_add(ht, size, hash);
3eca1b8c
MD
1653}
1654
14044b37 1655struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1656 unsigned long hash,
0422d92c 1657 cds_lfht_match_fct match,
996ff57c 1658 const void *key,
48ed1c18 1659 struct cds_lfht_node *node)
3eca1b8c 1660{
0422d92c 1661 unsigned long size;
83beee94 1662 struct cds_lfht_iter iter;
3eca1b8c 1663
709bacf9 1664 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1665 size = rcu_dereference(ht->size);
91a75cc5 1666 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1667 if (iter.node == node)
14360f1c 1668 ht_count_add(ht, size, hash);
83beee94 1669 return iter.node;
2ed95849
MD
1670}
1671
9357c415 1672struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1673 unsigned long hash,
0422d92c 1674 cds_lfht_match_fct match,
996ff57c 1675 const void *key,
48ed1c18
MD
1676 struct cds_lfht_node *node)
1677{
0422d92c 1678 unsigned long size;
83beee94 1679 struct cds_lfht_iter iter;
48ed1c18 1680
709bacf9 1681 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1682 size = rcu_dereference(ht->size);
83beee94 1683 for (;;) {
91a75cc5 1684 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1685 if (iter.node == node) {
14360f1c 1686 ht_count_add(ht, size, hash);
83beee94
MD
1687 return NULL;
1688 }
1689
1690 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1691 return iter.node;
1692 }
48ed1c18
MD
1693}
1694
2e79c445
MD
1695int cds_lfht_replace(struct cds_lfht *ht,
1696 struct cds_lfht_iter *old_iter,
1697 unsigned long hash,
1698 cds_lfht_match_fct match,
1699 const void *key,
9357c415
MD
1700 struct cds_lfht_node *new_node)
1701{
1702 unsigned long size;
1703
709bacf9 1704 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1705 if (!old_iter->node)
1706 return -ENOENT;
1707 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1708 return -EINVAL;
1709 if (caa_unlikely(!match(old_iter->node, key)))
1710 return -EINVAL;
7b3893e4 1711 size = rcu_dereference(ht->size);
9357c415
MD
1712 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1713 new_node);
1714}
1715
bc8c3c74 1716int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1717{
95bc7fb9 1718 unsigned long size;
df44348d 1719 int ret;
abc490a1 1720
7b3893e4 1721 size = rcu_dereference(ht->size);
bc8c3c74 1722 ret = _cds_lfht_del(ht, size, node);
14360f1c 1723 if (!ret) {
95bc7fb9
MD
1724 unsigned long hash;
1725
bc8c3c74 1726 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1727 ht_count_del(ht, size, hash);
1728 }
df44348d 1729 return ret;
2ed95849 1730}
ab7d5fc6 1731
df55172a
MD
1732int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1733{
a85eff52 1734 return is_removed(CMM_LOAD_SHARED(node->next));
df55172a
MD
1735}
1736
abc490a1 1737static
1ee8f000 1738int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1739{
14044b37 1740 struct cds_lfht_node *node;
4105056a 1741 unsigned long order, i, size;
674f7a69 1742
abc490a1 1743 /* Check that the table is empty */
9d72a73f 1744 node = bucket_at(ht, 0);
abc490a1 1745 do {
04db56f8 1746 node = clear_flag(node)->next;
1ee8f000 1747 if (!is_bucket(node))
abc490a1 1748 return -EPERM;
273399de 1749 assert(!is_removed(node));
2f943cd7 1750 assert(!is_removal_owner(node));
bb7b2f26 1751 } while (!is_end(node));
4105056a
MD
1752 /*
1753 * size accessed without rcu_dereference because hash table is
1754 * being destroyed.
1755 */
7b3893e4 1756 size = ht->size;
1f67ba50 1757 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1758 for (i = 0; i < size; i++) {
1759 node = bucket_at(ht, i);
1760 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1761 i, i, bit_reverse_ulong(node->reverse_hash));
1762 assert(is_bucket(node->next));
1763 }
24365af7 1764
5bc6b66f 1765 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1766 cds_lfht_free_bucket_table(ht, order);
5488222b 1767
abc490a1 1768 return 0;
674f7a69
MD
1769}
1770
1771/*
1772 * Should only be called when no more concurrent readers nor writers can
1773 * possibly access the table.
1774 */
b7d619b0 1775int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1776{
a1e5e232 1777 int ret, was_online;
5e28c532 1778
848d4088 1779 /* Wait for in-flight resize operations to complete */
24953e08
MD
1780 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1781 cmm_smp_mb(); /* Store destroy before load resize */
a1e5e232
MD
1782 was_online = ht->flavor->read_ongoing();
1783 if (was_online)
1784 ht->flavor->thread_offline();
10e68472
MD
1785 /* Calling with RCU read-side held is an error. */
1786 if (ht->flavor->read_ongoing()) {
1787 ret = -EINVAL;
1788 if (was_online)
1789 ht->flavor->thread_online();
1790 goto end;
1791 }
848d4088
MD
1792 while (uatomic_read(&ht->in_progress_resize))
1793 poll(NULL, 0, 100); /* wait for 100ms */
a1e5e232
MD
1794 if (was_online)
1795 ht->flavor->thread_online();
1ee8f000 1796 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1797 if (ret)
1798 return ret;
5afadd12 1799 free_split_items_count(ht);
b7d619b0
MD
1800 if (attr)
1801 *attr = ht->resize_attr;
98808fb1 1802 poison_free(ht);
10e68472 1803end:
5e28c532 1804 return ret;
674f7a69
MD
1805}
1806
14044b37 1807void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1808 long *approx_before,
273399de 1809 unsigned long *count,
d933dd0e 1810 long *approx_after)
273399de 1811{
14044b37 1812 struct cds_lfht_node *node, *next;
caf3653d 1813 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1814
7ed7682f 1815 *approx_before = 0;
5afadd12 1816 if (ht->split_count) {
973e5e1b
MD
1817 int i;
1818
4c42f1b8
LJ
1819 for (i = 0; i < split_count_mask + 1; i++) {
1820 *approx_before += uatomic_read(&ht->split_count[i].add);
1821 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1822 }
1823 }
1824
273399de 1825 *count = 0;
273399de 1826
1ee8f000 1827 /* Count non-bucket nodes in the table */
9d72a73f 1828 node = bucket_at(ht, 0);
273399de 1829 do {
04db56f8 1830 next = rcu_dereference(node->next);
b198f0fd 1831 if (is_removed(next)) {
1ee8f000 1832 if (!is_bucket(next))
caf3653d 1833 (nr_removed)++;
973e5e1b 1834 else
1ee8f000
LJ
1835 (nr_bucket)++;
1836 } else if (!is_bucket(next))
273399de 1837 (*count)++;
24365af7 1838 else
1ee8f000 1839 (nr_bucket)++;
273399de 1840 node = clear_flag(next);
bb7b2f26 1841 } while (!is_end(node));
caf3653d 1842 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1843 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1844 *approx_after = 0;
5afadd12 1845 if (ht->split_count) {
973e5e1b
MD
1846 int i;
1847
4c42f1b8
LJ
1848 for (i = 0; i < split_count_mask + 1; i++) {
1849 *approx_after += uatomic_read(&ht->split_count[i].add);
1850 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1851 }
1852 }
273399de
MD
1853}
1854
1475579c 1855/* called with resize mutex held */
abc490a1 1856static
4105056a 1857void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1858 unsigned long old_size, unsigned long new_size)
abc490a1 1859{
1475579c 1860 unsigned long old_order, new_order;
1475579c 1861
5bc6b66f
MD
1862 old_order = cds_lfht_get_count_order_ulong(old_size);
1863 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1864 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1865 old_size, old_order, new_size, new_order);
1475579c 1866 assert(new_size > old_size);
93d46c39 1867 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1868}
1869
1870/* called with resize mutex held */
1871static
4105056a 1872void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1873 unsigned long old_size, unsigned long new_size)
464a1ec9 1874{
1475579c 1875 unsigned long old_order, new_order;
464a1ec9 1876
d0d8f9aa 1877 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1878 old_order = cds_lfht_get_count_order_ulong(old_size);
1879 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1880 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1881 old_size, old_order, new_size, new_order);
1475579c 1882 assert(new_size < old_size);
1475579c 1883
1ee8f000 1884 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1885 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1886}
1887
1475579c
MD
1888
1889/* called with resize mutex held */
1890static
1891void _do_cds_lfht_resize(struct cds_lfht *ht)
1892{
1893 unsigned long new_size, old_size;
4105056a
MD
1894
1895 /*
1896 * Resize table, re-do if the target size has changed under us.
1897 */
1898 do {
d2be3620
MD
1899 assert(uatomic_read(&ht->in_progress_resize));
1900 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1901 break;
7b3893e4
LJ
1902 ht->resize_initiated = 1;
1903 old_size = ht->size;
1904 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1905 if (old_size < new_size)
1906 _do_cds_lfht_grow(ht, old_size, new_size);
1907 else if (old_size > new_size)
1908 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 1909 ht->resize_initiated = 0;
4105056a
MD
1910 /* write resize_initiated before read resize_target */
1911 cmm_smp_mb();
7b3893e4 1912 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
1913}
1914
abc490a1 1915static
ab65b890 1916unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1917{
7b3893e4 1918 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
1919}
1920
1475579c 1921static
4105056a 1922void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1923 unsigned long count)
1475579c 1924{
d0d8f9aa 1925 count = max(count, MIN_TABLE_SIZE);
747d725c 1926 count = min(count, ht->max_nr_buckets);
7b3893e4 1927 uatomic_set(&ht->resize_target, count);
1475579c
MD
1928}
1929
1930void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1931{
a1e5e232
MD
1932 int was_online;
1933
a1e5e232
MD
1934 was_online = ht->flavor->read_ongoing();
1935 if (was_online)
1936 ht->flavor->thread_offline();
10e68472
MD
1937 /* Calling with RCU read-side held is an error. */
1938 if (ht->flavor->read_ongoing()) {
1939 static int print_once;
1940
1941 if (!CMM_LOAD_SHARED(print_once))
1942 fprintf(stderr, "[error] rculfhash: cds_lfht_resize "
1943 "called with RCU read-side lock held.\n");
1944 CMM_STORE_SHARED(print_once, 1);
1945 assert(0);
1946 goto end;
1947 }
1948 resize_target_update_count(ht, new_size);
1949 CMM_STORE_SHARED(ht->resize_initiated, 1);
1475579c
MD
1950 pthread_mutex_lock(&ht->resize_mutex);
1951 _do_cds_lfht_resize(ht);
1952 pthread_mutex_unlock(&ht->resize_mutex);
10e68472 1953end:
a1e5e232
MD
1954 if (was_online)
1955 ht->flavor->thread_online();
abc490a1 1956}
464a1ec9 1957
abc490a1
MD
1958static
1959void do_resize_cb(struct rcu_head *head)
1960{
1961 struct rcu_resize_work *work =
1962 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1963 struct cds_lfht *ht = work->ht;
abc490a1 1964
7b17c13e 1965 ht->flavor->thread_offline();
abc490a1 1966 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1967 _do_cds_lfht_resize(ht);
abc490a1 1968 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1969 ht->flavor->thread_online();
98808fb1 1970 poison_free(work);
848d4088
MD
1971 cmm_smp_mb(); /* finish resize before decrement */
1972 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1973}
1974
abc490a1 1975static
f1f119ee 1976void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1977{
abc490a1
MD
1978 struct rcu_resize_work *work;
1979
4105056a
MD
1980 /* Store resize_target before read resize_initiated */
1981 cmm_smp_mb();
7b3893e4 1982 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
848d4088 1983 uatomic_inc(&ht->in_progress_resize);
59290e9d 1984 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1985 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1986 uatomic_dec(&ht->in_progress_resize);
59290e9d 1987 return;
ed35e6d8 1988 }
f9830efd 1989 work = malloc(sizeof(*work));
741f378e
MD
1990 if (work == NULL) {
1991 dbg_printf("error allocating resize work, bailing out\n");
1992 uatomic_dec(&ht->in_progress_resize);
1993 return;
1994 }
f9830efd 1995 work->ht = ht;
7b17c13e 1996 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
7b3893e4 1997 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 1998 }
ab7d5fc6 1999}
3171717f 2000
f1f119ee
LJ
2001static
2002void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2003{
2004 unsigned long target_size = size << growth;
2005
747d725c 2006 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
2007 if (resize_target_grow(ht, target_size) >= target_size)
2008 return;
2009
2010 __cds_lfht_resize_lazy_launch(ht);
2011}
2012
89bb121d
LJ
2013/*
2014 * We favor grow operations over shrink. A shrink operation never occurs
2015 * if a grow operation is queued for lazy execution. A grow operation
2016 * cancels any pending shrink lazy execution.
2017 */
3171717f 2018static
4105056a 2019void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
2020 unsigned long count)
2021{
b8af5011
MD
2022 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2023 return;
d0d8f9aa 2024 count = max(count, MIN_TABLE_SIZE);
747d725c 2025 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
2026 if (count == size)
2027 return; /* Already the right size, no resize needed */
2028 if (count > size) { /* lazy grow */
2029 if (resize_target_grow(ht, count) >= count)
2030 return;
2031 } else { /* lazy shrink */
2032 for (;;) {
2033 unsigned long s;
2034
7b3893e4 2035 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
2036 if (s == size)
2037 break; /* no resize needed */
2038 if (s > size)
2039 return; /* growing is/(was just) in progress */
2040 if (s <= count)
2041 return; /* some other thread do shrink */
2042 size = s;
2043 }
2044 }
f1f119ee 2045 __cds_lfht_resize_lazy_launch(ht);
3171717f 2046}
This page took 0.15942 seconds and 4 git commands to generate.