Port: Add Solaris getcpu 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>
e0ba718a 267
15cfbec7 268#include "config.h"
a47dd11c 269#include "compat-getcpu.h"
2ed95849 270#include <urcu.h>
abc490a1 271#include <urcu-call-rcu.h>
7b17c13e 272#include <urcu-flavor.h>
a42cc659
MD
273#include <urcu/arch.h>
274#include <urcu/uatomic.h>
a42cc659 275#include <urcu/compiler.h>
abc490a1 276#include <urcu/rculfhash.h>
0b6aa001 277#include <rculfhash-internal.h>
5e28c532 278#include <stdio.h>
464a1ec9 279#include <pthread.h>
44395fb7 280
f8994aee 281/*
4c42f1b8 282 * Split-counters lazily update the global counter each 1024
f8994aee
MD
283 * addition/removal. It automatically keeps track of resize required.
284 * We use the bucket length as indicator for need to expand for small
ffa11a18 285 * tables and machines lacking per-cpu data support.
f8994aee
MD
286 */
287#define COUNT_COMMIT_ORDER 10
4ddbb355 288#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
289#define CHAIN_LEN_TARGET 1
290#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 291
cd95516d 292/*
76a73da8 293 * Define the minimum table size.
cd95516d 294 */
d0d8f9aa
LJ
295#define MIN_TABLE_ORDER 0
296#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 297
b7d619b0 298/*
1ee8f000 299 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 300 */
6083a889
MD
301#define MIN_PARTITION_PER_THREAD_ORDER 12
302#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 303
d95bd160
MD
304/*
305 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 306 * It indicates that no node must attach to the node scheduled for
b198f0fd 307 * removal, and that node garbage collection must be performed.
1ee8f000 308 * The bucket flag does not require to be updated atomically with the
d95bd160 309 * pointer, but it is added as a pointer low bit flag to save space.
1f67ba50
MD
310 * The "removal owner" flag is used to detect which of the "del"
311 * operation that has set the "removed flag" gets to return the removed
312 * node to its caller. Note that the replace operation does not need to
313 * iteract with the "removal owner" flag, because it validates that
314 * the "removed" flag is not set before performing its cmpxchg.
d95bd160 315 */
d37166c6 316#define REMOVED_FLAG (1UL << 0)
1ee8f000 317#define BUCKET_FLAG (1UL << 1)
db00ccc3
MD
318#define REMOVAL_OWNER_FLAG (1UL << 2)
319#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 320
bb7b2f26 321/* Value of the end pointer. Should not interact with flags. */
f9c80341 322#define END_VALUE NULL
bb7b2f26 323
7f52427b
MD
324/*
325 * ht_items_count: Split-counters counting the number of node addition
326 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
327 * is set at hash table creation.
328 *
329 * These are free-running counters, never reset to zero. They count the
330 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
331 * operations to update the global counter. We choose a power-of-2 value
332 * for the trigger to deal with 32 or 64-bit overflow of the counter.
333 */
df44348d 334struct ht_items_count {
860d07e8 335 unsigned long add, del;
df44348d
MD
336} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
337
7f52427b
MD
338/*
339 * rcu_resize_work: Contains arguments passed to RCU worker thread
340 * responsible for performing lazy resize.
341 */
abc490a1
MD
342struct rcu_resize_work {
343 struct rcu_head head;
14044b37 344 struct cds_lfht *ht;
abc490a1 345};
2ed95849 346
7f52427b
MD
347/*
348 * partition_resize_work: Contains arguments passed to worker threads
349 * executing the hash table resize on partitions of the hash table
350 * assigned to each processor's worker thread.
351 */
b7d619b0 352struct partition_resize_work {
1af6e26e 353 pthread_t thread_id;
b7d619b0
MD
354 struct cds_lfht *ht;
355 unsigned long i, start, len;
356 void (*fct)(struct cds_lfht *ht, unsigned long i,
357 unsigned long start, unsigned long len);
358};
359
abc490a1
MD
360/*
361 * Algorithm to reverse bits in a word by lookup table, extended to
362 * 64-bit words.
f9830efd 363 * Source:
abc490a1 364 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 365 * Originally from Public Domain.
abc490a1
MD
366 */
367
67ecffc0 368static const uint8_t BitReverseTable256[256] =
2ed95849 369{
abc490a1
MD
370#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
371#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
372#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
373 R6(0), R6(2), R6(1), R6(3)
374};
375#undef R2
376#undef R4
377#undef R6
2ed95849 378
abc490a1
MD
379static
380uint8_t bit_reverse_u8(uint8_t v)
381{
382 return BitReverseTable256[v];
383}
ab7d5fc6 384
95bc7fb9
MD
385#if (CAA_BITS_PER_LONG == 32)
386static
abc490a1
MD
387uint32_t bit_reverse_u32(uint32_t v)
388{
67ecffc0
MD
389 return ((uint32_t) bit_reverse_u8(v) << 24) |
390 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
391 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
abc490a1 392 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849 393}
95bc7fb9
MD
394#else
395static
abc490a1 396uint64_t bit_reverse_u64(uint64_t v)
2ed95849 397{
67ecffc0
MD
398 return ((uint64_t) bit_reverse_u8(v) << 56) |
399 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
abc490a1
MD
400 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
401 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
67ecffc0
MD
402 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
403 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
abc490a1
MD
404 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
405 ((uint64_t) bit_reverse_u8(v >> 56));
406}
95bc7fb9 407#endif
abc490a1
MD
408
409static
410unsigned long bit_reverse_ulong(unsigned long v)
411{
412#if (CAA_BITS_PER_LONG == 32)
413 return bit_reverse_u32(v);
414#else
415 return bit_reverse_u64(v);
416#endif
417}
418
f9830efd 419/*
24365af7
MD
420 * fls: returns the position of the most significant bit.
421 * Returns 0 if no bit is set, else returns the position of the most
422 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 423 */
24365af7
MD
424#if defined(__i386) || defined(__x86_64)
425static inline
426unsigned int fls_u32(uint32_t x)
f9830efd 427{
24365af7
MD
428 int r;
429
e1789ce2 430 __asm__ ("bsrl %1,%0\n\t"
24365af7
MD
431 "jnz 1f\n\t"
432 "movl $-1,%0\n\t"
433 "1:\n\t"
434 : "=r" (r) : "rm" (x));
435 return r + 1;
436}
437#define HAS_FLS_U32
438#endif
439
440#if defined(__x86_64)
441static inline
442unsigned int fls_u64(uint64_t x)
443{
444 long r;
445
e1789ce2 446 __asm__ ("bsrq %1,%0\n\t"
24365af7
MD
447 "jnz 1f\n\t"
448 "movq $-1,%0\n\t"
449 "1:\n\t"
450 : "=r" (r) : "rm" (x));
451 return r + 1;
452}
453#define HAS_FLS_U64
454#endif
455
456#ifndef HAS_FLS_U64
457static __attribute__((unused))
458unsigned int fls_u64(uint64_t x)
459{
460 unsigned int r = 64;
461
462 if (!x)
463 return 0;
464
465 if (!(x & 0xFFFFFFFF00000000ULL)) {
466 x <<= 32;
467 r -= 32;
468 }
469 if (!(x & 0xFFFF000000000000ULL)) {
470 x <<= 16;
471 r -= 16;
472 }
473 if (!(x & 0xFF00000000000000ULL)) {
474 x <<= 8;
475 r -= 8;
476 }
477 if (!(x & 0xF000000000000000ULL)) {
478 x <<= 4;
479 r -= 4;
480 }
481 if (!(x & 0xC000000000000000ULL)) {
482 x <<= 2;
483 r -= 2;
484 }
485 if (!(x & 0x8000000000000000ULL)) {
486 x <<= 1;
487 r -= 1;
488 }
489 return r;
490}
491#endif
492
493#ifndef HAS_FLS_U32
494static __attribute__((unused))
495unsigned int fls_u32(uint32_t x)
496{
497 unsigned int r = 32;
f9830efd 498
24365af7
MD
499 if (!x)
500 return 0;
501 if (!(x & 0xFFFF0000U)) {
502 x <<= 16;
503 r -= 16;
504 }
505 if (!(x & 0xFF000000U)) {
506 x <<= 8;
507 r -= 8;
508 }
509 if (!(x & 0xF0000000U)) {
510 x <<= 4;
511 r -= 4;
512 }
513 if (!(x & 0xC0000000U)) {
514 x <<= 2;
515 r -= 2;
516 }
517 if (!(x & 0x80000000U)) {
518 x <<= 1;
519 r -= 1;
520 }
521 return r;
522}
523#endif
524
5bc6b66f 525unsigned int cds_lfht_fls_ulong(unsigned long x)
f9830efd 526{
6887cc5e 527#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
528 return fls_u32(x);
529#else
530 return fls_u64(x);
531#endif
532}
f9830efd 533
920f8ef6
LJ
534/*
535 * Return the minimum order for which x <= (1UL << order).
536 * Return -1 if x is 0.
537 */
5bc6b66f 538int cds_lfht_get_count_order_u32(uint32_t x)
24365af7 539{
920f8ef6
LJ
540 if (!x)
541 return -1;
24365af7 542
920f8ef6 543 return fls_u32(x - 1);
24365af7
MD
544}
545
920f8ef6
LJ
546/*
547 * Return the minimum order for which x <= (1UL << order).
548 * Return -1 if x is 0.
549 */
5bc6b66f 550int cds_lfht_get_count_order_ulong(unsigned long x)
24365af7 551{
920f8ef6
LJ
552 if (!x)
553 return -1;
24365af7 554
5bc6b66f 555 return cds_lfht_fls_ulong(x - 1);
f9830efd
MD
556}
557
558static
ab65b890 559void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 560
f8994aee 561static
4105056a 562void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
563 unsigned long count);
564
df44348d 565static long nr_cpus_mask = -1;
4c42f1b8 566static long split_count_mask = -1;
e53ab1eb 567static int split_count_order = -1;
4c42f1b8 568
4ddbb355 569#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
570static void ht_init_nr_cpus_mask(void)
571{
572 long maxcpus;
573
574 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
575 if (maxcpus <= 0) {
576 nr_cpus_mask = -2;
577 return;
578 }
579 /*
580 * round up number of CPUs to next power of two, so we
581 * can use & for modulo.
582 */
5bc6b66f 583 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
584 nr_cpus_mask = maxcpus - 1;
585}
4ddbb355
LJ
586#else /* #if defined(HAVE_SYSCONF) */
587static void ht_init_nr_cpus_mask(void)
588{
589 nr_cpus_mask = -2;
590}
591#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
592
593static
5afadd12 594void alloc_split_items_count(struct cds_lfht *ht)
df44348d 595{
4c42f1b8
LJ
596 if (nr_cpus_mask == -1) {
597 ht_init_nr_cpus_mask();
4ddbb355
LJ
598 if (nr_cpus_mask < 0)
599 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
600 else
601 split_count_mask = nr_cpus_mask;
e53ab1eb
MD
602 split_count_order =
603 cds_lfht_get_count_order_ulong(split_count_mask + 1);
df44348d 604 }
4c42f1b8 605
4ddbb355 606 assert(split_count_mask >= 0);
5afadd12
LJ
607
608 if (ht->flags & CDS_LFHT_ACCOUNTING) {
95bc7fb9
MD
609 ht->split_count = calloc(split_count_mask + 1,
610 sizeof(struct ht_items_count));
5afadd12
LJ
611 assert(ht->split_count);
612 } else {
613 ht->split_count = NULL;
614 }
df44348d
MD
615}
616
617static
5afadd12 618void free_split_items_count(struct cds_lfht *ht)
df44348d 619{
5afadd12 620 poison_free(ht->split_count);
df44348d
MD
621}
622
623static
14360f1c 624int ht_get_split_count_index(unsigned long hash)
df44348d
MD
625{
626 int cpu;
627
4c42f1b8 628 assert(split_count_mask >= 0);
a47dd11c 629 cpu = urcu_sched_getcpu();
8ed51e04 630 if (caa_unlikely(cpu < 0))
14360f1c 631 return hash & split_count_mask;
df44348d 632 else
4c42f1b8 633 return cpu & split_count_mask;
df44348d
MD
634}
635
636static
14360f1c 637void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 638{
4c42f1b8
LJ
639 unsigned long split_count;
640 int index;
314558bf 641 long count;
df44348d 642
8ed51e04 643 if (caa_unlikely(!ht->split_count))
3171717f 644 return;
14360f1c 645 index = ht_get_split_count_index(hash);
4c42f1b8 646 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
647 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
648 return;
649 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
650
651 dbg_printf("add split count %lu\n", split_count);
652 count = uatomic_add_return(&ht->count,
653 1UL << COUNT_COMMIT_ORDER);
4c299dcb 654 if (caa_likely(count & (count - 1)))
314558bf
MD
655 return;
656 /* Only if global count is power of 2 */
657
658 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
659 return;
660 dbg_printf("add set global %ld\n", count);
661 cds_lfht_resize_lazy_count(ht, size,
662 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
663}
664
665static
14360f1c 666void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 667{
4c42f1b8
LJ
668 unsigned long split_count;
669 int index;
314558bf 670 long count;
df44348d 671
8ed51e04 672 if (caa_unlikely(!ht->split_count))
3171717f 673 return;
14360f1c 674 index = ht_get_split_count_index(hash);
4c42f1b8 675 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
676 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
677 return;
678 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
679
680 dbg_printf("del split count %lu\n", split_count);
681 count = uatomic_add_return(&ht->count,
682 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 683 if (caa_likely(count & (count - 1)))
314558bf
MD
684 return;
685 /* Only if global count is power of 2 */
686
687 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
688 return;
689 dbg_printf("del set global %ld\n", count);
690 /*
691 * Don't shrink table if the number of nodes is below a
692 * certain threshold.
693 */
694 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
695 return;
696 cds_lfht_resize_lazy_count(ht, size,
697 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
698}
699
f9830efd 700static
4105056a 701void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 702{
f8994aee
MD
703 unsigned long count;
704
b8af5011
MD
705 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
706 return;
f8994aee
MD
707 count = uatomic_read(&ht->count);
708 /*
709 * Use bucket-local length for small table expand and for
710 * environments lacking per-cpu data support.
711 */
e53ab1eb 712 if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order)))
f8994aee 713 return;
24365af7 714 if (chain_len > 100)
f0c29ed7 715 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 716 chain_len);
e53ab1eb
MD
717 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) {
718 int growth;
719
720 /*
721 * Ideal growth calculated based on chain length.
722 */
723 growth = cds_lfht_get_count_order_u32(chain_len
724 - (CHAIN_LEN_TARGET - 1));
725 if ((ht->flags & CDS_LFHT_ACCOUNTING)
726 && (size << growth)
727 >= (1UL << (COUNT_COMMIT_ORDER
728 + split_count_order))) {
729 /*
730 * If ideal growth expands the hash table size
731 * beyond the "small hash table" sizes, use the
732 * maximum small hash table size to attempt
733 * expanding the hash table. This only applies
734 * when node accounting is available, otherwise
735 * the chain length is used to expand the hash
736 * table in every case.
737 */
738 growth = COUNT_COMMIT_ORDER + split_count_order
739 - cds_lfht_get_count_order_ulong(size);
740 if (growth <= 0)
741 return;
742 }
743 cds_lfht_resize_lazy_grow(ht, size, growth);
744 }
f9830efd
MD
745}
746
abc490a1 747static
14044b37 748struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 749{
14044b37 750 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
751}
752
753static
14044b37 754int is_removed(struct cds_lfht_node *node)
abc490a1 755{
d37166c6 756 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
757}
758
f5596c94 759static
1ee8f000 760int is_bucket(struct cds_lfht_node *node)
f5596c94 761{
1ee8f000 762 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
763}
764
765static
1ee8f000 766struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 767{
1ee8f000 768 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 769}
bb7b2f26 770
db00ccc3
MD
771static
772int is_removal_owner(struct cds_lfht_node *node)
773{
774 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
775}
776
777static
778struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
779{
780 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
781}
782
71bb3aca
MD
783static
784struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node)
785{
786 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG);
787}
788
bb7b2f26
MD
789static
790struct cds_lfht_node *get_end(void)
791{
792 return (struct cds_lfht_node *) END_VALUE;
793}
794
795static
796int is_end(struct cds_lfht_node *node)
797{
798 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
799}
800
abc490a1 801static
ab65b890
LJ
802unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
803 unsigned long v)
abc490a1
MD
804{
805 unsigned long old1, old2;
806
807 old1 = uatomic_read(ptr);
808 do {
809 old2 = old1;
810 if (old2 >= v)
f9830efd 811 return old2;
abc490a1 812 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 813 return old2;
abc490a1
MD
814}
815
48f1b16d
LJ
816static
817void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
818{
0b6aa001 819 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
820}
821
822/*
823 * cds_lfht_free_bucket_table() should be called with decreasing order.
824 * When cds_lfht_free_bucket_table(0) is called, it means the whole
825 * lfht is destroyed.
826 */
827static
828void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
829{
0b6aa001 830 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
831}
832
9d72a73f
LJ
833static inline
834struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 835{
0b6aa001 836 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
837}
838
9d72a73f
LJ
839static inline
840struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
841 unsigned long hash)
842{
843 assert(size > 0);
844 return bucket_at(ht, hash & (size - 1));
845}
846
273399de
MD
847/*
848 * Remove all logically deleted nodes from a bucket up to a certain node key.
849 */
850static
1ee8f000 851void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 852{
14044b37 853 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 854
1ee8f000
LJ
855 assert(!is_bucket(bucket));
856 assert(!is_removed(bucket));
2f943cd7 857 assert(!is_removal_owner(bucket));
1ee8f000 858 assert(!is_bucket(node));
c90201ac 859 assert(!is_removed(node));
2f943cd7 860 assert(!is_removal_owner(node));
273399de 861 for (;;) {
1ee8f000
LJ
862 iter_prev = bucket;
863 /* We can always skip the bucket node initially */
04db56f8 864 iter = rcu_dereference(iter_prev->next);
b4cb483f 865 assert(!is_removed(iter));
2f943cd7 866 assert(!is_removal_owner(iter));
04db56f8 867 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 868 /*
1ee8f000 869 * We should never be called with bucket (start of chain)
bd4db153
MD
870 * and logically removed node (end of path compression
871 * marker) being the actual same node. This would be a
872 * bug in the algorithm implementation.
873 */
1ee8f000 874 assert(bucket != node);
273399de 875 for (;;) {
8ed51e04 876 if (caa_unlikely(is_end(iter)))
f9c80341 877 return;
04db56f8 878 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 879 return;
04db56f8 880 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 881 if (caa_likely(is_removed(next)))
273399de 882 break;
b453eae1 883 iter_prev = clear_flag(iter);
273399de
MD
884 iter = next;
885 }
b198f0fd 886 assert(!is_removed(iter));
2f943cd7 887 assert(!is_removal_owner(iter));
1ee8f000
LJ
888 if (is_bucket(iter))
889 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
890 else
891 new_next = clear_flag(next);
04db56f8 892 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
893 }
894}
895
9357c415
MD
896static
897int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
898 struct cds_lfht_node *old_node,
3fb86f26 899 struct cds_lfht_node *old_next,
9357c415
MD
900 struct cds_lfht_node *new_node)
901{
04db56f8 902 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
903
904 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 905 return -ENOENT;
9357c415
MD
906
907 assert(!is_removed(old_node));
2f943cd7 908 assert(!is_removal_owner(old_node));
1ee8f000 909 assert(!is_bucket(old_node));
9357c415 910 assert(!is_removed(new_node));
2f943cd7 911 assert(!is_removal_owner(new_node));
1ee8f000 912 assert(!is_bucket(new_node));
9357c415 913 assert(new_node != old_node);
3fb86f26 914 for (;;) {
9357c415 915 /* Insert after node to be replaced */
9357c415
MD
916 if (is_removed(old_next)) {
917 /*
918 * Too late, the old node has been removed under us
919 * between lookup and replace. Fail.
920 */
7801dadd 921 return -ENOENT;
9357c415 922 }
feda2722
LJ
923 assert(old_next == clear_flag(old_next));
924 assert(new_node != old_next);
71bb3aca
MD
925 /*
926 * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED
927 * flag. It is either set atomically at the same time
928 * (replace) or after (del).
929 */
930 assert(!is_removal_owner(old_next));
feda2722 931 new_node->next = old_next;
9357c415
MD
932 /*
933 * Here is the whole trick for lock-free replace: we add
934 * the replacement node _after_ the node we want to
935 * replace by atomically setting its next pointer at the
936 * same time we set its removal flag. Given that
937 * the lookups/get next use an iterator aware of the
938 * next pointer, they will either skip the old node due
939 * to the removal flag and see the new node, or use
940 * the old node, but will not see the new one.
db00ccc3
MD
941 * This is a replacement of a node with another node
942 * that has the same value: we are therefore not
71bb3aca
MD
943 * removing a value from the hash table. We set both the
944 * REMOVED and REMOVAL_OWNER flags atomically so we own
945 * the node after successful cmpxchg.
9357c415 946 */
04db56f8 947 ret_next = uatomic_cmpxchg(&old_node->next,
71bb3aca 948 old_next, flag_removed_or_removal_owner(new_node));
3fb86f26 949 if (ret_next == old_next)
7801dadd 950 break; /* We performed the replacement. */
3fb86f26
LJ
951 old_next = ret_next;
952 }
9357c415 953
9357c415
MD
954 /*
955 * Ensure that the old node is not visible to readers anymore:
956 * lookup for the node, and remove it (along with any other
957 * logically removed node) if found.
958 */
04db56f8
LJ
959 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
960 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 961
a85eff52 962 assert(is_removed(CMM_LOAD_SHARED(old_node->next)));
7801dadd 963 return 0;
9357c415
MD
964}
965
83beee94
MD
966/*
967 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
968 * mode. A NULL unique_ret allows creation of duplicate keys.
969 */
abc490a1 970static
83beee94 971void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 972 unsigned long hash,
0422d92c 973 cds_lfht_match_fct match,
996ff57c 974 const void *key,
83beee94
MD
975 unsigned long size,
976 struct cds_lfht_node *node,
977 struct cds_lfht_iter *unique_ret,
1ee8f000 978 int bucket_flag)
abc490a1 979{
14044b37 980 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 981 *return_node;
04db56f8 982 struct cds_lfht_node *bucket;
abc490a1 983
1ee8f000 984 assert(!is_bucket(node));
c90201ac 985 assert(!is_removed(node));
2f943cd7 986 assert(!is_removal_owner(node));
91a75cc5 987 bucket = lookup_bucket(ht, size, hash);
abc490a1 988 for (;;) {
adc0de68 989 uint32_t chain_len = 0;
abc490a1 990
11519af6
MD
991 /*
992 * iter_prev points to the non-removed node prior to the
993 * insert location.
11519af6 994 */
04db56f8 995 iter_prev = bucket;
1ee8f000 996 /* We can always skip the bucket node initially */
04db56f8
LJ
997 iter = rcu_dereference(iter_prev->next);
998 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 999 for (;;) {
8ed51e04 1000 if (caa_unlikely(is_end(iter)))
273399de 1001 goto insert;
04db56f8 1002 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 1003 goto insert;
238cc06e 1004
1ee8f000
LJ
1005 /* bucket node is the first node of the identical-hash-value chain */
1006 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 1007 goto insert;
238cc06e 1008
04db56f8 1009 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 1010 if (caa_unlikely(is_removed(next)))
9dba85be 1011 goto gc_node;
238cc06e
LJ
1012
1013 /* uniquely add */
83beee94 1014 if (unique_ret
1ee8f000 1015 && !is_bucket(next)
04db56f8 1016 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
1017 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
1018
1019 /*
1020 * uniquely adding inserts the node as the first
1021 * node of the identical-hash-value node chain.
1022 *
1023 * This semantic ensures no duplicated keys
1024 * should ever be observable in the table
1f67ba50
MD
1025 * (including traversing the table node by
1026 * node by forward iterations)
238cc06e 1027 */
04db56f8 1028 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
1029 if (!d_iter.node)
1030 goto insert;
1031
1032 *unique_ret = d_iter;
83beee94 1033 return;
48ed1c18 1034 }
238cc06e 1035
11519af6 1036 /* Only account for identical reverse hash once */
04db56f8 1037 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1038 && !is_bucket(next))
4105056a 1039 check_resize(ht, size, ++chain_len);
11519af6 1040 iter_prev = clear_flag(iter);
273399de 1041 iter = next;
abc490a1 1042 }
48ed1c18 1043
273399de 1044 insert:
7ec59d3b 1045 assert(node != clear_flag(iter));
11519af6 1046 assert(!is_removed(iter_prev));
2f943cd7 1047 assert(!is_removal_owner(iter_prev));
c90201ac 1048 assert(!is_removed(iter));
2f943cd7 1049 assert(!is_removal_owner(iter));
f000907d 1050 assert(iter_prev != node);
1ee8f000 1051 if (!bucket_flag)
04db56f8 1052 node->next = clear_flag(iter);
f9c80341 1053 else
1ee8f000
LJ
1054 node->next = flag_bucket(clear_flag(iter));
1055 if (is_bucket(iter))
1056 new_node = flag_bucket(node);
f5596c94
MD
1057 else
1058 new_node = node;
04db56f8 1059 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1060 new_node) != iter) {
273399de 1061 continue; /* retry */
48ed1c18 1062 } else {
83beee94 1063 return_node = node;
960c9e4f 1064 goto end;
48ed1c18
MD
1065 }
1066
9dba85be
MD
1067 gc_node:
1068 assert(!is_removed(iter));
2f943cd7 1069 assert(!is_removal_owner(iter));
1ee8f000
LJ
1070 if (is_bucket(iter))
1071 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1072 else
1073 new_next = clear_flag(next);
04db56f8 1074 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1075 /* retry */
464a1ec9 1076 }
9357c415 1077end:
83beee94
MD
1078 if (unique_ret) {
1079 unique_ret->node = return_node;
1080 /* unique_ret->next left unset, never used. */
1081 }
abc490a1 1082}
464a1ec9 1083
abc490a1 1084static
860d07e8 1085int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1086 struct cds_lfht_node *node)
abc490a1 1087{
db00ccc3 1088 struct cds_lfht_node *bucket, *next;
5e28c532 1089
9357c415 1090 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1091 return -ENOENT;
9357c415 1092
7ec59d3b 1093 /* logically delete the node */
1ee8f000 1094 assert(!is_bucket(node));
c90201ac 1095 assert(!is_removed(node));
db00ccc3 1096 assert(!is_removal_owner(node));
48ed1c18 1097
db00ccc3
MD
1098 /*
1099 * We are first checking if the node had previously been
1100 * logically removed (this check is not atomic with setting the
1101 * logical removal flag). Return -ENOENT if the node had
1102 * previously been removed.
1103 */
a85eff52 1104 next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */
db00ccc3
MD
1105 if (caa_unlikely(is_removed(next)))
1106 return -ENOENT;
b65ec430 1107 assert(!is_bucket(next));
196f4fab
MD
1108 /*
1109 * The del operation semantic guarantees a full memory barrier
1110 * before the uatomic_or atomic commit of the deletion flag.
1111 */
1112 cmm_smp_mb__before_uatomic_or();
db00ccc3
MD
1113 /*
1114 * We set the REMOVED_FLAG unconditionally. Note that there may
1115 * be more than one concurrent thread setting this flag.
1116 * Knowing which wins the race will be known after the garbage
1117 * collection phase, stay tuned!
1118 */
1119 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1120 /* We performed the (logical) deletion. */
7ec59d3b
MD
1121
1122 /*
1123 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1124 * the node, and remove it (along with any other logically removed node)
1125 * if found.
11519af6 1126 */
04db56f8
LJ
1127 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1128 _cds_lfht_gc_bucket(bucket, node);
743f9143 1129
a85eff52 1130 assert(is_removed(CMM_LOAD_SHARED(node->next)));
db00ccc3
MD
1131 /*
1132 * Last phase: atomically exchange node->next with a version
1133 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1134 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1135 * the node and win the removal race.
1136 * It is interesting to note that all "add" paths are forbidden
1137 * to change the next pointer starting from the point where the
1138 * REMOVED_FLAG is set, so here using a read, followed by a
1139 * xchg() suffice to guarantee that the xchg() will ever only
1140 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1141 * was already set).
1142 */
1143 if (!is_removal_owner(uatomic_xchg(&node->next,
1144 flag_removal_owner(node->next))))
1145 return 0;
1146 else
1147 return -ENOENT;
abc490a1 1148}
2ed95849 1149
b7d619b0
MD
1150static
1151void *partition_resize_thread(void *arg)
1152{
1153 struct partition_resize_work *work = arg;
1154
7b17c13e 1155 work->ht->flavor->register_thread();
b7d619b0 1156 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1157 work->ht->flavor->unregister_thread();
b7d619b0
MD
1158 return NULL;
1159}
1160
1161static
1162void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1163 unsigned long len,
1164 void (*fct)(struct cds_lfht *ht, unsigned long i,
1165 unsigned long start, unsigned long len))
1166{
e54ec2f5 1167 unsigned long partition_len, start = 0;
b7d619b0 1168 struct partition_resize_work *work;
6083a889
MD
1169 int thread, ret;
1170 unsigned long nr_threads;
b7d619b0 1171
d7f3ba4c
EW
1172 assert(nr_cpus_mask != -1);
1173 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD)
1174 goto fallback;
1175
6083a889
MD
1176 /*
1177 * Note: nr_cpus_mask + 1 is always power of 2.
1178 * We spawn just the number of threads we need to satisfy the minimum
1179 * partition size, up to the number of CPUs in the system.
1180 */
91452a6a
MD
1181 if (nr_cpus_mask > 0) {
1182 nr_threads = min(nr_cpus_mask + 1,
1183 len >> MIN_PARTITION_PER_THREAD_ORDER);
1184 } else {
1185 nr_threads = 1;
1186 }
5bc6b66f 1187 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1188 work = calloc(nr_threads, sizeof(*work));
7c75d498
EW
1189 if (!work) {
1190 dbg_printf("error allocating for resize, single-threading\n");
1191 goto fallback;
1192 }
6083a889
MD
1193 for (thread = 0; thread < nr_threads; thread++) {
1194 work[thread].ht = ht;
1195 work[thread].i = i;
1196 work[thread].len = partition_len;
1197 work[thread].start = thread * partition_len;
1198 work[thread].fct = fct;
1af6e26e 1199 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1200 partition_resize_thread, &work[thread]);
e54ec2f5
EW
1201 if (ret == EAGAIN) {
1202 /*
1203 * Out of resources: wait and join the threads
1204 * we've created, then handle leftovers.
1205 */
1206 dbg_printf("error spawning for resize, single-threading\n");
1207 start = work[thread].start;
1208 len -= start;
1209 nr_threads = thread;
1210 break;
1211 }
b7d619b0
MD
1212 assert(!ret);
1213 }
6083a889 1214 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1215 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1216 assert(!ret);
1217 }
1218 free(work);
e54ec2f5
EW
1219
1220 /*
1221 * A pthread_create failure above will either lead in us having
1222 * no threads to join or starting at a non-zero offset,
1223 * fallback to single thread processing of leftovers.
1224 */
1225 if (start == 0 && nr_threads > 0)
1226 return;
7c75d498
EW
1227fallback:
1228 ht->flavor->thread_online();
e54ec2f5 1229 fct(ht, i, start, len);
7c75d498 1230 ht->flavor->thread_offline();
b7d619b0
MD
1231}
1232
e8de508e
MD
1233/*
1234 * Holding RCU read lock to protect _cds_lfht_add against memory
1235 * reclaim that could be performed by other call_rcu worker threads (ABA
1236 * problem).
9ee0fc9a 1237 *
b7d619b0 1238 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1239 * many worker threads, based on the number of CPUs available in the system.
1240 * This should therefore take care of not having the expand lagging behind too
1241 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1242 * schedule bucket node population fairly with insertions.
e8de508e 1243 */
4105056a 1244static
b7d619b0
MD
1245void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1246 unsigned long start, unsigned long len)
4105056a 1247{
9d72a73f 1248 unsigned long j, size = 1UL << (i - 1);
4105056a 1249
d0d8f9aa 1250 assert(i > MIN_TABLE_ORDER);
7b17c13e 1251 ht->flavor->read_lock();
9d72a73f
LJ
1252 for (j = size + start; j < size + start + len; j++) {
1253 struct cds_lfht_node *new_node = bucket_at(ht, j);
1254
1255 assert(j >= size && j < (size << 1));
1256 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1257 i, j, j);
1258 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1259 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1260 }
7b17c13e 1261 ht->flavor->read_unlock();
b7d619b0
MD
1262}
1263
1264static
1265void init_table_populate(struct cds_lfht *ht, unsigned long i,
1266 unsigned long len)
1267{
b7d619b0 1268 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1269}
1270
abc490a1 1271static
4105056a 1272void init_table(struct cds_lfht *ht,
93d46c39 1273 unsigned long first_order, unsigned long last_order)
24365af7 1274{
93d46c39 1275 unsigned long i;
24365af7 1276
93d46c39
LJ
1277 dbg_printf("init table: first_order %lu last_order %lu\n",
1278 first_order, last_order);
d0d8f9aa 1279 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1280 for (i = first_order; i <= last_order; i++) {
4105056a 1281 unsigned long len;
24365af7 1282
4f6e90b7 1283 len = 1UL << (i - 1);
f0c29ed7 1284 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1285
1286 /* Stop expand if the resize target changes under us */
7b3893e4 1287 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1288 break;
1289
48f1b16d 1290 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1291
4105056a 1292 /*
1ee8f000
LJ
1293 * Set all bucket nodes reverse hash values for a level and
1294 * link all bucket nodes into the table.
4105056a 1295 */
dc1da8f6 1296 init_table_populate(ht, i, len);
4105056a 1297
f9c80341
MD
1298 /*
1299 * Update table size.
1300 */
1301 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1302 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1303
4f6e90b7 1304 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1305 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1306 break;
1307 }
1308}
1309
e8de508e
MD
1310/*
1311 * Holding RCU read lock to protect _cds_lfht_remove against memory
1312 * reclaim that could be performed by other call_rcu worker threads (ABA
1313 * problem).
1314 * For a single level, we logically remove and garbage collect each node.
1315 *
1316 * As a design choice, we perform logical removal and garbage collection on a
1317 * node-per-node basis to simplify this algorithm. We also assume keeping good
1318 * cache locality of the operation would overweight possible performance gain
1319 * that could be achieved by batching garbage collection for multiple levels.
1320 * However, this would have to be justified by benchmarks.
1321 *
1322 * Concurrent removal and add operations are helping us perform garbage
1323 * collection of logically removed nodes. We guarantee that all logically
1324 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1325 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e 1326 *
1f67ba50
MD
1327 * Logical removal and garbage collection can therefore be done in batch
1328 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1329 *
b7d619b0
MD
1330 * When we reach a certain length, we can split this removal over many worker
1331 * threads, based on the number of CPUs available in the system. This should
1332 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1333 * updater threads actively inserting into the hash table.
e8de508e 1334 */
4105056a 1335static
b7d619b0
MD
1336void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1337 unsigned long start, unsigned long len)
4105056a 1338{
9d72a73f 1339 unsigned long j, size = 1UL << (i - 1);
4105056a 1340
d0d8f9aa 1341 assert(i > MIN_TABLE_ORDER);
7b17c13e 1342 ht->flavor->read_lock();
9d72a73f 1343 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1344 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1345 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1346
1347 assert(j >= size && j < (size << 1));
1348 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1349 i, j, j);
2e2ce1e9
LJ
1350 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1351 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1352 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1353 }
7b17c13e 1354 ht->flavor->read_unlock();
b7d619b0
MD
1355}
1356
1357static
1358void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1359{
b7d619b0 1360 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1361}
1362
61adb337
MD
1363/*
1364 * fini_table() is never called for first_order == 0, which is why
1365 * free_by_rcu_order == 0 can be used as criterion to know if free must
1366 * be called.
1367 */
1475579c 1368static
4105056a 1369void fini_table(struct cds_lfht *ht,
93d46c39 1370 unsigned long first_order, unsigned long last_order)
1475579c 1371{
93d46c39 1372 long i;
48f1b16d 1373 unsigned long free_by_rcu_order = 0;
1475579c 1374
93d46c39
LJ
1375 dbg_printf("fini table: first_order %lu last_order %lu\n",
1376 first_order, last_order);
d0d8f9aa 1377 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1378 for (i = last_order; i >= first_order; i--) {
4105056a 1379 unsigned long len;
1475579c 1380
4f6e90b7 1381 len = 1UL << (i - 1);
e15df1cc 1382 dbg_printf("fini order %ld len: %lu\n", i, len);
4105056a 1383
4d676753 1384 /* Stop shrink if the resize target changes under us */
7b3893e4 1385 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1386 break;
1387
1388 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1389 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1390
1391 /*
1392 * We need to wait for all add operations to reach Q.S. (and
1393 * thus use the new table for lookups) before we can start
1ee8f000 1394 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1395 * return a logically removed node as insert position.
1396 */
7b17c13e 1397 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1398 if (free_by_rcu_order)
1399 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1400
21263e21 1401 /*
1ee8f000
LJ
1402 * Set "removed" flag in bucket nodes about to be removed.
1403 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1404 * Concurrent add/remove operation are helping us doing
1405 * the gc.
21263e21 1406 */
4105056a
MD
1407 remove_table(ht, i, len);
1408
48f1b16d 1409 free_by_rcu_order = i;
4105056a
MD
1410
1411 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1412 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1413 break;
1414 }
0d14ceb2 1415
48f1b16d 1416 if (free_by_rcu_order) {
7b17c13e 1417 ht->flavor->update_synchronize_rcu();
48f1b16d 1418 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1419 }
1475579c
MD
1420}
1421
ff0d69de 1422static
1ee8f000 1423void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1424{
04db56f8 1425 struct cds_lfht_node *prev, *node;
9d72a73f 1426 unsigned long order, len, i;
ff0d69de 1427
48f1b16d 1428 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1429
9d72a73f
LJ
1430 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1431 node = bucket_at(ht, 0);
1432 node->next = flag_bucket(get_end());
1433 node->reverse_hash = 0;
ff0d69de 1434
5bc6b66f 1435 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
ff0d69de 1436 len = 1UL << (order - 1);
48f1b16d 1437 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1438
9d72a73f
LJ
1439 for (i = 0; i < len; i++) {
1440 /*
1441 * Now, we are trying to init the node with the
1442 * hash=(len+i) (which is also a bucket with the
1443 * index=(len+i)) and insert it into the hash table,
1444 * so this node has to be inserted after the bucket
1445 * with the index=(len+i)&(len-1)=i. And because there
1446 * is no other non-bucket node nor bucket node with
1447 * larger index/hash inserted, so the bucket node
1448 * being inserted should be inserted directly linked
1449 * after the bucket node with index=i.
1450 */
1451 prev = bucket_at(ht, i);
1452 node = bucket_at(ht, len + i);
ff0d69de 1453
1ee8f000 1454 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1455 order, len + i, len + i);
1456 node->reverse_hash = bit_reverse_ulong(len + i);
1457
1458 /* insert after prev */
1459 assert(is_bucket(prev->next));
ff0d69de 1460 node->next = prev->next;
1ee8f000 1461 prev->next = flag_bucket(node);
ff0d69de
LJ
1462 }
1463 }
1464}
1465
0422d92c 1466struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1467 unsigned long min_nr_alloc_buckets,
747d725c 1468 unsigned long max_nr_buckets,
b8af5011 1469 int flags,
0b6aa001 1470 const struct cds_lfht_mm_type *mm,
7b17c13e 1471 const struct rcu_flavor_struct *flavor,
b7d619b0 1472 pthread_attr_t *attr)
abc490a1 1473{
14044b37 1474 struct cds_lfht *ht;
24365af7 1475 unsigned long order;
abc490a1 1476
0722081a
LJ
1477 /* min_nr_alloc_buckets must be power of two */
1478 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1479 return NULL;
747d725c 1480
8129be4e 1481 /* init_size must be power of two */
5488222b 1482 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1483 return NULL;
747d725c 1484
c1888f3a
MD
1485 /*
1486 * Memory management plugin default.
1487 */
1488 if (!mm) {
5a2141a7
MD
1489 if (CAA_BITS_PER_LONG > 32
1490 && max_nr_buckets
c1888f3a
MD
1491 && max_nr_buckets <= (1ULL << 32)) {
1492 /*
1493 * For 64-bit architectures, with max number of
1494 * buckets small enough not to use the entire
1495 * 64-bit memory mapping space (and allowing a
1496 * fair number of hash table instances), use the
1497 * mmap allocator, which is faster than the
1498 * order allocator.
1499 */
1500 mm = &cds_lfht_mm_mmap;
1501 } else {
1502 /*
1503 * The fallback is to use the order allocator.
1504 */
1505 mm = &cds_lfht_mm_order;
1506 }
1507 }
1508
0b6aa001
LJ
1509 /* max_nr_buckets == 0 for order based mm means infinite */
1510 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1511 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1512
1513 /* max_nr_buckets must be power of two */
1514 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1515 return NULL;
1516
0722081a 1517 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1518 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1519 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1520 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1521
1522 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1523 assert(ht);
0b6aa001
LJ
1524 assert(ht->mm == mm);
1525 assert(ht->bucket_at == mm->bucket_at);
1526
b5d6b20f 1527 ht->flags = flags;
7b17c13e 1528 ht->flavor = flavor;
b7d619b0 1529 ht->resize_attr = attr;
5afadd12 1530 alloc_split_items_count(ht);
abc490a1
MD
1531 /* this mutex should not nest in read-side C.S. */
1532 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1533 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1534 ht->resize_target = 1UL << order;
1ee8f000 1535 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1536 ht->size = 1UL << order;
abc490a1
MD
1537 return ht;
1538}
1539
6f554439 1540void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1541 cds_lfht_match_fct match, const void *key,
6f554439 1542 struct cds_lfht_iter *iter)
2ed95849 1543{
04db56f8 1544 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1545 unsigned long reverse_hash, size;
2ed95849 1546
abc490a1 1547 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1548
7b3893e4 1549 size = rcu_dereference(ht->size);
04db56f8 1550 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1551 /* We can always skip the bucket node initially */
04db56f8 1552 node = rcu_dereference(bucket->next);
bb7b2f26 1553 node = clear_flag(node);
2ed95849 1554 for (;;) {
8ed51e04 1555 if (caa_unlikely(is_end(node))) {
96ad1112 1556 node = next = NULL;
abc490a1 1557 break;
bb7b2f26 1558 }
04db56f8 1559 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1560 node = next = NULL;
abc490a1 1561 break;
2ed95849 1562 }
04db56f8 1563 next = rcu_dereference(node->next);
7f52427b 1564 assert(node == clear_flag(node));
8ed51e04 1565 if (caa_likely(!is_removed(next))
1ee8f000 1566 && !is_bucket(next)
04db56f8 1567 && node->reverse_hash == reverse_hash
0422d92c 1568 && caa_likely(match(node, key))) {
273399de 1569 break;
2ed95849 1570 }
1b81fe1a 1571 node = clear_flag(next);
2ed95849 1572 }
a85eff52 1573 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1574 iter->node = node;
1575 iter->next = next;
abc490a1 1576}
e0ba718a 1577
0422d92c 1578void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1579 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1580{
adc0de68 1581 struct cds_lfht_node *node, *next;
a481e5ff 1582 unsigned long reverse_hash;
a481e5ff 1583
adc0de68 1584 node = iter->node;
04db56f8 1585 reverse_hash = node->reverse_hash;
adc0de68 1586 next = iter->next;
a481e5ff
MD
1587 node = clear_flag(next);
1588
1589 for (;;) {
8ed51e04 1590 if (caa_unlikely(is_end(node))) {
96ad1112 1591 node = next = NULL;
a481e5ff 1592 break;
bb7b2f26 1593 }
04db56f8 1594 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1595 node = next = NULL;
a481e5ff
MD
1596 break;
1597 }
04db56f8 1598 next = rcu_dereference(node->next);
8ed51e04 1599 if (caa_likely(!is_removed(next))
1ee8f000 1600 && !is_bucket(next)
04db56f8 1601 && caa_likely(match(node, key))) {
a481e5ff
MD
1602 break;
1603 }
1604 node = clear_flag(next);
1605 }
a85eff52 1606 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
adc0de68
MD
1607 iter->node = node;
1608 iter->next = next;
a481e5ff
MD
1609}
1610
4e9b9fbf
MD
1611void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1612{
1613 struct cds_lfht_node *node, *next;
1614
853395e1 1615 node = clear_flag(iter->next);
4e9b9fbf 1616 for (;;) {
8ed51e04 1617 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1618 node = next = NULL;
1619 break;
1620 }
04db56f8 1621 next = rcu_dereference(node->next);
8ed51e04 1622 if (caa_likely(!is_removed(next))
1ee8f000 1623 && !is_bucket(next)) {
4e9b9fbf
MD
1624 break;
1625 }
1626 node = clear_flag(next);
1627 }
a85eff52 1628 assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next)));
4e9b9fbf
MD
1629 iter->node = node;
1630 iter->next = next;
1631}
1632
1633void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1634{
4e9b9fbf 1635 /*
1ee8f000 1636 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1637 * first node of the linked list.
1638 */
9d72a73f 1639 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1640 cds_lfht_next(ht, iter);
1641}
1642
0422d92c
MD
1643void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1644 struct cds_lfht_node *node)
abc490a1 1645{
0422d92c 1646 unsigned long size;
ab7d5fc6 1647
709bacf9 1648 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1649 size = rcu_dereference(ht->size);
91a75cc5 1650 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1651 ht_count_add(ht, size, hash);
3eca1b8c
MD
1652}
1653
14044b37 1654struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1655 unsigned long hash,
0422d92c 1656 cds_lfht_match_fct match,
996ff57c 1657 const void *key,
48ed1c18 1658 struct cds_lfht_node *node)
3eca1b8c 1659{
0422d92c 1660 unsigned long size;
83beee94 1661 struct cds_lfht_iter iter;
3eca1b8c 1662
709bacf9 1663 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1664 size = rcu_dereference(ht->size);
91a75cc5 1665 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1666 if (iter.node == node)
14360f1c 1667 ht_count_add(ht, size, hash);
83beee94 1668 return iter.node;
2ed95849
MD
1669}
1670
9357c415 1671struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1672 unsigned long hash,
0422d92c 1673 cds_lfht_match_fct match,
996ff57c 1674 const void *key,
48ed1c18
MD
1675 struct cds_lfht_node *node)
1676{
0422d92c 1677 unsigned long size;
83beee94 1678 struct cds_lfht_iter iter;
48ed1c18 1679
709bacf9 1680 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1681 size = rcu_dereference(ht->size);
83beee94 1682 for (;;) {
91a75cc5 1683 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1684 if (iter.node == node) {
14360f1c 1685 ht_count_add(ht, size, hash);
83beee94
MD
1686 return NULL;
1687 }
1688
1689 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1690 return iter.node;
1691 }
48ed1c18
MD
1692}
1693
2e79c445
MD
1694int cds_lfht_replace(struct cds_lfht *ht,
1695 struct cds_lfht_iter *old_iter,
1696 unsigned long hash,
1697 cds_lfht_match_fct match,
1698 const void *key,
9357c415
MD
1699 struct cds_lfht_node *new_node)
1700{
1701 unsigned long size;
1702
709bacf9 1703 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1704 if (!old_iter->node)
1705 return -ENOENT;
1706 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1707 return -EINVAL;
1708 if (caa_unlikely(!match(old_iter->node, key)))
1709 return -EINVAL;
7b3893e4 1710 size = rcu_dereference(ht->size);
9357c415
MD
1711 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1712 new_node);
1713}
1714
bc8c3c74 1715int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1716{
95bc7fb9 1717 unsigned long size;
df44348d 1718 int ret;
abc490a1 1719
7b3893e4 1720 size = rcu_dereference(ht->size);
bc8c3c74 1721 ret = _cds_lfht_del(ht, size, node);
14360f1c 1722 if (!ret) {
95bc7fb9
MD
1723 unsigned long hash;
1724
bc8c3c74 1725 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1726 ht_count_del(ht, size, hash);
1727 }
df44348d 1728 return ret;
2ed95849 1729}
ab7d5fc6 1730
df55172a
MD
1731int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1732{
a85eff52 1733 return is_removed(CMM_LOAD_SHARED(node->next));
df55172a
MD
1734}
1735
abc490a1 1736static
1ee8f000 1737int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1738{
14044b37 1739 struct cds_lfht_node *node;
4105056a 1740 unsigned long order, i, size;
674f7a69 1741
abc490a1 1742 /* Check that the table is empty */
9d72a73f 1743 node = bucket_at(ht, 0);
abc490a1 1744 do {
04db56f8 1745 node = clear_flag(node)->next;
1ee8f000 1746 if (!is_bucket(node))
abc490a1 1747 return -EPERM;
273399de 1748 assert(!is_removed(node));
2f943cd7 1749 assert(!is_removal_owner(node));
bb7b2f26 1750 } while (!is_end(node));
4105056a
MD
1751 /*
1752 * size accessed without rcu_dereference because hash table is
1753 * being destroyed.
1754 */
7b3893e4 1755 size = ht->size;
1f67ba50 1756 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1757 for (i = 0; i < size; i++) {
1758 node = bucket_at(ht, i);
1759 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1760 i, i, bit_reverse_ulong(node->reverse_hash));
1761 assert(is_bucket(node->next));
1762 }
24365af7 1763
5bc6b66f 1764 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1765 cds_lfht_free_bucket_table(ht, order);
5488222b 1766
abc490a1 1767 return 0;
674f7a69
MD
1768}
1769
1770/*
1771 * Should only be called when no more concurrent readers nor writers can
1772 * possibly access the table.
1773 */
b7d619b0 1774int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1775{
a1e5e232 1776 int ret, was_online;
5e28c532 1777
848d4088 1778 /* Wait for in-flight resize operations to complete */
24953e08
MD
1779 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1780 cmm_smp_mb(); /* Store destroy before load resize */
a1e5e232
MD
1781 was_online = ht->flavor->read_ongoing();
1782 if (was_online)
1783 ht->flavor->thread_offline();
10e68472
MD
1784 /* Calling with RCU read-side held is an error. */
1785 if (ht->flavor->read_ongoing()) {
1786 ret = -EINVAL;
1787 if (was_online)
1788 ht->flavor->thread_online();
1789 goto end;
1790 }
848d4088
MD
1791 while (uatomic_read(&ht->in_progress_resize))
1792 poll(NULL, 0, 100); /* wait for 100ms */
a1e5e232
MD
1793 if (was_online)
1794 ht->flavor->thread_online();
1ee8f000 1795 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1796 if (ret)
1797 return ret;
5afadd12 1798 free_split_items_count(ht);
b7d619b0
MD
1799 if (attr)
1800 *attr = ht->resize_attr;
98808fb1 1801 poison_free(ht);
10e68472 1802end:
5e28c532 1803 return ret;
674f7a69
MD
1804}
1805
14044b37 1806void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1807 long *approx_before,
273399de 1808 unsigned long *count,
d933dd0e 1809 long *approx_after)
273399de 1810{
14044b37 1811 struct cds_lfht_node *node, *next;
caf3653d 1812 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1813
7ed7682f 1814 *approx_before = 0;
5afadd12 1815 if (ht->split_count) {
973e5e1b
MD
1816 int i;
1817
4c42f1b8
LJ
1818 for (i = 0; i < split_count_mask + 1; i++) {
1819 *approx_before += uatomic_read(&ht->split_count[i].add);
1820 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1821 }
1822 }
1823
273399de 1824 *count = 0;
273399de 1825
1ee8f000 1826 /* Count non-bucket nodes in the table */
9d72a73f 1827 node = bucket_at(ht, 0);
273399de 1828 do {
04db56f8 1829 next = rcu_dereference(node->next);
b198f0fd 1830 if (is_removed(next)) {
1ee8f000 1831 if (!is_bucket(next))
caf3653d 1832 (nr_removed)++;
973e5e1b 1833 else
1ee8f000
LJ
1834 (nr_bucket)++;
1835 } else if (!is_bucket(next))
273399de 1836 (*count)++;
24365af7 1837 else
1ee8f000 1838 (nr_bucket)++;
273399de 1839 node = clear_flag(next);
bb7b2f26 1840 } while (!is_end(node));
caf3653d 1841 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1842 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1843 *approx_after = 0;
5afadd12 1844 if (ht->split_count) {
973e5e1b
MD
1845 int i;
1846
4c42f1b8
LJ
1847 for (i = 0; i < split_count_mask + 1; i++) {
1848 *approx_after += uatomic_read(&ht->split_count[i].add);
1849 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1850 }
1851 }
273399de
MD
1852}
1853
1475579c 1854/* called with resize mutex held */
abc490a1 1855static
4105056a 1856void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1857 unsigned long old_size, unsigned long new_size)
abc490a1 1858{
1475579c 1859 unsigned long old_order, new_order;
1475579c 1860
5bc6b66f
MD
1861 old_order = cds_lfht_get_count_order_ulong(old_size);
1862 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1863 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1864 old_size, old_order, new_size, new_order);
1475579c 1865 assert(new_size > old_size);
93d46c39 1866 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1867}
1868
1869/* called with resize mutex held */
1870static
4105056a 1871void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1872 unsigned long old_size, unsigned long new_size)
464a1ec9 1873{
1475579c 1874 unsigned long old_order, new_order;
464a1ec9 1875
d0d8f9aa 1876 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1877 old_order = cds_lfht_get_count_order_ulong(old_size);
1878 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1879 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1880 old_size, old_order, new_size, new_order);
1475579c 1881 assert(new_size < old_size);
1475579c 1882
1ee8f000 1883 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1884 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1885}
1886
1475579c
MD
1887
1888/* called with resize mutex held */
1889static
1890void _do_cds_lfht_resize(struct cds_lfht *ht)
1891{
1892 unsigned long new_size, old_size;
4105056a
MD
1893
1894 /*
1895 * Resize table, re-do if the target size has changed under us.
1896 */
1897 do {
d2be3620
MD
1898 assert(uatomic_read(&ht->in_progress_resize));
1899 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1900 break;
7b3893e4
LJ
1901 ht->resize_initiated = 1;
1902 old_size = ht->size;
1903 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1904 if (old_size < new_size)
1905 _do_cds_lfht_grow(ht, old_size, new_size);
1906 else if (old_size > new_size)
1907 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 1908 ht->resize_initiated = 0;
4105056a
MD
1909 /* write resize_initiated before read resize_target */
1910 cmm_smp_mb();
7b3893e4 1911 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
1912}
1913
abc490a1 1914static
ab65b890 1915unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1916{
7b3893e4 1917 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
1918}
1919
1475579c 1920static
4105056a 1921void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1922 unsigned long count)
1475579c 1923{
d0d8f9aa 1924 count = max(count, MIN_TABLE_SIZE);
747d725c 1925 count = min(count, ht->max_nr_buckets);
7b3893e4 1926 uatomic_set(&ht->resize_target, count);
1475579c
MD
1927}
1928
1929void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1930{
a1e5e232
MD
1931 int was_online;
1932
a1e5e232
MD
1933 was_online = ht->flavor->read_ongoing();
1934 if (was_online)
1935 ht->flavor->thread_offline();
10e68472
MD
1936 /* Calling with RCU read-side held is an error. */
1937 if (ht->flavor->read_ongoing()) {
1938 static int print_once;
1939
1940 if (!CMM_LOAD_SHARED(print_once))
1941 fprintf(stderr, "[error] rculfhash: cds_lfht_resize "
1942 "called with RCU read-side lock held.\n");
1943 CMM_STORE_SHARED(print_once, 1);
1944 assert(0);
1945 goto end;
1946 }
1947 resize_target_update_count(ht, new_size);
1948 CMM_STORE_SHARED(ht->resize_initiated, 1);
1475579c
MD
1949 pthread_mutex_lock(&ht->resize_mutex);
1950 _do_cds_lfht_resize(ht);
1951 pthread_mutex_unlock(&ht->resize_mutex);
10e68472 1952end:
a1e5e232
MD
1953 if (was_online)
1954 ht->flavor->thread_online();
abc490a1 1955}
464a1ec9 1956
abc490a1
MD
1957static
1958void do_resize_cb(struct rcu_head *head)
1959{
1960 struct rcu_resize_work *work =
1961 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1962 struct cds_lfht *ht = work->ht;
abc490a1 1963
7b17c13e 1964 ht->flavor->thread_offline();
abc490a1 1965 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1966 _do_cds_lfht_resize(ht);
abc490a1 1967 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1968 ht->flavor->thread_online();
98808fb1 1969 poison_free(work);
848d4088
MD
1970 cmm_smp_mb(); /* finish resize before decrement */
1971 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1972}
1973
abc490a1 1974static
f1f119ee 1975void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1976{
abc490a1
MD
1977 struct rcu_resize_work *work;
1978
4105056a
MD
1979 /* Store resize_target before read resize_initiated */
1980 cmm_smp_mb();
7b3893e4 1981 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
848d4088 1982 uatomic_inc(&ht->in_progress_resize);
59290e9d 1983 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1984 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1985 uatomic_dec(&ht->in_progress_resize);
59290e9d 1986 return;
ed35e6d8 1987 }
f9830efd 1988 work = malloc(sizeof(*work));
741f378e
MD
1989 if (work == NULL) {
1990 dbg_printf("error allocating resize work, bailing out\n");
1991 uatomic_dec(&ht->in_progress_resize);
1992 return;
1993 }
f9830efd 1994 work->ht = ht;
7b17c13e 1995 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
7b3893e4 1996 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 1997 }
ab7d5fc6 1998}
3171717f 1999
f1f119ee
LJ
2000static
2001void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
2002{
2003 unsigned long target_size = size << growth;
2004
747d725c 2005 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
2006 if (resize_target_grow(ht, target_size) >= target_size)
2007 return;
2008
2009 __cds_lfht_resize_lazy_launch(ht);
2010}
2011
89bb121d
LJ
2012/*
2013 * We favor grow operations over shrink. A shrink operation never occurs
2014 * if a grow operation is queued for lazy execution. A grow operation
2015 * cancels any pending shrink lazy execution.
2016 */
3171717f 2017static
4105056a 2018void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
2019 unsigned long count)
2020{
b8af5011
MD
2021 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
2022 return;
d0d8f9aa 2023 count = max(count, MIN_TABLE_SIZE);
747d725c 2024 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
2025 if (count == size)
2026 return; /* Already the right size, no resize needed */
2027 if (count > size) { /* lazy grow */
2028 if (resize_target_grow(ht, count) >= count)
2029 return;
2030 } else { /* lazy shrink */
2031 for (;;) {
2032 unsigned long s;
2033
7b3893e4 2034 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
2035 if (s == size)
2036 break; /* no resize needed */
2037 if (s > size)
2038 return; /* growing is/(was just) in progress */
2039 if (s <= count)
2040 return; /* some other thread do shrink */
2041 size = s;
2042 }
2043 }
f1f119ee 2044 __cds_lfht_resize_lazy_launch(ht);
3171717f 2045}
This page took 0.26087 seconds and 4 git commands to generate.