rculfhash: use do {} while (0) for dbg_printf()
[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
120 * cds_lfht_add_unique (failure).
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
MD
131 *
132 * We define "write" operations as any of cds_lfht_add,
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
MD
227 * A bit of ascii art explanation:
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.
1475579c
MD
231 *
232 * This shows the nodes for a small table ordered by reversed bits:
233 *
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
243 *
244 * This shows the nodes in order of non-reversed bits, linked by
245 * reversed-bit order.
246 *
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"
2ed95849 269#include <urcu.h>
abc490a1 270#include <urcu-call-rcu.h>
7b17c13e 271#include <urcu-flavor.h>
a42cc659
MD
272#include <urcu/arch.h>
273#include <urcu/uatomic.h>
a42cc659 274#include <urcu/compiler.h>
abc490a1 275#include <urcu/rculfhash.h>
0b6aa001 276#include <rculfhash-internal.h>
5e28c532 277#include <stdio.h>
464a1ec9 278#include <pthread.h>
44395fb7 279
f8994aee 280/*
4c42f1b8 281 * Split-counters lazily update the global counter each 1024
f8994aee
MD
282 * addition/removal. It automatically keeps track of resize required.
283 * We use the bucket length as indicator for need to expand for small
284 * tables and machines lacking per-cpu data suppport.
285 */
286#define COUNT_COMMIT_ORDER 10
4ddbb355 287#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
288#define CHAIN_LEN_TARGET 1
289#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 290
cd95516d 291/*
76a73da8 292 * Define the minimum table size.
cd95516d 293 */
d0d8f9aa
LJ
294#define MIN_TABLE_ORDER 0
295#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 296
b7d619b0 297/*
1ee8f000 298 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 299 */
6083a889
MD
300#define MIN_PARTITION_PER_THREAD_ORDER 12
301#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 302
d95bd160
MD
303/*
304 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 305 * It indicates that no node must attach to the node scheduled for
b198f0fd 306 * removal, and that node garbage collection must be performed.
1ee8f000 307 * The bucket flag does not require to be updated atomically with the
d95bd160 308 * pointer, but it is added as a pointer low bit flag to save space.
1f67ba50
MD
309 * The "removal owner" flag is used to detect which of the "del"
310 * operation that has set the "removed flag" gets to return the removed
311 * node to its caller. Note that the replace operation does not need to
312 * iteract with the "removal owner" flag, because it validates that
313 * the "removed" flag is not set before performing its cmpxchg.
d95bd160 314 */
d37166c6 315#define REMOVED_FLAG (1UL << 0)
1ee8f000 316#define BUCKET_FLAG (1UL << 1)
db00ccc3
MD
317#define REMOVAL_OWNER_FLAG (1UL << 2)
318#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 319
bb7b2f26 320/* Value of the end pointer. Should not interact with flags. */
f9c80341 321#define END_VALUE NULL
bb7b2f26 322
7f52427b
MD
323/*
324 * ht_items_count: Split-counters counting the number of node addition
325 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
326 * is set at hash table creation.
327 *
328 * These are free-running counters, never reset to zero. They count the
329 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
330 * operations to update the global counter. We choose a power-of-2 value
331 * for the trigger to deal with 32 or 64-bit overflow of the counter.
332 */
df44348d 333struct ht_items_count {
860d07e8 334 unsigned long add, del;
df44348d
MD
335} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
336
7f52427b
MD
337/*
338 * rcu_resize_work: Contains arguments passed to RCU worker thread
339 * responsible for performing lazy resize.
340 */
abc490a1
MD
341struct rcu_resize_work {
342 struct rcu_head head;
14044b37 343 struct cds_lfht *ht;
abc490a1 344};
2ed95849 345
7f52427b
MD
346/*
347 * partition_resize_work: Contains arguments passed to worker threads
348 * executing the hash table resize on partitions of the hash table
349 * assigned to each processor's worker thread.
350 */
b7d619b0 351struct partition_resize_work {
1af6e26e 352 pthread_t thread_id;
b7d619b0
MD
353 struct cds_lfht *ht;
354 unsigned long i, start, len;
355 void (*fct)(struct cds_lfht *ht, unsigned long i,
356 unsigned long start, unsigned long len);
357};
358
abc490a1
MD
359/*
360 * Algorithm to reverse bits in a word by lookup table, extended to
361 * 64-bit words.
f9830efd 362 * Source:
abc490a1 363 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 364 * Originally from Public Domain.
abc490a1
MD
365 */
366
367static const uint8_t BitReverseTable256[256] =
2ed95849 368{
abc490a1
MD
369#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
370#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
371#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
372 R6(0), R6(2), R6(1), R6(3)
373};
374#undef R2
375#undef R4
376#undef R6
2ed95849 377
abc490a1
MD
378static
379uint8_t bit_reverse_u8(uint8_t v)
380{
381 return BitReverseTable256[v];
382}
ab7d5fc6 383
abc490a1
MD
384static __attribute__((unused))
385uint32_t bit_reverse_u32(uint32_t v)
386{
387 return ((uint32_t) bit_reverse_u8(v) << 24) |
388 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
389 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
390 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
391}
392
abc490a1
MD
393static __attribute__((unused))
394uint64_t bit_reverse_u64(uint64_t v)
2ed95849 395{
abc490a1
MD
396 return ((uint64_t) bit_reverse_u8(v) << 56) |
397 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
398 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
399 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
400 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
401 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
402 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
403 ((uint64_t) bit_reverse_u8(v >> 56));
404}
405
406static
407unsigned long bit_reverse_ulong(unsigned long v)
408{
409#if (CAA_BITS_PER_LONG == 32)
410 return bit_reverse_u32(v);
411#else
412 return bit_reverse_u64(v);
413#endif
414}
415
f9830efd 416/*
24365af7
MD
417 * fls: returns the position of the most significant bit.
418 * Returns 0 if no bit is set, else returns the position of the most
419 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 420 */
24365af7
MD
421#if defined(__i386) || defined(__x86_64)
422static inline
423unsigned int fls_u32(uint32_t x)
f9830efd 424{
24365af7
MD
425 int r;
426
427 asm("bsrl %1,%0\n\t"
428 "jnz 1f\n\t"
429 "movl $-1,%0\n\t"
430 "1:\n\t"
431 : "=r" (r) : "rm" (x));
432 return r + 1;
433}
434#define HAS_FLS_U32
435#endif
436
437#if defined(__x86_64)
438static inline
439unsigned int fls_u64(uint64_t x)
440{
441 long r;
442
443 asm("bsrq %1,%0\n\t"
444 "jnz 1f\n\t"
445 "movq $-1,%0\n\t"
446 "1:\n\t"
447 : "=r" (r) : "rm" (x));
448 return r + 1;
449}
450#define HAS_FLS_U64
451#endif
452
453#ifndef HAS_FLS_U64
454static __attribute__((unused))
455unsigned int fls_u64(uint64_t x)
456{
457 unsigned int r = 64;
458
459 if (!x)
460 return 0;
461
462 if (!(x & 0xFFFFFFFF00000000ULL)) {
463 x <<= 32;
464 r -= 32;
465 }
466 if (!(x & 0xFFFF000000000000ULL)) {
467 x <<= 16;
468 r -= 16;
469 }
470 if (!(x & 0xFF00000000000000ULL)) {
471 x <<= 8;
472 r -= 8;
473 }
474 if (!(x & 0xF000000000000000ULL)) {
475 x <<= 4;
476 r -= 4;
477 }
478 if (!(x & 0xC000000000000000ULL)) {
479 x <<= 2;
480 r -= 2;
481 }
482 if (!(x & 0x8000000000000000ULL)) {
483 x <<= 1;
484 r -= 1;
485 }
486 return r;
487}
488#endif
489
490#ifndef HAS_FLS_U32
491static __attribute__((unused))
492unsigned int fls_u32(uint32_t x)
493{
494 unsigned int r = 32;
f9830efd 495
24365af7
MD
496 if (!x)
497 return 0;
498 if (!(x & 0xFFFF0000U)) {
499 x <<= 16;
500 r -= 16;
501 }
502 if (!(x & 0xFF000000U)) {
503 x <<= 8;
504 r -= 8;
505 }
506 if (!(x & 0xF0000000U)) {
507 x <<= 4;
508 r -= 4;
509 }
510 if (!(x & 0xC0000000U)) {
511 x <<= 2;
512 r -= 2;
513 }
514 if (!(x & 0x80000000U)) {
515 x <<= 1;
516 r -= 1;
517 }
518 return r;
519}
520#endif
521
5bc6b66f 522unsigned int cds_lfht_fls_ulong(unsigned long x)
f9830efd 523{
6887cc5e 524#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
525 return fls_u32(x);
526#else
527 return fls_u64(x);
528#endif
529}
f9830efd 530
920f8ef6
LJ
531/*
532 * Return the minimum order for which x <= (1UL << order).
533 * Return -1 if x is 0.
534 */
5bc6b66f 535int cds_lfht_get_count_order_u32(uint32_t x)
24365af7 536{
920f8ef6
LJ
537 if (!x)
538 return -1;
24365af7 539
920f8ef6 540 return fls_u32(x - 1);
24365af7
MD
541}
542
920f8ef6
LJ
543/*
544 * Return the minimum order for which x <= (1UL << order).
545 * Return -1 if x is 0.
546 */
5bc6b66f 547int cds_lfht_get_count_order_ulong(unsigned long x)
24365af7 548{
920f8ef6
LJ
549 if (!x)
550 return -1;
24365af7 551
5bc6b66f 552 return cds_lfht_fls_ulong(x - 1);
f9830efd
MD
553}
554
555static
ab65b890 556void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 557
f8994aee 558static
4105056a 559void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
560 unsigned long count);
561
df44348d 562static long nr_cpus_mask = -1;
4c42f1b8
LJ
563static long split_count_mask = -1;
564
4ddbb355 565#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
566static void ht_init_nr_cpus_mask(void)
567{
568 long maxcpus;
569
570 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
571 if (maxcpus <= 0) {
572 nr_cpus_mask = -2;
573 return;
574 }
575 /*
576 * round up number of CPUs to next power of two, so we
577 * can use & for modulo.
578 */
5bc6b66f 579 maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
4c42f1b8
LJ
580 nr_cpus_mask = maxcpus - 1;
581}
4ddbb355
LJ
582#else /* #if defined(HAVE_SYSCONF) */
583static void ht_init_nr_cpus_mask(void)
584{
585 nr_cpus_mask = -2;
586}
587#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
588
589static
5afadd12 590void alloc_split_items_count(struct cds_lfht *ht)
df44348d
MD
591{
592 struct ht_items_count *count;
593
4c42f1b8
LJ
594 if (nr_cpus_mask == -1) {
595 ht_init_nr_cpus_mask();
4ddbb355
LJ
596 if (nr_cpus_mask < 0)
597 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
598 else
599 split_count_mask = nr_cpus_mask;
df44348d 600 }
4c42f1b8 601
4ddbb355 602 assert(split_count_mask >= 0);
5afadd12
LJ
603
604 if (ht->flags & CDS_LFHT_ACCOUNTING) {
605 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
606 assert(ht->split_count);
607 } else {
608 ht->split_count = NULL;
609 }
df44348d
MD
610}
611
612static
5afadd12 613void free_split_items_count(struct cds_lfht *ht)
df44348d 614{
5afadd12 615 poison_free(ht->split_count);
df44348d
MD
616}
617
14360f1c 618#if defined(HAVE_SCHED_GETCPU)
df44348d 619static
14360f1c 620int ht_get_split_count_index(unsigned long hash)
df44348d
MD
621{
622 int cpu;
623
4c42f1b8 624 assert(split_count_mask >= 0);
df44348d 625 cpu = sched_getcpu();
8ed51e04 626 if (caa_unlikely(cpu < 0))
14360f1c 627 return hash & split_count_mask;
df44348d 628 else
4c42f1b8 629 return cpu & split_count_mask;
df44348d 630}
14360f1c
LJ
631#else /* #if defined(HAVE_SCHED_GETCPU) */
632static
633int ht_get_split_count_index(unsigned long hash)
634{
635 return hash & split_count_mask;
636}
637#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
638
639static
14360f1c 640void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 641{
4c42f1b8
LJ
642 unsigned long split_count;
643 int index;
314558bf 644 long count;
df44348d 645
8ed51e04 646 if (caa_unlikely(!ht->split_count))
3171717f 647 return;
14360f1c 648 index = ht_get_split_count_index(hash);
4c42f1b8 649 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
314558bf
MD
650 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
651 return;
652 /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */
653
654 dbg_printf("add split count %lu\n", split_count);
655 count = uatomic_add_return(&ht->count,
656 1UL << COUNT_COMMIT_ORDER);
4c299dcb 657 if (caa_likely(count & (count - 1)))
314558bf
MD
658 return;
659 /* Only if global count is power of 2 */
660
661 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
662 return;
663 dbg_printf("add set global %ld\n", count);
664 cds_lfht_resize_lazy_count(ht, size,
665 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
666}
667
668static
14360f1c 669void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 670{
4c42f1b8
LJ
671 unsigned long split_count;
672 int index;
314558bf 673 long count;
df44348d 674
8ed51e04 675 if (caa_unlikely(!ht->split_count))
3171717f 676 return;
14360f1c 677 index = ht_get_split_count_index(hash);
4c42f1b8 678 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
314558bf
MD
679 if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))
680 return;
681 /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */
682
683 dbg_printf("del split count %lu\n", split_count);
684 count = uatomic_add_return(&ht->count,
685 -(1UL << COUNT_COMMIT_ORDER));
4c299dcb 686 if (caa_likely(count & (count - 1)))
314558bf
MD
687 return;
688 /* Only if global count is power of 2 */
689
690 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
691 return;
692 dbg_printf("del set global %ld\n", count);
693 /*
694 * Don't shrink table if the number of nodes is below a
695 * certain threshold.
696 */
697 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
698 return;
699 cds_lfht_resize_lazy_count(ht, size,
700 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
701}
702
f9830efd 703static
4105056a 704void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 705{
f8994aee
MD
706 unsigned long count;
707
b8af5011
MD
708 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
709 return;
f8994aee
MD
710 count = uatomic_read(&ht->count);
711 /*
712 * Use bucket-local length for small table expand and for
713 * environments lacking per-cpu data support.
714 */
715 if (count >= (1UL << COUNT_COMMIT_ORDER))
716 return;
24365af7 717 if (chain_len > 100)
f0c29ed7 718 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 719 chain_len);
3390d470 720 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
ab65b890 721 cds_lfht_resize_lazy_grow(ht, size,
5bc6b66f 722 cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
723}
724
abc490a1 725static
14044b37 726struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 727{
14044b37 728 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
729}
730
731static
14044b37 732int is_removed(struct cds_lfht_node *node)
abc490a1 733{
d37166c6 734 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
735}
736
737static
14044b37 738struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 739{
14044b37 740 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
741}
742
f5596c94 743static
1ee8f000 744int is_bucket(struct cds_lfht_node *node)
f5596c94 745{
1ee8f000 746 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
747}
748
749static
1ee8f000 750struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 751{
1ee8f000 752 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 753}
bb7b2f26 754
db00ccc3
MD
755static
756int is_removal_owner(struct cds_lfht_node *node)
757{
758 return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
759}
760
761static
762struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
763{
764 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
765}
766
bb7b2f26
MD
767static
768struct cds_lfht_node *get_end(void)
769{
770 return (struct cds_lfht_node *) END_VALUE;
771}
772
773static
774int is_end(struct cds_lfht_node *node)
775{
776 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
777}
778
abc490a1 779static
ab65b890
LJ
780unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
781 unsigned long v)
abc490a1
MD
782{
783 unsigned long old1, old2;
784
785 old1 = uatomic_read(ptr);
786 do {
787 old2 = old1;
788 if (old2 >= v)
f9830efd 789 return old2;
abc490a1 790 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 791 return old2;
abc490a1
MD
792}
793
48f1b16d
LJ
794static
795void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
796{
0b6aa001 797 return ht->mm->alloc_bucket_table(ht, order);
48f1b16d
LJ
798}
799
800/*
801 * cds_lfht_free_bucket_table() should be called with decreasing order.
802 * When cds_lfht_free_bucket_table(0) is called, it means the whole
803 * lfht is destroyed.
804 */
805static
806void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
807{
0b6aa001 808 return ht->mm->free_bucket_table(ht, order);
48f1b16d
LJ
809}
810
9d72a73f
LJ
811static inline
812struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 813{
0b6aa001 814 return ht->bucket_at(ht, index);
f4a9cc0b
LJ
815}
816
9d72a73f
LJ
817static inline
818struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
819 unsigned long hash)
820{
821 assert(size > 0);
822 return bucket_at(ht, hash & (size - 1));
823}
824
273399de
MD
825/*
826 * Remove all logically deleted nodes from a bucket up to a certain node key.
827 */
828static
1ee8f000 829void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 830{
14044b37 831 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 832
1ee8f000
LJ
833 assert(!is_bucket(bucket));
834 assert(!is_removed(bucket));
835 assert(!is_bucket(node));
c90201ac 836 assert(!is_removed(node));
273399de 837 for (;;) {
1ee8f000
LJ
838 iter_prev = bucket;
839 /* We can always skip the bucket node initially */
04db56f8 840 iter = rcu_dereference(iter_prev->next);
b4cb483f 841 assert(!is_removed(iter));
04db56f8 842 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 843 /*
1ee8f000 844 * We should never be called with bucket (start of chain)
bd4db153
MD
845 * and logically removed node (end of path compression
846 * marker) being the actual same node. This would be a
847 * bug in the algorithm implementation.
848 */
1ee8f000 849 assert(bucket != node);
273399de 850 for (;;) {
8ed51e04 851 if (caa_unlikely(is_end(iter)))
f9c80341 852 return;
04db56f8 853 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 854 return;
04db56f8 855 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 856 if (caa_likely(is_removed(next)))
273399de 857 break;
b453eae1 858 iter_prev = clear_flag(iter);
273399de
MD
859 iter = next;
860 }
b198f0fd 861 assert(!is_removed(iter));
1ee8f000
LJ
862 if (is_bucket(iter))
863 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
864 else
865 new_next = clear_flag(next);
04db56f8 866 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de
MD
867 }
868}
869
9357c415
MD
870static
871int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
872 struct cds_lfht_node *old_node,
3fb86f26 873 struct cds_lfht_node *old_next,
9357c415
MD
874 struct cds_lfht_node *new_node)
875{
04db56f8 876 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
877
878 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 879 return -ENOENT;
9357c415
MD
880
881 assert(!is_removed(old_node));
1ee8f000 882 assert(!is_bucket(old_node));
9357c415 883 assert(!is_removed(new_node));
1ee8f000 884 assert(!is_bucket(new_node));
9357c415 885 assert(new_node != old_node);
3fb86f26 886 for (;;) {
9357c415 887 /* Insert after node to be replaced */
9357c415
MD
888 if (is_removed(old_next)) {
889 /*
890 * Too late, the old node has been removed under us
891 * between lookup and replace. Fail.
892 */
7801dadd 893 return -ENOENT;
9357c415 894 }
feda2722
LJ
895 assert(old_next == clear_flag(old_next));
896 assert(new_node != old_next);
897 new_node->next = old_next;
9357c415
MD
898 /*
899 * Here is the whole trick for lock-free replace: we add
900 * the replacement node _after_ the node we want to
901 * replace by atomically setting its next pointer at the
902 * same time we set its removal flag. Given that
903 * the lookups/get next use an iterator aware of the
904 * next pointer, they will either skip the old node due
905 * to the removal flag and see the new node, or use
906 * the old node, but will not see the new one.
db00ccc3
MD
907 * This is a replacement of a node with another node
908 * that has the same value: we are therefore not
909 * removing a value from the hash table.
9357c415 910 */
04db56f8 911 ret_next = uatomic_cmpxchg(&old_node->next,
9357c415 912 old_next, flag_removed(new_node));
3fb86f26 913 if (ret_next == old_next)
7801dadd 914 break; /* We performed the replacement. */
3fb86f26
LJ
915 old_next = ret_next;
916 }
9357c415 917
9357c415
MD
918 /*
919 * Ensure that the old node is not visible to readers anymore:
920 * lookup for the node, and remove it (along with any other
921 * logically removed node) if found.
922 */
04db56f8
LJ
923 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
924 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 925
04db56f8 926 assert(is_removed(rcu_dereference(old_node->next)));
7801dadd 927 return 0;
9357c415
MD
928}
929
83beee94
MD
930/*
931 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
932 * mode. A NULL unique_ret allows creation of duplicate keys.
933 */
abc490a1 934static
83beee94 935void _cds_lfht_add(struct cds_lfht *ht,
91a75cc5 936 unsigned long hash,
0422d92c 937 cds_lfht_match_fct match,
996ff57c 938 const void *key,
83beee94
MD
939 unsigned long size,
940 struct cds_lfht_node *node,
941 struct cds_lfht_iter *unique_ret,
1ee8f000 942 int bucket_flag)
abc490a1 943{
14044b37 944 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 945 *return_node;
04db56f8 946 struct cds_lfht_node *bucket;
abc490a1 947
1ee8f000 948 assert(!is_bucket(node));
c90201ac 949 assert(!is_removed(node));
91a75cc5 950 bucket = lookup_bucket(ht, size, hash);
abc490a1 951 for (;;) {
adc0de68 952 uint32_t chain_len = 0;
abc490a1 953
11519af6
MD
954 /*
955 * iter_prev points to the non-removed node prior to the
956 * insert location.
11519af6 957 */
04db56f8 958 iter_prev = bucket;
1ee8f000 959 /* We can always skip the bucket node initially */
04db56f8
LJ
960 iter = rcu_dereference(iter_prev->next);
961 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 962 for (;;) {
8ed51e04 963 if (caa_unlikely(is_end(iter)))
273399de 964 goto insert;
04db56f8 965 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 966 goto insert;
238cc06e 967
1ee8f000
LJ
968 /* bucket node is the first node of the identical-hash-value chain */
969 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 970 goto insert;
238cc06e 971
04db56f8 972 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 973 if (caa_unlikely(is_removed(next)))
9dba85be 974 goto gc_node;
238cc06e
LJ
975
976 /* uniquely add */
83beee94 977 if (unique_ret
1ee8f000 978 && !is_bucket(next)
04db56f8 979 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
980 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
981
982 /*
983 * uniquely adding inserts the node as the first
984 * node of the identical-hash-value node chain.
985 *
986 * This semantic ensures no duplicated keys
987 * should ever be observable in the table
1f67ba50
MD
988 * (including traversing the table node by
989 * node by forward iterations)
238cc06e 990 */
04db56f8 991 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
992 if (!d_iter.node)
993 goto insert;
994
995 *unique_ret = d_iter;
83beee94 996 return;
48ed1c18 997 }
238cc06e 998
11519af6 999 /* Only account for identical reverse hash once */
04db56f8 1000 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 1001 && !is_bucket(next))
4105056a 1002 check_resize(ht, size, ++chain_len);
11519af6 1003 iter_prev = clear_flag(iter);
273399de 1004 iter = next;
abc490a1 1005 }
48ed1c18 1006
273399de 1007 insert:
7ec59d3b 1008 assert(node != clear_flag(iter));
11519af6 1009 assert(!is_removed(iter_prev));
c90201ac 1010 assert(!is_removed(iter));
f000907d 1011 assert(iter_prev != node);
1ee8f000 1012 if (!bucket_flag)
04db56f8 1013 node->next = clear_flag(iter);
f9c80341 1014 else
1ee8f000
LJ
1015 node->next = flag_bucket(clear_flag(iter));
1016 if (is_bucket(iter))
1017 new_node = flag_bucket(node);
f5596c94
MD
1018 else
1019 new_node = node;
04db56f8 1020 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1021 new_node) != iter) {
273399de 1022 continue; /* retry */
48ed1c18 1023 } else {
83beee94 1024 return_node = node;
960c9e4f 1025 goto end;
48ed1c18
MD
1026 }
1027
9dba85be
MD
1028 gc_node:
1029 assert(!is_removed(iter));
1ee8f000
LJ
1030 if (is_bucket(iter))
1031 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1032 else
1033 new_next = clear_flag(next);
04db56f8 1034 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1035 /* retry */
464a1ec9 1036 }
9357c415 1037end:
83beee94
MD
1038 if (unique_ret) {
1039 unique_ret->node = return_node;
1040 /* unique_ret->next left unset, never used. */
1041 }
abc490a1 1042}
464a1ec9 1043
abc490a1 1044static
860d07e8 1045int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
b65ec430 1046 struct cds_lfht_node *node)
abc490a1 1047{
db00ccc3 1048 struct cds_lfht_node *bucket, *next;
5e28c532 1049
9357c415 1050 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1051 return -ENOENT;
9357c415 1052
7ec59d3b 1053 /* logically delete the node */
1ee8f000 1054 assert(!is_bucket(node));
c90201ac 1055 assert(!is_removed(node));
db00ccc3 1056 assert(!is_removal_owner(node));
48ed1c18 1057
db00ccc3
MD
1058 /*
1059 * We are first checking if the node had previously been
1060 * logically removed (this check is not atomic with setting the
1061 * logical removal flag). Return -ENOENT if the node had
1062 * previously been removed.
1063 */
1064 next = rcu_dereference(node->next);
1065 if (caa_unlikely(is_removed(next)))
1066 return -ENOENT;
b65ec430 1067 assert(!is_bucket(next));
db00ccc3
MD
1068 /*
1069 * We set the REMOVED_FLAG unconditionally. Note that there may
1070 * be more than one concurrent thread setting this flag.
1071 * Knowing which wins the race will be known after the garbage
1072 * collection phase, stay tuned!
1073 */
1074 uatomic_or(&node->next, REMOVED_FLAG);
7ec59d3b 1075 /* We performed the (logical) deletion. */
7ec59d3b
MD
1076
1077 /*
1078 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1079 * the node, and remove it (along with any other logically removed node)
1080 * if found.
11519af6 1081 */
04db56f8
LJ
1082 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1083 _cds_lfht_gc_bucket(bucket, node);
743f9143 1084
04db56f8 1085 assert(is_removed(rcu_dereference(node->next)));
db00ccc3
MD
1086 /*
1087 * Last phase: atomically exchange node->next with a version
1088 * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
1089 * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
1090 * the node and win the removal race.
1091 * It is interesting to note that all "add" paths are forbidden
1092 * to change the next pointer starting from the point where the
1093 * REMOVED_FLAG is set, so here using a read, followed by a
1094 * xchg() suffice to guarantee that the xchg() will ever only
1095 * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
1096 * was already set).
1097 */
1098 if (!is_removal_owner(uatomic_xchg(&node->next,
1099 flag_removal_owner(node->next))))
1100 return 0;
1101 else
1102 return -ENOENT;
abc490a1 1103}
2ed95849 1104
b7d619b0
MD
1105static
1106void *partition_resize_thread(void *arg)
1107{
1108 struct partition_resize_work *work = arg;
1109
7b17c13e 1110 work->ht->flavor->register_thread();
b7d619b0 1111 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1112 work->ht->flavor->unregister_thread();
b7d619b0
MD
1113 return NULL;
1114}
1115
1116static
1117void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1118 unsigned long len,
1119 void (*fct)(struct cds_lfht *ht, unsigned long i,
1120 unsigned long start, unsigned long len))
1121{
1122 unsigned long partition_len;
1123 struct partition_resize_work *work;
6083a889
MD
1124 int thread, ret;
1125 unsigned long nr_threads;
b7d619b0 1126
6083a889
MD
1127 /*
1128 * Note: nr_cpus_mask + 1 is always power of 2.
1129 * We spawn just the number of threads we need to satisfy the minimum
1130 * partition size, up to the number of CPUs in the system.
1131 */
91452a6a
MD
1132 if (nr_cpus_mask > 0) {
1133 nr_threads = min(nr_cpus_mask + 1,
1134 len >> MIN_PARTITION_PER_THREAD_ORDER);
1135 } else {
1136 nr_threads = 1;
1137 }
5bc6b66f 1138 partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
6083a889 1139 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1140 assert(work);
6083a889
MD
1141 for (thread = 0; thread < nr_threads; thread++) {
1142 work[thread].ht = ht;
1143 work[thread].i = i;
1144 work[thread].len = partition_len;
1145 work[thread].start = thread * partition_len;
1146 work[thread].fct = fct;
1af6e26e 1147 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1148 partition_resize_thread, &work[thread]);
b7d619b0
MD
1149 assert(!ret);
1150 }
6083a889 1151 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1152 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1153 assert(!ret);
1154 }
1155 free(work);
b7d619b0
MD
1156}
1157
e8de508e
MD
1158/*
1159 * Holding RCU read lock to protect _cds_lfht_add against memory
1160 * reclaim that could be performed by other call_rcu worker threads (ABA
1161 * problem).
9ee0fc9a 1162 *
b7d619b0 1163 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1164 * many worker threads, based on the number of CPUs available in the system.
1165 * This should therefore take care of not having the expand lagging behind too
1166 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1167 * schedule bucket node population fairly with insertions.
e8de508e 1168 */
4105056a 1169static
b7d619b0
MD
1170void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1171 unsigned long start, unsigned long len)
4105056a 1172{
9d72a73f 1173 unsigned long j, size = 1UL << (i - 1);
4105056a 1174
d0d8f9aa 1175 assert(i > MIN_TABLE_ORDER);
7b17c13e 1176 ht->flavor->read_lock();
9d72a73f
LJ
1177 for (j = size + start; j < size + start + len; j++) {
1178 struct cds_lfht_node *new_node = bucket_at(ht, j);
1179
1180 assert(j >= size && j < (size << 1));
1181 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1182 i, j, j);
1183 new_node->reverse_hash = bit_reverse_ulong(j);
91a75cc5 1184 _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
4105056a 1185 }
7b17c13e 1186 ht->flavor->read_unlock();
b7d619b0
MD
1187}
1188
1189static
1190void init_table_populate(struct cds_lfht *ht, unsigned long i,
1191 unsigned long len)
1192{
1193 assert(nr_cpus_mask != -1);
6083a889 1194 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
7b17c13e 1195 ht->flavor->thread_online();
b7d619b0 1196 init_table_populate_partition(ht, i, 0, len);
7b17c13e 1197 ht->flavor->thread_offline();
b7d619b0
MD
1198 return;
1199 }
1200 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1201}
1202
abc490a1 1203static
4105056a 1204void init_table(struct cds_lfht *ht,
93d46c39 1205 unsigned long first_order, unsigned long last_order)
24365af7 1206{
93d46c39 1207 unsigned long i;
24365af7 1208
93d46c39
LJ
1209 dbg_printf("init table: first_order %lu last_order %lu\n",
1210 first_order, last_order);
d0d8f9aa 1211 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1212 for (i = first_order; i <= last_order; i++) {
4105056a 1213 unsigned long len;
24365af7 1214
4f6e90b7 1215 len = 1UL << (i - 1);
f0c29ed7 1216 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1217
1218 /* Stop expand if the resize target changes under us */
7b3893e4 1219 if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i))
4d676753
MD
1220 break;
1221
48f1b16d 1222 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1223
4105056a 1224 /*
1ee8f000
LJ
1225 * Set all bucket nodes reverse hash values for a level and
1226 * link all bucket nodes into the table.
4105056a 1227 */
dc1da8f6 1228 init_table_populate(ht, i, len);
4105056a 1229
f9c80341
MD
1230 /*
1231 * Update table size.
1232 */
1233 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1234 CMM_STORE_SHARED(ht->size, 1UL << i);
f9c80341 1235
4f6e90b7 1236 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1237 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1238 break;
1239 }
1240}
1241
e8de508e
MD
1242/*
1243 * Holding RCU read lock to protect _cds_lfht_remove against memory
1244 * reclaim that could be performed by other call_rcu worker threads (ABA
1245 * problem).
1246 * For a single level, we logically remove and garbage collect each node.
1247 *
1248 * As a design choice, we perform logical removal and garbage collection on a
1249 * node-per-node basis to simplify this algorithm. We also assume keeping good
1250 * cache locality of the operation would overweight possible performance gain
1251 * that could be achieved by batching garbage collection for multiple levels.
1252 * However, this would have to be justified by benchmarks.
1253 *
1254 * Concurrent removal and add operations are helping us perform garbage
1255 * collection of logically removed nodes. We guarantee that all logically
1256 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1257 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e 1258 *
1f67ba50
MD
1259 * Logical removal and garbage collection can therefore be done in batch
1260 * or on a node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1261 *
b7d619b0
MD
1262 * When we reach a certain length, we can split this removal over many worker
1263 * threads, based on the number of CPUs available in the system. This should
1264 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1265 * updater threads actively inserting into the hash table.
e8de508e 1266 */
4105056a 1267static
b7d619b0
MD
1268void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1269 unsigned long start, unsigned long len)
4105056a 1270{
9d72a73f 1271 unsigned long j, size = 1UL << (i - 1);
4105056a 1272
d0d8f9aa 1273 assert(i > MIN_TABLE_ORDER);
7b17c13e 1274 ht->flavor->read_lock();
9d72a73f 1275 for (j = size + start; j < size + start + len; j++) {
2e2ce1e9
LJ
1276 struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
1277 struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
9d72a73f
LJ
1278
1279 assert(j >= size && j < (size << 1));
1280 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1281 i, j, j);
2e2ce1e9
LJ
1282 /* Set the REMOVED_FLAG to freeze the ->next for gc */
1283 uatomic_or(&fini_bucket->next, REMOVED_FLAG);
1284 _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
abc490a1 1285 }
7b17c13e 1286 ht->flavor->read_unlock();
b7d619b0
MD
1287}
1288
1289static
1290void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1291{
1292
1293 assert(nr_cpus_mask != -1);
6083a889 1294 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
7b17c13e 1295 ht->flavor->thread_online();
b7d619b0 1296 remove_table_partition(ht, i, 0, len);
7b17c13e 1297 ht->flavor->thread_offline();
b7d619b0
MD
1298 return;
1299 }
1300 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1301}
1302
61adb337
MD
1303/*
1304 * fini_table() is never called for first_order == 0, which is why
1305 * free_by_rcu_order == 0 can be used as criterion to know if free must
1306 * be called.
1307 */
1475579c 1308static
4105056a 1309void fini_table(struct cds_lfht *ht,
93d46c39 1310 unsigned long first_order, unsigned long last_order)
1475579c 1311{
93d46c39 1312 long i;
48f1b16d 1313 unsigned long free_by_rcu_order = 0;
1475579c 1314
93d46c39
LJ
1315 dbg_printf("fini table: first_order %lu last_order %lu\n",
1316 first_order, last_order);
d0d8f9aa 1317 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1318 for (i = last_order; i >= first_order; i--) {
4105056a 1319 unsigned long len;
1475579c 1320
4f6e90b7 1321 len = 1UL << (i - 1);
1475579c 1322 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1323
4d676753 1324 /* Stop shrink if the resize target changes under us */
7b3893e4 1325 if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1)))
4d676753
MD
1326 break;
1327
1328 cmm_smp_wmb(); /* populate data before RCU size */
7b3893e4 1329 CMM_STORE_SHARED(ht->size, 1UL << (i - 1));
4d676753
MD
1330
1331 /*
1332 * We need to wait for all add operations to reach Q.S. (and
1333 * thus use the new table for lookups) before we can start
1ee8f000 1334 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1335 * return a logically removed node as insert position.
1336 */
7b17c13e 1337 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1338 if (free_by_rcu_order)
1339 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1340
21263e21 1341 /*
1ee8f000
LJ
1342 * Set "removed" flag in bucket nodes about to be removed.
1343 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1344 * Concurrent add/remove operation are helping us doing
1345 * the gc.
21263e21 1346 */
4105056a
MD
1347 remove_table(ht, i, len);
1348
48f1b16d 1349 free_by_rcu_order = i;
4105056a
MD
1350
1351 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1352 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1353 break;
1354 }
0d14ceb2 1355
48f1b16d 1356 if (free_by_rcu_order) {
7b17c13e 1357 ht->flavor->update_synchronize_rcu();
48f1b16d 1358 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1359 }
1475579c
MD
1360}
1361
ff0d69de 1362static
1ee8f000 1363void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1364{
04db56f8 1365 struct cds_lfht_node *prev, *node;
9d72a73f 1366 unsigned long order, len, i;
ff0d69de 1367
48f1b16d 1368 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1369
9d72a73f
LJ
1370 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1371 node = bucket_at(ht, 0);
1372 node->next = flag_bucket(get_end());
1373 node->reverse_hash = 0;
ff0d69de 1374
5bc6b66f 1375 for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
ff0d69de 1376 len = 1UL << (order - 1);
48f1b16d 1377 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1378
9d72a73f
LJ
1379 for (i = 0; i < len; i++) {
1380 /*
1381 * Now, we are trying to init the node with the
1382 * hash=(len+i) (which is also a bucket with the
1383 * index=(len+i)) and insert it into the hash table,
1384 * so this node has to be inserted after the bucket
1385 * with the index=(len+i)&(len-1)=i. And because there
1386 * is no other non-bucket node nor bucket node with
1387 * larger index/hash inserted, so the bucket node
1388 * being inserted should be inserted directly linked
1389 * after the bucket node with index=i.
1390 */
1391 prev = bucket_at(ht, i);
1392 node = bucket_at(ht, len + i);
ff0d69de 1393
1ee8f000 1394 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1395 order, len + i, len + i);
1396 node->reverse_hash = bit_reverse_ulong(len + i);
1397
1398 /* insert after prev */
1399 assert(is_bucket(prev->next));
ff0d69de 1400 node->next = prev->next;
1ee8f000 1401 prev->next = flag_bucket(node);
ff0d69de
LJ
1402 }
1403 }
1404}
1405
0422d92c 1406struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1407 unsigned long min_nr_alloc_buckets,
747d725c 1408 unsigned long max_nr_buckets,
b8af5011 1409 int flags,
0b6aa001 1410 const struct cds_lfht_mm_type *mm,
7b17c13e 1411 const struct rcu_flavor_struct *flavor,
b7d619b0 1412 pthread_attr_t *attr)
abc490a1 1413{
14044b37 1414 struct cds_lfht *ht;
24365af7 1415 unsigned long order;
abc490a1 1416
0722081a
LJ
1417 /* min_nr_alloc_buckets must be power of two */
1418 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1419 return NULL;
747d725c 1420
8129be4e 1421 /* init_size must be power of two */
5488222b 1422 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1423 return NULL;
747d725c 1424
c1888f3a
MD
1425 /*
1426 * Memory management plugin default.
1427 */
1428 if (!mm) {
5a2141a7
MD
1429 if (CAA_BITS_PER_LONG > 32
1430 && max_nr_buckets
c1888f3a
MD
1431 && max_nr_buckets <= (1ULL << 32)) {
1432 /*
1433 * For 64-bit architectures, with max number of
1434 * buckets small enough not to use the entire
1435 * 64-bit memory mapping space (and allowing a
1436 * fair number of hash table instances), use the
1437 * mmap allocator, which is faster than the
1438 * order allocator.
1439 */
1440 mm = &cds_lfht_mm_mmap;
1441 } else {
1442 /*
1443 * The fallback is to use the order allocator.
1444 */
1445 mm = &cds_lfht_mm_order;
1446 }
1447 }
1448
0b6aa001
LJ
1449 /* max_nr_buckets == 0 for order based mm means infinite */
1450 if (mm == &cds_lfht_mm_order && !max_nr_buckets)
747d725c
LJ
1451 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1452
1453 /* max_nr_buckets must be power of two */
1454 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1455 return NULL;
1456
0722081a 1457 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1458 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1459 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1460 init_size = min(init_size, max_nr_buckets);
0b6aa001
LJ
1461
1462 ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets);
b7d619b0 1463 assert(ht);
0b6aa001
LJ
1464 assert(ht->mm == mm);
1465 assert(ht->bucket_at == mm->bucket_at);
1466
b5d6b20f 1467 ht->flags = flags;
7b17c13e 1468 ht->flavor = flavor;
b7d619b0 1469 ht->resize_attr = attr;
5afadd12 1470 alloc_split_items_count(ht);
abc490a1
MD
1471 /* this mutex should not nest in read-side C.S. */
1472 pthread_mutex_init(&ht->resize_mutex, NULL);
5bc6b66f 1473 order = cds_lfht_get_count_order_ulong(init_size);
7b3893e4 1474 ht->resize_target = 1UL << order;
1ee8f000 1475 cds_lfht_create_bucket(ht, 1UL << order);
7b3893e4 1476 ht->size = 1UL << order;
abc490a1
MD
1477 return ht;
1478}
1479
6f554439 1480void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1481 cds_lfht_match_fct match, const void *key,
6f554439 1482 struct cds_lfht_iter *iter)
2ed95849 1483{
04db56f8 1484 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1485 unsigned long reverse_hash, size;
2ed95849 1486
abc490a1 1487 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1488
7b3893e4 1489 size = rcu_dereference(ht->size);
04db56f8 1490 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1491 /* We can always skip the bucket node initially */
04db56f8 1492 node = rcu_dereference(bucket->next);
bb7b2f26 1493 node = clear_flag(node);
2ed95849 1494 for (;;) {
8ed51e04 1495 if (caa_unlikely(is_end(node))) {
96ad1112 1496 node = next = NULL;
abc490a1 1497 break;
bb7b2f26 1498 }
04db56f8 1499 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1500 node = next = NULL;
abc490a1 1501 break;
2ed95849 1502 }
04db56f8 1503 next = rcu_dereference(node->next);
7f52427b 1504 assert(node == clear_flag(node));
8ed51e04 1505 if (caa_likely(!is_removed(next))
1ee8f000 1506 && !is_bucket(next)
04db56f8 1507 && node->reverse_hash == reverse_hash
0422d92c 1508 && caa_likely(match(node, key))) {
273399de 1509 break;
2ed95849 1510 }
1b81fe1a 1511 node = clear_flag(next);
2ed95849 1512 }
1ee8f000 1513 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1514 iter->node = node;
1515 iter->next = next;
abc490a1 1516}
e0ba718a 1517
0422d92c 1518void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1519 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1520{
adc0de68 1521 struct cds_lfht_node *node, *next;
a481e5ff 1522 unsigned long reverse_hash;
a481e5ff 1523
adc0de68 1524 node = iter->node;
04db56f8 1525 reverse_hash = node->reverse_hash;
adc0de68 1526 next = iter->next;
a481e5ff
MD
1527 node = clear_flag(next);
1528
1529 for (;;) {
8ed51e04 1530 if (caa_unlikely(is_end(node))) {
96ad1112 1531 node = next = NULL;
a481e5ff 1532 break;
bb7b2f26 1533 }
04db56f8 1534 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1535 node = next = NULL;
a481e5ff
MD
1536 break;
1537 }
04db56f8 1538 next = rcu_dereference(node->next);
8ed51e04 1539 if (caa_likely(!is_removed(next))
1ee8f000 1540 && !is_bucket(next)
04db56f8 1541 && caa_likely(match(node, key))) {
a481e5ff
MD
1542 break;
1543 }
1544 node = clear_flag(next);
1545 }
1ee8f000 1546 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1547 iter->node = node;
1548 iter->next = next;
a481e5ff
MD
1549}
1550
4e9b9fbf
MD
1551void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1552{
1553 struct cds_lfht_node *node, *next;
1554
853395e1 1555 node = clear_flag(iter->next);
4e9b9fbf 1556 for (;;) {
8ed51e04 1557 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1558 node = next = NULL;
1559 break;
1560 }
04db56f8 1561 next = rcu_dereference(node->next);
8ed51e04 1562 if (caa_likely(!is_removed(next))
1ee8f000 1563 && !is_bucket(next)) {
4e9b9fbf
MD
1564 break;
1565 }
1566 node = clear_flag(next);
1567 }
1ee8f000 1568 assert(!node || !is_bucket(rcu_dereference(node->next)));
4e9b9fbf
MD
1569 iter->node = node;
1570 iter->next = next;
1571}
1572
1573void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1574{
4e9b9fbf 1575 /*
1ee8f000 1576 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1577 * first node of the linked list.
1578 */
9d72a73f 1579 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1580 cds_lfht_next(ht, iter);
1581}
1582
0422d92c
MD
1583void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1584 struct cds_lfht_node *node)
abc490a1 1585{
0422d92c 1586 unsigned long size;
ab7d5fc6 1587
709bacf9 1588 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1589 size = rcu_dereference(ht->size);
91a75cc5 1590 _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
14360f1c 1591 ht_count_add(ht, size, hash);
3eca1b8c
MD
1592}
1593
14044b37 1594struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1595 unsigned long hash,
0422d92c 1596 cds_lfht_match_fct match,
996ff57c 1597 const void *key,
48ed1c18 1598 struct cds_lfht_node *node)
3eca1b8c 1599{
0422d92c 1600 unsigned long size;
83beee94 1601 struct cds_lfht_iter iter;
3eca1b8c 1602
709bacf9 1603 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1604 size = rcu_dereference(ht->size);
91a75cc5 1605 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1606 if (iter.node == node)
14360f1c 1607 ht_count_add(ht, size, hash);
83beee94 1608 return iter.node;
2ed95849
MD
1609}
1610
9357c415 1611struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1612 unsigned long hash,
0422d92c 1613 cds_lfht_match_fct match,
996ff57c 1614 const void *key,
48ed1c18
MD
1615 struct cds_lfht_node *node)
1616{
0422d92c 1617 unsigned long size;
83beee94 1618 struct cds_lfht_iter iter;
48ed1c18 1619
709bacf9 1620 node->reverse_hash = bit_reverse_ulong(hash);
7b3893e4 1621 size = rcu_dereference(ht->size);
83beee94 1622 for (;;) {
91a75cc5 1623 _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
83beee94 1624 if (iter.node == node) {
14360f1c 1625 ht_count_add(ht, size, hash);
83beee94
MD
1626 return NULL;
1627 }
1628
1629 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1630 return iter.node;
1631 }
48ed1c18
MD
1632}
1633
2e79c445
MD
1634int cds_lfht_replace(struct cds_lfht *ht,
1635 struct cds_lfht_iter *old_iter,
1636 unsigned long hash,
1637 cds_lfht_match_fct match,
1638 const void *key,
9357c415
MD
1639 struct cds_lfht_node *new_node)
1640{
1641 unsigned long size;
1642
709bacf9 1643 new_node->reverse_hash = bit_reverse_ulong(hash);
2e79c445
MD
1644 if (!old_iter->node)
1645 return -ENOENT;
1646 if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
1647 return -EINVAL;
1648 if (caa_unlikely(!match(old_iter->node, key)))
1649 return -EINVAL;
7b3893e4 1650 size = rcu_dereference(ht->size);
9357c415
MD
1651 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1652 new_node);
1653}
1654
bc8c3c74 1655int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1656{
14360f1c 1657 unsigned long size, hash;
df44348d 1658 int ret;
abc490a1 1659
7b3893e4 1660 size = rcu_dereference(ht->size);
bc8c3c74 1661 ret = _cds_lfht_del(ht, size, node);
14360f1c 1662 if (!ret) {
bc8c3c74 1663 hash = bit_reverse_ulong(node->reverse_hash);
14360f1c
LJ
1664 ht_count_del(ht, size, hash);
1665 }
df44348d 1666 return ret;
2ed95849 1667}
ab7d5fc6 1668
df55172a
MD
1669int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
1670{
1671 return is_removed(rcu_dereference(node->next));
1672}
1673
abc490a1 1674static
1ee8f000 1675int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1676{
14044b37 1677 struct cds_lfht_node *node;
4105056a 1678 unsigned long order, i, size;
674f7a69 1679
abc490a1 1680 /* Check that the table is empty */
9d72a73f 1681 node = bucket_at(ht, 0);
abc490a1 1682 do {
04db56f8 1683 node = clear_flag(node)->next;
1ee8f000 1684 if (!is_bucket(node))
abc490a1 1685 return -EPERM;
273399de 1686 assert(!is_removed(node));
bb7b2f26 1687 } while (!is_end(node));
4105056a
MD
1688 /*
1689 * size accessed without rcu_dereference because hash table is
1690 * being destroyed.
1691 */
7b3893e4 1692 size = ht->size;
1f67ba50 1693 /* Internal sanity check: all nodes left should be buckets */
48f1b16d
LJ
1694 for (i = 0; i < size; i++) {
1695 node = bucket_at(ht, i);
1696 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1697 i, i, bit_reverse_ulong(node->reverse_hash));
1698 assert(is_bucket(node->next));
1699 }
24365af7 1700
5bc6b66f 1701 for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
48f1b16d 1702 cds_lfht_free_bucket_table(ht, order);
5488222b 1703
abc490a1 1704 return 0;
674f7a69
MD
1705}
1706
1707/*
1708 * Should only be called when no more concurrent readers nor writers can
1709 * possibly access the table.
1710 */
b7d619b0 1711int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1712{
5e28c532
MD
1713 int ret;
1714
848d4088 1715 /* Wait for in-flight resize operations to complete */
24953e08
MD
1716 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1717 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1718 while (uatomic_read(&ht->in_progress_resize))
1719 poll(NULL, 0, 100); /* wait for 100ms */
1ee8f000 1720 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1721 if (ret)
1722 return ret;
5afadd12 1723 free_split_items_count(ht);
b7d619b0
MD
1724 if (attr)
1725 *attr = ht->resize_attr;
98808fb1 1726 poison_free(ht);
5e28c532 1727 return ret;
674f7a69
MD
1728}
1729
14044b37 1730void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1731 long *approx_before,
273399de 1732 unsigned long *count,
d933dd0e 1733 long *approx_after)
273399de 1734{
14044b37 1735 struct cds_lfht_node *node, *next;
caf3653d 1736 unsigned long nr_bucket = 0, nr_removed = 0;
273399de 1737
7ed7682f 1738 *approx_before = 0;
5afadd12 1739 if (ht->split_count) {
973e5e1b
MD
1740 int i;
1741
4c42f1b8
LJ
1742 for (i = 0; i < split_count_mask + 1; i++) {
1743 *approx_before += uatomic_read(&ht->split_count[i].add);
1744 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1745 }
1746 }
1747
273399de 1748 *count = 0;
273399de 1749
1ee8f000 1750 /* Count non-bucket nodes in the table */
9d72a73f 1751 node = bucket_at(ht, 0);
273399de 1752 do {
04db56f8 1753 next = rcu_dereference(node->next);
b198f0fd 1754 if (is_removed(next)) {
1ee8f000 1755 if (!is_bucket(next))
caf3653d 1756 (nr_removed)++;
973e5e1b 1757 else
1ee8f000
LJ
1758 (nr_bucket)++;
1759 } else if (!is_bucket(next))
273399de 1760 (*count)++;
24365af7 1761 else
1ee8f000 1762 (nr_bucket)++;
273399de 1763 node = clear_flag(next);
bb7b2f26 1764 } while (!is_end(node));
caf3653d 1765 dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
1ee8f000 1766 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1767 *approx_after = 0;
5afadd12 1768 if (ht->split_count) {
973e5e1b
MD
1769 int i;
1770
4c42f1b8
LJ
1771 for (i = 0; i < split_count_mask + 1; i++) {
1772 *approx_after += uatomic_read(&ht->split_count[i].add);
1773 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1774 }
1775 }
273399de
MD
1776}
1777
1475579c 1778/* called with resize mutex held */
abc490a1 1779static
4105056a 1780void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1781 unsigned long old_size, unsigned long new_size)
abc490a1 1782{
1475579c 1783 unsigned long old_order, new_order;
1475579c 1784
5bc6b66f
MD
1785 old_order = cds_lfht_get_count_order_ulong(old_size);
1786 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1787 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1788 old_size, old_order, new_size, new_order);
1475579c 1789 assert(new_size > old_size);
93d46c39 1790 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1791}
1792
1793/* called with resize mutex held */
1794static
4105056a 1795void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1796 unsigned long old_size, unsigned long new_size)
464a1ec9 1797{
1475579c 1798 unsigned long old_order, new_order;
464a1ec9 1799
d0d8f9aa 1800 new_size = max(new_size, MIN_TABLE_SIZE);
5bc6b66f
MD
1801 old_order = cds_lfht_get_count_order_ulong(old_size);
1802 new_order = cds_lfht_get_count_order_ulong(new_size);
1a401918
LJ
1803 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1804 old_size, old_order, new_size, new_order);
1475579c 1805 assert(new_size < old_size);
1475579c 1806
1ee8f000 1807 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1808 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1809}
1810
1475579c
MD
1811
1812/* called with resize mutex held */
1813static
1814void _do_cds_lfht_resize(struct cds_lfht *ht)
1815{
1816 unsigned long new_size, old_size;
4105056a
MD
1817
1818 /*
1819 * Resize table, re-do if the target size has changed under us.
1820 */
1821 do {
d2be3620
MD
1822 assert(uatomic_read(&ht->in_progress_resize));
1823 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1824 break;
7b3893e4
LJ
1825 ht->resize_initiated = 1;
1826 old_size = ht->size;
1827 new_size = CMM_LOAD_SHARED(ht->resize_target);
4105056a
MD
1828 if (old_size < new_size)
1829 _do_cds_lfht_grow(ht, old_size, new_size);
1830 else if (old_size > new_size)
1831 _do_cds_lfht_shrink(ht, old_size, new_size);
7b3893e4 1832 ht->resize_initiated = 0;
4105056a
MD
1833 /* write resize_initiated before read resize_target */
1834 cmm_smp_mb();
7b3893e4 1835 } while (ht->size != CMM_LOAD_SHARED(ht->resize_target));
1475579c
MD
1836}
1837
abc490a1 1838static
ab65b890 1839unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1840{
7b3893e4 1841 return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size);
464a1ec9
MD
1842}
1843
1475579c 1844static
4105056a 1845void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1846 unsigned long count)
1475579c 1847{
d0d8f9aa 1848 count = max(count, MIN_TABLE_SIZE);
747d725c 1849 count = min(count, ht->max_nr_buckets);
7b3893e4 1850 uatomic_set(&ht->resize_target, count);
1475579c
MD
1851}
1852
1853void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1854{
4105056a 1855 resize_target_update_count(ht, new_size);
7b3893e4 1856 CMM_STORE_SHARED(ht->resize_initiated, 1);
7b17c13e 1857 ht->flavor->thread_offline();
1475579c
MD
1858 pthread_mutex_lock(&ht->resize_mutex);
1859 _do_cds_lfht_resize(ht);
1860 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1861 ht->flavor->thread_online();
abc490a1 1862}
464a1ec9 1863
abc490a1
MD
1864static
1865void do_resize_cb(struct rcu_head *head)
1866{
1867 struct rcu_resize_work *work =
1868 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1869 struct cds_lfht *ht = work->ht;
abc490a1 1870
7b17c13e 1871 ht->flavor->thread_offline();
abc490a1 1872 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1873 _do_cds_lfht_resize(ht);
abc490a1 1874 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1875 ht->flavor->thread_online();
98808fb1 1876 poison_free(work);
848d4088
MD
1877 cmm_smp_mb(); /* finish resize before decrement */
1878 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1879}
1880
abc490a1 1881static
f1f119ee 1882void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1883{
abc490a1
MD
1884 struct rcu_resize_work *work;
1885
4105056a
MD
1886 /* Store resize_target before read resize_initiated */
1887 cmm_smp_mb();
7b3893e4 1888 if (!CMM_LOAD_SHARED(ht->resize_initiated)) {
848d4088 1889 uatomic_inc(&ht->in_progress_resize);
59290e9d 1890 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1891 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1892 uatomic_dec(&ht->in_progress_resize);
59290e9d 1893 return;
ed35e6d8 1894 }
f9830efd 1895 work = malloc(sizeof(*work));
741f378e
MD
1896 if (work == NULL) {
1897 dbg_printf("error allocating resize work, bailing out\n");
1898 uatomic_dec(&ht->in_progress_resize);
1899 return;
1900 }
f9830efd 1901 work->ht = ht;
7b17c13e 1902 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
7b3893e4 1903 CMM_STORE_SHARED(ht->resize_initiated, 1);
f9830efd 1904 }
ab7d5fc6 1905}
3171717f 1906
f1f119ee
LJ
1907static
1908void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1909{
1910 unsigned long target_size = size << growth;
1911
747d725c 1912 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
1913 if (resize_target_grow(ht, target_size) >= target_size)
1914 return;
1915
1916 __cds_lfht_resize_lazy_launch(ht);
1917}
1918
89bb121d
LJ
1919/*
1920 * We favor grow operations over shrink. A shrink operation never occurs
1921 * if a grow operation is queued for lazy execution. A grow operation
1922 * cancels any pending shrink lazy execution.
1923 */
3171717f 1924static
4105056a 1925void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1926 unsigned long count)
1927{
b8af5011
MD
1928 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1929 return;
d0d8f9aa 1930 count = max(count, MIN_TABLE_SIZE);
747d725c 1931 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
1932 if (count == size)
1933 return; /* Already the right size, no resize needed */
1934 if (count > size) { /* lazy grow */
1935 if (resize_target_grow(ht, count) >= count)
1936 return;
1937 } else { /* lazy shrink */
1938 for (;;) {
1939 unsigned long s;
1940
7b3893e4 1941 s = uatomic_cmpxchg(&ht->resize_target, size, count);
89bb121d
LJ
1942 if (s == size)
1943 break; /* no resize needed */
1944 if (s > size)
1945 return; /* growing is/(was just) in progress */
1946 if (s <= count)
1947 return; /* some other thread do shrink */
1948 size = s;
1949 }
1950 }
f1f119ee 1951 __cds_lfht_resize_lazy_launch(ht);
3171717f 1952}
This page took 0.142401 seconds and 4 git commands to generate.