use rcu_flavor for rculfhash
[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
37 * table lookups and use the returned objects safely by delaying
38 * memory reclaim of a grace period.
39 * - Add and remove operations are lock-free, and do not need to
40 * allocate memory. They need to be executed within RCU read-side
41 * critical section to ensure the objects they read are valid and to
42 * deal with the cmpxchg ABA problem.
43 * - add and add_unique operations are supported. add_unique checks if
44 * the node key already exists in the hash table. It ensures no key
45 * duplicata exists.
46 * - The resize operation executes concurrently with add/remove/lookup.
47 * - Hash table nodes are contained within a split-ordered list. This
48 * list is ordered by incrementing reversed-bits-hash value.
1ee8f000 49 * - An index of bucket nodes is kept. These bucket nodes are the hash
e753ff5a
MD
50 * table "buckets", and they are also chained together in the
51 * split-ordered list, which allows recursive expansion.
1475579c
MD
52 * - The resize operation for small tables only allows expanding the hash table.
53 * It is triggered automatically by detecting long chains in the add
54 * operation.
55 * - The resize operation for larger tables (and available through an
56 * API) allows both expanding and shrinking the hash table.
4c42f1b8 57 * - Split-counters are used to keep track of the number of
1475579c 58 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
59 * - Resize operation initiated by long chain detection is executed by a
60 * call_rcu thread, which keeps lock-freedom of add and remove.
61 * - Resize operations are protected by a mutex.
62 * - The removal operation is split in two parts: first, a "removed"
63 * flag is set in the next pointer within the node to remove. Then,
64 * a "garbage collection" is performed in the bucket containing the
65 * removed node (from the start of the bucket up to the removed node).
66 * All encountered nodes with "removed" flag set in their next
67 * pointers are removed from the linked-list. If the cmpxchg used for
68 * removal fails (due to concurrent garbage-collection or concurrent
69 * add), we retry from the beginning of the bucket. This ensures that
70 * the node with "removed" flag set is removed from the hash table
71 * (not visible to lookups anymore) before the RCU read-side critical
72 * section held across removal ends. Furthermore, this ensures that
73 * the node with "removed" flag set is removed from the linked-list
74 * before its memory is reclaimed. Only the thread which removal
75 * successfully set the "removed" flag (with a cmpxchg) into a node's
76 * next pointer is considered to have succeeded its removal (and thus
77 * owns the node to reclaim). Because we garbage-collect starting from
1ee8f000 78 * an invariant node (the start-of-bucket bucket node) up to the
e753ff5a
MD
79 * "removed" node (or find a reverse-hash that is higher), we are sure
80 * that a successful traversal of the chain leads to a chain that is
81 * present in the linked-list (the start node is never removed) and
82 * that is does not contain the "removed" node anymore, even if
83 * concurrent delete/add operations are changing the structure of the
84 * list concurrently.
29e669f6
MD
85 * - The add operation performs gargage collection of buckets if it
86 * encounters nodes with removed flag set in the bucket where it wants
87 * to add its new node. This ensures lock-freedom of add operation by
88 * helping the remover unlink nodes from the list rather than to wait
89 * for it do to so.
e753ff5a
MD
90 * - A RCU "order table" indexed by log2(hash index) is copied and
91 * expanded by the resize operation. This order table allows finding
1ee8f000
LJ
92 * the "bucket node" tables.
93 * - There is one bucket node table per hash index order. The size of
94 * each bucket node table is half the number of hashes contained in
93d46c39 95 * this order (except for order 0).
1ee8f000
LJ
96 * - synchronzie_rcu is used to garbage-collect the old bucket node table.
97 * - The per-order bucket node tables contain a compact version of the
e753ff5a
MD
98 * hash table nodes. These tables are invariant after they are
99 * populated into the hash table.
93d46c39 100 *
1ee8f000 101 * Bucket node tables:
93d46c39 102 *
1ee8f000
LJ
103 * hash table hash table the last all bucket node tables
104 * order size bucket node 0 1 2 3 4 5 6(index)
93d46c39
LJ
105 * table size
106 * 0 1 1 1
107 * 1 2 1 1 1
108 * 2 4 2 1 1 2
109 * 3 8 4 1 1 2 4
110 * 4 16 8 1 1 2 4 8
111 * 5 32 16 1 1 2 4 8 16
112 * 6 64 32 1 1 2 4 8 16 32
113 *
1ee8f000 114 * When growing/shrinking, we only focus on the last bucket node table
93d46c39
LJ
115 * which size is (!order ? 1 : (1 << (order -1))).
116 *
117 * Example for growing/shrinking:
1ee8f000
LJ
118 * grow hash table from order 5 to 6: init the index=6 bucket node table
119 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
93d46c39 120 *
1475579c
MD
121 * A bit of ascii art explanation:
122 *
123 * Order index is the off-by-one compare to the actual power of 2 because
124 * we use index 0 to deal with the 0 special-case.
125 *
126 * This shows the nodes for a small table ordered by reversed bits:
127 *
128 * bits reverse
129 * 0 000 000
130 * 4 100 001
131 * 2 010 010
132 * 6 110 011
133 * 1 001 100
134 * 5 101 101
135 * 3 011 110
136 * 7 111 111
137 *
138 * This shows the nodes in order of non-reversed bits, linked by
139 * reversed-bit order.
140 *
141 * order bits reverse
142 * 0 0 000 000
0adc36a8
LJ
143 * 1 | 1 001 100 <-
144 * 2 | | 2 010 010 <- |
f6fdd688 145 * | | | 3 011 110 | <- |
1475579c
MD
146 * 3 -> | | | 4 100 001 | |
147 * -> | | 5 101 101 |
148 * -> | 6 110 011
149 * -> 7 111 111
e753ff5a
MD
150 */
151
2ed95849
MD
152#define _LGPL_SOURCE
153#include <stdlib.h>
e0ba718a
MD
154#include <errno.h>
155#include <assert.h>
156#include <stdio.h>
abc490a1 157#include <stdint.h>
f000907d 158#include <string.h>
e0ba718a 159
15cfbec7 160#include "config.h"
2ed95849 161#include <urcu.h>
abc490a1 162#include <urcu-call-rcu.h>
7b17c13e 163#include <urcu-flavor.h>
a42cc659
MD
164#include <urcu/arch.h>
165#include <urcu/uatomic.h>
a42cc659 166#include <urcu/compiler.h>
abc490a1 167#include <urcu/rculfhash.h>
5e28c532 168#include <stdio.h>
464a1ec9 169#include <pthread.h>
44395fb7 170
f9830efd 171#ifdef DEBUG
f0c29ed7 172#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 173#else
e753ff5a 174#define dbg_printf(fmt, args...)
f9830efd
MD
175#endif
176
f8994aee 177/*
4c42f1b8 178 * Split-counters lazily update the global counter each 1024
f8994aee
MD
179 * addition/removal. It automatically keeps track of resize required.
180 * We use the bucket length as indicator for need to expand for small
181 * tables and machines lacking per-cpu data suppport.
182 */
183#define COUNT_COMMIT_ORDER 10
4ddbb355 184#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
185#define CHAIN_LEN_TARGET 1
186#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 187
cd95516d 188/*
76a73da8 189 * Define the minimum table size.
cd95516d 190 */
d0d8f9aa
LJ
191#define MIN_TABLE_ORDER 0
192#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 193
4105056a
MD
194#if (CAA_BITS_PER_LONG == 32)
195#define MAX_TABLE_ORDER 32
196#else
197#define MAX_TABLE_ORDER 64
198#endif
199
b7d619b0 200/*
1ee8f000 201 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 202 */
6083a889
MD
203#define MIN_PARTITION_PER_THREAD_ORDER 12
204#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 205
4105056a
MD
206#ifndef min
207#define min(a, b) ((a) < (b) ? (a) : (b))
208#endif
209
abc490a1
MD
210#ifndef max
211#define max(a, b) ((a) > (b) ? (a) : (b))
212#endif
2ed95849 213
d95bd160
MD
214/*
215 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 216 * It indicates that no node must attach to the node scheduled for
b198f0fd 217 * removal, and that node garbage collection must be performed.
1ee8f000 218 * The bucket flag does not require to be updated atomically with the
d95bd160
MD
219 * pointer, but it is added as a pointer low bit flag to save space.
220 */
d37166c6 221#define REMOVED_FLAG (1UL << 0)
1ee8f000 222#define BUCKET_FLAG (1UL << 1)
b198f0fd 223#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 224
bb7b2f26 225/* Value of the end pointer. Should not interact with flags. */
f9c80341 226#define END_VALUE NULL
bb7b2f26 227
7f52427b
MD
228/*
229 * ht_items_count: Split-counters counting the number of node addition
230 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
231 * is set at hash table creation.
232 *
233 * These are free-running counters, never reset to zero. They count the
234 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
235 * operations to update the global counter. We choose a power-of-2 value
236 * for the trigger to deal with 32 or 64-bit overflow of the counter.
237 */
df44348d 238struct ht_items_count {
860d07e8 239 unsigned long add, del;
df44348d
MD
240} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
241
7f52427b
MD
242/*
243 * rcu_table: Contains the size and desired new size if a resize
244 * operation is in progress, as well as the statically-sized array of
dcc26ef4 245 * bucket table pointers.
7f52427b 246 */
395270b6 247struct rcu_table {
4105056a 248 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 249 unsigned long resize_target;
11519af6 250 int resize_initiated;
dcc26ef4
LJ
251
252 /*
253 * Contains the per order-index-level bucket node table. The size
254 * of each bucket node table is half the number of hashes contained
255 * in this order (except for order 0). The minimum allocation size
256 * parameter allows combining the bucket node arrays of the lowermost
257 * levels to improve cache locality for small index orders.
258 */
259 struct cds_lfht_node *tbl[MAX_TABLE_ORDER];
395270b6
MD
260};
261
7f52427b
MD
262/*
263 * cds_lfht: Top-level data structure representing a lock-free hash
264 * table. Defined in the implementation file to make it be an opaque
265 * cookie to users.
266 */
14044b37 267struct cds_lfht {
4105056a 268 struct rcu_table t;
0722081a
LJ
269 unsigned long min_alloc_buckets_order;
270 unsigned long min_nr_alloc_buckets;
747d725c 271 unsigned long max_nr_buckets;
b8af5011 272 int flags;
5f511391
MD
273 /*
274 * We need to put the work threads offline (QSBR) when taking this
275 * mutex, because we use synchronize_rcu within this mutex critical
276 * section, which waits on read-side critical sections, and could
277 * therefore cause grace-period deadlock if we hold off RCU G.P.
278 * completion.
279 */
464a1ec9 280 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 281 unsigned int in_progress_resize, in_progress_destroy;
7b17c13e 282 const struct rcu_flavor_struct *flavor;
b7d619b0 283 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 284 long count; /* global approximate item count */
4c42f1b8 285 struct ht_items_count *split_count; /* split item count */
2ed95849
MD
286};
287
7f52427b
MD
288/*
289 * rcu_resize_work: Contains arguments passed to RCU worker thread
290 * responsible for performing lazy resize.
291 */
abc490a1
MD
292struct rcu_resize_work {
293 struct rcu_head head;
14044b37 294 struct cds_lfht *ht;
abc490a1 295};
2ed95849 296
7f52427b
MD
297/*
298 * partition_resize_work: Contains arguments passed to worker threads
299 * executing the hash table resize on partitions of the hash table
300 * assigned to each processor's worker thread.
301 */
b7d619b0 302struct partition_resize_work {
1af6e26e 303 pthread_t thread_id;
b7d619b0
MD
304 struct cds_lfht *ht;
305 unsigned long i, start, len;
306 void (*fct)(struct cds_lfht *ht, unsigned long i,
307 unsigned long start, unsigned long len);
308};
309
76a73da8 310static
83beee94 311void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 312 cds_lfht_match_fct match,
996ff57c 313 const void *key,
83beee94
MD
314 unsigned long size,
315 struct cds_lfht_node *node,
316 struct cds_lfht_iter *unique_ret,
1ee8f000 317 int bucket);
48ed1c18 318
abc490a1
MD
319/*
320 * Algorithm to reverse bits in a word by lookup table, extended to
321 * 64-bit words.
f9830efd 322 * Source:
abc490a1 323 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 324 * Originally from Public Domain.
abc490a1
MD
325 */
326
327static const uint8_t BitReverseTable256[256] =
2ed95849 328{
abc490a1
MD
329#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
330#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
331#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
332 R6(0), R6(2), R6(1), R6(3)
333};
334#undef R2
335#undef R4
336#undef R6
2ed95849 337
abc490a1
MD
338static
339uint8_t bit_reverse_u8(uint8_t v)
340{
341 return BitReverseTable256[v];
342}
ab7d5fc6 343
abc490a1
MD
344static __attribute__((unused))
345uint32_t bit_reverse_u32(uint32_t v)
346{
347 return ((uint32_t) bit_reverse_u8(v) << 24) |
348 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
349 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
350 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
351}
352
abc490a1
MD
353static __attribute__((unused))
354uint64_t bit_reverse_u64(uint64_t v)
2ed95849 355{
abc490a1
MD
356 return ((uint64_t) bit_reverse_u8(v) << 56) |
357 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
358 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
359 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
360 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
361 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
362 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
363 ((uint64_t) bit_reverse_u8(v >> 56));
364}
365
366static
367unsigned long bit_reverse_ulong(unsigned long v)
368{
369#if (CAA_BITS_PER_LONG == 32)
370 return bit_reverse_u32(v);
371#else
372 return bit_reverse_u64(v);
373#endif
374}
375
f9830efd 376/*
24365af7
MD
377 * fls: returns the position of the most significant bit.
378 * Returns 0 if no bit is set, else returns the position of the most
379 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 380 */
24365af7
MD
381#if defined(__i386) || defined(__x86_64)
382static inline
383unsigned int fls_u32(uint32_t x)
f9830efd 384{
24365af7
MD
385 int r;
386
387 asm("bsrl %1,%0\n\t"
388 "jnz 1f\n\t"
389 "movl $-1,%0\n\t"
390 "1:\n\t"
391 : "=r" (r) : "rm" (x));
392 return r + 1;
393}
394#define HAS_FLS_U32
395#endif
396
397#if defined(__x86_64)
398static inline
399unsigned int fls_u64(uint64_t x)
400{
401 long r;
402
403 asm("bsrq %1,%0\n\t"
404 "jnz 1f\n\t"
405 "movq $-1,%0\n\t"
406 "1:\n\t"
407 : "=r" (r) : "rm" (x));
408 return r + 1;
409}
410#define HAS_FLS_U64
411#endif
412
413#ifndef HAS_FLS_U64
414static __attribute__((unused))
415unsigned int fls_u64(uint64_t x)
416{
417 unsigned int r = 64;
418
419 if (!x)
420 return 0;
421
422 if (!(x & 0xFFFFFFFF00000000ULL)) {
423 x <<= 32;
424 r -= 32;
425 }
426 if (!(x & 0xFFFF000000000000ULL)) {
427 x <<= 16;
428 r -= 16;
429 }
430 if (!(x & 0xFF00000000000000ULL)) {
431 x <<= 8;
432 r -= 8;
433 }
434 if (!(x & 0xF000000000000000ULL)) {
435 x <<= 4;
436 r -= 4;
437 }
438 if (!(x & 0xC000000000000000ULL)) {
439 x <<= 2;
440 r -= 2;
441 }
442 if (!(x & 0x8000000000000000ULL)) {
443 x <<= 1;
444 r -= 1;
445 }
446 return r;
447}
448#endif
449
450#ifndef HAS_FLS_U32
451static __attribute__((unused))
452unsigned int fls_u32(uint32_t x)
453{
454 unsigned int r = 32;
f9830efd 455
24365af7
MD
456 if (!x)
457 return 0;
458 if (!(x & 0xFFFF0000U)) {
459 x <<= 16;
460 r -= 16;
461 }
462 if (!(x & 0xFF000000U)) {
463 x <<= 8;
464 r -= 8;
465 }
466 if (!(x & 0xF0000000U)) {
467 x <<= 4;
468 r -= 4;
469 }
470 if (!(x & 0xC0000000U)) {
471 x <<= 2;
472 r -= 2;
473 }
474 if (!(x & 0x80000000U)) {
475 x <<= 1;
476 r -= 1;
477 }
478 return r;
479}
480#endif
481
482unsigned int fls_ulong(unsigned long x)
f9830efd 483{
6887cc5e 484#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
485 return fls_u32(x);
486#else
487 return fls_u64(x);
488#endif
489}
f9830efd 490
920f8ef6
LJ
491/*
492 * Return the minimum order for which x <= (1UL << order).
493 * Return -1 if x is 0.
494 */
24365af7
MD
495int get_count_order_u32(uint32_t x)
496{
920f8ef6
LJ
497 if (!x)
498 return -1;
24365af7 499
920f8ef6 500 return fls_u32(x - 1);
24365af7
MD
501}
502
920f8ef6
LJ
503/*
504 * Return the minimum order for which x <= (1UL << order).
505 * Return -1 if x is 0.
506 */
24365af7
MD
507int get_count_order_ulong(unsigned long x)
508{
920f8ef6
LJ
509 if (!x)
510 return -1;
24365af7 511
920f8ef6 512 return fls_ulong(x - 1);
f9830efd
MD
513}
514
98808fb1 515#ifdef POISON_FREE
5afadd12
LJ
516#define poison_free(ptr) \
517 do { \
518 if (ptr) { \
519 memset(ptr, 0x42, sizeof(*(ptr))); \
520 free(ptr); \
521 } \
98808fb1
MD
522 } while (0)
523#else
524#define poison_free(ptr) free(ptr)
525#endif
526
f9830efd 527static
ab65b890 528void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 529
f8994aee 530static
4105056a 531void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
532 unsigned long count);
533
df44348d 534static long nr_cpus_mask = -1;
4c42f1b8
LJ
535static long split_count_mask = -1;
536
4ddbb355 537#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
538static void ht_init_nr_cpus_mask(void)
539{
540 long maxcpus;
541
542 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
543 if (maxcpus <= 0) {
544 nr_cpus_mask = -2;
545 return;
546 }
547 /*
548 * round up number of CPUs to next power of two, so we
549 * can use & for modulo.
550 */
551 maxcpus = 1UL << get_count_order_ulong(maxcpus);
552 nr_cpus_mask = maxcpus - 1;
553}
4ddbb355
LJ
554#else /* #if defined(HAVE_SYSCONF) */
555static void ht_init_nr_cpus_mask(void)
556{
557 nr_cpus_mask = -2;
558}
559#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
560
561static
5afadd12 562void alloc_split_items_count(struct cds_lfht *ht)
df44348d
MD
563{
564 struct ht_items_count *count;
565
4c42f1b8
LJ
566 if (nr_cpus_mask == -1) {
567 ht_init_nr_cpus_mask();
4ddbb355
LJ
568 if (nr_cpus_mask < 0)
569 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
570 else
571 split_count_mask = nr_cpus_mask;
df44348d 572 }
4c42f1b8 573
4ddbb355 574 assert(split_count_mask >= 0);
5afadd12
LJ
575
576 if (ht->flags & CDS_LFHT_ACCOUNTING) {
577 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
578 assert(ht->split_count);
579 } else {
580 ht->split_count = NULL;
581 }
df44348d
MD
582}
583
584static
5afadd12 585void free_split_items_count(struct cds_lfht *ht)
df44348d 586{
5afadd12 587 poison_free(ht->split_count);
df44348d
MD
588}
589
14360f1c 590#if defined(HAVE_SCHED_GETCPU)
df44348d 591static
14360f1c 592int ht_get_split_count_index(unsigned long hash)
df44348d
MD
593{
594 int cpu;
595
4c42f1b8 596 assert(split_count_mask >= 0);
df44348d 597 cpu = sched_getcpu();
8ed51e04 598 if (caa_unlikely(cpu < 0))
14360f1c 599 return hash & split_count_mask;
df44348d 600 else
4c42f1b8 601 return cpu & split_count_mask;
df44348d 602}
14360f1c
LJ
603#else /* #if defined(HAVE_SCHED_GETCPU) */
604static
605int ht_get_split_count_index(unsigned long hash)
606{
607 return hash & split_count_mask;
608}
609#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
610
611static
14360f1c 612void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 613{
4c42f1b8
LJ
614 unsigned long split_count;
615 int index;
df44348d 616
8ed51e04 617 if (caa_unlikely(!ht->split_count))
3171717f 618 return;
14360f1c 619 index = ht_get_split_count_index(hash);
4c42f1b8 620 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
8ed51e04 621 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 622 long count;
df44348d 623
4c42f1b8 624 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
625 count = uatomic_add_return(&ht->count,
626 1UL << COUNT_COMMIT_ORDER);
627 /* If power of 2 */
628 if (!(count & (count - 1))) {
4105056a 629 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 630 return;
e3ecefd6 631 dbg_printf("add set global %ld\n", count);
4105056a 632 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 633 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
634 }
635 }
636}
637
638static
14360f1c 639void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 640{
4c42f1b8
LJ
641 unsigned long split_count;
642 int index;
df44348d 643
8ed51e04 644 if (caa_unlikely(!ht->split_count))
3171717f 645 return;
14360f1c 646 index = ht_get_split_count_index(hash);
4c42f1b8 647 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
8ed51e04 648 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 649 long count;
df44348d 650
4c42f1b8 651 dbg_printf("del split count %lu\n", split_count);
df44348d 652 count = uatomic_add_return(&ht->count,
3171717f 653 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
654 /* If power of 2 */
655 if (!(count & (count - 1))) {
4105056a 656 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 657 return;
e3ecefd6
MD
658 dbg_printf("del set global %ld\n", count);
659 /*
c941bb9e 660 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
661 * certain threshold.
662 */
4c42f1b8 663 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 664 return;
4105056a 665 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 666 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
667 }
668 }
669}
670
f9830efd 671static
4105056a 672void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 673{
f8994aee
MD
674 unsigned long count;
675
b8af5011
MD
676 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
677 return;
f8994aee
MD
678 count = uatomic_read(&ht->count);
679 /*
680 * Use bucket-local length for small table expand and for
681 * environments lacking per-cpu data support.
682 */
683 if (count >= (1UL << COUNT_COMMIT_ORDER))
684 return;
24365af7 685 if (chain_len > 100)
f0c29ed7 686 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 687 chain_len);
3390d470 688 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
ab65b890 689 cds_lfht_resize_lazy_grow(ht, size,
01370f0b 690 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
691}
692
abc490a1 693static
14044b37 694struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 695{
14044b37 696 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
697}
698
699static
14044b37 700int is_removed(struct cds_lfht_node *node)
abc490a1 701{
d37166c6 702 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
703}
704
705static
14044b37 706struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 707{
14044b37 708 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
709}
710
f5596c94 711static
1ee8f000 712int is_bucket(struct cds_lfht_node *node)
f5596c94 713{
1ee8f000 714 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
715}
716
717static
1ee8f000 718struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 719{
1ee8f000 720 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 721}
bb7b2f26
MD
722
723static
724struct cds_lfht_node *get_end(void)
725{
726 return (struct cds_lfht_node *) END_VALUE;
727}
728
729static
730int is_end(struct cds_lfht_node *node)
731{
732 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
733}
734
abc490a1 735static
ab65b890
LJ
736unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
737 unsigned long v)
abc490a1
MD
738{
739 unsigned long old1, old2;
740
741 old1 = uatomic_read(ptr);
742 do {
743 old2 = old1;
744 if (old2 >= v)
f9830efd 745 return old2;
abc490a1 746 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 747 return old2;
abc490a1
MD
748}
749
48f1b16d
LJ
750static
751void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
752{
753 if (order == 0) {
0722081a 754 ht->t.tbl[0] = calloc(ht->min_nr_alloc_buckets,
48f1b16d
LJ
755 sizeof(struct cds_lfht_node));
756 assert(ht->t.tbl[0]);
0722081a 757 } else if (order > ht->min_alloc_buckets_order) {
48f1b16d
LJ
758 ht->t.tbl[order] = calloc(1UL << (order -1),
759 sizeof(struct cds_lfht_node));
760 assert(ht->t.tbl[order]);
761 }
0722081a 762 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
48f1b16d
LJ
763}
764
765/*
766 * cds_lfht_free_bucket_table() should be called with decreasing order.
767 * When cds_lfht_free_bucket_table(0) is called, it means the whole
768 * lfht is destroyed.
769 */
770static
771void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
772{
773 if (order == 0)
774 poison_free(ht->t.tbl[0]);
0722081a 775 else if (order > ht->min_alloc_buckets_order)
48f1b16d 776 poison_free(ht->t.tbl[order]);
0722081a 777 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
48f1b16d
LJ
778}
779
9d72a73f
LJ
780static inline
781struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 782{
9d72a73f 783 unsigned long order;
ef6e6171 784
9d72a73f 785 if ((__builtin_constant_p(index) && index == 0)
0722081a 786 || index < ht->min_nr_alloc_buckets) {
9d72a73f 787 dbg_printf("bucket index %lu order 0 aridx 0\n", index);
dcc26ef4 788 return &ht->t.tbl[0][index];
ef6e6171 789 }
a4ea2223
LJ
790 /*
791 * equivalent to get_count_order_ulong(index + 1), but optimizes
792 * away the non-existing 0 special-case for
793 * get_count_order_ulong.
794 */
795 order = fls_ulong(index);
9d72a73f
LJ
796 dbg_printf("bucket index %lu order %lu aridx %lu\n",
797 index, order, index & ((1UL << (order - 1)) - 1));
dcc26ef4 798 return &ht->t.tbl[order][index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
799}
800
9d72a73f
LJ
801static inline
802struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
803 unsigned long hash)
804{
805 assert(size > 0);
806 return bucket_at(ht, hash & (size - 1));
807}
808
273399de
MD
809/*
810 * Remove all logically deleted nodes from a bucket up to a certain node key.
811 */
812static
1ee8f000 813void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 814{
14044b37 815 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 816
1ee8f000
LJ
817 assert(!is_bucket(bucket));
818 assert(!is_removed(bucket));
819 assert(!is_bucket(node));
c90201ac 820 assert(!is_removed(node));
273399de 821 for (;;) {
1ee8f000
LJ
822 iter_prev = bucket;
823 /* We can always skip the bucket node initially */
04db56f8 824 iter = rcu_dereference(iter_prev->next);
b4cb483f 825 assert(!is_removed(iter));
04db56f8 826 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 827 /*
1ee8f000 828 * We should never be called with bucket (start of chain)
bd4db153
MD
829 * and logically removed node (end of path compression
830 * marker) being the actual same node. This would be a
831 * bug in the algorithm implementation.
832 */
1ee8f000 833 assert(bucket != node);
273399de 834 for (;;) {
8ed51e04 835 if (caa_unlikely(is_end(iter)))
f9c80341 836 return;
04db56f8 837 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 838 return;
04db56f8 839 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 840 if (caa_likely(is_removed(next)))
273399de 841 break;
b453eae1 842 iter_prev = clear_flag(iter);
273399de
MD
843 iter = next;
844 }
b198f0fd 845 assert(!is_removed(iter));
1ee8f000
LJ
846 if (is_bucket(iter))
847 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
848 else
849 new_next = clear_flag(next);
04db56f8 850 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 851 }
f9c80341 852 return;
273399de
MD
853}
854
9357c415
MD
855static
856int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
857 struct cds_lfht_node *old_node,
3fb86f26 858 struct cds_lfht_node *old_next,
9357c415
MD
859 struct cds_lfht_node *new_node)
860{
04db56f8 861 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
862
863 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 864 return -ENOENT;
9357c415
MD
865
866 assert(!is_removed(old_node));
1ee8f000 867 assert(!is_bucket(old_node));
9357c415 868 assert(!is_removed(new_node));
1ee8f000 869 assert(!is_bucket(new_node));
9357c415 870 assert(new_node != old_node);
3fb86f26 871 for (;;) {
9357c415 872 /* Insert after node to be replaced */
9357c415
MD
873 if (is_removed(old_next)) {
874 /*
875 * Too late, the old node has been removed under us
876 * between lookup and replace. Fail.
877 */
7801dadd 878 return -ENOENT;
9357c415 879 }
1ee8f000 880 assert(!is_bucket(old_next));
9357c415 881 assert(new_node != clear_flag(old_next));
04db56f8 882 new_node->next = clear_flag(old_next);
9357c415
MD
883 /*
884 * Here is the whole trick for lock-free replace: we add
885 * the replacement node _after_ the node we want to
886 * replace by atomically setting its next pointer at the
887 * same time we set its removal flag. Given that
888 * the lookups/get next use an iterator aware of the
889 * next pointer, they will either skip the old node due
890 * to the removal flag and see the new node, or use
891 * the old node, but will not see the new one.
892 */
04db56f8 893 ret_next = uatomic_cmpxchg(&old_node->next,
9357c415 894 old_next, flag_removed(new_node));
3fb86f26 895 if (ret_next == old_next)
7801dadd 896 break; /* We performed the replacement. */
3fb86f26
LJ
897 old_next = ret_next;
898 }
9357c415 899
9357c415
MD
900 /*
901 * Ensure that the old node is not visible to readers anymore:
902 * lookup for the node, and remove it (along with any other
903 * logically removed node) if found.
904 */
04db56f8
LJ
905 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
906 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 907
04db56f8 908 assert(is_removed(rcu_dereference(old_node->next)));
7801dadd 909 return 0;
9357c415
MD
910}
911
83beee94
MD
912/*
913 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
914 * mode. A NULL unique_ret allows creation of duplicate keys.
915 */
abc490a1 916static
83beee94 917void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 918 cds_lfht_match_fct match,
996ff57c 919 const void *key,
83beee94
MD
920 unsigned long size,
921 struct cds_lfht_node *node,
922 struct cds_lfht_iter *unique_ret,
1ee8f000 923 int bucket_flag)
abc490a1 924{
14044b37 925 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 926 *return_node;
04db56f8 927 struct cds_lfht_node *bucket;
abc490a1 928
1ee8f000 929 assert(!is_bucket(node));
c90201ac 930 assert(!is_removed(node));
04db56f8 931 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
abc490a1 932 for (;;) {
adc0de68 933 uint32_t chain_len = 0;
abc490a1 934
11519af6
MD
935 /*
936 * iter_prev points to the non-removed node prior to the
937 * insert location.
11519af6 938 */
04db56f8 939 iter_prev = bucket;
1ee8f000 940 /* We can always skip the bucket node initially */
04db56f8
LJ
941 iter = rcu_dereference(iter_prev->next);
942 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 943 for (;;) {
8ed51e04 944 if (caa_unlikely(is_end(iter)))
273399de 945 goto insert;
04db56f8 946 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 947 goto insert;
238cc06e 948
1ee8f000
LJ
949 /* bucket node is the first node of the identical-hash-value chain */
950 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 951 goto insert;
238cc06e 952
04db56f8 953 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 954 if (caa_unlikely(is_removed(next)))
9dba85be 955 goto gc_node;
238cc06e
LJ
956
957 /* uniquely add */
83beee94 958 if (unique_ret
1ee8f000 959 && !is_bucket(next)
04db56f8 960 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
961 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
962
963 /*
964 * uniquely adding inserts the node as the first
965 * node of the identical-hash-value node chain.
966 *
967 * This semantic ensures no duplicated keys
968 * should ever be observable in the table
969 * (including observe one node by one node
970 * by forward iterations)
971 */
04db56f8 972 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
973 if (!d_iter.node)
974 goto insert;
975
976 *unique_ret = d_iter;
83beee94 977 return;
48ed1c18 978 }
238cc06e 979
11519af6 980 /* Only account for identical reverse hash once */
04db56f8 981 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 982 && !is_bucket(next))
4105056a 983 check_resize(ht, size, ++chain_len);
11519af6 984 iter_prev = clear_flag(iter);
273399de 985 iter = next;
abc490a1 986 }
48ed1c18 987
273399de 988 insert:
7ec59d3b 989 assert(node != clear_flag(iter));
11519af6 990 assert(!is_removed(iter_prev));
c90201ac 991 assert(!is_removed(iter));
f000907d 992 assert(iter_prev != node);
1ee8f000 993 if (!bucket_flag)
04db56f8 994 node->next = clear_flag(iter);
f9c80341 995 else
1ee8f000
LJ
996 node->next = flag_bucket(clear_flag(iter));
997 if (is_bucket(iter))
998 new_node = flag_bucket(node);
f5596c94
MD
999 else
1000 new_node = node;
04db56f8 1001 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1002 new_node) != iter) {
273399de 1003 continue; /* retry */
48ed1c18 1004 } else {
83beee94 1005 return_node = node;
960c9e4f 1006 goto end;
48ed1c18
MD
1007 }
1008
9dba85be
MD
1009 gc_node:
1010 assert(!is_removed(iter));
1ee8f000
LJ
1011 if (is_bucket(iter))
1012 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1013 else
1014 new_next = clear_flag(next);
04db56f8 1015 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1016 /* retry */
464a1ec9 1017 }
9357c415 1018end:
83beee94
MD
1019 if (unique_ret) {
1020 unique_ret->node = return_node;
1021 /* unique_ret->next left unset, never used. */
1022 }
abc490a1 1023}
464a1ec9 1024
abc490a1 1025static
860d07e8 1026int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 1027 struct cds_lfht_node *node,
1ee8f000 1028 int bucket_removal)
abc490a1 1029{
04db56f8 1030 struct cds_lfht_node *bucket, *next, *old;
5e28c532 1031
9357c415 1032 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1033 return -ENOENT;
9357c415 1034
7ec59d3b 1035 /* logically delete the node */
1ee8f000 1036 assert(!is_bucket(node));
c90201ac 1037 assert(!is_removed(node));
04db56f8 1038 old = rcu_dereference(node->next);
7ec59d3b 1039 do {
48ed1c18
MD
1040 struct cds_lfht_node *new_next;
1041
7ec59d3b 1042 next = old;
8ed51e04 1043 if (caa_unlikely(is_removed(next)))
743f9143 1044 return -ENOENT;
1ee8f000
LJ
1045 if (bucket_removal)
1046 assert(is_bucket(next));
1475579c 1047 else
1ee8f000 1048 assert(!is_bucket(next));
48ed1c18 1049 new_next = flag_removed(next);
04db56f8 1050 old = uatomic_cmpxchg(&node->next, next, new_next);
7ec59d3b 1051 } while (old != next);
7ec59d3b 1052 /* We performed the (logical) deletion. */
7ec59d3b
MD
1053
1054 /*
1055 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1056 * the node, and remove it (along with any other logically removed node)
1057 * if found.
11519af6 1058 */
04db56f8
LJ
1059 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1060 _cds_lfht_gc_bucket(bucket, node);
743f9143 1061
04db56f8 1062 assert(is_removed(rcu_dereference(node->next)));
743f9143 1063 return 0;
abc490a1 1064}
2ed95849 1065
b7d619b0
MD
1066static
1067void *partition_resize_thread(void *arg)
1068{
1069 struct partition_resize_work *work = arg;
1070
7b17c13e 1071 work->ht->flavor->register_thread();
b7d619b0 1072 work->fct(work->ht, work->i, work->start, work->len);
7b17c13e 1073 work->ht->flavor->unregister_thread();
b7d619b0
MD
1074 return NULL;
1075}
1076
1077static
1078void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1079 unsigned long len,
1080 void (*fct)(struct cds_lfht *ht, unsigned long i,
1081 unsigned long start, unsigned long len))
1082{
1083 unsigned long partition_len;
1084 struct partition_resize_work *work;
6083a889
MD
1085 int thread, ret;
1086 unsigned long nr_threads;
b7d619b0 1087
6083a889
MD
1088 /*
1089 * Note: nr_cpus_mask + 1 is always power of 2.
1090 * We spawn just the number of threads we need to satisfy the minimum
1091 * partition size, up to the number of CPUs in the system.
1092 */
91452a6a
MD
1093 if (nr_cpus_mask > 0) {
1094 nr_threads = min(nr_cpus_mask + 1,
1095 len >> MIN_PARTITION_PER_THREAD_ORDER);
1096 } else {
1097 nr_threads = 1;
1098 }
6083a889
MD
1099 partition_len = len >> get_count_order_ulong(nr_threads);
1100 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1101 assert(work);
6083a889
MD
1102 for (thread = 0; thread < nr_threads; thread++) {
1103 work[thread].ht = ht;
1104 work[thread].i = i;
1105 work[thread].len = partition_len;
1106 work[thread].start = thread * partition_len;
1107 work[thread].fct = fct;
1af6e26e 1108 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1109 partition_resize_thread, &work[thread]);
b7d619b0
MD
1110 assert(!ret);
1111 }
6083a889 1112 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1113 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1114 assert(!ret);
1115 }
1116 free(work);
b7d619b0
MD
1117}
1118
e8de508e
MD
1119/*
1120 * Holding RCU read lock to protect _cds_lfht_add against memory
1121 * reclaim that could be performed by other call_rcu worker threads (ABA
1122 * problem).
9ee0fc9a 1123 *
b7d619b0 1124 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1125 * many worker threads, based on the number of CPUs available in the system.
1126 * This should therefore take care of not having the expand lagging behind too
1127 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1128 * schedule bucket node population fairly with insertions.
e8de508e 1129 */
4105056a 1130static
b7d619b0
MD
1131void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1132 unsigned long start, unsigned long len)
4105056a 1133{
9d72a73f 1134 unsigned long j, size = 1UL << (i - 1);
4105056a 1135
d0d8f9aa 1136 assert(i > MIN_TABLE_ORDER);
7b17c13e 1137 ht->flavor->read_lock();
9d72a73f
LJ
1138 for (j = size + start; j < size + start + len; j++) {
1139 struct cds_lfht_node *new_node = bucket_at(ht, j);
1140
1141 assert(j >= size && j < (size << 1));
1142 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1143 i, j, j);
1144 new_node->reverse_hash = bit_reverse_ulong(j);
1145 _cds_lfht_add(ht, NULL, NULL, size, new_node, NULL, 1);
4105056a 1146 }
7b17c13e 1147 ht->flavor->read_unlock();
b7d619b0
MD
1148}
1149
1150static
1151void init_table_populate(struct cds_lfht *ht, unsigned long i,
1152 unsigned long len)
1153{
1154 assert(nr_cpus_mask != -1);
6083a889 1155 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
7b17c13e 1156 ht->flavor->thread_online();
b7d619b0 1157 init_table_populate_partition(ht, i, 0, len);
7b17c13e 1158 ht->flavor->thread_offline();
b7d619b0
MD
1159 return;
1160 }
1161 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1162}
1163
abc490a1 1164static
4105056a 1165void init_table(struct cds_lfht *ht,
93d46c39 1166 unsigned long first_order, unsigned long last_order)
24365af7 1167{
93d46c39 1168 unsigned long i;
24365af7 1169
93d46c39
LJ
1170 dbg_printf("init table: first_order %lu last_order %lu\n",
1171 first_order, last_order);
d0d8f9aa 1172 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1173 for (i = first_order; i <= last_order; i++) {
4105056a 1174 unsigned long len;
24365af7 1175
4f6e90b7 1176 len = 1UL << (i - 1);
f0c29ed7 1177 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1178
1179 /* Stop expand if the resize target changes under us */
4f6e90b7 1180 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1181 break;
1182
48f1b16d 1183 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1184
4105056a 1185 /*
1ee8f000
LJ
1186 * Set all bucket nodes reverse hash values for a level and
1187 * link all bucket nodes into the table.
4105056a 1188 */
dc1da8f6 1189 init_table_populate(ht, i, len);
4105056a 1190
f9c80341
MD
1191 /*
1192 * Update table size.
1193 */
1194 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1195 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1196
4f6e90b7 1197 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1198 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1199 break;
1200 }
1201}
1202
e8de508e
MD
1203/*
1204 * Holding RCU read lock to protect _cds_lfht_remove against memory
1205 * reclaim that could be performed by other call_rcu worker threads (ABA
1206 * problem).
1207 * For a single level, we logically remove and garbage collect each node.
1208 *
1209 * As a design choice, we perform logical removal and garbage collection on a
1210 * node-per-node basis to simplify this algorithm. We also assume keeping good
1211 * cache locality of the operation would overweight possible performance gain
1212 * that could be achieved by batching garbage collection for multiple levels.
1213 * However, this would have to be justified by benchmarks.
1214 *
1215 * Concurrent removal and add operations are helping us perform garbage
1216 * collection of logically removed nodes. We guarantee that all logically
1217 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1218 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e
MD
1219 *
1220 * Logical removal and garbage collection can therefore be done in batch or on a
1221 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1222 *
b7d619b0
MD
1223 * When we reach a certain length, we can split this removal over many worker
1224 * threads, based on the number of CPUs available in the system. This should
1225 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1226 * updater threads actively inserting into the hash table.
e8de508e 1227 */
4105056a 1228static
b7d619b0
MD
1229void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1230 unsigned long start, unsigned long len)
4105056a 1231{
9d72a73f 1232 unsigned long j, size = 1UL << (i - 1);
4105056a 1233
d0d8f9aa 1234 assert(i > MIN_TABLE_ORDER);
7b17c13e 1235 ht->flavor->read_lock();
9d72a73f
LJ
1236 for (j = size + start; j < size + start + len; j++) {
1237 struct cds_lfht_node *fini_node = bucket_at(ht, j);
1238
1239 assert(j >= size && j < (size << 1));
1240 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1241 i, j, j);
1242 fini_node->reverse_hash = bit_reverse_ulong(j);
1243 (void) _cds_lfht_del(ht, size, fini_node, 1);
abc490a1 1244 }
7b17c13e 1245 ht->flavor->read_unlock();
b7d619b0
MD
1246}
1247
1248static
1249void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1250{
1251
1252 assert(nr_cpus_mask != -1);
6083a889 1253 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
7b17c13e 1254 ht->flavor->thread_online();
b7d619b0 1255 remove_table_partition(ht, i, 0, len);
7b17c13e 1256 ht->flavor->thread_offline();
b7d619b0
MD
1257 return;
1258 }
1259 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1260}
1261
1475579c 1262static
4105056a 1263void fini_table(struct cds_lfht *ht,
93d46c39 1264 unsigned long first_order, unsigned long last_order)
1475579c 1265{
93d46c39 1266 long i;
48f1b16d 1267 unsigned long free_by_rcu_order = 0;
1475579c 1268
93d46c39
LJ
1269 dbg_printf("fini table: first_order %lu last_order %lu\n",
1270 first_order, last_order);
d0d8f9aa 1271 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1272 for (i = last_order; i >= first_order; i--) {
4105056a 1273 unsigned long len;
1475579c 1274
4f6e90b7 1275 len = 1UL << (i - 1);
1475579c 1276 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1277
4d676753
MD
1278 /* Stop shrink if the resize target changes under us */
1279 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1280 break;
1281
1282 cmm_smp_wmb(); /* populate data before RCU size */
1283 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1284
1285 /*
1286 * We need to wait for all add operations to reach Q.S. (and
1287 * thus use the new table for lookups) before we can start
1ee8f000 1288 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1289 * return a logically removed node as insert position.
1290 */
7b17c13e 1291 ht->flavor->update_synchronize_rcu();
48f1b16d
LJ
1292 if (free_by_rcu_order)
1293 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1294
21263e21 1295 /*
1ee8f000
LJ
1296 * Set "removed" flag in bucket nodes about to be removed.
1297 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1298 * Concurrent add/remove operation are helping us doing
1299 * the gc.
21263e21 1300 */
4105056a
MD
1301 remove_table(ht, i, len);
1302
48f1b16d 1303 free_by_rcu_order = i;
4105056a
MD
1304
1305 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1306 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1307 break;
1308 }
0d14ceb2 1309
48f1b16d 1310 if (free_by_rcu_order) {
7b17c13e 1311 ht->flavor->update_synchronize_rcu();
48f1b16d 1312 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1313 }
1475579c
MD
1314}
1315
ff0d69de 1316static
1ee8f000 1317void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1318{
04db56f8 1319 struct cds_lfht_node *prev, *node;
9d72a73f 1320 unsigned long order, len, i;
ff0d69de 1321
48f1b16d 1322 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1323
9d72a73f
LJ
1324 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1325 node = bucket_at(ht, 0);
1326 node->next = flag_bucket(get_end());
1327 node->reverse_hash = 0;
ff0d69de
LJ
1328
1329 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1330 len = 1UL << (order - 1);
48f1b16d 1331 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1332
9d72a73f
LJ
1333 for (i = 0; i < len; i++) {
1334 /*
1335 * Now, we are trying to init the node with the
1336 * hash=(len+i) (which is also a bucket with the
1337 * index=(len+i)) and insert it into the hash table,
1338 * so this node has to be inserted after the bucket
1339 * with the index=(len+i)&(len-1)=i. And because there
1340 * is no other non-bucket node nor bucket node with
1341 * larger index/hash inserted, so the bucket node
1342 * being inserted should be inserted directly linked
1343 * after the bucket node with index=i.
1344 */
1345 prev = bucket_at(ht, i);
1346 node = bucket_at(ht, len + i);
ff0d69de 1347
1ee8f000 1348 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1349 order, len + i, len + i);
1350 node->reverse_hash = bit_reverse_ulong(len + i);
1351
1352 /* insert after prev */
1353 assert(is_bucket(prev->next));
ff0d69de 1354 node->next = prev->next;
1ee8f000 1355 prev->next = flag_bucket(node);
ff0d69de
LJ
1356 }
1357 }
1358}
1359
0422d92c 1360struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1361 unsigned long min_nr_alloc_buckets,
747d725c 1362 unsigned long max_nr_buckets,
b8af5011 1363 int flags,
7b17c13e 1364 const struct rcu_flavor_struct *flavor,
b7d619b0 1365 pthread_attr_t *attr)
abc490a1 1366{
14044b37 1367 struct cds_lfht *ht;
24365af7 1368 unsigned long order;
abc490a1 1369
0722081a
LJ
1370 /* min_nr_alloc_buckets must be power of two */
1371 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1372 return NULL;
747d725c 1373
8129be4e 1374 /* init_size must be power of two */
5488222b 1375 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1376 return NULL;
747d725c
LJ
1377
1378 if (!max_nr_buckets)
1379 max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1);
1380
1381 /* max_nr_buckets must be power of two */
1382 if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1)))
1383 return NULL;
1384
0722081a 1385 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1386 init_size = max(init_size, MIN_TABLE_SIZE);
747d725c
LJ
1387 max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets);
1388 init_size = min(init_size, max_nr_buckets);
14044b37 1389 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1390 assert(ht);
b5d6b20f 1391 ht->flags = flags;
7b17c13e 1392 ht->flavor = flavor;
b7d619b0 1393 ht->resize_attr = attr;
5afadd12 1394 alloc_split_items_count(ht);
abc490a1
MD
1395 /* this mutex should not nest in read-side C.S. */
1396 pthread_mutex_init(&ht->resize_mutex, NULL);
5488222b 1397 order = get_count_order_ulong(init_size);
93d46c39 1398 ht->t.resize_target = 1UL << order;
0722081a
LJ
1399 ht->min_nr_alloc_buckets = min_nr_alloc_buckets;
1400 ht->min_alloc_buckets_order = get_count_order_ulong(min_nr_alloc_buckets);
747d725c 1401 ht->max_nr_buckets = max_nr_buckets;
1ee8f000 1402 cds_lfht_create_bucket(ht, 1UL << order);
bcbd36fc 1403 ht->t.size = 1UL << order;
abc490a1
MD
1404 return ht;
1405}
1406
6f554439 1407void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1408 cds_lfht_match_fct match, const void *key,
6f554439 1409 struct cds_lfht_iter *iter)
2ed95849 1410{
04db56f8 1411 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1412 unsigned long reverse_hash, size;
2ed95849 1413
abc490a1 1414 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1415
4105056a 1416 size = rcu_dereference(ht->t.size);
04db56f8 1417 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1418 /* We can always skip the bucket node initially */
04db56f8 1419 node = rcu_dereference(bucket->next);
bb7b2f26 1420 node = clear_flag(node);
2ed95849 1421 for (;;) {
8ed51e04 1422 if (caa_unlikely(is_end(node))) {
96ad1112 1423 node = next = NULL;
abc490a1 1424 break;
bb7b2f26 1425 }
04db56f8 1426 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1427 node = next = NULL;
abc490a1 1428 break;
2ed95849 1429 }
04db56f8 1430 next = rcu_dereference(node->next);
7f52427b 1431 assert(node == clear_flag(node));
8ed51e04 1432 if (caa_likely(!is_removed(next))
1ee8f000 1433 && !is_bucket(next)
04db56f8 1434 && node->reverse_hash == reverse_hash
0422d92c 1435 && caa_likely(match(node, key))) {
273399de 1436 break;
2ed95849 1437 }
1b81fe1a 1438 node = clear_flag(next);
2ed95849 1439 }
1ee8f000 1440 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1441 iter->node = node;
1442 iter->next = next;
abc490a1 1443}
e0ba718a 1444
0422d92c 1445void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1446 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1447{
adc0de68 1448 struct cds_lfht_node *node, *next;
a481e5ff 1449 unsigned long reverse_hash;
a481e5ff 1450
adc0de68 1451 node = iter->node;
04db56f8 1452 reverse_hash = node->reverse_hash;
adc0de68 1453 next = iter->next;
a481e5ff
MD
1454 node = clear_flag(next);
1455
1456 for (;;) {
8ed51e04 1457 if (caa_unlikely(is_end(node))) {
96ad1112 1458 node = next = NULL;
a481e5ff 1459 break;
bb7b2f26 1460 }
04db56f8 1461 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1462 node = next = NULL;
a481e5ff
MD
1463 break;
1464 }
04db56f8 1465 next = rcu_dereference(node->next);
8ed51e04 1466 if (caa_likely(!is_removed(next))
1ee8f000 1467 && !is_bucket(next)
04db56f8 1468 && caa_likely(match(node, key))) {
a481e5ff
MD
1469 break;
1470 }
1471 node = clear_flag(next);
1472 }
1ee8f000 1473 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1474 iter->node = node;
1475 iter->next = next;
a481e5ff
MD
1476}
1477
4e9b9fbf
MD
1478void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1479{
1480 struct cds_lfht_node *node, *next;
1481
853395e1 1482 node = clear_flag(iter->next);
4e9b9fbf 1483 for (;;) {
8ed51e04 1484 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1485 node = next = NULL;
1486 break;
1487 }
04db56f8 1488 next = rcu_dereference(node->next);
8ed51e04 1489 if (caa_likely(!is_removed(next))
1ee8f000 1490 && !is_bucket(next)) {
4e9b9fbf
MD
1491 break;
1492 }
1493 node = clear_flag(next);
1494 }
1ee8f000 1495 assert(!node || !is_bucket(rcu_dereference(node->next)));
4e9b9fbf
MD
1496 iter->node = node;
1497 iter->next = next;
1498}
1499
1500void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1501{
4e9b9fbf 1502 /*
1ee8f000 1503 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1504 * first node of the linked list.
1505 */
9d72a73f 1506 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1507 cds_lfht_next(ht, iter);
1508}
1509
0422d92c
MD
1510void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1511 struct cds_lfht_node *node)
abc490a1 1512{
0422d92c 1513 unsigned long size;
ab7d5fc6 1514
04db56f8 1515 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1516 size = rcu_dereference(ht->t.size);
04db56f8 1517 _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0);
14360f1c 1518 ht_count_add(ht, size, hash);
3eca1b8c
MD
1519}
1520
14044b37 1521struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1522 unsigned long hash,
0422d92c 1523 cds_lfht_match_fct match,
996ff57c 1524 const void *key,
48ed1c18 1525 struct cds_lfht_node *node)
3eca1b8c 1526{
0422d92c 1527 unsigned long size;
83beee94 1528 struct cds_lfht_iter iter;
3eca1b8c 1529
04db56f8 1530 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1531 size = rcu_dereference(ht->t.size);
04db56f8 1532 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1533 if (iter.node == node)
14360f1c 1534 ht_count_add(ht, size, hash);
83beee94 1535 return iter.node;
2ed95849
MD
1536}
1537
9357c415 1538struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1539 unsigned long hash,
0422d92c 1540 cds_lfht_match_fct match,
996ff57c 1541 const void *key,
48ed1c18
MD
1542 struct cds_lfht_node *node)
1543{
0422d92c 1544 unsigned long size;
83beee94 1545 struct cds_lfht_iter iter;
48ed1c18 1546
04db56f8 1547 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
48ed1c18 1548 size = rcu_dereference(ht->t.size);
83beee94 1549 for (;;) {
04db56f8 1550 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1551 if (iter.node == node) {
14360f1c 1552 ht_count_add(ht, size, hash);
83beee94
MD
1553 return NULL;
1554 }
1555
1556 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1557 return iter.node;
1558 }
48ed1c18
MD
1559}
1560
9357c415
MD
1561int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1562 struct cds_lfht_node *new_node)
1563{
1564 unsigned long size;
1565
1566 size = rcu_dereference(ht->t.size);
1567 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1568 new_node);
1569}
1570
1571int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1572{
14360f1c 1573 unsigned long size, hash;
df44348d 1574 int ret;
abc490a1 1575
4105056a 1576 size = rcu_dereference(ht->t.size);
9357c415 1577 ret = _cds_lfht_del(ht, size, iter->node, 0);
14360f1c 1578 if (!ret) {
04db56f8 1579 hash = bit_reverse_ulong(iter->node->reverse_hash);
14360f1c
LJ
1580 ht_count_del(ht, size, hash);
1581 }
df44348d 1582 return ret;
2ed95849 1583}
ab7d5fc6 1584
abc490a1 1585static
1ee8f000 1586int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1587{
14044b37 1588 struct cds_lfht_node *node;
4105056a 1589 unsigned long order, i, size;
674f7a69 1590
abc490a1 1591 /* Check that the table is empty */
9d72a73f 1592 node = bucket_at(ht, 0);
abc490a1 1593 do {
04db56f8 1594 node = clear_flag(node)->next;
1ee8f000 1595 if (!is_bucket(node))
abc490a1 1596 return -EPERM;
273399de 1597 assert(!is_removed(node));
bb7b2f26 1598 } while (!is_end(node));
4105056a
MD
1599 /*
1600 * size accessed without rcu_dereference because hash table is
1601 * being destroyed.
1602 */
1603 size = ht->t.size;
1ee8f000 1604 /* Internal sanity check: all nodes left should be bucket */
48f1b16d
LJ
1605 for (i = 0; i < size; i++) {
1606 node = bucket_at(ht, i);
1607 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1608 i, i, bit_reverse_ulong(node->reverse_hash));
1609 assert(is_bucket(node->next));
1610 }
24365af7 1611
48f1b16d
LJ
1612 for (order = get_count_order_ulong(size); (long)order >= 0; order--)
1613 cds_lfht_free_bucket_table(ht, order);
5488222b 1614
abc490a1 1615 return 0;
674f7a69
MD
1616}
1617
1618/*
1619 * Should only be called when no more concurrent readers nor writers can
1620 * possibly access the table.
1621 */
b7d619b0 1622int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1623{
5e28c532
MD
1624 int ret;
1625
848d4088 1626 /* Wait for in-flight resize operations to complete */
24953e08
MD
1627 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1628 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1629 while (uatomic_read(&ht->in_progress_resize))
1630 poll(NULL, 0, 100); /* wait for 100ms */
1ee8f000 1631 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1632 if (ret)
1633 return ret;
5afadd12 1634 free_split_items_count(ht);
b7d619b0
MD
1635 if (attr)
1636 *attr = ht->resize_attr;
98808fb1 1637 poison_free(ht);
5e28c532 1638 return ret;
674f7a69
MD
1639}
1640
14044b37 1641void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1642 long *approx_before,
273399de 1643 unsigned long *count,
973e5e1b 1644 unsigned long *removed,
d933dd0e 1645 long *approx_after)
273399de 1646{
14044b37 1647 struct cds_lfht_node *node, *next;
1ee8f000 1648 unsigned long nr_bucket = 0;
273399de 1649
7ed7682f 1650 *approx_before = 0;
5afadd12 1651 if (ht->split_count) {
973e5e1b
MD
1652 int i;
1653
4c42f1b8
LJ
1654 for (i = 0; i < split_count_mask + 1; i++) {
1655 *approx_before += uatomic_read(&ht->split_count[i].add);
1656 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1657 }
1658 }
1659
273399de
MD
1660 *count = 0;
1661 *removed = 0;
1662
1ee8f000 1663 /* Count non-bucket nodes in the table */
9d72a73f 1664 node = bucket_at(ht, 0);
273399de 1665 do {
04db56f8 1666 next = rcu_dereference(node->next);
b198f0fd 1667 if (is_removed(next)) {
1ee8f000 1668 if (!is_bucket(next))
973e5e1b
MD
1669 (*removed)++;
1670 else
1ee8f000
LJ
1671 (nr_bucket)++;
1672 } else if (!is_bucket(next))
273399de 1673 (*count)++;
24365af7 1674 else
1ee8f000 1675 (nr_bucket)++;
273399de 1676 node = clear_flag(next);
bb7b2f26 1677 } while (!is_end(node));
1ee8f000 1678 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1679 *approx_after = 0;
5afadd12 1680 if (ht->split_count) {
973e5e1b
MD
1681 int i;
1682
4c42f1b8
LJ
1683 for (i = 0; i < split_count_mask + 1; i++) {
1684 *approx_after += uatomic_read(&ht->split_count[i].add);
1685 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1686 }
1687 }
273399de
MD
1688}
1689
1475579c 1690/* called with resize mutex held */
abc490a1 1691static
4105056a 1692void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1693 unsigned long old_size, unsigned long new_size)
abc490a1 1694{
1475579c 1695 unsigned long old_order, new_order;
1475579c 1696
93d46c39
LJ
1697 old_order = get_count_order_ulong(old_size);
1698 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1699 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1700 old_size, old_order, new_size, new_order);
1475579c 1701 assert(new_size > old_size);
93d46c39 1702 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1703}
1704
1705/* called with resize mutex held */
1706static
4105056a 1707void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1708 unsigned long old_size, unsigned long new_size)
464a1ec9 1709{
1475579c 1710 unsigned long old_order, new_order;
464a1ec9 1711
d0d8f9aa 1712 new_size = max(new_size, MIN_TABLE_SIZE);
93d46c39
LJ
1713 old_order = get_count_order_ulong(old_size);
1714 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1715 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1716 old_size, old_order, new_size, new_order);
1475579c 1717 assert(new_size < old_size);
1475579c 1718
1ee8f000 1719 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1720 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1721}
1722
1475579c
MD
1723
1724/* called with resize mutex held */
1725static
1726void _do_cds_lfht_resize(struct cds_lfht *ht)
1727{
1728 unsigned long new_size, old_size;
4105056a
MD
1729
1730 /*
1731 * Resize table, re-do if the target size has changed under us.
1732 */
1733 do {
d2be3620
MD
1734 assert(uatomic_read(&ht->in_progress_resize));
1735 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1736 break;
4105056a
MD
1737 ht->t.resize_initiated = 1;
1738 old_size = ht->t.size;
1739 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1740 if (old_size < new_size)
1741 _do_cds_lfht_grow(ht, old_size, new_size);
1742 else if (old_size > new_size)
1743 _do_cds_lfht_shrink(ht, old_size, new_size);
1744 ht->t.resize_initiated = 0;
1745 /* write resize_initiated before read resize_target */
1746 cmm_smp_mb();
4d676753 1747 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1748}
1749
abc490a1 1750static
ab65b890 1751unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1752{
ab65b890 1753 return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size);
464a1ec9
MD
1754}
1755
1475579c 1756static
4105056a 1757void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1758 unsigned long count)
1475579c 1759{
d0d8f9aa 1760 count = max(count, MIN_TABLE_SIZE);
747d725c 1761 count = min(count, ht->max_nr_buckets);
4105056a 1762 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1763}
1764
1765void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1766{
4105056a
MD
1767 resize_target_update_count(ht, new_size);
1768 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
7b17c13e 1769 ht->flavor->thread_offline();
1475579c
MD
1770 pthread_mutex_lock(&ht->resize_mutex);
1771 _do_cds_lfht_resize(ht);
1772 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1773 ht->flavor->thread_online();
abc490a1 1774}
464a1ec9 1775
abc490a1
MD
1776static
1777void do_resize_cb(struct rcu_head *head)
1778{
1779 struct rcu_resize_work *work =
1780 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1781 struct cds_lfht *ht = work->ht;
abc490a1 1782
7b17c13e 1783 ht->flavor->thread_offline();
abc490a1 1784 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1785 _do_cds_lfht_resize(ht);
abc490a1 1786 pthread_mutex_unlock(&ht->resize_mutex);
7b17c13e 1787 ht->flavor->thread_online();
98808fb1 1788 poison_free(work);
848d4088
MD
1789 cmm_smp_mb(); /* finish resize before decrement */
1790 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1791}
1792
abc490a1 1793static
f1f119ee 1794void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1795{
abc490a1
MD
1796 struct rcu_resize_work *work;
1797
4105056a
MD
1798 /* Store resize_target before read resize_initiated */
1799 cmm_smp_mb();
ab65b890 1800 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
848d4088 1801 uatomic_inc(&ht->in_progress_resize);
59290e9d 1802 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1803 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1804 uatomic_dec(&ht->in_progress_resize);
59290e9d 1805 return;
ed35e6d8 1806 }
f9830efd
MD
1807 work = malloc(sizeof(*work));
1808 work->ht = ht;
7b17c13e 1809 ht->flavor->update_call_rcu(&work->head, do_resize_cb);
4105056a 1810 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1811 }
ab7d5fc6 1812}
3171717f 1813
f1f119ee
LJ
1814static
1815void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1816{
1817 unsigned long target_size = size << growth;
1818
747d725c 1819 target_size = min(target_size, ht->max_nr_buckets);
f1f119ee
LJ
1820 if (resize_target_grow(ht, target_size) >= target_size)
1821 return;
1822
1823 __cds_lfht_resize_lazy_launch(ht);
1824}
1825
89bb121d
LJ
1826/*
1827 * We favor grow operations over shrink. A shrink operation never occurs
1828 * if a grow operation is queued for lazy execution. A grow operation
1829 * cancels any pending shrink lazy execution.
1830 */
3171717f 1831static
4105056a 1832void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1833 unsigned long count)
1834{
b8af5011
MD
1835 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1836 return;
d0d8f9aa 1837 count = max(count, MIN_TABLE_SIZE);
747d725c 1838 count = min(count, ht->max_nr_buckets);
89bb121d
LJ
1839 if (count == size)
1840 return; /* Already the right size, no resize needed */
1841 if (count > size) { /* lazy grow */
1842 if (resize_target_grow(ht, count) >= count)
1843 return;
1844 } else { /* lazy shrink */
1845 for (;;) {
1846 unsigned long s;
1847
1848 s = uatomic_cmpxchg(&ht->t.resize_target, size, count);
1849 if (s == size)
1850 break; /* no resize needed */
1851 if (s > size)
1852 return; /* growing is/(was just) in progress */
1853 if (s <= count)
1854 return; /* some other thread do shrink */
1855 size = s;
1856 }
1857 }
f1f119ee 1858 __cds_lfht_resize_lazy_launch(ht);
3171717f 1859}
This page took 0.226734 seconds and 4 git commands to generate.