Rename the parameter name of _cds_lfht_replace()
[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>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
5e28c532
MD
21 */
22
e753ff5a
MD
23/*
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
30 * (2002), 73-82.
31 *
1475579c 32 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
33 * implementation:
34 *
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
44 * duplicata exists.
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
1475579c
MD
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
53 * operation.
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
83 * list concurrently.
29e669f6
MD
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
88 * for it do to so.
e753ff5a
MD
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
94 * this order.
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
1475579c
MD
99 *
100 * A bit of ascii art explanation:
101 *
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
104 *
105 * This shows the nodes for a small table ordered by reversed bits:
106 *
107 * bits reverse
108 * 0 000 000
109 * 4 100 001
110 * 2 010 010
111 * 6 110 011
112 * 1 001 100
113 * 5 101 101
114 * 3 011 110
115 * 7 111 111
116 *
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
119 *
120 * order bits reverse
121 * 0 0 000 000
122 * |
f6fdd688
MD
123 * 1 | 1 001 100 <- <-
124 * | | | |
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
127 * | | | | | | |
1475579c
MD
128 * 3 -> | | | 4 100 001 | |
129 * -> | | 5 101 101 |
130 * -> | 6 110 011
131 * -> 7 111 111
e753ff5a
MD
132 */
133
2ed95849
MD
134#define _LGPL_SOURCE
135#include <stdlib.h>
e0ba718a
MD
136#include <errno.h>
137#include <assert.h>
138#include <stdio.h>
abc490a1 139#include <stdint.h>
f000907d 140#include <string.h>
e0ba718a 141
15cfbec7 142#include "config.h"
2ed95849 143#include <urcu.h>
abc490a1 144#include <urcu-call-rcu.h>
a42cc659
MD
145#include <urcu/arch.h>
146#include <urcu/uatomic.h>
a42cc659 147#include <urcu/compiler.h>
abc490a1 148#include <urcu/rculfhash.h>
5e28c532 149#include <stdio.h>
464a1ec9 150#include <pthread.h>
44395fb7 151
f9830efd 152#ifdef DEBUG
f0c29ed7 153#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 154#else
e753ff5a 155#define dbg_printf(fmt, args...)
f9830efd
MD
156#endif
157
f8994aee
MD
158/*
159 * Per-CPU split-counters lazily update the global counter each 1024
160 * addition/removal. It automatically keeps track of resize required.
161 * We use the bucket length as indicator for need to expand for small
162 * tables and machines lacking per-cpu data suppport.
163 */
164#define COUNT_COMMIT_ORDER 10
6ea6bc67
MD
165#define CHAIN_LEN_TARGET 1
166#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 167
cd95516d 168/*
76a73da8 169 * Define the minimum table size.
cd95516d 170 */
c9edd44a 171#define MIN_TABLE_SIZE 1
cd95516d 172
4105056a
MD
173#if (CAA_BITS_PER_LONG == 32)
174#define MAX_TABLE_ORDER 32
175#else
176#define MAX_TABLE_ORDER 64
177#endif
178
b7d619b0
MD
179/*
180 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
181 */
6083a889
MD
182#define MIN_PARTITION_PER_THREAD_ORDER 12
183#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 184
4105056a
MD
185#ifndef min
186#define min(a, b) ((a) < (b) ? (a) : (b))
187#endif
188
abc490a1
MD
189#ifndef max
190#define max(a, b) ((a) > (b) ? (a) : (b))
191#endif
2ed95849 192
d95bd160
MD
193/*
194 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 195 * It indicates that no node must attach to the node scheduled for
b198f0fd 196 * removal, and that node garbage collection must be performed.
d95bd160
MD
197 * The dummy flag does not require to be updated atomically with the
198 * pointer, but it is added as a pointer low bit flag to save space.
199 */
d37166c6 200#define REMOVED_FLAG (1UL << 0)
b198f0fd
MD
201#define DUMMY_FLAG (1UL << 1)
202#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 203
bb7b2f26 204/* Value of the end pointer. Should not interact with flags. */
f9c80341 205#define END_VALUE NULL
bb7b2f26 206
df44348d 207struct ht_items_count {
860d07e8 208 unsigned long add, del;
df44348d
MD
209} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
210
1475579c 211struct rcu_level {
0d14ceb2 212 /* Note: manually update allocation length when adding a field */
1475579c
MD
213 struct _cds_lfht_node nodes[0];
214};
215
395270b6 216struct rcu_table {
4105056a 217 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 218 unsigned long resize_target;
11519af6 219 int resize_initiated;
4105056a 220 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
221};
222
14044b37 223struct cds_lfht {
4105056a 224 struct rcu_table t;
14044b37
MD
225 cds_lfht_hash_fct hash_fct;
226 cds_lfht_compare_fct compare_fct;
732ad076 227 unsigned long hash_seed;
b8af5011 228 int flags;
5f511391
MD
229 /*
230 * We need to put the work threads offline (QSBR) when taking this
231 * mutex, because we use synchronize_rcu within this mutex critical
232 * section, which waits on read-side critical sections, and could
233 * therefore cause grace-period deadlock if we hold off RCU G.P.
234 * completion.
235 */
464a1ec9 236 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 237 unsigned int in_progress_resize, in_progress_destroy;
14044b37 238 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 239 void (*func)(struct rcu_head *head));
1475579c 240 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
241 void (*cds_lfht_rcu_read_lock)(void);
242 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
243 void (*cds_lfht_rcu_thread_offline)(void);
244 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
245 void (*cds_lfht_rcu_register_thread)(void);
246 void (*cds_lfht_rcu_unregister_thread)(void);
247 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 248 long count; /* global approximate item count */
df44348d 249 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
250};
251
abc490a1
MD
252struct rcu_resize_work {
253 struct rcu_head head;
14044b37 254 struct cds_lfht *ht;
abc490a1 255};
2ed95849 256
b7d619b0 257struct partition_resize_work {
1af6e26e 258 pthread_t thread_id;
b7d619b0
MD
259 struct cds_lfht *ht;
260 unsigned long i, start, len;
261 void (*fct)(struct cds_lfht *ht, unsigned long i,
262 unsigned long start, unsigned long len);
263};
264
48ed1c18
MD
265enum add_mode {
266 ADD_DEFAULT = 0,
267 ADD_UNIQUE = 1,
268 ADD_REPLACE = 2,
269};
270
76a73da8
MD
271static
272struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
273 unsigned long size,
274 struct cds_lfht_node *node,
48ed1c18
MD
275 enum add_mode mode, int dummy);
276
abc490a1
MD
277/*
278 * Algorithm to reverse bits in a word by lookup table, extended to
279 * 64-bit words.
f9830efd 280 * Source:
abc490a1 281 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 282 * Originally from Public Domain.
abc490a1
MD
283 */
284
285static const uint8_t BitReverseTable256[256] =
2ed95849 286{
abc490a1
MD
287#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
288#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
289#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
290 R6(0), R6(2), R6(1), R6(3)
291};
292#undef R2
293#undef R4
294#undef R6
2ed95849 295
abc490a1
MD
296static
297uint8_t bit_reverse_u8(uint8_t v)
298{
299 return BitReverseTable256[v];
300}
ab7d5fc6 301
abc490a1
MD
302static __attribute__((unused))
303uint32_t bit_reverse_u32(uint32_t v)
304{
305 return ((uint32_t) bit_reverse_u8(v) << 24) |
306 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
307 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
308 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
309}
310
abc490a1
MD
311static __attribute__((unused))
312uint64_t bit_reverse_u64(uint64_t v)
2ed95849 313{
abc490a1
MD
314 return ((uint64_t) bit_reverse_u8(v) << 56) |
315 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
316 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
317 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
318 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
319 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
320 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
321 ((uint64_t) bit_reverse_u8(v >> 56));
322}
323
324static
325unsigned long bit_reverse_ulong(unsigned long v)
326{
327#if (CAA_BITS_PER_LONG == 32)
328 return bit_reverse_u32(v);
329#else
330 return bit_reverse_u64(v);
331#endif
332}
333
f9830efd 334/*
24365af7
MD
335 * fls: returns the position of the most significant bit.
336 * Returns 0 if no bit is set, else returns the position of the most
337 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 338 */
24365af7
MD
339#if defined(__i386) || defined(__x86_64)
340static inline
341unsigned int fls_u32(uint32_t x)
f9830efd 342{
24365af7
MD
343 int r;
344
345 asm("bsrl %1,%0\n\t"
346 "jnz 1f\n\t"
347 "movl $-1,%0\n\t"
348 "1:\n\t"
349 : "=r" (r) : "rm" (x));
350 return r + 1;
351}
352#define HAS_FLS_U32
353#endif
354
355#if defined(__x86_64)
356static inline
357unsigned int fls_u64(uint64_t x)
358{
359 long r;
360
361 asm("bsrq %1,%0\n\t"
362 "jnz 1f\n\t"
363 "movq $-1,%0\n\t"
364 "1:\n\t"
365 : "=r" (r) : "rm" (x));
366 return r + 1;
367}
368#define HAS_FLS_U64
369#endif
370
371#ifndef HAS_FLS_U64
372static __attribute__((unused))
373unsigned int fls_u64(uint64_t x)
374{
375 unsigned int r = 64;
376
377 if (!x)
378 return 0;
379
380 if (!(x & 0xFFFFFFFF00000000ULL)) {
381 x <<= 32;
382 r -= 32;
383 }
384 if (!(x & 0xFFFF000000000000ULL)) {
385 x <<= 16;
386 r -= 16;
387 }
388 if (!(x & 0xFF00000000000000ULL)) {
389 x <<= 8;
390 r -= 8;
391 }
392 if (!(x & 0xF000000000000000ULL)) {
393 x <<= 4;
394 r -= 4;
395 }
396 if (!(x & 0xC000000000000000ULL)) {
397 x <<= 2;
398 r -= 2;
399 }
400 if (!(x & 0x8000000000000000ULL)) {
401 x <<= 1;
402 r -= 1;
403 }
404 return r;
405}
406#endif
407
408#ifndef HAS_FLS_U32
409static __attribute__((unused))
410unsigned int fls_u32(uint32_t x)
411{
412 unsigned int r = 32;
f9830efd 413
24365af7
MD
414 if (!x)
415 return 0;
416 if (!(x & 0xFFFF0000U)) {
417 x <<= 16;
418 r -= 16;
419 }
420 if (!(x & 0xFF000000U)) {
421 x <<= 8;
422 r -= 8;
423 }
424 if (!(x & 0xF0000000U)) {
425 x <<= 4;
426 r -= 4;
427 }
428 if (!(x & 0xC0000000U)) {
429 x <<= 2;
430 r -= 2;
431 }
432 if (!(x & 0x80000000U)) {
433 x <<= 1;
434 r -= 1;
435 }
436 return r;
437}
438#endif
439
440unsigned int fls_ulong(unsigned long x)
f9830efd 441{
24365af7
MD
442#if (CAA_BITS_PER_lONG == 32)
443 return fls_u32(x);
444#else
445 return fls_u64(x);
446#endif
447}
f9830efd 448
920f8ef6
LJ
449/*
450 * Return the minimum order for which x <= (1UL << order).
451 * Return -1 if x is 0.
452 */
24365af7
MD
453int get_count_order_u32(uint32_t x)
454{
920f8ef6
LJ
455 if (!x)
456 return -1;
24365af7 457
920f8ef6 458 return fls_u32(x - 1);
24365af7
MD
459}
460
920f8ef6
LJ
461/*
462 * Return the minimum order for which x <= (1UL << order).
463 * Return -1 if x is 0.
464 */
24365af7
MD
465int get_count_order_ulong(unsigned long x)
466{
920f8ef6
LJ
467 if (!x)
468 return -1;
24365af7 469
920f8ef6 470 return fls_ulong(x - 1);
f9830efd
MD
471}
472
98808fb1
MD
473#ifdef POISON_FREE
474#define poison_free(ptr) \
475 do { \
476 memset(ptr, 0x42, sizeof(*(ptr))); \
477 free(ptr); \
478 } while (0)
479#else
480#define poison_free(ptr) free(ptr)
481#endif
482
f9830efd 483static
4105056a 484void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 485
df44348d
MD
486/*
487 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
488 * available, then we support hash table item accounting.
489 * In the unfortunate event the number of CPUs reported would be
490 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
491 */
df44348d
MD
492#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
493
f8994aee 494static
4105056a 495void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
496 unsigned long count);
497
df44348d
MD
498static long nr_cpus_mask = -1;
499
500static
501struct ht_items_count *alloc_per_cpu_items_count(void)
502{
503 struct ht_items_count *count;
504
505 switch (nr_cpus_mask) {
506 case -2:
507 return NULL;
508 case -1:
509 {
510 long maxcpus;
511
512 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
513 if (maxcpus <= 0) {
514 nr_cpus_mask = -2;
515 return NULL;
516 }
517 /*
518 * round up number of CPUs to next power of two, so we
519 * can use & for modulo.
520 */
521 maxcpus = 1UL << get_count_order_ulong(maxcpus);
522 nr_cpus_mask = maxcpus - 1;
523 }
524 /* Fall-through */
525 default:
526 return calloc(nr_cpus_mask + 1, sizeof(*count));
527 }
528}
529
530static
531void free_per_cpu_items_count(struct ht_items_count *count)
532{
98808fb1 533 poison_free(count);
df44348d
MD
534}
535
536static
537int ht_get_cpu(void)
538{
539 int cpu;
540
541 assert(nr_cpus_mask >= 0);
542 cpu = sched_getcpu();
543 if (unlikely(cpu < 0))
544 return cpu;
545 else
546 return cpu & nr_cpus_mask;
547}
548
549static
4105056a 550void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 551{
3171717f 552 unsigned long percpu_count;
df44348d
MD
553 int cpu;
554
555 if (unlikely(!ht->percpu_count))
3171717f 556 return;
df44348d
MD
557 cpu = ht_get_cpu();
558 if (unlikely(cpu < 0))
3171717f
MD
559 return;
560 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 561 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 562 long count;
df44348d
MD
563
564 dbg_printf("add percpu %lu\n", percpu_count);
565 count = uatomic_add_return(&ht->count,
566 1UL << COUNT_COMMIT_ORDER);
567 /* If power of 2 */
568 if (!(count & (count - 1))) {
4105056a 569 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 570 return;
e3ecefd6 571 dbg_printf("add set global %ld\n", count);
4105056a 572 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 573 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
574 }
575 }
576}
577
578static
860d07e8 579void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
580{
581 unsigned long percpu_count;
3171717f 582 int cpu;
df44348d 583
3171717f
MD
584 if (unlikely(!ht->percpu_count))
585 return;
586 cpu = ht_get_cpu();
587 if (unlikely(cpu < 0))
588 return;
80d90c06 589 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 590 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 591 long count;
df44348d 592
860d07e8 593 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 594 count = uatomic_add_return(&ht->count,
3171717f 595 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
596 /* If power of 2 */
597 if (!(count & (count - 1))) {
4105056a 598 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 599 return;
e3ecefd6
MD
600 dbg_printf("del set global %ld\n", count);
601 /*
c941bb9e 602 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
603 * certain threshold.
604 */
c941bb9e 605 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 606 return;
4105056a 607 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 608 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
609 }
610 }
611}
612
613#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
614
91452a6a 615static const long nr_cpus_mask = -2;
df44348d
MD
616
617static
618struct ht_items_count *alloc_per_cpu_items_count(void)
619{
620 return NULL;
621}
622
623static
624void free_per_cpu_items_count(struct ht_items_count *count)
625{
626}
627
628static
4105056a 629void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
630{
631}
632
633static
860d07e8 634void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
635{
636}
637
638#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
639
640
f9830efd 641static
4105056a 642void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 643{
f8994aee
MD
644 unsigned long count;
645
b8af5011
MD
646 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
647 return;
f8994aee
MD
648 count = uatomic_read(&ht->count);
649 /*
650 * Use bucket-local length for small table expand and for
651 * environments lacking per-cpu data support.
652 */
653 if (count >= (1UL << COUNT_COMMIT_ORDER))
654 return;
24365af7 655 if (chain_len > 100)
f0c29ed7 656 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 657 chain_len);
3390d470 658 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 659 cds_lfht_resize_lazy(ht, size,
01370f0b 660 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
661}
662
abc490a1 663static
14044b37 664struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 665{
14044b37 666 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
667}
668
669static
14044b37 670int is_removed(struct cds_lfht_node *node)
abc490a1 671{
d37166c6 672 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
673}
674
675static
14044b37 676struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 677{
14044b37 678 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
679}
680
f5596c94 681static
14044b37 682int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
683{
684 return ((unsigned long) node) & DUMMY_FLAG;
685}
686
687static
14044b37 688struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 689{
14044b37 690 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 691}
bb7b2f26
MD
692
693static
694struct cds_lfht_node *get_end(void)
695{
696 return (struct cds_lfht_node *) END_VALUE;
697}
698
699static
700int is_end(struct cds_lfht_node *node)
701{
702 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
703}
704
abc490a1 705static
f9830efd 706unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
707{
708 unsigned long old1, old2;
709
710 old1 = uatomic_read(ptr);
711 do {
712 old2 = old1;
713 if (old2 >= v)
f9830efd 714 return old2;
abc490a1 715 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 716 return v;
abc490a1
MD
717}
718
f4a9cc0b
LJ
719static
720struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
721 unsigned long hash)
722{
723 unsigned long index, order;
724
725 assert(size > 0);
726 index = hash & (size - 1);
a4ea2223
LJ
727 /*
728 * equivalent to get_count_order_ulong(index + 1), but optimizes
729 * away the non-existing 0 special-case for
730 * get_count_order_ulong.
731 */
732 order = fls_ulong(index);
f4a9cc0b
LJ
733
734 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
735 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
736
737 return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
738}
739
273399de
MD
740/*
741 * Remove all logically deleted nodes from a bucket up to a certain node key.
742 */
743static
f9c80341 744void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 745{
14044b37 746 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 747
c90201ac
MD
748 assert(!is_dummy(dummy));
749 assert(!is_removed(dummy));
750 assert(!is_dummy(node));
751 assert(!is_removed(node));
273399de
MD
752 for (;;) {
753 iter_prev = dummy;
754 /* We can always skip the dummy node initially */
cc4fcb10
MD
755 iter = rcu_dereference(iter_prev->p.next);
756 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
757 /*
758 * We should never be called with dummy (start of chain)
759 * and logically removed node (end of path compression
760 * marker) being the actual same node. This would be a
761 * bug in the algorithm implementation.
762 */
763 assert(dummy != node);
273399de 764 for (;;) {
bb7b2f26 765 if (unlikely(is_end(iter)))
f9c80341 766 return;
76412f24 767 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 768 return;
cc4fcb10 769 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 770 if (likely(is_removed(next)))
273399de 771 break;
b453eae1 772 iter_prev = clear_flag(iter);
273399de
MD
773 iter = next;
774 }
b198f0fd 775 assert(!is_removed(iter));
f5596c94
MD
776 if (is_dummy(iter))
777 new_next = flag_dummy(clear_flag(next));
778 else
779 new_next = clear_flag(next);
48ed1c18
MD
780 if (is_removed(iter))
781 new_next = flag_removed(new_next);
f5596c94 782 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 783 }
f9c80341 784 return;
273399de
MD
785}
786
9357c415
MD
787static
788int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
789 struct cds_lfht_node *old_node,
3fb86f26 790 struct cds_lfht_node *old_next,
9357c415
MD
791 struct cds_lfht_node *new_node)
792{
3fb86f26 793 struct cds_lfht_node *dummy, *ret_next;
9357c415
MD
794 struct _cds_lfht_node *lookup;
795 int flagged = 0;
9357c415
MD
796
797 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
798 goto end;
799
800 assert(!is_removed(old_node));
801 assert(!is_dummy(old_node));
802 assert(!is_removed(new_node));
803 assert(!is_dummy(new_node));
804 assert(new_node != old_node);
3fb86f26 805 for (;;) {
9357c415 806 /* Insert after node to be replaced */
9357c415
MD
807 if (is_removed(old_next)) {
808 /*
809 * Too late, the old node has been removed under us
810 * between lookup and replace. Fail.
811 */
812 goto end;
813 }
814 assert(!is_dummy(old_next));
815 assert(new_node != clear_flag(old_next));
816 new_node->p.next = clear_flag(old_next);
817 /*
818 * Here is the whole trick for lock-free replace: we add
819 * the replacement node _after_ the node we want to
820 * replace by atomically setting its next pointer at the
821 * same time we set its removal flag. Given that
822 * the lookups/get next use an iterator aware of the
823 * next pointer, they will either skip the old node due
824 * to the removal flag and see the new node, or use
825 * the old node, but will not see the new one.
826 */
827 ret_next = uatomic_cmpxchg(&old_node->p.next,
828 old_next, flag_removed(new_node));
3fb86f26
LJ
829 if (ret_next == old_next)
830 break;
831 old_next = ret_next;
832 }
9357c415
MD
833
834 /* We performed the replacement. */
835 flagged = 1;
836
837 /*
838 * Ensure that the old node is not visible to readers anymore:
839 * lookup for the node, and remove it (along with any other
840 * logically removed node) if found.
841 */
f4a9cc0b 842 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
843 dummy = (struct cds_lfht_node *) lookup;
844 _cds_lfht_gc_bucket(dummy, new_node);
845end:
846 /*
847 * Only the flagging action indicated that we (and no other)
848 * replaced the node from the hash table.
849 */
850 if (flagged) {
851 assert(is_removed(rcu_dereference(old_node->p.next)));
852 return 0;
853 } else {
854 return -ENOENT;
855 }
856}
857
abc490a1 858static
4105056a
MD
859struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
860 unsigned long size,
861 struct cds_lfht_node *node,
48ed1c18 862 enum add_mode mode, int dummy)
abc490a1 863{
14044b37 864 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 865 *return_node;
14044b37 866 struct _cds_lfht_node *lookup;
abc490a1 867
c90201ac
MD
868 assert(!is_dummy(node));
869 assert(!is_removed(node));
4105056a 870 if (!size) {
f5596c94 871 assert(dummy);
bb7b2f26 872 node->p.next = flag_dummy(get_end());
18117871
MD
873 return node; /* Initial first add (head) */
874 }
f4a9cc0b 875 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 876 for (;;) {
adc0de68 877 uint32_t chain_len = 0;
abc490a1 878
11519af6
MD
879 /*
880 * iter_prev points to the non-removed node prior to the
881 * insert location.
11519af6 882 */
14044b37 883 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 884 /* We can always skip the dummy node initially */
cc4fcb10
MD
885 iter = rcu_dereference(iter_prev->p.next);
886 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 887 for (;;) {
bb7b2f26 888 if (unlikely(is_end(iter)))
273399de 889 goto insert;
76412f24 890 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 891 goto insert;
194fdbd1
LJ
892 /* dummy node is the first node of the identical-hash-value chain */
893 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
894 goto insert;
cc4fcb10 895 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 896 if (unlikely(is_removed(next)))
9dba85be 897 goto gc_node;
48ed1c18 898 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
1b81fe1a 899 && !is_dummy(next)
ae450da3 900 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash
e43f23f8
MD
901 && !ht->compare_fct(node->key, node->key_len,
902 clear_flag(iter)->key,
48ed1c18
MD
903 clear_flag(iter)->key_len)) {
904 if (mode == ADD_UNIQUE)
905 return clear_flag(iter);
906 else /* mode == ADD_REPLACE */
907 goto replace;
908 }
11519af6 909 /* Only account for identical reverse hash once */
24365af7
MD
910 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
911 && !is_dummy(next))
4105056a 912 check_resize(ht, size, ++chain_len);
11519af6 913 iter_prev = clear_flag(iter);
273399de 914 iter = next;
abc490a1 915 }
48ed1c18 916
273399de 917 insert:
7ec59d3b 918 assert(node != clear_flag(iter));
11519af6 919 assert(!is_removed(iter_prev));
c90201ac 920 assert(!is_removed(iter));
f000907d 921 assert(iter_prev != node);
f9c80341 922 if (!dummy)
1b81fe1a 923 node->p.next = clear_flag(iter);
f9c80341
MD
924 else
925 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
926 if (is_dummy(iter))
927 new_node = flag_dummy(node);
928 else
929 new_node = node;
cc4fcb10 930 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 931 new_node) != iter) {
273399de 932 continue; /* retry */
48ed1c18
MD
933 } else {
934 if (mode == ADD_REPLACE)
935 return_node = NULL;
936 else /* ADD_DEFAULT and ADD_UNIQUE */
937 return_node = node;
960c9e4f 938 goto end;
48ed1c18
MD
939 }
940
941 replace:
9357c415
MD
942
943 if (!_cds_lfht_replace(ht, size, clear_flag(iter), next,
944 node)) {
945 return_node = clear_flag(iter);
946 goto end; /* gc already done */
48ed1c18 947 } else {
9357c415 948 continue; /* retry */
48ed1c18
MD
949 }
950
9dba85be
MD
951 gc_node:
952 assert(!is_removed(iter));
f5596c94
MD
953 if (is_dummy(iter))
954 new_next = flag_dummy(clear_flag(next));
955 else
956 new_next = clear_flag(next);
957 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 958 /* retry */
464a1ec9 959 }
9357c415 960end:
48ed1c18 961 return return_node;
abc490a1 962}
464a1ec9 963
abc490a1 964static
860d07e8 965int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 966 struct cds_lfht_node *node,
b198f0fd 967 int dummy_removal)
abc490a1 968{
14044b37
MD
969 struct cds_lfht_node *dummy, *next, *old;
970 struct _cds_lfht_node *lookup;
abc490a1 971 int flagged = 0;
5e28c532 972
9357c415
MD
973 if (!node) /* Return -ENOENT if asked to delete NULL node */
974 goto end;
975
7ec59d3b 976 /* logically delete the node */
c90201ac
MD
977 assert(!is_dummy(node));
978 assert(!is_removed(node));
cc4fcb10 979 old = rcu_dereference(node->p.next);
7ec59d3b 980 do {
48ed1c18
MD
981 struct cds_lfht_node *new_next;
982
7ec59d3b 983 next = old;
76412f24 984 if (unlikely(is_removed(next)))
7ec59d3b 985 goto end;
1475579c
MD
986 if (dummy_removal)
987 assert(is_dummy(next));
988 else
989 assert(!is_dummy(next));
48ed1c18 990 new_next = flag_removed(next);
48ed1c18 991 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
992 } while (old != next);
993
994 /* We performed the (logical) deletion. */
995 flagged = 1;
996
997 /*
998 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
999 * the node, and remove it (along with any other logically removed node)
1000 * if found.
11519af6 1001 */
f4a9cc0b 1002 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 1003 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1004 _cds_lfht_gc_bucket(dummy, node);
2ed95849 1005end:
11519af6
MD
1006 /*
1007 * Only the flagging action indicated that we (and no other)
1008 * removed the node from the hash.
1009 */
7ec59d3b 1010 if (flagged) {
cc4fcb10 1011 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 1012 return 0;
9357c415 1013 } else {
11519af6 1014 return -ENOENT;
9357c415 1015 }
abc490a1 1016}
2ed95849 1017
b7d619b0
MD
1018static
1019void *partition_resize_thread(void *arg)
1020{
1021 struct partition_resize_work *work = arg;
1022
1023 work->ht->cds_lfht_rcu_register_thread();
1024 work->fct(work->ht, work->i, work->start, work->len);
1025 work->ht->cds_lfht_rcu_unregister_thread();
1026 return NULL;
1027}
1028
1029static
1030void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1031 unsigned long len,
1032 void (*fct)(struct cds_lfht *ht, unsigned long i,
1033 unsigned long start, unsigned long len))
1034{
1035 unsigned long partition_len;
1036 struct partition_resize_work *work;
6083a889
MD
1037 int thread, ret;
1038 unsigned long nr_threads;
b7d619b0 1039
6083a889
MD
1040 /*
1041 * Note: nr_cpus_mask + 1 is always power of 2.
1042 * We spawn just the number of threads we need to satisfy the minimum
1043 * partition size, up to the number of CPUs in the system.
1044 */
91452a6a
MD
1045 if (nr_cpus_mask > 0) {
1046 nr_threads = min(nr_cpus_mask + 1,
1047 len >> MIN_PARTITION_PER_THREAD_ORDER);
1048 } else {
1049 nr_threads = 1;
1050 }
6083a889
MD
1051 partition_len = len >> get_count_order_ulong(nr_threads);
1052 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1053 assert(work);
6083a889
MD
1054 for (thread = 0; thread < nr_threads; thread++) {
1055 work[thread].ht = ht;
1056 work[thread].i = i;
1057 work[thread].len = partition_len;
1058 work[thread].start = thread * partition_len;
1059 work[thread].fct = fct;
1af6e26e 1060 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1061 partition_resize_thread, &work[thread]);
b7d619b0
MD
1062 assert(!ret);
1063 }
6083a889 1064 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1065 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1066 assert(!ret);
1067 }
1068 free(work);
b7d619b0
MD
1069}
1070
e8de508e
MD
1071/*
1072 * Holding RCU read lock to protect _cds_lfht_add against memory
1073 * reclaim that could be performed by other call_rcu worker threads (ABA
1074 * problem).
9ee0fc9a 1075 *
b7d619b0 1076 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1077 * many worker threads, based on the number of CPUs available in the system.
1078 * This should therefore take care of not having the expand lagging behind too
1079 * many concurrent insertion threads by using the scheduler's ability to
1080 * schedule dummy node population fairly with insertions.
e8de508e 1081 */
4105056a 1082static
b7d619b0
MD
1083void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1084 unsigned long start, unsigned long len)
4105056a
MD
1085{
1086 unsigned long j;
1087
1088 ht->cds_lfht_rcu_read_lock();
b7d619b0 1089 for (j = start; j < start + len; j++) {
4105056a
MD
1090 struct cds_lfht_node *new_node =
1091 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1092
dc1da8f6 1093 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1094 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1095 new_node->p.reverse_hash =
1096 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a 1097 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1098 new_node, ADD_DEFAULT, 1);
4105056a
MD
1099 }
1100 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1101}
1102
1103static
1104void init_table_populate(struct cds_lfht *ht, unsigned long i,
1105 unsigned long len)
1106{
1107 assert(nr_cpus_mask != -1);
6083a889 1108 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1109 ht->cds_lfht_rcu_thread_online();
1110 init_table_populate_partition(ht, i, 0, len);
1111 ht->cds_lfht_rcu_thread_offline();
1112 return;
1113 }
1114 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1115}
1116
abc490a1 1117static
4105056a 1118void init_table(struct cds_lfht *ht,
24365af7
MD
1119 unsigned long first_order, unsigned long len_order)
1120{
1121 unsigned long i, end_order;
1122
f0c29ed7 1123 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1124 first_order, first_order + len_order);
1125 end_order = first_order + len_order;
24365af7 1126 for (i = first_order; i < end_order; i++) {
4105056a 1127 unsigned long len;
24365af7
MD
1128
1129 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1130 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1131
1132 /* Stop expand if the resize target changes under us */
1133 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1134 break;
1135
0d14ceb2 1136 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1137 assert(ht->t.tbl[i]);
4105056a 1138
4105056a 1139 /*
dc1da8f6
MD
1140 * Set all dummy nodes reverse hash values for a level and
1141 * link all dummy nodes into the table.
4105056a 1142 */
dc1da8f6 1143 init_table_populate(ht, i, len);
4105056a 1144
f9c80341
MD
1145 /*
1146 * Update table size.
1147 */
1148 cmm_smp_wmb(); /* populate data before RCU size */
1149 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1150
4105056a
MD
1151 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1152 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1153 break;
1154 }
1155}
1156
e8de508e
MD
1157/*
1158 * Holding RCU read lock to protect _cds_lfht_remove against memory
1159 * reclaim that could be performed by other call_rcu worker threads (ABA
1160 * problem).
1161 * For a single level, we logically remove and garbage collect each node.
1162 *
1163 * As a design choice, we perform logical removal and garbage collection on a
1164 * node-per-node basis to simplify this algorithm. We also assume keeping good
1165 * cache locality of the operation would overweight possible performance gain
1166 * that could be achieved by batching garbage collection for multiple levels.
1167 * However, this would have to be justified by benchmarks.
1168 *
1169 * Concurrent removal and add operations are helping us perform garbage
1170 * collection of logically removed nodes. We guarantee that all logically
1171 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1172 * invoked to free a hole level of dummy nodes (after a grace period).
1173 *
1174 * Logical removal and garbage collection can therefore be done in batch or on a
1175 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1176 *
b7d619b0
MD
1177 * When we reach a certain length, we can split this removal over many worker
1178 * threads, based on the number of CPUs available in the system. This should
1179 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1180 * updater threads actively inserting into the hash table.
e8de508e 1181 */
4105056a 1182static
b7d619b0
MD
1183void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1184 unsigned long start, unsigned long len)
4105056a
MD
1185{
1186 unsigned long j;
1187
1188 ht->cds_lfht_rcu_read_lock();
b7d619b0 1189 for (j = start; j < start + len; j++) {
4105056a
MD
1190 struct cds_lfht_node *fini_node =
1191 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1192
1193 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1194 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1195 fini_node->p.reverse_hash =
1196 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1197 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1198 fini_node, 1);
abc490a1 1199 }
4105056a 1200 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1201}
1202
1203static
1204void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1205{
1206
1207 assert(nr_cpus_mask != -1);
6083a889 1208 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1209 ht->cds_lfht_rcu_thread_online();
1210 remove_table_partition(ht, i, 0, len);
1211 ht->cds_lfht_rcu_thread_offline();
1212 return;
1213 }
1214 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1215}
1216
1475579c 1217static
4105056a 1218void fini_table(struct cds_lfht *ht,
1475579c
MD
1219 unsigned long first_order, unsigned long len_order)
1220{
1221 long i, end_order;
0d14ceb2 1222 void *free_by_rcu = NULL;
1475579c
MD
1223
1224 dbg_printf("fini table: first_order %lu end_order %lu\n",
1225 first_order, first_order + len_order);
1226 end_order = first_order + len_order;
1227 assert(first_order > 0);
1475579c 1228 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1229 unsigned long len;
1475579c
MD
1230
1231 len = !i ? 1 : 1UL << (i - 1);
1232 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1233
4d676753
MD
1234 /* Stop shrink if the resize target changes under us */
1235 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1236 break;
1237
1238 cmm_smp_wmb(); /* populate data before RCU size */
1239 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1240
1241 /*
1242 * We need to wait for all add operations to reach Q.S. (and
1243 * thus use the new table for lookups) before we can start
1244 * releasing the old dummy nodes. Otherwise their lookup will
1245 * return a logically removed node as insert position.
1246 */
1247 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1248 if (free_by_rcu)
1249 free(free_by_rcu);
4d676753 1250
21263e21 1251 /*
4105056a
MD
1252 * Set "removed" flag in dummy nodes about to be removed.
1253 * Unlink all now-logically-removed dummy node pointers.
1254 * Concurrent add/remove operation are helping us doing
1255 * the gc.
21263e21 1256 */
4105056a
MD
1257 remove_table(ht, i, len);
1258
0d14ceb2 1259 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1260
1261 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1262 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1263 break;
1264 }
0d14ceb2
LJ
1265
1266 if (free_by_rcu) {
1267 ht->cds_lfht_synchronize_rcu();
1268 free(free_by_rcu);
1269 }
1475579c
MD
1270}
1271
7a9dcf9b 1272struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1273 cds_lfht_compare_fct compare_fct,
1274 unsigned long hash_seed,
1275 unsigned long init_size,
b8af5011 1276 int flags,
14044b37 1277 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1278 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1279 void (*cds_lfht_synchronize_rcu)(void),
1280 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1281 void (*cds_lfht_rcu_read_unlock)(void),
1282 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1283 void (*cds_lfht_rcu_thread_online)(void),
1284 void (*cds_lfht_rcu_register_thread)(void),
1285 void (*cds_lfht_rcu_unregister_thread)(void),
1286 pthread_attr_t *attr)
abc490a1 1287{
14044b37 1288 struct cds_lfht *ht;
24365af7 1289 unsigned long order;
abc490a1 1290
8129be4e 1291 /* init_size must be power of two */
49619ea0 1292 if (init_size && (init_size & (init_size - 1)))
8129be4e 1293 return NULL;
14044b37 1294 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1295 assert(ht);
abc490a1 1296 ht->hash_fct = hash_fct;
732ad076
MD
1297 ht->compare_fct = compare_fct;
1298 ht->hash_seed = hash_seed;
14044b37 1299 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1300 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1301 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1302 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1303 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1304 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1305 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1306 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1307 ht->resize_attr = attr;
df44348d 1308 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1309 /* this mutex should not nest in read-side C.S. */
1310 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1311 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1312 ht->flags = flags;
5f511391 1313 ht->cds_lfht_rcu_thread_offline();
f000907d 1314 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1315 ht->t.resize_target = 1UL << (order - 1);
4105056a 1316 init_table(ht, 0, order);
f000907d 1317 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1318 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1319 return ht;
1320}
1321
adc0de68
MD
1322void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1323 struct cds_lfht_iter *iter)
2ed95849 1324{
bb7b2f26 1325 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1326 struct _cds_lfht_node *lookup;
f4a9cc0b 1327 unsigned long hash, reverse_hash, size;
2ed95849 1328
732ad076 1329 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1330 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1331
4105056a 1332 size = rcu_dereference(ht->t.size);
f4a9cc0b 1333 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1334 dummy_node = (struct cds_lfht_node *) lookup;
1335 /* We can always skip the dummy node initially */
1336 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1337 node = clear_flag(node);
2ed95849 1338 for (;;) {
bb7b2f26 1339 if (unlikely(is_end(node))) {
96ad1112 1340 node = next = NULL;
abc490a1 1341 break;
bb7b2f26 1342 }
cc4fcb10 1343 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1344 node = next = NULL;
abc490a1 1345 break;
2ed95849 1346 }
1b81fe1a 1347 next = rcu_dereference(node->p.next);
adc0de68 1348 if (likely(!is_removed(next))
1b81fe1a 1349 && !is_dummy(next)
ae450da3 1350 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1351 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1352 break;
2ed95849 1353 }
1b81fe1a 1354 node = clear_flag(next);
2ed95849 1355 }
1b81fe1a 1356 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1357 iter->node = node;
1358 iter->next = next;
abc490a1 1359}
e0ba718a 1360
3883c0e5 1361void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1362{
adc0de68 1363 struct cds_lfht_node *node, *next;
a481e5ff
MD
1364 unsigned long reverse_hash;
1365 void *key;
1366 size_t key_len;
1367
adc0de68 1368 node = iter->node;
a481e5ff
MD
1369 reverse_hash = node->p.reverse_hash;
1370 key = node->key;
1371 key_len = node->key_len;
adc0de68 1372 next = iter->next;
a481e5ff
MD
1373 node = clear_flag(next);
1374
1375 for (;;) {
bb7b2f26 1376 if (unlikely(is_end(node))) {
96ad1112 1377 node = next = NULL;
a481e5ff 1378 break;
bb7b2f26 1379 }
a481e5ff 1380 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1381 node = next = NULL;
a481e5ff
MD
1382 break;
1383 }
1384 next = rcu_dereference(node->p.next);
adc0de68 1385 if (likely(!is_removed(next))
a481e5ff
MD
1386 && !is_dummy(next)
1387 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1388 break;
1389 }
1390 node = clear_flag(next);
1391 }
1392 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1393 iter->node = node;
1394 iter->next = next;
a481e5ff
MD
1395}
1396
4e9b9fbf
MD
1397void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1398{
1399 struct cds_lfht_node *node, *next;
1400
853395e1 1401 node = clear_flag(iter->next);
4e9b9fbf
MD
1402 for (;;) {
1403 if (unlikely(is_end(node))) {
1404 node = next = NULL;
1405 break;
1406 }
1407 next = rcu_dereference(node->p.next);
1408 if (likely(!is_removed(next))
1409 && !is_dummy(next)) {
1410 break;
1411 }
1412 node = clear_flag(next);
1413 }
1414 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1415 iter->node = node;
1416 iter->next = next;
1417}
1418
1419void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1420{
1421 struct _cds_lfht_node *lookup;
1422
1423 /*
1424 * Get next after first dummy node. The first dummy node is the
1425 * first node of the linked list.
1426 */
1427 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1428 iter->next = lookup->next;
4e9b9fbf
MD
1429 cds_lfht_next(ht, iter);
1430}
1431
14044b37 1432void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1433{
4105056a 1434 unsigned long hash, size;
ab7d5fc6 1435
49c2e2d6 1436 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1437 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1438
4105056a 1439 size = rcu_dereference(ht->t.size);
48ed1c18 1440 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
4105056a 1441 ht_count_add(ht, size);
3eca1b8c
MD
1442}
1443
14044b37 1444struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1445 struct cds_lfht_node *node)
3eca1b8c 1446{
4105056a 1447 unsigned long hash, size;
df44348d 1448 struct cds_lfht_node *ret;
3eca1b8c 1449
49c2e2d6 1450 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1451 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1452
4105056a 1453 size = rcu_dereference(ht->t.size);
48ed1c18 1454 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
17f31d1b 1455 if (ret == node)
4105056a 1456 ht_count_add(ht, size);
df44348d 1457 return ret;
2ed95849
MD
1458}
1459
9357c415 1460struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1461 struct cds_lfht_node *node)
1462{
1463 unsigned long hash, size;
1464 struct cds_lfht_node *ret;
1465
1466 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1467 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1468
1469 size = rcu_dereference(ht->t.size);
1470 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1471 if (ret == NULL)
1472 ht_count_add(ht, size);
1473 return ret;
1474}
1475
9357c415
MD
1476int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1477 struct cds_lfht_node *new_node)
1478{
1479 unsigned long size;
1480
1481 size = rcu_dereference(ht->t.size);
1482 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1483 new_node);
1484}
1485
1486int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1487{
4105056a 1488 unsigned long size;
df44348d 1489 int ret;
abc490a1 1490
4105056a 1491 size = rcu_dereference(ht->t.size);
9357c415 1492 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1493 if (!ret)
860d07e8 1494 ht_count_del(ht, size);
df44348d 1495 return ret;
2ed95849 1496}
ab7d5fc6 1497
abc490a1 1498static
14044b37 1499int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1500{
14044b37
MD
1501 struct cds_lfht_node *node;
1502 struct _cds_lfht_node *lookup;
4105056a 1503 unsigned long order, i, size;
674f7a69 1504
abc490a1 1505 /* Check that the table is empty */
4105056a 1506 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1507 node = (struct cds_lfht_node *) lookup;
abc490a1 1508 do {
1b81fe1a
MD
1509 node = clear_flag(node)->p.next;
1510 if (!is_dummy(node))
abc490a1 1511 return -EPERM;
273399de 1512 assert(!is_removed(node));
bb7b2f26 1513 } while (!is_end(node));
4105056a
MD
1514 /*
1515 * size accessed without rcu_dereference because hash table is
1516 * being destroyed.
1517 */
1518 size = ht->t.size;
abc490a1 1519 /* Internal sanity check: all nodes left should be dummy */
4105056a 1520 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1521 unsigned long len;
1522
1523 len = !order ? 1 : 1UL << (order - 1);
1524 for (i = 0; i < len; i++) {
f0c29ed7 1525 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1526 order, i,
4105056a
MD
1527 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1528 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1529 }
4105056a 1530 poison_free(ht->t.tbl[order]);
674f7a69 1531 }
abc490a1 1532 return 0;
674f7a69
MD
1533}
1534
1535/*
1536 * Should only be called when no more concurrent readers nor writers can
1537 * possibly access the table.
1538 */
b7d619b0 1539int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1540{
5e28c532
MD
1541 int ret;
1542
848d4088 1543 /* Wait for in-flight resize operations to complete */
24953e08
MD
1544 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1545 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1546 while (uatomic_read(&ht->in_progress_resize))
1547 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1548 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1549 if (ret)
1550 return ret;
df44348d 1551 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1552 if (attr)
1553 *attr = ht->resize_attr;
98808fb1 1554 poison_free(ht);
5e28c532 1555 return ret;
674f7a69
MD
1556}
1557
14044b37 1558void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1559 long *approx_before,
273399de 1560 unsigned long *count,
973e5e1b 1561 unsigned long *removed,
d933dd0e 1562 long *approx_after)
273399de 1563{
14044b37
MD
1564 struct cds_lfht_node *node, *next;
1565 struct _cds_lfht_node *lookup;
24365af7 1566 unsigned long nr_dummy = 0;
273399de 1567
7ed7682f 1568 *approx_before = 0;
973e5e1b
MD
1569 if (nr_cpus_mask >= 0) {
1570 int i;
1571
1572 for (i = 0; i < nr_cpus_mask + 1; i++) {
1573 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1574 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1575 }
1576 }
1577
273399de
MD
1578 *count = 0;
1579 *removed = 0;
1580
24365af7 1581 /* Count non-dummy nodes in the table */
4105056a 1582 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1583 node = (struct cds_lfht_node *) lookup;
273399de 1584 do {
cc4fcb10 1585 next = rcu_dereference(node->p.next);
b198f0fd 1586 if (is_removed(next)) {
973e5e1b
MD
1587 if (!is_dummy(next))
1588 (*removed)++;
1589 else
1590 (nr_dummy)++;
1b81fe1a 1591 } else if (!is_dummy(next))
273399de 1592 (*count)++;
24365af7
MD
1593 else
1594 (nr_dummy)++;
273399de 1595 node = clear_flag(next);
bb7b2f26 1596 } while (!is_end(node));
f0c29ed7 1597 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1598 *approx_after = 0;
973e5e1b
MD
1599 if (nr_cpus_mask >= 0) {
1600 int i;
1601
1602 for (i = 0; i < nr_cpus_mask + 1; i++) {
1603 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1604 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1605 }
1606 }
273399de
MD
1607}
1608
1475579c 1609/* called with resize mutex held */
abc490a1 1610static
4105056a 1611void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1612 unsigned long old_size, unsigned long new_size)
abc490a1 1613{
1475579c 1614 unsigned long old_order, new_order;
1475579c
MD
1615
1616 old_order = get_count_order_ulong(old_size) + 1;
1617 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1618 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1619 old_size, old_order, new_size, new_order);
1475579c 1620 assert(new_size > old_size);
4105056a 1621 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1622}
1623
1624/* called with resize mutex held */
1625static
4105056a 1626void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1627 unsigned long old_size, unsigned long new_size)
464a1ec9 1628{
1475579c 1629 unsigned long old_order, new_order;
464a1ec9 1630
cd95516d 1631 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1632 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1633 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1634 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1635 old_size, old_order, new_size, new_order);
1475579c 1636 assert(new_size < old_size);
1475579c 1637
4105056a
MD
1638 /* Remove and unlink all dummy nodes to remove. */
1639 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1640}
1641
1475579c
MD
1642
1643/* called with resize mutex held */
1644static
1645void _do_cds_lfht_resize(struct cds_lfht *ht)
1646{
1647 unsigned long new_size, old_size;
4105056a
MD
1648
1649 /*
1650 * Resize table, re-do if the target size has changed under us.
1651 */
1652 do {
d2be3620
MD
1653 assert(uatomic_read(&ht->in_progress_resize));
1654 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1655 break;
4105056a
MD
1656 ht->t.resize_initiated = 1;
1657 old_size = ht->t.size;
1658 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1659 if (old_size < new_size)
1660 _do_cds_lfht_grow(ht, old_size, new_size);
1661 else if (old_size > new_size)
1662 _do_cds_lfht_shrink(ht, old_size, new_size);
1663 ht->t.resize_initiated = 0;
1664 /* write resize_initiated before read resize_target */
1665 cmm_smp_mb();
4d676753 1666 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1667}
1668
abc490a1 1669static
4105056a 1670unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1671 int growth_order)
464a1ec9 1672{
4105056a
MD
1673 return _uatomic_max(&ht->t.resize_target,
1674 size << growth_order);
464a1ec9
MD
1675}
1676
1475579c 1677static
4105056a 1678void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1679 unsigned long count)
1475579c 1680{
cd95516d 1681 count = max(count, MIN_TABLE_SIZE);
4105056a 1682 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1683}
1684
1685void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1686{
4105056a
MD
1687 resize_target_update_count(ht, new_size);
1688 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1689 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1690 pthread_mutex_lock(&ht->resize_mutex);
1691 _do_cds_lfht_resize(ht);
1692 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1693 ht->cds_lfht_rcu_thread_online();
abc490a1 1694}
464a1ec9 1695
abc490a1
MD
1696static
1697void do_resize_cb(struct rcu_head *head)
1698{
1699 struct rcu_resize_work *work =
1700 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1701 struct cds_lfht *ht = work->ht;
abc490a1 1702
5f511391 1703 ht->cds_lfht_rcu_thread_offline();
abc490a1 1704 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1705 _do_cds_lfht_resize(ht);
abc490a1 1706 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1707 ht->cds_lfht_rcu_thread_online();
98808fb1 1708 poison_free(work);
848d4088
MD
1709 cmm_smp_mb(); /* finish resize before decrement */
1710 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1711}
1712
abc490a1 1713static
4105056a 1714void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1715{
abc490a1 1716 struct rcu_resize_work *work;
f9830efd 1717 unsigned long target_size;
abc490a1 1718
4105056a
MD
1719 target_size = resize_target_update(ht, size, growth);
1720 /* Store resize_target before read resize_initiated */
1721 cmm_smp_mb();
1722 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1723 uatomic_inc(&ht->in_progress_resize);
59290e9d 1724 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1725 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1726 uatomic_dec(&ht->in_progress_resize);
59290e9d 1727 return;
ed35e6d8 1728 }
f9830efd
MD
1729 work = malloc(sizeof(*work));
1730 work->ht = ht;
14044b37 1731 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1732 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1733 }
ab7d5fc6 1734}
3171717f 1735
f8994aee
MD
1736#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1737
3171717f 1738static
4105056a 1739void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1740 unsigned long count)
1741{
1742 struct rcu_resize_work *work;
3171717f 1743
b8af5011
MD
1744 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1745 return;
4105056a
MD
1746 resize_target_update_count(ht, count);
1747 /* Store resize_target before read resize_initiated */
1748 cmm_smp_mb();
1749 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1750 uatomic_inc(&ht->in_progress_resize);
59290e9d 1751 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1752 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1753 uatomic_dec(&ht->in_progress_resize);
59290e9d 1754 return;
ed35e6d8 1755 }
3171717f
MD
1756 work = malloc(sizeof(*work));
1757 work->ht = ht;
1758 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1759 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1760 }
1761}
f8994aee
MD
1762
1763#endif
This page took 0.200177 seconds and 4 git commands to generate.