Move replace code out from _cds_lfht_add()
[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
76a73da8 265static
83beee94
MD
266void _cds_lfht_add(struct cds_lfht *ht,
267 unsigned long size,
268 struct cds_lfht_node *node,
269 struct cds_lfht_iter *unique_ret,
270 int dummy);
48ed1c18 271
abc490a1
MD
272/*
273 * Algorithm to reverse bits in a word by lookup table, extended to
274 * 64-bit words.
f9830efd 275 * Source:
abc490a1 276 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 277 * Originally from Public Domain.
abc490a1
MD
278 */
279
280static const uint8_t BitReverseTable256[256] =
2ed95849 281{
abc490a1
MD
282#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
283#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
284#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
285 R6(0), R6(2), R6(1), R6(3)
286};
287#undef R2
288#undef R4
289#undef R6
2ed95849 290
abc490a1
MD
291static
292uint8_t bit_reverse_u8(uint8_t v)
293{
294 return BitReverseTable256[v];
295}
ab7d5fc6 296
abc490a1
MD
297static __attribute__((unused))
298uint32_t bit_reverse_u32(uint32_t v)
299{
300 return ((uint32_t) bit_reverse_u8(v) << 24) |
301 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
302 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
303 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
304}
305
abc490a1
MD
306static __attribute__((unused))
307uint64_t bit_reverse_u64(uint64_t v)
2ed95849 308{
abc490a1
MD
309 return ((uint64_t) bit_reverse_u8(v) << 56) |
310 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
311 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
312 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
313 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
314 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
315 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
316 ((uint64_t) bit_reverse_u8(v >> 56));
317}
318
319static
320unsigned long bit_reverse_ulong(unsigned long v)
321{
322#if (CAA_BITS_PER_LONG == 32)
323 return bit_reverse_u32(v);
324#else
325 return bit_reverse_u64(v);
326#endif
327}
328
f9830efd 329/*
24365af7
MD
330 * fls: returns the position of the most significant bit.
331 * Returns 0 if no bit is set, else returns the position of the most
332 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 333 */
24365af7
MD
334#if defined(__i386) || defined(__x86_64)
335static inline
336unsigned int fls_u32(uint32_t x)
f9830efd 337{
24365af7
MD
338 int r;
339
340 asm("bsrl %1,%0\n\t"
341 "jnz 1f\n\t"
342 "movl $-1,%0\n\t"
343 "1:\n\t"
344 : "=r" (r) : "rm" (x));
345 return r + 1;
346}
347#define HAS_FLS_U32
348#endif
349
350#if defined(__x86_64)
351static inline
352unsigned int fls_u64(uint64_t x)
353{
354 long r;
355
356 asm("bsrq %1,%0\n\t"
357 "jnz 1f\n\t"
358 "movq $-1,%0\n\t"
359 "1:\n\t"
360 : "=r" (r) : "rm" (x));
361 return r + 1;
362}
363#define HAS_FLS_U64
364#endif
365
366#ifndef HAS_FLS_U64
367static __attribute__((unused))
368unsigned int fls_u64(uint64_t x)
369{
370 unsigned int r = 64;
371
372 if (!x)
373 return 0;
374
375 if (!(x & 0xFFFFFFFF00000000ULL)) {
376 x <<= 32;
377 r -= 32;
378 }
379 if (!(x & 0xFFFF000000000000ULL)) {
380 x <<= 16;
381 r -= 16;
382 }
383 if (!(x & 0xFF00000000000000ULL)) {
384 x <<= 8;
385 r -= 8;
386 }
387 if (!(x & 0xF000000000000000ULL)) {
388 x <<= 4;
389 r -= 4;
390 }
391 if (!(x & 0xC000000000000000ULL)) {
392 x <<= 2;
393 r -= 2;
394 }
395 if (!(x & 0x8000000000000000ULL)) {
396 x <<= 1;
397 r -= 1;
398 }
399 return r;
400}
401#endif
402
403#ifndef HAS_FLS_U32
404static __attribute__((unused))
405unsigned int fls_u32(uint32_t x)
406{
407 unsigned int r = 32;
f9830efd 408
24365af7
MD
409 if (!x)
410 return 0;
411 if (!(x & 0xFFFF0000U)) {
412 x <<= 16;
413 r -= 16;
414 }
415 if (!(x & 0xFF000000U)) {
416 x <<= 8;
417 r -= 8;
418 }
419 if (!(x & 0xF0000000U)) {
420 x <<= 4;
421 r -= 4;
422 }
423 if (!(x & 0xC0000000U)) {
424 x <<= 2;
425 r -= 2;
426 }
427 if (!(x & 0x80000000U)) {
428 x <<= 1;
429 r -= 1;
430 }
431 return r;
432}
433#endif
434
435unsigned int fls_ulong(unsigned long x)
f9830efd 436{
24365af7
MD
437#if (CAA_BITS_PER_lONG == 32)
438 return fls_u32(x);
439#else
440 return fls_u64(x);
441#endif
442}
f9830efd 443
920f8ef6
LJ
444/*
445 * Return the minimum order for which x <= (1UL << order).
446 * Return -1 if x is 0.
447 */
24365af7
MD
448int get_count_order_u32(uint32_t x)
449{
920f8ef6
LJ
450 if (!x)
451 return -1;
24365af7 452
920f8ef6 453 return fls_u32(x - 1);
24365af7
MD
454}
455
920f8ef6
LJ
456/*
457 * Return the minimum order for which x <= (1UL << order).
458 * Return -1 if x is 0.
459 */
24365af7
MD
460int get_count_order_ulong(unsigned long x)
461{
920f8ef6
LJ
462 if (!x)
463 return -1;
24365af7 464
920f8ef6 465 return fls_ulong(x - 1);
f9830efd
MD
466}
467
98808fb1
MD
468#ifdef POISON_FREE
469#define poison_free(ptr) \
470 do { \
471 memset(ptr, 0x42, sizeof(*(ptr))); \
472 free(ptr); \
473 } while (0)
474#else
475#define poison_free(ptr) free(ptr)
476#endif
477
f9830efd 478static
4105056a 479void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 480
df44348d
MD
481/*
482 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
483 * available, then we support hash table item accounting.
484 * In the unfortunate event the number of CPUs reported would be
485 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
486 */
df44348d
MD
487#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
488
f8994aee 489static
4105056a 490void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
491 unsigned long count);
492
df44348d
MD
493static long nr_cpus_mask = -1;
494
495static
496struct ht_items_count *alloc_per_cpu_items_count(void)
497{
498 struct ht_items_count *count;
499
500 switch (nr_cpus_mask) {
501 case -2:
502 return NULL;
503 case -1:
504 {
505 long maxcpus;
506
507 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
508 if (maxcpus <= 0) {
509 nr_cpus_mask = -2;
510 return NULL;
511 }
512 /*
513 * round up number of CPUs to next power of two, so we
514 * can use & for modulo.
515 */
516 maxcpus = 1UL << get_count_order_ulong(maxcpus);
517 nr_cpus_mask = maxcpus - 1;
518 }
519 /* Fall-through */
520 default:
521 return calloc(nr_cpus_mask + 1, sizeof(*count));
522 }
523}
524
525static
526void free_per_cpu_items_count(struct ht_items_count *count)
527{
98808fb1 528 poison_free(count);
df44348d
MD
529}
530
531static
532int ht_get_cpu(void)
533{
534 int cpu;
535
536 assert(nr_cpus_mask >= 0);
537 cpu = sched_getcpu();
538 if (unlikely(cpu < 0))
539 return cpu;
540 else
541 return cpu & nr_cpus_mask;
542}
543
544static
4105056a 545void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 546{
3171717f 547 unsigned long percpu_count;
df44348d
MD
548 int cpu;
549
550 if (unlikely(!ht->percpu_count))
3171717f 551 return;
df44348d
MD
552 cpu = ht_get_cpu();
553 if (unlikely(cpu < 0))
3171717f
MD
554 return;
555 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 556 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 557 long count;
df44348d
MD
558
559 dbg_printf("add percpu %lu\n", percpu_count);
560 count = uatomic_add_return(&ht->count,
561 1UL << COUNT_COMMIT_ORDER);
562 /* If power of 2 */
563 if (!(count & (count - 1))) {
4105056a 564 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 565 return;
e3ecefd6 566 dbg_printf("add set global %ld\n", count);
4105056a 567 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 568 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
569 }
570 }
571}
572
573static
860d07e8 574void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
575{
576 unsigned long percpu_count;
3171717f 577 int cpu;
df44348d 578
3171717f
MD
579 if (unlikely(!ht->percpu_count))
580 return;
581 cpu = ht_get_cpu();
582 if (unlikely(cpu < 0))
583 return;
80d90c06 584 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 585 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 586 long count;
df44348d 587
860d07e8 588 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 589 count = uatomic_add_return(&ht->count,
3171717f 590 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
591 /* If power of 2 */
592 if (!(count & (count - 1))) {
4105056a 593 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 594 return;
e3ecefd6
MD
595 dbg_printf("del set global %ld\n", count);
596 /*
c941bb9e 597 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
598 * certain threshold.
599 */
c941bb9e 600 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 601 return;
4105056a 602 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 603 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
604 }
605 }
606}
607
608#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
609
91452a6a 610static const long nr_cpus_mask = -2;
df44348d
MD
611
612static
613struct ht_items_count *alloc_per_cpu_items_count(void)
614{
615 return NULL;
616}
617
618static
619void free_per_cpu_items_count(struct ht_items_count *count)
620{
621}
622
623static
4105056a 624void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
625{
626}
627
628static
860d07e8 629void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
630{
631}
632
633#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
634
635
f9830efd 636static
4105056a 637void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 638{
f8994aee
MD
639 unsigned long count;
640
b8af5011
MD
641 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
642 return;
f8994aee
MD
643 count = uatomic_read(&ht->count);
644 /*
645 * Use bucket-local length for small table expand and for
646 * environments lacking per-cpu data support.
647 */
648 if (count >= (1UL << COUNT_COMMIT_ORDER))
649 return;
24365af7 650 if (chain_len > 100)
f0c29ed7 651 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 652 chain_len);
3390d470 653 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 654 cds_lfht_resize_lazy(ht, size,
01370f0b 655 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
656}
657
abc490a1 658static
14044b37 659struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 660{
14044b37 661 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
662}
663
664static
14044b37 665int is_removed(struct cds_lfht_node *node)
abc490a1 666{
d37166c6 667 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
668}
669
670static
14044b37 671struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 672{
14044b37 673 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
674}
675
f5596c94 676static
14044b37 677int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
678{
679 return ((unsigned long) node) & DUMMY_FLAG;
680}
681
682static
14044b37 683struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 684{
14044b37 685 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 686}
bb7b2f26
MD
687
688static
689struct cds_lfht_node *get_end(void)
690{
691 return (struct cds_lfht_node *) END_VALUE;
692}
693
694static
695int is_end(struct cds_lfht_node *node)
696{
697 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
698}
699
abc490a1 700static
f9830efd 701unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
702{
703 unsigned long old1, old2;
704
705 old1 = uatomic_read(ptr);
706 do {
707 old2 = old1;
708 if (old2 >= v)
f9830efd 709 return old2;
abc490a1 710 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 711 return v;
abc490a1
MD
712}
713
f4a9cc0b
LJ
714static
715struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
716 unsigned long hash)
717{
718 unsigned long index, order;
719
720 assert(size > 0);
721 index = hash & (size - 1);
a4ea2223
LJ
722 /*
723 * equivalent to get_count_order_ulong(index + 1), but optimizes
724 * away the non-existing 0 special-case for
725 * get_count_order_ulong.
726 */
727 order = fls_ulong(index);
f4a9cc0b
LJ
728
729 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
730 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
731
732 return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
733}
734
273399de
MD
735/*
736 * Remove all logically deleted nodes from a bucket up to a certain node key.
737 */
738static
f9c80341 739void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 740{
14044b37 741 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 742
c90201ac
MD
743 assert(!is_dummy(dummy));
744 assert(!is_removed(dummy));
745 assert(!is_dummy(node));
746 assert(!is_removed(node));
273399de
MD
747 for (;;) {
748 iter_prev = dummy;
749 /* We can always skip the dummy node initially */
cc4fcb10
MD
750 iter = rcu_dereference(iter_prev->p.next);
751 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
752 /*
753 * We should never be called with dummy (start of chain)
754 * and logically removed node (end of path compression
755 * marker) being the actual same node. This would be a
756 * bug in the algorithm implementation.
757 */
758 assert(dummy != node);
273399de 759 for (;;) {
bb7b2f26 760 if (unlikely(is_end(iter)))
f9c80341 761 return;
76412f24 762 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 763 return;
cc4fcb10 764 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 765 if (likely(is_removed(next)))
273399de 766 break;
b453eae1 767 iter_prev = clear_flag(iter);
273399de
MD
768 iter = next;
769 }
b198f0fd 770 assert(!is_removed(iter));
f5596c94
MD
771 if (is_dummy(iter))
772 new_next = flag_dummy(clear_flag(next));
773 else
774 new_next = clear_flag(next);
48ed1c18
MD
775 if (is_removed(iter))
776 new_next = flag_removed(new_next);
f5596c94 777 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 778 }
f9c80341 779 return;
273399de
MD
780}
781
9357c415
MD
782static
783int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
784 struct cds_lfht_node *old_node,
3fb86f26 785 struct cds_lfht_node *old_next,
9357c415
MD
786 struct cds_lfht_node *new_node)
787{
3fb86f26 788 struct cds_lfht_node *dummy, *ret_next;
9357c415
MD
789 struct _cds_lfht_node *lookup;
790 int flagged = 0;
9357c415
MD
791
792 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
793 goto end;
794
795 assert(!is_removed(old_node));
796 assert(!is_dummy(old_node));
797 assert(!is_removed(new_node));
798 assert(!is_dummy(new_node));
799 assert(new_node != old_node);
3fb86f26 800 for (;;) {
9357c415 801 /* Insert after node to be replaced */
9357c415
MD
802 if (is_removed(old_next)) {
803 /*
804 * Too late, the old node has been removed under us
805 * between lookup and replace. Fail.
806 */
807 goto end;
808 }
809 assert(!is_dummy(old_next));
810 assert(new_node != clear_flag(old_next));
811 new_node->p.next = clear_flag(old_next);
812 /*
813 * Here is the whole trick for lock-free replace: we add
814 * the replacement node _after_ the node we want to
815 * replace by atomically setting its next pointer at the
816 * same time we set its removal flag. Given that
817 * the lookups/get next use an iterator aware of the
818 * next pointer, they will either skip the old node due
819 * to the removal flag and see the new node, or use
820 * the old node, but will not see the new one.
821 */
822 ret_next = uatomic_cmpxchg(&old_node->p.next,
823 old_next, flag_removed(new_node));
3fb86f26
LJ
824 if (ret_next == old_next)
825 break;
826 old_next = ret_next;
827 }
9357c415
MD
828
829 /* We performed the replacement. */
830 flagged = 1;
831
832 /*
833 * Ensure that the old node is not visible to readers anymore:
834 * lookup for the node, and remove it (along with any other
835 * logically removed node) if found.
836 */
f4a9cc0b 837 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
838 dummy = (struct cds_lfht_node *) lookup;
839 _cds_lfht_gc_bucket(dummy, new_node);
840end:
841 /*
842 * Only the flagging action indicated that we (and no other)
843 * replaced the node from the hash table.
844 */
845 if (flagged) {
846 assert(is_removed(rcu_dereference(old_node->p.next)));
847 return 0;
848 } else {
849 return -ENOENT;
850 }
851}
852
83beee94
MD
853/*
854 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
855 * mode. A NULL unique_ret allows creation of duplicate keys.
856 */
abc490a1 857static
83beee94
MD
858void _cds_lfht_add(struct cds_lfht *ht,
859 unsigned long size,
860 struct cds_lfht_node *node,
861 struct cds_lfht_iter *unique_ret,
862 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);
83beee94 872 assert(!unique_ret);
bb7b2f26 873 node->p.next = flag_dummy(get_end());
83beee94 874 return; /* Initial first add (head) */
18117871 875 }
f4a9cc0b 876 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 877 for (;;) {
adc0de68 878 uint32_t chain_len = 0;
abc490a1 879
11519af6
MD
880 /*
881 * iter_prev points to the non-removed node prior to the
882 * insert location.
11519af6 883 */
14044b37 884 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 885 /* We can always skip the dummy node initially */
cc4fcb10
MD
886 iter = rcu_dereference(iter_prev->p.next);
887 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 888 for (;;) {
bb7b2f26 889 if (unlikely(is_end(iter)))
273399de 890 goto insert;
76412f24 891 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 892 goto insert;
194fdbd1
LJ
893 /* dummy node is the first node of the identical-hash-value chain */
894 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
895 goto insert;
cc4fcb10 896 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 897 if (unlikely(is_removed(next)))
9dba85be 898 goto gc_node;
83beee94 899 if (unique_ret
1b81fe1a 900 && !is_dummy(next)
ae450da3 901 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash
e43f23f8
MD
902 && !ht->compare_fct(node->key, node->key_len,
903 clear_flag(iter)->key,
48ed1c18 904 clear_flag(iter)->key_len)) {
83beee94
MD
905 unique_ret->node = clear_flag(iter);
906 unique_ret->next = next;
907 return;
48ed1c18 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 933 } else {
83beee94 934 return_node = node;
960c9e4f 935 goto end;
48ed1c18
MD
936 }
937
9dba85be
MD
938 gc_node:
939 assert(!is_removed(iter));
f5596c94
MD
940 if (is_dummy(iter))
941 new_next = flag_dummy(clear_flag(next));
942 else
943 new_next = clear_flag(next);
944 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 945 /* retry */
464a1ec9 946 }
9357c415 947end:
83beee94
MD
948 if (unique_ret) {
949 unique_ret->node = return_node;
950 /* unique_ret->next left unset, never used. */
951 }
abc490a1 952}
464a1ec9 953
abc490a1 954static
860d07e8 955int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 956 struct cds_lfht_node *node,
b198f0fd 957 int dummy_removal)
abc490a1 958{
14044b37
MD
959 struct cds_lfht_node *dummy, *next, *old;
960 struct _cds_lfht_node *lookup;
abc490a1 961 int flagged = 0;
5e28c532 962
9357c415
MD
963 if (!node) /* Return -ENOENT if asked to delete NULL node */
964 goto end;
965
7ec59d3b 966 /* logically delete the node */
c90201ac
MD
967 assert(!is_dummy(node));
968 assert(!is_removed(node));
cc4fcb10 969 old = rcu_dereference(node->p.next);
7ec59d3b 970 do {
48ed1c18
MD
971 struct cds_lfht_node *new_next;
972
7ec59d3b 973 next = old;
76412f24 974 if (unlikely(is_removed(next)))
7ec59d3b 975 goto end;
1475579c
MD
976 if (dummy_removal)
977 assert(is_dummy(next));
978 else
979 assert(!is_dummy(next));
48ed1c18 980 new_next = flag_removed(next);
48ed1c18 981 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
982 } while (old != next);
983
984 /* We performed the (logical) deletion. */
985 flagged = 1;
986
987 /*
988 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
989 * the node, and remove it (along with any other logically removed node)
990 * if found.
11519af6 991 */
f4a9cc0b 992 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 993 dummy = (struct cds_lfht_node *) lookup;
f9c80341 994 _cds_lfht_gc_bucket(dummy, node);
2ed95849 995end:
11519af6
MD
996 /*
997 * Only the flagging action indicated that we (and no other)
998 * removed the node from the hash.
999 */
7ec59d3b 1000 if (flagged) {
cc4fcb10 1001 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 1002 return 0;
9357c415 1003 } else {
11519af6 1004 return -ENOENT;
9357c415 1005 }
abc490a1 1006}
2ed95849 1007
b7d619b0
MD
1008static
1009void *partition_resize_thread(void *arg)
1010{
1011 struct partition_resize_work *work = arg;
1012
1013 work->ht->cds_lfht_rcu_register_thread();
1014 work->fct(work->ht, work->i, work->start, work->len);
1015 work->ht->cds_lfht_rcu_unregister_thread();
1016 return NULL;
1017}
1018
1019static
1020void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1021 unsigned long len,
1022 void (*fct)(struct cds_lfht *ht, unsigned long i,
1023 unsigned long start, unsigned long len))
1024{
1025 unsigned long partition_len;
1026 struct partition_resize_work *work;
6083a889
MD
1027 int thread, ret;
1028 unsigned long nr_threads;
b7d619b0 1029
6083a889
MD
1030 /*
1031 * Note: nr_cpus_mask + 1 is always power of 2.
1032 * We spawn just the number of threads we need to satisfy the minimum
1033 * partition size, up to the number of CPUs in the system.
1034 */
91452a6a
MD
1035 if (nr_cpus_mask > 0) {
1036 nr_threads = min(nr_cpus_mask + 1,
1037 len >> MIN_PARTITION_PER_THREAD_ORDER);
1038 } else {
1039 nr_threads = 1;
1040 }
6083a889
MD
1041 partition_len = len >> get_count_order_ulong(nr_threads);
1042 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1043 assert(work);
6083a889
MD
1044 for (thread = 0; thread < nr_threads; thread++) {
1045 work[thread].ht = ht;
1046 work[thread].i = i;
1047 work[thread].len = partition_len;
1048 work[thread].start = thread * partition_len;
1049 work[thread].fct = fct;
1af6e26e 1050 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1051 partition_resize_thread, &work[thread]);
b7d619b0
MD
1052 assert(!ret);
1053 }
6083a889 1054 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1055 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1056 assert(!ret);
1057 }
1058 free(work);
b7d619b0
MD
1059}
1060
e8de508e
MD
1061/*
1062 * Holding RCU read lock to protect _cds_lfht_add against memory
1063 * reclaim that could be performed by other call_rcu worker threads (ABA
1064 * problem).
9ee0fc9a 1065 *
b7d619b0 1066 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1067 * many worker threads, based on the number of CPUs available in the system.
1068 * This should therefore take care of not having the expand lagging behind too
1069 * many concurrent insertion threads by using the scheduler's ability to
1070 * schedule dummy node population fairly with insertions.
e8de508e 1071 */
4105056a 1072static
b7d619b0
MD
1073void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1074 unsigned long start, unsigned long len)
4105056a
MD
1075{
1076 unsigned long j;
1077
1078 ht->cds_lfht_rcu_read_lock();
b7d619b0 1079 for (j = start; j < start + len; j++) {
4105056a
MD
1080 struct cds_lfht_node *new_node =
1081 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1082
dc1da8f6 1083 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1084 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1085 new_node->p.reverse_hash =
1086 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
83beee94
MD
1087 _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
1088 new_node, NULL, 1);
4105056a
MD
1089 }
1090 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1091}
1092
1093static
1094void init_table_populate(struct cds_lfht *ht, unsigned long i,
1095 unsigned long len)
1096{
1097 assert(nr_cpus_mask != -1);
6083a889 1098 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1099 ht->cds_lfht_rcu_thread_online();
1100 init_table_populate_partition(ht, i, 0, len);
1101 ht->cds_lfht_rcu_thread_offline();
1102 return;
1103 }
1104 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1105}
1106
abc490a1 1107static
4105056a 1108void init_table(struct cds_lfht *ht,
24365af7
MD
1109 unsigned long first_order, unsigned long len_order)
1110{
1111 unsigned long i, end_order;
1112
f0c29ed7 1113 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1114 first_order, first_order + len_order);
1115 end_order = first_order + len_order;
24365af7 1116 for (i = first_order; i < end_order; i++) {
4105056a 1117 unsigned long len;
24365af7
MD
1118
1119 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1120 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1121
1122 /* Stop expand if the resize target changes under us */
1123 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1124 break;
1125
0d14ceb2 1126 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1127 assert(ht->t.tbl[i]);
4105056a 1128
4105056a 1129 /*
dc1da8f6
MD
1130 * Set all dummy nodes reverse hash values for a level and
1131 * link all dummy nodes into the table.
4105056a 1132 */
dc1da8f6 1133 init_table_populate(ht, i, len);
4105056a 1134
f9c80341
MD
1135 /*
1136 * Update table size.
1137 */
1138 cmm_smp_wmb(); /* populate data before RCU size */
1139 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1140
4105056a
MD
1141 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1142 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1143 break;
1144 }
1145}
1146
e8de508e
MD
1147/*
1148 * Holding RCU read lock to protect _cds_lfht_remove against memory
1149 * reclaim that could be performed by other call_rcu worker threads (ABA
1150 * problem).
1151 * For a single level, we logically remove and garbage collect each node.
1152 *
1153 * As a design choice, we perform logical removal and garbage collection on a
1154 * node-per-node basis to simplify this algorithm. We also assume keeping good
1155 * cache locality of the operation would overweight possible performance gain
1156 * that could be achieved by batching garbage collection for multiple levels.
1157 * However, this would have to be justified by benchmarks.
1158 *
1159 * Concurrent removal and add operations are helping us perform garbage
1160 * collection of logically removed nodes. We guarantee that all logically
1161 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1162 * invoked to free a hole level of dummy nodes (after a grace period).
1163 *
1164 * Logical removal and garbage collection can therefore be done in batch or on a
1165 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1166 *
b7d619b0
MD
1167 * When we reach a certain length, we can split this removal over many worker
1168 * threads, based on the number of CPUs available in the system. This should
1169 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1170 * updater threads actively inserting into the hash table.
e8de508e 1171 */
4105056a 1172static
b7d619b0
MD
1173void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1174 unsigned long start, unsigned long len)
4105056a
MD
1175{
1176 unsigned long j;
1177
1178 ht->cds_lfht_rcu_read_lock();
b7d619b0 1179 for (j = start; j < start + len; j++) {
4105056a
MD
1180 struct cds_lfht_node *fini_node =
1181 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1182
1183 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1184 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1185 fini_node->p.reverse_hash =
1186 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1187 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1188 fini_node, 1);
abc490a1 1189 }
4105056a 1190 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1191}
1192
1193static
1194void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1195{
1196
1197 assert(nr_cpus_mask != -1);
6083a889 1198 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1199 ht->cds_lfht_rcu_thread_online();
1200 remove_table_partition(ht, i, 0, len);
1201 ht->cds_lfht_rcu_thread_offline();
1202 return;
1203 }
1204 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1205}
1206
1475579c 1207static
4105056a 1208void fini_table(struct cds_lfht *ht,
1475579c
MD
1209 unsigned long first_order, unsigned long len_order)
1210{
1211 long i, end_order;
0d14ceb2 1212 void *free_by_rcu = NULL;
1475579c
MD
1213
1214 dbg_printf("fini table: first_order %lu end_order %lu\n",
1215 first_order, first_order + len_order);
1216 end_order = first_order + len_order;
1217 assert(first_order > 0);
1475579c 1218 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1219 unsigned long len;
1475579c
MD
1220
1221 len = !i ? 1 : 1UL << (i - 1);
1222 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1223
4d676753
MD
1224 /* Stop shrink if the resize target changes under us */
1225 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1226 break;
1227
1228 cmm_smp_wmb(); /* populate data before RCU size */
1229 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1230
1231 /*
1232 * We need to wait for all add operations to reach Q.S. (and
1233 * thus use the new table for lookups) before we can start
1234 * releasing the old dummy nodes. Otherwise their lookup will
1235 * return a logically removed node as insert position.
1236 */
1237 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1238 if (free_by_rcu)
1239 free(free_by_rcu);
4d676753 1240
21263e21 1241 /*
4105056a
MD
1242 * Set "removed" flag in dummy nodes about to be removed.
1243 * Unlink all now-logically-removed dummy node pointers.
1244 * Concurrent add/remove operation are helping us doing
1245 * the gc.
21263e21 1246 */
4105056a
MD
1247 remove_table(ht, i, len);
1248
0d14ceb2 1249 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1250
1251 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1252 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1253 break;
1254 }
0d14ceb2
LJ
1255
1256 if (free_by_rcu) {
1257 ht->cds_lfht_synchronize_rcu();
1258 free(free_by_rcu);
1259 }
1475579c
MD
1260}
1261
7a9dcf9b 1262struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1263 cds_lfht_compare_fct compare_fct,
1264 unsigned long hash_seed,
1265 unsigned long init_size,
b8af5011 1266 int flags,
14044b37 1267 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1268 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1269 void (*cds_lfht_synchronize_rcu)(void),
1270 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1271 void (*cds_lfht_rcu_read_unlock)(void),
1272 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1273 void (*cds_lfht_rcu_thread_online)(void),
1274 void (*cds_lfht_rcu_register_thread)(void),
1275 void (*cds_lfht_rcu_unregister_thread)(void),
1276 pthread_attr_t *attr)
abc490a1 1277{
14044b37 1278 struct cds_lfht *ht;
24365af7 1279 unsigned long order;
abc490a1 1280
8129be4e 1281 /* init_size must be power of two */
49619ea0 1282 if (init_size && (init_size & (init_size - 1)))
8129be4e 1283 return NULL;
14044b37 1284 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1285 assert(ht);
abc490a1 1286 ht->hash_fct = hash_fct;
732ad076
MD
1287 ht->compare_fct = compare_fct;
1288 ht->hash_seed = hash_seed;
14044b37 1289 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1290 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1291 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1292 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1293 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1294 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1295 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1296 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1297 ht->resize_attr = attr;
df44348d 1298 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1299 /* this mutex should not nest in read-side C.S. */
1300 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1301 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1302 ht->flags = flags;
5f511391 1303 ht->cds_lfht_rcu_thread_offline();
f000907d 1304 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1305 ht->t.resize_target = 1UL << (order - 1);
4105056a 1306 init_table(ht, 0, order);
f000907d 1307 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1308 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1309 return ht;
1310}
1311
adc0de68
MD
1312void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1313 struct cds_lfht_iter *iter)
2ed95849 1314{
bb7b2f26 1315 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1316 struct _cds_lfht_node *lookup;
f4a9cc0b 1317 unsigned long hash, reverse_hash, size;
2ed95849 1318
732ad076 1319 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1320 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1321
4105056a 1322 size = rcu_dereference(ht->t.size);
f4a9cc0b 1323 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1324 dummy_node = (struct cds_lfht_node *) lookup;
1325 /* We can always skip the dummy node initially */
1326 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1327 node = clear_flag(node);
2ed95849 1328 for (;;) {
bb7b2f26 1329 if (unlikely(is_end(node))) {
96ad1112 1330 node = next = NULL;
abc490a1 1331 break;
bb7b2f26 1332 }
cc4fcb10 1333 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1334 node = next = NULL;
abc490a1 1335 break;
2ed95849 1336 }
1b81fe1a 1337 next = rcu_dereference(node->p.next);
adc0de68 1338 if (likely(!is_removed(next))
1b81fe1a 1339 && !is_dummy(next)
ae450da3 1340 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1341 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1342 break;
2ed95849 1343 }
1b81fe1a 1344 node = clear_flag(next);
2ed95849 1345 }
1b81fe1a 1346 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1347 iter->node = node;
1348 iter->next = next;
abc490a1 1349}
e0ba718a 1350
3883c0e5 1351void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1352{
adc0de68 1353 struct cds_lfht_node *node, *next;
a481e5ff
MD
1354 unsigned long reverse_hash;
1355 void *key;
1356 size_t key_len;
1357
adc0de68 1358 node = iter->node;
a481e5ff
MD
1359 reverse_hash = node->p.reverse_hash;
1360 key = node->key;
1361 key_len = node->key_len;
adc0de68 1362 next = iter->next;
a481e5ff
MD
1363 node = clear_flag(next);
1364
1365 for (;;) {
bb7b2f26 1366 if (unlikely(is_end(node))) {
96ad1112 1367 node = next = NULL;
a481e5ff 1368 break;
bb7b2f26 1369 }
a481e5ff 1370 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1371 node = next = NULL;
a481e5ff
MD
1372 break;
1373 }
1374 next = rcu_dereference(node->p.next);
adc0de68 1375 if (likely(!is_removed(next))
a481e5ff
MD
1376 && !is_dummy(next)
1377 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1378 break;
1379 }
1380 node = clear_flag(next);
1381 }
1382 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1383 iter->node = node;
1384 iter->next = next;
a481e5ff
MD
1385}
1386
4e9b9fbf
MD
1387void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1388{
1389 struct cds_lfht_node *node, *next;
1390
853395e1 1391 node = clear_flag(iter->next);
4e9b9fbf
MD
1392 for (;;) {
1393 if (unlikely(is_end(node))) {
1394 node = next = NULL;
1395 break;
1396 }
1397 next = rcu_dereference(node->p.next);
1398 if (likely(!is_removed(next))
1399 && !is_dummy(next)) {
1400 break;
1401 }
1402 node = clear_flag(next);
1403 }
1404 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1405 iter->node = node;
1406 iter->next = next;
1407}
1408
1409void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1410{
1411 struct _cds_lfht_node *lookup;
1412
1413 /*
1414 * Get next after first dummy node. The first dummy node is the
1415 * first node of the linked list.
1416 */
1417 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1418 iter->next = lookup->next;
4e9b9fbf
MD
1419 cds_lfht_next(ht, iter);
1420}
1421
14044b37 1422void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1423{
4105056a 1424 unsigned long hash, size;
ab7d5fc6 1425
49c2e2d6 1426 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1427 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1428
4105056a 1429 size = rcu_dereference(ht->t.size);
83beee94 1430 _cds_lfht_add(ht, size, node, NULL, 0);
4105056a 1431 ht_count_add(ht, size);
3eca1b8c
MD
1432}
1433
14044b37 1434struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1435 struct cds_lfht_node *node)
3eca1b8c 1436{
4105056a 1437 unsigned long hash, size;
83beee94 1438 struct cds_lfht_iter iter;
3eca1b8c 1439
49c2e2d6 1440 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1441 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1442
4105056a 1443 size = rcu_dereference(ht->t.size);
83beee94
MD
1444 _cds_lfht_add(ht, size, node, &iter, 0);
1445 if (iter.node == node)
4105056a 1446 ht_count_add(ht, size);
83beee94 1447 return iter.node;
2ed95849
MD
1448}
1449
9357c415 1450struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1451 struct cds_lfht_node *node)
1452{
1453 unsigned long hash, size;
83beee94 1454 struct cds_lfht_iter iter;
48ed1c18
MD
1455
1456 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1457 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1458
1459 size = rcu_dereference(ht->t.size);
83beee94
MD
1460 for (;;) {
1461 _cds_lfht_add(ht, size, node, &iter, 0);
1462 if (iter.node == node) {
1463 ht_count_add(ht, size);
1464 return NULL;
1465 }
1466
1467 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1468 return iter.node;
1469 }
48ed1c18
MD
1470}
1471
9357c415
MD
1472int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1473 struct cds_lfht_node *new_node)
1474{
1475 unsigned long size;
1476
1477 size = rcu_dereference(ht->t.size);
1478 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1479 new_node);
1480}
1481
1482int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1483{
4105056a 1484 unsigned long size;
df44348d 1485 int ret;
abc490a1 1486
4105056a 1487 size = rcu_dereference(ht->t.size);
9357c415 1488 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1489 if (!ret)
860d07e8 1490 ht_count_del(ht, size);
df44348d 1491 return ret;
2ed95849 1492}
ab7d5fc6 1493
abc490a1 1494static
14044b37 1495int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1496{
14044b37
MD
1497 struct cds_lfht_node *node;
1498 struct _cds_lfht_node *lookup;
4105056a 1499 unsigned long order, i, size;
674f7a69 1500
abc490a1 1501 /* Check that the table is empty */
4105056a 1502 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1503 node = (struct cds_lfht_node *) lookup;
abc490a1 1504 do {
1b81fe1a
MD
1505 node = clear_flag(node)->p.next;
1506 if (!is_dummy(node))
abc490a1 1507 return -EPERM;
273399de 1508 assert(!is_removed(node));
bb7b2f26 1509 } while (!is_end(node));
4105056a
MD
1510 /*
1511 * size accessed without rcu_dereference because hash table is
1512 * being destroyed.
1513 */
1514 size = ht->t.size;
abc490a1 1515 /* Internal sanity check: all nodes left should be dummy */
4105056a 1516 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1517 unsigned long len;
1518
1519 len = !order ? 1 : 1UL << (order - 1);
1520 for (i = 0; i < len; i++) {
f0c29ed7 1521 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1522 order, i,
4105056a
MD
1523 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1524 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1525 }
4105056a 1526 poison_free(ht->t.tbl[order]);
674f7a69 1527 }
abc490a1 1528 return 0;
674f7a69
MD
1529}
1530
1531/*
1532 * Should only be called when no more concurrent readers nor writers can
1533 * possibly access the table.
1534 */
b7d619b0 1535int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1536{
5e28c532
MD
1537 int ret;
1538
848d4088 1539 /* Wait for in-flight resize operations to complete */
24953e08
MD
1540 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1541 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1542 while (uatomic_read(&ht->in_progress_resize))
1543 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1544 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1545 if (ret)
1546 return ret;
df44348d 1547 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1548 if (attr)
1549 *attr = ht->resize_attr;
98808fb1 1550 poison_free(ht);
5e28c532 1551 return ret;
674f7a69
MD
1552}
1553
14044b37 1554void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1555 long *approx_before,
273399de 1556 unsigned long *count,
973e5e1b 1557 unsigned long *removed,
d933dd0e 1558 long *approx_after)
273399de 1559{
14044b37
MD
1560 struct cds_lfht_node *node, *next;
1561 struct _cds_lfht_node *lookup;
24365af7 1562 unsigned long nr_dummy = 0;
273399de 1563
7ed7682f 1564 *approx_before = 0;
973e5e1b
MD
1565 if (nr_cpus_mask >= 0) {
1566 int i;
1567
1568 for (i = 0; i < nr_cpus_mask + 1; i++) {
1569 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1570 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1571 }
1572 }
1573
273399de
MD
1574 *count = 0;
1575 *removed = 0;
1576
24365af7 1577 /* Count non-dummy nodes in the table */
4105056a 1578 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1579 node = (struct cds_lfht_node *) lookup;
273399de 1580 do {
cc4fcb10 1581 next = rcu_dereference(node->p.next);
b198f0fd 1582 if (is_removed(next)) {
973e5e1b
MD
1583 if (!is_dummy(next))
1584 (*removed)++;
1585 else
1586 (nr_dummy)++;
1b81fe1a 1587 } else if (!is_dummy(next))
273399de 1588 (*count)++;
24365af7
MD
1589 else
1590 (nr_dummy)++;
273399de 1591 node = clear_flag(next);
bb7b2f26 1592 } while (!is_end(node));
f0c29ed7 1593 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1594 *approx_after = 0;
973e5e1b
MD
1595 if (nr_cpus_mask >= 0) {
1596 int i;
1597
1598 for (i = 0; i < nr_cpus_mask + 1; i++) {
1599 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1600 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1601 }
1602 }
273399de
MD
1603}
1604
1475579c 1605/* called with resize mutex held */
abc490a1 1606static
4105056a 1607void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1608 unsigned long old_size, unsigned long new_size)
abc490a1 1609{
1475579c 1610 unsigned long old_order, new_order;
1475579c
MD
1611
1612 old_order = get_count_order_ulong(old_size) + 1;
1613 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1614 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1615 old_size, old_order, new_size, new_order);
1475579c 1616 assert(new_size > old_size);
4105056a 1617 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1618}
1619
1620/* called with resize mutex held */
1621static
4105056a 1622void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1623 unsigned long old_size, unsigned long new_size)
464a1ec9 1624{
1475579c 1625 unsigned long old_order, new_order;
464a1ec9 1626
cd95516d 1627 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1628 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1629 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1630 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1631 old_size, old_order, new_size, new_order);
1475579c 1632 assert(new_size < old_size);
1475579c 1633
4105056a
MD
1634 /* Remove and unlink all dummy nodes to remove. */
1635 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1636}
1637
1475579c
MD
1638
1639/* called with resize mutex held */
1640static
1641void _do_cds_lfht_resize(struct cds_lfht *ht)
1642{
1643 unsigned long new_size, old_size;
4105056a
MD
1644
1645 /*
1646 * Resize table, re-do if the target size has changed under us.
1647 */
1648 do {
d2be3620
MD
1649 assert(uatomic_read(&ht->in_progress_resize));
1650 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1651 break;
4105056a
MD
1652 ht->t.resize_initiated = 1;
1653 old_size = ht->t.size;
1654 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1655 if (old_size < new_size)
1656 _do_cds_lfht_grow(ht, old_size, new_size);
1657 else if (old_size > new_size)
1658 _do_cds_lfht_shrink(ht, old_size, new_size);
1659 ht->t.resize_initiated = 0;
1660 /* write resize_initiated before read resize_target */
1661 cmm_smp_mb();
4d676753 1662 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1663}
1664
abc490a1 1665static
4105056a 1666unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1667 int growth_order)
464a1ec9 1668{
4105056a
MD
1669 return _uatomic_max(&ht->t.resize_target,
1670 size << growth_order);
464a1ec9
MD
1671}
1672
1475579c 1673static
4105056a 1674void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1675 unsigned long count)
1475579c 1676{
cd95516d 1677 count = max(count, MIN_TABLE_SIZE);
4105056a 1678 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1679}
1680
1681void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1682{
4105056a
MD
1683 resize_target_update_count(ht, new_size);
1684 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1685 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1686 pthread_mutex_lock(&ht->resize_mutex);
1687 _do_cds_lfht_resize(ht);
1688 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1689 ht->cds_lfht_rcu_thread_online();
abc490a1 1690}
464a1ec9 1691
abc490a1
MD
1692static
1693void do_resize_cb(struct rcu_head *head)
1694{
1695 struct rcu_resize_work *work =
1696 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1697 struct cds_lfht *ht = work->ht;
abc490a1 1698
5f511391 1699 ht->cds_lfht_rcu_thread_offline();
abc490a1 1700 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1701 _do_cds_lfht_resize(ht);
abc490a1 1702 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1703 ht->cds_lfht_rcu_thread_online();
98808fb1 1704 poison_free(work);
848d4088
MD
1705 cmm_smp_mb(); /* finish resize before decrement */
1706 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1707}
1708
abc490a1 1709static
4105056a 1710void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1711{
abc490a1 1712 struct rcu_resize_work *work;
f9830efd 1713 unsigned long target_size;
abc490a1 1714
4105056a
MD
1715 target_size = resize_target_update(ht, size, growth);
1716 /* Store resize_target before read resize_initiated */
1717 cmm_smp_mb();
1718 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1719 uatomic_inc(&ht->in_progress_resize);
59290e9d 1720 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1721 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1722 uatomic_dec(&ht->in_progress_resize);
59290e9d 1723 return;
ed35e6d8 1724 }
f9830efd
MD
1725 work = malloc(sizeof(*work));
1726 work->ht = ht;
14044b37 1727 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1728 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1729 }
ab7d5fc6 1730}
3171717f 1731
f8994aee
MD
1732#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1733
3171717f 1734static
4105056a 1735void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1736 unsigned long count)
1737{
1738 struct rcu_resize_work *work;
3171717f 1739
b8af5011
MD
1740 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1741 return;
4105056a
MD
1742 resize_target_update_count(ht, count);
1743 /* Store resize_target before read resize_initiated */
1744 cmm_smp_mb();
1745 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1746 uatomic_inc(&ht->in_progress_resize);
59290e9d 1747 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1748 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1749 uatomic_dec(&ht->in_progress_resize);
59290e9d 1750 return;
ed35e6d8 1751 }
3171717f
MD
1752 work = malloc(sizeof(*work));
1753 work->ht = ht;
1754 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1755 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1756 }
1757}
f8994aee
MD
1758
1759#endif
This page took 0.121561 seconds and 4 git commands to generate.