Remove dead code in _cds_lfht_gc_bucket()
[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 750 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 751 assert(!is_removed(iter));
cc4fcb10 752 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
753 /*
754 * We should never be called with dummy (start of chain)
755 * and logically removed node (end of path compression
756 * marker) being the actual same node. This would be a
757 * bug in the algorithm implementation.
758 */
759 assert(dummy != node);
273399de 760 for (;;) {
bb7b2f26 761 if (unlikely(is_end(iter)))
f9c80341 762 return;
76412f24 763 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 764 return;
cc4fcb10 765 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 766 if (likely(is_removed(next)))
273399de 767 break;
b453eae1 768 iter_prev = clear_flag(iter);
273399de
MD
769 iter = next;
770 }
b198f0fd 771 assert(!is_removed(iter));
f5596c94
MD
772 if (is_dummy(iter))
773 new_next = flag_dummy(clear_flag(next));
774 else
775 new_next = clear_flag(next);
776 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 777 }
f9c80341 778 return;
273399de
MD
779}
780
9357c415
MD
781static
782int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
783 struct cds_lfht_node *old_node,
3fb86f26 784 struct cds_lfht_node *old_next,
9357c415
MD
785 struct cds_lfht_node *new_node)
786{
3fb86f26 787 struct cds_lfht_node *dummy, *ret_next;
9357c415
MD
788 struct _cds_lfht_node *lookup;
789 int flagged = 0;
9357c415
MD
790
791 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
792 goto end;
793
794 assert(!is_removed(old_node));
795 assert(!is_dummy(old_node));
796 assert(!is_removed(new_node));
797 assert(!is_dummy(new_node));
798 assert(new_node != old_node);
3fb86f26 799 for (;;) {
9357c415 800 /* Insert after node to be replaced */
9357c415
MD
801 if (is_removed(old_next)) {
802 /*
803 * Too late, the old node has been removed under us
804 * between lookup and replace. Fail.
805 */
806 goto end;
807 }
808 assert(!is_dummy(old_next));
809 assert(new_node != clear_flag(old_next));
810 new_node->p.next = clear_flag(old_next);
811 /*
812 * Here is the whole trick for lock-free replace: we add
813 * the replacement node _after_ the node we want to
814 * replace by atomically setting its next pointer at the
815 * same time we set its removal flag. Given that
816 * the lookups/get next use an iterator aware of the
817 * next pointer, they will either skip the old node due
818 * to the removal flag and see the new node, or use
819 * the old node, but will not see the new one.
820 */
821 ret_next = uatomic_cmpxchg(&old_node->p.next,
822 old_next, flag_removed(new_node));
3fb86f26
LJ
823 if (ret_next == old_next)
824 break;
825 old_next = ret_next;
826 }
9357c415
MD
827
828 /* We performed the replacement. */
829 flagged = 1;
830
831 /*
832 * Ensure that the old node is not visible to readers anymore:
833 * lookup for the node, and remove it (along with any other
834 * logically removed node) if found.
835 */
f4a9cc0b 836 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
837 dummy = (struct cds_lfht_node *) lookup;
838 _cds_lfht_gc_bucket(dummy, new_node);
839end:
840 /*
841 * Only the flagging action indicated that we (and no other)
842 * replaced the node from the hash table.
843 */
844 if (flagged) {
845 assert(is_removed(rcu_dereference(old_node->p.next)));
846 return 0;
847 } else {
848 return -ENOENT;
849 }
850}
851
83beee94
MD
852/*
853 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
854 * mode. A NULL unique_ret allows creation of duplicate keys.
855 */
abc490a1 856static
83beee94
MD
857void _cds_lfht_add(struct cds_lfht *ht,
858 unsigned long size,
859 struct cds_lfht_node *node,
860 struct cds_lfht_iter *unique_ret,
861 int dummy)
abc490a1 862{
14044b37 863 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 864 *return_node;
14044b37 865 struct _cds_lfht_node *lookup;
abc490a1 866
c90201ac
MD
867 assert(!is_dummy(node));
868 assert(!is_removed(node));
4105056a 869 if (!size) {
f5596c94 870 assert(dummy);
83beee94 871 assert(!unique_ret);
bb7b2f26 872 node->p.next = flag_dummy(get_end());
83beee94 873 return; /* Initial first add (head) */
18117871 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;
83beee94 898 if (unique_ret
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 903 clear_flag(iter)->key_len)) {
83beee94
MD
904 unique_ret->node = clear_flag(iter);
905 unique_ret->next = next;
906 return;
48ed1c18 907 }
11519af6 908 /* Only account for identical reverse hash once */
24365af7
MD
909 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
910 && !is_dummy(next))
4105056a 911 check_resize(ht, size, ++chain_len);
11519af6 912 iter_prev = clear_flag(iter);
273399de 913 iter = next;
abc490a1 914 }
48ed1c18 915
273399de 916 insert:
7ec59d3b 917 assert(node != clear_flag(iter));
11519af6 918 assert(!is_removed(iter_prev));
c90201ac 919 assert(!is_removed(iter));
f000907d 920 assert(iter_prev != node);
f9c80341 921 if (!dummy)
1b81fe1a 922 node->p.next = clear_flag(iter);
f9c80341
MD
923 else
924 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
925 if (is_dummy(iter))
926 new_node = flag_dummy(node);
927 else
928 new_node = node;
cc4fcb10 929 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 930 new_node) != iter) {
273399de 931 continue; /* retry */
48ed1c18 932 } else {
83beee94 933 return_node = node;
960c9e4f 934 goto end;
48ed1c18
MD
935 }
936
9dba85be
MD
937 gc_node:
938 assert(!is_removed(iter));
f5596c94
MD
939 if (is_dummy(iter))
940 new_next = flag_dummy(clear_flag(next));
941 else
942 new_next = clear_flag(next);
943 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 944 /* retry */
464a1ec9 945 }
9357c415 946end:
83beee94
MD
947 if (unique_ret) {
948 unique_ret->node = return_node;
949 /* unique_ret->next left unset, never used. */
950 }
abc490a1 951}
464a1ec9 952
abc490a1 953static
860d07e8 954int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 955 struct cds_lfht_node *node,
b198f0fd 956 int dummy_removal)
abc490a1 957{
14044b37
MD
958 struct cds_lfht_node *dummy, *next, *old;
959 struct _cds_lfht_node *lookup;
abc490a1 960 int flagged = 0;
5e28c532 961
9357c415
MD
962 if (!node) /* Return -ENOENT if asked to delete NULL node */
963 goto end;
964
7ec59d3b 965 /* logically delete the node */
c90201ac
MD
966 assert(!is_dummy(node));
967 assert(!is_removed(node));
cc4fcb10 968 old = rcu_dereference(node->p.next);
7ec59d3b 969 do {
48ed1c18
MD
970 struct cds_lfht_node *new_next;
971
7ec59d3b 972 next = old;
76412f24 973 if (unlikely(is_removed(next)))
7ec59d3b 974 goto end;
1475579c
MD
975 if (dummy_removal)
976 assert(is_dummy(next));
977 else
978 assert(!is_dummy(next));
48ed1c18 979 new_next = flag_removed(next);
48ed1c18 980 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
981 } while (old != next);
982
983 /* We performed the (logical) deletion. */
984 flagged = 1;
985
986 /*
987 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
988 * the node, and remove it (along with any other logically removed node)
989 * if found.
11519af6 990 */
f4a9cc0b 991 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 992 dummy = (struct cds_lfht_node *) lookup;
f9c80341 993 _cds_lfht_gc_bucket(dummy, node);
2ed95849 994end:
11519af6
MD
995 /*
996 * Only the flagging action indicated that we (and no other)
997 * removed the node from the hash.
998 */
7ec59d3b 999 if (flagged) {
cc4fcb10 1000 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 1001 return 0;
9357c415 1002 } else {
11519af6 1003 return -ENOENT;
9357c415 1004 }
abc490a1 1005}
2ed95849 1006
b7d619b0
MD
1007static
1008void *partition_resize_thread(void *arg)
1009{
1010 struct partition_resize_work *work = arg;
1011
1012 work->ht->cds_lfht_rcu_register_thread();
1013 work->fct(work->ht, work->i, work->start, work->len);
1014 work->ht->cds_lfht_rcu_unregister_thread();
1015 return NULL;
1016}
1017
1018static
1019void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1020 unsigned long len,
1021 void (*fct)(struct cds_lfht *ht, unsigned long i,
1022 unsigned long start, unsigned long len))
1023{
1024 unsigned long partition_len;
1025 struct partition_resize_work *work;
6083a889
MD
1026 int thread, ret;
1027 unsigned long nr_threads;
b7d619b0 1028
6083a889
MD
1029 /*
1030 * Note: nr_cpus_mask + 1 is always power of 2.
1031 * We spawn just the number of threads we need to satisfy the minimum
1032 * partition size, up to the number of CPUs in the system.
1033 */
91452a6a
MD
1034 if (nr_cpus_mask > 0) {
1035 nr_threads = min(nr_cpus_mask + 1,
1036 len >> MIN_PARTITION_PER_THREAD_ORDER);
1037 } else {
1038 nr_threads = 1;
1039 }
6083a889
MD
1040 partition_len = len >> get_count_order_ulong(nr_threads);
1041 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1042 assert(work);
6083a889
MD
1043 for (thread = 0; thread < nr_threads; thread++) {
1044 work[thread].ht = ht;
1045 work[thread].i = i;
1046 work[thread].len = partition_len;
1047 work[thread].start = thread * partition_len;
1048 work[thread].fct = fct;
1af6e26e 1049 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1050 partition_resize_thread, &work[thread]);
b7d619b0
MD
1051 assert(!ret);
1052 }
6083a889 1053 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1054 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1055 assert(!ret);
1056 }
1057 free(work);
b7d619b0
MD
1058}
1059
e8de508e
MD
1060/*
1061 * Holding RCU read lock to protect _cds_lfht_add against memory
1062 * reclaim that could be performed by other call_rcu worker threads (ABA
1063 * problem).
9ee0fc9a 1064 *
b7d619b0 1065 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1066 * many worker threads, based on the number of CPUs available in the system.
1067 * This should therefore take care of not having the expand lagging behind too
1068 * many concurrent insertion threads by using the scheduler's ability to
1069 * schedule dummy node population fairly with insertions.
e8de508e 1070 */
4105056a 1071static
b7d619b0
MD
1072void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1073 unsigned long start, unsigned long len)
4105056a
MD
1074{
1075 unsigned long j;
1076
1077 ht->cds_lfht_rcu_read_lock();
b7d619b0 1078 for (j = start; j < start + len; j++) {
4105056a
MD
1079 struct cds_lfht_node *new_node =
1080 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1081
dc1da8f6 1082 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1083 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1084 new_node->p.reverse_hash =
1085 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
83beee94
MD
1086 _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
1087 new_node, NULL, 1);
4105056a
MD
1088 }
1089 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1090}
1091
1092static
1093void init_table_populate(struct cds_lfht *ht, unsigned long i,
1094 unsigned long len)
1095{
1096 assert(nr_cpus_mask != -1);
6083a889 1097 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1098 ht->cds_lfht_rcu_thread_online();
1099 init_table_populate_partition(ht, i, 0, len);
1100 ht->cds_lfht_rcu_thread_offline();
1101 return;
1102 }
1103 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1104}
1105
abc490a1 1106static
4105056a 1107void init_table(struct cds_lfht *ht,
24365af7
MD
1108 unsigned long first_order, unsigned long len_order)
1109{
1110 unsigned long i, end_order;
1111
f0c29ed7 1112 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1113 first_order, first_order + len_order);
1114 end_order = first_order + len_order;
24365af7 1115 for (i = first_order; i < end_order; i++) {
4105056a 1116 unsigned long len;
24365af7
MD
1117
1118 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1119 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1120
1121 /* Stop expand if the resize target changes under us */
1122 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1123 break;
1124
0d14ceb2 1125 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1126 assert(ht->t.tbl[i]);
4105056a 1127
4105056a 1128 /*
dc1da8f6
MD
1129 * Set all dummy nodes reverse hash values for a level and
1130 * link all dummy nodes into the table.
4105056a 1131 */
dc1da8f6 1132 init_table_populate(ht, i, len);
4105056a 1133
f9c80341
MD
1134 /*
1135 * Update table size.
1136 */
1137 cmm_smp_wmb(); /* populate data before RCU size */
1138 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1139
4105056a
MD
1140 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1141 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1142 break;
1143 }
1144}
1145
e8de508e
MD
1146/*
1147 * Holding RCU read lock to protect _cds_lfht_remove against memory
1148 * reclaim that could be performed by other call_rcu worker threads (ABA
1149 * problem).
1150 * For a single level, we logically remove and garbage collect each node.
1151 *
1152 * As a design choice, we perform logical removal and garbage collection on a
1153 * node-per-node basis to simplify this algorithm. We also assume keeping good
1154 * cache locality of the operation would overweight possible performance gain
1155 * that could be achieved by batching garbage collection for multiple levels.
1156 * However, this would have to be justified by benchmarks.
1157 *
1158 * Concurrent removal and add operations are helping us perform garbage
1159 * collection of logically removed nodes. We guarantee that all logically
1160 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1161 * invoked to free a hole level of dummy nodes (after a grace period).
1162 *
1163 * Logical removal and garbage collection can therefore be done in batch or on a
1164 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1165 *
b7d619b0
MD
1166 * When we reach a certain length, we can split this removal over many worker
1167 * threads, based on the number of CPUs available in the system. This should
1168 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1169 * updater threads actively inserting into the hash table.
e8de508e 1170 */
4105056a 1171static
b7d619b0
MD
1172void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1173 unsigned long start, unsigned long len)
4105056a
MD
1174{
1175 unsigned long j;
1176
1177 ht->cds_lfht_rcu_read_lock();
b7d619b0 1178 for (j = start; j < start + len; j++) {
4105056a
MD
1179 struct cds_lfht_node *fini_node =
1180 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1181
1182 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1183 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1184 fini_node->p.reverse_hash =
1185 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1186 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1187 fini_node, 1);
abc490a1 1188 }
4105056a 1189 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1190}
1191
1192static
1193void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1194{
1195
1196 assert(nr_cpus_mask != -1);
6083a889 1197 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1198 ht->cds_lfht_rcu_thread_online();
1199 remove_table_partition(ht, i, 0, len);
1200 ht->cds_lfht_rcu_thread_offline();
1201 return;
1202 }
1203 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1204}
1205
1475579c 1206static
4105056a 1207void fini_table(struct cds_lfht *ht,
1475579c
MD
1208 unsigned long first_order, unsigned long len_order)
1209{
1210 long i, end_order;
0d14ceb2 1211 void *free_by_rcu = NULL;
1475579c
MD
1212
1213 dbg_printf("fini table: first_order %lu end_order %lu\n",
1214 first_order, first_order + len_order);
1215 end_order = first_order + len_order;
1216 assert(first_order > 0);
1475579c 1217 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1218 unsigned long len;
1475579c
MD
1219
1220 len = !i ? 1 : 1UL << (i - 1);
1221 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1222
4d676753
MD
1223 /* Stop shrink if the resize target changes under us */
1224 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1225 break;
1226
1227 cmm_smp_wmb(); /* populate data before RCU size */
1228 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1229
1230 /*
1231 * We need to wait for all add operations to reach Q.S. (and
1232 * thus use the new table for lookups) before we can start
1233 * releasing the old dummy nodes. Otherwise their lookup will
1234 * return a logically removed node as insert position.
1235 */
1236 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1237 if (free_by_rcu)
1238 free(free_by_rcu);
4d676753 1239
21263e21 1240 /*
4105056a
MD
1241 * Set "removed" flag in dummy nodes about to be removed.
1242 * Unlink all now-logically-removed dummy node pointers.
1243 * Concurrent add/remove operation are helping us doing
1244 * the gc.
21263e21 1245 */
4105056a
MD
1246 remove_table(ht, i, len);
1247
0d14ceb2 1248 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1249
1250 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1251 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1252 break;
1253 }
0d14ceb2
LJ
1254
1255 if (free_by_rcu) {
1256 ht->cds_lfht_synchronize_rcu();
1257 free(free_by_rcu);
1258 }
1475579c
MD
1259}
1260
7a9dcf9b 1261struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1262 cds_lfht_compare_fct compare_fct,
1263 unsigned long hash_seed,
1264 unsigned long init_size,
b8af5011 1265 int flags,
14044b37 1266 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1267 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1268 void (*cds_lfht_synchronize_rcu)(void),
1269 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1270 void (*cds_lfht_rcu_read_unlock)(void),
1271 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1272 void (*cds_lfht_rcu_thread_online)(void),
1273 void (*cds_lfht_rcu_register_thread)(void),
1274 void (*cds_lfht_rcu_unregister_thread)(void),
1275 pthread_attr_t *attr)
abc490a1 1276{
14044b37 1277 struct cds_lfht *ht;
24365af7 1278 unsigned long order;
abc490a1 1279
8129be4e 1280 /* init_size must be power of two */
49619ea0 1281 if (init_size && (init_size & (init_size - 1)))
8129be4e 1282 return NULL;
14044b37 1283 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1284 assert(ht);
abc490a1 1285 ht->hash_fct = hash_fct;
732ad076
MD
1286 ht->compare_fct = compare_fct;
1287 ht->hash_seed = hash_seed;
14044b37 1288 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1289 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1290 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1291 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1292 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1293 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1294 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1295 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1296 ht->resize_attr = attr;
df44348d 1297 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1298 /* this mutex should not nest in read-side C.S. */
1299 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1300 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1301 ht->flags = flags;
5f511391 1302 ht->cds_lfht_rcu_thread_offline();
f000907d 1303 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1304 ht->t.resize_target = 1UL << (order - 1);
4105056a 1305 init_table(ht, 0, order);
f000907d 1306 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1307 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1308 return ht;
1309}
1310
adc0de68
MD
1311void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1312 struct cds_lfht_iter *iter)
2ed95849 1313{
bb7b2f26 1314 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1315 struct _cds_lfht_node *lookup;
f4a9cc0b 1316 unsigned long hash, reverse_hash, size;
2ed95849 1317
732ad076 1318 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1319 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1320
4105056a 1321 size = rcu_dereference(ht->t.size);
f4a9cc0b 1322 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1323 dummy_node = (struct cds_lfht_node *) lookup;
1324 /* We can always skip the dummy node initially */
1325 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1326 node = clear_flag(node);
2ed95849 1327 for (;;) {
bb7b2f26 1328 if (unlikely(is_end(node))) {
96ad1112 1329 node = next = NULL;
abc490a1 1330 break;
bb7b2f26 1331 }
cc4fcb10 1332 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1333 node = next = NULL;
abc490a1 1334 break;
2ed95849 1335 }
1b81fe1a 1336 next = rcu_dereference(node->p.next);
adc0de68 1337 if (likely(!is_removed(next))
1b81fe1a 1338 && !is_dummy(next)
ae450da3 1339 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1340 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1341 break;
2ed95849 1342 }
1b81fe1a 1343 node = clear_flag(next);
2ed95849 1344 }
1b81fe1a 1345 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1346 iter->node = node;
1347 iter->next = next;
abc490a1 1348}
e0ba718a 1349
3883c0e5 1350void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1351{
adc0de68 1352 struct cds_lfht_node *node, *next;
a481e5ff
MD
1353 unsigned long reverse_hash;
1354 void *key;
1355 size_t key_len;
1356
adc0de68 1357 node = iter->node;
a481e5ff
MD
1358 reverse_hash = node->p.reverse_hash;
1359 key = node->key;
1360 key_len = node->key_len;
adc0de68 1361 next = iter->next;
a481e5ff
MD
1362 node = clear_flag(next);
1363
1364 for (;;) {
bb7b2f26 1365 if (unlikely(is_end(node))) {
96ad1112 1366 node = next = NULL;
a481e5ff 1367 break;
bb7b2f26 1368 }
a481e5ff 1369 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1370 node = next = NULL;
a481e5ff
MD
1371 break;
1372 }
1373 next = rcu_dereference(node->p.next);
adc0de68 1374 if (likely(!is_removed(next))
a481e5ff
MD
1375 && !is_dummy(next)
1376 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1377 break;
1378 }
1379 node = clear_flag(next);
1380 }
1381 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1382 iter->node = node;
1383 iter->next = next;
a481e5ff
MD
1384}
1385
4e9b9fbf
MD
1386void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1387{
1388 struct cds_lfht_node *node, *next;
1389
853395e1 1390 node = clear_flag(iter->next);
4e9b9fbf
MD
1391 for (;;) {
1392 if (unlikely(is_end(node))) {
1393 node = next = NULL;
1394 break;
1395 }
1396 next = rcu_dereference(node->p.next);
1397 if (likely(!is_removed(next))
1398 && !is_dummy(next)) {
1399 break;
1400 }
1401 node = clear_flag(next);
1402 }
1403 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1404 iter->node = node;
1405 iter->next = next;
1406}
1407
1408void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1409{
1410 struct _cds_lfht_node *lookup;
1411
1412 /*
1413 * Get next after first dummy node. The first dummy node is the
1414 * first node of the linked list.
1415 */
1416 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1417 iter->next = lookup->next;
4e9b9fbf
MD
1418 cds_lfht_next(ht, iter);
1419}
1420
14044b37 1421void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1422{
4105056a 1423 unsigned long hash, size;
ab7d5fc6 1424
49c2e2d6 1425 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1426 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1427
4105056a 1428 size = rcu_dereference(ht->t.size);
83beee94 1429 _cds_lfht_add(ht, size, node, NULL, 0);
4105056a 1430 ht_count_add(ht, size);
3eca1b8c
MD
1431}
1432
14044b37 1433struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1434 struct cds_lfht_node *node)
3eca1b8c 1435{
4105056a 1436 unsigned long hash, size;
83beee94 1437 struct cds_lfht_iter iter;
3eca1b8c 1438
49c2e2d6 1439 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1440 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1441
4105056a 1442 size = rcu_dereference(ht->t.size);
83beee94
MD
1443 _cds_lfht_add(ht, size, node, &iter, 0);
1444 if (iter.node == node)
4105056a 1445 ht_count_add(ht, size);
83beee94 1446 return iter.node;
2ed95849
MD
1447}
1448
9357c415 1449struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1450 struct cds_lfht_node *node)
1451{
1452 unsigned long hash, size;
83beee94 1453 struct cds_lfht_iter iter;
48ed1c18
MD
1454
1455 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1456 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1457
1458 size = rcu_dereference(ht->t.size);
83beee94
MD
1459 for (;;) {
1460 _cds_lfht_add(ht, size, node, &iter, 0);
1461 if (iter.node == node) {
1462 ht_count_add(ht, size);
1463 return NULL;
1464 }
1465
1466 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1467 return iter.node;
1468 }
48ed1c18
MD
1469}
1470
9357c415
MD
1471int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1472 struct cds_lfht_node *new_node)
1473{
1474 unsigned long size;
1475
1476 size = rcu_dereference(ht->t.size);
1477 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1478 new_node);
1479}
1480
1481int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1482{
4105056a 1483 unsigned long size;
df44348d 1484 int ret;
abc490a1 1485
4105056a 1486 size = rcu_dereference(ht->t.size);
9357c415 1487 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1488 if (!ret)
860d07e8 1489 ht_count_del(ht, size);
df44348d 1490 return ret;
2ed95849 1491}
ab7d5fc6 1492
abc490a1 1493static
14044b37 1494int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1495{
14044b37
MD
1496 struct cds_lfht_node *node;
1497 struct _cds_lfht_node *lookup;
4105056a 1498 unsigned long order, i, size;
674f7a69 1499
abc490a1 1500 /* Check that the table is empty */
4105056a 1501 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1502 node = (struct cds_lfht_node *) lookup;
abc490a1 1503 do {
1b81fe1a
MD
1504 node = clear_flag(node)->p.next;
1505 if (!is_dummy(node))
abc490a1 1506 return -EPERM;
273399de 1507 assert(!is_removed(node));
bb7b2f26 1508 } while (!is_end(node));
4105056a
MD
1509 /*
1510 * size accessed without rcu_dereference because hash table is
1511 * being destroyed.
1512 */
1513 size = ht->t.size;
abc490a1 1514 /* Internal sanity check: all nodes left should be dummy */
4105056a 1515 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1516 unsigned long len;
1517
1518 len = !order ? 1 : 1UL << (order - 1);
1519 for (i = 0; i < len; i++) {
f0c29ed7 1520 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1521 order, i,
4105056a
MD
1522 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1523 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1524 }
4105056a 1525 poison_free(ht->t.tbl[order]);
674f7a69 1526 }
abc490a1 1527 return 0;
674f7a69
MD
1528}
1529
1530/*
1531 * Should only be called when no more concurrent readers nor writers can
1532 * possibly access the table.
1533 */
b7d619b0 1534int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1535{
5e28c532
MD
1536 int ret;
1537
848d4088 1538 /* Wait for in-flight resize operations to complete */
24953e08
MD
1539 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1540 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1541 while (uatomic_read(&ht->in_progress_resize))
1542 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1543 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1544 if (ret)
1545 return ret;
df44348d 1546 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1547 if (attr)
1548 *attr = ht->resize_attr;
98808fb1 1549 poison_free(ht);
5e28c532 1550 return ret;
674f7a69
MD
1551}
1552
14044b37 1553void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1554 long *approx_before,
273399de 1555 unsigned long *count,
973e5e1b 1556 unsigned long *removed,
d933dd0e 1557 long *approx_after)
273399de 1558{
14044b37
MD
1559 struct cds_lfht_node *node, *next;
1560 struct _cds_lfht_node *lookup;
24365af7 1561 unsigned long nr_dummy = 0;
273399de 1562
7ed7682f 1563 *approx_before = 0;
973e5e1b
MD
1564 if (nr_cpus_mask >= 0) {
1565 int i;
1566
1567 for (i = 0; i < nr_cpus_mask + 1; i++) {
1568 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1569 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1570 }
1571 }
1572
273399de
MD
1573 *count = 0;
1574 *removed = 0;
1575
24365af7 1576 /* Count non-dummy nodes in the table */
4105056a 1577 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1578 node = (struct cds_lfht_node *) lookup;
273399de 1579 do {
cc4fcb10 1580 next = rcu_dereference(node->p.next);
b198f0fd 1581 if (is_removed(next)) {
973e5e1b
MD
1582 if (!is_dummy(next))
1583 (*removed)++;
1584 else
1585 (nr_dummy)++;
1b81fe1a 1586 } else if (!is_dummy(next))
273399de 1587 (*count)++;
24365af7
MD
1588 else
1589 (nr_dummy)++;
273399de 1590 node = clear_flag(next);
bb7b2f26 1591 } while (!is_end(node));
f0c29ed7 1592 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1593 *approx_after = 0;
973e5e1b
MD
1594 if (nr_cpus_mask >= 0) {
1595 int i;
1596
1597 for (i = 0; i < nr_cpus_mask + 1; i++) {
1598 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1599 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1600 }
1601 }
273399de
MD
1602}
1603
1475579c 1604/* called with resize mutex held */
abc490a1 1605static
4105056a 1606void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1607 unsigned long old_size, unsigned long new_size)
abc490a1 1608{
1475579c 1609 unsigned long old_order, new_order;
1475579c
MD
1610
1611 old_order = get_count_order_ulong(old_size) + 1;
1612 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1613 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1614 old_size, old_order, new_size, new_order);
1475579c 1615 assert(new_size > old_size);
4105056a 1616 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1617}
1618
1619/* called with resize mutex held */
1620static
4105056a 1621void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1622 unsigned long old_size, unsigned long new_size)
464a1ec9 1623{
1475579c 1624 unsigned long old_order, new_order;
464a1ec9 1625
cd95516d 1626 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1627 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1628 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1629 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1630 old_size, old_order, new_size, new_order);
1475579c 1631 assert(new_size < old_size);
1475579c 1632
4105056a
MD
1633 /* Remove and unlink all dummy nodes to remove. */
1634 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1635}
1636
1475579c
MD
1637
1638/* called with resize mutex held */
1639static
1640void _do_cds_lfht_resize(struct cds_lfht *ht)
1641{
1642 unsigned long new_size, old_size;
4105056a
MD
1643
1644 /*
1645 * Resize table, re-do if the target size has changed under us.
1646 */
1647 do {
d2be3620
MD
1648 assert(uatomic_read(&ht->in_progress_resize));
1649 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1650 break;
4105056a
MD
1651 ht->t.resize_initiated = 1;
1652 old_size = ht->t.size;
1653 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1654 if (old_size < new_size)
1655 _do_cds_lfht_grow(ht, old_size, new_size);
1656 else if (old_size > new_size)
1657 _do_cds_lfht_shrink(ht, old_size, new_size);
1658 ht->t.resize_initiated = 0;
1659 /* write resize_initiated before read resize_target */
1660 cmm_smp_mb();
4d676753 1661 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1662}
1663
abc490a1 1664static
4105056a 1665unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1666 int growth_order)
464a1ec9 1667{
4105056a
MD
1668 return _uatomic_max(&ht->t.resize_target,
1669 size << growth_order);
464a1ec9
MD
1670}
1671
1475579c 1672static
4105056a 1673void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1674 unsigned long count)
1475579c 1675{
cd95516d 1676 count = max(count, MIN_TABLE_SIZE);
4105056a 1677 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1678}
1679
1680void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1681{
4105056a
MD
1682 resize_target_update_count(ht, new_size);
1683 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1684 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1685 pthread_mutex_lock(&ht->resize_mutex);
1686 _do_cds_lfht_resize(ht);
1687 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1688 ht->cds_lfht_rcu_thread_online();
abc490a1 1689}
464a1ec9 1690
abc490a1
MD
1691static
1692void do_resize_cb(struct rcu_head *head)
1693{
1694 struct rcu_resize_work *work =
1695 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1696 struct cds_lfht *ht = work->ht;
abc490a1 1697
5f511391 1698 ht->cds_lfht_rcu_thread_offline();
abc490a1 1699 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1700 _do_cds_lfht_resize(ht);
abc490a1 1701 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1702 ht->cds_lfht_rcu_thread_online();
98808fb1 1703 poison_free(work);
848d4088
MD
1704 cmm_smp_mb(); /* finish resize before decrement */
1705 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1706}
1707
abc490a1 1708static
4105056a 1709void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1710{
abc490a1 1711 struct rcu_resize_work *work;
f9830efd 1712 unsigned long target_size;
abc490a1 1713
4105056a
MD
1714 target_size = resize_target_update(ht, size, growth);
1715 /* Store resize_target before read resize_initiated */
1716 cmm_smp_mb();
1717 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1718 uatomic_inc(&ht->in_progress_resize);
59290e9d 1719 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1720 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1721 uatomic_dec(&ht->in_progress_resize);
59290e9d 1722 return;
ed35e6d8 1723 }
f9830efd
MD
1724 work = malloc(sizeof(*work));
1725 work->ht = ht;
14044b37 1726 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1727 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1728 }
ab7d5fc6 1729}
3171717f 1730
f8994aee
MD
1731#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1732
3171717f 1733static
4105056a 1734void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1735 unsigned long count)
1736{
1737 struct rcu_resize_work *work;
3171717f 1738
b8af5011
MD
1739 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1740 return;
4105056a
MD
1741 resize_target_update_count(ht, count);
1742 /* Store resize_target before read resize_initiated */
1743 cmm_smp_mb();
1744 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1745 uatomic_inc(&ht->in_progress_resize);
59290e9d 1746 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1747 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1748 uatomic_dec(&ht->in_progress_resize);
59290e9d 1749 return;
ed35e6d8 1750 }
3171717f
MD
1751 work = malloc(sizeof(*work));
1752 work->ht = ht;
1753 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1754 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1755 }
1756}
f8994aee
MD
1757
1758#endif
This page took 0.202044 seconds and 4 git commands to generate.