rculfhash: use dbg_printf() for grow/shrink printout
[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
MD
211struct rcu_level {
212 struct rcu_head head;
213 struct _cds_lfht_node nodes[0];
214};
215
395270b6 216struct rcu_table {
4105056a 217 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 218 unsigned long resize_target;
11519af6 219 int resize_initiated;
4105056a 220 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
221};
222
14044b37 223struct cds_lfht {
4105056a 224 struct rcu_table t;
14044b37
MD
225 cds_lfht_hash_fct hash_fct;
226 cds_lfht_compare_fct compare_fct;
732ad076 227 unsigned long hash_seed;
b8af5011 228 int flags;
5f511391
MD
229 /*
230 * We need to put the work threads offline (QSBR) when taking this
231 * mutex, because we use synchronize_rcu within this mutex critical
232 * section, which waits on read-side critical sections, and could
233 * therefore cause grace-period deadlock if we hold off RCU G.P.
234 * completion.
235 */
464a1ec9 236 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 237 unsigned int in_progress_resize, in_progress_destroy;
14044b37 238 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 239 void (*func)(struct rcu_head *head));
1475579c 240 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
241 void (*cds_lfht_rcu_read_lock)(void);
242 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
243 void (*cds_lfht_rcu_thread_offline)(void);
244 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
245 void (*cds_lfht_rcu_register_thread)(void);
246 void (*cds_lfht_rcu_unregister_thread)(void);
247 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 248 long count; /* global approximate item count */
df44348d 249 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
250};
251
abc490a1
MD
252struct rcu_resize_work {
253 struct rcu_head head;
14044b37 254 struct cds_lfht *ht;
abc490a1 255};
2ed95849 256
b7d619b0 257struct partition_resize_work {
1af6e26e 258 pthread_t thread_id;
b7d619b0
MD
259 struct cds_lfht *ht;
260 unsigned long i, start, len;
261 void (*fct)(struct cds_lfht *ht, unsigned long i,
262 unsigned long start, unsigned long len);
263};
264
48ed1c18
MD
265enum add_mode {
266 ADD_DEFAULT = 0,
267 ADD_UNIQUE = 1,
268 ADD_REPLACE = 2,
269};
270
76a73da8
MD
271static
272struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
273 unsigned long size,
274 struct cds_lfht_node *node,
48ed1c18
MD
275 enum add_mode mode, int dummy);
276
abc490a1
MD
277/*
278 * Algorithm to reverse bits in a word by lookup table, extended to
279 * 64-bit words.
f9830efd 280 * Source:
abc490a1 281 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 282 * Originally from Public Domain.
abc490a1
MD
283 */
284
285static const uint8_t BitReverseTable256[256] =
2ed95849 286{
abc490a1
MD
287#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
288#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
289#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
290 R6(0), R6(2), R6(1), R6(3)
291};
292#undef R2
293#undef R4
294#undef R6
2ed95849 295
abc490a1
MD
296static
297uint8_t bit_reverse_u8(uint8_t v)
298{
299 return BitReverseTable256[v];
300}
ab7d5fc6 301
abc490a1
MD
302static __attribute__((unused))
303uint32_t bit_reverse_u32(uint32_t v)
304{
305 return ((uint32_t) bit_reverse_u8(v) << 24) |
306 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
307 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
308 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
309}
310
abc490a1
MD
311static __attribute__((unused))
312uint64_t bit_reverse_u64(uint64_t v)
2ed95849 313{
abc490a1
MD
314 return ((uint64_t) bit_reverse_u8(v) << 56) |
315 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
316 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
317 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
318 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
319 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
320 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
321 ((uint64_t) bit_reverse_u8(v >> 56));
322}
323
324static
325unsigned long bit_reverse_ulong(unsigned long v)
326{
327#if (CAA_BITS_PER_LONG == 32)
328 return bit_reverse_u32(v);
329#else
330 return bit_reverse_u64(v);
331#endif
332}
333
f9830efd 334/*
24365af7
MD
335 * fls: returns the position of the most significant bit.
336 * Returns 0 if no bit is set, else returns the position of the most
337 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 338 */
24365af7
MD
339#if defined(__i386) || defined(__x86_64)
340static inline
341unsigned int fls_u32(uint32_t x)
f9830efd 342{
24365af7
MD
343 int r;
344
345 asm("bsrl %1,%0\n\t"
346 "jnz 1f\n\t"
347 "movl $-1,%0\n\t"
348 "1:\n\t"
349 : "=r" (r) : "rm" (x));
350 return r + 1;
351}
352#define HAS_FLS_U32
353#endif
354
355#if defined(__x86_64)
356static inline
357unsigned int fls_u64(uint64_t x)
358{
359 long r;
360
361 asm("bsrq %1,%0\n\t"
362 "jnz 1f\n\t"
363 "movq $-1,%0\n\t"
364 "1:\n\t"
365 : "=r" (r) : "rm" (x));
366 return r + 1;
367}
368#define HAS_FLS_U64
369#endif
370
371#ifndef HAS_FLS_U64
372static __attribute__((unused))
373unsigned int fls_u64(uint64_t x)
374{
375 unsigned int r = 64;
376
377 if (!x)
378 return 0;
379
380 if (!(x & 0xFFFFFFFF00000000ULL)) {
381 x <<= 32;
382 r -= 32;
383 }
384 if (!(x & 0xFFFF000000000000ULL)) {
385 x <<= 16;
386 r -= 16;
387 }
388 if (!(x & 0xFF00000000000000ULL)) {
389 x <<= 8;
390 r -= 8;
391 }
392 if (!(x & 0xF000000000000000ULL)) {
393 x <<= 4;
394 r -= 4;
395 }
396 if (!(x & 0xC000000000000000ULL)) {
397 x <<= 2;
398 r -= 2;
399 }
400 if (!(x & 0x8000000000000000ULL)) {
401 x <<= 1;
402 r -= 1;
403 }
404 return r;
405}
406#endif
407
408#ifndef HAS_FLS_U32
409static __attribute__((unused))
410unsigned int fls_u32(uint32_t x)
411{
412 unsigned int r = 32;
f9830efd 413
24365af7
MD
414 if (!x)
415 return 0;
416 if (!(x & 0xFFFF0000U)) {
417 x <<= 16;
418 r -= 16;
419 }
420 if (!(x & 0xFF000000U)) {
421 x <<= 8;
422 r -= 8;
423 }
424 if (!(x & 0xF0000000U)) {
425 x <<= 4;
426 r -= 4;
427 }
428 if (!(x & 0xC0000000U)) {
429 x <<= 2;
430 r -= 2;
431 }
432 if (!(x & 0x80000000U)) {
433 x <<= 1;
434 r -= 1;
435 }
436 return r;
437}
438#endif
439
440unsigned int fls_ulong(unsigned long x)
f9830efd 441{
24365af7
MD
442#if (CAA_BITS_PER_lONG == 32)
443 return fls_u32(x);
444#else
445 return fls_u64(x);
446#endif
447}
f9830efd 448
24365af7
MD
449int get_count_order_u32(uint32_t x)
450{
451 int order;
452
453 order = fls_u32(x) - 1;
454 if (x & (x - 1))
455 order++;
456 return order;
457}
458
459int get_count_order_ulong(unsigned long x)
460{
461 int order;
462
463 order = fls_ulong(x) - 1;
464 if (x & (x - 1))
465 order++;
466 return order;
f9830efd
MD
467}
468
98808fb1
MD
469#ifdef POISON_FREE
470#define poison_free(ptr) \
471 do { \
472 memset(ptr, 0x42, sizeof(*(ptr))); \
473 free(ptr); \
474 } while (0)
475#else
476#define poison_free(ptr) free(ptr)
477#endif
478
f9830efd 479static
4105056a 480void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 481
df44348d
MD
482/*
483 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
484 * available, then we support hash table item accounting.
485 * In the unfortunate event the number of CPUs reported would be
486 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
487 */
df44348d
MD
488#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
489
f8994aee 490static
4105056a 491void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
492 unsigned long count);
493
df44348d
MD
494static long nr_cpus_mask = -1;
495
496static
497struct ht_items_count *alloc_per_cpu_items_count(void)
498{
499 struct ht_items_count *count;
500
501 switch (nr_cpus_mask) {
502 case -2:
503 return NULL;
504 case -1:
505 {
506 long maxcpus;
507
508 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
509 if (maxcpus <= 0) {
510 nr_cpus_mask = -2;
511 return NULL;
512 }
513 /*
514 * round up number of CPUs to next power of two, so we
515 * can use & for modulo.
516 */
517 maxcpus = 1UL << get_count_order_ulong(maxcpus);
518 nr_cpus_mask = maxcpus - 1;
519 }
520 /* Fall-through */
521 default:
522 return calloc(nr_cpus_mask + 1, sizeof(*count));
523 }
524}
525
526static
527void free_per_cpu_items_count(struct ht_items_count *count)
528{
98808fb1 529 poison_free(count);
df44348d
MD
530}
531
532static
533int ht_get_cpu(void)
534{
535 int cpu;
536
537 assert(nr_cpus_mask >= 0);
538 cpu = sched_getcpu();
539 if (unlikely(cpu < 0))
540 return cpu;
541 else
542 return cpu & nr_cpus_mask;
543}
544
545static
4105056a 546void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 547{
3171717f 548 unsigned long percpu_count;
df44348d
MD
549 int cpu;
550
551 if (unlikely(!ht->percpu_count))
3171717f 552 return;
df44348d
MD
553 cpu = ht_get_cpu();
554 if (unlikely(cpu < 0))
3171717f
MD
555 return;
556 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 557 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 558 long count;
df44348d
MD
559
560 dbg_printf("add percpu %lu\n", percpu_count);
561 count = uatomic_add_return(&ht->count,
562 1UL << COUNT_COMMIT_ORDER);
563 /* If power of 2 */
564 if (!(count & (count - 1))) {
4105056a 565 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 566 return;
e3ecefd6 567 dbg_printf("add set global %ld\n", count);
4105056a 568 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 569 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
570 }
571 }
572}
573
574static
860d07e8 575void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
576{
577 unsigned long percpu_count;
3171717f 578 int cpu;
df44348d 579
3171717f
MD
580 if (unlikely(!ht->percpu_count))
581 return;
582 cpu = ht_get_cpu();
583 if (unlikely(cpu < 0))
584 return;
80d90c06 585 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 586 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 587 long count;
df44348d 588
860d07e8 589 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 590 count = uatomic_add_return(&ht->count,
3171717f 591 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
592 /* If power of 2 */
593 if (!(count & (count - 1))) {
4105056a 594 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 595 return;
e3ecefd6
MD
596 dbg_printf("del set global %ld\n", count);
597 /*
c941bb9e 598 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
599 * certain threshold.
600 */
c941bb9e 601 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 602 return;
4105056a 603 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 604 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
605 }
606 }
607}
608
609#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
610
91452a6a 611static const long nr_cpus_mask = -2;
df44348d
MD
612
613static
614struct ht_items_count *alloc_per_cpu_items_count(void)
615{
616 return NULL;
617}
618
619static
620void free_per_cpu_items_count(struct ht_items_count *count)
621{
622}
623
624static
4105056a 625void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
626{
627}
628
629static
860d07e8 630void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
631{
632}
633
634#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
635
636
f9830efd 637static
4105056a 638void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 639{
f8994aee
MD
640 unsigned long count;
641
b8af5011
MD
642 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
643 return;
f8994aee
MD
644 count = uatomic_read(&ht->count);
645 /*
646 * Use bucket-local length for small table expand and for
647 * environments lacking per-cpu data support.
648 */
649 if (count >= (1UL << COUNT_COMMIT_ORDER))
650 return;
24365af7 651 if (chain_len > 100)
f0c29ed7 652 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 653 chain_len);
3390d470 654 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 655 cds_lfht_resize_lazy(ht, size,
01370f0b 656 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
657}
658
abc490a1 659static
14044b37 660struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 661{
14044b37 662 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
663}
664
665static
14044b37 666int is_removed(struct cds_lfht_node *node)
abc490a1 667{
d37166c6 668 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
669}
670
671static
14044b37 672struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 673{
14044b37 674 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
675}
676
f5596c94 677static
14044b37 678int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
679{
680 return ((unsigned long) node) & DUMMY_FLAG;
681}
682
683static
14044b37 684struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 685{
14044b37 686 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 687}
bb7b2f26
MD
688
689static
690struct cds_lfht_node *get_end(void)
691{
692 return (struct cds_lfht_node *) END_VALUE;
693}
694
695static
696int is_end(struct cds_lfht_node *node)
697{
698 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
699}
700
abc490a1 701static
f9830efd 702unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
703{
704 unsigned long old1, old2;
705
706 old1 = uatomic_read(ptr);
707 do {
708 old2 = old1;
709 if (old2 >= v)
f9830efd 710 return old2;
abc490a1 711 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 712 return v;
abc490a1
MD
713}
714
1475579c
MD
715static
716void cds_lfht_free_level(struct rcu_head *head)
717{
718 struct rcu_level *l =
719 caa_container_of(head, struct rcu_level, head);
98808fb1 720 poison_free(l);
1475579c
MD
721}
722
273399de
MD
723/*
724 * Remove all logically deleted nodes from a bucket up to a certain node key.
725 */
726static
f9c80341 727void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 728{
14044b37 729 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 730
c90201ac
MD
731 assert(!is_dummy(dummy));
732 assert(!is_removed(dummy));
733 assert(!is_dummy(node));
734 assert(!is_removed(node));
273399de
MD
735 for (;;) {
736 iter_prev = dummy;
737 /* We can always skip the dummy node initially */
cc4fcb10
MD
738 iter = rcu_dereference(iter_prev->p.next);
739 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
740 /*
741 * We should never be called with dummy (start of chain)
742 * and logically removed node (end of path compression
743 * marker) being the actual same node. This would be a
744 * bug in the algorithm implementation.
745 */
746 assert(dummy != node);
273399de 747 for (;;) {
bb7b2f26 748 if (unlikely(is_end(iter)))
f9c80341 749 return;
76412f24 750 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 751 return;
cc4fcb10 752 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 753 if (likely(is_removed(next)))
273399de 754 break;
b453eae1 755 iter_prev = clear_flag(iter);
273399de
MD
756 iter = next;
757 }
b198f0fd 758 assert(!is_removed(iter));
f5596c94
MD
759 if (is_dummy(iter))
760 new_next = flag_dummy(clear_flag(next));
761 else
762 new_next = clear_flag(next);
48ed1c18
MD
763 if (is_removed(iter))
764 new_next = flag_removed(new_next);
f5596c94 765 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 766 }
f9c80341 767 return;
273399de
MD
768}
769
9357c415
MD
770static
771int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
772 struct cds_lfht_node *old_node,
773 struct cds_lfht_node *ret_next,
774 struct cds_lfht_node *new_node)
775{
776 struct cds_lfht_node *dummy, *old_next;
777 struct _cds_lfht_node *lookup;
778 int flagged = 0;
779 unsigned long hash, index, order;
780
781 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
782 goto end;
783
784 assert(!is_removed(old_node));
785 assert(!is_dummy(old_node));
786 assert(!is_removed(new_node));
787 assert(!is_dummy(new_node));
788 assert(new_node != old_node);
789 do {
790 /* Insert after node to be replaced */
791 old_next = ret_next;
792 if (is_removed(old_next)) {
793 /*
794 * Too late, the old node has been removed under us
795 * between lookup and replace. Fail.
796 */
797 goto end;
798 }
799 assert(!is_dummy(old_next));
800 assert(new_node != clear_flag(old_next));
801 new_node->p.next = clear_flag(old_next);
802 /*
803 * Here is the whole trick for lock-free replace: we add
804 * the replacement node _after_ the node we want to
805 * replace by atomically setting its next pointer at the
806 * same time we set its removal flag. Given that
807 * the lookups/get next use an iterator aware of the
808 * next pointer, they will either skip the old node due
809 * to the removal flag and see the new node, or use
810 * the old node, but will not see the new one.
811 */
812 ret_next = uatomic_cmpxchg(&old_node->p.next,
813 old_next, flag_removed(new_node));
814 } while (ret_next != old_next);
815
816 /* We performed the replacement. */
817 flagged = 1;
818
819 /*
820 * Ensure that the old node is not visible to readers anymore:
821 * lookup for the node, and remove it (along with any other
822 * logically removed node) if found.
823 */
824 hash = bit_reverse_ulong(old_node->p.reverse_hash);
825 assert(size > 0);
826 index = hash & (size - 1);
827 order = get_count_order_ulong(index + 1);
828 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
829 dummy = (struct cds_lfht_node *) lookup;
830 _cds_lfht_gc_bucket(dummy, new_node);
831end:
832 /*
833 * Only the flagging action indicated that we (and no other)
834 * replaced the node from the hash table.
835 */
836 if (flagged) {
837 assert(is_removed(rcu_dereference(old_node->p.next)));
838 return 0;
839 } else {
840 return -ENOENT;
841 }
842}
843
abc490a1 844static
4105056a
MD
845struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
846 unsigned long size,
847 struct cds_lfht_node *node,
48ed1c18 848 enum add_mode mode, int dummy)
abc490a1 849{
14044b37 850 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
adc0de68 851 *dummy_node, *return_node;
14044b37 852 struct _cds_lfht_node *lookup;
24365af7 853 unsigned long hash, index, order;
abc490a1 854
c90201ac
MD
855 assert(!is_dummy(node));
856 assert(!is_removed(node));
4105056a 857 if (!size) {
f5596c94 858 assert(dummy);
bb7b2f26 859 node->p.next = flag_dummy(get_end());
18117871
MD
860 return node; /* Initial first add (head) */
861 }
cc4fcb10 862 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 863 for (;;) {
adc0de68 864 uint32_t chain_len = 0;
abc490a1 865
11519af6
MD
866 /*
867 * iter_prev points to the non-removed node prior to the
868 * insert location.
11519af6 869 */
4105056a 870 index = hash & (size - 1);
24365af7 871 order = get_count_order_ulong(index + 1);
4105056a 872 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 873 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 874 /* We can always skip the dummy node initially */
cc4fcb10
MD
875 iter = rcu_dereference(iter_prev->p.next);
876 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 877 for (;;) {
bb7b2f26 878 if (unlikely(is_end(iter)))
273399de 879 goto insert;
76412f24 880 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 881 goto insert;
cc4fcb10 882 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 883 if (unlikely(is_removed(next)))
9dba85be 884 goto gc_node;
48ed1c18 885 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
1b81fe1a 886 && !is_dummy(next)
e43f23f8
MD
887 && !ht->compare_fct(node->key, node->key_len,
888 clear_flag(iter)->key,
48ed1c18
MD
889 clear_flag(iter)->key_len)) {
890 if (mode == ADD_UNIQUE)
891 return clear_flag(iter);
892 else /* mode == ADD_REPLACE */
893 goto replace;
894 }
11519af6 895 /* Only account for identical reverse hash once */
24365af7
MD
896 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
897 && !is_dummy(next))
4105056a 898 check_resize(ht, size, ++chain_len);
11519af6 899 iter_prev = clear_flag(iter);
273399de 900 iter = next;
abc490a1 901 }
48ed1c18 902
273399de 903 insert:
7ec59d3b 904 assert(node != clear_flag(iter));
11519af6 905 assert(!is_removed(iter_prev));
c90201ac 906 assert(!is_removed(iter));
f000907d 907 assert(iter_prev != node);
f9c80341 908 if (!dummy)
1b81fe1a 909 node->p.next = clear_flag(iter);
f9c80341
MD
910 else
911 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
912 if (is_dummy(iter))
913 new_node = flag_dummy(node);
914 else
915 new_node = node;
cc4fcb10 916 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 917 new_node) != iter) {
273399de 918 continue; /* retry */
48ed1c18
MD
919 } else {
920 if (mode == ADD_REPLACE)
921 return_node = NULL;
922 else /* ADD_DEFAULT and ADD_UNIQUE */
923 return_node = node;
273399de 924 goto gc_end;
48ed1c18
MD
925 }
926
927 replace:
9357c415
MD
928
929 if (!_cds_lfht_replace(ht, size, clear_flag(iter), next,
930 node)) {
931 return_node = clear_flag(iter);
932 goto end; /* gc already done */
48ed1c18 933 } else {
9357c415 934 continue; /* retry */
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 }
273399de
MD
946gc_end:
947 /* Garbage collect logically removed nodes in the bucket */
4105056a 948 index = hash & (size - 1);
24365af7 949 order = get_count_order_ulong(index + 1);
4105056a 950 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 951 dummy_node = (struct cds_lfht_node *) lookup;
f9c80341 952 _cds_lfht_gc_bucket(dummy_node, node);
9357c415 953end:
48ed1c18 954 return return_node;
abc490a1 955}
464a1ec9 956
abc490a1 957static
860d07e8 958int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 959 struct cds_lfht_node *node,
b198f0fd 960 int dummy_removal)
abc490a1 961{
14044b37
MD
962 struct cds_lfht_node *dummy, *next, *old;
963 struct _cds_lfht_node *lookup;
abc490a1 964 int flagged = 0;
24365af7 965 unsigned long hash, index, order;
5e28c532 966
9357c415
MD
967 if (!node) /* Return -ENOENT if asked to delete NULL node */
968 goto end;
969
7ec59d3b 970 /* logically delete the node */
c90201ac
MD
971 assert(!is_dummy(node));
972 assert(!is_removed(node));
cc4fcb10 973 old = rcu_dereference(node->p.next);
7ec59d3b 974 do {
48ed1c18
MD
975 struct cds_lfht_node *new_next;
976
7ec59d3b 977 next = old;
76412f24 978 if (unlikely(is_removed(next)))
7ec59d3b 979 goto end;
1475579c
MD
980 if (dummy_removal)
981 assert(is_dummy(next));
982 else
983 assert(!is_dummy(next));
48ed1c18 984 new_next = flag_removed(next);
48ed1c18 985 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
986 } while (old != next);
987
988 /* We performed the (logical) deletion. */
989 flagged = 1;
990
991 /*
992 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
993 * the node, and remove it (along with any other logically removed node)
994 * if found.
11519af6 995 */
cc4fcb10 996 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
997 assert(size > 0);
998 index = hash & (size - 1);
24365af7 999 order = get_count_order_ulong(index + 1);
4105056a 1000 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 1001 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1002 _cds_lfht_gc_bucket(dummy, node);
2ed95849 1003end:
11519af6
MD
1004 /*
1005 * Only the flagging action indicated that we (and no other)
1006 * removed the node from the hash.
1007 */
7ec59d3b 1008 if (flagged) {
cc4fcb10 1009 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 1010 return 0;
9357c415 1011 } else {
11519af6 1012 return -ENOENT;
9357c415 1013 }
abc490a1 1014}
2ed95849 1015
b7d619b0
MD
1016static
1017void *partition_resize_thread(void *arg)
1018{
1019 struct partition_resize_work *work = arg;
1020
1021 work->ht->cds_lfht_rcu_register_thread();
1022 work->fct(work->ht, work->i, work->start, work->len);
1023 work->ht->cds_lfht_rcu_unregister_thread();
1024 return NULL;
1025}
1026
1027static
1028void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1029 unsigned long len,
1030 void (*fct)(struct cds_lfht *ht, unsigned long i,
1031 unsigned long start, unsigned long len))
1032{
1033 unsigned long partition_len;
1034 struct partition_resize_work *work;
6083a889
MD
1035 int thread, ret;
1036 unsigned long nr_threads;
b7d619b0 1037
6083a889
MD
1038 /*
1039 * Note: nr_cpus_mask + 1 is always power of 2.
1040 * We spawn just the number of threads we need to satisfy the minimum
1041 * partition size, up to the number of CPUs in the system.
1042 */
91452a6a
MD
1043 if (nr_cpus_mask > 0) {
1044 nr_threads = min(nr_cpus_mask + 1,
1045 len >> MIN_PARTITION_PER_THREAD_ORDER);
1046 } else {
1047 nr_threads = 1;
1048 }
6083a889
MD
1049 partition_len = len >> get_count_order_ulong(nr_threads);
1050 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1051 assert(work);
6083a889
MD
1052 for (thread = 0; thread < nr_threads; thread++) {
1053 work[thread].ht = ht;
1054 work[thread].i = i;
1055 work[thread].len = partition_len;
1056 work[thread].start = thread * partition_len;
1057 work[thread].fct = fct;
1af6e26e 1058 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1059 partition_resize_thread, &work[thread]);
b7d619b0
MD
1060 assert(!ret);
1061 }
6083a889 1062 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1063 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1064 assert(!ret);
1065 }
1066 free(work);
b7d619b0
MD
1067}
1068
e8de508e
MD
1069/*
1070 * Holding RCU read lock to protect _cds_lfht_add against memory
1071 * reclaim that could be performed by other call_rcu worker threads (ABA
1072 * problem).
9ee0fc9a 1073 *
b7d619b0 1074 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1075 * many worker threads, based on the number of CPUs available in the system.
1076 * This should therefore take care of not having the expand lagging behind too
1077 * many concurrent insertion threads by using the scheduler's ability to
1078 * schedule dummy node population fairly with insertions.
e8de508e 1079 */
4105056a 1080static
b7d619b0
MD
1081void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1082 unsigned long start, unsigned long len)
4105056a
MD
1083{
1084 unsigned long j;
1085
1086 ht->cds_lfht_rcu_read_lock();
b7d619b0 1087 for (j = start; j < start + len; j++) {
4105056a
MD
1088 struct cds_lfht_node *new_node =
1089 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1090
dc1da8f6 1091 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1092 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1093 new_node->p.reverse_hash =
1094 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a 1095 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1096 new_node, ADD_DEFAULT, 1);
4105056a
MD
1097 }
1098 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1099}
1100
1101static
1102void init_table_populate(struct cds_lfht *ht, unsigned long i,
1103 unsigned long len)
1104{
1105 assert(nr_cpus_mask != -1);
6083a889 1106 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1107 ht->cds_lfht_rcu_thread_online();
1108 init_table_populate_partition(ht, i, 0, len);
1109 ht->cds_lfht_rcu_thread_offline();
1110 return;
1111 }
1112 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1113}
1114
abc490a1 1115static
4105056a 1116void init_table(struct cds_lfht *ht,
24365af7
MD
1117 unsigned long first_order, unsigned long len_order)
1118{
1119 unsigned long i, end_order;
1120
f0c29ed7 1121 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1122 first_order, first_order + len_order);
1123 end_order = first_order + len_order;
24365af7 1124 for (i = first_order; i < end_order; i++) {
4105056a 1125 unsigned long len;
24365af7
MD
1126
1127 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1128 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1129
1130 /* Stop expand if the resize target changes under us */
1131 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1132 break;
1133
4105056a 1134 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1135 + (len * sizeof(struct _cds_lfht_node)));
b7d619b0 1136 assert(ht->t.tbl[i]);
4105056a 1137
4105056a 1138 /*
dc1da8f6
MD
1139 * Set all dummy nodes reverse hash values for a level and
1140 * link all dummy nodes into the table.
4105056a 1141 */
dc1da8f6 1142 init_table_populate(ht, i, len);
4105056a 1143
f9c80341
MD
1144 /*
1145 * Update table size.
1146 */
1147 cmm_smp_wmb(); /* populate data before RCU size */
1148 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1149
4105056a
MD
1150 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1151 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1152 break;
1153 }
1154}
1155
e8de508e
MD
1156/*
1157 * Holding RCU read lock to protect _cds_lfht_remove against memory
1158 * reclaim that could be performed by other call_rcu worker threads (ABA
1159 * problem).
1160 * For a single level, we logically remove and garbage collect each node.
1161 *
1162 * As a design choice, we perform logical removal and garbage collection on a
1163 * node-per-node basis to simplify this algorithm. We also assume keeping good
1164 * cache locality of the operation would overweight possible performance gain
1165 * that could be achieved by batching garbage collection for multiple levels.
1166 * However, this would have to be justified by benchmarks.
1167 *
1168 * Concurrent removal and add operations are helping us perform garbage
1169 * collection of logically removed nodes. We guarantee that all logically
1170 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1171 * invoked to free a hole level of dummy nodes (after a grace period).
1172 *
1173 * Logical removal and garbage collection can therefore be done in batch or on a
1174 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1175 *
b7d619b0
MD
1176 * When we reach a certain length, we can split this removal over many worker
1177 * threads, based on the number of CPUs available in the system. This should
1178 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1179 * updater threads actively inserting into the hash table.
e8de508e 1180 */
4105056a 1181static
b7d619b0
MD
1182void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1183 unsigned long start, unsigned long len)
4105056a
MD
1184{
1185 unsigned long j;
1186
1187 ht->cds_lfht_rcu_read_lock();
b7d619b0 1188 for (j = start; j < start + len; j++) {
4105056a
MD
1189 struct cds_lfht_node *fini_node =
1190 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1191
1192 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1193 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1194 fini_node->p.reverse_hash =
1195 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1196 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1197 fini_node, 1);
abc490a1 1198 }
4105056a 1199 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1200}
1201
1202static
1203void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1204{
1205
1206 assert(nr_cpus_mask != -1);
6083a889 1207 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1208 ht->cds_lfht_rcu_thread_online();
1209 remove_table_partition(ht, i, 0, len);
1210 ht->cds_lfht_rcu_thread_offline();
1211 return;
1212 }
1213 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1214}
1215
1475579c 1216static
4105056a 1217void fini_table(struct cds_lfht *ht,
1475579c
MD
1218 unsigned long first_order, unsigned long len_order)
1219{
1220 long i, end_order;
1221
1222 dbg_printf("fini table: first_order %lu end_order %lu\n",
1223 first_order, first_order + len_order);
1224 end_order = first_order + len_order;
1225 assert(first_order > 0);
1475579c 1226 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1227 unsigned long len;
1475579c
MD
1228
1229 len = !i ? 1 : 1UL << (i - 1);
1230 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1231
4d676753
MD
1232 /* Stop shrink if the resize target changes under us */
1233 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1234 break;
1235
1236 cmm_smp_wmb(); /* populate data before RCU size */
1237 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1238
1239 /*
1240 * We need to wait for all add operations to reach Q.S. (and
1241 * thus use the new table for lookups) before we can start
1242 * releasing the old dummy nodes. Otherwise their lookup will
1243 * return a logically removed node as insert position.
1244 */
1245 ht->cds_lfht_synchronize_rcu();
1246
21263e21 1247 /*
4105056a
MD
1248 * Set "removed" flag in dummy nodes about to be removed.
1249 * Unlink all now-logically-removed dummy node pointers.
1250 * Concurrent add/remove operation are helping us doing
1251 * the gc.
21263e21 1252 */
4105056a
MD
1253 remove_table(ht, i, len);
1254
1255 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1256
1257 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1258 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1259 break;
1260 }
1475579c
MD
1261}
1262
7a9dcf9b 1263struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1264 cds_lfht_compare_fct compare_fct,
1265 unsigned long hash_seed,
1266 unsigned long init_size,
b8af5011 1267 int flags,
14044b37 1268 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1269 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1270 void (*cds_lfht_synchronize_rcu)(void),
1271 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1272 void (*cds_lfht_rcu_read_unlock)(void),
1273 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1274 void (*cds_lfht_rcu_thread_online)(void),
1275 void (*cds_lfht_rcu_register_thread)(void),
1276 void (*cds_lfht_rcu_unregister_thread)(void),
1277 pthread_attr_t *attr)
abc490a1 1278{
14044b37 1279 struct cds_lfht *ht;
24365af7 1280 unsigned long order;
abc490a1 1281
8129be4e 1282 /* init_size must be power of two */
49619ea0 1283 if (init_size && (init_size & (init_size - 1)))
8129be4e 1284 return NULL;
14044b37 1285 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1286 assert(ht);
abc490a1 1287 ht->hash_fct = hash_fct;
732ad076
MD
1288 ht->compare_fct = compare_fct;
1289 ht->hash_seed = hash_seed;
14044b37 1290 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1291 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1292 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1293 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1294 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1295 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1296 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1297 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1298 ht->resize_attr = attr;
df44348d 1299 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1300 /* this mutex should not nest in read-side C.S. */
1301 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1302 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1303 ht->flags = flags;
5f511391 1304 ht->cds_lfht_rcu_thread_offline();
f000907d 1305 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1306 ht->t.resize_target = 1UL << (order - 1);
4105056a 1307 init_table(ht, 0, order);
f000907d 1308 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1309 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1310 return ht;
1311}
1312
adc0de68
MD
1313void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1314 struct cds_lfht_iter *iter)
2ed95849 1315{
bb7b2f26 1316 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1317 struct _cds_lfht_node *lookup;
4105056a 1318 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1319
732ad076 1320 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1321 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1322
4105056a
MD
1323 size = rcu_dereference(ht->t.size);
1324 index = hash & (size - 1);
24365af7 1325 order = get_count_order_ulong(index + 1);
4105056a 1326 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1327 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1328 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1329 dummy_node = (struct cds_lfht_node *) lookup;
1330 /* We can always skip the dummy node initially */
1331 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1332 node = clear_flag(node);
2ed95849 1333 for (;;) {
bb7b2f26 1334 if (unlikely(is_end(node))) {
96ad1112 1335 node = next = NULL;
abc490a1 1336 break;
bb7b2f26 1337 }
cc4fcb10 1338 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1339 node = next = NULL;
abc490a1 1340 break;
2ed95849 1341 }
1b81fe1a 1342 next = rcu_dereference(node->p.next);
adc0de68 1343 if (likely(!is_removed(next))
1b81fe1a 1344 && !is_dummy(next)
49c2e2d6 1345 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1346 break;
2ed95849 1347 }
1b81fe1a 1348 node = clear_flag(next);
2ed95849 1349 }
1b81fe1a 1350 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1351 iter->node = node;
1352 iter->next = next;
abc490a1 1353}
e0ba718a 1354
3883c0e5 1355void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1356{
adc0de68 1357 struct cds_lfht_node *node, *next;
a481e5ff
MD
1358 unsigned long reverse_hash;
1359 void *key;
1360 size_t key_len;
1361
adc0de68 1362 node = iter->node;
a481e5ff
MD
1363 reverse_hash = node->p.reverse_hash;
1364 key = node->key;
1365 key_len = node->key_len;
adc0de68 1366 next = iter->next;
a481e5ff
MD
1367 node = clear_flag(next);
1368
1369 for (;;) {
bb7b2f26 1370 if (unlikely(is_end(node))) {
96ad1112 1371 node = next = NULL;
a481e5ff 1372 break;
bb7b2f26 1373 }
a481e5ff 1374 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1375 node = next = NULL;
a481e5ff
MD
1376 break;
1377 }
1378 next = rcu_dereference(node->p.next);
adc0de68 1379 if (likely(!is_removed(next))
a481e5ff
MD
1380 && !is_dummy(next)
1381 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1382 break;
1383 }
1384 node = clear_flag(next);
1385 }
1386 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1387 iter->node = node;
1388 iter->next = next;
a481e5ff
MD
1389}
1390
4e9b9fbf
MD
1391void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1392{
1393 struct cds_lfht_node *node, *next;
1394
853395e1 1395 node = clear_flag(iter->next);
4e9b9fbf
MD
1396 for (;;) {
1397 if (unlikely(is_end(node))) {
1398 node = next = NULL;
1399 break;
1400 }
1401 next = rcu_dereference(node->p.next);
1402 if (likely(!is_removed(next))
1403 && !is_dummy(next)) {
1404 break;
1405 }
1406 node = clear_flag(next);
1407 }
1408 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1409 iter->node = node;
1410 iter->next = next;
1411}
1412
1413void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1414{
1415 struct _cds_lfht_node *lookup;
1416
1417 /*
1418 * Get next after first dummy node. The first dummy node is the
1419 * first node of the linked list.
1420 */
1421 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1422 iter->next = lookup->next;
4e9b9fbf
MD
1423 cds_lfht_next(ht, iter);
1424}
1425
14044b37 1426void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1427{
4105056a 1428 unsigned long hash, size;
ab7d5fc6 1429
49c2e2d6 1430 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1431 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1432
4105056a 1433 size = rcu_dereference(ht->t.size);
48ed1c18 1434 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
4105056a 1435 ht_count_add(ht, size);
3eca1b8c
MD
1436}
1437
14044b37 1438struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1439 struct cds_lfht_node *node)
3eca1b8c 1440{
4105056a 1441 unsigned long hash, size;
df44348d 1442 struct cds_lfht_node *ret;
3eca1b8c 1443
49c2e2d6 1444 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1445 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1446
4105056a 1447 size = rcu_dereference(ht->t.size);
48ed1c18 1448 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
17f31d1b 1449 if (ret == node)
4105056a 1450 ht_count_add(ht, size);
df44348d 1451 return ret;
2ed95849
MD
1452}
1453
9357c415 1454struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1455 struct cds_lfht_node *node)
1456{
1457 unsigned long hash, size;
1458 struct cds_lfht_node *ret;
1459
1460 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1461 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1462
1463 size = rcu_dereference(ht->t.size);
1464 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1465 if (ret == NULL)
1466 ht_count_add(ht, size);
1467 return ret;
1468}
1469
9357c415
MD
1470int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1471 struct cds_lfht_node *new_node)
1472{
1473 unsigned long size;
1474
1475 size = rcu_dereference(ht->t.size);
1476 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1477 new_node);
1478}
1479
1480int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1481{
4105056a 1482 unsigned long size;
df44348d 1483 int ret;
abc490a1 1484
4105056a 1485 size = rcu_dereference(ht->t.size);
9357c415 1486 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1487 if (!ret)
860d07e8 1488 ht_count_del(ht, size);
df44348d 1489 return ret;
2ed95849 1490}
ab7d5fc6 1491
abc490a1 1492static
14044b37 1493int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1494{
14044b37
MD
1495 struct cds_lfht_node *node;
1496 struct _cds_lfht_node *lookup;
4105056a 1497 unsigned long order, i, size;
674f7a69 1498
abc490a1 1499 /* Check that the table is empty */
4105056a 1500 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1501 node = (struct cds_lfht_node *) lookup;
abc490a1 1502 do {
1b81fe1a
MD
1503 node = clear_flag(node)->p.next;
1504 if (!is_dummy(node))
abc490a1 1505 return -EPERM;
273399de 1506 assert(!is_removed(node));
bb7b2f26 1507 } while (!is_end(node));
4105056a
MD
1508 /*
1509 * size accessed without rcu_dereference because hash table is
1510 * being destroyed.
1511 */
1512 size = ht->t.size;
abc490a1 1513 /* Internal sanity check: all nodes left should be dummy */
4105056a 1514 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1515 unsigned long len;
1516
1517 len = !order ? 1 : 1UL << (order - 1);
1518 for (i = 0; i < len; i++) {
f0c29ed7 1519 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1520 order, i,
4105056a
MD
1521 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1522 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1523 }
4105056a 1524 poison_free(ht->t.tbl[order]);
674f7a69 1525 }
abc490a1 1526 return 0;
674f7a69
MD
1527}
1528
1529/*
1530 * Should only be called when no more concurrent readers nor writers can
1531 * possibly access the table.
1532 */
b7d619b0 1533int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1534{
5e28c532
MD
1535 int ret;
1536
848d4088 1537 /* Wait for in-flight resize operations to complete */
24953e08
MD
1538 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1539 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1540 while (uatomic_read(&ht->in_progress_resize))
1541 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1542 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1543 if (ret)
1544 return ret;
df44348d 1545 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1546 if (attr)
1547 *attr = ht->resize_attr;
98808fb1 1548 poison_free(ht);
5e28c532 1549 return ret;
674f7a69
MD
1550}
1551
14044b37 1552void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1553 long *approx_before,
273399de 1554 unsigned long *count,
973e5e1b 1555 unsigned long *removed,
d933dd0e 1556 long *approx_after)
273399de 1557{
14044b37
MD
1558 struct cds_lfht_node *node, *next;
1559 struct _cds_lfht_node *lookup;
24365af7 1560 unsigned long nr_dummy = 0;
273399de 1561
7ed7682f 1562 *approx_before = 0;
973e5e1b
MD
1563 if (nr_cpus_mask >= 0) {
1564 int i;
1565
1566 for (i = 0; i < nr_cpus_mask + 1; i++) {
1567 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1568 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1569 }
1570 }
1571
273399de
MD
1572 *count = 0;
1573 *removed = 0;
1574
24365af7 1575 /* Count non-dummy nodes in the table */
4105056a 1576 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1577 node = (struct cds_lfht_node *) lookup;
273399de 1578 do {
cc4fcb10 1579 next = rcu_dereference(node->p.next);
b198f0fd 1580 if (is_removed(next)) {
973e5e1b
MD
1581 if (!is_dummy(next))
1582 (*removed)++;
1583 else
1584 (nr_dummy)++;
1b81fe1a 1585 } else if (!is_dummy(next))
273399de 1586 (*count)++;
24365af7
MD
1587 else
1588 (nr_dummy)++;
273399de 1589 node = clear_flag(next);
bb7b2f26 1590 } while (!is_end(node));
f0c29ed7 1591 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1592 *approx_after = 0;
973e5e1b
MD
1593 if (nr_cpus_mask >= 0) {
1594 int i;
1595
1596 for (i = 0; i < nr_cpus_mask + 1; i++) {
1597 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1598 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1599 }
1600 }
273399de
MD
1601}
1602
1475579c 1603/* called with resize mutex held */
abc490a1 1604static
4105056a 1605void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1606 unsigned long old_size, unsigned long new_size)
abc490a1 1607{
1475579c 1608 unsigned long old_order, new_order;
1475579c
MD
1609
1610 old_order = get_count_order_ulong(old_size) + 1;
1611 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1612 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1613 old_size, old_order, new_size, new_order);
1475579c 1614 assert(new_size > old_size);
4105056a 1615 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1616}
1617
1618/* called with resize mutex held */
1619static
4105056a 1620void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1621 unsigned long old_size, unsigned long new_size)
464a1ec9 1622{
1475579c 1623 unsigned long old_order, new_order;
464a1ec9 1624
cd95516d 1625 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1626 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1627 new_order = get_count_order_ulong(new_size) + 1;
1a401918
LJ
1628 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1629 old_size, old_order, new_size, new_order);
1475579c 1630 assert(new_size < old_size);
1475579c 1631
4105056a
MD
1632 /* Remove and unlink all dummy nodes to remove. */
1633 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1634}
1635
1475579c
MD
1636
1637/* called with resize mutex held */
1638static
1639void _do_cds_lfht_resize(struct cds_lfht *ht)
1640{
1641 unsigned long new_size, old_size;
4105056a
MD
1642
1643 /*
1644 * Resize table, re-do if the target size has changed under us.
1645 */
1646 do {
d2be3620
MD
1647 assert(uatomic_read(&ht->in_progress_resize));
1648 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1649 break;
4105056a
MD
1650 ht->t.resize_initiated = 1;
1651 old_size = ht->t.size;
1652 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1653 if (old_size < new_size)
1654 _do_cds_lfht_grow(ht, old_size, new_size);
1655 else if (old_size > new_size)
1656 _do_cds_lfht_shrink(ht, old_size, new_size);
1657 ht->t.resize_initiated = 0;
1658 /* write resize_initiated before read resize_target */
1659 cmm_smp_mb();
4d676753 1660 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1661}
1662
abc490a1 1663static
4105056a 1664unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1665 int growth_order)
464a1ec9 1666{
4105056a
MD
1667 return _uatomic_max(&ht->t.resize_target,
1668 size << growth_order);
464a1ec9
MD
1669}
1670
1475579c 1671static
4105056a 1672void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1673 unsigned long count)
1475579c 1674{
cd95516d 1675 count = max(count, MIN_TABLE_SIZE);
4105056a 1676 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1677}
1678
1679void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1680{
4105056a
MD
1681 resize_target_update_count(ht, new_size);
1682 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1683 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1684 pthread_mutex_lock(&ht->resize_mutex);
1685 _do_cds_lfht_resize(ht);
1686 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1687 ht->cds_lfht_rcu_thread_online();
abc490a1 1688}
464a1ec9 1689
abc490a1
MD
1690static
1691void do_resize_cb(struct rcu_head *head)
1692{
1693 struct rcu_resize_work *work =
1694 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1695 struct cds_lfht *ht = work->ht;
abc490a1 1696
5f511391 1697 ht->cds_lfht_rcu_thread_offline();
abc490a1 1698 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1699 _do_cds_lfht_resize(ht);
abc490a1 1700 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1701 ht->cds_lfht_rcu_thread_online();
98808fb1 1702 poison_free(work);
848d4088
MD
1703 cmm_smp_mb(); /* finish resize before decrement */
1704 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1705}
1706
abc490a1 1707static
4105056a 1708void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1709{
abc490a1 1710 struct rcu_resize_work *work;
f9830efd 1711 unsigned long target_size;
abc490a1 1712
4105056a
MD
1713 target_size = resize_target_update(ht, size, growth);
1714 /* Store resize_target before read resize_initiated */
1715 cmm_smp_mb();
1716 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1717 uatomic_inc(&ht->in_progress_resize);
59290e9d 1718 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1719 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1720 uatomic_dec(&ht->in_progress_resize);
59290e9d 1721 return;
ed35e6d8 1722 }
f9830efd
MD
1723 work = malloc(sizeof(*work));
1724 work->ht = ht;
14044b37 1725 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1726 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1727 }
ab7d5fc6 1728}
3171717f 1729
f8994aee
MD
1730#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1731
3171717f 1732static
4105056a 1733void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1734 unsigned long count)
1735{
1736 struct rcu_resize_work *work;
3171717f 1737
b8af5011
MD
1738 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1739 return;
4105056a
MD
1740 resize_target_update_count(ht, count);
1741 /* Store resize_target before read resize_initiated */
1742 cmm_smp_mb();
1743 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1744 uatomic_inc(&ht->in_progress_resize);
59290e9d 1745 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1746 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1747 uatomic_dec(&ht->in_progress_resize);
59290e9d 1748 return;
ed35e6d8 1749 }
3171717f
MD
1750 work = malloc(sizeof(*work));
1751 work->ht = ht;
1752 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1753 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1754 }
1755}
f8994aee
MD
1756
1757#endif
This page took 0.196121 seconds and 4 git commands to generate.