rculfhash: Remove unneed branches
[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
93d46c39
LJ
94 * this order (except for order 0).
95 * - synchronzie_rcu is used to garbage-collect the old dummy node table.
e753ff5a
MD
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.
93d46c39
LJ
99 *
100 * Dummy node tables:
101 *
102 * hash table hash table the last all dummy node tables
103 * order size dummy node 0 1 2 3 4 5 6(index)
104 * table size
105 * 0 1 1 1
106 * 1 2 1 1 1
107 * 2 4 2 1 1 2
108 * 3 8 4 1 1 2 4
109 * 4 16 8 1 1 2 4 8
110 * 5 32 16 1 1 2 4 8 16
111 * 6 64 32 1 1 2 4 8 16 32
112 *
113 * When growing/shrinking, we only focus on the last dummy node table
114 * which size is (!order ? 1 : (1 << (order -1))).
115 *
116 * Example for growing/shrinking:
117 * grow hash table from order 5 to 6: init the index=6 dummy node table
118 * shrink hash table from order 6 to 5: fini the index=6 dummy node table
119 *
1475579c
MD
120 * A bit of ascii art explanation:
121 *
122 * Order index is the off-by-one compare to the actual power of 2 because
123 * we use index 0 to deal with the 0 special-case.
124 *
125 * This shows the nodes for a small table ordered by reversed bits:
126 *
127 * bits reverse
128 * 0 000 000
129 * 4 100 001
130 * 2 010 010
131 * 6 110 011
132 * 1 001 100
133 * 5 101 101
134 * 3 011 110
135 * 7 111 111
136 *
137 * This shows the nodes in order of non-reversed bits, linked by
138 * reversed-bit order.
139 *
140 * order bits reverse
141 * 0 0 000 000
0adc36a8
LJ
142 * 1 | 1 001 100 <-
143 * 2 | | 2 010 010 <- |
f6fdd688 144 * | | | 3 011 110 | <- |
1475579c
MD
145 * 3 -> | | | 4 100 001 | |
146 * -> | | 5 101 101 |
147 * -> | 6 110 011
148 * -> 7 111 111
e753ff5a
MD
149 */
150
2ed95849
MD
151#define _LGPL_SOURCE
152#include <stdlib.h>
e0ba718a
MD
153#include <errno.h>
154#include <assert.h>
155#include <stdio.h>
abc490a1 156#include <stdint.h>
f000907d 157#include <string.h>
e0ba718a 158
15cfbec7 159#include "config.h"
2ed95849 160#include <urcu.h>
abc490a1 161#include <urcu-call-rcu.h>
a42cc659
MD
162#include <urcu/arch.h>
163#include <urcu/uatomic.h>
a42cc659 164#include <urcu/compiler.h>
abc490a1 165#include <urcu/rculfhash.h>
5e28c532 166#include <stdio.h>
464a1ec9 167#include <pthread.h>
44395fb7 168
f9830efd 169#ifdef DEBUG
f0c29ed7 170#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 171#else
e753ff5a 172#define dbg_printf(fmt, args...)
f9830efd
MD
173#endif
174
f8994aee
MD
175/*
176 * Per-CPU split-counters lazily update the global counter each 1024
177 * addition/removal. It automatically keeps track of resize required.
178 * We use the bucket length as indicator for need to expand for small
179 * tables and machines lacking per-cpu data suppport.
180 */
181#define COUNT_COMMIT_ORDER 10
6ea6bc67
MD
182#define CHAIN_LEN_TARGET 1
183#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 184
cd95516d 185/*
76a73da8 186 * Define the minimum table size.
cd95516d 187 */
c9edd44a 188#define MIN_TABLE_SIZE 1
cd95516d 189
4105056a
MD
190#if (CAA_BITS_PER_LONG == 32)
191#define MAX_TABLE_ORDER 32
192#else
193#define MAX_TABLE_ORDER 64
194#endif
195
b7d619b0
MD
196/*
197 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
198 */
6083a889
MD
199#define MIN_PARTITION_PER_THREAD_ORDER 12
200#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 201
4105056a
MD
202#ifndef min
203#define min(a, b) ((a) < (b) ? (a) : (b))
204#endif
205
abc490a1
MD
206#ifndef max
207#define max(a, b) ((a) > (b) ? (a) : (b))
208#endif
2ed95849 209
d95bd160
MD
210/*
211 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 212 * It indicates that no node must attach to the node scheduled for
b198f0fd 213 * removal, and that node garbage collection must be performed.
d95bd160
MD
214 * The dummy flag does not require to be updated atomically with the
215 * pointer, but it is added as a pointer low bit flag to save space.
216 */
d37166c6 217#define REMOVED_FLAG (1UL << 0)
b198f0fd
MD
218#define DUMMY_FLAG (1UL << 1)
219#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 220
bb7b2f26 221/* Value of the end pointer. Should not interact with flags. */
f9c80341 222#define END_VALUE NULL
bb7b2f26 223
df44348d 224struct ht_items_count {
860d07e8 225 unsigned long add, del;
df44348d
MD
226} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
227
1475579c 228struct rcu_level {
0d14ceb2 229 /* Note: manually update allocation length when adding a field */
1475579c
MD
230 struct _cds_lfht_node nodes[0];
231};
232
395270b6 233struct rcu_table {
4105056a 234 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 235 unsigned long resize_target;
11519af6 236 int resize_initiated;
4105056a 237 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
238};
239
14044b37 240struct cds_lfht {
4105056a 241 struct rcu_table t;
14044b37
MD
242 cds_lfht_hash_fct hash_fct;
243 cds_lfht_compare_fct compare_fct;
732ad076 244 unsigned long hash_seed;
b8af5011 245 int flags;
5f511391
MD
246 /*
247 * We need to put the work threads offline (QSBR) when taking this
248 * mutex, because we use synchronize_rcu within this mutex critical
249 * section, which waits on read-side critical sections, and could
250 * therefore cause grace-period deadlock if we hold off RCU G.P.
251 * completion.
252 */
464a1ec9 253 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 254 unsigned int in_progress_resize, in_progress_destroy;
14044b37 255 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 256 void (*func)(struct rcu_head *head));
1475579c 257 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
258 void (*cds_lfht_rcu_read_lock)(void);
259 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
260 void (*cds_lfht_rcu_thread_offline)(void);
261 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
262 void (*cds_lfht_rcu_register_thread)(void);
263 void (*cds_lfht_rcu_unregister_thread)(void);
264 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 265 long count; /* global approximate item count */
df44348d 266 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
267};
268
abc490a1
MD
269struct rcu_resize_work {
270 struct rcu_head head;
14044b37 271 struct cds_lfht *ht;
abc490a1 272};
2ed95849 273
b7d619b0 274struct partition_resize_work {
1af6e26e 275 pthread_t thread_id;
b7d619b0
MD
276 struct cds_lfht *ht;
277 unsigned long i, start, len;
278 void (*fct)(struct cds_lfht *ht, unsigned long i,
279 unsigned long start, unsigned long len);
280};
281
76a73da8 282static
83beee94
MD
283void _cds_lfht_add(struct cds_lfht *ht,
284 unsigned long size,
285 struct cds_lfht_node *node,
286 struct cds_lfht_iter *unique_ret,
287 int dummy);
48ed1c18 288
abc490a1
MD
289/*
290 * Algorithm to reverse bits in a word by lookup table, extended to
291 * 64-bit words.
f9830efd 292 * Source:
abc490a1 293 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 294 * Originally from Public Domain.
abc490a1
MD
295 */
296
297static const uint8_t BitReverseTable256[256] =
2ed95849 298{
abc490a1
MD
299#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
300#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
301#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
302 R6(0), R6(2), R6(1), R6(3)
303};
304#undef R2
305#undef R4
306#undef R6
2ed95849 307
abc490a1
MD
308static
309uint8_t bit_reverse_u8(uint8_t v)
310{
311 return BitReverseTable256[v];
312}
ab7d5fc6 313
abc490a1
MD
314static __attribute__((unused))
315uint32_t bit_reverse_u32(uint32_t v)
316{
317 return ((uint32_t) bit_reverse_u8(v) << 24) |
318 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
319 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
320 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
321}
322
abc490a1
MD
323static __attribute__((unused))
324uint64_t bit_reverse_u64(uint64_t v)
2ed95849 325{
abc490a1
MD
326 return ((uint64_t) bit_reverse_u8(v) << 56) |
327 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
328 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
329 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
330 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
331 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
332 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
333 ((uint64_t) bit_reverse_u8(v >> 56));
334}
335
336static
337unsigned long bit_reverse_ulong(unsigned long v)
338{
339#if (CAA_BITS_PER_LONG == 32)
340 return bit_reverse_u32(v);
341#else
342 return bit_reverse_u64(v);
343#endif
344}
345
f9830efd 346/*
24365af7
MD
347 * fls: returns the position of the most significant bit.
348 * Returns 0 if no bit is set, else returns the position of the most
349 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 350 */
24365af7
MD
351#if defined(__i386) || defined(__x86_64)
352static inline
353unsigned int fls_u32(uint32_t x)
f9830efd 354{
24365af7
MD
355 int r;
356
357 asm("bsrl %1,%0\n\t"
358 "jnz 1f\n\t"
359 "movl $-1,%0\n\t"
360 "1:\n\t"
361 : "=r" (r) : "rm" (x));
362 return r + 1;
363}
364#define HAS_FLS_U32
365#endif
366
367#if defined(__x86_64)
368static inline
369unsigned int fls_u64(uint64_t x)
370{
371 long r;
372
373 asm("bsrq %1,%0\n\t"
374 "jnz 1f\n\t"
375 "movq $-1,%0\n\t"
376 "1:\n\t"
377 : "=r" (r) : "rm" (x));
378 return r + 1;
379}
380#define HAS_FLS_U64
381#endif
382
383#ifndef HAS_FLS_U64
384static __attribute__((unused))
385unsigned int fls_u64(uint64_t x)
386{
387 unsigned int r = 64;
388
389 if (!x)
390 return 0;
391
392 if (!(x & 0xFFFFFFFF00000000ULL)) {
393 x <<= 32;
394 r -= 32;
395 }
396 if (!(x & 0xFFFF000000000000ULL)) {
397 x <<= 16;
398 r -= 16;
399 }
400 if (!(x & 0xFF00000000000000ULL)) {
401 x <<= 8;
402 r -= 8;
403 }
404 if (!(x & 0xF000000000000000ULL)) {
405 x <<= 4;
406 r -= 4;
407 }
408 if (!(x & 0xC000000000000000ULL)) {
409 x <<= 2;
410 r -= 2;
411 }
412 if (!(x & 0x8000000000000000ULL)) {
413 x <<= 1;
414 r -= 1;
415 }
416 return r;
417}
418#endif
419
420#ifndef HAS_FLS_U32
421static __attribute__((unused))
422unsigned int fls_u32(uint32_t x)
423{
424 unsigned int r = 32;
f9830efd 425
24365af7
MD
426 if (!x)
427 return 0;
428 if (!(x & 0xFFFF0000U)) {
429 x <<= 16;
430 r -= 16;
431 }
432 if (!(x & 0xFF000000U)) {
433 x <<= 8;
434 r -= 8;
435 }
436 if (!(x & 0xF0000000U)) {
437 x <<= 4;
438 r -= 4;
439 }
440 if (!(x & 0xC0000000U)) {
441 x <<= 2;
442 r -= 2;
443 }
444 if (!(x & 0x80000000U)) {
445 x <<= 1;
446 r -= 1;
447 }
448 return r;
449}
450#endif
451
452unsigned int fls_ulong(unsigned long x)
f9830efd 453{
24365af7
MD
454#if (CAA_BITS_PER_lONG == 32)
455 return fls_u32(x);
456#else
457 return fls_u64(x);
458#endif
459}
f9830efd 460
920f8ef6
LJ
461/*
462 * Return the minimum order for which x <= (1UL << order).
463 * Return -1 if x is 0.
464 */
24365af7
MD
465int get_count_order_u32(uint32_t x)
466{
920f8ef6
LJ
467 if (!x)
468 return -1;
24365af7 469
920f8ef6 470 return fls_u32(x - 1);
24365af7
MD
471}
472
920f8ef6
LJ
473/*
474 * Return the minimum order for which x <= (1UL << order).
475 * Return -1 if x is 0.
476 */
24365af7
MD
477int get_count_order_ulong(unsigned long x)
478{
920f8ef6
LJ
479 if (!x)
480 return -1;
24365af7 481
920f8ef6 482 return fls_ulong(x - 1);
f9830efd
MD
483}
484
98808fb1
MD
485#ifdef POISON_FREE
486#define poison_free(ptr) \
487 do { \
488 memset(ptr, 0x42, sizeof(*(ptr))); \
489 free(ptr); \
490 } while (0)
491#else
492#define poison_free(ptr) free(ptr)
493#endif
494
f9830efd 495static
4105056a 496void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 497
df44348d
MD
498/*
499 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
500 * available, then we support hash table item accounting.
501 * In the unfortunate event the number of CPUs reported would be
502 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
503 */
df44348d
MD
504#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
505
f8994aee 506static
4105056a 507void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
508 unsigned long count);
509
df44348d
MD
510static long nr_cpus_mask = -1;
511
512static
513struct ht_items_count *alloc_per_cpu_items_count(void)
514{
515 struct ht_items_count *count;
516
517 switch (nr_cpus_mask) {
518 case -2:
519 return NULL;
520 case -1:
521 {
522 long maxcpus;
523
524 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
525 if (maxcpus <= 0) {
526 nr_cpus_mask = -2;
527 return NULL;
528 }
529 /*
530 * round up number of CPUs to next power of two, so we
531 * can use & for modulo.
532 */
533 maxcpus = 1UL << get_count_order_ulong(maxcpus);
534 nr_cpus_mask = maxcpus - 1;
535 }
536 /* Fall-through */
537 default:
538 return calloc(nr_cpus_mask + 1, sizeof(*count));
539 }
540}
541
542static
543void free_per_cpu_items_count(struct ht_items_count *count)
544{
98808fb1 545 poison_free(count);
df44348d
MD
546}
547
548static
549int ht_get_cpu(void)
550{
551 int cpu;
552
553 assert(nr_cpus_mask >= 0);
554 cpu = sched_getcpu();
555 if (unlikely(cpu < 0))
556 return cpu;
557 else
558 return cpu & nr_cpus_mask;
559}
560
561static
4105056a 562void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 563{
3171717f 564 unsigned long percpu_count;
df44348d
MD
565 int cpu;
566
567 if (unlikely(!ht->percpu_count))
3171717f 568 return;
df44348d
MD
569 cpu = ht_get_cpu();
570 if (unlikely(cpu < 0))
3171717f
MD
571 return;
572 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 573 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 574 long count;
df44348d
MD
575
576 dbg_printf("add percpu %lu\n", percpu_count);
577 count = uatomic_add_return(&ht->count,
578 1UL << COUNT_COMMIT_ORDER);
579 /* If power of 2 */
580 if (!(count & (count - 1))) {
4105056a 581 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 582 return;
e3ecefd6 583 dbg_printf("add set global %ld\n", count);
4105056a 584 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 585 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
586 }
587 }
588}
589
590static
860d07e8 591void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
592{
593 unsigned long percpu_count;
3171717f 594 int cpu;
df44348d 595
3171717f
MD
596 if (unlikely(!ht->percpu_count))
597 return;
598 cpu = ht_get_cpu();
599 if (unlikely(cpu < 0))
600 return;
80d90c06 601 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 602 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 603 long count;
df44348d 604
860d07e8 605 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 606 count = uatomic_add_return(&ht->count,
3171717f 607 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
608 /* If power of 2 */
609 if (!(count & (count - 1))) {
4105056a 610 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 611 return;
e3ecefd6
MD
612 dbg_printf("del set global %ld\n", count);
613 /*
c941bb9e 614 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
615 * certain threshold.
616 */
c941bb9e 617 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 618 return;
4105056a 619 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 620 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
621 }
622 }
623}
624
625#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
626
91452a6a 627static const long nr_cpus_mask = -2;
df44348d
MD
628
629static
630struct ht_items_count *alloc_per_cpu_items_count(void)
631{
632 return NULL;
633}
634
635static
636void free_per_cpu_items_count(struct ht_items_count *count)
637{
638}
639
640static
4105056a 641void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
642{
643}
644
645static
860d07e8 646void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
647{
648}
649
650#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
651
652
f9830efd 653static
4105056a 654void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 655{
f8994aee
MD
656 unsigned long count;
657
b8af5011
MD
658 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
659 return;
f8994aee
MD
660 count = uatomic_read(&ht->count);
661 /*
662 * Use bucket-local length for small table expand and for
663 * environments lacking per-cpu data support.
664 */
665 if (count >= (1UL << COUNT_COMMIT_ORDER))
666 return;
24365af7 667 if (chain_len > 100)
f0c29ed7 668 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 669 chain_len);
3390d470 670 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 671 cds_lfht_resize_lazy(ht, size,
01370f0b 672 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
673}
674
abc490a1 675static
14044b37 676struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 677{
14044b37 678 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
679}
680
681static
14044b37 682int is_removed(struct cds_lfht_node *node)
abc490a1 683{
d37166c6 684 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
685}
686
687static
14044b37 688struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 689{
14044b37 690 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
691}
692
f5596c94 693static
14044b37 694int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
695{
696 return ((unsigned long) node) & DUMMY_FLAG;
697}
698
699static
14044b37 700struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 701{
14044b37 702 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 703}
bb7b2f26
MD
704
705static
706struct cds_lfht_node *get_end(void)
707{
708 return (struct cds_lfht_node *) END_VALUE;
709}
710
711static
712int is_end(struct cds_lfht_node *node)
713{
714 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
715}
716
abc490a1 717static
f9830efd 718unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
719{
720 unsigned long old1, old2;
721
722 old1 = uatomic_read(ptr);
723 do {
724 old2 = old1;
725 if (old2 >= v)
f9830efd 726 return old2;
abc490a1 727 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 728 return v;
abc490a1
MD
729}
730
f4a9cc0b
LJ
731static
732struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
733 unsigned long hash)
734{
735 unsigned long index, order;
736
737 assert(size > 0);
738 index = hash & (size - 1);
a4ea2223
LJ
739 /*
740 * equivalent to get_count_order_ulong(index + 1), but optimizes
741 * away the non-existing 0 special-case for
742 * get_count_order_ulong.
743 */
744 order = fls_ulong(index);
f4a9cc0b
LJ
745
746 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
747 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
748
749 return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
750}
751
273399de
MD
752/*
753 * Remove all logically deleted nodes from a bucket up to a certain node key.
754 */
755static
f9c80341 756void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 757{
14044b37 758 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 759
c90201ac
MD
760 assert(!is_dummy(dummy));
761 assert(!is_removed(dummy));
762 assert(!is_dummy(node));
763 assert(!is_removed(node));
273399de
MD
764 for (;;) {
765 iter_prev = dummy;
766 /* We can always skip the dummy node initially */
cc4fcb10 767 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 768 assert(!is_removed(iter));
cc4fcb10 769 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
770 /*
771 * We should never be called with dummy (start of chain)
772 * and logically removed node (end of path compression
773 * marker) being the actual same node. This would be a
774 * bug in the algorithm implementation.
775 */
776 assert(dummy != node);
273399de 777 for (;;) {
bb7b2f26 778 if (unlikely(is_end(iter)))
f9c80341 779 return;
76412f24 780 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 781 return;
cc4fcb10 782 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 783 if (likely(is_removed(next)))
273399de 784 break;
b453eae1 785 iter_prev = clear_flag(iter);
273399de
MD
786 iter = next;
787 }
b198f0fd 788 assert(!is_removed(iter));
f5596c94
MD
789 if (is_dummy(iter))
790 new_next = flag_dummy(clear_flag(next));
791 else
792 new_next = clear_flag(next);
793 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 794 }
f9c80341 795 return;
273399de
MD
796}
797
9357c415
MD
798static
799int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
800 struct cds_lfht_node *old_node,
3fb86f26 801 struct cds_lfht_node *old_next,
9357c415
MD
802 struct cds_lfht_node *new_node)
803{
3fb86f26 804 struct cds_lfht_node *dummy, *ret_next;
9357c415 805 struct _cds_lfht_node *lookup;
9357c415
MD
806
807 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 808 return -ENOENT;
9357c415
MD
809
810 assert(!is_removed(old_node));
811 assert(!is_dummy(old_node));
812 assert(!is_removed(new_node));
813 assert(!is_dummy(new_node));
814 assert(new_node != old_node);
3fb86f26 815 for (;;) {
9357c415 816 /* Insert after node to be replaced */
9357c415
MD
817 if (is_removed(old_next)) {
818 /*
819 * Too late, the old node has been removed under us
820 * between lookup and replace. Fail.
821 */
7801dadd 822 return -ENOENT;
9357c415
MD
823 }
824 assert(!is_dummy(old_next));
825 assert(new_node != clear_flag(old_next));
826 new_node->p.next = clear_flag(old_next);
827 /*
828 * Here is the whole trick for lock-free replace: we add
829 * the replacement node _after_ the node we want to
830 * replace by atomically setting its next pointer at the
831 * same time we set its removal flag. Given that
832 * the lookups/get next use an iterator aware of the
833 * next pointer, they will either skip the old node due
834 * to the removal flag and see the new node, or use
835 * the old node, but will not see the new one.
836 */
837 ret_next = uatomic_cmpxchg(&old_node->p.next,
838 old_next, flag_removed(new_node));
3fb86f26 839 if (ret_next == old_next)
7801dadd 840 break; /* We performed the replacement. */
3fb86f26
LJ
841 old_next = ret_next;
842 }
9357c415 843
9357c415
MD
844 /*
845 * Ensure that the old node is not visible to readers anymore:
846 * lookup for the node, and remove it (along with any other
847 * logically removed node) if found.
848 */
f4a9cc0b 849 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
850 dummy = (struct cds_lfht_node *) lookup;
851 _cds_lfht_gc_bucket(dummy, new_node);
7801dadd
LJ
852
853 assert(is_removed(rcu_dereference(old_node->p.next)));
854 return 0;
9357c415
MD
855}
856
83beee94
MD
857/*
858 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
859 * mode. A NULL unique_ret allows creation of duplicate keys.
860 */
abc490a1 861static
83beee94
MD
862void _cds_lfht_add(struct cds_lfht *ht,
863 unsigned long size,
864 struct cds_lfht_node *node,
865 struct cds_lfht_iter *unique_ret,
866 int dummy)
abc490a1 867{
14044b37 868 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 869 *return_node;
14044b37 870 struct _cds_lfht_node *lookup;
abc490a1 871
c90201ac
MD
872 assert(!is_dummy(node));
873 assert(!is_removed(node));
f4a9cc0b 874 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 875 for (;;) {
adc0de68 876 uint32_t chain_len = 0;
abc490a1 877
11519af6
MD
878 /*
879 * iter_prev points to the non-removed node prior to the
880 * insert location.
11519af6 881 */
14044b37 882 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 883 /* We can always skip the dummy node initially */
cc4fcb10
MD
884 iter = rcu_dereference(iter_prev->p.next);
885 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 886 for (;;) {
bb7b2f26 887 if (unlikely(is_end(iter)))
273399de 888 goto insert;
76412f24 889 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 890 goto insert;
238cc06e 891
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;
238cc06e 895
cc4fcb10 896 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 897 if (unlikely(is_removed(next)))
9dba85be 898 goto gc_node;
238cc06e
LJ
899
900 /* uniquely add */
83beee94 901 if (unique_ret
1b81fe1a 902 && !is_dummy(next)
238cc06e
LJ
903 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
904 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
905
906 /*
907 * uniquely adding inserts the node as the first
908 * node of the identical-hash-value node chain.
909 *
910 * This semantic ensures no duplicated keys
911 * should ever be observable in the table
912 * (including observe one node by one node
913 * by forward iterations)
914 */
915 cds_lfht_next_duplicate(ht, &d_iter);
916 if (!d_iter.node)
917 goto insert;
918
919 *unique_ret = d_iter;
83beee94 920 return;
48ed1c18 921 }
238cc06e 922
11519af6 923 /* Only account for identical reverse hash once */
24365af7
MD
924 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
925 && !is_dummy(next))
4105056a 926 check_resize(ht, size, ++chain_len);
11519af6 927 iter_prev = clear_flag(iter);
273399de 928 iter = next;
abc490a1 929 }
48ed1c18 930
273399de 931 insert:
7ec59d3b 932 assert(node != clear_flag(iter));
11519af6 933 assert(!is_removed(iter_prev));
c90201ac 934 assert(!is_removed(iter));
f000907d 935 assert(iter_prev != node);
f9c80341 936 if (!dummy)
1b81fe1a 937 node->p.next = clear_flag(iter);
f9c80341
MD
938 else
939 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
940 if (is_dummy(iter))
941 new_node = flag_dummy(node);
942 else
943 new_node = node;
cc4fcb10 944 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 945 new_node) != iter) {
273399de 946 continue; /* retry */
48ed1c18 947 } else {
83beee94 948 return_node = node;
960c9e4f 949 goto end;
48ed1c18
MD
950 }
951
9dba85be
MD
952 gc_node:
953 assert(!is_removed(iter));
f5596c94
MD
954 if (is_dummy(iter))
955 new_next = flag_dummy(clear_flag(next));
956 else
957 new_next = clear_flag(next);
958 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 959 /* retry */
464a1ec9 960 }
9357c415 961end:
83beee94
MD
962 if (unique_ret) {
963 unique_ret->node = return_node;
964 /* unique_ret->next left unset, never used. */
965 }
abc490a1 966}
464a1ec9 967
abc490a1 968static
860d07e8 969int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 970 struct cds_lfht_node *node,
b198f0fd 971 int dummy_removal)
abc490a1 972{
14044b37
MD
973 struct cds_lfht_node *dummy, *next, *old;
974 struct _cds_lfht_node *lookup;
5e28c532 975
9357c415 976 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 977 return -ENOENT;
9357c415 978
7ec59d3b 979 /* logically delete the node */
c90201ac
MD
980 assert(!is_dummy(node));
981 assert(!is_removed(node));
cc4fcb10 982 old = rcu_dereference(node->p.next);
7ec59d3b 983 do {
48ed1c18
MD
984 struct cds_lfht_node *new_next;
985
7ec59d3b 986 next = old;
76412f24 987 if (unlikely(is_removed(next)))
743f9143 988 return -ENOENT;
1475579c
MD
989 if (dummy_removal)
990 assert(is_dummy(next));
991 else
992 assert(!is_dummy(next));
48ed1c18 993 new_next = flag_removed(next);
48ed1c18 994 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b 995 } while (old != next);
7ec59d3b 996 /* We performed the (logical) deletion. */
7ec59d3b
MD
997
998 /*
999 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1000 * the node, and remove it (along with any other logically removed node)
1001 * if found.
11519af6 1002 */
f4a9cc0b 1003 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 1004 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1005 _cds_lfht_gc_bucket(dummy, node);
743f9143
LJ
1006
1007 assert(is_removed(rcu_dereference(node->p.next)));
1008 return 0;
abc490a1 1009}
2ed95849 1010
b7d619b0
MD
1011static
1012void *partition_resize_thread(void *arg)
1013{
1014 struct partition_resize_work *work = arg;
1015
1016 work->ht->cds_lfht_rcu_register_thread();
1017 work->fct(work->ht, work->i, work->start, work->len);
1018 work->ht->cds_lfht_rcu_unregister_thread();
1019 return NULL;
1020}
1021
1022static
1023void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1024 unsigned long len,
1025 void (*fct)(struct cds_lfht *ht, unsigned long i,
1026 unsigned long start, unsigned long len))
1027{
1028 unsigned long partition_len;
1029 struct partition_resize_work *work;
6083a889
MD
1030 int thread, ret;
1031 unsigned long nr_threads;
b7d619b0 1032
6083a889
MD
1033 /*
1034 * Note: nr_cpus_mask + 1 is always power of 2.
1035 * We spawn just the number of threads we need to satisfy the minimum
1036 * partition size, up to the number of CPUs in the system.
1037 */
91452a6a
MD
1038 if (nr_cpus_mask > 0) {
1039 nr_threads = min(nr_cpus_mask + 1,
1040 len >> MIN_PARTITION_PER_THREAD_ORDER);
1041 } else {
1042 nr_threads = 1;
1043 }
6083a889
MD
1044 partition_len = len >> get_count_order_ulong(nr_threads);
1045 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1046 assert(work);
6083a889
MD
1047 for (thread = 0; thread < nr_threads; thread++) {
1048 work[thread].ht = ht;
1049 work[thread].i = i;
1050 work[thread].len = partition_len;
1051 work[thread].start = thread * partition_len;
1052 work[thread].fct = fct;
1af6e26e 1053 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1054 partition_resize_thread, &work[thread]);
b7d619b0
MD
1055 assert(!ret);
1056 }
6083a889 1057 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1058 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1059 assert(!ret);
1060 }
1061 free(work);
b7d619b0
MD
1062}
1063
e8de508e
MD
1064/*
1065 * Holding RCU read lock to protect _cds_lfht_add against memory
1066 * reclaim that could be performed by other call_rcu worker threads (ABA
1067 * problem).
9ee0fc9a 1068 *
b7d619b0 1069 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1070 * many worker threads, based on the number of CPUs available in the system.
1071 * This should therefore take care of not having the expand lagging behind too
1072 * many concurrent insertion threads by using the scheduler's ability to
1073 * schedule dummy node population fairly with insertions.
e8de508e 1074 */
4105056a 1075static
b7d619b0
MD
1076void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1077 unsigned long start, unsigned long len)
4105056a
MD
1078{
1079 unsigned long j;
1080
4f6e90b7 1081 assert(i > 0);
4105056a 1082 ht->cds_lfht_rcu_read_lock();
b7d619b0 1083 for (j = start; j < start + len; j++) {
4105056a
MD
1084 struct cds_lfht_node *new_node =
1085 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1086
dc1da8f6 1087 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1088 i, j, (1UL << (i - 1)) + j);
dc1da8f6 1089 new_node->p.reverse_hash =
4f6e90b7
LJ
1090 bit_reverse_ulong((1UL << (i - 1)) + j);
1091 _cds_lfht_add(ht, 1UL << (i - 1),
83beee94 1092 new_node, NULL, 1);
4105056a
MD
1093 }
1094 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1095}
1096
1097static
1098void init_table_populate(struct cds_lfht *ht, unsigned long i,
1099 unsigned long len)
1100{
1101 assert(nr_cpus_mask != -1);
6083a889 1102 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1103 ht->cds_lfht_rcu_thread_online();
1104 init_table_populate_partition(ht, i, 0, len);
1105 ht->cds_lfht_rcu_thread_offline();
1106 return;
1107 }
1108 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1109}
1110
abc490a1 1111static
4105056a 1112void init_table(struct cds_lfht *ht,
93d46c39 1113 unsigned long first_order, unsigned long last_order)
24365af7 1114{
93d46c39 1115 unsigned long i;
24365af7 1116
93d46c39
LJ
1117 dbg_printf("init table: first_order %lu last_order %lu\n",
1118 first_order, last_order);
4f6e90b7 1119 assert(first_order > 0);
93d46c39 1120 for (i = first_order; i <= last_order; i++) {
4105056a 1121 unsigned long len;
24365af7 1122
4f6e90b7 1123 len = 1UL << (i - 1);
f0c29ed7 1124 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1125
1126 /* Stop expand if the resize target changes under us */
4f6e90b7 1127 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1128 break;
1129
0d14ceb2 1130 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1131 assert(ht->t.tbl[i]);
4105056a 1132
4105056a 1133 /*
dc1da8f6
MD
1134 * Set all dummy nodes reverse hash values for a level and
1135 * link all dummy nodes into the table.
4105056a 1136 */
dc1da8f6 1137 init_table_populate(ht, i, len);
4105056a 1138
f9c80341
MD
1139 /*
1140 * Update table size.
1141 */
1142 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1143 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1144
4f6e90b7 1145 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1146 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1147 break;
1148 }
1149}
1150
e8de508e
MD
1151/*
1152 * Holding RCU read lock to protect _cds_lfht_remove against memory
1153 * reclaim that could be performed by other call_rcu worker threads (ABA
1154 * problem).
1155 * For a single level, we logically remove and garbage collect each node.
1156 *
1157 * As a design choice, we perform logical removal and garbage collection on a
1158 * node-per-node basis to simplify this algorithm. We also assume keeping good
1159 * cache locality of the operation would overweight possible performance gain
1160 * that could be achieved by batching garbage collection for multiple levels.
1161 * However, this would have to be justified by benchmarks.
1162 *
1163 * Concurrent removal and add operations are helping us perform garbage
1164 * collection of logically removed nodes. We guarantee that all logically
1165 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1166 * invoked to free a hole level of dummy nodes (after a grace period).
1167 *
1168 * Logical removal and garbage collection can therefore be done in batch or on a
1169 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1170 *
b7d619b0
MD
1171 * When we reach a certain length, we can split this removal over many worker
1172 * threads, based on the number of CPUs available in the system. This should
1173 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1174 * updater threads actively inserting into the hash table.
e8de508e 1175 */
4105056a 1176static
b7d619b0
MD
1177void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1178 unsigned long start, unsigned long len)
4105056a
MD
1179{
1180 unsigned long j;
1181
4f6e90b7 1182 assert(i > 0);
4105056a 1183 ht->cds_lfht_rcu_read_lock();
b7d619b0 1184 for (j = start; j < start + len; j++) {
4105056a
MD
1185 struct cds_lfht_node *fini_node =
1186 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1187
1188 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1189 i, j, (1UL << (i - 1)) + j);
4105056a 1190 fini_node->p.reverse_hash =
4f6e90b7
LJ
1191 bit_reverse_ulong((1UL << (i - 1)) + j);
1192 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1193 }
4105056a 1194 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1195}
1196
1197static
1198void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1199{
1200
1201 assert(nr_cpus_mask != -1);
6083a889 1202 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1203 ht->cds_lfht_rcu_thread_online();
1204 remove_table_partition(ht, i, 0, len);
1205 ht->cds_lfht_rcu_thread_offline();
1206 return;
1207 }
1208 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1209}
1210
1475579c 1211static
4105056a 1212void fini_table(struct cds_lfht *ht,
93d46c39 1213 unsigned long first_order, unsigned long last_order)
1475579c 1214{
93d46c39 1215 long i;
0d14ceb2 1216 void *free_by_rcu = NULL;
1475579c 1217
93d46c39
LJ
1218 dbg_printf("fini table: first_order %lu last_order %lu\n",
1219 first_order, last_order);
1475579c 1220 assert(first_order > 0);
93d46c39 1221 for (i = last_order; i >= first_order; i--) {
4105056a 1222 unsigned long len;
1475579c 1223
4f6e90b7 1224 len = 1UL << (i - 1);
1475579c 1225 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1226
4d676753
MD
1227 /* Stop shrink if the resize target changes under us */
1228 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1229 break;
1230
1231 cmm_smp_wmb(); /* populate data before RCU size */
1232 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1233
1234 /*
1235 * We need to wait for all add operations to reach Q.S. (and
1236 * thus use the new table for lookups) before we can start
1237 * releasing the old dummy nodes. Otherwise their lookup will
1238 * return a logically removed node as insert position.
1239 */
1240 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1241 if (free_by_rcu)
1242 free(free_by_rcu);
4d676753 1243
21263e21 1244 /*
4105056a
MD
1245 * Set "removed" flag in dummy nodes about to be removed.
1246 * Unlink all now-logically-removed dummy node pointers.
1247 * Concurrent add/remove operation are helping us doing
1248 * the gc.
21263e21 1249 */
4105056a
MD
1250 remove_table(ht, i, len);
1251
0d14ceb2 1252 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1253
1254 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1255 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1256 break;
1257 }
0d14ceb2
LJ
1258
1259 if (free_by_rcu) {
1260 ht->cds_lfht_synchronize_rcu();
1261 free(free_by_rcu);
1262 }
1475579c
MD
1263}
1264
ff0d69de
LJ
1265static
1266void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
1267{
1268 struct _cds_lfht_node *prev, *node;
1269 unsigned long order, len, i, j;
1270
1271 ht->t.tbl[0] = calloc(1, sizeof(struct _cds_lfht_node));
1272 assert(ht->t.tbl[0]);
1273
1274 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1275 ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
1276 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1277
1278 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1279 len = 1UL << (order - 1);
1280 ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
1281 assert(ht->t.tbl[order]);
1282
1283 i = 0;
1284 prev = ht->t.tbl[i]->nodes;
1285 for (j = 0; j < len; j++) {
1286 if (j & (j - 1)) { /* Between power of 2 */
1287 prev++;
1288 } else if (j) { /* At each power of 2 */
1289 i++;
1290 prev = ht->t.tbl[i]->nodes;
1291 }
1292
1293 node = &ht->t.tbl[order]->nodes[j];
1294 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1295 order, j, j + len);
1296 node->next = prev->next;
1297 assert(is_dummy(node->next));
1298 node->reverse_hash = bit_reverse_ulong(j + len);
1299 prev->next = flag_dummy((struct cds_lfht_node *)node);
1300 }
1301 }
1302}
1303
7a9dcf9b 1304struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1305 cds_lfht_compare_fct compare_fct,
1306 unsigned long hash_seed,
1307 unsigned long init_size,
b8af5011 1308 int flags,
14044b37 1309 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1310 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1311 void (*cds_lfht_synchronize_rcu)(void),
1312 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1313 void (*cds_lfht_rcu_read_unlock)(void),
1314 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1315 void (*cds_lfht_rcu_thread_online)(void),
1316 void (*cds_lfht_rcu_register_thread)(void),
1317 void (*cds_lfht_rcu_unregister_thread)(void),
1318 pthread_attr_t *attr)
abc490a1 1319{
14044b37 1320 struct cds_lfht *ht;
24365af7 1321 unsigned long order;
abc490a1 1322
8129be4e 1323 /* init_size must be power of two */
49619ea0 1324 if (init_size && (init_size & (init_size - 1)))
8129be4e 1325 return NULL;
14044b37 1326 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1327 assert(ht);
abc490a1 1328 ht->hash_fct = hash_fct;
732ad076
MD
1329 ht->compare_fct = compare_fct;
1330 ht->hash_seed = hash_seed;
14044b37 1331 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1332 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1333 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1334 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1335 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1336 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1337 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1338 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1339 ht->resize_attr = attr;
df44348d 1340 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1341 /* this mutex should not nest in read-side C.S. */
1342 pthread_mutex_init(&ht->resize_mutex, NULL);
b8af5011 1343 ht->flags = flags;
ff0d69de 1344 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE));
93d46c39 1345 ht->t.resize_target = 1UL << order;
ff0d69de
LJ
1346 cds_lfht_create_dummy(ht, 1UL << order);
1347 ht->t.size = 1UL << order;
abc490a1
MD
1348 return ht;
1349}
1350
adc0de68
MD
1351void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1352 struct cds_lfht_iter *iter)
2ed95849 1353{
bb7b2f26 1354 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1355 struct _cds_lfht_node *lookup;
f4a9cc0b 1356 unsigned long hash, reverse_hash, size;
2ed95849 1357
732ad076 1358 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1359 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1360
4105056a 1361 size = rcu_dereference(ht->t.size);
f4a9cc0b 1362 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1363 dummy_node = (struct cds_lfht_node *) lookup;
1364 /* We can always skip the dummy node initially */
1365 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1366 node = clear_flag(node);
2ed95849 1367 for (;;) {
bb7b2f26 1368 if (unlikely(is_end(node))) {
96ad1112 1369 node = next = NULL;
abc490a1 1370 break;
bb7b2f26 1371 }
cc4fcb10 1372 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1373 node = next = NULL;
abc490a1 1374 break;
2ed95849 1375 }
1b81fe1a 1376 next = rcu_dereference(node->p.next);
adc0de68 1377 if (likely(!is_removed(next))
1b81fe1a 1378 && !is_dummy(next)
ae450da3 1379 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1380 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1381 break;
2ed95849 1382 }
1b81fe1a 1383 node = clear_flag(next);
2ed95849 1384 }
1b81fe1a 1385 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1386 iter->node = node;
1387 iter->next = next;
abc490a1 1388}
e0ba718a 1389
3883c0e5 1390void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1391{
adc0de68 1392 struct cds_lfht_node *node, *next;
a481e5ff
MD
1393 unsigned long reverse_hash;
1394 void *key;
1395 size_t key_len;
1396
adc0de68 1397 node = iter->node;
a481e5ff
MD
1398 reverse_hash = node->p.reverse_hash;
1399 key = node->key;
1400 key_len = node->key_len;
adc0de68 1401 next = iter->next;
a481e5ff
MD
1402 node = clear_flag(next);
1403
1404 for (;;) {
bb7b2f26 1405 if (unlikely(is_end(node))) {
96ad1112 1406 node = next = NULL;
a481e5ff 1407 break;
bb7b2f26 1408 }
a481e5ff 1409 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1410 node = next = NULL;
a481e5ff
MD
1411 break;
1412 }
1413 next = rcu_dereference(node->p.next);
adc0de68 1414 if (likely(!is_removed(next))
a481e5ff
MD
1415 && !is_dummy(next)
1416 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1417 break;
1418 }
1419 node = clear_flag(next);
1420 }
1421 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1422 iter->node = node;
1423 iter->next = next;
a481e5ff
MD
1424}
1425
4e9b9fbf
MD
1426void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1427{
1428 struct cds_lfht_node *node, *next;
1429
853395e1 1430 node = clear_flag(iter->next);
4e9b9fbf
MD
1431 for (;;) {
1432 if (unlikely(is_end(node))) {
1433 node = next = NULL;
1434 break;
1435 }
1436 next = rcu_dereference(node->p.next);
1437 if (likely(!is_removed(next))
1438 && !is_dummy(next)) {
1439 break;
1440 }
1441 node = clear_flag(next);
1442 }
1443 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1444 iter->node = node;
1445 iter->next = next;
1446}
1447
1448void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1449{
1450 struct _cds_lfht_node *lookup;
1451
1452 /*
1453 * Get next after first dummy node. The first dummy node is the
1454 * first node of the linked list.
1455 */
1456 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1457 iter->next = lookup->next;
4e9b9fbf
MD
1458 cds_lfht_next(ht, iter);
1459}
1460
14044b37 1461void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1462{
4105056a 1463 unsigned long hash, size;
ab7d5fc6 1464
49c2e2d6 1465 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1466 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1467
4105056a 1468 size = rcu_dereference(ht->t.size);
83beee94 1469 _cds_lfht_add(ht, size, node, NULL, 0);
4105056a 1470 ht_count_add(ht, size);
3eca1b8c
MD
1471}
1472
14044b37 1473struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1474 struct cds_lfht_node *node)
3eca1b8c 1475{
4105056a 1476 unsigned long hash, size;
83beee94 1477 struct cds_lfht_iter iter;
3eca1b8c 1478
49c2e2d6 1479 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1480 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1481
4105056a 1482 size = rcu_dereference(ht->t.size);
83beee94
MD
1483 _cds_lfht_add(ht, size, node, &iter, 0);
1484 if (iter.node == node)
4105056a 1485 ht_count_add(ht, size);
83beee94 1486 return iter.node;
2ed95849
MD
1487}
1488
9357c415 1489struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1490 struct cds_lfht_node *node)
1491{
1492 unsigned long hash, size;
83beee94 1493 struct cds_lfht_iter iter;
48ed1c18
MD
1494
1495 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1496 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1497
1498 size = rcu_dereference(ht->t.size);
83beee94
MD
1499 for (;;) {
1500 _cds_lfht_add(ht, size, node, &iter, 0);
1501 if (iter.node == node) {
1502 ht_count_add(ht, size);
1503 return NULL;
1504 }
1505
1506 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1507 return iter.node;
1508 }
48ed1c18
MD
1509}
1510
9357c415
MD
1511int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1512 struct cds_lfht_node *new_node)
1513{
1514 unsigned long size;
1515
1516 size = rcu_dereference(ht->t.size);
1517 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1518 new_node);
1519}
1520
1521int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1522{
4105056a 1523 unsigned long size;
df44348d 1524 int ret;
abc490a1 1525
4105056a 1526 size = rcu_dereference(ht->t.size);
9357c415 1527 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1528 if (!ret)
860d07e8 1529 ht_count_del(ht, size);
df44348d 1530 return ret;
2ed95849 1531}
ab7d5fc6 1532
abc490a1 1533static
14044b37 1534int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1535{
14044b37
MD
1536 struct cds_lfht_node *node;
1537 struct _cds_lfht_node *lookup;
4105056a 1538 unsigned long order, i, size;
674f7a69 1539
abc490a1 1540 /* Check that the table is empty */
4105056a 1541 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1542 node = (struct cds_lfht_node *) lookup;
abc490a1 1543 do {
1b81fe1a
MD
1544 node = clear_flag(node)->p.next;
1545 if (!is_dummy(node))
abc490a1 1546 return -EPERM;
273399de 1547 assert(!is_removed(node));
bb7b2f26 1548 } while (!is_end(node));
4105056a
MD
1549 /*
1550 * size accessed without rcu_dereference because hash table is
1551 * being destroyed.
1552 */
1553 size = ht->t.size;
abc490a1 1554 /* Internal sanity check: all nodes left should be dummy */
4105056a 1555 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1556 unsigned long len;
1557
1558 len = !order ? 1 : 1UL << (order - 1);
1559 for (i = 0; i < len; i++) {
f0c29ed7 1560 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1561 order, i,
4105056a
MD
1562 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1563 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1564 }
4105056a 1565 poison_free(ht->t.tbl[order]);
674f7a69 1566 }
abc490a1 1567 return 0;
674f7a69
MD
1568}
1569
1570/*
1571 * Should only be called when no more concurrent readers nor writers can
1572 * possibly access the table.
1573 */
b7d619b0 1574int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1575{
5e28c532
MD
1576 int ret;
1577
848d4088 1578 /* Wait for in-flight resize operations to complete */
24953e08
MD
1579 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1580 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1581 while (uatomic_read(&ht->in_progress_resize))
1582 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1583 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1584 if (ret)
1585 return ret;
df44348d 1586 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1587 if (attr)
1588 *attr = ht->resize_attr;
98808fb1 1589 poison_free(ht);
5e28c532 1590 return ret;
674f7a69
MD
1591}
1592
14044b37 1593void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1594 long *approx_before,
273399de 1595 unsigned long *count,
973e5e1b 1596 unsigned long *removed,
d933dd0e 1597 long *approx_after)
273399de 1598{
14044b37
MD
1599 struct cds_lfht_node *node, *next;
1600 struct _cds_lfht_node *lookup;
24365af7 1601 unsigned long nr_dummy = 0;
273399de 1602
7ed7682f 1603 *approx_before = 0;
973e5e1b
MD
1604 if (nr_cpus_mask >= 0) {
1605 int i;
1606
1607 for (i = 0; i < nr_cpus_mask + 1; i++) {
1608 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1609 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1610 }
1611 }
1612
273399de
MD
1613 *count = 0;
1614 *removed = 0;
1615
24365af7 1616 /* Count non-dummy nodes in the table */
4105056a 1617 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1618 node = (struct cds_lfht_node *) lookup;
273399de 1619 do {
cc4fcb10 1620 next = rcu_dereference(node->p.next);
b198f0fd 1621 if (is_removed(next)) {
973e5e1b
MD
1622 if (!is_dummy(next))
1623 (*removed)++;
1624 else
1625 (nr_dummy)++;
1b81fe1a 1626 } else if (!is_dummy(next))
273399de 1627 (*count)++;
24365af7
MD
1628 else
1629 (nr_dummy)++;
273399de 1630 node = clear_flag(next);
bb7b2f26 1631 } while (!is_end(node));
f0c29ed7 1632 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1633 *approx_after = 0;
973e5e1b
MD
1634 if (nr_cpus_mask >= 0) {
1635 int i;
1636
1637 for (i = 0; i < nr_cpus_mask + 1; i++) {
1638 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1639 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1640 }
1641 }
273399de
MD
1642}
1643
1475579c 1644/* called with resize mutex held */
abc490a1 1645static
4105056a 1646void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1647 unsigned long old_size, unsigned long new_size)
abc490a1 1648{
1475579c 1649 unsigned long old_order, new_order;
1475579c 1650
93d46c39
LJ
1651 old_order = get_count_order_ulong(old_size);
1652 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1653 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1654 old_size, old_order, new_size, new_order);
1475579c 1655 assert(new_size > old_size);
93d46c39 1656 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1657}
1658
1659/* called with resize mutex held */
1660static
4105056a 1661void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1662 unsigned long old_size, unsigned long new_size)
464a1ec9 1663{
1475579c 1664 unsigned long old_order, new_order;
464a1ec9 1665
cd95516d 1666 new_size = max(new_size, MIN_TABLE_SIZE);
93d46c39
LJ
1667 old_order = get_count_order_ulong(old_size);
1668 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1669 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1670 old_size, old_order, new_size, new_order);
1475579c 1671 assert(new_size < old_size);
1475579c 1672
4105056a 1673 /* Remove and unlink all dummy nodes to remove. */
93d46c39 1674 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1675}
1676
1475579c
MD
1677
1678/* called with resize mutex held */
1679static
1680void _do_cds_lfht_resize(struct cds_lfht *ht)
1681{
1682 unsigned long new_size, old_size;
4105056a
MD
1683
1684 /*
1685 * Resize table, re-do if the target size has changed under us.
1686 */
1687 do {
d2be3620
MD
1688 assert(uatomic_read(&ht->in_progress_resize));
1689 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1690 break;
4105056a
MD
1691 ht->t.resize_initiated = 1;
1692 old_size = ht->t.size;
1693 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1694 if (old_size < new_size)
1695 _do_cds_lfht_grow(ht, old_size, new_size);
1696 else if (old_size > new_size)
1697 _do_cds_lfht_shrink(ht, old_size, new_size);
1698 ht->t.resize_initiated = 0;
1699 /* write resize_initiated before read resize_target */
1700 cmm_smp_mb();
4d676753 1701 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1702}
1703
abc490a1 1704static
4105056a 1705unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1706 int growth_order)
464a1ec9 1707{
4105056a
MD
1708 return _uatomic_max(&ht->t.resize_target,
1709 size << growth_order);
464a1ec9
MD
1710}
1711
1475579c 1712static
4105056a 1713void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1714 unsigned long count)
1475579c 1715{
cd95516d 1716 count = max(count, MIN_TABLE_SIZE);
4105056a 1717 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1718}
1719
1720void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1721{
4105056a
MD
1722 resize_target_update_count(ht, new_size);
1723 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1724 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1725 pthread_mutex_lock(&ht->resize_mutex);
1726 _do_cds_lfht_resize(ht);
1727 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1728 ht->cds_lfht_rcu_thread_online();
abc490a1 1729}
464a1ec9 1730
abc490a1
MD
1731static
1732void do_resize_cb(struct rcu_head *head)
1733{
1734 struct rcu_resize_work *work =
1735 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1736 struct cds_lfht *ht = work->ht;
abc490a1 1737
5f511391 1738 ht->cds_lfht_rcu_thread_offline();
abc490a1 1739 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1740 _do_cds_lfht_resize(ht);
abc490a1 1741 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1742 ht->cds_lfht_rcu_thread_online();
98808fb1 1743 poison_free(work);
848d4088
MD
1744 cmm_smp_mb(); /* finish resize before decrement */
1745 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1746}
1747
abc490a1 1748static
4105056a 1749void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1750{
abc490a1 1751 struct rcu_resize_work *work;
f9830efd 1752 unsigned long target_size;
abc490a1 1753
4105056a
MD
1754 target_size = resize_target_update(ht, size, growth);
1755 /* Store resize_target before read resize_initiated */
1756 cmm_smp_mb();
1757 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1758 uatomic_inc(&ht->in_progress_resize);
59290e9d 1759 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1760 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1761 uatomic_dec(&ht->in_progress_resize);
59290e9d 1762 return;
ed35e6d8 1763 }
f9830efd
MD
1764 work = malloc(sizeof(*work));
1765 work->ht = ht;
14044b37 1766 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1767 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1768 }
ab7d5fc6 1769}
3171717f 1770
f8994aee
MD
1771#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1772
3171717f 1773static
4105056a 1774void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1775 unsigned long count)
1776{
1777 struct rcu_resize_work *work;
3171717f 1778
b8af5011
MD
1779 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1780 return;
4105056a
MD
1781 resize_target_update_count(ht, count);
1782 /* Store resize_target before read resize_initiated */
1783 cmm_smp_mb();
1784 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1785 uatomic_inc(&ht->in_progress_resize);
59290e9d 1786 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1787 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1788 uatomic_dec(&ht->in_progress_resize);
59290e9d 1789 return;
ed35e6d8 1790 }
3171717f
MD
1791 work = malloc(sizeof(*work));
1792 work->ht = ht;
1793 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1794 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1795 }
1796}
f8994aee
MD
1797
1798#endif
This page took 0.206603 seconds and 4 git commands to generate.