rculfhash: Cast to struct rcu_level * instead of void * (cleanup)
[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;
5488222b
LJ
244 unsigned long min_alloc_order;
245 unsigned long min_alloc_size;
732ad076 246 unsigned long hash_seed;
b8af5011 247 int flags;
5f511391
MD
248 /*
249 * We need to put the work threads offline (QSBR) when taking this
250 * mutex, because we use synchronize_rcu within this mutex critical
251 * section, which waits on read-side critical sections, and could
252 * therefore cause grace-period deadlock if we hold off RCU G.P.
253 * completion.
254 */
464a1ec9 255 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 256 unsigned int in_progress_resize, in_progress_destroy;
14044b37 257 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 258 void (*func)(struct rcu_head *head));
1475579c 259 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
260 void (*cds_lfht_rcu_read_lock)(void);
261 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
262 void (*cds_lfht_rcu_thread_offline)(void);
263 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
264 void (*cds_lfht_rcu_register_thread)(void);
265 void (*cds_lfht_rcu_unregister_thread)(void);
266 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 267 long count; /* global approximate item count */
df44348d 268 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
269};
270
abc490a1
MD
271struct rcu_resize_work {
272 struct rcu_head head;
14044b37 273 struct cds_lfht *ht;
abc490a1 274};
2ed95849 275
b7d619b0 276struct partition_resize_work {
1af6e26e 277 pthread_t thread_id;
b7d619b0
MD
278 struct cds_lfht *ht;
279 unsigned long i, start, len;
280 void (*fct)(struct cds_lfht *ht, unsigned long i,
281 unsigned long start, unsigned long len);
282};
283
76a73da8 284static
83beee94
MD
285void _cds_lfht_add(struct cds_lfht *ht,
286 unsigned long size,
287 struct cds_lfht_node *node,
288 struct cds_lfht_iter *unique_ret,
289 int dummy);
48ed1c18 290
abc490a1
MD
291/*
292 * Algorithm to reverse bits in a word by lookup table, extended to
293 * 64-bit words.
f9830efd 294 * Source:
abc490a1 295 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 296 * Originally from Public Domain.
abc490a1
MD
297 */
298
299static const uint8_t BitReverseTable256[256] =
2ed95849 300{
abc490a1
MD
301#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
302#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
303#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
304 R6(0), R6(2), R6(1), R6(3)
305};
306#undef R2
307#undef R4
308#undef R6
2ed95849 309
abc490a1
MD
310static
311uint8_t bit_reverse_u8(uint8_t v)
312{
313 return BitReverseTable256[v];
314}
ab7d5fc6 315
abc490a1
MD
316static __attribute__((unused))
317uint32_t bit_reverse_u32(uint32_t v)
318{
319 return ((uint32_t) bit_reverse_u8(v) << 24) |
320 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
321 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
322 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
323}
324
abc490a1
MD
325static __attribute__((unused))
326uint64_t bit_reverse_u64(uint64_t v)
2ed95849 327{
abc490a1
MD
328 return ((uint64_t) bit_reverse_u8(v) << 56) |
329 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
330 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
331 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
332 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
333 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
334 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
335 ((uint64_t) bit_reverse_u8(v >> 56));
336}
337
338static
339unsigned long bit_reverse_ulong(unsigned long v)
340{
341#if (CAA_BITS_PER_LONG == 32)
342 return bit_reverse_u32(v);
343#else
344 return bit_reverse_u64(v);
345#endif
346}
347
f9830efd 348/*
24365af7
MD
349 * fls: returns the position of the most significant bit.
350 * Returns 0 if no bit is set, else returns the position of the most
351 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 352 */
24365af7
MD
353#if defined(__i386) || defined(__x86_64)
354static inline
355unsigned int fls_u32(uint32_t x)
f9830efd 356{
24365af7
MD
357 int r;
358
359 asm("bsrl %1,%0\n\t"
360 "jnz 1f\n\t"
361 "movl $-1,%0\n\t"
362 "1:\n\t"
363 : "=r" (r) : "rm" (x));
364 return r + 1;
365}
366#define HAS_FLS_U32
367#endif
368
369#if defined(__x86_64)
370static inline
371unsigned int fls_u64(uint64_t x)
372{
373 long r;
374
375 asm("bsrq %1,%0\n\t"
376 "jnz 1f\n\t"
377 "movq $-1,%0\n\t"
378 "1:\n\t"
379 : "=r" (r) : "rm" (x));
380 return r + 1;
381}
382#define HAS_FLS_U64
383#endif
384
385#ifndef HAS_FLS_U64
386static __attribute__((unused))
387unsigned int fls_u64(uint64_t x)
388{
389 unsigned int r = 64;
390
391 if (!x)
392 return 0;
393
394 if (!(x & 0xFFFFFFFF00000000ULL)) {
395 x <<= 32;
396 r -= 32;
397 }
398 if (!(x & 0xFFFF000000000000ULL)) {
399 x <<= 16;
400 r -= 16;
401 }
402 if (!(x & 0xFF00000000000000ULL)) {
403 x <<= 8;
404 r -= 8;
405 }
406 if (!(x & 0xF000000000000000ULL)) {
407 x <<= 4;
408 r -= 4;
409 }
410 if (!(x & 0xC000000000000000ULL)) {
411 x <<= 2;
412 r -= 2;
413 }
414 if (!(x & 0x8000000000000000ULL)) {
415 x <<= 1;
416 r -= 1;
417 }
418 return r;
419}
420#endif
421
422#ifndef HAS_FLS_U32
423static __attribute__((unused))
424unsigned int fls_u32(uint32_t x)
425{
426 unsigned int r = 32;
f9830efd 427
24365af7
MD
428 if (!x)
429 return 0;
430 if (!(x & 0xFFFF0000U)) {
431 x <<= 16;
432 r -= 16;
433 }
434 if (!(x & 0xFF000000U)) {
435 x <<= 8;
436 r -= 8;
437 }
438 if (!(x & 0xF0000000U)) {
439 x <<= 4;
440 r -= 4;
441 }
442 if (!(x & 0xC0000000U)) {
443 x <<= 2;
444 r -= 2;
445 }
446 if (!(x & 0x80000000U)) {
447 x <<= 1;
448 r -= 1;
449 }
450 return r;
451}
452#endif
453
454unsigned int fls_ulong(unsigned long x)
f9830efd 455{
24365af7
MD
456#if (CAA_BITS_PER_lONG == 32)
457 return fls_u32(x);
458#else
459 return fls_u64(x);
460#endif
461}
f9830efd 462
920f8ef6
LJ
463/*
464 * Return the minimum order for which x <= (1UL << order).
465 * Return -1 if x is 0.
466 */
24365af7
MD
467int get_count_order_u32(uint32_t x)
468{
920f8ef6
LJ
469 if (!x)
470 return -1;
24365af7 471
920f8ef6 472 return fls_u32(x - 1);
24365af7
MD
473}
474
920f8ef6
LJ
475/*
476 * Return the minimum order for which x <= (1UL << order).
477 * Return -1 if x is 0.
478 */
24365af7
MD
479int get_count_order_ulong(unsigned long x)
480{
920f8ef6
LJ
481 if (!x)
482 return -1;
24365af7 483
920f8ef6 484 return fls_ulong(x - 1);
f9830efd
MD
485}
486
98808fb1
MD
487#ifdef POISON_FREE
488#define poison_free(ptr) \
489 do { \
490 memset(ptr, 0x42, sizeof(*(ptr))); \
491 free(ptr); \
492 } while (0)
493#else
494#define poison_free(ptr) free(ptr)
495#endif
496
f9830efd 497static
4105056a 498void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 499
df44348d
MD
500/*
501 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
502 * available, then we support hash table item accounting.
503 * In the unfortunate event the number of CPUs reported would be
504 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
505 */
df44348d
MD
506#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
507
f8994aee 508static
4105056a 509void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
510 unsigned long count);
511
df44348d
MD
512static long nr_cpus_mask = -1;
513
514static
515struct ht_items_count *alloc_per_cpu_items_count(void)
516{
517 struct ht_items_count *count;
518
519 switch (nr_cpus_mask) {
520 case -2:
521 return NULL;
522 case -1:
523 {
524 long maxcpus;
525
526 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
527 if (maxcpus <= 0) {
528 nr_cpus_mask = -2;
529 return NULL;
530 }
531 /*
532 * round up number of CPUs to next power of two, so we
533 * can use & for modulo.
534 */
535 maxcpus = 1UL << get_count_order_ulong(maxcpus);
536 nr_cpus_mask = maxcpus - 1;
537 }
538 /* Fall-through */
539 default:
540 return calloc(nr_cpus_mask + 1, sizeof(*count));
541 }
542}
543
544static
545void free_per_cpu_items_count(struct ht_items_count *count)
546{
98808fb1 547 poison_free(count);
df44348d
MD
548}
549
550static
551int ht_get_cpu(void)
552{
553 int cpu;
554
555 assert(nr_cpus_mask >= 0);
556 cpu = sched_getcpu();
557 if (unlikely(cpu < 0))
558 return cpu;
559 else
560 return cpu & nr_cpus_mask;
561}
562
563static
4105056a 564void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 565{
3171717f 566 unsigned long percpu_count;
df44348d
MD
567 int cpu;
568
569 if (unlikely(!ht->percpu_count))
3171717f 570 return;
df44348d
MD
571 cpu = ht_get_cpu();
572 if (unlikely(cpu < 0))
3171717f
MD
573 return;
574 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 575 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 576 long count;
df44348d
MD
577
578 dbg_printf("add percpu %lu\n", percpu_count);
579 count = uatomic_add_return(&ht->count,
580 1UL << COUNT_COMMIT_ORDER);
581 /* If power of 2 */
582 if (!(count & (count - 1))) {
4105056a 583 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 584 return;
e3ecefd6 585 dbg_printf("add set global %ld\n", count);
4105056a 586 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 587 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
588 }
589 }
590}
591
592static
860d07e8 593void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
594{
595 unsigned long percpu_count;
3171717f 596 int cpu;
df44348d 597
3171717f
MD
598 if (unlikely(!ht->percpu_count))
599 return;
600 cpu = ht_get_cpu();
601 if (unlikely(cpu < 0))
602 return;
80d90c06 603 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 604 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 605 long count;
df44348d 606
860d07e8 607 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 608 count = uatomic_add_return(&ht->count,
3171717f 609 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
610 /* If power of 2 */
611 if (!(count & (count - 1))) {
4105056a 612 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 613 return;
e3ecefd6
MD
614 dbg_printf("del set global %ld\n", count);
615 /*
c941bb9e 616 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
617 * certain threshold.
618 */
c941bb9e 619 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 620 return;
4105056a 621 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 622 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
623 }
624 }
625}
626
627#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
628
91452a6a 629static const long nr_cpus_mask = -2;
df44348d
MD
630
631static
632struct ht_items_count *alloc_per_cpu_items_count(void)
633{
634 return NULL;
635}
636
637static
638void free_per_cpu_items_count(struct ht_items_count *count)
639{
640}
641
642static
4105056a 643void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
644{
645}
646
647static
860d07e8 648void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
649{
650}
651
652#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
653
654
f9830efd 655static
4105056a 656void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 657{
f8994aee
MD
658 unsigned long count;
659
b8af5011
MD
660 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
661 return;
f8994aee
MD
662 count = uatomic_read(&ht->count);
663 /*
664 * Use bucket-local length for small table expand and for
665 * environments lacking per-cpu data support.
666 */
667 if (count >= (1UL << COUNT_COMMIT_ORDER))
668 return;
24365af7 669 if (chain_len > 100)
f0c29ed7 670 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 671 chain_len);
3390d470 672 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 673 cds_lfht_resize_lazy(ht, size,
01370f0b 674 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
675}
676
abc490a1 677static
14044b37 678struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 679{
14044b37 680 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
681}
682
683static
14044b37 684int is_removed(struct cds_lfht_node *node)
abc490a1 685{
d37166c6 686 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
687}
688
689static
14044b37 690struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 691{
14044b37 692 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
693}
694
f5596c94 695static
14044b37 696int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
697{
698 return ((unsigned long) node) & DUMMY_FLAG;
699}
700
701static
14044b37 702struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 703{
14044b37 704 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 705}
bb7b2f26
MD
706
707static
708struct cds_lfht_node *get_end(void)
709{
710 return (struct cds_lfht_node *) END_VALUE;
711}
712
713static
714int is_end(struct cds_lfht_node *node)
715{
716 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
717}
718
abc490a1 719static
f9830efd 720unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
721{
722 unsigned long old1, old2;
723
724 old1 = uatomic_read(ptr);
725 do {
726 old2 = old1;
727 if (old2 >= v)
f9830efd 728 return old2;
abc490a1 729 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 730 return v;
abc490a1
MD
731}
732
f4a9cc0b
LJ
733static
734struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
735 unsigned long hash)
736{
737 unsigned long index, order;
738
739 assert(size > 0);
740 index = hash & (size - 1);
ef6e6171
LJ
741
742 if (index < ht->min_alloc_size) {
743 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
744 hash, index);
745 return &ht->t.tbl[0]->nodes[index];
746 }
a4ea2223
LJ
747 /*
748 * equivalent to get_count_order_ulong(index + 1), but optimizes
749 * away the non-existing 0 special-case for
750 * get_count_order_ulong.
751 */
752 order = fls_ulong(index);
f4a9cc0b 753 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
ef6e6171
LJ
754 hash, index, order, index & ((1UL << (order - 1)) - 1));
755 return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
756}
757
273399de
MD
758/*
759 * Remove all logically deleted nodes from a bucket up to a certain node key.
760 */
761static
f9c80341 762void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 763{
14044b37 764 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 765
c90201ac
MD
766 assert(!is_dummy(dummy));
767 assert(!is_removed(dummy));
768 assert(!is_dummy(node));
769 assert(!is_removed(node));
273399de
MD
770 for (;;) {
771 iter_prev = dummy;
772 /* We can always skip the dummy node initially */
cc4fcb10 773 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 774 assert(!is_removed(iter));
cc4fcb10 775 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
776 /*
777 * We should never be called with dummy (start of chain)
778 * and logically removed node (end of path compression
779 * marker) being the actual same node. This would be a
780 * bug in the algorithm implementation.
781 */
782 assert(dummy != node);
273399de 783 for (;;) {
bb7b2f26 784 if (unlikely(is_end(iter)))
f9c80341 785 return;
76412f24 786 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 787 return;
cc4fcb10 788 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 789 if (likely(is_removed(next)))
273399de 790 break;
b453eae1 791 iter_prev = clear_flag(iter);
273399de
MD
792 iter = next;
793 }
b198f0fd 794 assert(!is_removed(iter));
f5596c94
MD
795 if (is_dummy(iter))
796 new_next = flag_dummy(clear_flag(next));
797 else
798 new_next = clear_flag(next);
799 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 800 }
f9c80341 801 return;
273399de
MD
802}
803
9357c415
MD
804static
805int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
806 struct cds_lfht_node *old_node,
3fb86f26 807 struct cds_lfht_node *old_next,
9357c415
MD
808 struct cds_lfht_node *new_node)
809{
3fb86f26 810 struct cds_lfht_node *dummy, *ret_next;
9357c415 811 struct _cds_lfht_node *lookup;
9357c415
MD
812
813 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 814 return -ENOENT;
9357c415
MD
815
816 assert(!is_removed(old_node));
817 assert(!is_dummy(old_node));
818 assert(!is_removed(new_node));
819 assert(!is_dummy(new_node));
820 assert(new_node != old_node);
3fb86f26 821 for (;;) {
9357c415 822 /* Insert after node to be replaced */
9357c415
MD
823 if (is_removed(old_next)) {
824 /*
825 * Too late, the old node has been removed under us
826 * between lookup and replace. Fail.
827 */
7801dadd 828 return -ENOENT;
9357c415
MD
829 }
830 assert(!is_dummy(old_next));
831 assert(new_node != clear_flag(old_next));
832 new_node->p.next = clear_flag(old_next);
833 /*
834 * Here is the whole trick for lock-free replace: we add
835 * the replacement node _after_ the node we want to
836 * replace by atomically setting its next pointer at the
837 * same time we set its removal flag. Given that
838 * the lookups/get next use an iterator aware of the
839 * next pointer, they will either skip the old node due
840 * to the removal flag and see the new node, or use
841 * the old node, but will not see the new one.
842 */
843 ret_next = uatomic_cmpxchg(&old_node->p.next,
844 old_next, flag_removed(new_node));
3fb86f26 845 if (ret_next == old_next)
7801dadd 846 break; /* We performed the replacement. */
3fb86f26
LJ
847 old_next = ret_next;
848 }
9357c415 849
9357c415
MD
850 /*
851 * Ensure that the old node is not visible to readers anymore:
852 * lookup for the node, and remove it (along with any other
853 * logically removed node) if found.
854 */
f4a9cc0b 855 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
856 dummy = (struct cds_lfht_node *) lookup;
857 _cds_lfht_gc_bucket(dummy, new_node);
7801dadd
LJ
858
859 assert(is_removed(rcu_dereference(old_node->p.next)));
860 return 0;
9357c415
MD
861}
862
83beee94
MD
863/*
864 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
865 * mode. A NULL unique_ret allows creation of duplicate keys.
866 */
abc490a1 867static
83beee94
MD
868void _cds_lfht_add(struct cds_lfht *ht,
869 unsigned long size,
870 struct cds_lfht_node *node,
871 struct cds_lfht_iter *unique_ret,
872 int dummy)
abc490a1 873{
14044b37 874 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 875 *return_node;
14044b37 876 struct _cds_lfht_node *lookup;
abc490a1 877
c90201ac
MD
878 assert(!is_dummy(node));
879 assert(!is_removed(node));
f4a9cc0b 880 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 881 for (;;) {
adc0de68 882 uint32_t chain_len = 0;
abc490a1 883
11519af6
MD
884 /*
885 * iter_prev points to the non-removed node prior to the
886 * insert location.
11519af6 887 */
14044b37 888 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 889 /* We can always skip the dummy node initially */
cc4fcb10
MD
890 iter = rcu_dereference(iter_prev->p.next);
891 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 892 for (;;) {
bb7b2f26 893 if (unlikely(is_end(iter)))
273399de 894 goto insert;
76412f24 895 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 896 goto insert;
238cc06e 897
194fdbd1
LJ
898 /* dummy node is the first node of the identical-hash-value chain */
899 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
900 goto insert;
238cc06e 901
cc4fcb10 902 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 903 if (unlikely(is_removed(next)))
9dba85be 904 goto gc_node;
238cc06e
LJ
905
906 /* uniquely add */
83beee94 907 if (unique_ret
1b81fe1a 908 && !is_dummy(next)
238cc06e
LJ
909 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
910 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
911
912 /*
913 * uniquely adding inserts the node as the first
914 * node of the identical-hash-value node chain.
915 *
916 * This semantic ensures no duplicated keys
917 * should ever be observable in the table
918 * (including observe one node by one node
919 * by forward iterations)
920 */
921 cds_lfht_next_duplicate(ht, &d_iter);
922 if (!d_iter.node)
923 goto insert;
924
925 *unique_ret = d_iter;
83beee94 926 return;
48ed1c18 927 }
238cc06e 928
11519af6 929 /* Only account for identical reverse hash once */
24365af7
MD
930 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
931 && !is_dummy(next))
4105056a 932 check_resize(ht, size, ++chain_len);
11519af6 933 iter_prev = clear_flag(iter);
273399de 934 iter = next;
abc490a1 935 }
48ed1c18 936
273399de 937 insert:
7ec59d3b 938 assert(node != clear_flag(iter));
11519af6 939 assert(!is_removed(iter_prev));
c90201ac 940 assert(!is_removed(iter));
f000907d 941 assert(iter_prev != node);
f9c80341 942 if (!dummy)
1b81fe1a 943 node->p.next = clear_flag(iter);
f9c80341
MD
944 else
945 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
946 if (is_dummy(iter))
947 new_node = flag_dummy(node);
948 else
949 new_node = node;
cc4fcb10 950 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 951 new_node) != iter) {
273399de 952 continue; /* retry */
48ed1c18 953 } else {
83beee94 954 return_node = node;
960c9e4f 955 goto end;
48ed1c18
MD
956 }
957
9dba85be
MD
958 gc_node:
959 assert(!is_removed(iter));
f5596c94
MD
960 if (is_dummy(iter))
961 new_next = flag_dummy(clear_flag(next));
962 else
963 new_next = clear_flag(next);
964 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 965 /* retry */
464a1ec9 966 }
9357c415 967end:
83beee94
MD
968 if (unique_ret) {
969 unique_ret->node = return_node;
970 /* unique_ret->next left unset, never used. */
971 }
abc490a1 972}
464a1ec9 973
abc490a1 974static
860d07e8 975int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 976 struct cds_lfht_node *node,
b198f0fd 977 int dummy_removal)
abc490a1 978{
14044b37
MD
979 struct cds_lfht_node *dummy, *next, *old;
980 struct _cds_lfht_node *lookup;
5e28c532 981
9357c415 982 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 983 return -ENOENT;
9357c415 984
7ec59d3b 985 /* logically delete the node */
c90201ac
MD
986 assert(!is_dummy(node));
987 assert(!is_removed(node));
cc4fcb10 988 old = rcu_dereference(node->p.next);
7ec59d3b 989 do {
48ed1c18
MD
990 struct cds_lfht_node *new_next;
991
7ec59d3b 992 next = old;
76412f24 993 if (unlikely(is_removed(next)))
743f9143 994 return -ENOENT;
1475579c
MD
995 if (dummy_removal)
996 assert(is_dummy(next));
997 else
998 assert(!is_dummy(next));
48ed1c18 999 new_next = flag_removed(next);
48ed1c18 1000 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b 1001 } while (old != next);
7ec59d3b 1002 /* We performed the (logical) deletion. */
7ec59d3b
MD
1003
1004 /*
1005 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1006 * the node, and remove it (along with any other logically removed node)
1007 * if found.
11519af6 1008 */
f4a9cc0b 1009 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 1010 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1011 _cds_lfht_gc_bucket(dummy, node);
743f9143
LJ
1012
1013 assert(is_removed(rcu_dereference(node->p.next)));
1014 return 0;
abc490a1 1015}
2ed95849 1016
b7d619b0
MD
1017static
1018void *partition_resize_thread(void *arg)
1019{
1020 struct partition_resize_work *work = arg;
1021
1022 work->ht->cds_lfht_rcu_register_thread();
1023 work->fct(work->ht, work->i, work->start, work->len);
1024 work->ht->cds_lfht_rcu_unregister_thread();
1025 return NULL;
1026}
1027
1028static
1029void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1030 unsigned long len,
1031 void (*fct)(struct cds_lfht *ht, unsigned long i,
1032 unsigned long start, unsigned long len))
1033{
1034 unsigned long partition_len;
1035 struct partition_resize_work *work;
6083a889
MD
1036 int thread, ret;
1037 unsigned long nr_threads;
b7d619b0 1038
6083a889
MD
1039 /*
1040 * Note: nr_cpus_mask + 1 is always power of 2.
1041 * We spawn just the number of threads we need to satisfy the minimum
1042 * partition size, up to the number of CPUs in the system.
1043 */
91452a6a
MD
1044 if (nr_cpus_mask > 0) {
1045 nr_threads = min(nr_cpus_mask + 1,
1046 len >> MIN_PARTITION_PER_THREAD_ORDER);
1047 } else {
1048 nr_threads = 1;
1049 }
6083a889
MD
1050 partition_len = len >> get_count_order_ulong(nr_threads);
1051 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1052 assert(work);
6083a889
MD
1053 for (thread = 0; thread < nr_threads; thread++) {
1054 work[thread].ht = ht;
1055 work[thread].i = i;
1056 work[thread].len = partition_len;
1057 work[thread].start = thread * partition_len;
1058 work[thread].fct = fct;
1af6e26e 1059 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1060 partition_resize_thread, &work[thread]);
b7d619b0
MD
1061 assert(!ret);
1062 }
6083a889 1063 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1064 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1065 assert(!ret);
1066 }
1067 free(work);
b7d619b0
MD
1068}
1069
e8de508e
MD
1070/*
1071 * Holding RCU read lock to protect _cds_lfht_add against memory
1072 * reclaim that could be performed by other call_rcu worker threads (ABA
1073 * problem).
9ee0fc9a 1074 *
b7d619b0 1075 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1076 * many worker threads, based on the number of CPUs available in the system.
1077 * This should therefore take care of not having the expand lagging behind too
1078 * many concurrent insertion threads by using the scheduler's ability to
1079 * schedule dummy node population fairly with insertions.
e8de508e 1080 */
4105056a 1081static
b7d619b0
MD
1082void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1083 unsigned long start, unsigned long len)
4105056a
MD
1084{
1085 unsigned long j;
1086
5488222b 1087 assert(i > ht->min_alloc_order);
4105056a 1088 ht->cds_lfht_rcu_read_lock();
b7d619b0 1089 for (j = start; j < start + len; j++) {
4105056a
MD
1090 struct cds_lfht_node *new_node =
1091 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1092
dc1da8f6 1093 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1094 i, j, (1UL << (i - 1)) + j);
dc1da8f6 1095 new_node->p.reverse_hash =
4f6e90b7
LJ
1096 bit_reverse_ulong((1UL << (i - 1)) + j);
1097 _cds_lfht_add(ht, 1UL << (i - 1),
83beee94 1098 new_node, NULL, 1);
4105056a
MD
1099 }
1100 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1101}
1102
1103static
1104void init_table_populate(struct cds_lfht *ht, unsigned long i,
1105 unsigned long len)
1106{
1107 assert(nr_cpus_mask != -1);
6083a889 1108 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1109 ht->cds_lfht_rcu_thread_online();
1110 init_table_populate_partition(ht, i, 0, len);
1111 ht->cds_lfht_rcu_thread_offline();
1112 return;
1113 }
1114 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1115}
1116
abc490a1 1117static
4105056a 1118void init_table(struct cds_lfht *ht,
93d46c39 1119 unsigned long first_order, unsigned long last_order)
24365af7 1120{
93d46c39 1121 unsigned long i;
24365af7 1122
93d46c39
LJ
1123 dbg_printf("init table: first_order %lu last_order %lu\n",
1124 first_order, last_order);
5488222b 1125 assert(first_order > ht->min_alloc_order);
93d46c39 1126 for (i = first_order; i <= last_order; i++) {
4105056a 1127 unsigned long len;
24365af7 1128
4f6e90b7 1129 len = 1UL << (i - 1);
f0c29ed7 1130 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1131
1132 /* Stop expand if the resize target changes under us */
4f6e90b7 1133 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1134 break;
1135
0d14ceb2 1136 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1137 assert(ht->t.tbl[i]);
4105056a 1138
4105056a 1139 /*
dc1da8f6
MD
1140 * Set all dummy nodes reverse hash values for a level and
1141 * link all dummy nodes into the table.
4105056a 1142 */
dc1da8f6 1143 init_table_populate(ht, i, len);
4105056a 1144
f9c80341
MD
1145 /*
1146 * Update table size.
1147 */
1148 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1149 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1150
4f6e90b7 1151 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1152 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1153 break;
1154 }
1155}
1156
e8de508e
MD
1157/*
1158 * Holding RCU read lock to protect _cds_lfht_remove against memory
1159 * reclaim that could be performed by other call_rcu worker threads (ABA
1160 * problem).
1161 * For a single level, we logically remove and garbage collect each node.
1162 *
1163 * As a design choice, we perform logical removal and garbage collection on a
1164 * node-per-node basis to simplify this algorithm. We also assume keeping good
1165 * cache locality of the operation would overweight possible performance gain
1166 * that could be achieved by batching garbage collection for multiple levels.
1167 * However, this would have to be justified by benchmarks.
1168 *
1169 * Concurrent removal and add operations are helping us perform garbage
1170 * collection of logically removed nodes. We guarantee that all logically
1171 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1172 * invoked to free a hole level of dummy nodes (after a grace period).
1173 *
1174 * Logical removal and garbage collection can therefore be done in batch or on a
1175 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1176 *
b7d619b0
MD
1177 * When we reach a certain length, we can split this removal over many worker
1178 * threads, based on the number of CPUs available in the system. This should
1179 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1180 * updater threads actively inserting into the hash table.
e8de508e 1181 */
4105056a 1182static
b7d619b0
MD
1183void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1184 unsigned long start, unsigned long len)
4105056a
MD
1185{
1186 unsigned long j;
1187
5488222b 1188 assert(i > ht->min_alloc_order);
4105056a 1189 ht->cds_lfht_rcu_read_lock();
b7d619b0 1190 for (j = start; j < start + len; j++) {
4105056a
MD
1191 struct cds_lfht_node *fini_node =
1192 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1193
1194 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1195 i, j, (1UL << (i - 1)) + j);
4105056a 1196 fini_node->p.reverse_hash =
4f6e90b7
LJ
1197 bit_reverse_ulong((1UL << (i - 1)) + j);
1198 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1199 }
4105056a 1200 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1201}
1202
1203static
1204void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1205{
1206
1207 assert(nr_cpus_mask != -1);
6083a889 1208 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1209 ht->cds_lfht_rcu_thread_online();
1210 remove_table_partition(ht, i, 0, len);
1211 ht->cds_lfht_rcu_thread_offline();
1212 return;
1213 }
1214 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1215}
1216
1475579c 1217static
4105056a 1218void fini_table(struct cds_lfht *ht,
93d46c39 1219 unsigned long first_order, unsigned long last_order)
1475579c 1220{
93d46c39 1221 long i;
0d14ceb2 1222 void *free_by_rcu = NULL;
1475579c 1223
93d46c39
LJ
1224 dbg_printf("fini table: first_order %lu last_order %lu\n",
1225 first_order, last_order);
5488222b 1226 assert(first_order > ht->min_alloc_order);
93d46c39 1227 for (i = last_order; i >= first_order; i--) {
4105056a 1228 unsigned long len;
1475579c 1229
4f6e90b7 1230 len = 1UL << (i - 1);
1475579c 1231 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1232
4d676753
MD
1233 /* Stop shrink if the resize target changes under us */
1234 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1235 break;
1236
1237 cmm_smp_wmb(); /* populate data before RCU size */
1238 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1239
1240 /*
1241 * We need to wait for all add operations to reach Q.S. (and
1242 * thus use the new table for lookups) before we can start
1243 * releasing the old dummy nodes. Otherwise their lookup will
1244 * return a logically removed node as insert position.
1245 */
1246 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1247 if (free_by_rcu)
1248 free(free_by_rcu);
4d676753 1249
21263e21 1250 /*
4105056a
MD
1251 * Set "removed" flag in dummy nodes about to be removed.
1252 * Unlink all now-logically-removed dummy node pointers.
1253 * Concurrent add/remove operation are helping us doing
1254 * the gc.
21263e21 1255 */
4105056a
MD
1256 remove_table(ht, i, len);
1257
0d14ceb2 1258 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1259
1260 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1261 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1262 break;
1263 }
0d14ceb2
LJ
1264
1265 if (free_by_rcu) {
1266 ht->cds_lfht_synchronize_rcu();
1267 free(free_by_rcu);
1268 }
1475579c
MD
1269}
1270
ff0d69de
LJ
1271static
1272void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
1273{
1274 struct _cds_lfht_node *prev, *node;
1275 unsigned long order, len, i, j;
1276
5488222b 1277 ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node));
ff0d69de
LJ
1278 assert(ht->t.tbl[0]);
1279
1280 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1281 ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
1282 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1283
1284 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1285 len = 1UL << (order - 1);
5488222b 1286 if (order <= ht->min_alloc_order) {
eb631bf2 1287 ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
5488222b
LJ
1288 } else {
1289 ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
1290 assert(ht->t.tbl[order]);
1291 }
ff0d69de
LJ
1292
1293 i = 0;
1294 prev = ht->t.tbl[i]->nodes;
1295 for (j = 0; j < len; j++) {
1296 if (j & (j - 1)) { /* Between power of 2 */
1297 prev++;
1298 } else if (j) { /* At each power of 2 */
1299 i++;
1300 prev = ht->t.tbl[i]->nodes;
1301 }
1302
1303 node = &ht->t.tbl[order]->nodes[j];
1304 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1305 order, j, j + len);
1306 node->next = prev->next;
1307 assert(is_dummy(node->next));
1308 node->reverse_hash = bit_reverse_ulong(j + len);
1309 prev->next = flag_dummy((struct cds_lfht_node *)node);
1310 }
1311 }
1312}
1313
7a9dcf9b 1314struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1315 cds_lfht_compare_fct compare_fct,
1316 unsigned long hash_seed,
1317 unsigned long init_size,
5488222b 1318 unsigned long min_alloc_size,
b8af5011 1319 int flags,
14044b37 1320 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1321 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1322 void (*cds_lfht_synchronize_rcu)(void),
1323 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1324 void (*cds_lfht_rcu_read_unlock)(void),
1325 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1326 void (*cds_lfht_rcu_thread_online)(void),
1327 void (*cds_lfht_rcu_register_thread)(void),
1328 void (*cds_lfht_rcu_unregister_thread)(void),
1329 pthread_attr_t *attr)
abc490a1 1330{
14044b37 1331 struct cds_lfht *ht;
24365af7 1332 unsigned long order;
abc490a1 1333
5488222b
LJ
1334 /* min_alloc_size must be power of two */
1335 if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
1336 return NULL;
8129be4e 1337 /* init_size must be power of two */
5488222b 1338 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1339 return NULL;
5488222b
LJ
1340 min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
1341 init_size = max(init_size, min_alloc_size);
14044b37 1342 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1343 assert(ht);
abc490a1 1344 ht->hash_fct = hash_fct;
732ad076
MD
1345 ht->compare_fct = compare_fct;
1346 ht->hash_seed = hash_seed;
14044b37 1347 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1348 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1349 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1350 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1351 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1352 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1353 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1354 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1355 ht->resize_attr = attr;
df44348d 1356 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1357 /* this mutex should not nest in read-side C.S. */
1358 pthread_mutex_init(&ht->resize_mutex, NULL);
b8af5011 1359 ht->flags = flags;
5488222b 1360 order = get_count_order_ulong(init_size);
93d46c39 1361 ht->t.resize_target = 1UL << order;
ff0d69de
LJ
1362 cds_lfht_create_dummy(ht, 1UL << order);
1363 ht->t.size = 1UL << order;
5488222b
LJ
1364 ht->min_alloc_size = min_alloc_size;
1365 ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
abc490a1
MD
1366 return ht;
1367}
1368
adc0de68
MD
1369void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1370 struct cds_lfht_iter *iter)
2ed95849 1371{
bb7b2f26 1372 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1373 struct _cds_lfht_node *lookup;
f4a9cc0b 1374 unsigned long hash, reverse_hash, size;
2ed95849 1375
732ad076 1376 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1377 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1378
4105056a 1379 size = rcu_dereference(ht->t.size);
f4a9cc0b 1380 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1381 dummy_node = (struct cds_lfht_node *) lookup;
1382 /* We can always skip the dummy node initially */
1383 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1384 node = clear_flag(node);
2ed95849 1385 for (;;) {
bb7b2f26 1386 if (unlikely(is_end(node))) {
96ad1112 1387 node = next = NULL;
abc490a1 1388 break;
bb7b2f26 1389 }
cc4fcb10 1390 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1391 node = next = NULL;
abc490a1 1392 break;
2ed95849 1393 }
1b81fe1a 1394 next = rcu_dereference(node->p.next);
adc0de68 1395 if (likely(!is_removed(next))
1b81fe1a 1396 && !is_dummy(next)
ae450da3 1397 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1398 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1399 break;
2ed95849 1400 }
1b81fe1a 1401 node = clear_flag(next);
2ed95849 1402 }
1b81fe1a 1403 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1404 iter->node = node;
1405 iter->next = next;
abc490a1 1406}
e0ba718a 1407
3883c0e5 1408void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1409{
adc0de68 1410 struct cds_lfht_node *node, *next;
a481e5ff
MD
1411 unsigned long reverse_hash;
1412 void *key;
1413 size_t key_len;
1414
adc0de68 1415 node = iter->node;
a481e5ff
MD
1416 reverse_hash = node->p.reverse_hash;
1417 key = node->key;
1418 key_len = node->key_len;
adc0de68 1419 next = iter->next;
a481e5ff
MD
1420 node = clear_flag(next);
1421
1422 for (;;) {
bb7b2f26 1423 if (unlikely(is_end(node))) {
96ad1112 1424 node = next = NULL;
a481e5ff 1425 break;
bb7b2f26 1426 }
a481e5ff 1427 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1428 node = next = NULL;
a481e5ff
MD
1429 break;
1430 }
1431 next = rcu_dereference(node->p.next);
adc0de68 1432 if (likely(!is_removed(next))
a481e5ff
MD
1433 && !is_dummy(next)
1434 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1435 break;
1436 }
1437 node = clear_flag(next);
1438 }
1439 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1440 iter->node = node;
1441 iter->next = next;
a481e5ff
MD
1442}
1443
4e9b9fbf
MD
1444void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1445{
1446 struct cds_lfht_node *node, *next;
1447
853395e1 1448 node = clear_flag(iter->next);
4e9b9fbf
MD
1449 for (;;) {
1450 if (unlikely(is_end(node))) {
1451 node = next = NULL;
1452 break;
1453 }
1454 next = rcu_dereference(node->p.next);
1455 if (likely(!is_removed(next))
1456 && !is_dummy(next)) {
1457 break;
1458 }
1459 node = clear_flag(next);
1460 }
1461 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1462 iter->node = node;
1463 iter->next = next;
1464}
1465
1466void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1467{
1468 struct _cds_lfht_node *lookup;
1469
1470 /*
1471 * Get next after first dummy node. The first dummy node is the
1472 * first node of the linked list.
1473 */
1474 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1475 iter->next = lookup->next;
4e9b9fbf
MD
1476 cds_lfht_next(ht, iter);
1477}
1478
14044b37 1479void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1480{
4105056a 1481 unsigned long hash, size;
ab7d5fc6 1482
49c2e2d6 1483 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1484 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1485
4105056a 1486 size = rcu_dereference(ht->t.size);
83beee94 1487 _cds_lfht_add(ht, size, node, NULL, 0);
4105056a 1488 ht_count_add(ht, size);
3eca1b8c
MD
1489}
1490
14044b37 1491struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1492 struct cds_lfht_node *node)
3eca1b8c 1493{
4105056a 1494 unsigned long hash, size;
83beee94 1495 struct cds_lfht_iter iter;
3eca1b8c 1496
49c2e2d6 1497 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1498 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1499
4105056a 1500 size = rcu_dereference(ht->t.size);
83beee94
MD
1501 _cds_lfht_add(ht, size, node, &iter, 0);
1502 if (iter.node == node)
4105056a 1503 ht_count_add(ht, size);
83beee94 1504 return iter.node;
2ed95849
MD
1505}
1506
9357c415 1507struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1508 struct cds_lfht_node *node)
1509{
1510 unsigned long hash, size;
83beee94 1511 struct cds_lfht_iter iter;
48ed1c18
MD
1512
1513 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1514 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1515
1516 size = rcu_dereference(ht->t.size);
83beee94
MD
1517 for (;;) {
1518 _cds_lfht_add(ht, size, node, &iter, 0);
1519 if (iter.node == node) {
1520 ht_count_add(ht, size);
1521 return NULL;
1522 }
1523
1524 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1525 return iter.node;
1526 }
48ed1c18
MD
1527}
1528
9357c415
MD
1529int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1530 struct cds_lfht_node *new_node)
1531{
1532 unsigned long size;
1533
1534 size = rcu_dereference(ht->t.size);
1535 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1536 new_node);
1537}
1538
1539int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1540{
4105056a 1541 unsigned long size;
df44348d 1542 int ret;
abc490a1 1543
4105056a 1544 size = rcu_dereference(ht->t.size);
9357c415 1545 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1546 if (!ret)
860d07e8 1547 ht_count_del(ht, size);
df44348d 1548 return ret;
2ed95849 1549}
ab7d5fc6 1550
abc490a1 1551static
14044b37 1552int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1553{
14044b37
MD
1554 struct cds_lfht_node *node;
1555 struct _cds_lfht_node *lookup;
4105056a 1556 unsigned long order, i, size;
674f7a69 1557
abc490a1 1558 /* Check that the table is empty */
4105056a 1559 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1560 node = (struct cds_lfht_node *) lookup;
abc490a1 1561 do {
1b81fe1a
MD
1562 node = clear_flag(node)->p.next;
1563 if (!is_dummy(node))
abc490a1 1564 return -EPERM;
273399de 1565 assert(!is_removed(node));
bb7b2f26 1566 } while (!is_end(node));
4105056a
MD
1567 /*
1568 * size accessed without rcu_dereference because hash table is
1569 * being destroyed.
1570 */
1571 size = ht->t.size;
abc490a1 1572 /* Internal sanity check: all nodes left should be dummy */
4105056a 1573 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1574 unsigned long len;
1575
1576 len = !order ? 1 : 1UL << (order - 1);
1577 for (i = 0; i < len; i++) {
f0c29ed7 1578 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1579 order, i,
4105056a
MD
1580 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1581 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1582 }
5488222b
LJ
1583
1584 if (order == ht->min_alloc_order)
1585 poison_free(ht->t.tbl[0]);
1586 else if (order > ht->min_alloc_order)
1587 poison_free(ht->t.tbl[order]);
1588 /* Nothing to delete for order < ht->min_alloc_order */
674f7a69 1589 }
abc490a1 1590 return 0;
674f7a69
MD
1591}
1592
1593/*
1594 * Should only be called when no more concurrent readers nor writers can
1595 * possibly access the table.
1596 */
b7d619b0 1597int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1598{
5e28c532
MD
1599 int ret;
1600
848d4088 1601 /* Wait for in-flight resize operations to complete */
24953e08
MD
1602 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1603 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1604 while (uatomic_read(&ht->in_progress_resize))
1605 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1606 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1607 if (ret)
1608 return ret;
df44348d 1609 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1610 if (attr)
1611 *attr = ht->resize_attr;
98808fb1 1612 poison_free(ht);
5e28c532 1613 return ret;
674f7a69
MD
1614}
1615
14044b37 1616void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1617 long *approx_before,
273399de 1618 unsigned long *count,
973e5e1b 1619 unsigned long *removed,
d933dd0e 1620 long *approx_after)
273399de 1621{
14044b37
MD
1622 struct cds_lfht_node *node, *next;
1623 struct _cds_lfht_node *lookup;
24365af7 1624 unsigned long nr_dummy = 0;
273399de 1625
7ed7682f 1626 *approx_before = 0;
973e5e1b
MD
1627 if (nr_cpus_mask >= 0) {
1628 int i;
1629
1630 for (i = 0; i < nr_cpus_mask + 1; i++) {
1631 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1632 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1633 }
1634 }
1635
273399de
MD
1636 *count = 0;
1637 *removed = 0;
1638
24365af7 1639 /* Count non-dummy nodes in the table */
4105056a 1640 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1641 node = (struct cds_lfht_node *) lookup;
273399de 1642 do {
cc4fcb10 1643 next = rcu_dereference(node->p.next);
b198f0fd 1644 if (is_removed(next)) {
973e5e1b
MD
1645 if (!is_dummy(next))
1646 (*removed)++;
1647 else
1648 (nr_dummy)++;
1b81fe1a 1649 } else if (!is_dummy(next))
273399de 1650 (*count)++;
24365af7
MD
1651 else
1652 (nr_dummy)++;
273399de 1653 node = clear_flag(next);
bb7b2f26 1654 } while (!is_end(node));
f0c29ed7 1655 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1656 *approx_after = 0;
973e5e1b
MD
1657 if (nr_cpus_mask >= 0) {
1658 int i;
1659
1660 for (i = 0; i < nr_cpus_mask + 1; i++) {
1661 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1662 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1663 }
1664 }
273399de
MD
1665}
1666
1475579c 1667/* called with resize mutex held */
abc490a1 1668static
4105056a 1669void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1670 unsigned long old_size, unsigned long new_size)
abc490a1 1671{
1475579c 1672 unsigned long old_order, new_order;
1475579c 1673
93d46c39
LJ
1674 old_order = get_count_order_ulong(old_size);
1675 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1676 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1677 old_size, old_order, new_size, new_order);
1475579c 1678 assert(new_size > old_size);
93d46c39 1679 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1680}
1681
1682/* called with resize mutex held */
1683static
4105056a 1684void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1685 unsigned long old_size, unsigned long new_size)
464a1ec9 1686{
1475579c 1687 unsigned long old_order, new_order;
464a1ec9 1688
5488222b 1689 new_size = max(new_size, ht->min_alloc_size);
93d46c39
LJ
1690 old_order = get_count_order_ulong(old_size);
1691 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1692 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1693 old_size, old_order, new_size, new_order);
1475579c 1694 assert(new_size < old_size);
1475579c 1695
4105056a 1696 /* Remove and unlink all dummy nodes to remove. */
93d46c39 1697 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1698}
1699
1475579c
MD
1700
1701/* called with resize mutex held */
1702static
1703void _do_cds_lfht_resize(struct cds_lfht *ht)
1704{
1705 unsigned long new_size, old_size;
4105056a
MD
1706
1707 /*
1708 * Resize table, re-do if the target size has changed under us.
1709 */
1710 do {
d2be3620
MD
1711 assert(uatomic_read(&ht->in_progress_resize));
1712 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1713 break;
4105056a
MD
1714 ht->t.resize_initiated = 1;
1715 old_size = ht->t.size;
1716 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1717 if (old_size < new_size)
1718 _do_cds_lfht_grow(ht, old_size, new_size);
1719 else if (old_size > new_size)
1720 _do_cds_lfht_shrink(ht, old_size, new_size);
1721 ht->t.resize_initiated = 0;
1722 /* write resize_initiated before read resize_target */
1723 cmm_smp_mb();
4d676753 1724 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1725}
1726
abc490a1 1727static
4105056a 1728unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1729 int growth_order)
464a1ec9 1730{
4105056a
MD
1731 return _uatomic_max(&ht->t.resize_target,
1732 size << growth_order);
464a1ec9
MD
1733}
1734
1475579c 1735static
4105056a 1736void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1737 unsigned long count)
1475579c 1738{
5488222b 1739 count = max(count, ht->min_alloc_size);
4105056a 1740 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1741}
1742
1743void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1744{
4105056a
MD
1745 resize_target_update_count(ht, new_size);
1746 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1747 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1748 pthread_mutex_lock(&ht->resize_mutex);
1749 _do_cds_lfht_resize(ht);
1750 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1751 ht->cds_lfht_rcu_thread_online();
abc490a1 1752}
464a1ec9 1753
abc490a1
MD
1754static
1755void do_resize_cb(struct rcu_head *head)
1756{
1757 struct rcu_resize_work *work =
1758 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1759 struct cds_lfht *ht = work->ht;
abc490a1 1760
5f511391 1761 ht->cds_lfht_rcu_thread_offline();
abc490a1 1762 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1763 _do_cds_lfht_resize(ht);
abc490a1 1764 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1765 ht->cds_lfht_rcu_thread_online();
98808fb1 1766 poison_free(work);
848d4088
MD
1767 cmm_smp_mb(); /* finish resize before decrement */
1768 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1769}
1770
abc490a1 1771static
4105056a 1772void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1773{
abc490a1 1774 struct rcu_resize_work *work;
f9830efd 1775 unsigned long target_size;
abc490a1 1776
4105056a
MD
1777 target_size = resize_target_update(ht, size, growth);
1778 /* Store resize_target before read resize_initiated */
1779 cmm_smp_mb();
1780 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1781 uatomic_inc(&ht->in_progress_resize);
59290e9d 1782 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1783 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1784 uatomic_dec(&ht->in_progress_resize);
59290e9d 1785 return;
ed35e6d8 1786 }
f9830efd
MD
1787 work = malloc(sizeof(*work));
1788 work->ht = ht;
14044b37 1789 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1790 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1791 }
ab7d5fc6 1792}
3171717f 1793
f8994aee
MD
1794#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1795
3171717f 1796static
4105056a 1797void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1798 unsigned long count)
1799{
1800 struct rcu_resize_work *work;
3171717f 1801
b8af5011
MD
1802 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1803 return;
4105056a
MD
1804 resize_target_update_count(ht, count);
1805 /* Store resize_target before read resize_initiated */
1806 cmm_smp_mb();
1807 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1808 uatomic_inc(&ht->in_progress_resize);
59290e9d 1809 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1810 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1811 uatomic_dec(&ht->in_progress_resize);
59290e9d 1812 return;
ed35e6d8 1813 }
3171717f
MD
1814 work = malloc(sizeof(*work));
1815 work->ht = ht;
1816 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1817 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1818 }
1819}
f8994aee
MD
1820
1821#endif
This page took 0.207359 seconds and 4 git commands to generate.