rculfhash: rename percpu_count to split_count
[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.
4c42f1b8 56 * - Split-counters are used to keep track of the number of
1475579c 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 175/*
4c42f1b8 176 * Split-counters lazily update the global counter each 1024
f8994aee
MD
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 */
4c42f1b8 268 struct ht_items_count *split_count; /* split 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 512static long nr_cpus_mask = -1;
4c42f1b8
LJ
513static long split_count_mask = -1;
514
515static void ht_init_nr_cpus_mask(void)
516{
517 long maxcpus;
518
519 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
520 if (maxcpus <= 0) {
521 nr_cpus_mask = -2;
522 return;
523 }
524 /*
525 * round up number of CPUs to next power of two, so we
526 * can use & for modulo.
527 */
528 maxcpus = 1UL << get_count_order_ulong(maxcpus);
529 nr_cpus_mask = maxcpus - 1;
530}
df44348d
MD
531
532static
4c42f1b8 533struct ht_items_count *alloc_split_items_count(void)
df44348d
MD
534{
535 struct ht_items_count *count;
536
4c42f1b8
LJ
537 if (nr_cpus_mask == -1) {
538 ht_init_nr_cpus_mask();
539 split_count_mask = nr_cpus_mask;
df44348d 540 }
4c42f1b8
LJ
541
542 if (split_count_mask < 0)
543 return NULL;
544 else
545 return calloc(split_count_mask + 1, sizeof(*count));
df44348d
MD
546}
547
548static
4c42f1b8 549void free_split_items_count(struct ht_items_count *count)
df44348d 550{
98808fb1 551 poison_free(count);
df44348d
MD
552}
553
554static
4c42f1b8 555int ht_get_split_count_index(void)
df44348d
MD
556{
557 int cpu;
558
4c42f1b8 559 assert(split_count_mask >= 0);
df44348d
MD
560 cpu = sched_getcpu();
561 if (unlikely(cpu < 0))
562 return cpu;
563 else
4c42f1b8 564 return cpu & split_count_mask;
df44348d
MD
565}
566
567static
4105056a 568void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 569{
4c42f1b8
LJ
570 unsigned long split_count;
571 int index;
df44348d 572
4c42f1b8 573 if (unlikely(!ht->split_count))
3171717f 574 return;
4c42f1b8
LJ
575 index = ht_get_split_count_index();
576 if (unlikely(index < 0))
3171717f 577 return;
4c42f1b8
LJ
578 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
579 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 580 long count;
df44348d 581
4c42f1b8 582 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
583 count = uatomic_add_return(&ht->count,
584 1UL << COUNT_COMMIT_ORDER);
585 /* If power of 2 */
586 if (!(count & (count - 1))) {
4105056a 587 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 588 return;
e3ecefd6 589 dbg_printf("add set global %ld\n", count);
4105056a 590 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 591 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
592 }
593 }
594}
595
596static
860d07e8 597void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d 598{
4c42f1b8
LJ
599 unsigned long split_count;
600 int index;
df44348d 601
4c42f1b8 602 if (unlikely(!ht->split_count))
3171717f 603 return;
4c42f1b8
LJ
604 index = ht_get_split_count_index();
605 if (unlikely(index < 0))
3171717f 606 return;
4c42f1b8
LJ
607 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
608 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 609 long count;
df44348d 610
4c42f1b8 611 dbg_printf("del split count %lu\n", split_count);
df44348d 612 count = uatomic_add_return(&ht->count,
3171717f 613 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
614 /* If power of 2 */
615 if (!(count & (count - 1))) {
4105056a 616 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 617 return;
e3ecefd6
MD
618 dbg_printf("del set global %ld\n", count);
619 /*
c941bb9e 620 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
621 * certain threshold.
622 */
4c42f1b8 623 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 624 return;
4105056a 625 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 626 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
627 }
628 }
629}
630
631#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
632
91452a6a 633static const long nr_cpus_mask = -2;
4c42f1b8 634static const long split_count_mask = -2;
df44348d
MD
635
636static
4c42f1b8 637struct ht_items_count *alloc_split_items_count(void)
df44348d
MD
638{
639 return NULL;
640}
641
642static
4c42f1b8 643void free_split_items_count(struct ht_items_count *count)
df44348d
MD
644{
645}
646
647static
4105056a 648void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
649{
650}
651
652static
860d07e8 653void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
654{
655}
656
657#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
658
659
f9830efd 660static
4105056a 661void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 662{
f8994aee
MD
663 unsigned long count;
664
b8af5011
MD
665 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
666 return;
f8994aee
MD
667 count = uatomic_read(&ht->count);
668 /*
669 * Use bucket-local length for small table expand and for
670 * environments lacking per-cpu data support.
671 */
672 if (count >= (1UL << COUNT_COMMIT_ORDER))
673 return;
24365af7 674 if (chain_len > 100)
f0c29ed7 675 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 676 chain_len);
3390d470 677 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 678 cds_lfht_resize_lazy(ht, size,
01370f0b 679 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
680}
681
abc490a1 682static
14044b37 683struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 684{
14044b37 685 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
686}
687
688static
14044b37 689int is_removed(struct cds_lfht_node *node)
abc490a1 690{
d37166c6 691 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
692}
693
694static
14044b37 695struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 696{
14044b37 697 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
698}
699
f5596c94 700static
14044b37 701int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
702{
703 return ((unsigned long) node) & DUMMY_FLAG;
704}
705
706static
14044b37 707struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 708{
14044b37 709 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 710}
bb7b2f26
MD
711
712static
713struct cds_lfht_node *get_end(void)
714{
715 return (struct cds_lfht_node *) END_VALUE;
716}
717
718static
719int is_end(struct cds_lfht_node *node)
720{
721 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
722}
723
abc490a1 724static
f9830efd 725unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
726{
727 unsigned long old1, old2;
728
729 old1 = uatomic_read(ptr);
730 do {
731 old2 = old1;
732 if (old2 >= v)
f9830efd 733 return old2;
abc490a1 734 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 735 return v;
abc490a1
MD
736}
737
f4a9cc0b
LJ
738static
739struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
740 unsigned long hash)
741{
742 unsigned long index, order;
743
744 assert(size > 0);
745 index = hash & (size - 1);
ef6e6171
LJ
746
747 if (index < ht->min_alloc_size) {
748 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
749 hash, index);
750 return &ht->t.tbl[0]->nodes[index];
751 }
a4ea2223
LJ
752 /*
753 * equivalent to get_count_order_ulong(index + 1), but optimizes
754 * away the non-existing 0 special-case for
755 * get_count_order_ulong.
756 */
757 order = fls_ulong(index);
f4a9cc0b 758 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
ef6e6171
LJ
759 hash, index, order, index & ((1UL << (order - 1)) - 1));
760 return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
761}
762
273399de
MD
763/*
764 * Remove all logically deleted nodes from a bucket up to a certain node key.
765 */
766static
f9c80341 767void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 768{
14044b37 769 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 770
c90201ac
MD
771 assert(!is_dummy(dummy));
772 assert(!is_removed(dummy));
773 assert(!is_dummy(node));
774 assert(!is_removed(node));
273399de
MD
775 for (;;) {
776 iter_prev = dummy;
777 /* We can always skip the dummy node initially */
cc4fcb10 778 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 779 assert(!is_removed(iter));
cc4fcb10 780 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
781 /*
782 * We should never be called with dummy (start of chain)
783 * and logically removed node (end of path compression
784 * marker) being the actual same node. This would be a
785 * bug in the algorithm implementation.
786 */
787 assert(dummy != node);
273399de 788 for (;;) {
bb7b2f26 789 if (unlikely(is_end(iter)))
f9c80341 790 return;
76412f24 791 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 792 return;
cc4fcb10 793 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 794 if (likely(is_removed(next)))
273399de 795 break;
b453eae1 796 iter_prev = clear_flag(iter);
273399de
MD
797 iter = next;
798 }
b198f0fd 799 assert(!is_removed(iter));
f5596c94
MD
800 if (is_dummy(iter))
801 new_next = flag_dummy(clear_flag(next));
802 else
803 new_next = clear_flag(next);
804 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 805 }
f9c80341 806 return;
273399de
MD
807}
808
9357c415
MD
809static
810int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
811 struct cds_lfht_node *old_node,
3fb86f26 812 struct cds_lfht_node *old_next,
9357c415
MD
813 struct cds_lfht_node *new_node)
814{
3fb86f26 815 struct cds_lfht_node *dummy, *ret_next;
9357c415 816 struct _cds_lfht_node *lookup;
9357c415
MD
817
818 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 819 return -ENOENT;
9357c415
MD
820
821 assert(!is_removed(old_node));
822 assert(!is_dummy(old_node));
823 assert(!is_removed(new_node));
824 assert(!is_dummy(new_node));
825 assert(new_node != old_node);
3fb86f26 826 for (;;) {
9357c415 827 /* Insert after node to be replaced */
9357c415
MD
828 if (is_removed(old_next)) {
829 /*
830 * Too late, the old node has been removed under us
831 * between lookup and replace. Fail.
832 */
7801dadd 833 return -ENOENT;
9357c415
MD
834 }
835 assert(!is_dummy(old_next));
836 assert(new_node != clear_flag(old_next));
837 new_node->p.next = clear_flag(old_next);
838 /*
839 * Here is the whole trick for lock-free replace: we add
840 * the replacement node _after_ the node we want to
841 * replace by atomically setting its next pointer at the
842 * same time we set its removal flag. Given that
843 * the lookups/get next use an iterator aware of the
844 * next pointer, they will either skip the old node due
845 * to the removal flag and see the new node, or use
846 * the old node, but will not see the new one.
847 */
848 ret_next = uatomic_cmpxchg(&old_node->p.next,
849 old_next, flag_removed(new_node));
3fb86f26 850 if (ret_next == old_next)
7801dadd 851 break; /* We performed the replacement. */
3fb86f26
LJ
852 old_next = ret_next;
853 }
9357c415 854
9357c415
MD
855 /*
856 * Ensure that the old node is not visible to readers anymore:
857 * lookup for the node, and remove it (along with any other
858 * logically removed node) if found.
859 */
f4a9cc0b 860 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
861 dummy = (struct cds_lfht_node *) lookup;
862 _cds_lfht_gc_bucket(dummy, new_node);
7801dadd
LJ
863
864 assert(is_removed(rcu_dereference(old_node->p.next)));
865 return 0;
9357c415
MD
866}
867
83beee94
MD
868/*
869 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
870 * mode. A NULL unique_ret allows creation of duplicate keys.
871 */
abc490a1 872static
83beee94
MD
873void _cds_lfht_add(struct cds_lfht *ht,
874 unsigned long size,
875 struct cds_lfht_node *node,
876 struct cds_lfht_iter *unique_ret,
877 int dummy)
abc490a1 878{
14044b37 879 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 880 *return_node;
14044b37 881 struct _cds_lfht_node *lookup;
abc490a1 882
c90201ac
MD
883 assert(!is_dummy(node));
884 assert(!is_removed(node));
f4a9cc0b 885 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 886 for (;;) {
adc0de68 887 uint32_t chain_len = 0;
abc490a1 888
11519af6
MD
889 /*
890 * iter_prev points to the non-removed node prior to the
891 * insert location.
11519af6 892 */
14044b37 893 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 894 /* We can always skip the dummy node initially */
cc4fcb10
MD
895 iter = rcu_dereference(iter_prev->p.next);
896 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 897 for (;;) {
bb7b2f26 898 if (unlikely(is_end(iter)))
273399de 899 goto insert;
76412f24 900 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 901 goto insert;
238cc06e 902
194fdbd1
LJ
903 /* dummy node is the first node of the identical-hash-value chain */
904 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
905 goto insert;
238cc06e 906
cc4fcb10 907 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 908 if (unlikely(is_removed(next)))
9dba85be 909 goto gc_node;
238cc06e
LJ
910
911 /* uniquely add */
83beee94 912 if (unique_ret
1b81fe1a 913 && !is_dummy(next)
238cc06e
LJ
914 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
915 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
916
917 /*
918 * uniquely adding inserts the node as the first
919 * node of the identical-hash-value node chain.
920 *
921 * This semantic ensures no duplicated keys
922 * should ever be observable in the table
923 * (including observe one node by one node
924 * by forward iterations)
925 */
926 cds_lfht_next_duplicate(ht, &d_iter);
927 if (!d_iter.node)
928 goto insert;
929
930 *unique_ret = d_iter;
83beee94 931 return;
48ed1c18 932 }
238cc06e 933
11519af6 934 /* Only account for identical reverse hash once */
24365af7
MD
935 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
936 && !is_dummy(next))
4105056a 937 check_resize(ht, size, ++chain_len);
11519af6 938 iter_prev = clear_flag(iter);
273399de 939 iter = next;
abc490a1 940 }
48ed1c18 941
273399de 942 insert:
7ec59d3b 943 assert(node != clear_flag(iter));
11519af6 944 assert(!is_removed(iter_prev));
c90201ac 945 assert(!is_removed(iter));
f000907d 946 assert(iter_prev != node);
f9c80341 947 if (!dummy)
1b81fe1a 948 node->p.next = clear_flag(iter);
f9c80341
MD
949 else
950 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
951 if (is_dummy(iter))
952 new_node = flag_dummy(node);
953 else
954 new_node = node;
cc4fcb10 955 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 956 new_node) != iter) {
273399de 957 continue; /* retry */
48ed1c18 958 } else {
83beee94 959 return_node = node;
960c9e4f 960 goto end;
48ed1c18
MD
961 }
962
9dba85be
MD
963 gc_node:
964 assert(!is_removed(iter));
f5596c94
MD
965 if (is_dummy(iter))
966 new_next = flag_dummy(clear_flag(next));
967 else
968 new_next = clear_flag(next);
969 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 970 /* retry */
464a1ec9 971 }
9357c415 972end:
83beee94
MD
973 if (unique_ret) {
974 unique_ret->node = return_node;
975 /* unique_ret->next left unset, never used. */
976 }
abc490a1 977}
464a1ec9 978
abc490a1 979static
860d07e8 980int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 981 struct cds_lfht_node *node,
b198f0fd 982 int dummy_removal)
abc490a1 983{
14044b37
MD
984 struct cds_lfht_node *dummy, *next, *old;
985 struct _cds_lfht_node *lookup;
5e28c532 986
9357c415 987 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 988 return -ENOENT;
9357c415 989
7ec59d3b 990 /* logically delete the node */
c90201ac
MD
991 assert(!is_dummy(node));
992 assert(!is_removed(node));
cc4fcb10 993 old = rcu_dereference(node->p.next);
7ec59d3b 994 do {
48ed1c18
MD
995 struct cds_lfht_node *new_next;
996
7ec59d3b 997 next = old;
76412f24 998 if (unlikely(is_removed(next)))
743f9143 999 return -ENOENT;
1475579c
MD
1000 if (dummy_removal)
1001 assert(is_dummy(next));
1002 else
1003 assert(!is_dummy(next));
48ed1c18 1004 new_next = flag_removed(next);
48ed1c18 1005 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b 1006 } while (old != next);
7ec59d3b 1007 /* We performed the (logical) deletion. */
7ec59d3b
MD
1008
1009 /*
1010 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1011 * the node, and remove it (along with any other logically removed node)
1012 * if found.
11519af6 1013 */
f4a9cc0b 1014 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 1015 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1016 _cds_lfht_gc_bucket(dummy, node);
743f9143
LJ
1017
1018 assert(is_removed(rcu_dereference(node->p.next)));
1019 return 0;
abc490a1 1020}
2ed95849 1021
b7d619b0
MD
1022static
1023void *partition_resize_thread(void *arg)
1024{
1025 struct partition_resize_work *work = arg;
1026
1027 work->ht->cds_lfht_rcu_register_thread();
1028 work->fct(work->ht, work->i, work->start, work->len);
1029 work->ht->cds_lfht_rcu_unregister_thread();
1030 return NULL;
1031}
1032
1033static
1034void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1035 unsigned long len,
1036 void (*fct)(struct cds_lfht *ht, unsigned long i,
1037 unsigned long start, unsigned long len))
1038{
1039 unsigned long partition_len;
1040 struct partition_resize_work *work;
6083a889
MD
1041 int thread, ret;
1042 unsigned long nr_threads;
b7d619b0 1043
6083a889
MD
1044 /*
1045 * Note: nr_cpus_mask + 1 is always power of 2.
1046 * We spawn just the number of threads we need to satisfy the minimum
1047 * partition size, up to the number of CPUs in the system.
1048 */
91452a6a
MD
1049 if (nr_cpus_mask > 0) {
1050 nr_threads = min(nr_cpus_mask + 1,
1051 len >> MIN_PARTITION_PER_THREAD_ORDER);
1052 } else {
1053 nr_threads = 1;
1054 }
6083a889
MD
1055 partition_len = len >> get_count_order_ulong(nr_threads);
1056 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1057 assert(work);
6083a889
MD
1058 for (thread = 0; thread < nr_threads; thread++) {
1059 work[thread].ht = ht;
1060 work[thread].i = i;
1061 work[thread].len = partition_len;
1062 work[thread].start = thread * partition_len;
1063 work[thread].fct = fct;
1af6e26e 1064 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1065 partition_resize_thread, &work[thread]);
b7d619b0
MD
1066 assert(!ret);
1067 }
6083a889 1068 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1069 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1070 assert(!ret);
1071 }
1072 free(work);
b7d619b0
MD
1073}
1074
e8de508e
MD
1075/*
1076 * Holding RCU read lock to protect _cds_lfht_add against memory
1077 * reclaim that could be performed by other call_rcu worker threads (ABA
1078 * problem).
9ee0fc9a 1079 *
b7d619b0 1080 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1081 * many worker threads, based on the number of CPUs available in the system.
1082 * This should therefore take care of not having the expand lagging behind too
1083 * many concurrent insertion threads by using the scheduler's ability to
1084 * schedule dummy node population fairly with insertions.
e8de508e 1085 */
4105056a 1086static
b7d619b0
MD
1087void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1088 unsigned long start, unsigned long len)
4105056a
MD
1089{
1090 unsigned long j;
1091
5488222b 1092 assert(i > ht->min_alloc_order);
4105056a 1093 ht->cds_lfht_rcu_read_lock();
b7d619b0 1094 for (j = start; j < start + len; j++) {
4105056a
MD
1095 struct cds_lfht_node *new_node =
1096 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1097
dc1da8f6 1098 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1099 i, j, (1UL << (i - 1)) + j);
dc1da8f6 1100 new_node->p.reverse_hash =
4f6e90b7
LJ
1101 bit_reverse_ulong((1UL << (i - 1)) + j);
1102 _cds_lfht_add(ht, 1UL << (i - 1),
83beee94 1103 new_node, NULL, 1);
4105056a
MD
1104 }
1105 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1106}
1107
1108static
1109void init_table_populate(struct cds_lfht *ht, unsigned long i,
1110 unsigned long len)
1111{
1112 assert(nr_cpus_mask != -1);
6083a889 1113 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1114 ht->cds_lfht_rcu_thread_online();
1115 init_table_populate_partition(ht, i, 0, len);
1116 ht->cds_lfht_rcu_thread_offline();
1117 return;
1118 }
1119 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1120}
1121
abc490a1 1122static
4105056a 1123void init_table(struct cds_lfht *ht,
93d46c39 1124 unsigned long first_order, unsigned long last_order)
24365af7 1125{
93d46c39 1126 unsigned long i;
24365af7 1127
93d46c39
LJ
1128 dbg_printf("init table: first_order %lu last_order %lu\n",
1129 first_order, last_order);
5488222b 1130 assert(first_order > ht->min_alloc_order);
93d46c39 1131 for (i = first_order; i <= last_order; i++) {
4105056a 1132 unsigned long len;
24365af7 1133
4f6e90b7 1134 len = 1UL << (i - 1);
f0c29ed7 1135 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1136
1137 /* Stop expand if the resize target changes under us */
4f6e90b7 1138 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1139 break;
1140
0d14ceb2 1141 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1142 assert(ht->t.tbl[i]);
4105056a 1143
4105056a 1144 /*
dc1da8f6
MD
1145 * Set all dummy nodes reverse hash values for a level and
1146 * link all dummy nodes into the table.
4105056a 1147 */
dc1da8f6 1148 init_table_populate(ht, i, len);
4105056a 1149
f9c80341
MD
1150 /*
1151 * Update table size.
1152 */
1153 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1154 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1155
4f6e90b7 1156 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1157 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1158 break;
1159 }
1160}
1161
e8de508e
MD
1162/*
1163 * Holding RCU read lock to protect _cds_lfht_remove against memory
1164 * reclaim that could be performed by other call_rcu worker threads (ABA
1165 * problem).
1166 * For a single level, we logically remove and garbage collect each node.
1167 *
1168 * As a design choice, we perform logical removal and garbage collection on a
1169 * node-per-node basis to simplify this algorithm. We also assume keeping good
1170 * cache locality of the operation would overweight possible performance gain
1171 * that could be achieved by batching garbage collection for multiple levels.
1172 * However, this would have to be justified by benchmarks.
1173 *
1174 * Concurrent removal and add operations are helping us perform garbage
1175 * collection of logically removed nodes. We guarantee that all logically
1176 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1177 * invoked to free a hole level of dummy nodes (after a grace period).
1178 *
1179 * Logical removal and garbage collection can therefore be done in batch or on a
1180 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1181 *
b7d619b0
MD
1182 * When we reach a certain length, we can split this removal over many worker
1183 * threads, based on the number of CPUs available in the system. This should
1184 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1185 * updater threads actively inserting into the hash table.
e8de508e 1186 */
4105056a 1187static
b7d619b0
MD
1188void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1189 unsigned long start, unsigned long len)
4105056a
MD
1190{
1191 unsigned long j;
1192
5488222b 1193 assert(i > ht->min_alloc_order);
4105056a 1194 ht->cds_lfht_rcu_read_lock();
b7d619b0 1195 for (j = start; j < start + len; j++) {
4105056a
MD
1196 struct cds_lfht_node *fini_node =
1197 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1198
1199 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1200 i, j, (1UL << (i - 1)) + j);
4105056a 1201 fini_node->p.reverse_hash =
4f6e90b7
LJ
1202 bit_reverse_ulong((1UL << (i - 1)) + j);
1203 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1204 }
4105056a 1205 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1206}
1207
1208static
1209void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1210{
1211
1212 assert(nr_cpus_mask != -1);
6083a889 1213 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1214 ht->cds_lfht_rcu_thread_online();
1215 remove_table_partition(ht, i, 0, len);
1216 ht->cds_lfht_rcu_thread_offline();
1217 return;
1218 }
1219 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1220}
1221
1475579c 1222static
4105056a 1223void fini_table(struct cds_lfht *ht,
93d46c39 1224 unsigned long first_order, unsigned long last_order)
1475579c 1225{
93d46c39 1226 long i;
0d14ceb2 1227 void *free_by_rcu = NULL;
1475579c 1228
93d46c39
LJ
1229 dbg_printf("fini table: first_order %lu last_order %lu\n",
1230 first_order, last_order);
5488222b 1231 assert(first_order > ht->min_alloc_order);
93d46c39 1232 for (i = last_order; i >= first_order; i--) {
4105056a 1233 unsigned long len;
1475579c 1234
4f6e90b7 1235 len = 1UL << (i - 1);
1475579c 1236 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1237
4d676753
MD
1238 /* Stop shrink if the resize target changes under us */
1239 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1240 break;
1241
1242 cmm_smp_wmb(); /* populate data before RCU size */
1243 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1244
1245 /*
1246 * We need to wait for all add operations to reach Q.S. (and
1247 * thus use the new table for lookups) before we can start
1248 * releasing the old dummy nodes. Otherwise their lookup will
1249 * return a logically removed node as insert position.
1250 */
1251 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1252 if (free_by_rcu)
1253 free(free_by_rcu);
4d676753 1254
21263e21 1255 /*
4105056a
MD
1256 * Set "removed" flag in dummy nodes about to be removed.
1257 * Unlink all now-logically-removed dummy node pointers.
1258 * Concurrent add/remove operation are helping us doing
1259 * the gc.
21263e21 1260 */
4105056a
MD
1261 remove_table(ht, i, len);
1262
0d14ceb2 1263 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1264
1265 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1266 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1267 break;
1268 }
0d14ceb2
LJ
1269
1270 if (free_by_rcu) {
1271 ht->cds_lfht_synchronize_rcu();
1272 free(free_by_rcu);
1273 }
1475579c
MD
1274}
1275
ff0d69de
LJ
1276static
1277void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
1278{
1279 struct _cds_lfht_node *prev, *node;
1280 unsigned long order, len, i, j;
1281
5488222b 1282 ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node));
ff0d69de
LJ
1283 assert(ht->t.tbl[0]);
1284
1285 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1286 ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
1287 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1288
1289 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1290 len = 1UL << (order - 1);
5488222b 1291 if (order <= ht->min_alloc_order) {
eb631bf2 1292 ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
5488222b
LJ
1293 } else {
1294 ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
1295 assert(ht->t.tbl[order]);
1296 }
ff0d69de
LJ
1297
1298 i = 0;
1299 prev = ht->t.tbl[i]->nodes;
1300 for (j = 0; j < len; j++) {
1301 if (j & (j - 1)) { /* Between power of 2 */
1302 prev++;
1303 } else if (j) { /* At each power of 2 */
1304 i++;
1305 prev = ht->t.tbl[i]->nodes;
1306 }
1307
1308 node = &ht->t.tbl[order]->nodes[j];
1309 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1310 order, j, j + len);
1311 node->next = prev->next;
1312 assert(is_dummy(node->next));
1313 node->reverse_hash = bit_reverse_ulong(j + len);
1314 prev->next = flag_dummy((struct cds_lfht_node *)node);
1315 }
1316 }
1317}
1318
7a9dcf9b 1319struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1320 cds_lfht_compare_fct compare_fct,
1321 unsigned long hash_seed,
1322 unsigned long init_size,
5488222b 1323 unsigned long min_alloc_size,
b8af5011 1324 int flags,
14044b37 1325 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1326 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1327 void (*cds_lfht_synchronize_rcu)(void),
1328 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1329 void (*cds_lfht_rcu_read_unlock)(void),
1330 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1331 void (*cds_lfht_rcu_thread_online)(void),
1332 void (*cds_lfht_rcu_register_thread)(void),
1333 void (*cds_lfht_rcu_unregister_thread)(void),
1334 pthread_attr_t *attr)
abc490a1 1335{
14044b37 1336 struct cds_lfht *ht;
24365af7 1337 unsigned long order;
abc490a1 1338
5488222b
LJ
1339 /* min_alloc_size must be power of two */
1340 if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
1341 return NULL;
8129be4e 1342 /* init_size must be power of two */
5488222b 1343 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1344 return NULL;
5488222b
LJ
1345 min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
1346 init_size = max(init_size, min_alloc_size);
14044b37 1347 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1348 assert(ht);
abc490a1 1349 ht->hash_fct = hash_fct;
732ad076
MD
1350 ht->compare_fct = compare_fct;
1351 ht->hash_seed = hash_seed;
14044b37 1352 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1353 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1354 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1355 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1356 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1357 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1358 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1359 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1360 ht->resize_attr = attr;
4c42f1b8 1361 ht->split_count = alloc_split_items_count();
abc490a1
MD
1362 /* this mutex should not nest in read-side C.S. */
1363 pthread_mutex_init(&ht->resize_mutex, NULL);
b8af5011 1364 ht->flags = flags;
5488222b 1365 order = get_count_order_ulong(init_size);
93d46c39 1366 ht->t.resize_target = 1UL << order;
ff0d69de
LJ
1367 cds_lfht_create_dummy(ht, 1UL << order);
1368 ht->t.size = 1UL << order;
5488222b
LJ
1369 ht->min_alloc_size = min_alloc_size;
1370 ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
abc490a1
MD
1371 return ht;
1372}
1373
adc0de68
MD
1374void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1375 struct cds_lfht_iter *iter)
2ed95849 1376{
bb7b2f26 1377 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1378 struct _cds_lfht_node *lookup;
f4a9cc0b 1379 unsigned long hash, reverse_hash, size;
2ed95849 1380
732ad076 1381 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1382 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1383
4105056a 1384 size = rcu_dereference(ht->t.size);
f4a9cc0b 1385 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1386 dummy_node = (struct cds_lfht_node *) lookup;
1387 /* We can always skip the dummy node initially */
1388 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1389 node = clear_flag(node);
2ed95849 1390 for (;;) {
bb7b2f26 1391 if (unlikely(is_end(node))) {
96ad1112 1392 node = next = NULL;
abc490a1 1393 break;
bb7b2f26 1394 }
cc4fcb10 1395 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1396 node = next = NULL;
abc490a1 1397 break;
2ed95849 1398 }
1b81fe1a 1399 next = rcu_dereference(node->p.next);
adc0de68 1400 if (likely(!is_removed(next))
1b81fe1a 1401 && !is_dummy(next)
ae450da3 1402 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1403 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1404 break;
2ed95849 1405 }
1b81fe1a 1406 node = clear_flag(next);
2ed95849 1407 }
1b81fe1a 1408 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1409 iter->node = node;
1410 iter->next = next;
abc490a1 1411}
e0ba718a 1412
3883c0e5 1413void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1414{
adc0de68 1415 struct cds_lfht_node *node, *next;
a481e5ff
MD
1416 unsigned long reverse_hash;
1417 void *key;
1418 size_t key_len;
1419
adc0de68 1420 node = iter->node;
a481e5ff
MD
1421 reverse_hash = node->p.reverse_hash;
1422 key = node->key;
1423 key_len = node->key_len;
adc0de68 1424 next = iter->next;
a481e5ff
MD
1425 node = clear_flag(next);
1426
1427 for (;;) {
bb7b2f26 1428 if (unlikely(is_end(node))) {
96ad1112 1429 node = next = NULL;
a481e5ff 1430 break;
bb7b2f26 1431 }
a481e5ff 1432 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1433 node = next = NULL;
a481e5ff
MD
1434 break;
1435 }
1436 next = rcu_dereference(node->p.next);
adc0de68 1437 if (likely(!is_removed(next))
a481e5ff
MD
1438 && !is_dummy(next)
1439 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1440 break;
1441 }
1442 node = clear_flag(next);
1443 }
1444 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1445 iter->node = node;
1446 iter->next = next;
a481e5ff
MD
1447}
1448
4e9b9fbf
MD
1449void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1450{
1451 struct cds_lfht_node *node, *next;
1452
853395e1 1453 node = clear_flag(iter->next);
4e9b9fbf
MD
1454 for (;;) {
1455 if (unlikely(is_end(node))) {
1456 node = next = NULL;
1457 break;
1458 }
1459 next = rcu_dereference(node->p.next);
1460 if (likely(!is_removed(next))
1461 && !is_dummy(next)) {
1462 break;
1463 }
1464 node = clear_flag(next);
1465 }
1466 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1467 iter->node = node;
1468 iter->next = next;
1469}
1470
1471void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1472{
1473 struct _cds_lfht_node *lookup;
1474
1475 /*
1476 * Get next after first dummy node. The first dummy node is the
1477 * first node of the linked list.
1478 */
1479 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1480 iter->next = lookup->next;
4e9b9fbf
MD
1481 cds_lfht_next(ht, iter);
1482}
1483
14044b37 1484void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1485{
4105056a 1486 unsigned long hash, size;
ab7d5fc6 1487
49c2e2d6 1488 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1489 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1490
4105056a 1491 size = rcu_dereference(ht->t.size);
83beee94 1492 _cds_lfht_add(ht, size, node, NULL, 0);
4105056a 1493 ht_count_add(ht, size);
3eca1b8c
MD
1494}
1495
14044b37 1496struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1497 struct cds_lfht_node *node)
3eca1b8c 1498{
4105056a 1499 unsigned long hash, size;
83beee94 1500 struct cds_lfht_iter iter;
3eca1b8c 1501
49c2e2d6 1502 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1503 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1504
4105056a 1505 size = rcu_dereference(ht->t.size);
83beee94
MD
1506 _cds_lfht_add(ht, size, node, &iter, 0);
1507 if (iter.node == node)
4105056a 1508 ht_count_add(ht, size);
83beee94 1509 return iter.node;
2ed95849
MD
1510}
1511
9357c415 1512struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1513 struct cds_lfht_node *node)
1514{
1515 unsigned long hash, size;
83beee94 1516 struct cds_lfht_iter iter;
48ed1c18
MD
1517
1518 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1519 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1520
1521 size = rcu_dereference(ht->t.size);
83beee94
MD
1522 for (;;) {
1523 _cds_lfht_add(ht, size, node, &iter, 0);
1524 if (iter.node == node) {
1525 ht_count_add(ht, size);
1526 return NULL;
1527 }
1528
1529 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1530 return iter.node;
1531 }
48ed1c18
MD
1532}
1533
9357c415
MD
1534int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1535 struct cds_lfht_node *new_node)
1536{
1537 unsigned long size;
1538
1539 size = rcu_dereference(ht->t.size);
1540 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1541 new_node);
1542}
1543
1544int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1545{
4105056a 1546 unsigned long size;
df44348d 1547 int ret;
abc490a1 1548
4105056a 1549 size = rcu_dereference(ht->t.size);
9357c415 1550 ret = _cds_lfht_del(ht, size, iter->node, 0);
df44348d 1551 if (!ret)
860d07e8 1552 ht_count_del(ht, size);
df44348d 1553 return ret;
2ed95849 1554}
ab7d5fc6 1555
abc490a1 1556static
14044b37 1557int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1558{
14044b37
MD
1559 struct cds_lfht_node *node;
1560 struct _cds_lfht_node *lookup;
4105056a 1561 unsigned long order, i, size;
674f7a69 1562
abc490a1 1563 /* Check that the table is empty */
4105056a 1564 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1565 node = (struct cds_lfht_node *) lookup;
abc490a1 1566 do {
1b81fe1a
MD
1567 node = clear_flag(node)->p.next;
1568 if (!is_dummy(node))
abc490a1 1569 return -EPERM;
273399de 1570 assert(!is_removed(node));
bb7b2f26 1571 } while (!is_end(node));
4105056a
MD
1572 /*
1573 * size accessed without rcu_dereference because hash table is
1574 * being destroyed.
1575 */
1576 size = ht->t.size;
abc490a1 1577 /* Internal sanity check: all nodes left should be dummy */
4105056a 1578 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1579 unsigned long len;
1580
1581 len = !order ? 1 : 1UL << (order - 1);
1582 for (i = 0; i < len; i++) {
f0c29ed7 1583 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1584 order, i,
4105056a
MD
1585 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1586 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1587 }
5488222b
LJ
1588
1589 if (order == ht->min_alloc_order)
1590 poison_free(ht->t.tbl[0]);
1591 else if (order > ht->min_alloc_order)
1592 poison_free(ht->t.tbl[order]);
1593 /* Nothing to delete for order < ht->min_alloc_order */
674f7a69 1594 }
abc490a1 1595 return 0;
674f7a69
MD
1596}
1597
1598/*
1599 * Should only be called when no more concurrent readers nor writers can
1600 * possibly access the table.
1601 */
b7d619b0 1602int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1603{
5e28c532
MD
1604 int ret;
1605
848d4088 1606 /* Wait for in-flight resize operations to complete */
24953e08
MD
1607 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1608 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1609 while (uatomic_read(&ht->in_progress_resize))
1610 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1611 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1612 if (ret)
1613 return ret;
4c42f1b8 1614 free_split_items_count(ht->split_count);
b7d619b0
MD
1615 if (attr)
1616 *attr = ht->resize_attr;
98808fb1 1617 poison_free(ht);
5e28c532 1618 return ret;
674f7a69
MD
1619}
1620
14044b37 1621void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1622 long *approx_before,
273399de 1623 unsigned long *count,
973e5e1b 1624 unsigned long *removed,
d933dd0e 1625 long *approx_after)
273399de 1626{
14044b37
MD
1627 struct cds_lfht_node *node, *next;
1628 struct _cds_lfht_node *lookup;
24365af7 1629 unsigned long nr_dummy = 0;
273399de 1630
7ed7682f 1631 *approx_before = 0;
4c42f1b8 1632 if (split_count_mask >= 0) {
973e5e1b
MD
1633 int i;
1634
4c42f1b8
LJ
1635 for (i = 0; i < split_count_mask + 1; i++) {
1636 *approx_before += uatomic_read(&ht->split_count[i].add);
1637 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1638 }
1639 }
1640
273399de
MD
1641 *count = 0;
1642 *removed = 0;
1643
24365af7 1644 /* Count non-dummy nodes in the table */
4105056a 1645 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1646 node = (struct cds_lfht_node *) lookup;
273399de 1647 do {
cc4fcb10 1648 next = rcu_dereference(node->p.next);
b198f0fd 1649 if (is_removed(next)) {
973e5e1b
MD
1650 if (!is_dummy(next))
1651 (*removed)++;
1652 else
1653 (nr_dummy)++;
1b81fe1a 1654 } else if (!is_dummy(next))
273399de 1655 (*count)++;
24365af7
MD
1656 else
1657 (nr_dummy)++;
273399de 1658 node = clear_flag(next);
bb7b2f26 1659 } while (!is_end(node));
f0c29ed7 1660 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1661 *approx_after = 0;
4c42f1b8 1662 if (split_count_mask >= 0) {
973e5e1b
MD
1663 int i;
1664
4c42f1b8
LJ
1665 for (i = 0; i < split_count_mask + 1; i++) {
1666 *approx_after += uatomic_read(&ht->split_count[i].add);
1667 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1668 }
1669 }
273399de
MD
1670}
1671
1475579c 1672/* called with resize mutex held */
abc490a1 1673static
4105056a 1674void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1675 unsigned long old_size, unsigned long new_size)
abc490a1 1676{
1475579c 1677 unsigned long old_order, new_order;
1475579c 1678
93d46c39
LJ
1679 old_order = get_count_order_ulong(old_size);
1680 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1681 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1682 old_size, old_order, new_size, new_order);
1475579c 1683 assert(new_size > old_size);
93d46c39 1684 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1685}
1686
1687/* called with resize mutex held */
1688static
4105056a 1689void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1690 unsigned long old_size, unsigned long new_size)
464a1ec9 1691{
1475579c 1692 unsigned long old_order, new_order;
464a1ec9 1693
5488222b 1694 new_size = max(new_size, ht->min_alloc_size);
93d46c39
LJ
1695 old_order = get_count_order_ulong(old_size);
1696 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1697 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1698 old_size, old_order, new_size, new_order);
1475579c 1699 assert(new_size < old_size);
1475579c 1700
4105056a 1701 /* Remove and unlink all dummy nodes to remove. */
93d46c39 1702 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1703}
1704
1475579c
MD
1705
1706/* called with resize mutex held */
1707static
1708void _do_cds_lfht_resize(struct cds_lfht *ht)
1709{
1710 unsigned long new_size, old_size;
4105056a
MD
1711
1712 /*
1713 * Resize table, re-do if the target size has changed under us.
1714 */
1715 do {
d2be3620
MD
1716 assert(uatomic_read(&ht->in_progress_resize));
1717 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1718 break;
4105056a
MD
1719 ht->t.resize_initiated = 1;
1720 old_size = ht->t.size;
1721 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1722 if (old_size < new_size)
1723 _do_cds_lfht_grow(ht, old_size, new_size);
1724 else if (old_size > new_size)
1725 _do_cds_lfht_shrink(ht, old_size, new_size);
1726 ht->t.resize_initiated = 0;
1727 /* write resize_initiated before read resize_target */
1728 cmm_smp_mb();
4d676753 1729 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1730}
1731
abc490a1 1732static
4105056a 1733unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1734 int growth_order)
464a1ec9 1735{
4105056a
MD
1736 return _uatomic_max(&ht->t.resize_target,
1737 size << growth_order);
464a1ec9
MD
1738}
1739
1475579c 1740static
4105056a 1741void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1742 unsigned long count)
1475579c 1743{
5488222b 1744 count = max(count, ht->min_alloc_size);
4105056a 1745 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1746}
1747
1748void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1749{
4105056a
MD
1750 resize_target_update_count(ht, new_size);
1751 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1752 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1753 pthread_mutex_lock(&ht->resize_mutex);
1754 _do_cds_lfht_resize(ht);
1755 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1756 ht->cds_lfht_rcu_thread_online();
abc490a1 1757}
464a1ec9 1758
abc490a1
MD
1759static
1760void do_resize_cb(struct rcu_head *head)
1761{
1762 struct rcu_resize_work *work =
1763 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1764 struct cds_lfht *ht = work->ht;
abc490a1 1765
5f511391 1766 ht->cds_lfht_rcu_thread_offline();
abc490a1 1767 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1768 _do_cds_lfht_resize(ht);
abc490a1 1769 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1770 ht->cds_lfht_rcu_thread_online();
98808fb1 1771 poison_free(work);
848d4088
MD
1772 cmm_smp_mb(); /* finish resize before decrement */
1773 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1774}
1775
abc490a1 1776static
4105056a 1777void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1778{
abc490a1 1779 struct rcu_resize_work *work;
f9830efd 1780 unsigned long target_size;
abc490a1 1781
4105056a
MD
1782 target_size = resize_target_update(ht, size, growth);
1783 /* Store resize_target before read resize_initiated */
1784 cmm_smp_mb();
1785 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1786 uatomic_inc(&ht->in_progress_resize);
59290e9d 1787 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1788 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1789 uatomic_dec(&ht->in_progress_resize);
59290e9d 1790 return;
ed35e6d8 1791 }
f9830efd
MD
1792 work = malloc(sizeof(*work));
1793 work->ht = ht;
14044b37 1794 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1795 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1796 }
ab7d5fc6 1797}
3171717f 1798
f8994aee
MD
1799#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1800
3171717f 1801static
4105056a 1802void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1803 unsigned long count)
1804{
1805 struct rcu_resize_work *work;
3171717f 1806
b8af5011
MD
1807 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1808 return;
4105056a
MD
1809 resize_target_update_count(ht, count);
1810 /* Store resize_target before read resize_initiated */
1811 cmm_smp_mb();
1812 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1813 uatomic_inc(&ht->in_progress_resize);
59290e9d 1814 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1815 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1816 uatomic_dec(&ht->in_progress_resize);
59290e9d 1817 return;
ed35e6d8 1818 }
3171717f
MD
1819 work = malloc(sizeof(*work));
1820 work->ht = ht;
1821 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1822 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1823 }
1824}
f8994aee
MD
1825
1826#endif
This page took 0.217934 seconds and 4 git commands to generate.