rculfhash: use DEFAULT_SPLIT_COUNT_MASK for !HAVE_SYSCONF
[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
4ddbb355 182#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
183#define CHAIN_LEN_TARGET 1
184#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 185
cd95516d 186/*
76a73da8 187 * Define the minimum table size.
cd95516d 188 */
c9edd44a 189#define MIN_TABLE_SIZE 1
cd95516d 190
4105056a
MD
191#if (CAA_BITS_PER_LONG == 32)
192#define MAX_TABLE_ORDER 32
193#else
194#define MAX_TABLE_ORDER 64
195#endif
196
b7d619b0
MD
197/*
198 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
199 */
6083a889
MD
200#define MIN_PARTITION_PER_THREAD_ORDER 12
201#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 202
4105056a
MD
203#ifndef min
204#define min(a, b) ((a) < (b) ? (a) : (b))
205#endif
206
abc490a1
MD
207#ifndef max
208#define max(a, b) ((a) > (b) ? (a) : (b))
209#endif
2ed95849 210
d95bd160
MD
211/*
212 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 213 * It indicates that no node must attach to the node scheduled for
b198f0fd 214 * removal, and that node garbage collection must be performed.
d95bd160
MD
215 * The dummy flag does not require to be updated atomically with the
216 * pointer, but it is added as a pointer low bit flag to save space.
217 */
d37166c6 218#define REMOVED_FLAG (1UL << 0)
b198f0fd
MD
219#define DUMMY_FLAG (1UL << 1)
220#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 221
bb7b2f26 222/* Value of the end pointer. Should not interact with flags. */
f9c80341 223#define END_VALUE NULL
bb7b2f26 224
df44348d 225struct ht_items_count {
860d07e8 226 unsigned long add, del;
df44348d
MD
227} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
228
1475579c 229struct rcu_level {
0d14ceb2 230 /* Note: manually update allocation length when adding a field */
1475579c
MD
231 struct _cds_lfht_node nodes[0];
232};
233
395270b6 234struct rcu_table {
4105056a 235 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 236 unsigned long resize_target;
11519af6 237 int resize_initiated;
4105056a 238 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
239};
240
14044b37 241struct cds_lfht {
4105056a 242 struct rcu_table t;
14044b37
MD
243 cds_lfht_hash_fct hash_fct;
244 cds_lfht_compare_fct compare_fct;
5488222b
LJ
245 unsigned long min_alloc_order;
246 unsigned long min_alloc_size;
732ad076 247 unsigned long hash_seed;
b8af5011 248 int flags;
5f511391
MD
249 /*
250 * We need to put the work threads offline (QSBR) when taking this
251 * mutex, because we use synchronize_rcu within this mutex critical
252 * section, which waits on read-side critical sections, and could
253 * therefore cause grace-period deadlock if we hold off RCU G.P.
254 * completion.
255 */
464a1ec9 256 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 257 unsigned int in_progress_resize, in_progress_destroy;
14044b37 258 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 259 void (*func)(struct rcu_head *head));
1475579c 260 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
261 void (*cds_lfht_rcu_read_lock)(void);
262 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
263 void (*cds_lfht_rcu_thread_offline)(void);
264 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
265 void (*cds_lfht_rcu_register_thread)(void);
266 void (*cds_lfht_rcu_unregister_thread)(void);
267 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 268 long count; /* global approximate item count */
4c42f1b8 269 struct ht_items_count *split_count; /* split item count */
2ed95849
MD
270};
271
abc490a1
MD
272struct rcu_resize_work {
273 struct rcu_head head;
14044b37 274 struct cds_lfht *ht;
abc490a1 275};
2ed95849 276
b7d619b0 277struct partition_resize_work {
1af6e26e 278 pthread_t thread_id;
b7d619b0
MD
279 struct cds_lfht *ht;
280 unsigned long i, start, len;
281 void (*fct)(struct cds_lfht *ht, unsigned long i,
282 unsigned long start, unsigned long len);
283};
284
76a73da8 285static
83beee94
MD
286void _cds_lfht_add(struct cds_lfht *ht,
287 unsigned long size,
288 struct cds_lfht_node *node,
289 struct cds_lfht_iter *unique_ret,
290 int dummy);
48ed1c18 291
abc490a1
MD
292/*
293 * Algorithm to reverse bits in a word by lookup table, extended to
294 * 64-bit words.
f9830efd 295 * Source:
abc490a1 296 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 297 * Originally from Public Domain.
abc490a1
MD
298 */
299
300static const uint8_t BitReverseTable256[256] =
2ed95849 301{
abc490a1
MD
302#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
303#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
304#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
305 R6(0), R6(2), R6(1), R6(3)
306};
307#undef R2
308#undef R4
309#undef R6
2ed95849 310
abc490a1
MD
311static
312uint8_t bit_reverse_u8(uint8_t v)
313{
314 return BitReverseTable256[v];
315}
ab7d5fc6 316
abc490a1
MD
317static __attribute__((unused))
318uint32_t bit_reverse_u32(uint32_t v)
319{
320 return ((uint32_t) bit_reverse_u8(v) << 24) |
321 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
322 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
323 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
324}
325
abc490a1
MD
326static __attribute__((unused))
327uint64_t bit_reverse_u64(uint64_t v)
2ed95849 328{
abc490a1
MD
329 return ((uint64_t) bit_reverse_u8(v) << 56) |
330 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
331 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
332 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
333 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
334 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
335 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
336 ((uint64_t) bit_reverse_u8(v >> 56));
337}
338
339static
340unsigned long bit_reverse_ulong(unsigned long v)
341{
342#if (CAA_BITS_PER_LONG == 32)
343 return bit_reverse_u32(v);
344#else
345 return bit_reverse_u64(v);
346#endif
347}
348
f9830efd 349/*
24365af7
MD
350 * fls: returns the position of the most significant bit.
351 * Returns 0 if no bit is set, else returns the position of the most
352 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 353 */
24365af7
MD
354#if defined(__i386) || defined(__x86_64)
355static inline
356unsigned int fls_u32(uint32_t x)
f9830efd 357{
24365af7
MD
358 int r;
359
360 asm("bsrl %1,%0\n\t"
361 "jnz 1f\n\t"
362 "movl $-1,%0\n\t"
363 "1:\n\t"
364 : "=r" (r) : "rm" (x));
365 return r + 1;
366}
367#define HAS_FLS_U32
368#endif
369
370#if defined(__x86_64)
371static inline
372unsigned int fls_u64(uint64_t x)
373{
374 long r;
375
376 asm("bsrq %1,%0\n\t"
377 "jnz 1f\n\t"
378 "movq $-1,%0\n\t"
379 "1:\n\t"
380 : "=r" (r) : "rm" (x));
381 return r + 1;
382}
383#define HAS_FLS_U64
384#endif
385
386#ifndef HAS_FLS_U64
387static __attribute__((unused))
388unsigned int fls_u64(uint64_t x)
389{
390 unsigned int r = 64;
391
392 if (!x)
393 return 0;
394
395 if (!(x & 0xFFFFFFFF00000000ULL)) {
396 x <<= 32;
397 r -= 32;
398 }
399 if (!(x & 0xFFFF000000000000ULL)) {
400 x <<= 16;
401 r -= 16;
402 }
403 if (!(x & 0xFF00000000000000ULL)) {
404 x <<= 8;
405 r -= 8;
406 }
407 if (!(x & 0xF000000000000000ULL)) {
408 x <<= 4;
409 r -= 4;
410 }
411 if (!(x & 0xC000000000000000ULL)) {
412 x <<= 2;
413 r -= 2;
414 }
415 if (!(x & 0x8000000000000000ULL)) {
416 x <<= 1;
417 r -= 1;
418 }
419 return r;
420}
421#endif
422
423#ifndef HAS_FLS_U32
424static __attribute__((unused))
425unsigned int fls_u32(uint32_t x)
426{
427 unsigned int r = 32;
f9830efd 428
24365af7
MD
429 if (!x)
430 return 0;
431 if (!(x & 0xFFFF0000U)) {
432 x <<= 16;
433 r -= 16;
434 }
435 if (!(x & 0xFF000000U)) {
436 x <<= 8;
437 r -= 8;
438 }
439 if (!(x & 0xF0000000U)) {
440 x <<= 4;
441 r -= 4;
442 }
443 if (!(x & 0xC0000000U)) {
444 x <<= 2;
445 r -= 2;
446 }
447 if (!(x & 0x80000000U)) {
448 x <<= 1;
449 r -= 1;
450 }
451 return r;
452}
453#endif
454
455unsigned int fls_ulong(unsigned long x)
f9830efd 456{
24365af7
MD
457#if (CAA_BITS_PER_lONG == 32)
458 return fls_u32(x);
459#else
460 return fls_u64(x);
461#endif
462}
f9830efd 463
920f8ef6
LJ
464/*
465 * Return the minimum order for which x <= (1UL << order).
466 * Return -1 if x is 0.
467 */
24365af7
MD
468int get_count_order_u32(uint32_t x)
469{
920f8ef6
LJ
470 if (!x)
471 return -1;
24365af7 472
920f8ef6 473 return fls_u32(x - 1);
24365af7
MD
474}
475
920f8ef6
LJ
476/*
477 * Return the minimum order for which x <= (1UL << order).
478 * Return -1 if x is 0.
479 */
24365af7
MD
480int get_count_order_ulong(unsigned long x)
481{
920f8ef6
LJ
482 if (!x)
483 return -1;
24365af7 484
920f8ef6 485 return fls_ulong(x - 1);
f9830efd
MD
486}
487
98808fb1
MD
488#ifdef POISON_FREE
489#define poison_free(ptr) \
490 do { \
491 memset(ptr, 0x42, sizeof(*(ptr))); \
492 free(ptr); \
493 } while (0)
494#else
495#define poison_free(ptr) free(ptr)
496#endif
497
f9830efd 498static
4105056a 499void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 500
f8994aee 501static
4105056a 502void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
503 unsigned long count);
504
df44348d 505static long nr_cpus_mask = -1;
4c42f1b8
LJ
506static long split_count_mask = -1;
507
4ddbb355 508#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
509static void ht_init_nr_cpus_mask(void)
510{
511 long maxcpus;
512
513 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
514 if (maxcpus <= 0) {
515 nr_cpus_mask = -2;
516 return;
517 }
518 /*
519 * round up number of CPUs to next power of two, so we
520 * can use & for modulo.
521 */
522 maxcpus = 1UL << get_count_order_ulong(maxcpus);
523 nr_cpus_mask = maxcpus - 1;
524}
4ddbb355
LJ
525#else /* #if defined(HAVE_SYSCONF) */
526static void ht_init_nr_cpus_mask(void)
527{
528 nr_cpus_mask = -2;
529}
530#endif /* #else #if defined(HAVE_SYSCONF) */
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();
4ddbb355
LJ
539 if (nr_cpus_mask < 0)
540 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
541 else
542 split_count_mask = nr_cpus_mask;
df44348d 543 }
4c42f1b8 544
4ddbb355
LJ
545 assert(split_count_mask >= 0);
546 return calloc(split_count_mask + 1, sizeof(*count));
df44348d
MD
547}
548
549static
4c42f1b8 550void free_split_items_count(struct ht_items_count *count)
df44348d 551{
98808fb1 552 poison_free(count);
df44348d
MD
553}
554
14360f1c 555#if defined(HAVE_SCHED_GETCPU)
df44348d 556static
14360f1c 557int ht_get_split_count_index(unsigned long hash)
df44348d
MD
558{
559 int cpu;
560
4c42f1b8 561 assert(split_count_mask >= 0);
df44348d
MD
562 cpu = sched_getcpu();
563 if (unlikely(cpu < 0))
14360f1c 564 return hash & split_count_mask;
df44348d 565 else
4c42f1b8 566 return cpu & split_count_mask;
df44348d 567}
14360f1c
LJ
568#else /* #if defined(HAVE_SCHED_GETCPU) */
569static
570int ht_get_split_count_index(unsigned long hash)
571{
572 return hash & split_count_mask;
573}
574#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
575
576static
14360f1c 577void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 578{
4c42f1b8
LJ
579 unsigned long split_count;
580 int index;
df44348d 581
4c42f1b8 582 if (unlikely(!ht->split_count))
3171717f 583 return;
14360f1c 584 index = ht_get_split_count_index(hash);
4c42f1b8
LJ
585 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
586 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 587 long count;
df44348d 588
4c42f1b8 589 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
590 count = uatomic_add_return(&ht->count,
591 1UL << COUNT_COMMIT_ORDER);
592 /* If power of 2 */
593 if (!(count & (count - 1))) {
4105056a 594 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 595 return;
e3ecefd6 596 dbg_printf("add set global %ld\n", count);
4105056a 597 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 598 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
599 }
600 }
601}
602
603static
14360f1c 604void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 605{
4c42f1b8
LJ
606 unsigned long split_count;
607 int index;
df44348d 608
4c42f1b8 609 if (unlikely(!ht->split_count))
3171717f 610 return;
14360f1c 611 index = ht_get_split_count_index(hash);
4c42f1b8
LJ
612 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
613 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 614 long count;
df44348d 615
4c42f1b8 616 dbg_printf("del split count %lu\n", split_count);
df44348d 617 count = uatomic_add_return(&ht->count,
3171717f 618 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
619 /* If power of 2 */
620 if (!(count & (count - 1))) {
4105056a 621 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 622 return;
e3ecefd6
MD
623 dbg_printf("del set global %ld\n", count);
624 /*
c941bb9e 625 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
626 * certain threshold.
627 */
4c42f1b8 628 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 629 return;
4105056a 630 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 631 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
632 }
633 }
634}
635
f9830efd 636static
4105056a 637void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 638{
f8994aee
MD
639 unsigned long count;
640
b8af5011
MD
641 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
642 return;
f8994aee
MD
643 count = uatomic_read(&ht->count);
644 /*
645 * Use bucket-local length for small table expand and for
646 * environments lacking per-cpu data support.
647 */
648 if (count >= (1UL << COUNT_COMMIT_ORDER))
649 return;
24365af7 650 if (chain_len > 100)
f0c29ed7 651 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 652 chain_len);
3390d470 653 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 654 cds_lfht_resize_lazy(ht, size,
01370f0b 655 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
656}
657
abc490a1 658static
14044b37 659struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 660{
14044b37 661 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
662}
663
664static
14044b37 665int is_removed(struct cds_lfht_node *node)
abc490a1 666{
d37166c6 667 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
668}
669
670static
14044b37 671struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 672{
14044b37 673 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
674}
675
f5596c94 676static
14044b37 677int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
678{
679 return ((unsigned long) node) & DUMMY_FLAG;
680}
681
682static
14044b37 683struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 684{
14044b37 685 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 686}
bb7b2f26
MD
687
688static
689struct cds_lfht_node *get_end(void)
690{
691 return (struct cds_lfht_node *) END_VALUE;
692}
693
694static
695int is_end(struct cds_lfht_node *node)
696{
697 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
698}
699
abc490a1 700static
f9830efd 701unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
702{
703 unsigned long old1, old2;
704
705 old1 = uatomic_read(ptr);
706 do {
707 old2 = old1;
708 if (old2 >= v)
f9830efd 709 return old2;
abc490a1 710 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 711 return v;
abc490a1
MD
712}
713
f4a9cc0b
LJ
714static
715struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
716 unsigned long hash)
717{
718 unsigned long index, order;
719
720 assert(size > 0);
721 index = hash & (size - 1);
ef6e6171
LJ
722
723 if (index < ht->min_alloc_size) {
724 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
725 hash, index);
726 return &ht->t.tbl[0]->nodes[index];
727 }
a4ea2223
LJ
728 /*
729 * equivalent to get_count_order_ulong(index + 1), but optimizes
730 * away the non-existing 0 special-case for
731 * get_count_order_ulong.
732 */
733 order = fls_ulong(index);
f4a9cc0b 734 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
ef6e6171
LJ
735 hash, index, order, index & ((1UL << (order - 1)) - 1));
736 return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
737}
738
273399de
MD
739/*
740 * Remove all logically deleted nodes from a bucket up to a certain node key.
741 */
742static
f9c80341 743void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 744{
14044b37 745 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 746
c90201ac
MD
747 assert(!is_dummy(dummy));
748 assert(!is_removed(dummy));
749 assert(!is_dummy(node));
750 assert(!is_removed(node));
273399de
MD
751 for (;;) {
752 iter_prev = dummy;
753 /* We can always skip the dummy node initially */
cc4fcb10 754 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 755 assert(!is_removed(iter));
cc4fcb10 756 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
757 /*
758 * We should never be called with dummy (start of chain)
759 * and logically removed node (end of path compression
760 * marker) being the actual same node. This would be a
761 * bug in the algorithm implementation.
762 */
763 assert(dummy != node);
273399de 764 for (;;) {
bb7b2f26 765 if (unlikely(is_end(iter)))
f9c80341 766 return;
76412f24 767 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 768 return;
cc4fcb10 769 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 770 if (likely(is_removed(next)))
273399de 771 break;
b453eae1 772 iter_prev = clear_flag(iter);
273399de
MD
773 iter = next;
774 }
b198f0fd 775 assert(!is_removed(iter));
f5596c94
MD
776 if (is_dummy(iter))
777 new_next = flag_dummy(clear_flag(next));
778 else
779 new_next = clear_flag(next);
780 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 781 }
f9c80341 782 return;
273399de
MD
783}
784
9357c415
MD
785static
786int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
787 struct cds_lfht_node *old_node,
3fb86f26 788 struct cds_lfht_node *old_next,
9357c415
MD
789 struct cds_lfht_node *new_node)
790{
3fb86f26 791 struct cds_lfht_node *dummy, *ret_next;
9357c415 792 struct _cds_lfht_node *lookup;
9357c415
MD
793
794 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 795 return -ENOENT;
9357c415
MD
796
797 assert(!is_removed(old_node));
798 assert(!is_dummy(old_node));
799 assert(!is_removed(new_node));
800 assert(!is_dummy(new_node));
801 assert(new_node != old_node);
3fb86f26 802 for (;;) {
9357c415 803 /* Insert after node to be replaced */
9357c415
MD
804 if (is_removed(old_next)) {
805 /*
806 * Too late, the old node has been removed under us
807 * between lookup and replace. Fail.
808 */
7801dadd 809 return -ENOENT;
9357c415
MD
810 }
811 assert(!is_dummy(old_next));
812 assert(new_node != clear_flag(old_next));
813 new_node->p.next = clear_flag(old_next);
814 /*
815 * Here is the whole trick for lock-free replace: we add
816 * the replacement node _after_ the node we want to
817 * replace by atomically setting its next pointer at the
818 * same time we set its removal flag. Given that
819 * the lookups/get next use an iterator aware of the
820 * next pointer, they will either skip the old node due
821 * to the removal flag and see the new node, or use
822 * the old node, but will not see the new one.
823 */
824 ret_next = uatomic_cmpxchg(&old_node->p.next,
825 old_next, flag_removed(new_node));
3fb86f26 826 if (ret_next == old_next)
7801dadd 827 break; /* We performed the replacement. */
3fb86f26
LJ
828 old_next = ret_next;
829 }
9357c415 830
9357c415
MD
831 /*
832 * Ensure that the old node is not visible to readers anymore:
833 * lookup for the node, and remove it (along with any other
834 * logically removed node) if found.
835 */
f4a9cc0b 836 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
837 dummy = (struct cds_lfht_node *) lookup;
838 _cds_lfht_gc_bucket(dummy, new_node);
7801dadd
LJ
839
840 assert(is_removed(rcu_dereference(old_node->p.next)));
841 return 0;
9357c415
MD
842}
843
83beee94
MD
844/*
845 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
846 * mode. A NULL unique_ret allows creation of duplicate keys.
847 */
abc490a1 848static
83beee94
MD
849void _cds_lfht_add(struct cds_lfht *ht,
850 unsigned long size,
851 struct cds_lfht_node *node,
852 struct cds_lfht_iter *unique_ret,
853 int dummy)
abc490a1 854{
14044b37 855 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 856 *return_node;
14044b37 857 struct _cds_lfht_node *lookup;
abc490a1 858
c90201ac
MD
859 assert(!is_dummy(node));
860 assert(!is_removed(node));
f4a9cc0b 861 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 862 for (;;) {
adc0de68 863 uint32_t chain_len = 0;
abc490a1 864
11519af6
MD
865 /*
866 * iter_prev points to the non-removed node prior to the
867 * insert location.
11519af6 868 */
14044b37 869 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 870 /* We can always skip the dummy node initially */
cc4fcb10
MD
871 iter = rcu_dereference(iter_prev->p.next);
872 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 873 for (;;) {
bb7b2f26 874 if (unlikely(is_end(iter)))
273399de 875 goto insert;
76412f24 876 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 877 goto insert;
238cc06e 878
194fdbd1
LJ
879 /* dummy node is the first node of the identical-hash-value chain */
880 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
881 goto insert;
238cc06e 882
cc4fcb10 883 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 884 if (unlikely(is_removed(next)))
9dba85be 885 goto gc_node;
238cc06e
LJ
886
887 /* uniquely add */
83beee94 888 if (unique_ret
1b81fe1a 889 && !is_dummy(next)
238cc06e
LJ
890 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
891 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
892
893 /*
894 * uniquely adding inserts the node as the first
895 * node of the identical-hash-value node chain.
896 *
897 * This semantic ensures no duplicated keys
898 * should ever be observable in the table
899 * (including observe one node by one node
900 * by forward iterations)
901 */
902 cds_lfht_next_duplicate(ht, &d_iter);
903 if (!d_iter.node)
904 goto insert;
905
906 *unique_ret = d_iter;
83beee94 907 return;
48ed1c18 908 }
238cc06e 909
11519af6 910 /* Only account for identical reverse hash once */
24365af7
MD
911 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
912 && !is_dummy(next))
4105056a 913 check_resize(ht, size, ++chain_len);
11519af6 914 iter_prev = clear_flag(iter);
273399de 915 iter = next;
abc490a1 916 }
48ed1c18 917
273399de 918 insert:
7ec59d3b 919 assert(node != clear_flag(iter));
11519af6 920 assert(!is_removed(iter_prev));
c90201ac 921 assert(!is_removed(iter));
f000907d 922 assert(iter_prev != node);
f9c80341 923 if (!dummy)
1b81fe1a 924 node->p.next = clear_flag(iter);
f9c80341
MD
925 else
926 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
927 if (is_dummy(iter))
928 new_node = flag_dummy(node);
929 else
930 new_node = node;
cc4fcb10 931 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 932 new_node) != iter) {
273399de 933 continue; /* retry */
48ed1c18 934 } else {
83beee94 935 return_node = node;
960c9e4f 936 goto end;
48ed1c18
MD
937 }
938
9dba85be
MD
939 gc_node:
940 assert(!is_removed(iter));
f5596c94
MD
941 if (is_dummy(iter))
942 new_next = flag_dummy(clear_flag(next));
943 else
944 new_next = clear_flag(next);
945 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 946 /* retry */
464a1ec9 947 }
9357c415 948end:
83beee94
MD
949 if (unique_ret) {
950 unique_ret->node = return_node;
951 /* unique_ret->next left unset, never used. */
952 }
abc490a1 953}
464a1ec9 954
abc490a1 955static
860d07e8 956int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 957 struct cds_lfht_node *node,
b198f0fd 958 int dummy_removal)
abc490a1 959{
14044b37
MD
960 struct cds_lfht_node *dummy, *next, *old;
961 struct _cds_lfht_node *lookup;
5e28c532 962
9357c415 963 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 964 return -ENOENT;
9357c415 965
7ec59d3b 966 /* logically delete the node */
c90201ac
MD
967 assert(!is_dummy(node));
968 assert(!is_removed(node));
cc4fcb10 969 old = rcu_dereference(node->p.next);
7ec59d3b 970 do {
48ed1c18
MD
971 struct cds_lfht_node *new_next;
972
7ec59d3b 973 next = old;
76412f24 974 if (unlikely(is_removed(next)))
743f9143 975 return -ENOENT;
1475579c
MD
976 if (dummy_removal)
977 assert(is_dummy(next));
978 else
979 assert(!is_dummy(next));
48ed1c18 980 new_next = flag_removed(next);
48ed1c18 981 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b 982 } while (old != next);
7ec59d3b 983 /* We performed the (logical) deletion. */
7ec59d3b
MD
984
985 /*
986 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
987 * the node, and remove it (along with any other logically removed node)
988 * if found.
11519af6 989 */
f4a9cc0b 990 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 991 dummy = (struct cds_lfht_node *) lookup;
f9c80341 992 _cds_lfht_gc_bucket(dummy, node);
743f9143
LJ
993
994 assert(is_removed(rcu_dereference(node->p.next)));
995 return 0;
abc490a1 996}
2ed95849 997
b7d619b0
MD
998static
999void *partition_resize_thread(void *arg)
1000{
1001 struct partition_resize_work *work = arg;
1002
1003 work->ht->cds_lfht_rcu_register_thread();
1004 work->fct(work->ht, work->i, work->start, work->len);
1005 work->ht->cds_lfht_rcu_unregister_thread();
1006 return NULL;
1007}
1008
1009static
1010void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1011 unsigned long len,
1012 void (*fct)(struct cds_lfht *ht, unsigned long i,
1013 unsigned long start, unsigned long len))
1014{
1015 unsigned long partition_len;
1016 struct partition_resize_work *work;
6083a889
MD
1017 int thread, ret;
1018 unsigned long nr_threads;
b7d619b0 1019
6083a889
MD
1020 /*
1021 * Note: nr_cpus_mask + 1 is always power of 2.
1022 * We spawn just the number of threads we need to satisfy the minimum
1023 * partition size, up to the number of CPUs in the system.
1024 */
91452a6a
MD
1025 if (nr_cpus_mask > 0) {
1026 nr_threads = min(nr_cpus_mask + 1,
1027 len >> MIN_PARTITION_PER_THREAD_ORDER);
1028 } else {
1029 nr_threads = 1;
1030 }
6083a889
MD
1031 partition_len = len >> get_count_order_ulong(nr_threads);
1032 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1033 assert(work);
6083a889
MD
1034 for (thread = 0; thread < nr_threads; thread++) {
1035 work[thread].ht = ht;
1036 work[thread].i = i;
1037 work[thread].len = partition_len;
1038 work[thread].start = thread * partition_len;
1039 work[thread].fct = fct;
1af6e26e 1040 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1041 partition_resize_thread, &work[thread]);
b7d619b0
MD
1042 assert(!ret);
1043 }
6083a889 1044 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1045 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1046 assert(!ret);
1047 }
1048 free(work);
b7d619b0
MD
1049}
1050
e8de508e
MD
1051/*
1052 * Holding RCU read lock to protect _cds_lfht_add against memory
1053 * reclaim that could be performed by other call_rcu worker threads (ABA
1054 * problem).
9ee0fc9a 1055 *
b7d619b0 1056 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1057 * many worker threads, based on the number of CPUs available in the system.
1058 * This should therefore take care of not having the expand lagging behind too
1059 * many concurrent insertion threads by using the scheduler's ability to
1060 * schedule dummy node population fairly with insertions.
e8de508e 1061 */
4105056a 1062static
b7d619b0
MD
1063void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1064 unsigned long start, unsigned long len)
4105056a
MD
1065{
1066 unsigned long j;
1067
5488222b 1068 assert(i > ht->min_alloc_order);
4105056a 1069 ht->cds_lfht_rcu_read_lock();
b7d619b0 1070 for (j = start; j < start + len; j++) {
4105056a
MD
1071 struct cds_lfht_node *new_node =
1072 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1073
dc1da8f6 1074 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1075 i, j, (1UL << (i - 1)) + j);
dc1da8f6 1076 new_node->p.reverse_hash =
4f6e90b7
LJ
1077 bit_reverse_ulong((1UL << (i - 1)) + j);
1078 _cds_lfht_add(ht, 1UL << (i - 1),
83beee94 1079 new_node, NULL, 1);
4105056a
MD
1080 }
1081 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1082}
1083
1084static
1085void init_table_populate(struct cds_lfht *ht, unsigned long i,
1086 unsigned long len)
1087{
1088 assert(nr_cpus_mask != -1);
6083a889 1089 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1090 ht->cds_lfht_rcu_thread_online();
1091 init_table_populate_partition(ht, i, 0, len);
1092 ht->cds_lfht_rcu_thread_offline();
1093 return;
1094 }
1095 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1096}
1097
abc490a1 1098static
4105056a 1099void init_table(struct cds_lfht *ht,
93d46c39 1100 unsigned long first_order, unsigned long last_order)
24365af7 1101{
93d46c39 1102 unsigned long i;
24365af7 1103
93d46c39
LJ
1104 dbg_printf("init table: first_order %lu last_order %lu\n",
1105 first_order, last_order);
5488222b 1106 assert(first_order > ht->min_alloc_order);
93d46c39 1107 for (i = first_order; i <= last_order; i++) {
4105056a 1108 unsigned long len;
24365af7 1109
4f6e90b7 1110 len = 1UL << (i - 1);
f0c29ed7 1111 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1112
1113 /* Stop expand if the resize target changes under us */
4f6e90b7 1114 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1115 break;
1116
0d14ceb2 1117 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1118 assert(ht->t.tbl[i]);
4105056a 1119
4105056a 1120 /*
dc1da8f6
MD
1121 * Set all dummy nodes reverse hash values for a level and
1122 * link all dummy nodes into the table.
4105056a 1123 */
dc1da8f6 1124 init_table_populate(ht, i, len);
4105056a 1125
f9c80341
MD
1126 /*
1127 * Update table size.
1128 */
1129 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1130 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1131
4f6e90b7 1132 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1133 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1134 break;
1135 }
1136}
1137
e8de508e
MD
1138/*
1139 * Holding RCU read lock to protect _cds_lfht_remove against memory
1140 * reclaim that could be performed by other call_rcu worker threads (ABA
1141 * problem).
1142 * For a single level, we logically remove and garbage collect each node.
1143 *
1144 * As a design choice, we perform logical removal and garbage collection on a
1145 * node-per-node basis to simplify this algorithm. We also assume keeping good
1146 * cache locality of the operation would overweight possible performance gain
1147 * that could be achieved by batching garbage collection for multiple levels.
1148 * However, this would have to be justified by benchmarks.
1149 *
1150 * Concurrent removal and add operations are helping us perform garbage
1151 * collection of logically removed nodes. We guarantee that all logically
1152 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1153 * invoked to free a hole level of dummy nodes (after a grace period).
1154 *
1155 * Logical removal and garbage collection can therefore be done in batch or on a
1156 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1157 *
b7d619b0
MD
1158 * When we reach a certain length, we can split this removal over many worker
1159 * threads, based on the number of CPUs available in the system. This should
1160 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1161 * updater threads actively inserting into the hash table.
e8de508e 1162 */
4105056a 1163static
b7d619b0
MD
1164void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1165 unsigned long start, unsigned long len)
4105056a
MD
1166{
1167 unsigned long j;
1168
5488222b 1169 assert(i > ht->min_alloc_order);
4105056a 1170 ht->cds_lfht_rcu_read_lock();
b7d619b0 1171 for (j = start; j < start + len; j++) {
4105056a
MD
1172 struct cds_lfht_node *fini_node =
1173 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1174
1175 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1176 i, j, (1UL << (i - 1)) + j);
4105056a 1177 fini_node->p.reverse_hash =
4f6e90b7
LJ
1178 bit_reverse_ulong((1UL << (i - 1)) + j);
1179 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1180 }
4105056a 1181 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1182}
1183
1184static
1185void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1186{
1187
1188 assert(nr_cpus_mask != -1);
6083a889 1189 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1190 ht->cds_lfht_rcu_thread_online();
1191 remove_table_partition(ht, i, 0, len);
1192 ht->cds_lfht_rcu_thread_offline();
1193 return;
1194 }
1195 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1196}
1197
1475579c 1198static
4105056a 1199void fini_table(struct cds_lfht *ht,
93d46c39 1200 unsigned long first_order, unsigned long last_order)
1475579c 1201{
93d46c39 1202 long i;
0d14ceb2 1203 void *free_by_rcu = NULL;
1475579c 1204
93d46c39
LJ
1205 dbg_printf("fini table: first_order %lu last_order %lu\n",
1206 first_order, last_order);
5488222b 1207 assert(first_order > ht->min_alloc_order);
93d46c39 1208 for (i = last_order; i >= first_order; i--) {
4105056a 1209 unsigned long len;
1475579c 1210
4f6e90b7 1211 len = 1UL << (i - 1);
1475579c 1212 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1213
4d676753
MD
1214 /* Stop shrink if the resize target changes under us */
1215 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1216 break;
1217
1218 cmm_smp_wmb(); /* populate data before RCU size */
1219 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1220
1221 /*
1222 * We need to wait for all add operations to reach Q.S. (and
1223 * thus use the new table for lookups) before we can start
1224 * releasing the old dummy nodes. Otherwise their lookup will
1225 * return a logically removed node as insert position.
1226 */
1227 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1228 if (free_by_rcu)
1229 free(free_by_rcu);
4d676753 1230
21263e21 1231 /*
4105056a
MD
1232 * Set "removed" flag in dummy nodes about to be removed.
1233 * Unlink all now-logically-removed dummy node pointers.
1234 * Concurrent add/remove operation are helping us doing
1235 * the gc.
21263e21 1236 */
4105056a
MD
1237 remove_table(ht, i, len);
1238
0d14ceb2 1239 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1240
1241 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1242 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1243 break;
1244 }
0d14ceb2
LJ
1245
1246 if (free_by_rcu) {
1247 ht->cds_lfht_synchronize_rcu();
1248 free(free_by_rcu);
1249 }
1475579c
MD
1250}
1251
ff0d69de
LJ
1252static
1253void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
1254{
1255 struct _cds_lfht_node *prev, *node;
1256 unsigned long order, len, i, j;
1257
5488222b 1258 ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node));
ff0d69de
LJ
1259 assert(ht->t.tbl[0]);
1260
1261 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1262 ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
1263 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1264
1265 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1266 len = 1UL << (order - 1);
5488222b 1267 if (order <= ht->min_alloc_order) {
eb631bf2 1268 ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
5488222b
LJ
1269 } else {
1270 ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
1271 assert(ht->t.tbl[order]);
1272 }
ff0d69de
LJ
1273
1274 i = 0;
1275 prev = ht->t.tbl[i]->nodes;
1276 for (j = 0; j < len; j++) {
1277 if (j & (j - 1)) { /* Between power of 2 */
1278 prev++;
1279 } else if (j) { /* At each power of 2 */
1280 i++;
1281 prev = ht->t.tbl[i]->nodes;
1282 }
1283
1284 node = &ht->t.tbl[order]->nodes[j];
1285 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1286 order, j, j + len);
1287 node->next = prev->next;
1288 assert(is_dummy(node->next));
1289 node->reverse_hash = bit_reverse_ulong(j + len);
1290 prev->next = flag_dummy((struct cds_lfht_node *)node);
1291 }
1292 }
1293}
1294
7a9dcf9b 1295struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1296 cds_lfht_compare_fct compare_fct,
1297 unsigned long hash_seed,
1298 unsigned long init_size,
5488222b 1299 unsigned long min_alloc_size,
b8af5011 1300 int flags,
14044b37 1301 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1302 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1303 void (*cds_lfht_synchronize_rcu)(void),
1304 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1305 void (*cds_lfht_rcu_read_unlock)(void),
1306 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1307 void (*cds_lfht_rcu_thread_online)(void),
1308 void (*cds_lfht_rcu_register_thread)(void),
1309 void (*cds_lfht_rcu_unregister_thread)(void),
1310 pthread_attr_t *attr)
abc490a1 1311{
14044b37 1312 struct cds_lfht *ht;
24365af7 1313 unsigned long order;
abc490a1 1314
5488222b
LJ
1315 /* min_alloc_size must be power of two */
1316 if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
1317 return NULL;
8129be4e 1318 /* init_size must be power of two */
5488222b 1319 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1320 return NULL;
5488222b
LJ
1321 min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
1322 init_size = max(init_size, min_alloc_size);
14044b37 1323 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1324 assert(ht);
abc490a1 1325 ht->hash_fct = hash_fct;
732ad076
MD
1326 ht->compare_fct = compare_fct;
1327 ht->hash_seed = hash_seed;
14044b37 1328 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1329 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1330 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1331 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1332 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1333 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1334 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1335 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1336 ht->resize_attr = attr;
4c42f1b8 1337 ht->split_count = alloc_split_items_count();
abc490a1
MD
1338 /* this mutex should not nest in read-side C.S. */
1339 pthread_mutex_init(&ht->resize_mutex, NULL);
b8af5011 1340 ht->flags = flags;
5488222b 1341 order = get_count_order_ulong(init_size);
93d46c39 1342 ht->t.resize_target = 1UL << order;
ff0d69de
LJ
1343 cds_lfht_create_dummy(ht, 1UL << order);
1344 ht->t.size = 1UL << order;
5488222b
LJ
1345 ht->min_alloc_size = min_alloc_size;
1346 ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
abc490a1
MD
1347 return ht;
1348}
1349
adc0de68
MD
1350void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1351 struct cds_lfht_iter *iter)
2ed95849 1352{
bb7b2f26 1353 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1354 struct _cds_lfht_node *lookup;
f4a9cc0b 1355 unsigned long hash, reverse_hash, size;
2ed95849 1356
732ad076 1357 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1358 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1359
4105056a 1360 size = rcu_dereference(ht->t.size);
f4a9cc0b 1361 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1362 dummy_node = (struct cds_lfht_node *) lookup;
1363 /* We can always skip the dummy node initially */
1364 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1365 node = clear_flag(node);
2ed95849 1366 for (;;) {
bb7b2f26 1367 if (unlikely(is_end(node))) {
96ad1112 1368 node = next = NULL;
abc490a1 1369 break;
bb7b2f26 1370 }
cc4fcb10 1371 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1372 node = next = NULL;
abc490a1 1373 break;
2ed95849 1374 }
1b81fe1a 1375 next = rcu_dereference(node->p.next);
adc0de68 1376 if (likely(!is_removed(next))
1b81fe1a 1377 && !is_dummy(next)
ae450da3 1378 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1379 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1380 break;
2ed95849 1381 }
1b81fe1a 1382 node = clear_flag(next);
2ed95849 1383 }
1b81fe1a 1384 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1385 iter->node = node;
1386 iter->next = next;
abc490a1 1387}
e0ba718a 1388
3883c0e5 1389void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1390{
adc0de68 1391 struct cds_lfht_node *node, *next;
a481e5ff
MD
1392 unsigned long reverse_hash;
1393 void *key;
1394 size_t key_len;
1395
adc0de68 1396 node = iter->node;
a481e5ff
MD
1397 reverse_hash = node->p.reverse_hash;
1398 key = node->key;
1399 key_len = node->key_len;
adc0de68 1400 next = iter->next;
a481e5ff
MD
1401 node = clear_flag(next);
1402
1403 for (;;) {
bb7b2f26 1404 if (unlikely(is_end(node))) {
96ad1112 1405 node = next = NULL;
a481e5ff 1406 break;
bb7b2f26 1407 }
a481e5ff 1408 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1409 node = next = NULL;
a481e5ff
MD
1410 break;
1411 }
1412 next = rcu_dereference(node->p.next);
adc0de68 1413 if (likely(!is_removed(next))
a481e5ff
MD
1414 && !is_dummy(next)
1415 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1416 break;
1417 }
1418 node = clear_flag(next);
1419 }
1420 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1421 iter->node = node;
1422 iter->next = next;
a481e5ff
MD
1423}
1424
4e9b9fbf
MD
1425void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1426{
1427 struct cds_lfht_node *node, *next;
1428
853395e1 1429 node = clear_flag(iter->next);
4e9b9fbf
MD
1430 for (;;) {
1431 if (unlikely(is_end(node))) {
1432 node = next = NULL;
1433 break;
1434 }
1435 next = rcu_dereference(node->p.next);
1436 if (likely(!is_removed(next))
1437 && !is_dummy(next)) {
1438 break;
1439 }
1440 node = clear_flag(next);
1441 }
1442 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1443 iter->node = node;
1444 iter->next = next;
1445}
1446
1447void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1448{
1449 struct _cds_lfht_node *lookup;
1450
1451 /*
1452 * Get next after first dummy node. The first dummy node is the
1453 * first node of the linked list.
1454 */
1455 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1456 iter->next = lookup->next;
4e9b9fbf
MD
1457 cds_lfht_next(ht, iter);
1458}
1459
14044b37 1460void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1461{
4105056a 1462 unsigned long hash, size;
ab7d5fc6 1463
49c2e2d6 1464 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1465 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1466
4105056a 1467 size = rcu_dereference(ht->t.size);
83beee94 1468 _cds_lfht_add(ht, size, node, NULL, 0);
14360f1c 1469 ht_count_add(ht, size, hash);
3eca1b8c
MD
1470}
1471
14044b37 1472struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1473 struct cds_lfht_node *node)
3eca1b8c 1474{
4105056a 1475 unsigned long hash, size;
83beee94 1476 struct cds_lfht_iter iter;
3eca1b8c 1477
49c2e2d6 1478 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1479 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1480
4105056a 1481 size = rcu_dereference(ht->t.size);
83beee94
MD
1482 _cds_lfht_add(ht, size, node, &iter, 0);
1483 if (iter.node == node)
14360f1c 1484 ht_count_add(ht, size, hash);
83beee94 1485 return iter.node;
2ed95849
MD
1486}
1487
9357c415 1488struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1489 struct cds_lfht_node *node)
1490{
1491 unsigned long hash, size;
83beee94 1492 struct cds_lfht_iter iter;
48ed1c18
MD
1493
1494 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1495 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1496
1497 size = rcu_dereference(ht->t.size);
83beee94
MD
1498 for (;;) {
1499 _cds_lfht_add(ht, size, node, &iter, 0);
1500 if (iter.node == node) {
14360f1c 1501 ht_count_add(ht, size, hash);
83beee94
MD
1502 return NULL;
1503 }
1504
1505 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1506 return iter.node;
1507 }
48ed1c18
MD
1508}
1509
9357c415
MD
1510int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1511 struct cds_lfht_node *new_node)
1512{
1513 unsigned long size;
1514
1515 size = rcu_dereference(ht->t.size);
1516 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1517 new_node);
1518}
1519
1520int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1521{
14360f1c 1522 unsigned long size, hash;
df44348d 1523 int ret;
abc490a1 1524
4105056a 1525 size = rcu_dereference(ht->t.size);
9357c415 1526 ret = _cds_lfht_del(ht, size, iter->node, 0);
14360f1c
LJ
1527 if (!ret) {
1528 hash = bit_reverse_ulong(iter->node->p.reverse_hash);
1529 ht_count_del(ht, size, hash);
1530 }
df44348d 1531 return ret;
2ed95849 1532}
ab7d5fc6 1533
abc490a1 1534static
14044b37 1535int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1536{
14044b37
MD
1537 struct cds_lfht_node *node;
1538 struct _cds_lfht_node *lookup;
4105056a 1539 unsigned long order, i, size;
674f7a69 1540
abc490a1 1541 /* Check that the table is empty */
4105056a 1542 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1543 node = (struct cds_lfht_node *) lookup;
abc490a1 1544 do {
1b81fe1a
MD
1545 node = clear_flag(node)->p.next;
1546 if (!is_dummy(node))
abc490a1 1547 return -EPERM;
273399de 1548 assert(!is_removed(node));
bb7b2f26 1549 } while (!is_end(node));
4105056a
MD
1550 /*
1551 * size accessed without rcu_dereference because hash table is
1552 * being destroyed.
1553 */
1554 size = ht->t.size;
abc490a1 1555 /* Internal sanity check: all nodes left should be dummy */
4105056a 1556 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1557 unsigned long len;
1558
1559 len = !order ? 1 : 1UL << (order - 1);
1560 for (i = 0; i < len; i++) {
f0c29ed7 1561 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1562 order, i,
4105056a
MD
1563 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1564 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1565 }
5488222b
LJ
1566
1567 if (order == ht->min_alloc_order)
1568 poison_free(ht->t.tbl[0]);
1569 else if (order > ht->min_alloc_order)
1570 poison_free(ht->t.tbl[order]);
1571 /* Nothing to delete for order < ht->min_alloc_order */
674f7a69 1572 }
abc490a1 1573 return 0;
674f7a69
MD
1574}
1575
1576/*
1577 * Should only be called when no more concurrent readers nor writers can
1578 * possibly access the table.
1579 */
b7d619b0 1580int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1581{
5e28c532
MD
1582 int ret;
1583
848d4088 1584 /* Wait for in-flight resize operations to complete */
24953e08
MD
1585 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1586 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1587 while (uatomic_read(&ht->in_progress_resize))
1588 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1589 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1590 if (ret)
1591 return ret;
4c42f1b8 1592 free_split_items_count(ht->split_count);
b7d619b0
MD
1593 if (attr)
1594 *attr = ht->resize_attr;
98808fb1 1595 poison_free(ht);
5e28c532 1596 return ret;
674f7a69
MD
1597}
1598
14044b37 1599void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1600 long *approx_before,
273399de 1601 unsigned long *count,
973e5e1b 1602 unsigned long *removed,
d933dd0e 1603 long *approx_after)
273399de 1604{
14044b37
MD
1605 struct cds_lfht_node *node, *next;
1606 struct _cds_lfht_node *lookup;
24365af7 1607 unsigned long nr_dummy = 0;
273399de 1608
7ed7682f 1609 *approx_before = 0;
4c42f1b8 1610 if (split_count_mask >= 0) {
973e5e1b
MD
1611 int i;
1612
4c42f1b8
LJ
1613 for (i = 0; i < split_count_mask + 1; i++) {
1614 *approx_before += uatomic_read(&ht->split_count[i].add);
1615 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1616 }
1617 }
1618
273399de
MD
1619 *count = 0;
1620 *removed = 0;
1621
24365af7 1622 /* Count non-dummy nodes in the table */
4105056a 1623 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1624 node = (struct cds_lfht_node *) lookup;
273399de 1625 do {
cc4fcb10 1626 next = rcu_dereference(node->p.next);
b198f0fd 1627 if (is_removed(next)) {
973e5e1b
MD
1628 if (!is_dummy(next))
1629 (*removed)++;
1630 else
1631 (nr_dummy)++;
1b81fe1a 1632 } else if (!is_dummy(next))
273399de 1633 (*count)++;
24365af7
MD
1634 else
1635 (nr_dummy)++;
273399de 1636 node = clear_flag(next);
bb7b2f26 1637 } while (!is_end(node));
f0c29ed7 1638 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1639 *approx_after = 0;
4c42f1b8 1640 if (split_count_mask >= 0) {
973e5e1b
MD
1641 int i;
1642
4c42f1b8
LJ
1643 for (i = 0; i < split_count_mask + 1; i++) {
1644 *approx_after += uatomic_read(&ht->split_count[i].add);
1645 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1646 }
1647 }
273399de
MD
1648}
1649
1475579c 1650/* called with resize mutex held */
abc490a1 1651static
4105056a 1652void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1653 unsigned long old_size, unsigned long new_size)
abc490a1 1654{
1475579c 1655 unsigned long old_order, new_order;
1475579c 1656
93d46c39
LJ
1657 old_order = get_count_order_ulong(old_size);
1658 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1659 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1660 old_size, old_order, new_size, new_order);
1475579c 1661 assert(new_size > old_size);
93d46c39 1662 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1663}
1664
1665/* called with resize mutex held */
1666static
4105056a 1667void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1668 unsigned long old_size, unsigned long new_size)
464a1ec9 1669{
1475579c 1670 unsigned long old_order, new_order;
464a1ec9 1671
5488222b 1672 new_size = max(new_size, ht->min_alloc_size);
93d46c39
LJ
1673 old_order = get_count_order_ulong(old_size);
1674 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1675 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1676 old_size, old_order, new_size, new_order);
1475579c 1677 assert(new_size < old_size);
1475579c 1678
4105056a 1679 /* Remove and unlink all dummy nodes to remove. */
93d46c39 1680 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1681}
1682
1475579c
MD
1683
1684/* called with resize mutex held */
1685static
1686void _do_cds_lfht_resize(struct cds_lfht *ht)
1687{
1688 unsigned long new_size, old_size;
4105056a
MD
1689
1690 /*
1691 * Resize table, re-do if the target size has changed under us.
1692 */
1693 do {
d2be3620
MD
1694 assert(uatomic_read(&ht->in_progress_resize));
1695 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1696 break;
4105056a
MD
1697 ht->t.resize_initiated = 1;
1698 old_size = ht->t.size;
1699 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1700 if (old_size < new_size)
1701 _do_cds_lfht_grow(ht, old_size, new_size);
1702 else if (old_size > new_size)
1703 _do_cds_lfht_shrink(ht, old_size, new_size);
1704 ht->t.resize_initiated = 0;
1705 /* write resize_initiated before read resize_target */
1706 cmm_smp_mb();
4d676753 1707 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1708}
1709
abc490a1 1710static
4105056a 1711unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1712 int growth_order)
464a1ec9 1713{
4105056a
MD
1714 return _uatomic_max(&ht->t.resize_target,
1715 size << growth_order);
464a1ec9
MD
1716}
1717
1475579c 1718static
4105056a 1719void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1720 unsigned long count)
1475579c 1721{
5488222b 1722 count = max(count, ht->min_alloc_size);
4105056a 1723 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1724}
1725
1726void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1727{
4105056a
MD
1728 resize_target_update_count(ht, new_size);
1729 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1730 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1731 pthread_mutex_lock(&ht->resize_mutex);
1732 _do_cds_lfht_resize(ht);
1733 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1734 ht->cds_lfht_rcu_thread_online();
abc490a1 1735}
464a1ec9 1736
abc490a1
MD
1737static
1738void do_resize_cb(struct rcu_head *head)
1739{
1740 struct rcu_resize_work *work =
1741 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1742 struct cds_lfht *ht = work->ht;
abc490a1 1743
5f511391 1744 ht->cds_lfht_rcu_thread_offline();
abc490a1 1745 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1746 _do_cds_lfht_resize(ht);
abc490a1 1747 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1748 ht->cds_lfht_rcu_thread_online();
98808fb1 1749 poison_free(work);
848d4088
MD
1750 cmm_smp_mb(); /* finish resize before decrement */
1751 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1752}
1753
abc490a1 1754static
4105056a 1755void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1756{
abc490a1 1757 struct rcu_resize_work *work;
f9830efd 1758 unsigned long target_size;
abc490a1 1759
4105056a
MD
1760 target_size = resize_target_update(ht, size, growth);
1761 /* Store resize_target before read resize_initiated */
1762 cmm_smp_mb();
1763 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1764 uatomic_inc(&ht->in_progress_resize);
59290e9d 1765 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1766 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1767 uatomic_dec(&ht->in_progress_resize);
59290e9d 1768 return;
ed35e6d8 1769 }
f9830efd
MD
1770 work = malloc(sizeof(*work));
1771 work->ht = ht;
14044b37 1772 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1773 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1774 }
ab7d5fc6 1775}
3171717f
MD
1776
1777static
4105056a 1778void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1779 unsigned long count)
1780{
1781 struct rcu_resize_work *work;
3171717f 1782
b8af5011
MD
1783 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1784 return;
4105056a
MD
1785 resize_target_update_count(ht, count);
1786 /* Store resize_target before read resize_initiated */
1787 cmm_smp_mb();
1788 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1789 uatomic_inc(&ht->in_progress_resize);
59290e9d 1790 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1791 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1792 uatomic_dec(&ht->in_progress_resize);
59290e9d 1793 return;
ed35e6d8 1794 }
3171717f
MD
1795 work = malloc(sizeof(*work));
1796 work->ht = ht;
1797 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1798 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1799 }
1800}
This page took 0.213193 seconds and 4 git commands to generate.