Fix CAA_BITS_PER_lONG typo
[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{
6887cc5e 457#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
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 488#ifdef POISON_FREE
5afadd12
LJ
489#define poison_free(ptr) \
490 do { \
491 if (ptr) { \
492 memset(ptr, 0x42, sizeof(*(ptr))); \
493 free(ptr); \
494 } \
98808fb1
MD
495 } while (0)
496#else
497#define poison_free(ptr) free(ptr)
498#endif
499
f9830efd 500static
4105056a 501void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 502
f8994aee 503static
4105056a 504void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
505 unsigned long count);
506
df44348d 507static long nr_cpus_mask = -1;
4c42f1b8
LJ
508static long split_count_mask = -1;
509
4ddbb355 510#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
511static void ht_init_nr_cpus_mask(void)
512{
513 long maxcpus;
514
515 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
516 if (maxcpus <= 0) {
517 nr_cpus_mask = -2;
518 return;
519 }
520 /*
521 * round up number of CPUs to next power of two, so we
522 * can use & for modulo.
523 */
524 maxcpus = 1UL << get_count_order_ulong(maxcpus);
525 nr_cpus_mask = maxcpus - 1;
526}
4ddbb355
LJ
527#else /* #if defined(HAVE_SYSCONF) */
528static void ht_init_nr_cpus_mask(void)
529{
530 nr_cpus_mask = -2;
531}
532#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
533
534static
5afadd12 535void alloc_split_items_count(struct cds_lfht *ht)
df44348d
MD
536{
537 struct ht_items_count *count;
538
4c42f1b8
LJ
539 if (nr_cpus_mask == -1) {
540 ht_init_nr_cpus_mask();
4ddbb355
LJ
541 if (nr_cpus_mask < 0)
542 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
543 else
544 split_count_mask = nr_cpus_mask;
df44348d 545 }
4c42f1b8 546
4ddbb355 547 assert(split_count_mask >= 0);
5afadd12
LJ
548
549 if (ht->flags & CDS_LFHT_ACCOUNTING) {
550 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
551 assert(ht->split_count);
552 } else {
553 ht->split_count = NULL;
554 }
df44348d
MD
555}
556
557static
5afadd12 558void free_split_items_count(struct cds_lfht *ht)
df44348d 559{
5afadd12 560 poison_free(ht->split_count);
df44348d
MD
561}
562
14360f1c 563#if defined(HAVE_SCHED_GETCPU)
df44348d 564static
14360f1c 565int ht_get_split_count_index(unsigned long hash)
df44348d
MD
566{
567 int cpu;
568
4c42f1b8 569 assert(split_count_mask >= 0);
df44348d
MD
570 cpu = sched_getcpu();
571 if (unlikely(cpu < 0))
14360f1c 572 return hash & split_count_mask;
df44348d 573 else
4c42f1b8 574 return cpu & split_count_mask;
df44348d 575}
14360f1c
LJ
576#else /* #if defined(HAVE_SCHED_GETCPU) */
577static
578int ht_get_split_count_index(unsigned long hash)
579{
580 return hash & split_count_mask;
581}
582#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
583
584static
14360f1c 585void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 586{
4c42f1b8
LJ
587 unsigned long split_count;
588 int index;
df44348d 589
4c42f1b8 590 if (unlikely(!ht->split_count))
3171717f 591 return;
14360f1c 592 index = ht_get_split_count_index(hash);
4c42f1b8
LJ
593 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
594 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 595 long count;
df44348d 596
4c42f1b8 597 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
598 count = uatomic_add_return(&ht->count,
599 1UL << COUNT_COMMIT_ORDER);
600 /* If power of 2 */
601 if (!(count & (count - 1))) {
4105056a 602 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 603 return;
e3ecefd6 604 dbg_printf("add set global %ld\n", count);
4105056a 605 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 606 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
607 }
608 }
609}
610
611static
14360f1c 612void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 613{
4c42f1b8
LJ
614 unsigned long split_count;
615 int index;
df44348d 616
4c42f1b8 617 if (unlikely(!ht->split_count))
3171717f 618 return;
14360f1c 619 index = ht_get_split_count_index(hash);
4c42f1b8
LJ
620 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
621 if (unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 622 long count;
df44348d 623
4c42f1b8 624 dbg_printf("del split count %lu\n", split_count);
df44348d 625 count = uatomic_add_return(&ht->count,
3171717f 626 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
627 /* If power of 2 */
628 if (!(count & (count - 1))) {
4105056a 629 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 630 return;
e3ecefd6
MD
631 dbg_printf("del set global %ld\n", count);
632 /*
c941bb9e 633 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
634 * certain threshold.
635 */
4c42f1b8 636 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 637 return;
4105056a 638 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 639 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
640 }
641 }
642}
643
f9830efd 644static
4105056a 645void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 646{
f8994aee
MD
647 unsigned long count;
648
b8af5011
MD
649 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
650 return;
f8994aee
MD
651 count = uatomic_read(&ht->count);
652 /*
653 * Use bucket-local length for small table expand and for
654 * environments lacking per-cpu data support.
655 */
656 if (count >= (1UL << COUNT_COMMIT_ORDER))
657 return;
24365af7 658 if (chain_len > 100)
f0c29ed7 659 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 660 chain_len);
3390d470 661 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 662 cds_lfht_resize_lazy(ht, size,
01370f0b 663 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
664}
665
abc490a1 666static
14044b37 667struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 668{
14044b37 669 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
670}
671
672static
14044b37 673int is_removed(struct cds_lfht_node *node)
abc490a1 674{
d37166c6 675 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
676}
677
678static
14044b37 679struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 680{
14044b37 681 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
682}
683
f5596c94 684static
14044b37 685int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
686{
687 return ((unsigned long) node) & DUMMY_FLAG;
688}
689
690static
14044b37 691struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 692{
14044b37 693 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 694}
bb7b2f26
MD
695
696static
697struct cds_lfht_node *get_end(void)
698{
699 return (struct cds_lfht_node *) END_VALUE;
700}
701
702static
703int is_end(struct cds_lfht_node *node)
704{
705 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
706}
707
abc490a1 708static
f9830efd 709unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
710{
711 unsigned long old1, old2;
712
713 old1 = uatomic_read(ptr);
714 do {
715 old2 = old1;
716 if (old2 >= v)
f9830efd 717 return old2;
abc490a1 718 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 719 return v;
abc490a1
MD
720}
721
f4a9cc0b
LJ
722static
723struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
724 unsigned long hash)
725{
726 unsigned long index, order;
727
728 assert(size > 0);
729 index = hash & (size - 1);
ef6e6171
LJ
730
731 if (index < ht->min_alloc_size) {
732 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
733 hash, index);
734 return &ht->t.tbl[0]->nodes[index];
735 }
a4ea2223
LJ
736 /*
737 * equivalent to get_count_order_ulong(index + 1), but optimizes
738 * away the non-existing 0 special-case for
739 * get_count_order_ulong.
740 */
741 order = fls_ulong(index);
f4a9cc0b 742 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
ef6e6171
LJ
743 hash, index, order, index & ((1UL << (order - 1)) - 1));
744 return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
745}
746
273399de
MD
747/*
748 * Remove all logically deleted nodes from a bucket up to a certain node key.
749 */
750static
f9c80341 751void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 752{
14044b37 753 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 754
c90201ac
MD
755 assert(!is_dummy(dummy));
756 assert(!is_removed(dummy));
757 assert(!is_dummy(node));
758 assert(!is_removed(node));
273399de
MD
759 for (;;) {
760 iter_prev = dummy;
761 /* We can always skip the dummy node initially */
cc4fcb10 762 iter = rcu_dereference(iter_prev->p.next);
b4cb483f 763 assert(!is_removed(iter));
cc4fcb10 764 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
765 /*
766 * We should never be called with dummy (start of chain)
767 * and logically removed node (end of path compression
768 * marker) being the actual same node. This would be a
769 * bug in the algorithm implementation.
770 */
771 assert(dummy != node);
273399de 772 for (;;) {
bb7b2f26 773 if (unlikely(is_end(iter)))
f9c80341 774 return;
76412f24 775 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 776 return;
cc4fcb10 777 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 778 if (likely(is_removed(next)))
273399de 779 break;
b453eae1 780 iter_prev = clear_flag(iter);
273399de
MD
781 iter = next;
782 }
b198f0fd 783 assert(!is_removed(iter));
f5596c94
MD
784 if (is_dummy(iter))
785 new_next = flag_dummy(clear_flag(next));
786 else
787 new_next = clear_flag(next);
788 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 789 }
f9c80341 790 return;
273399de
MD
791}
792
9357c415
MD
793static
794int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
795 struct cds_lfht_node *old_node,
3fb86f26 796 struct cds_lfht_node *old_next,
9357c415
MD
797 struct cds_lfht_node *new_node)
798{
3fb86f26 799 struct cds_lfht_node *dummy, *ret_next;
9357c415 800 struct _cds_lfht_node *lookup;
9357c415
MD
801
802 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 803 return -ENOENT;
9357c415
MD
804
805 assert(!is_removed(old_node));
806 assert(!is_dummy(old_node));
807 assert(!is_removed(new_node));
808 assert(!is_dummy(new_node));
809 assert(new_node != old_node);
3fb86f26 810 for (;;) {
9357c415 811 /* Insert after node to be replaced */
9357c415
MD
812 if (is_removed(old_next)) {
813 /*
814 * Too late, the old node has been removed under us
815 * between lookup and replace. Fail.
816 */
7801dadd 817 return -ENOENT;
9357c415
MD
818 }
819 assert(!is_dummy(old_next));
820 assert(new_node != clear_flag(old_next));
821 new_node->p.next = clear_flag(old_next);
822 /*
823 * Here is the whole trick for lock-free replace: we add
824 * the replacement node _after_ the node we want to
825 * replace by atomically setting its next pointer at the
826 * same time we set its removal flag. Given that
827 * the lookups/get next use an iterator aware of the
828 * next pointer, they will either skip the old node due
829 * to the removal flag and see the new node, or use
830 * the old node, but will not see the new one.
831 */
832 ret_next = uatomic_cmpxchg(&old_node->p.next,
833 old_next, flag_removed(new_node));
3fb86f26 834 if (ret_next == old_next)
7801dadd 835 break; /* We performed the replacement. */
3fb86f26
LJ
836 old_next = ret_next;
837 }
9357c415 838
9357c415
MD
839 /*
840 * Ensure that the old node is not visible to readers anymore:
841 * lookup for the node, and remove it (along with any other
842 * logically removed node) if found.
843 */
f4a9cc0b 844 lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
9357c415
MD
845 dummy = (struct cds_lfht_node *) lookup;
846 _cds_lfht_gc_bucket(dummy, new_node);
7801dadd
LJ
847
848 assert(is_removed(rcu_dereference(old_node->p.next)));
849 return 0;
9357c415
MD
850}
851
83beee94
MD
852/*
853 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
854 * mode. A NULL unique_ret allows creation of duplicate keys.
855 */
abc490a1 856static
83beee94
MD
857void _cds_lfht_add(struct cds_lfht *ht,
858 unsigned long size,
859 struct cds_lfht_node *node,
860 struct cds_lfht_iter *unique_ret,
861 int dummy)
abc490a1 862{
14044b37 863 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 864 *return_node;
14044b37 865 struct _cds_lfht_node *lookup;
abc490a1 866
c90201ac
MD
867 assert(!is_dummy(node));
868 assert(!is_removed(node));
f4a9cc0b 869 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
abc490a1 870 for (;;) {
adc0de68 871 uint32_t chain_len = 0;
abc490a1 872
11519af6
MD
873 /*
874 * iter_prev points to the non-removed node prior to the
875 * insert location.
11519af6 876 */
14044b37 877 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 878 /* We can always skip the dummy node initially */
cc4fcb10
MD
879 iter = rcu_dereference(iter_prev->p.next);
880 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 881 for (;;) {
bb7b2f26 882 if (unlikely(is_end(iter)))
273399de 883 goto insert;
76412f24 884 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 885 goto insert;
238cc06e 886
194fdbd1
LJ
887 /* dummy node is the first node of the identical-hash-value chain */
888 if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
889 goto insert;
238cc06e 890
cc4fcb10 891 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 892 if (unlikely(is_removed(next)))
9dba85be 893 goto gc_node;
238cc06e
LJ
894
895 /* uniquely add */
83beee94 896 if (unique_ret
1b81fe1a 897 && !is_dummy(next)
238cc06e
LJ
898 && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
899 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
900
901 /*
902 * uniquely adding inserts the node as the first
903 * node of the identical-hash-value node chain.
904 *
905 * This semantic ensures no duplicated keys
906 * should ever be observable in the table
907 * (including observe one node by one node
908 * by forward iterations)
909 */
910 cds_lfht_next_duplicate(ht, &d_iter);
911 if (!d_iter.node)
912 goto insert;
913
914 *unique_ret = d_iter;
83beee94 915 return;
48ed1c18 916 }
238cc06e 917
11519af6 918 /* Only account for identical reverse hash once */
24365af7
MD
919 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
920 && !is_dummy(next))
4105056a 921 check_resize(ht, size, ++chain_len);
11519af6 922 iter_prev = clear_flag(iter);
273399de 923 iter = next;
abc490a1 924 }
48ed1c18 925
273399de 926 insert:
7ec59d3b 927 assert(node != clear_flag(iter));
11519af6 928 assert(!is_removed(iter_prev));
c90201ac 929 assert(!is_removed(iter));
f000907d 930 assert(iter_prev != node);
f9c80341 931 if (!dummy)
1b81fe1a 932 node->p.next = clear_flag(iter);
f9c80341
MD
933 else
934 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
935 if (is_dummy(iter))
936 new_node = flag_dummy(node);
937 else
938 new_node = node;
cc4fcb10 939 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 940 new_node) != iter) {
273399de 941 continue; /* retry */
48ed1c18 942 } else {
83beee94 943 return_node = node;
960c9e4f 944 goto end;
48ed1c18
MD
945 }
946
9dba85be
MD
947 gc_node:
948 assert(!is_removed(iter));
f5596c94
MD
949 if (is_dummy(iter))
950 new_next = flag_dummy(clear_flag(next));
951 else
952 new_next = clear_flag(next);
953 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 954 /* retry */
464a1ec9 955 }
9357c415 956end:
83beee94
MD
957 if (unique_ret) {
958 unique_ret->node = return_node;
959 /* unique_ret->next left unset, never used. */
960 }
abc490a1 961}
464a1ec9 962
abc490a1 963static
860d07e8 964int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 965 struct cds_lfht_node *node,
b198f0fd 966 int dummy_removal)
abc490a1 967{
14044b37
MD
968 struct cds_lfht_node *dummy, *next, *old;
969 struct _cds_lfht_node *lookup;
5e28c532 970
9357c415 971 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 972 return -ENOENT;
9357c415 973
7ec59d3b 974 /* logically delete the node */
c90201ac
MD
975 assert(!is_dummy(node));
976 assert(!is_removed(node));
cc4fcb10 977 old = rcu_dereference(node->p.next);
7ec59d3b 978 do {
48ed1c18
MD
979 struct cds_lfht_node *new_next;
980
7ec59d3b 981 next = old;
76412f24 982 if (unlikely(is_removed(next)))
743f9143 983 return -ENOENT;
1475579c
MD
984 if (dummy_removal)
985 assert(is_dummy(next));
986 else
987 assert(!is_dummy(next));
48ed1c18 988 new_next = flag_removed(next);
48ed1c18 989 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b 990 } while (old != next);
7ec59d3b 991 /* We performed the (logical) deletion. */
7ec59d3b
MD
992
993 /*
994 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
995 * the node, and remove it (along with any other logically removed node)
996 * if found.
11519af6 997 */
f4a9cc0b 998 lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
14044b37 999 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1000 _cds_lfht_gc_bucket(dummy, node);
743f9143
LJ
1001
1002 assert(is_removed(rcu_dereference(node->p.next)));
1003 return 0;
abc490a1 1004}
2ed95849 1005
b7d619b0
MD
1006static
1007void *partition_resize_thread(void *arg)
1008{
1009 struct partition_resize_work *work = arg;
1010
1011 work->ht->cds_lfht_rcu_register_thread();
1012 work->fct(work->ht, work->i, work->start, work->len);
1013 work->ht->cds_lfht_rcu_unregister_thread();
1014 return NULL;
1015}
1016
1017static
1018void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1019 unsigned long len,
1020 void (*fct)(struct cds_lfht *ht, unsigned long i,
1021 unsigned long start, unsigned long len))
1022{
1023 unsigned long partition_len;
1024 struct partition_resize_work *work;
6083a889
MD
1025 int thread, ret;
1026 unsigned long nr_threads;
b7d619b0 1027
6083a889
MD
1028 /*
1029 * Note: nr_cpus_mask + 1 is always power of 2.
1030 * We spawn just the number of threads we need to satisfy the minimum
1031 * partition size, up to the number of CPUs in the system.
1032 */
91452a6a
MD
1033 if (nr_cpus_mask > 0) {
1034 nr_threads = min(nr_cpus_mask + 1,
1035 len >> MIN_PARTITION_PER_THREAD_ORDER);
1036 } else {
1037 nr_threads = 1;
1038 }
6083a889
MD
1039 partition_len = len >> get_count_order_ulong(nr_threads);
1040 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1041 assert(work);
6083a889
MD
1042 for (thread = 0; thread < nr_threads; thread++) {
1043 work[thread].ht = ht;
1044 work[thread].i = i;
1045 work[thread].len = partition_len;
1046 work[thread].start = thread * partition_len;
1047 work[thread].fct = fct;
1af6e26e 1048 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1049 partition_resize_thread, &work[thread]);
b7d619b0
MD
1050 assert(!ret);
1051 }
6083a889 1052 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1053 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1054 assert(!ret);
1055 }
1056 free(work);
b7d619b0
MD
1057}
1058
e8de508e
MD
1059/*
1060 * Holding RCU read lock to protect _cds_lfht_add against memory
1061 * reclaim that could be performed by other call_rcu worker threads (ABA
1062 * problem).
9ee0fc9a 1063 *
b7d619b0 1064 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1065 * many worker threads, based on the number of CPUs available in the system.
1066 * This should therefore take care of not having the expand lagging behind too
1067 * many concurrent insertion threads by using the scheduler's ability to
1068 * schedule dummy node population fairly with insertions.
e8de508e 1069 */
4105056a 1070static
b7d619b0
MD
1071void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1072 unsigned long start, unsigned long len)
4105056a
MD
1073{
1074 unsigned long j;
1075
5488222b 1076 assert(i > ht->min_alloc_order);
4105056a 1077 ht->cds_lfht_rcu_read_lock();
b7d619b0 1078 for (j = start; j < start + len; j++) {
4105056a
MD
1079 struct cds_lfht_node *new_node =
1080 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1081
dc1da8f6 1082 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1083 i, j, (1UL << (i - 1)) + j);
dc1da8f6 1084 new_node->p.reverse_hash =
4f6e90b7
LJ
1085 bit_reverse_ulong((1UL << (i - 1)) + j);
1086 _cds_lfht_add(ht, 1UL << (i - 1),
83beee94 1087 new_node, NULL, 1);
4105056a
MD
1088 }
1089 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1090}
1091
1092static
1093void init_table_populate(struct cds_lfht *ht, unsigned long i,
1094 unsigned long len)
1095{
1096 assert(nr_cpus_mask != -1);
6083a889 1097 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1098 ht->cds_lfht_rcu_thread_online();
1099 init_table_populate_partition(ht, i, 0, len);
1100 ht->cds_lfht_rcu_thread_offline();
1101 return;
1102 }
1103 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1104}
1105
abc490a1 1106static
4105056a 1107void init_table(struct cds_lfht *ht,
93d46c39 1108 unsigned long first_order, unsigned long last_order)
24365af7 1109{
93d46c39 1110 unsigned long i;
24365af7 1111
93d46c39
LJ
1112 dbg_printf("init table: first_order %lu last_order %lu\n",
1113 first_order, last_order);
5488222b 1114 assert(first_order > ht->min_alloc_order);
93d46c39 1115 for (i = first_order; i <= last_order; i++) {
4105056a 1116 unsigned long len;
24365af7 1117
4f6e90b7 1118 len = 1UL << (i - 1);
f0c29ed7 1119 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1120
1121 /* Stop expand if the resize target changes under us */
4f6e90b7 1122 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1123 break;
1124
0d14ceb2 1125 ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
b7d619b0 1126 assert(ht->t.tbl[i]);
4105056a 1127
4105056a 1128 /*
dc1da8f6
MD
1129 * Set all dummy nodes reverse hash values for a level and
1130 * link all dummy nodes into the table.
4105056a 1131 */
dc1da8f6 1132 init_table_populate(ht, i, len);
4105056a 1133
f9c80341
MD
1134 /*
1135 * Update table size.
1136 */
1137 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1138 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1139
4f6e90b7 1140 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1141 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1142 break;
1143 }
1144}
1145
e8de508e
MD
1146/*
1147 * Holding RCU read lock to protect _cds_lfht_remove against memory
1148 * reclaim that could be performed by other call_rcu worker threads (ABA
1149 * problem).
1150 * For a single level, we logically remove and garbage collect each node.
1151 *
1152 * As a design choice, we perform logical removal and garbage collection on a
1153 * node-per-node basis to simplify this algorithm. We also assume keeping good
1154 * cache locality of the operation would overweight possible performance gain
1155 * that could be achieved by batching garbage collection for multiple levels.
1156 * However, this would have to be justified by benchmarks.
1157 *
1158 * Concurrent removal and add operations are helping us perform garbage
1159 * collection of logically removed nodes. We guarantee that all logically
1160 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1161 * invoked to free a hole level of dummy nodes (after a grace period).
1162 *
1163 * Logical removal and garbage collection can therefore be done in batch or on a
1164 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1165 *
b7d619b0
MD
1166 * When we reach a certain length, we can split this removal over many worker
1167 * threads, based on the number of CPUs available in the system. This should
1168 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1169 * updater threads actively inserting into the hash table.
e8de508e 1170 */
4105056a 1171static
b7d619b0
MD
1172void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1173 unsigned long start, unsigned long len)
4105056a
MD
1174{
1175 unsigned long j;
1176
5488222b 1177 assert(i > ht->min_alloc_order);
4105056a 1178 ht->cds_lfht_rcu_read_lock();
b7d619b0 1179 for (j = start; j < start + len; j++) {
4105056a
MD
1180 struct cds_lfht_node *fini_node =
1181 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1182
1183 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1184 i, j, (1UL << (i - 1)) + j);
4105056a 1185 fini_node->p.reverse_hash =
4f6e90b7
LJ
1186 bit_reverse_ulong((1UL << (i - 1)) + j);
1187 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1188 }
4105056a 1189 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1190}
1191
1192static
1193void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1194{
1195
1196 assert(nr_cpus_mask != -1);
6083a889 1197 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1198 ht->cds_lfht_rcu_thread_online();
1199 remove_table_partition(ht, i, 0, len);
1200 ht->cds_lfht_rcu_thread_offline();
1201 return;
1202 }
1203 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1204}
1205
1475579c 1206static
4105056a 1207void fini_table(struct cds_lfht *ht,
93d46c39 1208 unsigned long first_order, unsigned long last_order)
1475579c 1209{
93d46c39 1210 long i;
0d14ceb2 1211 void *free_by_rcu = NULL;
1475579c 1212
93d46c39
LJ
1213 dbg_printf("fini table: first_order %lu last_order %lu\n",
1214 first_order, last_order);
5488222b 1215 assert(first_order > ht->min_alloc_order);
93d46c39 1216 for (i = last_order; i >= first_order; i--) {
4105056a 1217 unsigned long len;
1475579c 1218
4f6e90b7 1219 len = 1UL << (i - 1);
1475579c 1220 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1221
4d676753
MD
1222 /* Stop shrink if the resize target changes under us */
1223 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1224 break;
1225
1226 cmm_smp_wmb(); /* populate data before RCU size */
1227 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1228
1229 /*
1230 * We need to wait for all add operations to reach Q.S. (and
1231 * thus use the new table for lookups) before we can start
1232 * releasing the old dummy nodes. Otherwise their lookup will
1233 * return a logically removed node as insert position.
1234 */
1235 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1236 if (free_by_rcu)
1237 free(free_by_rcu);
4d676753 1238
21263e21 1239 /*
4105056a
MD
1240 * Set "removed" flag in dummy nodes about to be removed.
1241 * Unlink all now-logically-removed dummy node pointers.
1242 * Concurrent add/remove operation are helping us doing
1243 * the gc.
21263e21 1244 */
4105056a
MD
1245 remove_table(ht, i, len);
1246
0d14ceb2 1247 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1248
1249 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1250 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1251 break;
1252 }
0d14ceb2
LJ
1253
1254 if (free_by_rcu) {
1255 ht->cds_lfht_synchronize_rcu();
1256 free(free_by_rcu);
1257 }
1475579c
MD
1258}
1259
ff0d69de
LJ
1260static
1261void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
1262{
1263 struct _cds_lfht_node *prev, *node;
1264 unsigned long order, len, i, j;
1265
5488222b 1266 ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct _cds_lfht_node));
ff0d69de
LJ
1267 assert(ht->t.tbl[0]);
1268
1269 dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
1270 ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end());
1271 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1272
1273 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1274 len = 1UL << (order - 1);
5488222b 1275 if (order <= ht->min_alloc_order) {
eb631bf2 1276 ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
5488222b
LJ
1277 } else {
1278 ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
1279 assert(ht->t.tbl[order]);
1280 }
ff0d69de
LJ
1281
1282 i = 0;
1283 prev = ht->t.tbl[i]->nodes;
1284 for (j = 0; j < len; j++) {
1285 if (j & (j - 1)) { /* Between power of 2 */
1286 prev++;
1287 } else if (j) { /* At each power of 2 */
1288 i++;
1289 prev = ht->t.tbl[i]->nodes;
1290 }
1291
1292 node = &ht->t.tbl[order]->nodes[j];
1293 dbg_printf("create dummy: order %lu index %lu hash %lu\n",
1294 order, j, j + len);
1295 node->next = prev->next;
1296 assert(is_dummy(node->next));
1297 node->reverse_hash = bit_reverse_ulong(j + len);
1298 prev->next = flag_dummy((struct cds_lfht_node *)node);
1299 }
1300 }
1301}
1302
7a9dcf9b 1303struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1304 cds_lfht_compare_fct compare_fct,
1305 unsigned long hash_seed,
1306 unsigned long init_size,
5488222b 1307 unsigned long min_alloc_size,
b8af5011 1308 int flags,
14044b37 1309 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1310 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1311 void (*cds_lfht_synchronize_rcu)(void),
1312 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1313 void (*cds_lfht_rcu_read_unlock)(void),
1314 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1315 void (*cds_lfht_rcu_thread_online)(void),
1316 void (*cds_lfht_rcu_register_thread)(void),
1317 void (*cds_lfht_rcu_unregister_thread)(void),
1318 pthread_attr_t *attr)
abc490a1 1319{
14044b37 1320 struct cds_lfht *ht;
24365af7 1321 unsigned long order;
abc490a1 1322
5488222b
LJ
1323 /* min_alloc_size must be power of two */
1324 if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
1325 return NULL;
8129be4e 1326 /* init_size must be power of two */
5488222b 1327 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1328 return NULL;
5488222b
LJ
1329 min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
1330 init_size = max(init_size, min_alloc_size);
14044b37 1331 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1332 assert(ht);
b5d6b20f 1333 ht->flags = flags;
abc490a1 1334 ht->hash_fct = hash_fct;
732ad076
MD
1335 ht->compare_fct = compare_fct;
1336 ht->hash_seed = hash_seed;
14044b37 1337 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1338 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1339 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1340 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1341 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1342 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1343 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1344 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1345 ht->resize_attr = attr;
5afadd12 1346 alloc_split_items_count(ht);
abc490a1
MD
1347 /* this mutex should not nest in read-side C.S. */
1348 pthread_mutex_init(&ht->resize_mutex, NULL);
5488222b 1349 order = get_count_order_ulong(init_size);
93d46c39 1350 ht->t.resize_target = 1UL << order;
5488222b
LJ
1351 ht->min_alloc_size = min_alloc_size;
1352 ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
bcbd36fc
LJ
1353 cds_lfht_create_dummy(ht, 1UL << order);
1354 ht->t.size = 1UL << order;
abc490a1
MD
1355 return ht;
1356}
1357
adc0de68
MD
1358void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1359 struct cds_lfht_iter *iter)
2ed95849 1360{
bb7b2f26 1361 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1362 struct _cds_lfht_node *lookup;
f4a9cc0b 1363 unsigned long hash, reverse_hash, size;
2ed95849 1364
732ad076 1365 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1366 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1367
4105056a 1368 size = rcu_dereference(ht->t.size);
f4a9cc0b 1369 lookup = lookup_bucket(ht, size, hash);
bb7b2f26
MD
1370 dummy_node = (struct cds_lfht_node *) lookup;
1371 /* We can always skip the dummy node initially */
1372 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1373 node = clear_flag(node);
2ed95849 1374 for (;;) {
bb7b2f26 1375 if (unlikely(is_end(node))) {
96ad1112 1376 node = next = NULL;
abc490a1 1377 break;
bb7b2f26 1378 }
cc4fcb10 1379 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1380 node = next = NULL;
abc490a1 1381 break;
2ed95849 1382 }
1b81fe1a 1383 next = rcu_dereference(node->p.next);
adc0de68 1384 if (likely(!is_removed(next))
1b81fe1a 1385 && !is_dummy(next)
ae450da3 1386 && clear_flag(node)->p.reverse_hash == reverse_hash
49c2e2d6 1387 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1388 break;
2ed95849 1389 }
1b81fe1a 1390 node = clear_flag(next);
2ed95849 1391 }
1b81fe1a 1392 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1393 iter->node = node;
1394 iter->next = next;
abc490a1 1395}
e0ba718a 1396
3883c0e5 1397void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1398{
adc0de68 1399 struct cds_lfht_node *node, *next;
a481e5ff
MD
1400 unsigned long reverse_hash;
1401 void *key;
1402 size_t key_len;
1403
adc0de68 1404 node = iter->node;
a481e5ff
MD
1405 reverse_hash = node->p.reverse_hash;
1406 key = node->key;
1407 key_len = node->key_len;
adc0de68 1408 next = iter->next;
a481e5ff
MD
1409 node = clear_flag(next);
1410
1411 for (;;) {
bb7b2f26 1412 if (unlikely(is_end(node))) {
96ad1112 1413 node = next = NULL;
a481e5ff 1414 break;
bb7b2f26 1415 }
a481e5ff 1416 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1417 node = next = NULL;
a481e5ff
MD
1418 break;
1419 }
1420 next = rcu_dereference(node->p.next);
adc0de68 1421 if (likely(!is_removed(next))
a481e5ff
MD
1422 && !is_dummy(next)
1423 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1424 break;
1425 }
1426 node = clear_flag(next);
1427 }
1428 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1429 iter->node = node;
1430 iter->next = next;
a481e5ff
MD
1431}
1432
4e9b9fbf
MD
1433void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1434{
1435 struct cds_lfht_node *node, *next;
1436
853395e1 1437 node = clear_flag(iter->next);
4e9b9fbf
MD
1438 for (;;) {
1439 if (unlikely(is_end(node))) {
1440 node = next = NULL;
1441 break;
1442 }
1443 next = rcu_dereference(node->p.next);
1444 if (likely(!is_removed(next))
1445 && !is_dummy(next)) {
1446 break;
1447 }
1448 node = clear_flag(next);
1449 }
1450 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1451 iter->node = node;
1452 iter->next = next;
1453}
1454
1455void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1456{
1457 struct _cds_lfht_node *lookup;
1458
1459 /*
1460 * Get next after first dummy node. The first dummy node is the
1461 * first node of the linked list.
1462 */
1463 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1464 iter->next = lookup->next;
4e9b9fbf
MD
1465 cds_lfht_next(ht, iter);
1466}
1467
14044b37 1468void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1469{
4105056a 1470 unsigned long hash, size;
ab7d5fc6 1471
49c2e2d6 1472 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1473 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1474
4105056a 1475 size = rcu_dereference(ht->t.size);
83beee94 1476 _cds_lfht_add(ht, size, node, NULL, 0);
14360f1c 1477 ht_count_add(ht, size, hash);
3eca1b8c
MD
1478}
1479
14044b37 1480struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1481 struct cds_lfht_node *node)
3eca1b8c 1482{
4105056a 1483 unsigned long hash, size;
83beee94 1484 struct cds_lfht_iter iter;
3eca1b8c 1485
49c2e2d6 1486 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1487 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1488
4105056a 1489 size = rcu_dereference(ht->t.size);
83beee94
MD
1490 _cds_lfht_add(ht, size, node, &iter, 0);
1491 if (iter.node == node)
14360f1c 1492 ht_count_add(ht, size, hash);
83beee94 1493 return iter.node;
2ed95849
MD
1494}
1495
9357c415 1496struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
48ed1c18
MD
1497 struct cds_lfht_node *node)
1498{
1499 unsigned long hash, size;
83beee94 1500 struct cds_lfht_iter iter;
48ed1c18
MD
1501
1502 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1503 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1504
1505 size = rcu_dereference(ht->t.size);
83beee94
MD
1506 for (;;) {
1507 _cds_lfht_add(ht, size, node, &iter, 0);
1508 if (iter.node == node) {
14360f1c 1509 ht_count_add(ht, size, hash);
83beee94
MD
1510 return NULL;
1511 }
1512
1513 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1514 return iter.node;
1515 }
48ed1c18
MD
1516}
1517
9357c415
MD
1518int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1519 struct cds_lfht_node *new_node)
1520{
1521 unsigned long size;
1522
1523 size = rcu_dereference(ht->t.size);
1524 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1525 new_node);
1526}
1527
1528int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1529{
14360f1c 1530 unsigned long size, hash;
df44348d 1531 int ret;
abc490a1 1532
4105056a 1533 size = rcu_dereference(ht->t.size);
9357c415 1534 ret = _cds_lfht_del(ht, size, iter->node, 0);
14360f1c
LJ
1535 if (!ret) {
1536 hash = bit_reverse_ulong(iter->node->p.reverse_hash);
1537 ht_count_del(ht, size, hash);
1538 }
df44348d 1539 return ret;
2ed95849 1540}
ab7d5fc6 1541
abc490a1 1542static
14044b37 1543int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1544{
14044b37
MD
1545 struct cds_lfht_node *node;
1546 struct _cds_lfht_node *lookup;
4105056a 1547 unsigned long order, i, size;
674f7a69 1548
abc490a1 1549 /* Check that the table is empty */
4105056a 1550 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1551 node = (struct cds_lfht_node *) lookup;
abc490a1 1552 do {
1b81fe1a
MD
1553 node = clear_flag(node)->p.next;
1554 if (!is_dummy(node))
abc490a1 1555 return -EPERM;
273399de 1556 assert(!is_removed(node));
bb7b2f26 1557 } while (!is_end(node));
4105056a
MD
1558 /*
1559 * size accessed without rcu_dereference because hash table is
1560 * being destroyed.
1561 */
1562 size = ht->t.size;
abc490a1 1563 /* Internal sanity check: all nodes left should be dummy */
4105056a 1564 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1565 unsigned long len;
1566
1567 len = !order ? 1 : 1UL << (order - 1);
1568 for (i = 0; i < len; i++) {
f0c29ed7 1569 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1570 order, i,
4105056a
MD
1571 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1572 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1573 }
5488222b
LJ
1574
1575 if (order == ht->min_alloc_order)
1576 poison_free(ht->t.tbl[0]);
1577 else if (order > ht->min_alloc_order)
1578 poison_free(ht->t.tbl[order]);
1579 /* Nothing to delete for order < ht->min_alloc_order */
674f7a69 1580 }
abc490a1 1581 return 0;
674f7a69
MD
1582}
1583
1584/*
1585 * Should only be called when no more concurrent readers nor writers can
1586 * possibly access the table.
1587 */
b7d619b0 1588int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1589{
5e28c532
MD
1590 int ret;
1591
848d4088 1592 /* Wait for in-flight resize operations to complete */
24953e08
MD
1593 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1594 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1595 while (uatomic_read(&ht->in_progress_resize))
1596 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1597 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1598 if (ret)
1599 return ret;
5afadd12 1600 free_split_items_count(ht);
b7d619b0
MD
1601 if (attr)
1602 *attr = ht->resize_attr;
98808fb1 1603 poison_free(ht);
5e28c532 1604 return ret;
674f7a69
MD
1605}
1606
14044b37 1607void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1608 long *approx_before,
273399de 1609 unsigned long *count,
973e5e1b 1610 unsigned long *removed,
d933dd0e 1611 long *approx_after)
273399de 1612{
14044b37
MD
1613 struct cds_lfht_node *node, *next;
1614 struct _cds_lfht_node *lookup;
24365af7 1615 unsigned long nr_dummy = 0;
273399de 1616
7ed7682f 1617 *approx_before = 0;
5afadd12 1618 if (ht->split_count) {
973e5e1b
MD
1619 int i;
1620
4c42f1b8
LJ
1621 for (i = 0; i < split_count_mask + 1; i++) {
1622 *approx_before += uatomic_read(&ht->split_count[i].add);
1623 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1624 }
1625 }
1626
273399de
MD
1627 *count = 0;
1628 *removed = 0;
1629
24365af7 1630 /* Count non-dummy nodes in the table */
4105056a 1631 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1632 node = (struct cds_lfht_node *) lookup;
273399de 1633 do {
cc4fcb10 1634 next = rcu_dereference(node->p.next);
b198f0fd 1635 if (is_removed(next)) {
973e5e1b
MD
1636 if (!is_dummy(next))
1637 (*removed)++;
1638 else
1639 (nr_dummy)++;
1b81fe1a 1640 } else if (!is_dummy(next))
273399de 1641 (*count)++;
24365af7
MD
1642 else
1643 (nr_dummy)++;
273399de 1644 node = clear_flag(next);
bb7b2f26 1645 } while (!is_end(node));
f0c29ed7 1646 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1647 *approx_after = 0;
5afadd12 1648 if (ht->split_count) {
973e5e1b
MD
1649 int i;
1650
4c42f1b8
LJ
1651 for (i = 0; i < split_count_mask + 1; i++) {
1652 *approx_after += uatomic_read(&ht->split_count[i].add);
1653 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1654 }
1655 }
273399de
MD
1656}
1657
1475579c 1658/* called with resize mutex held */
abc490a1 1659static
4105056a 1660void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1661 unsigned long old_size, unsigned long new_size)
abc490a1 1662{
1475579c 1663 unsigned long old_order, new_order;
1475579c 1664
93d46c39
LJ
1665 old_order = get_count_order_ulong(old_size);
1666 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1667 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1668 old_size, old_order, new_size, new_order);
1475579c 1669 assert(new_size > old_size);
93d46c39 1670 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1671}
1672
1673/* called with resize mutex held */
1674static
4105056a 1675void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1676 unsigned long old_size, unsigned long new_size)
464a1ec9 1677{
1475579c 1678 unsigned long old_order, new_order;
464a1ec9 1679
5488222b 1680 new_size = max(new_size, ht->min_alloc_size);
93d46c39
LJ
1681 old_order = get_count_order_ulong(old_size);
1682 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1683 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1684 old_size, old_order, new_size, new_order);
1475579c 1685 assert(new_size < old_size);
1475579c 1686
4105056a 1687 /* Remove and unlink all dummy nodes to remove. */
93d46c39 1688 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1689}
1690
1475579c
MD
1691
1692/* called with resize mutex held */
1693static
1694void _do_cds_lfht_resize(struct cds_lfht *ht)
1695{
1696 unsigned long new_size, old_size;
4105056a
MD
1697
1698 /*
1699 * Resize table, re-do if the target size has changed under us.
1700 */
1701 do {
d2be3620
MD
1702 assert(uatomic_read(&ht->in_progress_resize));
1703 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1704 break;
4105056a
MD
1705 ht->t.resize_initiated = 1;
1706 old_size = ht->t.size;
1707 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1708 if (old_size < new_size)
1709 _do_cds_lfht_grow(ht, old_size, new_size);
1710 else if (old_size > new_size)
1711 _do_cds_lfht_shrink(ht, old_size, new_size);
1712 ht->t.resize_initiated = 0;
1713 /* write resize_initiated before read resize_target */
1714 cmm_smp_mb();
4d676753 1715 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1716}
1717
abc490a1 1718static
4105056a 1719unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1720 int growth_order)
464a1ec9 1721{
4105056a
MD
1722 return _uatomic_max(&ht->t.resize_target,
1723 size << growth_order);
464a1ec9
MD
1724}
1725
1475579c 1726static
4105056a 1727void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1728 unsigned long count)
1475579c 1729{
5488222b 1730 count = max(count, ht->min_alloc_size);
4105056a 1731 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1732}
1733
1734void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1735{
4105056a
MD
1736 resize_target_update_count(ht, new_size);
1737 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1738 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1739 pthread_mutex_lock(&ht->resize_mutex);
1740 _do_cds_lfht_resize(ht);
1741 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1742 ht->cds_lfht_rcu_thread_online();
abc490a1 1743}
464a1ec9 1744
abc490a1
MD
1745static
1746void do_resize_cb(struct rcu_head *head)
1747{
1748 struct rcu_resize_work *work =
1749 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1750 struct cds_lfht *ht = work->ht;
abc490a1 1751
5f511391 1752 ht->cds_lfht_rcu_thread_offline();
abc490a1 1753 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1754 _do_cds_lfht_resize(ht);
abc490a1 1755 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1756 ht->cds_lfht_rcu_thread_online();
98808fb1 1757 poison_free(work);
848d4088
MD
1758 cmm_smp_mb(); /* finish resize before decrement */
1759 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1760}
1761
abc490a1 1762static
4105056a 1763void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1764{
abc490a1 1765 struct rcu_resize_work *work;
f9830efd 1766 unsigned long target_size;
abc490a1 1767
4105056a
MD
1768 target_size = resize_target_update(ht, size, growth);
1769 /* Store resize_target before read resize_initiated */
1770 cmm_smp_mb();
1771 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088 1772 uatomic_inc(&ht->in_progress_resize);
59290e9d 1773 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1774 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1775 uatomic_dec(&ht->in_progress_resize);
59290e9d 1776 return;
ed35e6d8 1777 }
f9830efd
MD
1778 work = malloc(sizeof(*work));
1779 work->ht = ht;
14044b37 1780 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1781 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1782 }
ab7d5fc6 1783}
3171717f
MD
1784
1785static
4105056a 1786void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1787 unsigned long count)
1788{
1789 struct rcu_resize_work *work;
3171717f 1790
b8af5011
MD
1791 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1792 return;
4105056a
MD
1793 resize_target_update_count(ht, count);
1794 /* Store resize_target before read resize_initiated */
1795 cmm_smp_mb();
1796 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f 1797 uatomic_inc(&ht->in_progress_resize);
59290e9d 1798 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1799 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1800 uatomic_dec(&ht->in_progress_resize);
59290e9d 1801 return;
ed35e6d8 1802 }
3171717f
MD
1803 work = malloc(sizeof(*work));
1804 work->ht = ht;
1805 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1806 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1807 }
1808}
This page took 0.220315 seconds and 4 git commands to generate.