Add Lai Jiangshan's copyright to rculfhash
[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>
0dcf4847 7 * Copyright 2011 - Lai Jiangshan <laijs@cn.fujitsu.com>
abc490a1
MD
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
5e28c532
MD
22 */
23
e753ff5a
MD
24/*
25 * Based on the following articles:
26 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
27 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
28 * - Michael, M. M. High performance dynamic lock-free hash tables
29 * and list-based sets. In Proceedings of the fourteenth annual ACM
30 * symposium on Parallel algorithms and architectures, ACM Press,
31 * (2002), 73-82.
32 *
1475579c 33 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
34 * implementation:
35 *
36 * - RCU read-side critical section allows readers to perform hash
37 * table lookups and use the returned objects safely by delaying
38 * memory reclaim of a grace period.
39 * - Add and remove operations are lock-free, and do not need to
40 * allocate memory. They need to be executed within RCU read-side
41 * critical section to ensure the objects they read are valid and to
42 * deal with the cmpxchg ABA problem.
43 * - add and add_unique operations are supported. add_unique checks if
44 * the node key already exists in the hash table. It ensures no key
45 * duplicata exists.
46 * - The resize operation executes concurrently with add/remove/lookup.
47 * - Hash table nodes are contained within a split-ordered list. This
48 * list is ordered by incrementing reversed-bits-hash value.
1ee8f000 49 * - An index of bucket nodes is kept. These bucket nodes are the hash
e753ff5a
MD
50 * table "buckets", and they are also chained together in the
51 * split-ordered list, which allows recursive expansion.
1475579c
MD
52 * - The resize operation for small tables only allows expanding the hash table.
53 * It is triggered automatically by detecting long chains in the add
54 * operation.
55 * - The resize operation for larger tables (and available through an
56 * API) allows both expanding and shrinking the hash table.
4c42f1b8 57 * - Split-counters are used to keep track of the number of
1475579c 58 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
59 * - Resize operation initiated by long chain detection is executed by a
60 * call_rcu thread, which keeps lock-freedom of add and remove.
61 * - Resize operations are protected by a mutex.
62 * - The removal operation is split in two parts: first, a "removed"
63 * flag is set in the next pointer within the node to remove. Then,
64 * a "garbage collection" is performed in the bucket containing the
65 * removed node (from the start of the bucket up to the removed node).
66 * All encountered nodes with "removed" flag set in their next
67 * pointers are removed from the linked-list. If the cmpxchg used for
68 * removal fails (due to concurrent garbage-collection or concurrent
69 * add), we retry from the beginning of the bucket. This ensures that
70 * the node with "removed" flag set is removed from the hash table
71 * (not visible to lookups anymore) before the RCU read-side critical
72 * section held across removal ends. Furthermore, this ensures that
73 * the node with "removed" flag set is removed from the linked-list
74 * before its memory is reclaimed. Only the thread which removal
75 * successfully set the "removed" flag (with a cmpxchg) into a node's
76 * next pointer is considered to have succeeded its removal (and thus
77 * owns the node to reclaim). Because we garbage-collect starting from
1ee8f000 78 * an invariant node (the start-of-bucket bucket node) up to the
e753ff5a
MD
79 * "removed" node (or find a reverse-hash that is higher), we are sure
80 * that a successful traversal of the chain leads to a chain that is
81 * present in the linked-list (the start node is never removed) and
82 * that is does not contain the "removed" node anymore, even if
83 * concurrent delete/add operations are changing the structure of the
84 * list concurrently.
29e669f6
MD
85 * - The add operation performs gargage collection of buckets if it
86 * encounters nodes with removed flag set in the bucket where it wants
87 * to add its new node. This ensures lock-freedom of add operation by
88 * helping the remover unlink nodes from the list rather than to wait
89 * for it do to so.
e753ff5a
MD
90 * - A RCU "order table" indexed by log2(hash index) is copied and
91 * expanded by the resize operation. This order table allows finding
1ee8f000
LJ
92 * the "bucket node" tables.
93 * - There is one bucket node table per hash index order. The size of
94 * each bucket node table is half the number of hashes contained in
93d46c39 95 * this order (except for order 0).
1ee8f000
LJ
96 * - synchronzie_rcu is used to garbage-collect the old bucket node table.
97 * - The per-order bucket node tables contain a compact version of the
e753ff5a
MD
98 * hash table nodes. These tables are invariant after they are
99 * populated into the hash table.
93d46c39 100 *
1ee8f000 101 * Bucket node tables:
93d46c39 102 *
1ee8f000
LJ
103 * hash table hash table the last all bucket node tables
104 * order size bucket node 0 1 2 3 4 5 6(index)
93d46c39
LJ
105 * table size
106 * 0 1 1 1
107 * 1 2 1 1 1
108 * 2 4 2 1 1 2
109 * 3 8 4 1 1 2 4
110 * 4 16 8 1 1 2 4 8
111 * 5 32 16 1 1 2 4 8 16
112 * 6 64 32 1 1 2 4 8 16 32
113 *
1ee8f000 114 * When growing/shrinking, we only focus on the last bucket node table
93d46c39
LJ
115 * which size is (!order ? 1 : (1 << (order -1))).
116 *
117 * Example for growing/shrinking:
1ee8f000
LJ
118 * grow hash table from order 5 to 6: init the index=6 bucket node table
119 * shrink hash table from order 6 to 5: fini the index=6 bucket node table
93d46c39 120 *
1475579c
MD
121 * A bit of ascii art explanation:
122 *
123 * Order index is the off-by-one compare to the actual power of 2 because
124 * we use index 0 to deal with the 0 special-case.
125 *
126 * This shows the nodes for a small table ordered by reversed bits:
127 *
128 * bits reverse
129 * 0 000 000
130 * 4 100 001
131 * 2 010 010
132 * 6 110 011
133 * 1 001 100
134 * 5 101 101
135 * 3 011 110
136 * 7 111 111
137 *
138 * This shows the nodes in order of non-reversed bits, linked by
139 * reversed-bit order.
140 *
141 * order bits reverse
142 * 0 0 000 000
0adc36a8
LJ
143 * 1 | 1 001 100 <-
144 * 2 | | 2 010 010 <- |
f6fdd688 145 * | | | 3 011 110 | <- |
1475579c
MD
146 * 3 -> | | | 4 100 001 | |
147 * -> | | 5 101 101 |
148 * -> | 6 110 011
149 * -> 7 111 111
e753ff5a
MD
150 */
151
2ed95849
MD
152#define _LGPL_SOURCE
153#include <stdlib.h>
e0ba718a
MD
154#include <errno.h>
155#include <assert.h>
156#include <stdio.h>
abc490a1 157#include <stdint.h>
f000907d 158#include <string.h>
e0ba718a 159
15cfbec7 160#include "config.h"
2ed95849 161#include <urcu.h>
abc490a1 162#include <urcu-call-rcu.h>
a42cc659
MD
163#include <urcu/arch.h>
164#include <urcu/uatomic.h>
a42cc659 165#include <urcu/compiler.h>
abc490a1 166#include <urcu/rculfhash.h>
5e28c532 167#include <stdio.h>
464a1ec9 168#include <pthread.h>
44395fb7 169
f9830efd 170#ifdef DEBUG
f0c29ed7 171#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 172#else
e753ff5a 173#define dbg_printf(fmt, args...)
f9830efd
MD
174#endif
175
f8994aee 176/*
4c42f1b8 177 * Split-counters lazily update the global counter each 1024
f8994aee
MD
178 * addition/removal. It automatically keeps track of resize required.
179 * We use the bucket length as indicator for need to expand for small
180 * tables and machines lacking per-cpu data suppport.
181 */
182#define COUNT_COMMIT_ORDER 10
4ddbb355 183#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
6ea6bc67
MD
184#define CHAIN_LEN_TARGET 1
185#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 186
cd95516d 187/*
76a73da8 188 * Define the minimum table size.
cd95516d 189 */
c9edd44a 190#define MIN_TABLE_SIZE 1
cd95516d 191
4105056a
MD
192#if (CAA_BITS_PER_LONG == 32)
193#define MAX_TABLE_ORDER 32
194#else
195#define MAX_TABLE_ORDER 64
196#endif
197
b7d619b0 198/*
1ee8f000 199 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 200 */
6083a889
MD
201#define MIN_PARTITION_PER_THREAD_ORDER 12
202#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 203
4105056a
MD
204#ifndef min
205#define min(a, b) ((a) < (b) ? (a) : (b))
206#endif
207
abc490a1
MD
208#ifndef max
209#define max(a, b) ((a) > (b) ? (a) : (b))
210#endif
2ed95849 211
d95bd160
MD
212/*
213 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 214 * It indicates that no node must attach to the node scheduled for
b198f0fd 215 * removal, and that node garbage collection must be performed.
1ee8f000 216 * The bucket flag does not require to be updated atomically with the
d95bd160
MD
217 * pointer, but it is added as a pointer low bit flag to save space.
218 */
d37166c6 219#define REMOVED_FLAG (1UL << 0)
1ee8f000 220#define BUCKET_FLAG (1UL << 1)
b198f0fd 221#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 222
bb7b2f26 223/* Value of the end pointer. Should not interact with flags. */
f9c80341 224#define END_VALUE NULL
bb7b2f26 225
7f52427b
MD
226/*
227 * ht_items_count: Split-counters counting the number of node addition
228 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
229 * is set at hash table creation.
230 *
231 * These are free-running counters, never reset to zero. They count the
232 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
233 * operations to update the global counter. We choose a power-of-2 value
234 * for the trigger to deal with 32 or 64-bit overflow of the counter.
235 */
df44348d 236struct ht_items_count {
860d07e8 237 unsigned long add, del;
df44348d
MD
238} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
239
7f52427b 240/*
1ee8f000
LJ
241 * rcu_level: Contains the per order-index-level bucket node table. The
242 * size of each bucket node table is half the number of hashes contained
7f52427b 243 * in this order (except for order 0). The minimum allocation size
1ee8f000 244 * parameter allows combining the bucket node arrays of the lowermost
7f52427b
MD
245 * levels to improve cache locality for small index orders.
246 */
1475579c 247struct rcu_level {
0d14ceb2 248 /* Note: manually update allocation length when adding a field */
04db56f8 249 struct cds_lfht_node nodes[0];
1475579c
MD
250};
251
7f52427b
MD
252/*
253 * rcu_table: Contains the size and desired new size if a resize
254 * operation is in progress, as well as the statically-sized array of
255 * rcu_level pointers.
256 */
395270b6 257struct rcu_table {
4105056a 258 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 259 unsigned long resize_target;
11519af6 260 int resize_initiated;
4105056a 261 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
262};
263
7f52427b
MD
264/*
265 * cds_lfht: Top-level data structure representing a lock-free hash
266 * table. Defined in the implementation file to make it be an opaque
267 * cookie to users.
268 */
14044b37 269struct cds_lfht {
4105056a 270 struct rcu_table t;
5488222b
LJ
271 unsigned long min_alloc_order;
272 unsigned long min_alloc_size;
b8af5011 273 int flags;
5f511391
MD
274 /*
275 * We need to put the work threads offline (QSBR) when taking this
276 * mutex, because we use synchronize_rcu within this mutex critical
277 * section, which waits on read-side critical sections, and could
278 * therefore cause grace-period deadlock if we hold off RCU G.P.
279 * completion.
280 */
464a1ec9 281 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 282 unsigned int in_progress_resize, in_progress_destroy;
14044b37 283 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 284 void (*func)(struct rcu_head *head));
1475579c 285 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
286 void (*cds_lfht_rcu_read_lock)(void);
287 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
288 void (*cds_lfht_rcu_thread_offline)(void);
289 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
290 void (*cds_lfht_rcu_register_thread)(void);
291 void (*cds_lfht_rcu_unregister_thread)(void);
292 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 293 long count; /* global approximate item count */
4c42f1b8 294 struct ht_items_count *split_count; /* split item count */
2ed95849
MD
295};
296
7f52427b
MD
297/*
298 * rcu_resize_work: Contains arguments passed to RCU worker thread
299 * responsible for performing lazy resize.
300 */
abc490a1
MD
301struct rcu_resize_work {
302 struct rcu_head head;
14044b37 303 struct cds_lfht *ht;
abc490a1 304};
2ed95849 305
7f52427b
MD
306/*
307 * partition_resize_work: Contains arguments passed to worker threads
308 * executing the hash table resize on partitions of the hash table
309 * assigned to each processor's worker thread.
310 */
b7d619b0 311struct partition_resize_work {
1af6e26e 312 pthread_t thread_id;
b7d619b0
MD
313 struct cds_lfht *ht;
314 unsigned long i, start, len;
315 void (*fct)(struct cds_lfht *ht, unsigned long i,
316 unsigned long start, unsigned long len);
317};
318
76a73da8 319static
83beee94 320void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 321 cds_lfht_match_fct match,
996ff57c 322 const void *key,
83beee94
MD
323 unsigned long size,
324 struct cds_lfht_node *node,
325 struct cds_lfht_iter *unique_ret,
1ee8f000 326 int bucket);
48ed1c18 327
abc490a1
MD
328/*
329 * Algorithm to reverse bits in a word by lookup table, extended to
330 * 64-bit words.
f9830efd 331 * Source:
abc490a1 332 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 333 * Originally from Public Domain.
abc490a1
MD
334 */
335
336static const uint8_t BitReverseTable256[256] =
2ed95849 337{
abc490a1
MD
338#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
339#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
340#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
341 R6(0), R6(2), R6(1), R6(3)
342};
343#undef R2
344#undef R4
345#undef R6
2ed95849 346
abc490a1
MD
347static
348uint8_t bit_reverse_u8(uint8_t v)
349{
350 return BitReverseTable256[v];
351}
ab7d5fc6 352
abc490a1
MD
353static __attribute__((unused))
354uint32_t bit_reverse_u32(uint32_t v)
355{
356 return ((uint32_t) bit_reverse_u8(v) << 24) |
357 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
358 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
359 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
360}
361
abc490a1
MD
362static __attribute__((unused))
363uint64_t bit_reverse_u64(uint64_t v)
2ed95849 364{
abc490a1
MD
365 return ((uint64_t) bit_reverse_u8(v) << 56) |
366 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
367 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
368 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
369 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
370 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
371 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
372 ((uint64_t) bit_reverse_u8(v >> 56));
373}
374
375static
376unsigned long bit_reverse_ulong(unsigned long v)
377{
378#if (CAA_BITS_PER_LONG == 32)
379 return bit_reverse_u32(v);
380#else
381 return bit_reverse_u64(v);
382#endif
383}
384
f9830efd 385/*
24365af7
MD
386 * fls: returns the position of the most significant bit.
387 * Returns 0 if no bit is set, else returns the position of the most
388 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 389 */
24365af7
MD
390#if defined(__i386) || defined(__x86_64)
391static inline
392unsigned int fls_u32(uint32_t x)
f9830efd 393{
24365af7
MD
394 int r;
395
396 asm("bsrl %1,%0\n\t"
397 "jnz 1f\n\t"
398 "movl $-1,%0\n\t"
399 "1:\n\t"
400 : "=r" (r) : "rm" (x));
401 return r + 1;
402}
403#define HAS_FLS_U32
404#endif
405
406#if defined(__x86_64)
407static inline
408unsigned int fls_u64(uint64_t x)
409{
410 long r;
411
412 asm("bsrq %1,%0\n\t"
413 "jnz 1f\n\t"
414 "movq $-1,%0\n\t"
415 "1:\n\t"
416 : "=r" (r) : "rm" (x));
417 return r + 1;
418}
419#define HAS_FLS_U64
420#endif
421
422#ifndef HAS_FLS_U64
423static __attribute__((unused))
424unsigned int fls_u64(uint64_t x)
425{
426 unsigned int r = 64;
427
428 if (!x)
429 return 0;
430
431 if (!(x & 0xFFFFFFFF00000000ULL)) {
432 x <<= 32;
433 r -= 32;
434 }
435 if (!(x & 0xFFFF000000000000ULL)) {
436 x <<= 16;
437 r -= 16;
438 }
439 if (!(x & 0xFF00000000000000ULL)) {
440 x <<= 8;
441 r -= 8;
442 }
443 if (!(x & 0xF000000000000000ULL)) {
444 x <<= 4;
445 r -= 4;
446 }
447 if (!(x & 0xC000000000000000ULL)) {
448 x <<= 2;
449 r -= 2;
450 }
451 if (!(x & 0x8000000000000000ULL)) {
452 x <<= 1;
453 r -= 1;
454 }
455 return r;
456}
457#endif
458
459#ifndef HAS_FLS_U32
460static __attribute__((unused))
461unsigned int fls_u32(uint32_t x)
462{
463 unsigned int r = 32;
f9830efd 464
24365af7
MD
465 if (!x)
466 return 0;
467 if (!(x & 0xFFFF0000U)) {
468 x <<= 16;
469 r -= 16;
470 }
471 if (!(x & 0xFF000000U)) {
472 x <<= 8;
473 r -= 8;
474 }
475 if (!(x & 0xF0000000U)) {
476 x <<= 4;
477 r -= 4;
478 }
479 if (!(x & 0xC0000000U)) {
480 x <<= 2;
481 r -= 2;
482 }
483 if (!(x & 0x80000000U)) {
484 x <<= 1;
485 r -= 1;
486 }
487 return r;
488}
489#endif
490
491unsigned int fls_ulong(unsigned long x)
f9830efd 492{
6887cc5e 493#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
494 return fls_u32(x);
495#else
496 return fls_u64(x);
497#endif
498}
f9830efd 499
920f8ef6
LJ
500/*
501 * Return the minimum order for which x <= (1UL << order).
502 * Return -1 if x is 0.
503 */
24365af7
MD
504int get_count_order_u32(uint32_t x)
505{
920f8ef6
LJ
506 if (!x)
507 return -1;
24365af7 508
920f8ef6 509 return fls_u32(x - 1);
24365af7
MD
510}
511
920f8ef6
LJ
512/*
513 * Return the minimum order for which x <= (1UL << order).
514 * Return -1 if x is 0.
515 */
24365af7
MD
516int get_count_order_ulong(unsigned long x)
517{
920f8ef6
LJ
518 if (!x)
519 return -1;
24365af7 520
920f8ef6 521 return fls_ulong(x - 1);
f9830efd
MD
522}
523
98808fb1 524#ifdef POISON_FREE
5afadd12
LJ
525#define poison_free(ptr) \
526 do { \
527 if (ptr) { \
528 memset(ptr, 0x42, sizeof(*(ptr))); \
529 free(ptr); \
530 } \
98808fb1
MD
531 } while (0)
532#else
533#define poison_free(ptr) free(ptr)
534#endif
535
f9830efd 536static
ab65b890 537void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 538
f8994aee 539static
4105056a 540void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
541 unsigned long count);
542
df44348d 543static long nr_cpus_mask = -1;
4c42f1b8
LJ
544static long split_count_mask = -1;
545
4ddbb355 546#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
547static void ht_init_nr_cpus_mask(void)
548{
549 long maxcpus;
550
551 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
552 if (maxcpus <= 0) {
553 nr_cpus_mask = -2;
554 return;
555 }
556 /*
557 * round up number of CPUs to next power of two, so we
558 * can use & for modulo.
559 */
560 maxcpus = 1UL << get_count_order_ulong(maxcpus);
561 nr_cpus_mask = maxcpus - 1;
562}
4ddbb355
LJ
563#else /* #if defined(HAVE_SYSCONF) */
564static void ht_init_nr_cpus_mask(void)
565{
566 nr_cpus_mask = -2;
567}
568#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
569
570static
5afadd12 571void alloc_split_items_count(struct cds_lfht *ht)
df44348d
MD
572{
573 struct ht_items_count *count;
574
4c42f1b8
LJ
575 if (nr_cpus_mask == -1) {
576 ht_init_nr_cpus_mask();
4ddbb355
LJ
577 if (nr_cpus_mask < 0)
578 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
579 else
580 split_count_mask = nr_cpus_mask;
df44348d 581 }
4c42f1b8 582
4ddbb355 583 assert(split_count_mask >= 0);
5afadd12
LJ
584
585 if (ht->flags & CDS_LFHT_ACCOUNTING) {
586 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
587 assert(ht->split_count);
588 } else {
589 ht->split_count = NULL;
590 }
df44348d
MD
591}
592
593static
5afadd12 594void free_split_items_count(struct cds_lfht *ht)
df44348d 595{
5afadd12 596 poison_free(ht->split_count);
df44348d
MD
597}
598
14360f1c 599#if defined(HAVE_SCHED_GETCPU)
df44348d 600static
14360f1c 601int ht_get_split_count_index(unsigned long hash)
df44348d
MD
602{
603 int cpu;
604
4c42f1b8 605 assert(split_count_mask >= 0);
df44348d 606 cpu = sched_getcpu();
8ed51e04 607 if (caa_unlikely(cpu < 0))
14360f1c 608 return hash & split_count_mask;
df44348d 609 else
4c42f1b8 610 return cpu & split_count_mask;
df44348d 611}
14360f1c
LJ
612#else /* #if defined(HAVE_SCHED_GETCPU) */
613static
614int ht_get_split_count_index(unsigned long hash)
615{
616 return hash & split_count_mask;
617}
618#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
619
620static
14360f1c 621void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 622{
4c42f1b8
LJ
623 unsigned long split_count;
624 int index;
df44348d 625
8ed51e04 626 if (caa_unlikely(!ht->split_count))
3171717f 627 return;
14360f1c 628 index = ht_get_split_count_index(hash);
4c42f1b8 629 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
8ed51e04 630 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 631 long count;
df44348d 632
4c42f1b8 633 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
634 count = uatomic_add_return(&ht->count,
635 1UL << COUNT_COMMIT_ORDER);
636 /* If power of 2 */
637 if (!(count & (count - 1))) {
4105056a 638 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 639 return;
e3ecefd6 640 dbg_printf("add set global %ld\n", count);
4105056a 641 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 642 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
643 }
644 }
645}
646
647static
14360f1c 648void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 649{
4c42f1b8
LJ
650 unsigned long split_count;
651 int index;
df44348d 652
8ed51e04 653 if (caa_unlikely(!ht->split_count))
3171717f 654 return;
14360f1c 655 index = ht_get_split_count_index(hash);
4c42f1b8 656 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
8ed51e04 657 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 658 long count;
df44348d 659
4c42f1b8 660 dbg_printf("del split count %lu\n", split_count);
df44348d 661 count = uatomic_add_return(&ht->count,
3171717f 662 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
663 /* If power of 2 */
664 if (!(count & (count - 1))) {
4105056a 665 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 666 return;
e3ecefd6
MD
667 dbg_printf("del set global %ld\n", count);
668 /*
c941bb9e 669 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
670 * certain threshold.
671 */
4c42f1b8 672 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 673 return;
4105056a 674 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 675 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
676 }
677 }
678}
679
f9830efd 680static
4105056a 681void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 682{
f8994aee
MD
683 unsigned long count;
684
b8af5011
MD
685 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
686 return;
f8994aee
MD
687 count = uatomic_read(&ht->count);
688 /*
689 * Use bucket-local length for small table expand and for
690 * environments lacking per-cpu data support.
691 */
692 if (count >= (1UL << COUNT_COMMIT_ORDER))
693 return;
24365af7 694 if (chain_len > 100)
f0c29ed7 695 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 696 chain_len);
3390d470 697 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
ab65b890 698 cds_lfht_resize_lazy_grow(ht, size,
01370f0b 699 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
700}
701
abc490a1 702static
14044b37 703struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 704{
14044b37 705 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
706}
707
708static
14044b37 709int is_removed(struct cds_lfht_node *node)
abc490a1 710{
d37166c6 711 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
712}
713
714static
14044b37 715struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 716{
14044b37 717 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
718}
719
f5596c94 720static
1ee8f000 721int is_bucket(struct cds_lfht_node *node)
f5596c94 722{
1ee8f000 723 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
724}
725
726static
1ee8f000 727struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 728{
1ee8f000 729 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 730}
bb7b2f26
MD
731
732static
733struct cds_lfht_node *get_end(void)
734{
735 return (struct cds_lfht_node *) END_VALUE;
736}
737
738static
739int is_end(struct cds_lfht_node *node)
740{
741 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
742}
743
abc490a1 744static
ab65b890
LJ
745unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
746 unsigned long v)
abc490a1
MD
747{
748 unsigned long old1, old2;
749
750 old1 = uatomic_read(ptr);
751 do {
752 old2 = old1;
753 if (old2 >= v)
f9830efd 754 return old2;
abc490a1 755 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 756 return old2;
abc490a1
MD
757}
758
f4a9cc0b 759static
04db56f8 760struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
f4a9cc0b
LJ
761 unsigned long hash)
762{
763 unsigned long index, order;
764
765 assert(size > 0);
766 index = hash & (size - 1);
ef6e6171
LJ
767
768 if (index < ht->min_alloc_size) {
769 dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
770 hash, index);
771 return &ht->t.tbl[0]->nodes[index];
772 }
a4ea2223
LJ
773 /*
774 * equivalent to get_count_order_ulong(index + 1), but optimizes
775 * away the non-existing 0 special-case for
776 * get_count_order_ulong.
777 */
778 order = fls_ulong(index);
f4a9cc0b 779 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
ef6e6171
LJ
780 hash, index, order, index & ((1UL << (order - 1)) - 1));
781 return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
782}
783
273399de
MD
784/*
785 * Remove all logically deleted nodes from a bucket up to a certain node key.
786 */
787static
1ee8f000 788void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 789{
14044b37 790 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 791
1ee8f000
LJ
792 assert(!is_bucket(bucket));
793 assert(!is_removed(bucket));
794 assert(!is_bucket(node));
c90201ac 795 assert(!is_removed(node));
273399de 796 for (;;) {
1ee8f000
LJ
797 iter_prev = bucket;
798 /* We can always skip the bucket node initially */
04db56f8 799 iter = rcu_dereference(iter_prev->next);
b4cb483f 800 assert(!is_removed(iter));
04db56f8 801 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 802 /*
1ee8f000 803 * We should never be called with bucket (start of chain)
bd4db153
MD
804 * and logically removed node (end of path compression
805 * marker) being the actual same node. This would be a
806 * bug in the algorithm implementation.
807 */
1ee8f000 808 assert(bucket != node);
273399de 809 for (;;) {
8ed51e04 810 if (caa_unlikely(is_end(iter)))
f9c80341 811 return;
04db56f8 812 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 813 return;
04db56f8 814 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 815 if (caa_likely(is_removed(next)))
273399de 816 break;
b453eae1 817 iter_prev = clear_flag(iter);
273399de
MD
818 iter = next;
819 }
b198f0fd 820 assert(!is_removed(iter));
1ee8f000
LJ
821 if (is_bucket(iter))
822 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
823 else
824 new_next = clear_flag(next);
04db56f8 825 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 826 }
f9c80341 827 return;
273399de
MD
828}
829
9357c415
MD
830static
831int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
832 struct cds_lfht_node *old_node,
3fb86f26 833 struct cds_lfht_node *old_next,
9357c415
MD
834 struct cds_lfht_node *new_node)
835{
04db56f8 836 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
837
838 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 839 return -ENOENT;
9357c415
MD
840
841 assert(!is_removed(old_node));
1ee8f000 842 assert(!is_bucket(old_node));
9357c415 843 assert(!is_removed(new_node));
1ee8f000 844 assert(!is_bucket(new_node));
9357c415 845 assert(new_node != old_node);
3fb86f26 846 for (;;) {
9357c415 847 /* Insert after node to be replaced */
9357c415
MD
848 if (is_removed(old_next)) {
849 /*
850 * Too late, the old node has been removed under us
851 * between lookup and replace. Fail.
852 */
7801dadd 853 return -ENOENT;
9357c415 854 }
1ee8f000 855 assert(!is_bucket(old_next));
9357c415 856 assert(new_node != clear_flag(old_next));
04db56f8 857 new_node->next = clear_flag(old_next);
9357c415
MD
858 /*
859 * Here is the whole trick for lock-free replace: we add
860 * the replacement node _after_ the node we want to
861 * replace by atomically setting its next pointer at the
862 * same time we set its removal flag. Given that
863 * the lookups/get next use an iterator aware of the
864 * next pointer, they will either skip the old node due
865 * to the removal flag and see the new node, or use
866 * the old node, but will not see the new one.
867 */
04db56f8 868 ret_next = uatomic_cmpxchg(&old_node->next,
9357c415 869 old_next, flag_removed(new_node));
3fb86f26 870 if (ret_next == old_next)
7801dadd 871 break; /* We performed the replacement. */
3fb86f26
LJ
872 old_next = ret_next;
873 }
9357c415 874
9357c415
MD
875 /*
876 * Ensure that the old node is not visible to readers anymore:
877 * lookup for the node, and remove it (along with any other
878 * logically removed node) if found.
879 */
04db56f8
LJ
880 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
881 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 882
04db56f8 883 assert(is_removed(rcu_dereference(old_node->next)));
7801dadd 884 return 0;
9357c415
MD
885}
886
83beee94
MD
887/*
888 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
889 * mode. A NULL unique_ret allows creation of duplicate keys.
890 */
abc490a1 891static
83beee94 892void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 893 cds_lfht_match_fct match,
996ff57c 894 const void *key,
83beee94
MD
895 unsigned long size,
896 struct cds_lfht_node *node,
897 struct cds_lfht_iter *unique_ret,
1ee8f000 898 int bucket_flag)
abc490a1 899{
14044b37 900 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 901 *return_node;
04db56f8 902 struct cds_lfht_node *bucket;
abc490a1 903
1ee8f000 904 assert(!is_bucket(node));
c90201ac 905 assert(!is_removed(node));
04db56f8 906 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
abc490a1 907 for (;;) {
adc0de68 908 uint32_t chain_len = 0;
abc490a1 909
11519af6
MD
910 /*
911 * iter_prev points to the non-removed node prior to the
912 * insert location.
11519af6 913 */
04db56f8 914 iter_prev = bucket;
1ee8f000 915 /* We can always skip the bucket node initially */
04db56f8
LJ
916 iter = rcu_dereference(iter_prev->next);
917 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 918 for (;;) {
8ed51e04 919 if (caa_unlikely(is_end(iter)))
273399de 920 goto insert;
04db56f8 921 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 922 goto insert;
238cc06e 923
1ee8f000
LJ
924 /* bucket node is the first node of the identical-hash-value chain */
925 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 926 goto insert;
238cc06e 927
04db56f8 928 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 929 if (caa_unlikely(is_removed(next)))
9dba85be 930 goto gc_node;
238cc06e
LJ
931
932 /* uniquely add */
83beee94 933 if (unique_ret
1ee8f000 934 && !is_bucket(next)
04db56f8 935 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
936 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
937
938 /*
939 * uniquely adding inserts the node as the first
940 * node of the identical-hash-value node chain.
941 *
942 * This semantic ensures no duplicated keys
943 * should ever be observable in the table
944 * (including observe one node by one node
945 * by forward iterations)
946 */
04db56f8 947 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
948 if (!d_iter.node)
949 goto insert;
950
951 *unique_ret = d_iter;
83beee94 952 return;
48ed1c18 953 }
238cc06e 954
11519af6 955 /* Only account for identical reverse hash once */
04db56f8 956 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 957 && !is_bucket(next))
4105056a 958 check_resize(ht, size, ++chain_len);
11519af6 959 iter_prev = clear_flag(iter);
273399de 960 iter = next;
abc490a1 961 }
48ed1c18 962
273399de 963 insert:
7ec59d3b 964 assert(node != clear_flag(iter));
11519af6 965 assert(!is_removed(iter_prev));
c90201ac 966 assert(!is_removed(iter));
f000907d 967 assert(iter_prev != node);
1ee8f000 968 if (!bucket_flag)
04db56f8 969 node->next = clear_flag(iter);
f9c80341 970 else
1ee8f000
LJ
971 node->next = flag_bucket(clear_flag(iter));
972 if (is_bucket(iter))
973 new_node = flag_bucket(node);
f5596c94
MD
974 else
975 new_node = node;
04db56f8 976 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 977 new_node) != iter) {
273399de 978 continue; /* retry */
48ed1c18 979 } else {
83beee94 980 return_node = node;
960c9e4f 981 goto end;
48ed1c18
MD
982 }
983
9dba85be
MD
984 gc_node:
985 assert(!is_removed(iter));
1ee8f000
LJ
986 if (is_bucket(iter))
987 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
988 else
989 new_next = clear_flag(next);
04db56f8 990 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 991 /* retry */
464a1ec9 992 }
9357c415 993end:
83beee94
MD
994 if (unique_ret) {
995 unique_ret->node = return_node;
996 /* unique_ret->next left unset, never used. */
997 }
abc490a1 998}
464a1ec9 999
abc490a1 1000static
860d07e8 1001int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 1002 struct cds_lfht_node *node,
1ee8f000 1003 int bucket_removal)
abc490a1 1004{
04db56f8 1005 struct cds_lfht_node *bucket, *next, *old;
5e28c532 1006
9357c415 1007 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1008 return -ENOENT;
9357c415 1009
7ec59d3b 1010 /* logically delete the node */
1ee8f000 1011 assert(!is_bucket(node));
c90201ac 1012 assert(!is_removed(node));
04db56f8 1013 old = rcu_dereference(node->next);
7ec59d3b 1014 do {
48ed1c18
MD
1015 struct cds_lfht_node *new_next;
1016
7ec59d3b 1017 next = old;
8ed51e04 1018 if (caa_unlikely(is_removed(next)))
743f9143 1019 return -ENOENT;
1ee8f000
LJ
1020 if (bucket_removal)
1021 assert(is_bucket(next));
1475579c 1022 else
1ee8f000 1023 assert(!is_bucket(next));
48ed1c18 1024 new_next = flag_removed(next);
04db56f8 1025 old = uatomic_cmpxchg(&node->next, next, new_next);
7ec59d3b 1026 } while (old != next);
7ec59d3b 1027 /* We performed the (logical) deletion. */
7ec59d3b
MD
1028
1029 /*
1030 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1031 * the node, and remove it (along with any other logically removed node)
1032 * if found.
11519af6 1033 */
04db56f8
LJ
1034 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1035 _cds_lfht_gc_bucket(bucket, node);
743f9143 1036
04db56f8 1037 assert(is_removed(rcu_dereference(node->next)));
743f9143 1038 return 0;
abc490a1 1039}
2ed95849 1040
b7d619b0
MD
1041static
1042void *partition_resize_thread(void *arg)
1043{
1044 struct partition_resize_work *work = arg;
1045
1046 work->ht->cds_lfht_rcu_register_thread();
1047 work->fct(work->ht, work->i, work->start, work->len);
1048 work->ht->cds_lfht_rcu_unregister_thread();
1049 return NULL;
1050}
1051
1052static
1053void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1054 unsigned long len,
1055 void (*fct)(struct cds_lfht *ht, unsigned long i,
1056 unsigned long start, unsigned long len))
1057{
1058 unsigned long partition_len;
1059 struct partition_resize_work *work;
6083a889
MD
1060 int thread, ret;
1061 unsigned long nr_threads;
b7d619b0 1062
6083a889
MD
1063 /*
1064 * Note: nr_cpus_mask + 1 is always power of 2.
1065 * We spawn just the number of threads we need to satisfy the minimum
1066 * partition size, up to the number of CPUs in the system.
1067 */
91452a6a
MD
1068 if (nr_cpus_mask > 0) {
1069 nr_threads = min(nr_cpus_mask + 1,
1070 len >> MIN_PARTITION_PER_THREAD_ORDER);
1071 } else {
1072 nr_threads = 1;
1073 }
6083a889
MD
1074 partition_len = len >> get_count_order_ulong(nr_threads);
1075 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1076 assert(work);
6083a889
MD
1077 for (thread = 0; thread < nr_threads; thread++) {
1078 work[thread].ht = ht;
1079 work[thread].i = i;
1080 work[thread].len = partition_len;
1081 work[thread].start = thread * partition_len;
1082 work[thread].fct = fct;
1af6e26e 1083 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1084 partition_resize_thread, &work[thread]);
b7d619b0
MD
1085 assert(!ret);
1086 }
6083a889 1087 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1088 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1089 assert(!ret);
1090 }
1091 free(work);
b7d619b0
MD
1092}
1093
e8de508e
MD
1094/*
1095 * Holding RCU read lock to protect _cds_lfht_add against memory
1096 * reclaim that could be performed by other call_rcu worker threads (ABA
1097 * problem).
9ee0fc9a 1098 *
b7d619b0 1099 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1100 * many worker threads, based on the number of CPUs available in the system.
1101 * This should therefore take care of not having the expand lagging behind too
1102 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1103 * schedule bucket node population fairly with insertions.
e8de508e 1104 */
4105056a 1105static
b7d619b0
MD
1106void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1107 unsigned long start, unsigned long len)
4105056a
MD
1108{
1109 unsigned long j;
1110
5488222b 1111 assert(i > ht->min_alloc_order);
4105056a 1112 ht->cds_lfht_rcu_read_lock();
b7d619b0 1113 for (j = start; j < start + len; j++) {
04db56f8 1114 struct cds_lfht_node *new_node = &ht->t.tbl[i]->nodes[j];
4105056a 1115
dc1da8f6 1116 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4f6e90b7 1117 i, j, (1UL << (i - 1)) + j);
04db56f8 1118 new_node->reverse_hash =
4f6e90b7 1119 bit_reverse_ulong((1UL << (i - 1)) + j);
04db56f8 1120 _cds_lfht_add(ht, NULL, NULL, 1UL << (i - 1),
83beee94 1121 new_node, NULL, 1);
4105056a
MD
1122 }
1123 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1124}
1125
1126static
1127void init_table_populate(struct cds_lfht *ht, unsigned long i,
1128 unsigned long len)
1129{
1130 assert(nr_cpus_mask != -1);
6083a889 1131 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1132 ht->cds_lfht_rcu_thread_online();
1133 init_table_populate_partition(ht, i, 0, len);
1134 ht->cds_lfht_rcu_thread_offline();
1135 return;
1136 }
1137 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1138}
1139
abc490a1 1140static
4105056a 1141void init_table(struct cds_lfht *ht,
93d46c39 1142 unsigned long first_order, unsigned long last_order)
24365af7 1143{
93d46c39 1144 unsigned long i;
24365af7 1145
93d46c39
LJ
1146 dbg_printf("init table: first_order %lu last_order %lu\n",
1147 first_order, last_order);
5488222b 1148 assert(first_order > ht->min_alloc_order);
93d46c39 1149 for (i = first_order; i <= last_order; i++) {
4105056a 1150 unsigned long len;
24365af7 1151
4f6e90b7 1152 len = 1UL << (i - 1);
f0c29ed7 1153 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1154
1155 /* Stop expand if the resize target changes under us */
4f6e90b7 1156 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1157 break;
1158
04db56f8 1159 ht->t.tbl[i] = calloc(1, len * sizeof(struct cds_lfht_node));
b7d619b0 1160 assert(ht->t.tbl[i]);
4105056a 1161
4105056a 1162 /*
1ee8f000
LJ
1163 * Set all bucket nodes reverse hash values for a level and
1164 * link all bucket nodes into the table.
4105056a 1165 */
dc1da8f6 1166 init_table_populate(ht, i, len);
4105056a 1167
f9c80341
MD
1168 /*
1169 * Update table size.
1170 */
1171 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1172 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1173
4f6e90b7 1174 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1175 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1176 break;
1177 }
1178}
1179
e8de508e
MD
1180/*
1181 * Holding RCU read lock to protect _cds_lfht_remove against memory
1182 * reclaim that could be performed by other call_rcu worker threads (ABA
1183 * problem).
1184 * For a single level, we logically remove and garbage collect each node.
1185 *
1186 * As a design choice, we perform logical removal and garbage collection on a
1187 * node-per-node basis to simplify this algorithm. We also assume keeping good
1188 * cache locality of the operation would overweight possible performance gain
1189 * that could be achieved by batching garbage collection for multiple levels.
1190 * However, this would have to be justified by benchmarks.
1191 *
1192 * Concurrent removal and add operations are helping us perform garbage
1193 * collection of logically removed nodes. We guarantee that all logically
1194 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1195 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e
MD
1196 *
1197 * Logical removal and garbage collection can therefore be done in batch or on a
1198 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1199 *
b7d619b0
MD
1200 * When we reach a certain length, we can split this removal over many worker
1201 * threads, based on the number of CPUs available in the system. This should
1202 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1203 * updater threads actively inserting into the hash table.
e8de508e 1204 */
4105056a 1205static
b7d619b0
MD
1206void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1207 unsigned long start, unsigned long len)
4105056a
MD
1208{
1209 unsigned long j;
1210
5488222b 1211 assert(i > ht->min_alloc_order);
4105056a 1212 ht->cds_lfht_rcu_read_lock();
b7d619b0 1213 for (j = start; j < start + len; j++) {
04db56f8 1214 struct cds_lfht_node *fini_node = &ht->t.tbl[i]->nodes[j];
4105056a
MD
1215
1216 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
4f6e90b7 1217 i, j, (1UL << (i - 1)) + j);
04db56f8 1218 fini_node->reverse_hash =
4f6e90b7
LJ
1219 bit_reverse_ulong((1UL << (i - 1)) + j);
1220 (void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
abc490a1 1221 }
4105056a 1222 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1223}
1224
1225static
1226void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1227{
1228
1229 assert(nr_cpus_mask != -1);
6083a889 1230 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1231 ht->cds_lfht_rcu_thread_online();
1232 remove_table_partition(ht, i, 0, len);
1233 ht->cds_lfht_rcu_thread_offline();
1234 return;
1235 }
1236 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1237}
1238
1475579c 1239static
4105056a 1240void fini_table(struct cds_lfht *ht,
93d46c39 1241 unsigned long first_order, unsigned long last_order)
1475579c 1242{
93d46c39 1243 long i;
0d14ceb2 1244 void *free_by_rcu = NULL;
1475579c 1245
93d46c39
LJ
1246 dbg_printf("fini table: first_order %lu last_order %lu\n",
1247 first_order, last_order);
5488222b 1248 assert(first_order > ht->min_alloc_order);
93d46c39 1249 for (i = last_order; i >= first_order; i--) {
4105056a 1250 unsigned long len;
1475579c 1251
4f6e90b7 1252 len = 1UL << (i - 1);
1475579c 1253 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1254
4d676753
MD
1255 /* Stop shrink if the resize target changes under us */
1256 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1257 break;
1258
1259 cmm_smp_wmb(); /* populate data before RCU size */
1260 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1261
1262 /*
1263 * We need to wait for all add operations to reach Q.S. (and
1264 * thus use the new table for lookups) before we can start
1ee8f000 1265 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1266 * return a logically removed node as insert position.
1267 */
1268 ht->cds_lfht_synchronize_rcu();
0d14ceb2
LJ
1269 if (free_by_rcu)
1270 free(free_by_rcu);
4d676753 1271
21263e21 1272 /*
1ee8f000
LJ
1273 * Set "removed" flag in bucket nodes about to be removed.
1274 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1275 * Concurrent add/remove operation are helping us doing
1276 * the gc.
21263e21 1277 */
4105056a
MD
1278 remove_table(ht, i, len);
1279
0d14ceb2 1280 free_by_rcu = ht->t.tbl[i];
4105056a
MD
1281
1282 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1283 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1284 break;
1285 }
0d14ceb2
LJ
1286
1287 if (free_by_rcu) {
1288 ht->cds_lfht_synchronize_rcu();
1289 free(free_by_rcu);
1290 }
1475579c
MD
1291}
1292
ff0d69de 1293static
1ee8f000 1294void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1295{
04db56f8 1296 struct cds_lfht_node *prev, *node;
ff0d69de
LJ
1297 unsigned long order, len, i, j;
1298
04db56f8 1299 ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct cds_lfht_node));
ff0d69de
LJ
1300 assert(ht->t.tbl[0]);
1301
1ee8f000
LJ
1302 dbg_printf("create bucket: order %lu index %lu hash %lu\n", 0, 0, 0);
1303 ht->t.tbl[0]->nodes[0].next = flag_bucket(get_end());
ff0d69de
LJ
1304 ht->t.tbl[0]->nodes[0].reverse_hash = 0;
1305
1306 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1307 len = 1UL << (order - 1);
5488222b 1308 if (order <= ht->min_alloc_order) {
eb631bf2 1309 ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
5488222b 1310 } else {
04db56f8 1311 ht->t.tbl[order] = calloc(1, len * sizeof(struct cds_lfht_node));
5488222b
LJ
1312 assert(ht->t.tbl[order]);
1313 }
ff0d69de
LJ
1314
1315 i = 0;
1316 prev = ht->t.tbl[i]->nodes;
1317 for (j = 0; j < len; j++) {
1318 if (j & (j - 1)) { /* Between power of 2 */
1319 prev++;
1320 } else if (j) { /* At each power of 2 */
1321 i++;
1322 prev = ht->t.tbl[i]->nodes;
1323 }
1324
1325 node = &ht->t.tbl[order]->nodes[j];
1ee8f000 1326 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
ff0d69de
LJ
1327 order, j, j + len);
1328 node->next = prev->next;
1ee8f000 1329 assert(is_bucket(node->next));
ff0d69de 1330 node->reverse_hash = bit_reverse_ulong(j + len);
1ee8f000 1331 prev->next = flag_bucket(node);
ff0d69de
LJ
1332 }
1333 }
1334}
1335
0422d92c 1336struct cds_lfht *_cds_lfht_new(unsigned long init_size,
5488222b 1337 unsigned long min_alloc_size,
b8af5011 1338 int flags,
14044b37 1339 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1340 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1341 void (*cds_lfht_synchronize_rcu)(void),
1342 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1343 void (*cds_lfht_rcu_read_unlock)(void),
1344 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1345 void (*cds_lfht_rcu_thread_online)(void),
1346 void (*cds_lfht_rcu_register_thread)(void),
1347 void (*cds_lfht_rcu_unregister_thread)(void),
1348 pthread_attr_t *attr)
abc490a1 1349{
14044b37 1350 struct cds_lfht *ht;
24365af7 1351 unsigned long order;
abc490a1 1352
5488222b
LJ
1353 /* min_alloc_size must be power of two */
1354 if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
1355 return NULL;
8129be4e 1356 /* init_size must be power of two */
5488222b 1357 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1358 return NULL;
5488222b
LJ
1359 min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
1360 init_size = max(init_size, min_alloc_size);
14044b37 1361 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1362 assert(ht);
b5d6b20f 1363 ht->flags = flags;
14044b37 1364 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1365 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1366 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1367 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1368 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1369 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1370 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1371 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1372 ht->resize_attr = attr;
5afadd12 1373 alloc_split_items_count(ht);
abc490a1
MD
1374 /* this mutex should not nest in read-side C.S. */
1375 pthread_mutex_init(&ht->resize_mutex, NULL);
5488222b 1376 order = get_count_order_ulong(init_size);
93d46c39 1377 ht->t.resize_target = 1UL << order;
5488222b
LJ
1378 ht->min_alloc_size = min_alloc_size;
1379 ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
1ee8f000 1380 cds_lfht_create_bucket(ht, 1UL << order);
bcbd36fc 1381 ht->t.size = 1UL << order;
abc490a1
MD
1382 return ht;
1383}
1384
6f554439 1385void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1386 cds_lfht_match_fct match, const void *key,
6f554439 1387 struct cds_lfht_iter *iter)
2ed95849 1388{
04db56f8 1389 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1390 unsigned long reverse_hash, size;
2ed95849 1391
abc490a1 1392 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1393
4105056a 1394 size = rcu_dereference(ht->t.size);
04db56f8 1395 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1396 /* We can always skip the bucket node initially */
04db56f8 1397 node = rcu_dereference(bucket->next);
bb7b2f26 1398 node = clear_flag(node);
2ed95849 1399 for (;;) {
8ed51e04 1400 if (caa_unlikely(is_end(node))) {
96ad1112 1401 node = next = NULL;
abc490a1 1402 break;
bb7b2f26 1403 }
04db56f8 1404 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1405 node = next = NULL;
abc490a1 1406 break;
2ed95849 1407 }
04db56f8 1408 next = rcu_dereference(node->next);
7f52427b 1409 assert(node == clear_flag(node));
8ed51e04 1410 if (caa_likely(!is_removed(next))
1ee8f000 1411 && !is_bucket(next)
04db56f8 1412 && node->reverse_hash == reverse_hash
0422d92c 1413 && caa_likely(match(node, key))) {
273399de 1414 break;
2ed95849 1415 }
1b81fe1a 1416 node = clear_flag(next);
2ed95849 1417 }
1ee8f000 1418 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1419 iter->node = node;
1420 iter->next = next;
abc490a1 1421}
e0ba718a 1422
0422d92c 1423void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1424 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1425{
adc0de68 1426 struct cds_lfht_node *node, *next;
a481e5ff 1427 unsigned long reverse_hash;
a481e5ff 1428
adc0de68 1429 node = iter->node;
04db56f8 1430 reverse_hash = node->reverse_hash;
adc0de68 1431 next = iter->next;
a481e5ff
MD
1432 node = clear_flag(next);
1433
1434 for (;;) {
8ed51e04 1435 if (caa_unlikely(is_end(node))) {
96ad1112 1436 node = next = NULL;
a481e5ff 1437 break;
bb7b2f26 1438 }
04db56f8 1439 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1440 node = next = NULL;
a481e5ff
MD
1441 break;
1442 }
04db56f8 1443 next = rcu_dereference(node->next);
8ed51e04 1444 if (caa_likely(!is_removed(next))
1ee8f000 1445 && !is_bucket(next)
04db56f8 1446 && caa_likely(match(node, key))) {
a481e5ff
MD
1447 break;
1448 }
1449 node = clear_flag(next);
1450 }
1ee8f000 1451 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1452 iter->node = node;
1453 iter->next = next;
a481e5ff
MD
1454}
1455
4e9b9fbf
MD
1456void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1457{
1458 struct cds_lfht_node *node, *next;
1459
853395e1 1460 node = clear_flag(iter->next);
4e9b9fbf 1461 for (;;) {
8ed51e04 1462 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1463 node = next = NULL;
1464 break;
1465 }
04db56f8 1466 next = rcu_dereference(node->next);
8ed51e04 1467 if (caa_likely(!is_removed(next))
1ee8f000 1468 && !is_bucket(next)) {
4e9b9fbf
MD
1469 break;
1470 }
1471 node = clear_flag(next);
1472 }
1ee8f000 1473 assert(!node || !is_bucket(rcu_dereference(node->next)));
4e9b9fbf
MD
1474 iter->node = node;
1475 iter->next = next;
1476}
1477
1478void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1479{
04db56f8 1480 struct cds_lfht_node *lookup;
4e9b9fbf
MD
1481
1482 /*
1ee8f000 1483 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1484 * first node of the linked list.
1485 */
1486 lookup = &ht->t.tbl[0]->nodes[0];
853395e1 1487 iter->next = lookup->next;
4e9b9fbf
MD
1488 cds_lfht_next(ht, iter);
1489}
1490
0422d92c
MD
1491void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1492 struct cds_lfht_node *node)
abc490a1 1493{
0422d92c 1494 unsigned long size;
ab7d5fc6 1495
04db56f8 1496 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1497 size = rcu_dereference(ht->t.size);
04db56f8 1498 _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0);
14360f1c 1499 ht_count_add(ht, size, hash);
3eca1b8c
MD
1500}
1501
14044b37 1502struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1503 unsigned long hash,
0422d92c 1504 cds_lfht_match_fct match,
996ff57c 1505 const void *key,
48ed1c18 1506 struct cds_lfht_node *node)
3eca1b8c 1507{
0422d92c 1508 unsigned long size;
83beee94 1509 struct cds_lfht_iter iter;
3eca1b8c 1510
04db56f8 1511 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1512 size = rcu_dereference(ht->t.size);
04db56f8 1513 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1514 if (iter.node == node)
14360f1c 1515 ht_count_add(ht, size, hash);
83beee94 1516 return iter.node;
2ed95849
MD
1517}
1518
9357c415 1519struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1520 unsigned long hash,
0422d92c 1521 cds_lfht_match_fct match,
996ff57c 1522 const void *key,
48ed1c18
MD
1523 struct cds_lfht_node *node)
1524{
0422d92c 1525 unsigned long size;
83beee94 1526 struct cds_lfht_iter iter;
48ed1c18 1527
04db56f8 1528 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
48ed1c18 1529 size = rcu_dereference(ht->t.size);
83beee94 1530 for (;;) {
04db56f8 1531 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1532 if (iter.node == node) {
14360f1c 1533 ht_count_add(ht, size, hash);
83beee94
MD
1534 return NULL;
1535 }
1536
1537 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1538 return iter.node;
1539 }
48ed1c18
MD
1540}
1541
9357c415
MD
1542int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1543 struct cds_lfht_node *new_node)
1544{
1545 unsigned long size;
1546
1547 size = rcu_dereference(ht->t.size);
1548 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1549 new_node);
1550}
1551
1552int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1553{
14360f1c 1554 unsigned long size, hash;
df44348d 1555 int ret;
abc490a1 1556
4105056a 1557 size = rcu_dereference(ht->t.size);
9357c415 1558 ret = _cds_lfht_del(ht, size, iter->node, 0);
14360f1c 1559 if (!ret) {
04db56f8 1560 hash = bit_reverse_ulong(iter->node->reverse_hash);
14360f1c
LJ
1561 ht_count_del(ht, size, hash);
1562 }
df44348d 1563 return ret;
2ed95849 1564}
ab7d5fc6 1565
abc490a1 1566static
1ee8f000 1567int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1568{
14044b37 1569 struct cds_lfht_node *node;
4105056a 1570 unsigned long order, i, size;
674f7a69 1571
abc490a1 1572 /* Check that the table is empty */
04db56f8 1573 node = &ht->t.tbl[0]->nodes[0];
abc490a1 1574 do {
04db56f8 1575 node = clear_flag(node)->next;
1ee8f000 1576 if (!is_bucket(node))
abc490a1 1577 return -EPERM;
273399de 1578 assert(!is_removed(node));
bb7b2f26 1579 } while (!is_end(node));
4105056a
MD
1580 /*
1581 * size accessed without rcu_dereference because hash table is
1582 * being destroyed.
1583 */
1584 size = ht->t.size;
1ee8f000 1585 /* Internal sanity check: all nodes left should be bucket */
4105056a 1586 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1587 unsigned long len;
1588
1589 len = !order ? 1 : 1UL << (order - 1);
1590 for (i = 0; i < len; i++) {
f0c29ed7 1591 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1592 order, i,
4105056a 1593 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1ee8f000 1594 assert(is_bucket(ht->t.tbl[order]->nodes[i].next));
24365af7 1595 }
5488222b
LJ
1596
1597 if (order == ht->min_alloc_order)
1598 poison_free(ht->t.tbl[0]);
1599 else if (order > ht->min_alloc_order)
1600 poison_free(ht->t.tbl[order]);
1601 /* Nothing to delete for order < ht->min_alloc_order */
674f7a69 1602 }
abc490a1 1603 return 0;
674f7a69
MD
1604}
1605
1606/*
1607 * Should only be called when no more concurrent readers nor writers can
1608 * possibly access the table.
1609 */
b7d619b0 1610int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1611{
5e28c532
MD
1612 int ret;
1613
848d4088 1614 /* Wait for in-flight resize operations to complete */
24953e08
MD
1615 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1616 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1617 while (uatomic_read(&ht->in_progress_resize))
1618 poll(NULL, 0, 100); /* wait for 100ms */
1ee8f000 1619 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1620 if (ret)
1621 return ret;
5afadd12 1622 free_split_items_count(ht);
b7d619b0
MD
1623 if (attr)
1624 *attr = ht->resize_attr;
98808fb1 1625 poison_free(ht);
5e28c532 1626 return ret;
674f7a69
MD
1627}
1628
14044b37 1629void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1630 long *approx_before,
273399de 1631 unsigned long *count,
973e5e1b 1632 unsigned long *removed,
d933dd0e 1633 long *approx_after)
273399de 1634{
14044b37 1635 struct cds_lfht_node *node, *next;
1ee8f000 1636 unsigned long nr_bucket = 0;
273399de 1637
7ed7682f 1638 *approx_before = 0;
5afadd12 1639 if (ht->split_count) {
973e5e1b
MD
1640 int i;
1641
4c42f1b8
LJ
1642 for (i = 0; i < split_count_mask + 1; i++) {
1643 *approx_before += uatomic_read(&ht->split_count[i].add);
1644 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1645 }
1646 }
1647
273399de
MD
1648 *count = 0;
1649 *removed = 0;
1650
1ee8f000 1651 /* Count non-bucket nodes in the table */
04db56f8 1652 node = &ht->t.tbl[0]->nodes[0];
273399de 1653 do {
04db56f8 1654 next = rcu_dereference(node->next);
b198f0fd 1655 if (is_removed(next)) {
1ee8f000 1656 if (!is_bucket(next))
973e5e1b
MD
1657 (*removed)++;
1658 else
1ee8f000
LJ
1659 (nr_bucket)++;
1660 } else if (!is_bucket(next))
273399de 1661 (*count)++;
24365af7 1662 else
1ee8f000 1663 (nr_bucket)++;
273399de 1664 node = clear_flag(next);
bb7b2f26 1665 } while (!is_end(node));
1ee8f000 1666 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1667 *approx_after = 0;
5afadd12 1668 if (ht->split_count) {
973e5e1b
MD
1669 int i;
1670
4c42f1b8
LJ
1671 for (i = 0; i < split_count_mask + 1; i++) {
1672 *approx_after += uatomic_read(&ht->split_count[i].add);
1673 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1674 }
1675 }
273399de
MD
1676}
1677
1475579c 1678/* called with resize mutex held */
abc490a1 1679static
4105056a 1680void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1681 unsigned long old_size, unsigned long new_size)
abc490a1 1682{
1475579c 1683 unsigned long old_order, new_order;
1475579c 1684
93d46c39
LJ
1685 old_order = get_count_order_ulong(old_size);
1686 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1687 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1688 old_size, old_order, new_size, new_order);
1475579c 1689 assert(new_size > old_size);
93d46c39 1690 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1691}
1692
1693/* called with resize mutex held */
1694static
4105056a 1695void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1696 unsigned long old_size, unsigned long new_size)
464a1ec9 1697{
1475579c 1698 unsigned long old_order, new_order;
464a1ec9 1699
5488222b 1700 new_size = max(new_size, ht->min_alloc_size);
93d46c39
LJ
1701 old_order = get_count_order_ulong(old_size);
1702 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1703 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1704 old_size, old_order, new_size, new_order);
1475579c 1705 assert(new_size < old_size);
1475579c 1706
1ee8f000 1707 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1708 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1709}
1710
1475579c
MD
1711
1712/* called with resize mutex held */
1713static
1714void _do_cds_lfht_resize(struct cds_lfht *ht)
1715{
1716 unsigned long new_size, old_size;
4105056a
MD
1717
1718 /*
1719 * Resize table, re-do if the target size has changed under us.
1720 */
1721 do {
d2be3620
MD
1722 assert(uatomic_read(&ht->in_progress_resize));
1723 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1724 break;
4105056a
MD
1725 ht->t.resize_initiated = 1;
1726 old_size = ht->t.size;
1727 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1728 if (old_size < new_size)
1729 _do_cds_lfht_grow(ht, old_size, new_size);
1730 else if (old_size > new_size)
1731 _do_cds_lfht_shrink(ht, old_size, new_size);
1732 ht->t.resize_initiated = 0;
1733 /* write resize_initiated before read resize_target */
1734 cmm_smp_mb();
4d676753 1735 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1736}
1737
abc490a1 1738static
ab65b890 1739unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1740{
ab65b890 1741 return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size);
464a1ec9
MD
1742}
1743
1475579c 1744static
4105056a 1745void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1746 unsigned long count)
1475579c 1747{
5488222b 1748 count = max(count, ht->min_alloc_size);
4105056a 1749 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1750}
1751
1752void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1753{
4105056a
MD
1754 resize_target_update_count(ht, new_size);
1755 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1756 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1757 pthread_mutex_lock(&ht->resize_mutex);
1758 _do_cds_lfht_resize(ht);
1759 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1760 ht->cds_lfht_rcu_thread_online();
abc490a1 1761}
464a1ec9 1762
abc490a1
MD
1763static
1764void do_resize_cb(struct rcu_head *head)
1765{
1766 struct rcu_resize_work *work =
1767 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1768 struct cds_lfht *ht = work->ht;
abc490a1 1769
5f511391 1770 ht->cds_lfht_rcu_thread_offline();
abc490a1 1771 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1772 _do_cds_lfht_resize(ht);
abc490a1 1773 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1774 ht->cds_lfht_rcu_thread_online();
98808fb1 1775 poison_free(work);
848d4088
MD
1776 cmm_smp_mb(); /* finish resize before decrement */
1777 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1778}
1779
abc490a1 1780static
f1f119ee 1781void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1782{
abc490a1
MD
1783 struct rcu_resize_work *work;
1784
4105056a
MD
1785 /* Store resize_target before read resize_initiated */
1786 cmm_smp_mb();
ab65b890 1787 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
848d4088 1788 uatomic_inc(&ht->in_progress_resize);
59290e9d 1789 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1790 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1791 uatomic_dec(&ht->in_progress_resize);
59290e9d 1792 return;
ed35e6d8 1793 }
f9830efd
MD
1794 work = malloc(sizeof(*work));
1795 work->ht = ht;
14044b37 1796 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1797 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1798 }
ab7d5fc6 1799}
3171717f 1800
f1f119ee
LJ
1801static
1802void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1803{
1804 unsigned long target_size = size << growth;
1805
1806 if (resize_target_grow(ht, target_size) >= target_size)
1807 return;
1808
1809 __cds_lfht_resize_lazy_launch(ht);
1810}
1811
89bb121d
LJ
1812/*
1813 * We favor grow operations over shrink. A shrink operation never occurs
1814 * if a grow operation is queued for lazy execution. A grow operation
1815 * cancels any pending shrink lazy execution.
1816 */
3171717f 1817static
4105056a 1818void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1819 unsigned long count)
1820{
b8af5011
MD
1821 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1822 return;
89bb121d
LJ
1823 count = max(count, ht->min_alloc_size);
1824 if (count == size)
1825 return; /* Already the right size, no resize needed */
1826 if (count > size) { /* lazy grow */
1827 if (resize_target_grow(ht, count) >= count)
1828 return;
1829 } else { /* lazy shrink */
1830 for (;;) {
1831 unsigned long s;
1832
1833 s = uatomic_cmpxchg(&ht->t.resize_target, size, count);
1834 if (s == size)
1835 break; /* no resize needed */
1836 if (s > size)
1837 return; /* growing is/(was just) in progress */
1838 if (s <= count)
1839 return; /* some other thread do shrink */
1840 size = s;
1841 }
1842 }
f1f119ee 1843 __cds_lfht_resize_lazy_launch(ht);
3171717f 1844}
This page took 0.222732 seconds and 4 git commands to generate.