rename min_alloc_size/min_alloc_order
[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 */
d0d8f9aa
LJ
190#define MIN_TABLE_ORDER 0
191#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER)
cd95516d 192
4105056a
MD
193#if (CAA_BITS_PER_LONG == 32)
194#define MAX_TABLE_ORDER 32
195#else
196#define MAX_TABLE_ORDER 64
197#endif
198
b7d619b0 199/*
1ee8f000 200 * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink.
b7d619b0 201 */
6083a889
MD
202#define MIN_PARTITION_PER_THREAD_ORDER 12
203#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 204
4105056a
MD
205#ifndef min
206#define min(a, b) ((a) < (b) ? (a) : (b))
207#endif
208
abc490a1
MD
209#ifndef max
210#define max(a, b) ((a) > (b) ? (a) : (b))
211#endif
2ed95849 212
d95bd160
MD
213/*
214 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 215 * It indicates that no node must attach to the node scheduled for
b198f0fd 216 * removal, and that node garbage collection must be performed.
1ee8f000 217 * The bucket flag does not require to be updated atomically with the
d95bd160
MD
218 * pointer, but it is added as a pointer low bit flag to save space.
219 */
d37166c6 220#define REMOVED_FLAG (1UL << 0)
1ee8f000 221#define BUCKET_FLAG (1UL << 1)
b198f0fd 222#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 223
bb7b2f26 224/* Value of the end pointer. Should not interact with flags. */
f9c80341 225#define END_VALUE NULL
bb7b2f26 226
7f52427b
MD
227/*
228 * ht_items_count: Split-counters counting the number of node addition
229 * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
230 * is set at hash table creation.
231 *
232 * These are free-running counters, never reset to zero. They count the
233 * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
234 * operations to update the global counter. We choose a power-of-2 value
235 * for the trigger to deal with 32 or 64-bit overflow of the counter.
236 */
df44348d 237struct ht_items_count {
860d07e8 238 unsigned long add, del;
df44348d
MD
239} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
240
7f52427b
MD
241/*
242 * rcu_table: Contains the size and desired new size if a resize
243 * operation is in progress, as well as the statically-sized array of
dcc26ef4 244 * bucket table pointers.
7f52427b 245 */
395270b6 246struct rcu_table {
4105056a 247 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 248 unsigned long resize_target;
11519af6 249 int resize_initiated;
dcc26ef4
LJ
250
251 /*
252 * Contains the per order-index-level bucket node table. The size
253 * of each bucket node table is half the number of hashes contained
254 * in this order (except for order 0). The minimum allocation size
255 * parameter allows combining the bucket node arrays of the lowermost
256 * levels to improve cache locality for small index orders.
257 */
258 struct cds_lfht_node *tbl[MAX_TABLE_ORDER];
395270b6
MD
259};
260
7f52427b
MD
261/*
262 * cds_lfht: Top-level data structure representing a lock-free hash
263 * table. Defined in the implementation file to make it be an opaque
264 * cookie to users.
265 */
14044b37 266struct cds_lfht {
4105056a 267 struct rcu_table t;
0722081a
LJ
268 unsigned long min_alloc_buckets_order;
269 unsigned long min_nr_alloc_buckets;
b8af5011 270 int flags;
5f511391
MD
271 /*
272 * We need to put the work threads offline (QSBR) when taking this
273 * mutex, because we use synchronize_rcu within this mutex critical
274 * section, which waits on read-side critical sections, and could
275 * therefore cause grace-period deadlock if we hold off RCU G.P.
276 * completion.
277 */
464a1ec9 278 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 279 unsigned int in_progress_resize, in_progress_destroy;
14044b37 280 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 281 void (*func)(struct rcu_head *head));
1475579c 282 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
283 void (*cds_lfht_rcu_read_lock)(void);
284 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
285 void (*cds_lfht_rcu_thread_offline)(void);
286 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
287 void (*cds_lfht_rcu_register_thread)(void);
288 void (*cds_lfht_rcu_unregister_thread)(void);
289 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 290 long count; /* global approximate item count */
4c42f1b8 291 struct ht_items_count *split_count; /* split item count */
2ed95849
MD
292};
293
7f52427b
MD
294/*
295 * rcu_resize_work: Contains arguments passed to RCU worker thread
296 * responsible for performing lazy resize.
297 */
abc490a1
MD
298struct rcu_resize_work {
299 struct rcu_head head;
14044b37 300 struct cds_lfht *ht;
abc490a1 301};
2ed95849 302
7f52427b
MD
303/*
304 * partition_resize_work: Contains arguments passed to worker threads
305 * executing the hash table resize on partitions of the hash table
306 * assigned to each processor's worker thread.
307 */
b7d619b0 308struct partition_resize_work {
1af6e26e 309 pthread_t thread_id;
b7d619b0
MD
310 struct cds_lfht *ht;
311 unsigned long i, start, len;
312 void (*fct)(struct cds_lfht *ht, unsigned long i,
313 unsigned long start, unsigned long len);
314};
315
76a73da8 316static
83beee94 317void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 318 cds_lfht_match_fct match,
996ff57c 319 const void *key,
83beee94
MD
320 unsigned long size,
321 struct cds_lfht_node *node,
322 struct cds_lfht_iter *unique_ret,
1ee8f000 323 int bucket);
48ed1c18 324
abc490a1
MD
325/*
326 * Algorithm to reverse bits in a word by lookup table, extended to
327 * 64-bit words.
f9830efd 328 * Source:
abc490a1 329 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 330 * Originally from Public Domain.
abc490a1
MD
331 */
332
333static const uint8_t BitReverseTable256[256] =
2ed95849 334{
abc490a1
MD
335#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
336#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
337#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
338 R6(0), R6(2), R6(1), R6(3)
339};
340#undef R2
341#undef R4
342#undef R6
2ed95849 343
abc490a1
MD
344static
345uint8_t bit_reverse_u8(uint8_t v)
346{
347 return BitReverseTable256[v];
348}
ab7d5fc6 349
abc490a1
MD
350static __attribute__((unused))
351uint32_t bit_reverse_u32(uint32_t v)
352{
353 return ((uint32_t) bit_reverse_u8(v) << 24) |
354 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
355 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
356 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
357}
358
abc490a1
MD
359static __attribute__((unused))
360uint64_t bit_reverse_u64(uint64_t v)
2ed95849 361{
abc490a1
MD
362 return ((uint64_t) bit_reverse_u8(v) << 56) |
363 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
364 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
365 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
366 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
367 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
368 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
369 ((uint64_t) bit_reverse_u8(v >> 56));
370}
371
372static
373unsigned long bit_reverse_ulong(unsigned long v)
374{
375#if (CAA_BITS_PER_LONG == 32)
376 return bit_reverse_u32(v);
377#else
378 return bit_reverse_u64(v);
379#endif
380}
381
f9830efd 382/*
24365af7
MD
383 * fls: returns the position of the most significant bit.
384 * Returns 0 if no bit is set, else returns the position of the most
385 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 386 */
24365af7
MD
387#if defined(__i386) || defined(__x86_64)
388static inline
389unsigned int fls_u32(uint32_t x)
f9830efd 390{
24365af7
MD
391 int r;
392
393 asm("bsrl %1,%0\n\t"
394 "jnz 1f\n\t"
395 "movl $-1,%0\n\t"
396 "1:\n\t"
397 : "=r" (r) : "rm" (x));
398 return r + 1;
399}
400#define HAS_FLS_U32
401#endif
402
403#if defined(__x86_64)
404static inline
405unsigned int fls_u64(uint64_t x)
406{
407 long r;
408
409 asm("bsrq %1,%0\n\t"
410 "jnz 1f\n\t"
411 "movq $-1,%0\n\t"
412 "1:\n\t"
413 : "=r" (r) : "rm" (x));
414 return r + 1;
415}
416#define HAS_FLS_U64
417#endif
418
419#ifndef HAS_FLS_U64
420static __attribute__((unused))
421unsigned int fls_u64(uint64_t x)
422{
423 unsigned int r = 64;
424
425 if (!x)
426 return 0;
427
428 if (!(x & 0xFFFFFFFF00000000ULL)) {
429 x <<= 32;
430 r -= 32;
431 }
432 if (!(x & 0xFFFF000000000000ULL)) {
433 x <<= 16;
434 r -= 16;
435 }
436 if (!(x & 0xFF00000000000000ULL)) {
437 x <<= 8;
438 r -= 8;
439 }
440 if (!(x & 0xF000000000000000ULL)) {
441 x <<= 4;
442 r -= 4;
443 }
444 if (!(x & 0xC000000000000000ULL)) {
445 x <<= 2;
446 r -= 2;
447 }
448 if (!(x & 0x8000000000000000ULL)) {
449 x <<= 1;
450 r -= 1;
451 }
452 return r;
453}
454#endif
455
456#ifndef HAS_FLS_U32
457static __attribute__((unused))
458unsigned int fls_u32(uint32_t x)
459{
460 unsigned int r = 32;
f9830efd 461
24365af7
MD
462 if (!x)
463 return 0;
464 if (!(x & 0xFFFF0000U)) {
465 x <<= 16;
466 r -= 16;
467 }
468 if (!(x & 0xFF000000U)) {
469 x <<= 8;
470 r -= 8;
471 }
472 if (!(x & 0xF0000000U)) {
473 x <<= 4;
474 r -= 4;
475 }
476 if (!(x & 0xC0000000U)) {
477 x <<= 2;
478 r -= 2;
479 }
480 if (!(x & 0x80000000U)) {
481 x <<= 1;
482 r -= 1;
483 }
484 return r;
485}
486#endif
487
488unsigned int fls_ulong(unsigned long x)
f9830efd 489{
6887cc5e 490#if (CAA_BITS_PER_LONG == 32)
24365af7
MD
491 return fls_u32(x);
492#else
493 return fls_u64(x);
494#endif
495}
f9830efd 496
920f8ef6
LJ
497/*
498 * Return the minimum order for which x <= (1UL << order).
499 * Return -1 if x is 0.
500 */
24365af7
MD
501int get_count_order_u32(uint32_t x)
502{
920f8ef6
LJ
503 if (!x)
504 return -1;
24365af7 505
920f8ef6 506 return fls_u32(x - 1);
24365af7
MD
507}
508
920f8ef6
LJ
509/*
510 * Return the minimum order for which x <= (1UL << order).
511 * Return -1 if x is 0.
512 */
24365af7
MD
513int get_count_order_ulong(unsigned long x)
514{
920f8ef6
LJ
515 if (!x)
516 return -1;
24365af7 517
920f8ef6 518 return fls_ulong(x - 1);
f9830efd
MD
519}
520
98808fb1 521#ifdef POISON_FREE
5afadd12
LJ
522#define poison_free(ptr) \
523 do { \
524 if (ptr) { \
525 memset(ptr, 0x42, sizeof(*(ptr))); \
526 free(ptr); \
527 } \
98808fb1
MD
528 } while (0)
529#else
530#define poison_free(ptr) free(ptr)
531#endif
532
f9830efd 533static
ab65b890 534void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 535
f8994aee 536static
4105056a 537void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
538 unsigned long count);
539
df44348d 540static long nr_cpus_mask = -1;
4c42f1b8
LJ
541static long split_count_mask = -1;
542
4ddbb355 543#if defined(HAVE_SYSCONF)
4c42f1b8
LJ
544static void ht_init_nr_cpus_mask(void)
545{
546 long maxcpus;
547
548 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
549 if (maxcpus <= 0) {
550 nr_cpus_mask = -2;
551 return;
552 }
553 /*
554 * round up number of CPUs to next power of two, so we
555 * can use & for modulo.
556 */
557 maxcpus = 1UL << get_count_order_ulong(maxcpus);
558 nr_cpus_mask = maxcpus - 1;
559}
4ddbb355
LJ
560#else /* #if defined(HAVE_SYSCONF) */
561static void ht_init_nr_cpus_mask(void)
562{
563 nr_cpus_mask = -2;
564}
565#endif /* #else #if defined(HAVE_SYSCONF) */
df44348d
MD
566
567static
5afadd12 568void alloc_split_items_count(struct cds_lfht *ht)
df44348d
MD
569{
570 struct ht_items_count *count;
571
4c42f1b8
LJ
572 if (nr_cpus_mask == -1) {
573 ht_init_nr_cpus_mask();
4ddbb355
LJ
574 if (nr_cpus_mask < 0)
575 split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
576 else
577 split_count_mask = nr_cpus_mask;
df44348d 578 }
4c42f1b8 579
4ddbb355 580 assert(split_count_mask >= 0);
5afadd12
LJ
581
582 if (ht->flags & CDS_LFHT_ACCOUNTING) {
583 ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
584 assert(ht->split_count);
585 } else {
586 ht->split_count = NULL;
587 }
df44348d
MD
588}
589
590static
5afadd12 591void free_split_items_count(struct cds_lfht *ht)
df44348d 592{
5afadd12 593 poison_free(ht->split_count);
df44348d
MD
594}
595
14360f1c 596#if defined(HAVE_SCHED_GETCPU)
df44348d 597static
14360f1c 598int ht_get_split_count_index(unsigned long hash)
df44348d
MD
599{
600 int cpu;
601
4c42f1b8 602 assert(split_count_mask >= 0);
df44348d 603 cpu = sched_getcpu();
8ed51e04 604 if (caa_unlikely(cpu < 0))
14360f1c 605 return hash & split_count_mask;
df44348d 606 else
4c42f1b8 607 return cpu & split_count_mask;
df44348d 608}
14360f1c
LJ
609#else /* #if defined(HAVE_SCHED_GETCPU) */
610static
611int ht_get_split_count_index(unsigned long hash)
612{
613 return hash & split_count_mask;
614}
615#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
df44348d
MD
616
617static
14360f1c 618void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 619{
4c42f1b8
LJ
620 unsigned long split_count;
621 int index;
df44348d 622
8ed51e04 623 if (caa_unlikely(!ht->split_count))
3171717f 624 return;
14360f1c 625 index = ht_get_split_count_index(hash);
4c42f1b8 626 split_count = uatomic_add_return(&ht->split_count[index].add, 1);
8ed51e04 627 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 628 long count;
df44348d 629
4c42f1b8 630 dbg_printf("add split count %lu\n", split_count);
df44348d
MD
631 count = uatomic_add_return(&ht->count,
632 1UL << COUNT_COMMIT_ORDER);
633 /* If power of 2 */
634 if (!(count & (count - 1))) {
4105056a 635 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 636 return;
e3ecefd6 637 dbg_printf("add set global %ld\n", count);
4105056a 638 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 639 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
640 }
641 }
642}
643
644static
14360f1c 645void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
df44348d 646{
4c42f1b8
LJ
647 unsigned long split_count;
648 int index;
df44348d 649
8ed51e04 650 if (caa_unlikely(!ht->split_count))
3171717f 651 return;
14360f1c 652 index = ht_get_split_count_index(hash);
4c42f1b8 653 split_count = uatomic_add_return(&ht->split_count[index].del, 1);
8ed51e04 654 if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 655 long count;
df44348d 656
4c42f1b8 657 dbg_printf("del split count %lu\n", split_count);
df44348d 658 count = uatomic_add_return(&ht->count,
3171717f 659 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
660 /* If power of 2 */
661 if (!(count & (count - 1))) {
4105056a 662 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 663 return;
e3ecefd6
MD
664 dbg_printf("del set global %ld\n", count);
665 /*
c941bb9e 666 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
667 * certain threshold.
668 */
4c42f1b8 669 if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
e3ecefd6 670 return;
4105056a 671 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 672 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
673 }
674 }
675}
676
f9830efd 677static
4105056a 678void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 679{
f8994aee
MD
680 unsigned long count;
681
b8af5011
MD
682 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
683 return;
f8994aee
MD
684 count = uatomic_read(&ht->count);
685 /*
686 * Use bucket-local length for small table expand and for
687 * environments lacking per-cpu data support.
688 */
689 if (count >= (1UL << COUNT_COMMIT_ORDER))
690 return;
24365af7 691 if (chain_len > 100)
f0c29ed7 692 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 693 chain_len);
3390d470 694 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
ab65b890 695 cds_lfht_resize_lazy_grow(ht, size,
01370f0b 696 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
697}
698
abc490a1 699static
14044b37 700struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 701{
14044b37 702 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
703}
704
705static
14044b37 706int is_removed(struct cds_lfht_node *node)
abc490a1 707{
d37166c6 708 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
709}
710
711static
14044b37 712struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 713{
14044b37 714 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
715}
716
f5596c94 717static
1ee8f000 718int is_bucket(struct cds_lfht_node *node)
f5596c94 719{
1ee8f000 720 return ((unsigned long) node) & BUCKET_FLAG;
f5596c94
MD
721}
722
723static
1ee8f000 724struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node)
f5596c94 725{
1ee8f000 726 return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
f5596c94 727}
bb7b2f26
MD
728
729static
730struct cds_lfht_node *get_end(void)
731{
732 return (struct cds_lfht_node *) END_VALUE;
733}
734
735static
736int is_end(struct cds_lfht_node *node)
737{
738 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
739}
740
abc490a1 741static
ab65b890
LJ
742unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
743 unsigned long v)
abc490a1
MD
744{
745 unsigned long old1, old2;
746
747 old1 = uatomic_read(ptr);
748 do {
749 old2 = old1;
750 if (old2 >= v)
f9830efd 751 return old2;
abc490a1 752 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
ab65b890 753 return old2;
abc490a1
MD
754}
755
48f1b16d
LJ
756static
757void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order)
758{
759 if (order == 0) {
0722081a 760 ht->t.tbl[0] = calloc(ht->min_nr_alloc_buckets,
48f1b16d
LJ
761 sizeof(struct cds_lfht_node));
762 assert(ht->t.tbl[0]);
0722081a 763 } else if (order > ht->min_alloc_buckets_order) {
48f1b16d
LJ
764 ht->t.tbl[order] = calloc(1UL << (order -1),
765 sizeof(struct cds_lfht_node));
766 assert(ht->t.tbl[order]);
767 }
0722081a 768 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
48f1b16d
LJ
769}
770
771/*
772 * cds_lfht_free_bucket_table() should be called with decreasing order.
773 * When cds_lfht_free_bucket_table(0) is called, it means the whole
774 * lfht is destroyed.
775 */
776static
777void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order)
778{
779 if (order == 0)
780 poison_free(ht->t.tbl[0]);
0722081a 781 else if (order > ht->min_alloc_buckets_order)
48f1b16d 782 poison_free(ht->t.tbl[order]);
0722081a 783 /* Nothing to do for 0 < order && order <= ht->min_alloc_buckets_order */
48f1b16d
LJ
784}
785
9d72a73f
LJ
786static inline
787struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index)
f4a9cc0b 788{
9d72a73f 789 unsigned long order;
ef6e6171 790
9d72a73f 791 if ((__builtin_constant_p(index) && index == 0)
0722081a 792 || index < ht->min_nr_alloc_buckets) {
9d72a73f 793 dbg_printf("bucket index %lu order 0 aridx 0\n", index);
dcc26ef4 794 return &ht->t.tbl[0][index];
ef6e6171 795 }
a4ea2223
LJ
796 /*
797 * equivalent to get_count_order_ulong(index + 1), but optimizes
798 * away the non-existing 0 special-case for
799 * get_count_order_ulong.
800 */
801 order = fls_ulong(index);
9d72a73f
LJ
802 dbg_printf("bucket index %lu order %lu aridx %lu\n",
803 index, order, index & ((1UL << (order - 1)) - 1));
dcc26ef4 804 return &ht->t.tbl[order][index & ((1UL << (order - 1)) - 1)];
f4a9cc0b
LJ
805}
806
9d72a73f
LJ
807static inline
808struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
809 unsigned long hash)
810{
811 assert(size > 0);
812 return bucket_at(ht, hash & (size - 1));
813}
814
273399de
MD
815/*
816 * Remove all logically deleted nodes from a bucket up to a certain node key.
817 */
818static
1ee8f000 819void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node)
273399de 820{
14044b37 821 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 822
1ee8f000
LJ
823 assert(!is_bucket(bucket));
824 assert(!is_removed(bucket));
825 assert(!is_bucket(node));
c90201ac 826 assert(!is_removed(node));
273399de 827 for (;;) {
1ee8f000
LJ
828 iter_prev = bucket;
829 /* We can always skip the bucket node initially */
04db56f8 830 iter = rcu_dereference(iter_prev->next);
b4cb483f 831 assert(!is_removed(iter));
04db56f8 832 assert(iter_prev->reverse_hash <= node->reverse_hash);
bd4db153 833 /*
1ee8f000 834 * We should never be called with bucket (start of chain)
bd4db153
MD
835 * and logically removed node (end of path compression
836 * marker) being the actual same node. This would be a
837 * bug in the algorithm implementation.
838 */
1ee8f000 839 assert(bucket != node);
273399de 840 for (;;) {
8ed51e04 841 if (caa_unlikely(is_end(iter)))
f9c80341 842 return;
04db56f8 843 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
f9c80341 844 return;
04db56f8 845 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 846 if (caa_likely(is_removed(next)))
273399de 847 break;
b453eae1 848 iter_prev = clear_flag(iter);
273399de
MD
849 iter = next;
850 }
b198f0fd 851 assert(!is_removed(iter));
1ee8f000
LJ
852 if (is_bucket(iter))
853 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
854 else
855 new_next = clear_flag(next);
04db56f8 856 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 857 }
f9c80341 858 return;
273399de
MD
859}
860
9357c415
MD
861static
862int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
863 struct cds_lfht_node *old_node,
3fb86f26 864 struct cds_lfht_node *old_next,
9357c415
MD
865 struct cds_lfht_node *new_node)
866{
04db56f8 867 struct cds_lfht_node *bucket, *ret_next;
9357c415
MD
868
869 if (!old_node) /* Return -ENOENT if asked to replace NULL node */
7801dadd 870 return -ENOENT;
9357c415
MD
871
872 assert(!is_removed(old_node));
1ee8f000 873 assert(!is_bucket(old_node));
9357c415 874 assert(!is_removed(new_node));
1ee8f000 875 assert(!is_bucket(new_node));
9357c415 876 assert(new_node != old_node);
3fb86f26 877 for (;;) {
9357c415 878 /* Insert after node to be replaced */
9357c415
MD
879 if (is_removed(old_next)) {
880 /*
881 * Too late, the old node has been removed under us
882 * between lookup and replace. Fail.
883 */
7801dadd 884 return -ENOENT;
9357c415 885 }
1ee8f000 886 assert(!is_bucket(old_next));
9357c415 887 assert(new_node != clear_flag(old_next));
04db56f8 888 new_node->next = clear_flag(old_next);
9357c415
MD
889 /*
890 * Here is the whole trick for lock-free replace: we add
891 * the replacement node _after_ the node we want to
892 * replace by atomically setting its next pointer at the
893 * same time we set its removal flag. Given that
894 * the lookups/get next use an iterator aware of the
895 * next pointer, they will either skip the old node due
896 * to the removal flag and see the new node, or use
897 * the old node, but will not see the new one.
898 */
04db56f8 899 ret_next = uatomic_cmpxchg(&old_node->next,
9357c415 900 old_next, flag_removed(new_node));
3fb86f26 901 if (ret_next == old_next)
7801dadd 902 break; /* We performed the replacement. */
3fb86f26
LJ
903 old_next = ret_next;
904 }
9357c415 905
9357c415
MD
906 /*
907 * Ensure that the old node is not visible to readers anymore:
908 * lookup for the node, and remove it (along with any other
909 * logically removed node) if found.
910 */
04db56f8
LJ
911 bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
912 _cds_lfht_gc_bucket(bucket, new_node);
7801dadd 913
04db56f8 914 assert(is_removed(rcu_dereference(old_node->next)));
7801dadd 915 return 0;
9357c415
MD
916}
917
83beee94
MD
918/*
919 * A non-NULL unique_ret pointer uses the "add unique" (or uniquify) add
920 * mode. A NULL unique_ret allows creation of duplicate keys.
921 */
abc490a1 922static
83beee94 923void _cds_lfht_add(struct cds_lfht *ht,
0422d92c 924 cds_lfht_match_fct match,
996ff57c 925 const void *key,
83beee94
MD
926 unsigned long size,
927 struct cds_lfht_node *node,
928 struct cds_lfht_iter *unique_ret,
1ee8f000 929 int bucket_flag)
abc490a1 930{
14044b37 931 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
960c9e4f 932 *return_node;
04db56f8 933 struct cds_lfht_node *bucket;
abc490a1 934
1ee8f000 935 assert(!is_bucket(node));
c90201ac 936 assert(!is_removed(node));
04db56f8 937 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
abc490a1 938 for (;;) {
adc0de68 939 uint32_t chain_len = 0;
abc490a1 940
11519af6
MD
941 /*
942 * iter_prev points to the non-removed node prior to the
943 * insert location.
11519af6 944 */
04db56f8 945 iter_prev = bucket;
1ee8f000 946 /* We can always skip the bucket node initially */
04db56f8
LJ
947 iter = rcu_dereference(iter_prev->next);
948 assert(iter_prev->reverse_hash <= node->reverse_hash);
abc490a1 949 for (;;) {
8ed51e04 950 if (caa_unlikely(is_end(iter)))
273399de 951 goto insert;
04db56f8 952 if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
273399de 953 goto insert;
238cc06e 954
1ee8f000
LJ
955 /* bucket node is the first node of the identical-hash-value chain */
956 if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash)
194fdbd1 957 goto insert;
238cc06e 958
04db56f8 959 next = rcu_dereference(clear_flag(iter)->next);
8ed51e04 960 if (caa_unlikely(is_removed(next)))
9dba85be 961 goto gc_node;
238cc06e
LJ
962
963 /* uniquely add */
83beee94 964 if (unique_ret
1ee8f000 965 && !is_bucket(next)
04db56f8 966 && clear_flag(iter)->reverse_hash == node->reverse_hash) {
238cc06e
LJ
967 struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
968
969 /*
970 * uniquely adding inserts the node as the first
971 * node of the identical-hash-value node chain.
972 *
973 * This semantic ensures no duplicated keys
974 * should ever be observable in the table
975 * (including observe one node by one node
976 * by forward iterations)
977 */
04db56f8 978 cds_lfht_next_duplicate(ht, match, key, &d_iter);
238cc06e
LJ
979 if (!d_iter.node)
980 goto insert;
981
982 *unique_ret = d_iter;
83beee94 983 return;
48ed1c18 984 }
238cc06e 985
11519af6 986 /* Only account for identical reverse hash once */
04db56f8 987 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
1ee8f000 988 && !is_bucket(next))
4105056a 989 check_resize(ht, size, ++chain_len);
11519af6 990 iter_prev = clear_flag(iter);
273399de 991 iter = next;
abc490a1 992 }
48ed1c18 993
273399de 994 insert:
7ec59d3b 995 assert(node != clear_flag(iter));
11519af6 996 assert(!is_removed(iter_prev));
c90201ac 997 assert(!is_removed(iter));
f000907d 998 assert(iter_prev != node);
1ee8f000 999 if (!bucket_flag)
04db56f8 1000 node->next = clear_flag(iter);
f9c80341 1001 else
1ee8f000
LJ
1002 node->next = flag_bucket(clear_flag(iter));
1003 if (is_bucket(iter))
1004 new_node = flag_bucket(node);
f5596c94
MD
1005 else
1006 new_node = node;
04db56f8 1007 if (uatomic_cmpxchg(&iter_prev->next, iter,
48ed1c18 1008 new_node) != iter) {
273399de 1009 continue; /* retry */
48ed1c18 1010 } else {
83beee94 1011 return_node = node;
960c9e4f 1012 goto end;
48ed1c18
MD
1013 }
1014
9dba85be
MD
1015 gc_node:
1016 assert(!is_removed(iter));
1ee8f000
LJ
1017 if (is_bucket(iter))
1018 new_next = flag_bucket(clear_flag(next));
f5596c94
MD
1019 else
1020 new_next = clear_flag(next);
04db56f8 1021 (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
273399de 1022 /* retry */
464a1ec9 1023 }
9357c415 1024end:
83beee94
MD
1025 if (unique_ret) {
1026 unique_ret->node = return_node;
1027 /* unique_ret->next left unset, never used. */
1028 }
abc490a1 1029}
464a1ec9 1030
abc490a1 1031static
860d07e8 1032int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 1033 struct cds_lfht_node *node,
1ee8f000 1034 int bucket_removal)
abc490a1 1035{
04db56f8 1036 struct cds_lfht_node *bucket, *next, *old;
5e28c532 1037
9357c415 1038 if (!node) /* Return -ENOENT if asked to delete NULL node */
743f9143 1039 return -ENOENT;
9357c415 1040
7ec59d3b 1041 /* logically delete the node */
1ee8f000 1042 assert(!is_bucket(node));
c90201ac 1043 assert(!is_removed(node));
04db56f8 1044 old = rcu_dereference(node->next);
7ec59d3b 1045 do {
48ed1c18
MD
1046 struct cds_lfht_node *new_next;
1047
7ec59d3b 1048 next = old;
8ed51e04 1049 if (caa_unlikely(is_removed(next)))
743f9143 1050 return -ENOENT;
1ee8f000
LJ
1051 if (bucket_removal)
1052 assert(is_bucket(next));
1475579c 1053 else
1ee8f000 1054 assert(!is_bucket(next));
48ed1c18 1055 new_next = flag_removed(next);
04db56f8 1056 old = uatomic_cmpxchg(&node->next, next, new_next);
7ec59d3b 1057 } while (old != next);
7ec59d3b 1058 /* We performed the (logical) deletion. */
7ec59d3b
MD
1059
1060 /*
1061 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1062 * the node, and remove it (along with any other logically removed node)
1063 * if found.
11519af6 1064 */
04db56f8
LJ
1065 bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
1066 _cds_lfht_gc_bucket(bucket, node);
743f9143 1067
04db56f8 1068 assert(is_removed(rcu_dereference(node->next)));
743f9143 1069 return 0;
abc490a1 1070}
2ed95849 1071
b7d619b0
MD
1072static
1073void *partition_resize_thread(void *arg)
1074{
1075 struct partition_resize_work *work = arg;
1076
1077 work->ht->cds_lfht_rcu_register_thread();
1078 work->fct(work->ht, work->i, work->start, work->len);
1079 work->ht->cds_lfht_rcu_unregister_thread();
1080 return NULL;
1081}
1082
1083static
1084void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1085 unsigned long len,
1086 void (*fct)(struct cds_lfht *ht, unsigned long i,
1087 unsigned long start, unsigned long len))
1088{
1089 unsigned long partition_len;
1090 struct partition_resize_work *work;
6083a889
MD
1091 int thread, ret;
1092 unsigned long nr_threads;
b7d619b0 1093
6083a889
MD
1094 /*
1095 * Note: nr_cpus_mask + 1 is always power of 2.
1096 * We spawn just the number of threads we need to satisfy the minimum
1097 * partition size, up to the number of CPUs in the system.
1098 */
91452a6a
MD
1099 if (nr_cpus_mask > 0) {
1100 nr_threads = min(nr_cpus_mask + 1,
1101 len >> MIN_PARTITION_PER_THREAD_ORDER);
1102 } else {
1103 nr_threads = 1;
1104 }
6083a889
MD
1105 partition_len = len >> get_count_order_ulong(nr_threads);
1106 work = calloc(nr_threads, sizeof(*work));
b7d619b0 1107 assert(work);
6083a889
MD
1108 for (thread = 0; thread < nr_threads; thread++) {
1109 work[thread].ht = ht;
1110 work[thread].i = i;
1111 work[thread].len = partition_len;
1112 work[thread].start = thread * partition_len;
1113 work[thread].fct = fct;
1af6e26e 1114 ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
6083a889 1115 partition_resize_thread, &work[thread]);
b7d619b0
MD
1116 assert(!ret);
1117 }
6083a889 1118 for (thread = 0; thread < nr_threads; thread++) {
1af6e26e 1119 ret = pthread_join(work[thread].thread_id, NULL);
b7d619b0
MD
1120 assert(!ret);
1121 }
1122 free(work);
b7d619b0
MD
1123}
1124
e8de508e
MD
1125/*
1126 * Holding RCU read lock to protect _cds_lfht_add against memory
1127 * reclaim that could be performed by other call_rcu worker threads (ABA
1128 * problem).
9ee0fc9a 1129 *
b7d619b0 1130 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1131 * many worker threads, based on the number of CPUs available in the system.
1132 * This should therefore take care of not having the expand lagging behind too
1133 * many concurrent insertion threads by using the scheduler's ability to
1ee8f000 1134 * schedule bucket node population fairly with insertions.
e8de508e 1135 */
4105056a 1136static
b7d619b0
MD
1137void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1138 unsigned long start, unsigned long len)
4105056a 1139{
9d72a73f 1140 unsigned long j, size = 1UL << (i - 1);
4105056a 1141
d0d8f9aa 1142 assert(i > MIN_TABLE_ORDER);
4105056a 1143 ht->cds_lfht_rcu_read_lock();
9d72a73f
LJ
1144 for (j = size + start; j < size + start + len; j++) {
1145 struct cds_lfht_node *new_node = bucket_at(ht, j);
1146
1147 assert(j >= size && j < (size << 1));
1148 dbg_printf("init populate: order %lu index %lu hash %lu\n",
1149 i, j, j);
1150 new_node->reverse_hash = bit_reverse_ulong(j);
1151 _cds_lfht_add(ht, NULL, NULL, size, new_node, NULL, 1);
4105056a
MD
1152 }
1153 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1154}
1155
1156static
1157void init_table_populate(struct cds_lfht *ht, unsigned long i,
1158 unsigned long len)
1159{
1160 assert(nr_cpus_mask != -1);
6083a889 1161 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1162 ht->cds_lfht_rcu_thread_online();
1163 init_table_populate_partition(ht, i, 0, len);
1164 ht->cds_lfht_rcu_thread_offline();
1165 return;
1166 }
1167 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1168}
1169
abc490a1 1170static
4105056a 1171void init_table(struct cds_lfht *ht,
93d46c39 1172 unsigned long first_order, unsigned long last_order)
24365af7 1173{
93d46c39 1174 unsigned long i;
24365af7 1175
93d46c39
LJ
1176 dbg_printf("init table: first_order %lu last_order %lu\n",
1177 first_order, last_order);
d0d8f9aa 1178 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1179 for (i = first_order; i <= last_order; i++) {
4105056a 1180 unsigned long len;
24365af7 1181
4f6e90b7 1182 len = 1UL << (i - 1);
f0c29ed7 1183 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1184
1185 /* Stop expand if the resize target changes under us */
4f6e90b7 1186 if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
4d676753
MD
1187 break;
1188
48f1b16d 1189 cds_lfht_alloc_bucket_table(ht, i);
4105056a 1190
4105056a 1191 /*
1ee8f000
LJ
1192 * Set all bucket nodes reverse hash values for a level and
1193 * link all bucket nodes into the table.
4105056a 1194 */
dc1da8f6 1195 init_table_populate(ht, i, len);
4105056a 1196
f9c80341
MD
1197 /*
1198 * Update table size.
1199 */
1200 cmm_smp_wmb(); /* populate data before RCU size */
4f6e90b7 1201 CMM_STORE_SHARED(ht->t.size, 1UL << i);
f9c80341 1202
4f6e90b7 1203 dbg_printf("init new size: %lu\n", 1UL << i);
4105056a
MD
1204 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1205 break;
1206 }
1207}
1208
e8de508e
MD
1209/*
1210 * Holding RCU read lock to protect _cds_lfht_remove against memory
1211 * reclaim that could be performed by other call_rcu worker threads (ABA
1212 * problem).
1213 * For a single level, we logically remove and garbage collect each node.
1214 *
1215 * As a design choice, we perform logical removal and garbage collection on a
1216 * node-per-node basis to simplify this algorithm. We also assume keeping good
1217 * cache locality of the operation would overweight possible performance gain
1218 * that could be achieved by batching garbage collection for multiple levels.
1219 * However, this would have to be justified by benchmarks.
1220 *
1221 * Concurrent removal and add operations are helping us perform garbage
1222 * collection of logically removed nodes. We guarantee that all logically
1223 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1ee8f000 1224 * invoked to free a hole level of bucket nodes (after a grace period).
e8de508e
MD
1225 *
1226 * Logical removal and garbage collection can therefore be done in batch or on a
1227 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1228 *
b7d619b0
MD
1229 * When we reach a certain length, we can split this removal over many worker
1230 * threads, based on the number of CPUs available in the system. This should
1231 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1232 * updater threads actively inserting into the hash table.
e8de508e 1233 */
4105056a 1234static
b7d619b0
MD
1235void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1236 unsigned long start, unsigned long len)
4105056a 1237{
9d72a73f 1238 unsigned long j, size = 1UL << (i - 1);
4105056a 1239
d0d8f9aa 1240 assert(i > MIN_TABLE_ORDER);
4105056a 1241 ht->cds_lfht_rcu_read_lock();
9d72a73f
LJ
1242 for (j = size + start; j < size + start + len; j++) {
1243 struct cds_lfht_node *fini_node = bucket_at(ht, j);
1244
1245 assert(j >= size && j < (size << 1));
1246 dbg_printf("remove entry: order %lu index %lu hash %lu\n",
1247 i, j, j);
1248 fini_node->reverse_hash = bit_reverse_ulong(j);
1249 (void) _cds_lfht_del(ht, size, fini_node, 1);
abc490a1 1250 }
4105056a 1251 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1252}
1253
1254static
1255void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1256{
1257
1258 assert(nr_cpus_mask != -1);
6083a889 1259 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1260 ht->cds_lfht_rcu_thread_online();
1261 remove_table_partition(ht, i, 0, len);
1262 ht->cds_lfht_rcu_thread_offline();
1263 return;
1264 }
1265 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1266}
1267
1475579c 1268static
4105056a 1269void fini_table(struct cds_lfht *ht,
93d46c39 1270 unsigned long first_order, unsigned long last_order)
1475579c 1271{
93d46c39 1272 long i;
48f1b16d 1273 unsigned long free_by_rcu_order = 0;
1475579c 1274
93d46c39
LJ
1275 dbg_printf("fini table: first_order %lu last_order %lu\n",
1276 first_order, last_order);
d0d8f9aa 1277 assert(first_order > MIN_TABLE_ORDER);
93d46c39 1278 for (i = last_order; i >= first_order; i--) {
4105056a 1279 unsigned long len;
1475579c 1280
4f6e90b7 1281 len = 1UL << (i - 1);
1475579c 1282 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1283
4d676753
MD
1284 /* Stop shrink if the resize target changes under us */
1285 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1286 break;
1287
1288 cmm_smp_wmb(); /* populate data before RCU size */
1289 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1290
1291 /*
1292 * We need to wait for all add operations to reach Q.S. (and
1293 * thus use the new table for lookups) before we can start
1ee8f000 1294 * releasing the old bucket nodes. Otherwise their lookup will
4d676753
MD
1295 * return a logically removed node as insert position.
1296 */
1297 ht->cds_lfht_synchronize_rcu();
48f1b16d
LJ
1298 if (free_by_rcu_order)
1299 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
4d676753 1300
21263e21 1301 /*
1ee8f000
LJ
1302 * Set "removed" flag in bucket nodes about to be removed.
1303 * Unlink all now-logically-removed bucket node pointers.
4105056a
MD
1304 * Concurrent add/remove operation are helping us doing
1305 * the gc.
21263e21 1306 */
4105056a
MD
1307 remove_table(ht, i, len);
1308
48f1b16d 1309 free_by_rcu_order = i;
4105056a
MD
1310
1311 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1312 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1313 break;
1314 }
0d14ceb2 1315
48f1b16d 1316 if (free_by_rcu_order) {
0d14ceb2 1317 ht->cds_lfht_synchronize_rcu();
48f1b16d 1318 cds_lfht_free_bucket_table(ht, free_by_rcu_order);
0d14ceb2 1319 }
1475579c
MD
1320}
1321
ff0d69de 1322static
1ee8f000 1323void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size)
ff0d69de 1324{
04db56f8 1325 struct cds_lfht_node *prev, *node;
9d72a73f 1326 unsigned long order, len, i;
ff0d69de 1327
48f1b16d 1328 cds_lfht_alloc_bucket_table(ht, 0);
ff0d69de 1329
9d72a73f
LJ
1330 dbg_printf("create bucket: order 0 index 0 hash 0\n");
1331 node = bucket_at(ht, 0);
1332 node->next = flag_bucket(get_end());
1333 node->reverse_hash = 0;
ff0d69de
LJ
1334
1335 for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
1336 len = 1UL << (order - 1);
48f1b16d 1337 cds_lfht_alloc_bucket_table(ht, order);
ff0d69de 1338
9d72a73f
LJ
1339 for (i = 0; i < len; i++) {
1340 /*
1341 * Now, we are trying to init the node with the
1342 * hash=(len+i) (which is also a bucket with the
1343 * index=(len+i)) and insert it into the hash table,
1344 * so this node has to be inserted after the bucket
1345 * with the index=(len+i)&(len-1)=i. And because there
1346 * is no other non-bucket node nor bucket node with
1347 * larger index/hash inserted, so the bucket node
1348 * being inserted should be inserted directly linked
1349 * after the bucket node with index=i.
1350 */
1351 prev = bucket_at(ht, i);
1352 node = bucket_at(ht, len + i);
ff0d69de 1353
1ee8f000 1354 dbg_printf("create bucket: order %lu index %lu hash %lu\n",
9d72a73f
LJ
1355 order, len + i, len + i);
1356 node->reverse_hash = bit_reverse_ulong(len + i);
1357
1358 /* insert after prev */
1359 assert(is_bucket(prev->next));
ff0d69de 1360 node->next = prev->next;
1ee8f000 1361 prev->next = flag_bucket(node);
ff0d69de
LJ
1362 }
1363 }
1364}
1365
0422d92c 1366struct cds_lfht *_cds_lfht_new(unsigned long init_size,
0722081a 1367 unsigned long min_nr_alloc_buckets,
b8af5011 1368 int flags,
14044b37 1369 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1370 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1371 void (*cds_lfht_synchronize_rcu)(void),
1372 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1373 void (*cds_lfht_rcu_read_unlock)(void),
1374 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1375 void (*cds_lfht_rcu_thread_online)(void),
1376 void (*cds_lfht_rcu_register_thread)(void),
1377 void (*cds_lfht_rcu_unregister_thread)(void),
1378 pthread_attr_t *attr)
abc490a1 1379{
14044b37 1380 struct cds_lfht *ht;
24365af7 1381 unsigned long order;
abc490a1 1382
0722081a
LJ
1383 /* min_nr_alloc_buckets must be power of two */
1384 if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1)))
5488222b 1385 return NULL;
8129be4e 1386 /* init_size must be power of two */
5488222b 1387 if (!init_size || (init_size & (init_size - 1)))
8129be4e 1388 return NULL;
0722081a 1389 min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE);
d0d8f9aa 1390 init_size = max(init_size, MIN_TABLE_SIZE);
14044b37 1391 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1392 assert(ht);
b5d6b20f 1393 ht->flags = flags;
14044b37 1394 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1395 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1396 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1397 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1398 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1399 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1400 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1401 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1402 ht->resize_attr = attr;
5afadd12 1403 alloc_split_items_count(ht);
abc490a1
MD
1404 /* this mutex should not nest in read-side C.S. */
1405 pthread_mutex_init(&ht->resize_mutex, NULL);
5488222b 1406 order = get_count_order_ulong(init_size);
93d46c39 1407 ht->t.resize_target = 1UL << order;
0722081a
LJ
1408 ht->min_nr_alloc_buckets = min_nr_alloc_buckets;
1409 ht->min_alloc_buckets_order = get_count_order_ulong(min_nr_alloc_buckets);
1ee8f000 1410 cds_lfht_create_bucket(ht, 1UL << order);
bcbd36fc 1411 ht->t.size = 1UL << order;
abc490a1
MD
1412 return ht;
1413}
1414
6f554439 1415void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash,
996ff57c 1416 cds_lfht_match_fct match, const void *key,
6f554439 1417 struct cds_lfht_iter *iter)
2ed95849 1418{
04db56f8 1419 struct cds_lfht_node *node, *next, *bucket;
0422d92c 1420 unsigned long reverse_hash, size;
2ed95849 1421
abc490a1 1422 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1423
4105056a 1424 size = rcu_dereference(ht->t.size);
04db56f8 1425 bucket = lookup_bucket(ht, size, hash);
1ee8f000 1426 /* We can always skip the bucket node initially */
04db56f8 1427 node = rcu_dereference(bucket->next);
bb7b2f26 1428 node = clear_flag(node);
2ed95849 1429 for (;;) {
8ed51e04 1430 if (caa_unlikely(is_end(node))) {
96ad1112 1431 node = next = NULL;
abc490a1 1432 break;
bb7b2f26 1433 }
04db56f8 1434 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1435 node = next = NULL;
abc490a1 1436 break;
2ed95849 1437 }
04db56f8 1438 next = rcu_dereference(node->next);
7f52427b 1439 assert(node == clear_flag(node));
8ed51e04 1440 if (caa_likely(!is_removed(next))
1ee8f000 1441 && !is_bucket(next)
04db56f8 1442 && node->reverse_hash == reverse_hash
0422d92c 1443 && caa_likely(match(node, key))) {
273399de 1444 break;
2ed95849 1445 }
1b81fe1a 1446 node = clear_flag(next);
2ed95849 1447 }
1ee8f000 1448 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1449 iter->node = node;
1450 iter->next = next;
abc490a1 1451}
e0ba718a 1452
0422d92c 1453void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
996ff57c 1454 const void *key, struct cds_lfht_iter *iter)
a481e5ff 1455{
adc0de68 1456 struct cds_lfht_node *node, *next;
a481e5ff 1457 unsigned long reverse_hash;
a481e5ff 1458
adc0de68 1459 node = iter->node;
04db56f8 1460 reverse_hash = node->reverse_hash;
adc0de68 1461 next = iter->next;
a481e5ff
MD
1462 node = clear_flag(next);
1463
1464 for (;;) {
8ed51e04 1465 if (caa_unlikely(is_end(node))) {
96ad1112 1466 node = next = NULL;
a481e5ff 1467 break;
bb7b2f26 1468 }
04db56f8 1469 if (caa_unlikely(node->reverse_hash > reverse_hash)) {
96ad1112 1470 node = next = NULL;
a481e5ff
MD
1471 break;
1472 }
04db56f8 1473 next = rcu_dereference(node->next);
8ed51e04 1474 if (caa_likely(!is_removed(next))
1ee8f000 1475 && !is_bucket(next)
04db56f8 1476 && caa_likely(match(node, key))) {
a481e5ff
MD
1477 break;
1478 }
1479 node = clear_flag(next);
1480 }
1ee8f000 1481 assert(!node || !is_bucket(rcu_dereference(node->next)));
adc0de68
MD
1482 iter->node = node;
1483 iter->next = next;
a481e5ff
MD
1484}
1485
4e9b9fbf
MD
1486void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1487{
1488 struct cds_lfht_node *node, *next;
1489
853395e1 1490 node = clear_flag(iter->next);
4e9b9fbf 1491 for (;;) {
8ed51e04 1492 if (caa_unlikely(is_end(node))) {
4e9b9fbf
MD
1493 node = next = NULL;
1494 break;
1495 }
04db56f8 1496 next = rcu_dereference(node->next);
8ed51e04 1497 if (caa_likely(!is_removed(next))
1ee8f000 1498 && !is_bucket(next)) {
4e9b9fbf
MD
1499 break;
1500 }
1501 node = clear_flag(next);
1502 }
1ee8f000 1503 assert(!node || !is_bucket(rcu_dereference(node->next)));
4e9b9fbf
MD
1504 iter->node = node;
1505 iter->next = next;
1506}
1507
1508void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
1509{
4e9b9fbf 1510 /*
1ee8f000 1511 * Get next after first bucket node. The first bucket node is the
4e9b9fbf
MD
1512 * first node of the linked list.
1513 */
9d72a73f 1514 iter->next = bucket_at(ht, 0)->next;
4e9b9fbf
MD
1515 cds_lfht_next(ht, iter);
1516}
1517
0422d92c
MD
1518void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
1519 struct cds_lfht_node *node)
abc490a1 1520{
0422d92c 1521 unsigned long size;
ab7d5fc6 1522
04db56f8 1523 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1524 size = rcu_dereference(ht->t.size);
04db56f8 1525 _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0);
14360f1c 1526 ht_count_add(ht, size, hash);
3eca1b8c
MD
1527}
1528
14044b37 1529struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
6f554439 1530 unsigned long hash,
0422d92c 1531 cds_lfht_match_fct match,
996ff57c 1532 const void *key,
48ed1c18 1533 struct cds_lfht_node *node)
3eca1b8c 1534{
0422d92c 1535 unsigned long size;
83beee94 1536 struct cds_lfht_iter iter;
3eca1b8c 1537
04db56f8 1538 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
4105056a 1539 size = rcu_dereference(ht->t.size);
04db56f8 1540 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1541 if (iter.node == node)
14360f1c 1542 ht_count_add(ht, size, hash);
83beee94 1543 return iter.node;
2ed95849
MD
1544}
1545
9357c415 1546struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
6f554439 1547 unsigned long hash,
0422d92c 1548 cds_lfht_match_fct match,
996ff57c 1549 const void *key,
48ed1c18
MD
1550 struct cds_lfht_node *node)
1551{
0422d92c 1552 unsigned long size;
83beee94 1553 struct cds_lfht_iter iter;
48ed1c18 1554
04db56f8 1555 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
48ed1c18 1556 size = rcu_dereference(ht->t.size);
83beee94 1557 for (;;) {
04db56f8 1558 _cds_lfht_add(ht, match, key, size, node, &iter, 0);
83beee94 1559 if (iter.node == node) {
14360f1c 1560 ht_count_add(ht, size, hash);
83beee94
MD
1561 return NULL;
1562 }
1563
1564 if (!_cds_lfht_replace(ht, size, iter.node, iter.next, node))
1565 return iter.node;
1566 }
48ed1c18
MD
1567}
1568
9357c415
MD
1569int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
1570 struct cds_lfht_node *new_node)
1571{
1572 unsigned long size;
1573
1574 size = rcu_dereference(ht->t.size);
1575 return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
1576 new_node);
1577}
1578
1579int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
2ed95849 1580{
14360f1c 1581 unsigned long size, hash;
df44348d 1582 int ret;
abc490a1 1583
4105056a 1584 size = rcu_dereference(ht->t.size);
9357c415 1585 ret = _cds_lfht_del(ht, size, iter->node, 0);
14360f1c 1586 if (!ret) {
04db56f8 1587 hash = bit_reverse_ulong(iter->node->reverse_hash);
14360f1c
LJ
1588 ht_count_del(ht, size, hash);
1589 }
df44348d 1590 return ret;
2ed95849 1591}
ab7d5fc6 1592
abc490a1 1593static
1ee8f000 1594int cds_lfht_delete_bucket(struct cds_lfht *ht)
674f7a69 1595{
14044b37 1596 struct cds_lfht_node *node;
4105056a 1597 unsigned long order, i, size;
674f7a69 1598
abc490a1 1599 /* Check that the table is empty */
9d72a73f 1600 node = bucket_at(ht, 0);
abc490a1 1601 do {
04db56f8 1602 node = clear_flag(node)->next;
1ee8f000 1603 if (!is_bucket(node))
abc490a1 1604 return -EPERM;
273399de 1605 assert(!is_removed(node));
bb7b2f26 1606 } while (!is_end(node));
4105056a
MD
1607 /*
1608 * size accessed without rcu_dereference because hash table is
1609 * being destroyed.
1610 */
1611 size = ht->t.size;
1ee8f000 1612 /* Internal sanity check: all nodes left should be bucket */
48f1b16d
LJ
1613 for (i = 0; i < size; i++) {
1614 node = bucket_at(ht, i);
1615 dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n",
1616 i, i, bit_reverse_ulong(node->reverse_hash));
1617 assert(is_bucket(node->next));
1618 }
24365af7 1619
48f1b16d
LJ
1620 for (order = get_count_order_ulong(size); (long)order >= 0; order--)
1621 cds_lfht_free_bucket_table(ht, order);
5488222b 1622
abc490a1 1623 return 0;
674f7a69
MD
1624}
1625
1626/*
1627 * Should only be called when no more concurrent readers nor writers can
1628 * possibly access the table.
1629 */
b7d619b0 1630int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1631{
5e28c532
MD
1632 int ret;
1633
848d4088 1634 /* Wait for in-flight resize operations to complete */
24953e08
MD
1635 _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
1636 cmm_smp_mb(); /* Store destroy before load resize */
848d4088
MD
1637 while (uatomic_read(&ht->in_progress_resize))
1638 poll(NULL, 0, 100); /* wait for 100ms */
1ee8f000 1639 ret = cds_lfht_delete_bucket(ht);
abc490a1
MD
1640 if (ret)
1641 return ret;
5afadd12 1642 free_split_items_count(ht);
b7d619b0
MD
1643 if (attr)
1644 *attr = ht->resize_attr;
98808fb1 1645 poison_free(ht);
5e28c532 1646 return ret;
674f7a69
MD
1647}
1648
14044b37 1649void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1650 long *approx_before,
273399de 1651 unsigned long *count,
973e5e1b 1652 unsigned long *removed,
d933dd0e 1653 long *approx_after)
273399de 1654{
14044b37 1655 struct cds_lfht_node *node, *next;
1ee8f000 1656 unsigned long nr_bucket = 0;
273399de 1657
7ed7682f 1658 *approx_before = 0;
5afadd12 1659 if (ht->split_count) {
973e5e1b
MD
1660 int i;
1661
4c42f1b8
LJ
1662 for (i = 0; i < split_count_mask + 1; i++) {
1663 *approx_before += uatomic_read(&ht->split_count[i].add);
1664 *approx_before -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1665 }
1666 }
1667
273399de
MD
1668 *count = 0;
1669 *removed = 0;
1670
1ee8f000 1671 /* Count non-bucket nodes in the table */
9d72a73f 1672 node = bucket_at(ht, 0);
273399de 1673 do {
04db56f8 1674 next = rcu_dereference(node->next);
b198f0fd 1675 if (is_removed(next)) {
1ee8f000 1676 if (!is_bucket(next))
973e5e1b
MD
1677 (*removed)++;
1678 else
1ee8f000
LJ
1679 (nr_bucket)++;
1680 } else if (!is_bucket(next))
273399de 1681 (*count)++;
24365af7 1682 else
1ee8f000 1683 (nr_bucket)++;
273399de 1684 node = clear_flag(next);
bb7b2f26 1685 } while (!is_end(node));
1ee8f000 1686 dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
7ed7682f 1687 *approx_after = 0;
5afadd12 1688 if (ht->split_count) {
973e5e1b
MD
1689 int i;
1690
4c42f1b8
LJ
1691 for (i = 0; i < split_count_mask + 1; i++) {
1692 *approx_after += uatomic_read(&ht->split_count[i].add);
1693 *approx_after -= uatomic_read(&ht->split_count[i].del);
973e5e1b
MD
1694 }
1695 }
273399de
MD
1696}
1697
1475579c 1698/* called with resize mutex held */
abc490a1 1699static
4105056a 1700void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1701 unsigned long old_size, unsigned long new_size)
abc490a1 1702{
1475579c 1703 unsigned long old_order, new_order;
1475579c 1704
93d46c39
LJ
1705 old_order = get_count_order_ulong(old_size);
1706 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1707 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1708 old_size, old_order, new_size, new_order);
1475579c 1709 assert(new_size > old_size);
93d46c39 1710 init_table(ht, old_order + 1, new_order);
abc490a1
MD
1711}
1712
1713/* called with resize mutex held */
1714static
4105056a 1715void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1716 unsigned long old_size, unsigned long new_size)
464a1ec9 1717{
1475579c 1718 unsigned long old_order, new_order;
464a1ec9 1719
d0d8f9aa 1720 new_size = max(new_size, MIN_TABLE_SIZE);
93d46c39
LJ
1721 old_order = get_count_order_ulong(old_size);
1722 new_order = get_count_order_ulong(new_size);
1a401918
LJ
1723 dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1724 old_size, old_order, new_size, new_order);
1475579c 1725 assert(new_size < old_size);
1475579c 1726
1ee8f000 1727 /* Remove and unlink all bucket nodes to remove. */
93d46c39 1728 fini_table(ht, new_order + 1, old_order);
464a1ec9
MD
1729}
1730
1475579c
MD
1731
1732/* called with resize mutex held */
1733static
1734void _do_cds_lfht_resize(struct cds_lfht *ht)
1735{
1736 unsigned long new_size, old_size;
4105056a
MD
1737
1738 /*
1739 * Resize table, re-do if the target size has changed under us.
1740 */
1741 do {
d2be3620
MD
1742 assert(uatomic_read(&ht->in_progress_resize));
1743 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1744 break;
4105056a
MD
1745 ht->t.resize_initiated = 1;
1746 old_size = ht->t.size;
1747 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1748 if (old_size < new_size)
1749 _do_cds_lfht_grow(ht, old_size, new_size);
1750 else if (old_size > new_size)
1751 _do_cds_lfht_shrink(ht, old_size, new_size);
1752 ht->t.resize_initiated = 0;
1753 /* write resize_initiated before read resize_target */
1754 cmm_smp_mb();
4d676753 1755 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1756}
1757
abc490a1 1758static
ab65b890 1759unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1760{
ab65b890 1761 return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size);
464a1ec9
MD
1762}
1763
1475579c 1764static
4105056a 1765void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1766 unsigned long count)
1475579c 1767{
d0d8f9aa 1768 count = max(count, MIN_TABLE_SIZE);
4105056a 1769 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1770}
1771
1772void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1773{
4105056a
MD
1774 resize_target_update_count(ht, new_size);
1775 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1776 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1777 pthread_mutex_lock(&ht->resize_mutex);
1778 _do_cds_lfht_resize(ht);
1779 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1780 ht->cds_lfht_rcu_thread_online();
abc490a1 1781}
464a1ec9 1782
abc490a1
MD
1783static
1784void do_resize_cb(struct rcu_head *head)
1785{
1786 struct rcu_resize_work *work =
1787 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1788 struct cds_lfht *ht = work->ht;
abc490a1 1789
5f511391 1790 ht->cds_lfht_rcu_thread_offline();
abc490a1 1791 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1792 _do_cds_lfht_resize(ht);
abc490a1 1793 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1794 ht->cds_lfht_rcu_thread_online();
98808fb1 1795 poison_free(work);
848d4088
MD
1796 cmm_smp_mb(); /* finish resize before decrement */
1797 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1798}
1799
abc490a1 1800static
f1f119ee 1801void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
ab7d5fc6 1802{
abc490a1
MD
1803 struct rcu_resize_work *work;
1804
4105056a
MD
1805 /* Store resize_target before read resize_initiated */
1806 cmm_smp_mb();
ab65b890 1807 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
848d4088 1808 uatomic_inc(&ht->in_progress_resize);
59290e9d 1809 cmm_smp_mb(); /* increment resize count before load destroy */
ed35e6d8
MD
1810 if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
1811 uatomic_dec(&ht->in_progress_resize);
59290e9d 1812 return;
ed35e6d8 1813 }
f9830efd
MD
1814 work = malloc(sizeof(*work));
1815 work->ht = ht;
14044b37 1816 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1817 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1818 }
ab7d5fc6 1819}
3171717f 1820
f1f119ee
LJ
1821static
1822void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
1823{
1824 unsigned long target_size = size << growth;
1825
1826 if (resize_target_grow(ht, target_size) >= target_size)
1827 return;
1828
1829 __cds_lfht_resize_lazy_launch(ht);
1830}
1831
89bb121d
LJ
1832/*
1833 * We favor grow operations over shrink. A shrink operation never occurs
1834 * if a grow operation is queued for lazy execution. A grow operation
1835 * cancels any pending shrink lazy execution.
1836 */
3171717f 1837static
4105056a 1838void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1839 unsigned long count)
1840{
b8af5011
MD
1841 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1842 return;
d0d8f9aa 1843 count = max(count, MIN_TABLE_SIZE);
89bb121d
LJ
1844 if (count == size)
1845 return; /* Already the right size, no resize needed */
1846 if (count > size) { /* lazy grow */
1847 if (resize_target_grow(ht, count) >= count)
1848 return;
1849 } else { /* lazy shrink */
1850 for (;;) {
1851 unsigned long s;
1852
1853 s = uatomic_cmpxchg(&ht->t.resize_target, size, count);
1854 if (s == size)
1855 break; /* no resize needed */
1856 if (s > size)
1857 return; /* growing is/(was just) in progress */
1858 if (s <= count)
1859 return; /* some other thread do shrink */
1860 size = s;
1861 }
1862 }
f1f119ee 1863 __cds_lfht_resize_lazy_launch(ht);
3171717f 1864}
This page took 0.222029 seconds and 4 git commands to generate.