rculfhash tests: make node count RCU aware
[urcu.git] / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
1475579c 4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
abc490a1
MD
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
5e28c532
MD
21 */
22
e753ff5a
MD
23/*
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
30 * (2002), 73-82.
31 *
1475579c 32 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
33 * implementation:
34 *
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
44 * duplicata exists.
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
1475579c
MD
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
53 * operation.
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
83 * list concurrently.
29e669f6
MD
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
88 * for it do to so.
e753ff5a
MD
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
94 * this order.
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
1475579c
MD
99 *
100 * A bit of ascii art explanation:
101 *
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
104 *
105 * This shows the nodes for a small table ordered by reversed bits:
106 *
107 * bits reverse
108 * 0 000 000
109 * 4 100 001
110 * 2 010 010
111 * 6 110 011
112 * 1 001 100
113 * 5 101 101
114 * 3 011 110
115 * 7 111 111
116 *
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
119 *
120 * order bits reverse
121 * 0 0 000 000
122 * |
f6fdd688
MD
123 * 1 | 1 001 100 <- <-
124 * | | | |
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
127 * | | | | | | |
1475579c
MD
128 * 3 -> | | | 4 100 001 | |
129 * -> | | 5 101 101 |
130 * -> | 6 110 011
131 * -> 7 111 111
e753ff5a
MD
132 */
133
2ed95849
MD
134#define _LGPL_SOURCE
135#include <stdlib.h>
e0ba718a
MD
136#include <errno.h>
137#include <assert.h>
138#include <stdio.h>
abc490a1 139#include <stdint.h>
f000907d 140#include <string.h>
e0ba718a 141
df44348d 142#include "config.h"
2ed95849 143#include <urcu.h>
abc490a1 144#include <urcu-call-rcu.h>
a42cc659
MD
145#include <urcu/arch.h>
146#include <urcu/uatomic.h>
674f7a69 147#include <urcu/jhash.h>
a42cc659 148#include <urcu/compiler.h>
abc490a1 149#include <urcu/rculfhash.h>
5e28c532 150#include <stdio.h>
464a1ec9 151#include <pthread.h>
44395fb7 152
f9830efd 153#ifdef DEBUG
f0c29ed7 154#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 155#else
e753ff5a 156#define dbg_printf(fmt, args...)
f9830efd
MD
157#endif
158
f8994aee
MD
159/*
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
164 */
165#define COUNT_COMMIT_ORDER 10
6ea6bc67
MD
166#define CHAIN_LEN_TARGET 1
167#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 168
cd95516d 169/*
76a73da8 170 * Define the minimum table size.
cd95516d 171 */
c9edd44a 172#define MIN_TABLE_SIZE 1
cd95516d 173
4105056a
MD
174#if (CAA_BITS_PER_LONG == 32)
175#define MAX_TABLE_ORDER 32
176#else
177#define MAX_TABLE_ORDER 64
178#endif
179
b7d619b0
MD
180/*
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 */
6083a889
MD
183#define MIN_PARTITION_PER_THREAD_ORDER 12
184#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 185
4105056a
MD
186#ifndef min
187#define min(a, b) ((a) < (b) ? (a) : (b))
188#endif
189
abc490a1
MD
190#ifndef max
191#define max(a, b) ((a) > (b) ? (a) : (b))
192#endif
2ed95849 193
d95bd160
MD
194/*
195 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 196 * It indicates that no node must attach to the node scheduled for
b198f0fd 197 * removal, and that node garbage collection must be performed.
d95bd160
MD
198 * The dummy flag does not require to be updated atomically with the
199 * pointer, but it is added as a pointer low bit flag to save space.
200 */
d37166c6 201#define REMOVED_FLAG (1UL << 0)
b198f0fd
MD
202#define DUMMY_FLAG (1UL << 1)
203#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 204
bb7b2f26 205/* Value of the end pointer. Should not interact with flags. */
f9c80341 206#define END_VALUE NULL
bb7b2f26 207
df44348d 208struct ht_items_count {
860d07e8 209 unsigned long add, del;
df44348d
MD
210} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
211
1475579c
MD
212struct rcu_level {
213 struct rcu_head head;
214 struct _cds_lfht_node nodes[0];
215};
216
395270b6 217struct rcu_table {
4105056a 218 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 219 unsigned long resize_target;
11519af6 220 int resize_initiated;
4105056a 221 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
222};
223
14044b37 224struct cds_lfht {
4105056a 225 struct rcu_table t;
14044b37
MD
226 cds_lfht_hash_fct hash_fct;
227 cds_lfht_compare_fct compare_fct;
732ad076 228 unsigned long hash_seed;
b8af5011 229 int flags;
5f511391
MD
230 /*
231 * We need to put the work threads offline (QSBR) when taking this
232 * mutex, because we use synchronize_rcu within this mutex critical
233 * section, which waits on read-side critical sections, and could
234 * therefore cause grace-period deadlock if we hold off RCU G.P.
235 * completion.
236 */
464a1ec9 237 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 238 unsigned int in_progress_resize, in_progress_destroy;
14044b37 239 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 240 void (*func)(struct rcu_head *head));
1475579c 241 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
242 void (*cds_lfht_rcu_read_lock)(void);
243 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
244 void (*cds_lfht_rcu_thread_offline)(void);
245 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
246 void (*cds_lfht_rcu_register_thread)(void);
247 void (*cds_lfht_rcu_unregister_thread)(void);
248 pthread_attr_t *resize_attr; /* Resize threads attributes */
7de5ccfd 249 long count; /* global approximate item count */
df44348d 250 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
251};
252
abc490a1
MD
253struct rcu_resize_work {
254 struct rcu_head head;
14044b37 255 struct cds_lfht *ht;
abc490a1 256};
2ed95849 257
b7d619b0
MD
258struct partition_resize_work {
259 struct rcu_head head;
260 struct cds_lfht *ht;
261 unsigned long i, start, len;
262 void (*fct)(struct cds_lfht *ht, unsigned long i,
263 unsigned long start, unsigned long len);
264};
265
48ed1c18
MD
266enum add_mode {
267 ADD_DEFAULT = 0,
268 ADD_UNIQUE = 1,
269 ADD_REPLACE = 2,
270};
271
76a73da8
MD
272static
273struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
274 unsigned long size,
275 struct cds_lfht_node *node,
48ed1c18
MD
276 enum add_mode mode, int dummy);
277
abc490a1
MD
278/*
279 * Algorithm to reverse bits in a word by lookup table, extended to
280 * 64-bit words.
f9830efd 281 * Source:
abc490a1 282 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 283 * Originally from Public Domain.
abc490a1
MD
284 */
285
286static const uint8_t BitReverseTable256[256] =
2ed95849 287{
abc490a1
MD
288#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
289#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
290#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
291 R6(0), R6(2), R6(1), R6(3)
292};
293#undef R2
294#undef R4
295#undef R6
2ed95849 296
abc490a1
MD
297static
298uint8_t bit_reverse_u8(uint8_t v)
299{
300 return BitReverseTable256[v];
301}
ab7d5fc6 302
abc490a1
MD
303static __attribute__((unused))
304uint32_t bit_reverse_u32(uint32_t v)
305{
306 return ((uint32_t) bit_reverse_u8(v) << 24) |
307 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
308 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
309 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
310}
311
abc490a1
MD
312static __attribute__((unused))
313uint64_t bit_reverse_u64(uint64_t v)
2ed95849 314{
abc490a1
MD
315 return ((uint64_t) bit_reverse_u8(v) << 56) |
316 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
317 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
318 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
319 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
320 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
321 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
322 ((uint64_t) bit_reverse_u8(v >> 56));
323}
324
325static
326unsigned long bit_reverse_ulong(unsigned long v)
327{
328#if (CAA_BITS_PER_LONG == 32)
329 return bit_reverse_u32(v);
330#else
331 return bit_reverse_u64(v);
332#endif
333}
334
f9830efd 335/*
24365af7
MD
336 * fls: returns the position of the most significant bit.
337 * Returns 0 if no bit is set, else returns the position of the most
338 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 339 */
24365af7
MD
340#if defined(__i386) || defined(__x86_64)
341static inline
342unsigned int fls_u32(uint32_t x)
f9830efd 343{
24365af7
MD
344 int r;
345
346 asm("bsrl %1,%0\n\t"
347 "jnz 1f\n\t"
348 "movl $-1,%0\n\t"
349 "1:\n\t"
350 : "=r" (r) : "rm" (x));
351 return r + 1;
352}
353#define HAS_FLS_U32
354#endif
355
356#if defined(__x86_64)
357static inline
358unsigned int fls_u64(uint64_t x)
359{
360 long r;
361
362 asm("bsrq %1,%0\n\t"
363 "jnz 1f\n\t"
364 "movq $-1,%0\n\t"
365 "1:\n\t"
366 : "=r" (r) : "rm" (x));
367 return r + 1;
368}
369#define HAS_FLS_U64
370#endif
371
372#ifndef HAS_FLS_U64
373static __attribute__((unused))
374unsigned int fls_u64(uint64_t x)
375{
376 unsigned int r = 64;
377
378 if (!x)
379 return 0;
380
381 if (!(x & 0xFFFFFFFF00000000ULL)) {
382 x <<= 32;
383 r -= 32;
384 }
385 if (!(x & 0xFFFF000000000000ULL)) {
386 x <<= 16;
387 r -= 16;
388 }
389 if (!(x & 0xFF00000000000000ULL)) {
390 x <<= 8;
391 r -= 8;
392 }
393 if (!(x & 0xF000000000000000ULL)) {
394 x <<= 4;
395 r -= 4;
396 }
397 if (!(x & 0xC000000000000000ULL)) {
398 x <<= 2;
399 r -= 2;
400 }
401 if (!(x & 0x8000000000000000ULL)) {
402 x <<= 1;
403 r -= 1;
404 }
405 return r;
406}
407#endif
408
409#ifndef HAS_FLS_U32
410static __attribute__((unused))
411unsigned int fls_u32(uint32_t x)
412{
413 unsigned int r = 32;
f9830efd 414
24365af7
MD
415 if (!x)
416 return 0;
417 if (!(x & 0xFFFF0000U)) {
418 x <<= 16;
419 r -= 16;
420 }
421 if (!(x & 0xFF000000U)) {
422 x <<= 8;
423 r -= 8;
424 }
425 if (!(x & 0xF0000000U)) {
426 x <<= 4;
427 r -= 4;
428 }
429 if (!(x & 0xC0000000U)) {
430 x <<= 2;
431 r -= 2;
432 }
433 if (!(x & 0x80000000U)) {
434 x <<= 1;
435 r -= 1;
436 }
437 return r;
438}
439#endif
440
441unsigned int fls_ulong(unsigned long x)
f9830efd 442{
24365af7
MD
443#if (CAA_BITS_PER_lONG == 32)
444 return fls_u32(x);
445#else
446 return fls_u64(x);
447#endif
448}
f9830efd 449
24365af7
MD
450int get_count_order_u32(uint32_t x)
451{
452 int order;
453
454 order = fls_u32(x) - 1;
455 if (x & (x - 1))
456 order++;
457 return order;
458}
459
460int get_count_order_ulong(unsigned long x)
461{
462 int order;
463
464 order = fls_ulong(x) - 1;
465 if (x & (x - 1))
466 order++;
467 return order;
f9830efd
MD
468}
469
98808fb1
MD
470#ifdef POISON_FREE
471#define poison_free(ptr) \
472 do { \
473 memset(ptr, 0x42, sizeof(*(ptr))); \
474 free(ptr); \
475 } while (0)
476#else
477#define poison_free(ptr) free(ptr)
478#endif
479
f9830efd 480static
4105056a 481void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 482
df44348d
MD
483/*
484 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
485 * available, then we support hash table item accounting.
486 * In the unfortunate event the number of CPUs reported would be
487 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
488 */
df44348d
MD
489#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
490
f8994aee 491static
4105056a 492void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
493 unsigned long count);
494
df44348d
MD
495static long nr_cpus_mask = -1;
496
497static
498struct ht_items_count *alloc_per_cpu_items_count(void)
499{
500 struct ht_items_count *count;
501
502 switch (nr_cpus_mask) {
503 case -2:
504 return NULL;
505 case -1:
506 {
507 long maxcpus;
508
509 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
510 if (maxcpus <= 0) {
511 nr_cpus_mask = -2;
512 return NULL;
513 }
514 /*
515 * round up number of CPUs to next power of two, so we
516 * can use & for modulo.
517 */
518 maxcpus = 1UL << get_count_order_ulong(maxcpus);
519 nr_cpus_mask = maxcpus - 1;
520 }
521 /* Fall-through */
522 default:
523 return calloc(nr_cpus_mask + 1, sizeof(*count));
524 }
525}
526
527static
528void free_per_cpu_items_count(struct ht_items_count *count)
529{
98808fb1 530 poison_free(count);
df44348d
MD
531}
532
533static
534int ht_get_cpu(void)
535{
536 int cpu;
537
538 assert(nr_cpus_mask >= 0);
539 cpu = sched_getcpu();
540 if (unlikely(cpu < 0))
541 return cpu;
542 else
543 return cpu & nr_cpus_mask;
544}
545
546static
4105056a 547void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 548{
3171717f 549 unsigned long percpu_count;
df44348d
MD
550 int cpu;
551
552 if (unlikely(!ht->percpu_count))
3171717f 553 return;
df44348d
MD
554 cpu = ht_get_cpu();
555 if (unlikely(cpu < 0))
3171717f
MD
556 return;
557 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d 558 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 559 long count;
df44348d
MD
560
561 dbg_printf("add percpu %lu\n", percpu_count);
562 count = uatomic_add_return(&ht->count,
563 1UL << COUNT_COMMIT_ORDER);
564 /* If power of 2 */
565 if (!(count & (count - 1))) {
4105056a 566 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee 567 return;
e3ecefd6 568 dbg_printf("add set global %ld\n", count);
4105056a 569 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 570 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
571 }
572 }
573}
574
575static
860d07e8 576void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
577{
578 unsigned long percpu_count;
3171717f 579 int cpu;
df44348d 580
3171717f
MD
581 if (unlikely(!ht->percpu_count))
582 return;
583 cpu = ht_get_cpu();
584 if (unlikely(cpu < 0))
585 return;
80d90c06 586 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d 587 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
e3ecefd6 588 long count;
df44348d 589
860d07e8 590 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 591 count = uatomic_add_return(&ht->count,
3171717f 592 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
593 /* If power of 2 */
594 if (!(count & (count - 1))) {
4105056a 595 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 596 return;
e3ecefd6
MD
597 dbg_printf("del set global %ld\n", count);
598 /*
c941bb9e 599 * Don't shrink table if the number of nodes is below a
e3ecefd6
MD
600 * certain threshold.
601 */
c941bb9e 602 if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
e3ecefd6 603 return;
4105056a 604 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 605 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
606 }
607 }
608}
609
610#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
611
612static const long nr_cpus_mask = -1;
613
614static
615struct ht_items_count *alloc_per_cpu_items_count(void)
616{
617 return NULL;
618}
619
620static
621void free_per_cpu_items_count(struct ht_items_count *count)
622{
623}
624
625static
4105056a 626void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
627{
628}
629
630static
860d07e8 631void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
632{
633}
634
635#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
636
637
f9830efd 638static
4105056a 639void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 640{
f8994aee
MD
641 unsigned long count;
642
b8af5011
MD
643 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
644 return;
f8994aee
MD
645 count = uatomic_read(&ht->count);
646 /*
647 * Use bucket-local length for small table expand and for
648 * environments lacking per-cpu data support.
649 */
650 if (count >= (1UL << COUNT_COMMIT_ORDER))
651 return;
24365af7 652 if (chain_len > 100)
f0c29ed7 653 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 654 chain_len);
3390d470 655 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 656 cds_lfht_resize_lazy(ht, size,
01370f0b 657 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
658}
659
abc490a1 660static
14044b37 661struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 662{
14044b37 663 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
664}
665
666static
14044b37 667int is_removed(struct cds_lfht_node *node)
abc490a1 668{
d37166c6 669 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
670}
671
672static
14044b37 673struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 674{
14044b37 675 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
676}
677
f5596c94 678static
14044b37 679int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
680{
681 return ((unsigned long) node) & DUMMY_FLAG;
682}
683
684static
14044b37 685struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 686{
14044b37 687 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 688}
bb7b2f26
MD
689
690static
691struct cds_lfht_node *get_end(void)
692{
693 return (struct cds_lfht_node *) END_VALUE;
694}
695
696static
697int is_end(struct cds_lfht_node *node)
698{
699 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
700}
701
abc490a1 702static
f9830efd 703unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
704{
705 unsigned long old1, old2;
706
707 old1 = uatomic_read(ptr);
708 do {
709 old2 = old1;
710 if (old2 >= v)
f9830efd 711 return old2;
abc490a1 712 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 713 return v;
abc490a1
MD
714}
715
1475579c
MD
716static
717void cds_lfht_free_level(struct rcu_head *head)
718{
719 struct rcu_level *l =
720 caa_container_of(head, struct rcu_level, head);
98808fb1 721 poison_free(l);
1475579c
MD
722}
723
273399de
MD
724/*
725 * Remove all logically deleted nodes from a bucket up to a certain node key.
726 */
727static
f9c80341 728void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 729{
14044b37 730 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 731
c90201ac
MD
732 assert(!is_dummy(dummy));
733 assert(!is_removed(dummy));
734 assert(!is_dummy(node));
735 assert(!is_removed(node));
273399de
MD
736 for (;;) {
737 iter_prev = dummy;
738 /* We can always skip the dummy node initially */
cc4fcb10
MD
739 iter = rcu_dereference(iter_prev->p.next);
740 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
741 /*
742 * We should never be called with dummy (start of chain)
743 * and logically removed node (end of path compression
744 * marker) being the actual same node. This would be a
745 * bug in the algorithm implementation.
746 */
747 assert(dummy != node);
273399de 748 for (;;) {
bb7b2f26 749 if (unlikely(is_end(iter)))
f9c80341 750 return;
76412f24 751 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 752 return;
cc4fcb10 753 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 754 if (likely(is_removed(next)))
273399de 755 break;
b453eae1 756 iter_prev = clear_flag(iter);
273399de
MD
757 iter = next;
758 }
b198f0fd 759 assert(!is_removed(iter));
f5596c94
MD
760 if (is_dummy(iter))
761 new_next = flag_dummy(clear_flag(next));
762 else
763 new_next = clear_flag(next);
48ed1c18
MD
764 if (is_removed(iter))
765 new_next = flag_removed(new_next);
f5596c94 766 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 767 }
f9c80341 768 return;
273399de
MD
769}
770
abc490a1 771static
4105056a
MD
772struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
773 unsigned long size,
774 struct cds_lfht_node *node,
48ed1c18 775 enum add_mode mode, int dummy)
abc490a1 776{
14044b37 777 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
adc0de68 778 *dummy_node, *return_node;
14044b37 779 struct _cds_lfht_node *lookup;
24365af7 780 unsigned long hash, index, order;
abc490a1 781
c90201ac
MD
782 assert(!is_dummy(node));
783 assert(!is_removed(node));
4105056a 784 if (!size) {
f5596c94 785 assert(dummy);
bb7b2f26 786 node->p.next = flag_dummy(get_end());
18117871
MD
787 return node; /* Initial first add (head) */
788 }
cc4fcb10 789 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 790 for (;;) {
adc0de68 791 uint32_t chain_len = 0;
abc490a1 792
11519af6
MD
793 /*
794 * iter_prev points to the non-removed node prior to the
795 * insert location.
11519af6 796 */
4105056a 797 index = hash & (size - 1);
24365af7 798 order = get_count_order_ulong(index + 1);
4105056a 799 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 800 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 801 /* We can always skip the dummy node initially */
cc4fcb10
MD
802 iter = rcu_dereference(iter_prev->p.next);
803 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 804 for (;;) {
bb7b2f26 805 if (unlikely(is_end(iter)))
273399de 806 goto insert;
76412f24 807 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 808 goto insert;
cc4fcb10 809 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 810 if (unlikely(is_removed(next)))
9dba85be 811 goto gc_node;
48ed1c18 812 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
1b81fe1a 813 && !is_dummy(next)
e43f23f8
MD
814 && !ht->compare_fct(node->key, node->key_len,
815 clear_flag(iter)->key,
48ed1c18
MD
816 clear_flag(iter)->key_len)) {
817 if (mode == ADD_UNIQUE)
818 return clear_flag(iter);
819 else /* mode == ADD_REPLACE */
820 goto replace;
821 }
11519af6 822 /* Only account for identical reverse hash once */
24365af7
MD
823 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
824 && !is_dummy(next))
4105056a 825 check_resize(ht, size, ++chain_len);
11519af6 826 iter_prev = clear_flag(iter);
273399de 827 iter = next;
abc490a1 828 }
48ed1c18 829
273399de 830 insert:
7ec59d3b 831 assert(node != clear_flag(iter));
11519af6 832 assert(!is_removed(iter_prev));
c90201ac 833 assert(!is_removed(iter));
f000907d 834 assert(iter_prev != node);
f9c80341 835 if (!dummy)
1b81fe1a 836 node->p.next = clear_flag(iter);
f9c80341
MD
837 else
838 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
839 if (is_dummy(iter))
840 new_node = flag_dummy(node);
841 else
842 new_node = node;
cc4fcb10 843 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 844 new_node) != iter) {
273399de 845 continue; /* retry */
48ed1c18
MD
846 } else {
847 if (mode == ADD_REPLACE)
848 return_node = NULL;
849 else /* ADD_DEFAULT and ADD_UNIQUE */
850 return_node = node;
273399de 851 goto gc_end;
48ed1c18
MD
852 }
853
854 replace:
adc0de68
MD
855 /* Insert after node to be replaced */
856 iter_prev = clear_flag(iter);
857 iter = next;
48ed1c18
MD
858 assert(node != clear_flag(iter));
859 assert(!is_removed(iter_prev));
860 assert(!is_removed(iter));
48ed1c18
MD
861 assert(iter_prev != node);
862 assert(!dummy);
adc0de68 863 node->p.next = clear_flag(iter);
48ed1c18
MD
864 if (is_dummy(iter))
865 new_node = flag_dummy(node);
866 else
867 new_node = node;
868 /*
adc0de68
MD
869 * Here is the whole trick for lock-free replace: we add
870 * the replacement node _after_ the node we want to
871 * replace by atomically setting its next pointer at the
b198f0fd 872 * same time we set its removal flag. Given that
adc0de68
MD
873 * the lookups/get next use an iterator aware of the
874 * next pointer, they will either skip the old node due
b198f0fd
MD
875 * to the removal flag and see the new node, or use
876 * the old node, but will not see the new one.
48ed1c18 877 */
adc0de68 878 new_node = flag_removed(new_node);
adc0de68
MD
879 if (uatomic_cmpxchg(&iter_prev->p.next,
880 iter, new_node) != iter) {
48ed1c18
MD
881 continue; /* retry */
882 } else {
adc0de68
MD
883 return_node = iter_prev;
884 goto gc_end;
48ed1c18
MD
885 }
886
9dba85be
MD
887 gc_node:
888 assert(!is_removed(iter));
f5596c94
MD
889 if (is_dummy(iter))
890 new_next = flag_dummy(clear_flag(next));
891 else
892 new_next = clear_flag(next);
893 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 894 /* retry */
464a1ec9 895 }
273399de
MD
896gc_end:
897 /* Garbage collect logically removed nodes in the bucket */
4105056a 898 index = hash & (size - 1);
24365af7 899 order = get_count_order_ulong(index + 1);
4105056a 900 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 901 dummy_node = (struct cds_lfht_node *) lookup;
f9c80341 902 _cds_lfht_gc_bucket(dummy_node, node);
48ed1c18 903 return return_node;
abc490a1 904}
464a1ec9 905
abc490a1 906static
860d07e8 907int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 908 struct cds_lfht_node *node,
b198f0fd 909 int dummy_removal)
abc490a1 910{
14044b37
MD
911 struct cds_lfht_node *dummy, *next, *old;
912 struct _cds_lfht_node *lookup;
abc490a1 913 int flagged = 0;
24365af7 914 unsigned long hash, index, order;
5e28c532 915
7ec59d3b 916 /* logically delete the node */
c90201ac
MD
917 assert(!is_dummy(node));
918 assert(!is_removed(node));
cc4fcb10 919 old = rcu_dereference(node->p.next);
7ec59d3b 920 do {
48ed1c18
MD
921 struct cds_lfht_node *new_next;
922
7ec59d3b 923 next = old;
76412f24 924 if (unlikely(is_removed(next)))
7ec59d3b 925 goto end;
1475579c
MD
926 if (dummy_removal)
927 assert(is_dummy(next));
928 else
929 assert(!is_dummy(next));
48ed1c18 930 new_next = flag_removed(next);
48ed1c18 931 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
932 } while (old != next);
933
934 /* We performed the (logical) deletion. */
935 flagged = 1;
936
937 /*
938 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
939 * the node, and remove it (along with any other logically removed node)
940 * if found.
11519af6 941 */
cc4fcb10 942 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
943 assert(size > 0);
944 index = hash & (size - 1);
24365af7 945 order = get_count_order_ulong(index + 1);
4105056a 946 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 947 dummy = (struct cds_lfht_node *) lookup;
f9c80341 948 _cds_lfht_gc_bucket(dummy, node);
2ed95849 949end:
11519af6
MD
950 /*
951 * Only the flagging action indicated that we (and no other)
952 * removed the node from the hash.
953 */
7ec59d3b 954 if (flagged) {
cc4fcb10 955 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 956 return 0;
7ec59d3b 957 } else
11519af6 958 return -ENOENT;
abc490a1 959}
2ed95849 960
b7d619b0
MD
961static
962void *partition_resize_thread(void *arg)
963{
964 struct partition_resize_work *work = arg;
965
966 work->ht->cds_lfht_rcu_register_thread();
967 work->fct(work->ht, work->i, work->start, work->len);
968 work->ht->cds_lfht_rcu_unregister_thread();
969 return NULL;
970}
971
972static
973void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
974 unsigned long len,
975 void (*fct)(struct cds_lfht *ht, unsigned long i,
976 unsigned long start, unsigned long len))
977{
978 unsigned long partition_len;
979 struct partition_resize_work *work;
6083a889
MD
980 int thread, ret;
981 unsigned long nr_threads;
b7d619b0
MD
982 pthread_t *thread_id;
983
6083a889
MD
984 /*
985 * Note: nr_cpus_mask + 1 is always power of 2.
986 * We spawn just the number of threads we need to satisfy the minimum
987 * partition size, up to the number of CPUs in the system.
988 */
989 nr_threads = min(nr_cpus_mask + 1,
990 len >> MIN_PARTITION_PER_THREAD_ORDER);
991 partition_len = len >> get_count_order_ulong(nr_threads);
992 work = calloc(nr_threads, sizeof(*work));
993 thread_id = calloc(nr_threads, sizeof(*thread_id));
b7d619b0 994 assert(work);
6083a889
MD
995 for (thread = 0; thread < nr_threads; thread++) {
996 work[thread].ht = ht;
997 work[thread].i = i;
998 work[thread].len = partition_len;
999 work[thread].start = thread * partition_len;
1000 work[thread].fct = fct;
1001 ret = pthread_create(&thread_id[thread], ht->resize_attr,
1002 partition_resize_thread, &work[thread]);
b7d619b0
MD
1003 assert(!ret);
1004 }
6083a889
MD
1005 for (thread = 0; thread < nr_threads; thread++) {
1006 ret = pthread_join(thread_id[thread], NULL);
b7d619b0
MD
1007 assert(!ret);
1008 }
1009 free(work);
1010 free(thread_id);
1011}
1012
e8de508e
MD
1013/*
1014 * Holding RCU read lock to protect _cds_lfht_add against memory
1015 * reclaim that could be performed by other call_rcu worker threads (ABA
1016 * problem).
9ee0fc9a 1017 *
b7d619b0 1018 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1019 * many worker threads, based on the number of CPUs available in the system.
1020 * This should therefore take care of not having the expand lagging behind too
1021 * many concurrent insertion threads by using the scheduler's ability to
1022 * schedule dummy node population fairly with insertions.
e8de508e 1023 */
4105056a 1024static
b7d619b0
MD
1025void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1026 unsigned long start, unsigned long len)
4105056a
MD
1027{
1028 unsigned long j;
1029
1030 ht->cds_lfht_rcu_read_lock();
b7d619b0 1031 for (j = start; j < start + len; j++) {
4105056a
MD
1032 struct cds_lfht_node *new_node =
1033 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1034
dc1da8f6 1035 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1036 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1037 new_node->p.reverse_hash =
1038 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a 1039 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1040 new_node, ADD_DEFAULT, 1);
4105056a
MD
1041 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1042 break;
1043 }
1044 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1045}
1046
1047static
1048void init_table_populate(struct cds_lfht *ht, unsigned long i,
1049 unsigned long len)
1050{
1051 assert(nr_cpus_mask != -1);
6083a889 1052 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1053 ht->cds_lfht_rcu_thread_online();
1054 init_table_populate_partition(ht, i, 0, len);
1055 ht->cds_lfht_rcu_thread_offline();
1056 return;
1057 }
1058 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1059}
1060
abc490a1 1061static
4105056a 1062void init_table(struct cds_lfht *ht,
24365af7
MD
1063 unsigned long first_order, unsigned long len_order)
1064{
1065 unsigned long i, end_order;
1066
f0c29ed7 1067 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1068 first_order, first_order + len_order);
1069 end_order = first_order + len_order;
24365af7 1070 for (i = first_order; i < end_order; i++) {
4105056a 1071 unsigned long len;
24365af7
MD
1072
1073 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1074 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1075
1076 /* Stop expand if the resize target changes under us */
1077 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1078 break;
1079
4105056a 1080 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1081 + (len * sizeof(struct _cds_lfht_node)));
b7d619b0 1082 assert(ht->t.tbl[i]);
4105056a 1083
4105056a 1084 /*
dc1da8f6
MD
1085 * Set all dummy nodes reverse hash values for a level and
1086 * link all dummy nodes into the table.
4105056a 1087 */
dc1da8f6 1088 init_table_populate(ht, i, len);
4105056a 1089
f9c80341
MD
1090 /*
1091 * Update table size.
1092 */
1093 cmm_smp_wmb(); /* populate data before RCU size */
1094 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1095
4105056a
MD
1096 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1097 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1098 break;
1099 }
1100}
1101
e8de508e
MD
1102/*
1103 * Holding RCU read lock to protect _cds_lfht_remove against memory
1104 * reclaim that could be performed by other call_rcu worker threads (ABA
1105 * problem).
1106 * For a single level, we logically remove and garbage collect each node.
1107 *
1108 * As a design choice, we perform logical removal and garbage collection on a
1109 * node-per-node basis to simplify this algorithm. We also assume keeping good
1110 * cache locality of the operation would overweight possible performance gain
1111 * that could be achieved by batching garbage collection for multiple levels.
1112 * However, this would have to be justified by benchmarks.
1113 *
1114 * Concurrent removal and add operations are helping us perform garbage
1115 * collection of logically removed nodes. We guarantee that all logically
1116 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1117 * invoked to free a hole level of dummy nodes (after a grace period).
1118 *
1119 * Logical removal and garbage collection can therefore be done in batch or on a
1120 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1121 *
b7d619b0
MD
1122 * When we reach a certain length, we can split this removal over many worker
1123 * threads, based on the number of CPUs available in the system. This should
1124 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1125 * updater threads actively inserting into the hash table.
e8de508e 1126 */
4105056a 1127static
b7d619b0
MD
1128void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1129 unsigned long start, unsigned long len)
4105056a
MD
1130{
1131 unsigned long j;
1132
1133 ht->cds_lfht_rcu_read_lock();
b7d619b0 1134 for (j = start; j < start + len; j++) {
4105056a
MD
1135 struct cds_lfht_node *fini_node =
1136 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1137
1138 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1139 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1140 fini_node->p.reverse_hash =
1141 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1142 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1143 fini_node, 1);
33c7c748
MD
1144 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1145 break;
abc490a1 1146 }
4105056a 1147 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1148}
1149
1150static
1151void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1152{
1153
1154 assert(nr_cpus_mask != -1);
6083a889 1155 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1156 ht->cds_lfht_rcu_thread_online();
1157 remove_table_partition(ht, i, 0, len);
1158 ht->cds_lfht_rcu_thread_offline();
1159 return;
1160 }
1161 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1162}
1163
1475579c 1164static
4105056a 1165void fini_table(struct cds_lfht *ht,
1475579c
MD
1166 unsigned long first_order, unsigned long len_order)
1167{
1168 long i, end_order;
1169
1170 dbg_printf("fini table: first_order %lu end_order %lu\n",
1171 first_order, first_order + len_order);
1172 end_order = first_order + len_order;
1173 assert(first_order > 0);
1475579c 1174 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1175 unsigned long len;
1475579c
MD
1176
1177 len = !i ? 1 : 1UL << (i - 1);
1178 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1179
4d676753
MD
1180 /* Stop shrink if the resize target changes under us */
1181 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1182 break;
1183
1184 cmm_smp_wmb(); /* populate data before RCU size */
1185 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1186
1187 /*
1188 * We need to wait for all add operations to reach Q.S. (and
1189 * thus use the new table for lookups) before we can start
1190 * releasing the old dummy nodes. Otherwise their lookup will
1191 * return a logically removed node as insert position.
1192 */
1193 ht->cds_lfht_synchronize_rcu();
1194
21263e21 1195 /*
4105056a
MD
1196 * Set "removed" flag in dummy nodes about to be removed.
1197 * Unlink all now-logically-removed dummy node pointers.
1198 * Concurrent add/remove operation are helping us doing
1199 * the gc.
21263e21 1200 */
4105056a
MD
1201 remove_table(ht, i, len);
1202
1203 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1204
1205 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1206 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1207 break;
1208 }
1475579c
MD
1209}
1210
7a9dcf9b 1211struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1212 cds_lfht_compare_fct compare_fct,
1213 unsigned long hash_seed,
1214 unsigned long init_size,
b8af5011 1215 int flags,
14044b37 1216 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1217 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1218 void (*cds_lfht_synchronize_rcu)(void),
1219 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1220 void (*cds_lfht_rcu_read_unlock)(void),
1221 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1222 void (*cds_lfht_rcu_thread_online)(void),
1223 void (*cds_lfht_rcu_register_thread)(void),
1224 void (*cds_lfht_rcu_unregister_thread)(void),
1225 pthread_attr_t *attr)
abc490a1 1226{
14044b37 1227 struct cds_lfht *ht;
24365af7 1228 unsigned long order;
abc490a1 1229
8129be4e 1230 /* init_size must be power of two */
49619ea0 1231 if (init_size && (init_size & (init_size - 1)))
8129be4e 1232 return NULL;
14044b37 1233 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1234 assert(ht);
abc490a1 1235 ht->hash_fct = hash_fct;
732ad076
MD
1236 ht->compare_fct = compare_fct;
1237 ht->hash_seed = hash_seed;
14044b37 1238 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1239 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1240 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1241 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1242 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1243 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1244 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1245 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1246 ht->resize_attr = attr;
df44348d 1247 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1248 /* this mutex should not nest in read-side C.S. */
1249 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1250 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1251 ht->flags = flags;
5f511391 1252 ht->cds_lfht_rcu_thread_offline();
f000907d 1253 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1254 ht->t.resize_target = 1UL << (order - 1);
4105056a 1255 init_table(ht, 0, order);
f000907d 1256 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1257 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1258 return ht;
1259}
1260
adc0de68
MD
1261void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1262 struct cds_lfht_iter *iter)
2ed95849 1263{
bb7b2f26 1264 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1265 struct _cds_lfht_node *lookup;
4105056a 1266 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1267
732ad076 1268 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1269 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1270
4105056a
MD
1271 size = rcu_dereference(ht->t.size);
1272 index = hash & (size - 1);
24365af7 1273 order = get_count_order_ulong(index + 1);
4105056a 1274 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1275 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1276 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1277 dummy_node = (struct cds_lfht_node *) lookup;
1278 /* We can always skip the dummy node initially */
1279 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1280 node = clear_flag(node);
2ed95849 1281 for (;;) {
bb7b2f26 1282 if (unlikely(is_end(node))) {
96ad1112 1283 node = next = NULL;
abc490a1 1284 break;
bb7b2f26 1285 }
cc4fcb10 1286 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1287 node = next = NULL;
abc490a1 1288 break;
2ed95849 1289 }
1b81fe1a 1290 next = rcu_dereference(node->p.next);
adc0de68 1291 if (likely(!is_removed(next))
1b81fe1a 1292 && !is_dummy(next)
49c2e2d6 1293 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1294 break;
2ed95849 1295 }
1b81fe1a 1296 node = clear_flag(next);
2ed95849 1297 }
1b81fe1a 1298 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1299 iter->node = node;
1300 iter->next = next;
abc490a1 1301}
e0ba718a 1302
adc0de68 1303void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1304{
adc0de68 1305 struct cds_lfht_node *node, *next;
a481e5ff
MD
1306 unsigned long reverse_hash;
1307 void *key;
1308 size_t key_len;
1309
adc0de68 1310 node = iter->node;
a481e5ff
MD
1311 reverse_hash = node->p.reverse_hash;
1312 key = node->key;
1313 key_len = node->key_len;
adc0de68 1314 next = iter->next;
a481e5ff
MD
1315 node = clear_flag(next);
1316
1317 for (;;) {
bb7b2f26 1318 if (unlikely(is_end(node))) {
96ad1112 1319 node = next = NULL;
a481e5ff 1320 break;
bb7b2f26 1321 }
a481e5ff 1322 if (unlikely(node->p.reverse_hash > reverse_hash)) {
96ad1112 1323 node = next = NULL;
a481e5ff
MD
1324 break;
1325 }
1326 next = rcu_dereference(node->p.next);
adc0de68 1327 if (likely(!is_removed(next))
a481e5ff
MD
1328 && !is_dummy(next)
1329 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1330 break;
1331 }
1332 node = clear_flag(next);
1333 }
1334 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1335 iter->node = node;
1336 iter->next = next;
a481e5ff
MD
1337}
1338
14044b37 1339void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1340{
4105056a 1341 unsigned long hash, size;
ab7d5fc6 1342
49c2e2d6 1343 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1344 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1345
4105056a 1346 size = rcu_dereference(ht->t.size);
48ed1c18 1347 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
4105056a 1348 ht_count_add(ht, size);
3eca1b8c
MD
1349}
1350
14044b37 1351struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1352 struct cds_lfht_node *node)
3eca1b8c 1353{
4105056a 1354 unsigned long hash, size;
df44348d 1355 struct cds_lfht_node *ret;
3eca1b8c 1356
49c2e2d6 1357 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1358 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1359
4105056a 1360 size = rcu_dereference(ht->t.size);
48ed1c18 1361 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
17f31d1b 1362 if (ret == node)
4105056a 1363 ht_count_add(ht, size);
df44348d 1364 return ret;
2ed95849
MD
1365}
1366
48ed1c18
MD
1367struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
1368 struct cds_lfht_node *node)
1369{
1370 unsigned long hash, size;
1371 struct cds_lfht_node *ret;
1372
1373 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1374 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1375
1376 size = rcu_dereference(ht->t.size);
1377 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1378 if (ret == NULL)
1379 ht_count_add(ht, size);
1380 return ret;
1381}
1382
860d07e8 1383int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1384{
4105056a 1385 unsigned long size;
df44348d 1386 int ret;
abc490a1 1387
4105056a 1388 size = rcu_dereference(ht->t.size);
b198f0fd 1389 ret = _cds_lfht_del(ht, size, node, 0);
df44348d 1390 if (!ret)
860d07e8 1391 ht_count_del(ht, size);
df44348d 1392 return ret;
2ed95849 1393}
ab7d5fc6 1394
abc490a1 1395static
14044b37 1396int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1397{
14044b37
MD
1398 struct cds_lfht_node *node;
1399 struct _cds_lfht_node *lookup;
4105056a 1400 unsigned long order, i, size;
674f7a69 1401
abc490a1 1402 /* Check that the table is empty */
4105056a 1403 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1404 node = (struct cds_lfht_node *) lookup;
abc490a1 1405 do {
1b81fe1a
MD
1406 node = clear_flag(node)->p.next;
1407 if (!is_dummy(node))
abc490a1 1408 return -EPERM;
273399de 1409 assert(!is_removed(node));
bb7b2f26 1410 } while (!is_end(node));
4105056a
MD
1411 /*
1412 * size accessed without rcu_dereference because hash table is
1413 * being destroyed.
1414 */
1415 size = ht->t.size;
abc490a1 1416 /* Internal sanity check: all nodes left should be dummy */
4105056a 1417 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1418 unsigned long len;
1419
1420 len = !order ? 1 : 1UL << (order - 1);
1421 for (i = 0; i < len; i++) {
f0c29ed7 1422 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1423 order, i,
4105056a
MD
1424 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1425 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1426 }
4105056a 1427 poison_free(ht->t.tbl[order]);
674f7a69 1428 }
abc490a1 1429 return 0;
674f7a69
MD
1430}
1431
1432/*
1433 * Should only be called when no more concurrent readers nor writers can
1434 * possibly access the table.
1435 */
b7d619b0 1436int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1437{
5e28c532
MD
1438 int ret;
1439
848d4088 1440 /* Wait for in-flight resize operations to complete */
33c7c748 1441 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
1442 while (uatomic_read(&ht->in_progress_resize))
1443 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1444 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1445 if (ret)
1446 return ret;
df44348d 1447 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1448 if (attr)
1449 *attr = ht->resize_attr;
98808fb1 1450 poison_free(ht);
5e28c532 1451 return ret;
674f7a69
MD
1452}
1453
14044b37 1454void cds_lfht_count_nodes(struct cds_lfht *ht,
d933dd0e 1455 long *approx_before,
273399de 1456 unsigned long *count,
973e5e1b 1457 unsigned long *removed,
d933dd0e 1458 long *approx_after)
273399de 1459{
14044b37
MD
1460 struct cds_lfht_node *node, *next;
1461 struct _cds_lfht_node *lookup;
24365af7 1462 unsigned long nr_dummy = 0;
273399de 1463
7ed7682f 1464 *approx_before = 0;
973e5e1b
MD
1465 if (nr_cpus_mask >= 0) {
1466 int i;
1467
1468 for (i = 0; i < nr_cpus_mask + 1; i++) {
1469 *approx_before += uatomic_read(&ht->percpu_count[i].add);
1470 *approx_before -= uatomic_read(&ht->percpu_count[i].del);
1471 }
1472 }
1473
273399de
MD
1474 *count = 0;
1475 *removed = 0;
1476
24365af7 1477 /* Count non-dummy nodes in the table */
4105056a 1478 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1479 node = (struct cds_lfht_node *) lookup;
273399de 1480 do {
cc4fcb10 1481 next = rcu_dereference(node->p.next);
b198f0fd 1482 if (is_removed(next)) {
973e5e1b
MD
1483 if (!is_dummy(next))
1484 (*removed)++;
1485 else
1486 (nr_dummy)++;
1b81fe1a 1487 } else if (!is_dummy(next))
273399de 1488 (*count)++;
24365af7
MD
1489 else
1490 (nr_dummy)++;
273399de 1491 node = clear_flag(next);
bb7b2f26 1492 } while (!is_end(node));
f0c29ed7 1493 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
7ed7682f 1494 *approx_after = 0;
973e5e1b
MD
1495 if (nr_cpus_mask >= 0) {
1496 int i;
1497
1498 for (i = 0; i < nr_cpus_mask + 1; i++) {
1499 *approx_after += uatomic_read(&ht->percpu_count[i].add);
1500 *approx_after -= uatomic_read(&ht->percpu_count[i].del);
1501 }
1502 }
273399de
MD
1503}
1504
1475579c 1505/* called with resize mutex held */
abc490a1 1506static
4105056a 1507void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1508 unsigned long old_size, unsigned long new_size)
abc490a1 1509{
1475579c 1510 unsigned long old_order, new_order;
1475579c
MD
1511
1512 old_order = get_count_order_ulong(old_size) + 1;
1513 new_order = get_count_order_ulong(new_size) + 1;
1514 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1515 old_size, old_order, new_size, new_order);
1475579c 1516 assert(new_size > old_size);
4105056a 1517 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1518}
1519
1520/* called with resize mutex held */
1521static
4105056a 1522void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1523 unsigned long old_size, unsigned long new_size)
464a1ec9 1524{
1475579c 1525 unsigned long old_order, new_order;
464a1ec9 1526
cd95516d 1527 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1528 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1529 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1530 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1531 old_size, old_order, new_size, new_order);
1475579c 1532 assert(new_size < old_size);
1475579c 1533
4105056a
MD
1534 /* Remove and unlink all dummy nodes to remove. */
1535 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1536}
1537
1475579c
MD
1538
1539/* called with resize mutex held */
1540static
1541void _do_cds_lfht_resize(struct cds_lfht *ht)
1542{
1543 unsigned long new_size, old_size;
4105056a
MD
1544
1545 /*
1546 * Resize table, re-do if the target size has changed under us.
1547 */
1548 do {
1549 ht->t.resize_initiated = 1;
1550 old_size = ht->t.size;
1551 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1552 if (old_size < new_size)
1553 _do_cds_lfht_grow(ht, old_size, new_size);
1554 else if (old_size > new_size)
1555 _do_cds_lfht_shrink(ht, old_size, new_size);
1556 ht->t.resize_initiated = 0;
1557 /* write resize_initiated before read resize_target */
1558 cmm_smp_mb();
4d676753 1559 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1560}
1561
abc490a1 1562static
4105056a 1563unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1564 int growth_order)
464a1ec9 1565{
4105056a
MD
1566 return _uatomic_max(&ht->t.resize_target,
1567 size << growth_order);
464a1ec9
MD
1568}
1569
1475579c 1570static
4105056a 1571void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1572 unsigned long count)
1475579c 1573{
cd95516d 1574 count = max(count, MIN_TABLE_SIZE);
4105056a 1575 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1576}
1577
1578void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1579{
4105056a
MD
1580 resize_target_update_count(ht, new_size);
1581 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1582 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1583 pthread_mutex_lock(&ht->resize_mutex);
1584 _do_cds_lfht_resize(ht);
1585 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1586 ht->cds_lfht_rcu_thread_online();
abc490a1 1587}
464a1ec9 1588
abc490a1
MD
1589static
1590void do_resize_cb(struct rcu_head *head)
1591{
1592 struct rcu_resize_work *work =
1593 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1594 struct cds_lfht *ht = work->ht;
abc490a1 1595
5f511391 1596 ht->cds_lfht_rcu_thread_offline();
abc490a1 1597 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1598 _do_cds_lfht_resize(ht);
abc490a1 1599 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1600 ht->cds_lfht_rcu_thread_online();
98808fb1 1601 poison_free(work);
848d4088
MD
1602 cmm_smp_mb(); /* finish resize before decrement */
1603 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1604}
1605
abc490a1 1606static
4105056a 1607void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1608{
abc490a1 1609 struct rcu_resize_work *work;
f9830efd 1610 unsigned long target_size;
abc490a1 1611
4105056a
MD
1612 target_size = resize_target_update(ht, size, growth);
1613 /* Store resize_target before read resize_initiated */
1614 cmm_smp_mb();
1615 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088
MD
1616 uatomic_inc(&ht->in_progress_resize);
1617 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1618 work = malloc(sizeof(*work));
1619 work->ht = ht;
14044b37 1620 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1621 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1622 }
ab7d5fc6 1623}
3171717f 1624
f8994aee
MD
1625#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1626
3171717f 1627static
4105056a 1628void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1629 unsigned long count)
1630{
1631 struct rcu_resize_work *work;
3171717f 1632
b8af5011
MD
1633 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1634 return;
4105056a
MD
1635 resize_target_update_count(ht, count);
1636 /* Store resize_target before read resize_initiated */
1637 cmm_smp_mb();
1638 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f
MD
1639 uatomic_inc(&ht->in_progress_resize);
1640 cmm_smp_mb(); /* increment resize count before calling it */
1641 work = malloc(sizeof(*work));
1642 work->ht = ht;
1643 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1644 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1645 }
1646}
f8994aee
MD
1647
1648#endif
This page took 0.188647 seconds and 4 git commands to generate.