call_rcu implementation: add missing static
[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
MD
172//#define MIN_TABLE_SIZE 128
173#define MIN_TABLE_SIZE 1
cd95516d 174
4105056a
MD
175#if (CAA_BITS_PER_LONG == 32)
176#define MAX_TABLE_ORDER 32
177#else
178#define MAX_TABLE_ORDER 64
179#endif
180
181#ifndef min
182#define min(a, b) ((a) < (b) ? (a) : (b))
183#endif
184
abc490a1
MD
185#ifndef max
186#define max(a, b) ((a) > (b) ? (a) : (b))
187#endif
2ed95849 188
d95bd160
MD
189/*
190 * The removed flag needs to be updated atomically with the pointer.
191 * The dummy flag does not require to be updated atomically with the
192 * pointer, but it is added as a pointer low bit flag to save space.
193 */
d37166c6 194#define REMOVED_FLAG (1UL << 0)
f5596c94
MD
195#define DUMMY_FLAG (1UL << 1)
196#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 197
bb7b2f26
MD
198/* Value of the end pointer. Should not interact with flags. */
199#define END_VALUE 0x4
200
df44348d 201struct ht_items_count {
3171717f 202 unsigned long add, remove;
df44348d
MD
203} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
204
1475579c
MD
205struct rcu_level {
206 struct rcu_head head;
207 struct _cds_lfht_node nodes[0];
208};
209
395270b6 210struct rcu_table {
4105056a 211 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 212 unsigned long resize_target;
11519af6 213 int resize_initiated;
4105056a 214 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
215};
216
14044b37 217struct cds_lfht {
4105056a 218 struct rcu_table t;
14044b37
MD
219 cds_lfht_hash_fct hash_fct;
220 cds_lfht_compare_fct compare_fct;
732ad076 221 unsigned long hash_seed;
b8af5011 222 int flags;
5f511391
MD
223 /*
224 * We need to put the work threads offline (QSBR) when taking this
225 * mutex, because we use synchronize_rcu within this mutex critical
226 * section, which waits on read-side critical sections, and could
227 * therefore cause grace-period deadlock if we hold off RCU G.P.
228 * completion.
229 */
464a1ec9 230 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 231 unsigned int in_progress_resize, in_progress_destroy;
14044b37 232 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 233 void (*func)(struct rcu_head *head));
1475579c 234 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
235 void (*cds_lfht_rcu_read_lock)(void);
236 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
237 void (*cds_lfht_rcu_thread_offline)(void);
238 void (*cds_lfht_rcu_thread_online)(void);
df44348d
MD
239 unsigned long count; /* global approximate item count */
240 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
241};
242
abc490a1
MD
243struct rcu_resize_work {
244 struct rcu_head head;
14044b37 245 struct cds_lfht *ht;
abc490a1 246};
2ed95849 247
76a73da8
MD
248static
249struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
250 unsigned long size,
251 struct cds_lfht_node *node,
252 int unique, int dummy);
253
abc490a1
MD
254/*
255 * Algorithm to reverse bits in a word by lookup table, extended to
256 * 64-bit words.
f9830efd 257 * Source:
abc490a1 258 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 259 * Originally from Public Domain.
abc490a1
MD
260 */
261
262static const uint8_t BitReverseTable256[256] =
2ed95849 263{
abc490a1
MD
264#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
265#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
266#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
267 R6(0), R6(2), R6(1), R6(3)
268};
269#undef R2
270#undef R4
271#undef R6
2ed95849 272
abc490a1
MD
273static
274uint8_t bit_reverse_u8(uint8_t v)
275{
276 return BitReverseTable256[v];
277}
ab7d5fc6 278
abc490a1
MD
279static __attribute__((unused))
280uint32_t bit_reverse_u32(uint32_t v)
281{
282 return ((uint32_t) bit_reverse_u8(v) << 24) |
283 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
284 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
285 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
286}
287
abc490a1
MD
288static __attribute__((unused))
289uint64_t bit_reverse_u64(uint64_t v)
2ed95849 290{
abc490a1
MD
291 return ((uint64_t) bit_reverse_u8(v) << 56) |
292 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
293 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
294 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
295 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
296 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
297 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
298 ((uint64_t) bit_reverse_u8(v >> 56));
299}
300
301static
302unsigned long bit_reverse_ulong(unsigned long v)
303{
304#if (CAA_BITS_PER_LONG == 32)
305 return bit_reverse_u32(v);
306#else
307 return bit_reverse_u64(v);
308#endif
309}
310
f9830efd 311/*
24365af7
MD
312 * fls: returns the position of the most significant bit.
313 * Returns 0 if no bit is set, else returns the position of the most
314 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 315 */
24365af7
MD
316#if defined(__i386) || defined(__x86_64)
317static inline
318unsigned int fls_u32(uint32_t x)
f9830efd 319{
24365af7
MD
320 int r;
321
322 asm("bsrl %1,%0\n\t"
323 "jnz 1f\n\t"
324 "movl $-1,%0\n\t"
325 "1:\n\t"
326 : "=r" (r) : "rm" (x));
327 return r + 1;
328}
329#define HAS_FLS_U32
330#endif
331
332#if defined(__x86_64)
333static inline
334unsigned int fls_u64(uint64_t x)
335{
336 long r;
337
338 asm("bsrq %1,%0\n\t"
339 "jnz 1f\n\t"
340 "movq $-1,%0\n\t"
341 "1:\n\t"
342 : "=r" (r) : "rm" (x));
343 return r + 1;
344}
345#define HAS_FLS_U64
346#endif
347
348#ifndef HAS_FLS_U64
349static __attribute__((unused))
350unsigned int fls_u64(uint64_t x)
351{
352 unsigned int r = 64;
353
354 if (!x)
355 return 0;
356
357 if (!(x & 0xFFFFFFFF00000000ULL)) {
358 x <<= 32;
359 r -= 32;
360 }
361 if (!(x & 0xFFFF000000000000ULL)) {
362 x <<= 16;
363 r -= 16;
364 }
365 if (!(x & 0xFF00000000000000ULL)) {
366 x <<= 8;
367 r -= 8;
368 }
369 if (!(x & 0xF000000000000000ULL)) {
370 x <<= 4;
371 r -= 4;
372 }
373 if (!(x & 0xC000000000000000ULL)) {
374 x <<= 2;
375 r -= 2;
376 }
377 if (!(x & 0x8000000000000000ULL)) {
378 x <<= 1;
379 r -= 1;
380 }
381 return r;
382}
383#endif
384
385#ifndef HAS_FLS_U32
386static __attribute__((unused))
387unsigned int fls_u32(uint32_t x)
388{
389 unsigned int r = 32;
f9830efd 390
24365af7
MD
391 if (!x)
392 return 0;
393 if (!(x & 0xFFFF0000U)) {
394 x <<= 16;
395 r -= 16;
396 }
397 if (!(x & 0xFF000000U)) {
398 x <<= 8;
399 r -= 8;
400 }
401 if (!(x & 0xF0000000U)) {
402 x <<= 4;
403 r -= 4;
404 }
405 if (!(x & 0xC0000000U)) {
406 x <<= 2;
407 r -= 2;
408 }
409 if (!(x & 0x80000000U)) {
410 x <<= 1;
411 r -= 1;
412 }
413 return r;
414}
415#endif
416
417unsigned int fls_ulong(unsigned long x)
f9830efd 418{
24365af7
MD
419#if (CAA_BITS_PER_lONG == 32)
420 return fls_u32(x);
421#else
422 return fls_u64(x);
423#endif
424}
f9830efd 425
24365af7
MD
426int get_count_order_u32(uint32_t x)
427{
428 int order;
429
430 order = fls_u32(x) - 1;
431 if (x & (x - 1))
432 order++;
433 return order;
434}
435
436int get_count_order_ulong(unsigned long x)
437{
438 int order;
439
440 order = fls_ulong(x) - 1;
441 if (x & (x - 1))
442 order++;
443 return order;
f9830efd
MD
444}
445
98808fb1
MD
446#ifdef POISON_FREE
447#define poison_free(ptr) \
448 do { \
449 memset(ptr, 0x42, sizeof(*(ptr))); \
450 free(ptr); \
451 } while (0)
452#else
453#define poison_free(ptr) free(ptr)
454#endif
455
f9830efd 456static
4105056a 457void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 458
df44348d
MD
459/*
460 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
461 * available, then we support hash table item accounting.
462 * In the unfortunate event the number of CPUs reported would be
463 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
464 */
df44348d
MD
465#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
466
f8994aee 467static
4105056a 468void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
469 unsigned long count);
470
df44348d
MD
471static long nr_cpus_mask = -1;
472
473static
474struct ht_items_count *alloc_per_cpu_items_count(void)
475{
476 struct ht_items_count *count;
477
478 switch (nr_cpus_mask) {
479 case -2:
480 return NULL;
481 case -1:
482 {
483 long maxcpus;
484
485 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
486 if (maxcpus <= 0) {
487 nr_cpus_mask = -2;
488 return NULL;
489 }
490 /*
491 * round up number of CPUs to next power of two, so we
492 * can use & for modulo.
493 */
494 maxcpus = 1UL << get_count_order_ulong(maxcpus);
495 nr_cpus_mask = maxcpus - 1;
496 }
497 /* Fall-through */
498 default:
499 return calloc(nr_cpus_mask + 1, sizeof(*count));
500 }
501}
502
503static
504void free_per_cpu_items_count(struct ht_items_count *count)
505{
98808fb1 506 poison_free(count);
df44348d
MD
507}
508
509static
510int ht_get_cpu(void)
511{
512 int cpu;
513
514 assert(nr_cpus_mask >= 0);
515 cpu = sched_getcpu();
516 if (unlikely(cpu < 0))
517 return cpu;
518 else
519 return cpu & nr_cpus_mask;
520}
521
522static
4105056a 523void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 524{
3171717f 525 unsigned long percpu_count;
df44348d
MD
526 int cpu;
527
528 if (unlikely(!ht->percpu_count))
3171717f 529 return;
df44348d
MD
530 cpu = ht_get_cpu();
531 if (unlikely(cpu < 0))
3171717f
MD
532 return;
533 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d
MD
534 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
535 unsigned long count;
536
537 dbg_printf("add percpu %lu\n", percpu_count);
538 count = uatomic_add_return(&ht->count,
539 1UL << COUNT_COMMIT_ORDER);
540 /* If power of 2 */
541 if (!(count & (count - 1))) {
4105056a 542 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee
MD
543 return;
544 dbg_printf("add set global %lu\n", count);
4105056a 545 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 546 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
547 }
548 }
549}
550
551static
4105056a 552void ht_count_remove(struct cds_lfht *ht, unsigned long size)
df44348d
MD
553{
554 unsigned long percpu_count;
3171717f 555 int cpu;
df44348d 556
3171717f
MD
557 if (unlikely(!ht->percpu_count))
558 return;
559 cpu = ht_get_cpu();
560 if (unlikely(cpu < 0))
561 return;
562 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1);
df44348d
MD
563 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
564 unsigned long count;
565
566 dbg_printf("remove percpu %lu\n", percpu_count);
567 count = uatomic_add_return(&ht->count,
3171717f 568 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
569 /* If power of 2 */
570 if (!(count & (count - 1))) {
4105056a 571 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee
MD
572 return;
573 dbg_printf("remove set global %lu\n", count);
4105056a 574 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 575 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
576 }
577 }
578}
579
580#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
581
582static const long nr_cpus_mask = -1;
583
584static
585struct ht_items_count *alloc_per_cpu_items_count(void)
586{
587 return NULL;
588}
589
590static
591void free_per_cpu_items_count(struct ht_items_count *count)
592{
593}
594
595static
4105056a 596void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
597{
598}
599
600static
4105056a 601void ht_count_remove(struct cds_lfht *ht, unsigned long size)
df44348d
MD
602{
603}
604
605#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
606
607
f9830efd 608static
4105056a 609void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 610{
f8994aee
MD
611 unsigned long count;
612
b8af5011
MD
613 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
614 return;
f8994aee
MD
615 count = uatomic_read(&ht->count);
616 /*
617 * Use bucket-local length for small table expand and for
618 * environments lacking per-cpu data support.
619 */
620 if (count >= (1UL << COUNT_COMMIT_ORDER))
621 return;
24365af7 622 if (chain_len > 100)
f0c29ed7 623 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 624 chain_len);
3390d470 625 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 626 cds_lfht_resize_lazy(ht, size,
01370f0b 627 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
628}
629
abc490a1 630static
14044b37 631struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 632{
14044b37 633 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
634}
635
636static
14044b37 637int is_removed(struct cds_lfht_node *node)
abc490a1 638{
d37166c6 639 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
640}
641
642static
14044b37 643struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 644{
14044b37 645 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
646}
647
f5596c94 648static
14044b37 649int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
650{
651 return ((unsigned long) node) & DUMMY_FLAG;
652}
653
654static
14044b37 655struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 656{
14044b37 657 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 658}
bb7b2f26
MD
659
660static
661struct cds_lfht_node *get_end(void)
662{
663 return (struct cds_lfht_node *) END_VALUE;
664}
665
666static
667int is_end(struct cds_lfht_node *node)
668{
669 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
670}
671
abc490a1 672static
f9830efd 673unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
674{
675 unsigned long old1, old2;
676
677 old1 = uatomic_read(ptr);
678 do {
679 old2 = old1;
680 if (old2 >= v)
f9830efd 681 return old2;
abc490a1 682 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 683 return v;
abc490a1
MD
684}
685
1475579c
MD
686static
687void cds_lfht_free_level(struct rcu_head *head)
688{
689 struct rcu_level *l =
690 caa_container_of(head, struct rcu_level, head);
98808fb1 691 poison_free(l);
1475579c
MD
692}
693
273399de
MD
694/*
695 * Remove all logically deleted nodes from a bucket up to a certain node key.
696 */
697static
76a73da8 698int _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 699{
14044b37 700 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 701
c90201ac
MD
702 assert(!is_dummy(dummy));
703 assert(!is_removed(dummy));
704 assert(!is_dummy(node));
705 assert(!is_removed(node));
273399de
MD
706 for (;;) {
707 iter_prev = dummy;
708 /* We can always skip the dummy node initially */
cc4fcb10 709 iter = rcu_dereference(iter_prev->p.next);
76a73da8
MD
710 if (unlikely(iter == NULL)) {
711 /*
712 * We are executing concurrently with a hash table
713 * expand, so we see a dummy node with NULL next value.
714 * Help expand by linking this node into the list and
715 * retry.
716 */
717 return 1;
718 }
cc4fcb10 719 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
720 /*
721 * We should never be called with dummy (start of chain)
722 * and logically removed node (end of path compression
723 * marker) being the actual same node. This would be a
724 * bug in the algorithm implementation.
725 */
726 assert(dummy != node);
273399de 727 for (;;) {
bb7b2f26
MD
728 assert(iter != NULL);
729 if (unlikely(is_end(iter)))
76a73da8 730 return 0;
76412f24 731 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
76a73da8 732 return 0;
cc4fcb10 733 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 734 if (likely(is_removed(next)))
273399de 735 break;
b453eae1 736 iter_prev = clear_flag(iter);
273399de
MD
737 iter = next;
738 }
739 assert(!is_removed(iter));
f5596c94
MD
740 if (is_dummy(iter))
741 new_next = flag_dummy(clear_flag(next));
742 else
743 new_next = clear_flag(next);
bb7b2f26 744 assert(new_next != NULL);
f5596c94 745 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 746 }
76a73da8 747 return 0;
273399de
MD
748}
749
abc490a1 750static
4105056a
MD
751struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
752 unsigned long size,
753 struct cds_lfht_node *node,
754 int unique, int dummy)
abc490a1 755{
14044b37 756 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
f5596c94 757 *dummy_node;
14044b37 758 struct _cds_lfht_node *lookup;
24365af7 759 unsigned long hash, index, order;
bb7b2f26 760 int force_dummy = 0;
abc490a1 761
c90201ac
MD
762 assert(!is_dummy(node));
763 assert(!is_removed(node));
4105056a 764 if (!size) {
f5596c94 765 assert(dummy);
bb7b2f26 766 node->p.next = flag_dummy(get_end());
18117871
MD
767 return node; /* Initial first add (head) */
768 }
cc4fcb10 769 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 770 for (;;) {
f9830efd 771 uint32_t chain_len = 0;
abc490a1 772
11519af6
MD
773 /*
774 * iter_prev points to the non-removed node prior to the
775 * insert location.
11519af6 776 */
4105056a 777 index = hash & (size - 1);
24365af7 778 order = get_count_order_ulong(index + 1);
4105056a 779 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 780 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 781 /* We can always skip the dummy node initially */
cc4fcb10 782 iter = rcu_dereference(iter_prev->p.next);
76a73da8
MD
783 if (unlikely(iter == NULL)) {
784 /*
785 * We are executing concurrently with a hash table
786 * expand, so we see a dummy node with NULL next value.
787 * Help expand by linking this node into the list and
788 * retry.
789 */
790 (void) _cds_lfht_add(ht, size >> 1, iter_prev, 0, 1);
791 continue; /* retry */
792 }
cc4fcb10 793 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 794 for (;;) {
bb7b2f26 795 assert(iter != NULL);
76a73da8
MD
796 /*
797 * When adding a dummy node, we allow concurrent
798 * add/removal to help. If we find the dummy node in
799 * place, skip its insertion.
800 */
801 if (unlikely(dummy && clear_flag(iter) == node))
802 return node;
bb7b2f26 803 if (unlikely(is_end(iter)))
273399de 804 goto insert;
76412f24 805 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 806 goto insert;
cc4fcb10 807 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 808 if (unlikely(is_removed(next)))
9dba85be 809 goto gc_node;
e43f23f8 810 if (unique
1b81fe1a 811 && !is_dummy(next)
e43f23f8
MD
812 && !ht->compare_fct(node->key, node->key_len,
813 clear_flag(iter)->key,
814 clear_flag(iter)->key_len))
18117871 815 return clear_flag(iter);
11519af6 816 /* Only account for identical reverse hash once */
24365af7
MD
817 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
818 && !is_dummy(next))
4105056a 819 check_resize(ht, size, ++chain_len);
11519af6 820 iter_prev = clear_flag(iter);
273399de 821 iter = next;
abc490a1 822 }
273399de 823 insert:
7ec59d3b 824 assert(node != clear_flag(iter));
11519af6 825 assert(!is_removed(iter_prev));
c90201ac 826 assert(!is_removed(iter));
f000907d 827 assert(iter_prev != node);
bb7b2f26 828 if (!dummy) {
1b81fe1a 829 node->p.next = clear_flag(iter);
bb7b2f26
MD
830 } else {
831 /*
832 * Dummy node insertion is performed concurrently (help
833 * scheme). We try to link its next node, and if this
834 * succeeds, it _means_ it's us who link this dummy node
835 * into the table. force_dummy is set as soon as we
836 * succeed this cmpxchg within this function.
837 */
838 if (!force_dummy) {
839 if (uatomic_cmpxchg(&node->p.next, NULL,
840 flag_dummy(clear_flag(iter))) != NULL) {
841 return NULL;
842 }
843 force_dummy = 1;
844 } else {
845 node->p.next = flag_dummy(clear_flag(iter));
846 }
847 }
f5596c94
MD
848 if (is_dummy(iter))
849 new_node = flag_dummy(node);
850 else
851 new_node = node;
bb7b2f26 852 assert(new_node != NULL);
cc4fcb10 853 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
f5596c94 854 new_node) != iter)
273399de 855 continue; /* retry */
11519af6 856 else
273399de 857 goto gc_end;
9dba85be
MD
858 gc_node:
859 assert(!is_removed(iter));
f5596c94
MD
860 if (is_dummy(iter))
861 new_next = flag_dummy(clear_flag(next));
862 else
863 new_next = clear_flag(next);
bb7b2f26 864 assert(new_next != NULL);
f5596c94 865 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 866 /* retry */
464a1ec9 867 }
273399de
MD
868gc_end:
869 /* Garbage collect logically removed nodes in the bucket */
4105056a 870 index = hash & (size - 1);
24365af7 871 order = get_count_order_ulong(index + 1);
4105056a 872 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 873 dummy_node = (struct cds_lfht_node *) lookup;
76a73da8
MD
874 if (_cds_lfht_gc_bucket(dummy_node, node)) {
875 /* Help expand */
876 (void) _cds_lfht_add(ht, size >> 1, dummy_node, 0, 1);
877 goto gc_end; /* retry */
878 }
18117871 879 return node;
abc490a1 880}
464a1ec9 881
abc490a1 882static
4105056a
MD
883int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
884 struct cds_lfht_node *node,
885 int dummy_removal)
abc490a1 886{
14044b37
MD
887 struct cds_lfht_node *dummy, *next, *old;
888 struct _cds_lfht_node *lookup;
abc490a1 889 int flagged = 0;
24365af7 890 unsigned long hash, index, order;
5e28c532 891
7ec59d3b 892 /* logically delete the node */
c90201ac
MD
893 assert(!is_dummy(node));
894 assert(!is_removed(node));
cc4fcb10 895 old = rcu_dereference(node->p.next);
7ec59d3b
MD
896 do {
897 next = old;
76412f24 898 if (unlikely(is_removed(next)))
7ec59d3b 899 goto end;
1475579c
MD
900 if (dummy_removal)
901 assert(is_dummy(next));
902 else
903 assert(!is_dummy(next));
bb7b2f26 904 assert(next != NULL);
cc4fcb10 905 old = uatomic_cmpxchg(&node->p.next, next,
7ec59d3b
MD
906 flag_removed(next));
907 } while (old != next);
908
909 /* We performed the (logical) deletion. */
910 flagged = 1;
911
912 /*
913 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
914 * the node, and remove it (along with any other logically removed node)
915 * if found.
11519af6 916 */
76a73da8 917gc_retry:
cc4fcb10 918 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
919 assert(size > 0);
920 index = hash & (size - 1);
24365af7 921 order = get_count_order_ulong(index + 1);
4105056a 922 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 923 dummy = (struct cds_lfht_node *) lookup;
76a73da8
MD
924 if (_cds_lfht_gc_bucket(dummy, node)) {
925 /* Help expand */
926 (void) _cds_lfht_add(ht, size >> 1, dummy, 0, 1);
927 goto gc_retry; /* retry */
928 }
2ed95849 929end:
11519af6
MD
930 /*
931 * Only the flagging action indicated that we (and no other)
932 * removed the node from the hash.
933 */
7ec59d3b 934 if (flagged) {
cc4fcb10 935 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 936 return 0;
7ec59d3b 937 } else
11519af6 938 return -ENOENT;
abc490a1 939}
2ed95849 940
4105056a
MD
941static
942void init_table_hash(struct cds_lfht *ht, unsigned long i,
943 unsigned long len)
944{
945 unsigned long j;
946
947 for (j = 0; j < len; j++) {
948 struct cds_lfht_node *new_node =
949 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
950
951 dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
952 i, j, !i ? 0 : (1UL << (i - 1)) + j);
953 new_node->p.reverse_hash =
954 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
955 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
956 break;
957 }
958}
959
e8de508e
MD
960/*
961 * Holding RCU read lock to protect _cds_lfht_add against memory
962 * reclaim that could be performed by other call_rcu worker threads (ABA
963 * problem).
964 */
4105056a
MD
965static
966void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
967{
968 unsigned long j;
969
5f511391 970 ht->cds_lfht_rcu_thread_online();
4105056a
MD
971 ht->cds_lfht_rcu_read_lock();
972 for (j = 0; j < len; j++) {
973 struct cds_lfht_node *new_node =
974 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
975
976 dbg_printf("init link: i %lu j %lu hash %lu\n",
977 i, j, !i ? 0 : (1UL << (i - 1)) + j);
978 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
979 new_node, 0, 1);
980 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
981 break;
982 }
983 ht->cds_lfht_rcu_read_unlock();
5f511391 984 ht->cds_lfht_rcu_thread_offline();
4105056a
MD
985}
986
abc490a1 987static
4105056a 988void init_table(struct cds_lfht *ht,
24365af7
MD
989 unsigned long first_order, unsigned long len_order)
990{
991 unsigned long i, end_order;
992
f0c29ed7 993 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
994 first_order, first_order + len_order);
995 end_order = first_order + len_order;
24365af7 996 for (i = first_order; i < end_order; i++) {
4105056a 997 unsigned long len;
24365af7
MD
998
999 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1000 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1001
1002 /* Stop expand if the resize target changes under us */
1003 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1004 break;
1005
4105056a 1006 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1007 + (len * sizeof(struct _cds_lfht_node)));
4105056a
MD
1008
1009 /* Set all dummy nodes reverse hash values for a level */
1010 init_table_hash(ht, i, len);
1011
76a73da8
MD
1012 /*
1013 * Update table size. At this point, concurrent add/remove see
1014 * dummy nodes with correctly initialized reverse hash value,
1015 * but with NULL next pointers. If they do, they can help us
1016 * link the dummy nodes into the list and retry.
1017 */
1018 cmm_smp_wmb(); /* populate data before RCU size */
1019 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1020
4105056a
MD
1021 /*
1022 * Link all dummy nodes into the table. Concurrent
1023 * add/remove are helping us.
1024 */
1025 init_table_link(ht, i, len);
1026
4105056a
MD
1027 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1028 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1029 break;
1030 }
1031}
1032
e8de508e
MD
1033/*
1034 * Holding RCU read lock to protect _cds_lfht_remove against memory
1035 * reclaim that could be performed by other call_rcu worker threads (ABA
1036 * problem).
1037 * For a single level, we logically remove and garbage collect each node.
1038 *
1039 * As a design choice, we perform logical removal and garbage collection on a
1040 * node-per-node basis to simplify this algorithm. We also assume keeping good
1041 * cache locality of the operation would overweight possible performance gain
1042 * that could be achieved by batching garbage collection for multiple levels.
1043 * However, this would have to be justified by benchmarks.
1044 *
1045 * Concurrent removal and add operations are helping us perform garbage
1046 * collection of logically removed nodes. We guarantee that all logically
1047 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1048 * invoked to free a hole level of dummy nodes (after a grace period).
1049 *
1050 * Logical removal and garbage collection can therefore be done in batch or on a
1051 * node-per-node basis, as long as the guarantee above holds.
1052 */
4105056a
MD
1053static
1054void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1055{
1056 unsigned long j;
1057
5f511391 1058 ht->cds_lfht_rcu_thread_online();
4105056a
MD
1059 ht->cds_lfht_rcu_read_lock();
1060 for (j = 0; j < len; j++) {
1061 struct cds_lfht_node *fini_node =
1062 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1063
1064 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1065 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1066 fini_node->p.reverse_hash =
1067 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
1068 (void) _cds_lfht_remove(ht, !i ? 0 : (1UL << (i - 1)),
1069 fini_node, 1);
33c7c748
MD
1070 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1071 break;
abc490a1 1072 }
4105056a 1073 ht->cds_lfht_rcu_read_unlock();
5f511391 1074 ht->cds_lfht_rcu_thread_offline();
2ed95849
MD
1075}
1076
1475579c 1077static
4105056a 1078void fini_table(struct cds_lfht *ht,
1475579c
MD
1079 unsigned long first_order, unsigned long len_order)
1080{
1081 long i, end_order;
1082
1083 dbg_printf("fini table: first_order %lu end_order %lu\n",
1084 first_order, first_order + len_order);
1085 end_order = first_order + len_order;
1086 assert(first_order > 0);
1475579c 1087 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1088 unsigned long len;
1475579c
MD
1089
1090 len = !i ? 1 : 1UL << (i - 1);
1091 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1092
4d676753
MD
1093 /* Stop shrink if the resize target changes under us */
1094 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1095 break;
1096
1097 cmm_smp_wmb(); /* populate data before RCU size */
1098 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1099
1100 /*
1101 * We need to wait for all add operations to reach Q.S. (and
1102 * thus use the new table for lookups) before we can start
1103 * releasing the old dummy nodes. Otherwise their lookup will
1104 * return a logically removed node as insert position.
1105 */
1106 ht->cds_lfht_synchronize_rcu();
1107
21263e21 1108 /*
4105056a
MD
1109 * Set "removed" flag in dummy nodes about to be removed.
1110 * Unlink all now-logically-removed dummy node pointers.
1111 * Concurrent add/remove operation are helping us doing
1112 * the gc.
21263e21 1113 */
4105056a
MD
1114 remove_table(ht, i, len);
1115
1116 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1117
1118 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1119 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1120 break;
1121 }
1475579c
MD
1122}
1123
14044b37
MD
1124struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
1125 cds_lfht_compare_fct compare_fct,
1126 unsigned long hash_seed,
1127 unsigned long init_size,
b8af5011 1128 int flags,
14044b37 1129 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1130 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1131 void (*cds_lfht_synchronize_rcu)(void),
1132 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1133 void (*cds_lfht_rcu_read_unlock)(void),
1134 void (*cds_lfht_rcu_thread_offline)(void),
1135 void (*cds_lfht_rcu_thread_online)(void))
abc490a1 1136{
14044b37 1137 struct cds_lfht *ht;
24365af7 1138 unsigned long order;
abc490a1 1139
8129be4e 1140 /* init_size must be power of two */
49619ea0 1141 if (init_size && (init_size & (init_size - 1)))
8129be4e 1142 return NULL;
14044b37 1143 ht = calloc(1, sizeof(struct cds_lfht));
abc490a1 1144 ht->hash_fct = hash_fct;
732ad076
MD
1145 ht->compare_fct = compare_fct;
1146 ht->hash_seed = hash_seed;
14044b37 1147 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1148 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1149 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1150 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1151 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1152 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
df44348d 1153 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1154 /* this mutex should not nest in read-side C.S. */
1155 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1156 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1157 ht->flags = flags;
5f511391 1158 ht->cds_lfht_rcu_thread_offline();
f000907d 1159 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1160 ht->t.resize_target = 1UL << (order - 1);
4105056a 1161 init_table(ht, 0, order);
f000907d 1162 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1163 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1164 return ht;
1165}
1166
14044b37 1167struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
2ed95849 1168{
bb7b2f26 1169 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1170 struct _cds_lfht_node *lookup;
4105056a 1171 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1172
732ad076 1173 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1174 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1175
bb7b2f26 1176restart:
4105056a
MD
1177 size = rcu_dereference(ht->t.size);
1178 index = hash & (size - 1);
24365af7 1179 order = get_count_order_ulong(index + 1);
4105056a 1180 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1181 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1182 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1183 dummy_node = (struct cds_lfht_node *) lookup;
1184 /* We can always skip the dummy node initially */
1185 node = rcu_dereference(dummy_node->p.next);
1186 if (unlikely(node == NULL)) {
1187 /*
1188 * We are executing concurrently with a hash table
1189 * expand, so we see a dummy node with NULL next value.
1190 * Help expand by linking this node into the list and
1191 * retry.
1192 */
1193 (void) _cds_lfht_add(ht, size >> 1, dummy_node, 0, 1);
1194 goto restart; /* retry */
1195 }
1196 node = clear_flag(node);
2ed95849 1197 for (;;) {
bb7b2f26
MD
1198 if (unlikely(is_end(node))) {
1199 node = NULL;
abc490a1 1200 break;
bb7b2f26 1201 }
cc4fcb10 1202 if (unlikely(node->p.reverse_hash > reverse_hash)) {
abc490a1
MD
1203 node = NULL;
1204 break;
2ed95849 1205 }
1b81fe1a
MD
1206 next = rcu_dereference(node->p.next);
1207 if (likely(!is_removed(next))
1208 && !is_dummy(next)
49c2e2d6 1209 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1210 break;
2ed95849 1211 }
1b81fe1a 1212 node = clear_flag(next);
2ed95849 1213 }
1b81fe1a 1214 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
abc490a1
MD
1215 return node;
1216}
e0ba718a 1217
a481e5ff
MD
1218struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
1219 struct cds_lfht_node *node)
1220{
1221 struct cds_lfht_node *next;
1222 unsigned long reverse_hash;
1223 void *key;
1224 size_t key_len;
1225
1226 reverse_hash = node->p.reverse_hash;
1227 key = node->key;
1228 key_len = node->key_len;
1229 next = rcu_dereference(node->p.next);
1230 node = clear_flag(next);
1231
1232 for (;;) {
bb7b2f26
MD
1233 if (unlikely(is_end(node))) {
1234 node = NULL;
a481e5ff 1235 break;
bb7b2f26 1236 }
a481e5ff
MD
1237 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1238 node = NULL;
1239 break;
1240 }
1241 next = rcu_dereference(node->p.next);
1242 if (likely(!is_removed(next))
1243 && !is_dummy(next)
1244 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1245 break;
1246 }
1247 node = clear_flag(next);
1248 }
1249 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1250 return node;
1251}
1252
14044b37 1253void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1254{
4105056a 1255 unsigned long hash, size;
ab7d5fc6 1256
49c2e2d6 1257 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1258 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1259
4105056a
MD
1260 size = rcu_dereference(ht->t.size);
1261 (void) _cds_lfht_add(ht, size, node, 0, 0);
1262 ht_count_add(ht, size);
3eca1b8c
MD
1263}
1264
14044b37
MD
1265struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1266 struct cds_lfht_node *node)
3eca1b8c 1267{
4105056a 1268 unsigned long hash, size;
df44348d 1269 struct cds_lfht_node *ret;
3eca1b8c 1270
49c2e2d6 1271 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1272 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1273
4105056a
MD
1274 size = rcu_dereference(ht->t.size);
1275 ret = _cds_lfht_add(ht, size, node, 1, 0);
17f31d1b 1276 if (ret == node)
4105056a 1277 ht_count_add(ht, size);
df44348d 1278 return ret;
2ed95849
MD
1279}
1280
14044b37 1281int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1282{
4105056a 1283 unsigned long size;
df44348d 1284 int ret;
abc490a1 1285
4105056a
MD
1286 size = rcu_dereference(ht->t.size);
1287 ret = _cds_lfht_remove(ht, size, node, 0);
df44348d 1288 if (!ret)
4105056a 1289 ht_count_remove(ht, size);
df44348d 1290 return ret;
2ed95849 1291}
ab7d5fc6 1292
abc490a1 1293static
14044b37 1294int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1295{
14044b37
MD
1296 struct cds_lfht_node *node;
1297 struct _cds_lfht_node *lookup;
4105056a 1298 unsigned long order, i, size;
674f7a69 1299
abc490a1 1300 /* Check that the table is empty */
4105056a 1301 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1302 node = (struct cds_lfht_node *) lookup;
abc490a1 1303 do {
1b81fe1a
MD
1304 node = clear_flag(node)->p.next;
1305 if (!is_dummy(node))
abc490a1 1306 return -EPERM;
273399de 1307 assert(!is_removed(node));
bb7b2f26 1308 } while (!is_end(node));
4105056a
MD
1309 /*
1310 * size accessed without rcu_dereference because hash table is
1311 * being destroyed.
1312 */
1313 size = ht->t.size;
abc490a1 1314 /* Internal sanity check: all nodes left should be dummy */
4105056a 1315 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1316 unsigned long len;
1317
1318 len = !order ? 1 : 1UL << (order - 1);
1319 for (i = 0; i < len; i++) {
f0c29ed7 1320 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1321 order, i,
4105056a
MD
1322 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1323 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1324 }
4105056a 1325 poison_free(ht->t.tbl[order]);
674f7a69 1326 }
abc490a1 1327 return 0;
674f7a69
MD
1328}
1329
1330/*
1331 * Should only be called when no more concurrent readers nor writers can
1332 * possibly access the table.
1333 */
14044b37 1334int cds_lfht_destroy(struct cds_lfht *ht)
674f7a69 1335{
5e28c532
MD
1336 int ret;
1337
848d4088 1338 /* Wait for in-flight resize operations to complete */
33c7c748 1339 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
1340 while (uatomic_read(&ht->in_progress_resize))
1341 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1342 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1343 if (ret)
1344 return ret;
df44348d 1345 free_per_cpu_items_count(ht->percpu_count);
98808fb1 1346 poison_free(ht);
5e28c532 1347 return ret;
674f7a69
MD
1348}
1349
14044b37 1350void cds_lfht_count_nodes(struct cds_lfht *ht,
273399de
MD
1351 unsigned long *count,
1352 unsigned long *removed)
1353{
14044b37
MD
1354 struct cds_lfht_node *node, *next;
1355 struct _cds_lfht_node *lookup;
24365af7 1356 unsigned long nr_dummy = 0;
273399de
MD
1357
1358 *count = 0;
1359 *removed = 0;
1360
24365af7 1361 /* Count non-dummy nodes in the table */
4105056a 1362 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1363 node = (struct cds_lfht_node *) lookup;
273399de 1364 do {
cc4fcb10 1365 next = rcu_dereference(node->p.next);
273399de 1366 if (is_removed(next)) {
1b81fe1a 1367 assert(!is_dummy(next));
273399de 1368 (*removed)++;
1b81fe1a 1369 } else if (!is_dummy(next))
273399de 1370 (*count)++;
24365af7
MD
1371 else
1372 (nr_dummy)++;
273399de 1373 node = clear_flag(next);
bb7b2f26 1374 } while (!is_end(node));
f0c29ed7 1375 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
273399de
MD
1376}
1377
1475579c 1378/* called with resize mutex held */
abc490a1 1379static
4105056a 1380void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1381 unsigned long old_size, unsigned long new_size)
abc490a1 1382{
1475579c 1383 unsigned long old_order, new_order;
1475579c
MD
1384
1385 old_order = get_count_order_ulong(old_size) + 1;
1386 new_order = get_count_order_ulong(new_size) + 1;
1387 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1388 old_size, old_order, new_size, new_order);
1475579c 1389 assert(new_size > old_size);
4105056a 1390 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1391}
1392
1393/* called with resize mutex held */
1394static
4105056a 1395void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1396 unsigned long old_size, unsigned long new_size)
464a1ec9 1397{
1475579c 1398 unsigned long old_order, new_order;
464a1ec9 1399
cd95516d 1400 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1401 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1402 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1403 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1404 old_size, old_order, new_size, new_order);
1475579c 1405 assert(new_size < old_size);
1475579c 1406
4105056a
MD
1407 /* Remove and unlink all dummy nodes to remove. */
1408 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1409}
1410
1475579c
MD
1411
1412/* called with resize mutex held */
1413static
1414void _do_cds_lfht_resize(struct cds_lfht *ht)
1415{
1416 unsigned long new_size, old_size;
4105056a
MD
1417
1418 /*
1419 * Resize table, re-do if the target size has changed under us.
1420 */
1421 do {
1422 ht->t.resize_initiated = 1;
1423 old_size = ht->t.size;
1424 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1425 if (old_size < new_size)
1426 _do_cds_lfht_grow(ht, old_size, new_size);
1427 else if (old_size > new_size)
1428 _do_cds_lfht_shrink(ht, old_size, new_size);
1429 ht->t.resize_initiated = 0;
1430 /* write resize_initiated before read resize_target */
1431 cmm_smp_mb();
4d676753 1432 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1433}
1434
abc490a1 1435static
4105056a 1436unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1437 int growth_order)
464a1ec9 1438{
4105056a
MD
1439 return _uatomic_max(&ht->t.resize_target,
1440 size << growth_order);
464a1ec9
MD
1441}
1442
1475579c 1443static
4105056a 1444void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1445 unsigned long count)
1475579c 1446{
cd95516d 1447 count = max(count, MIN_TABLE_SIZE);
4105056a 1448 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1449}
1450
1451void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1452{
4105056a
MD
1453 resize_target_update_count(ht, new_size);
1454 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1455 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1456 pthread_mutex_lock(&ht->resize_mutex);
1457 _do_cds_lfht_resize(ht);
1458 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1459 ht->cds_lfht_rcu_thread_online();
abc490a1 1460}
464a1ec9 1461
abc490a1
MD
1462static
1463void do_resize_cb(struct rcu_head *head)
1464{
1465 struct rcu_resize_work *work =
1466 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1467 struct cds_lfht *ht = work->ht;
abc490a1 1468
5f511391 1469 ht->cds_lfht_rcu_thread_offline();
abc490a1 1470 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1471 _do_cds_lfht_resize(ht);
abc490a1 1472 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1473 ht->cds_lfht_rcu_thread_online();
98808fb1 1474 poison_free(work);
848d4088
MD
1475 cmm_smp_mb(); /* finish resize before decrement */
1476 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1477}
1478
abc490a1 1479static
4105056a 1480void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1481{
abc490a1 1482 struct rcu_resize_work *work;
f9830efd 1483 unsigned long target_size;
abc490a1 1484
4105056a
MD
1485 target_size = resize_target_update(ht, size, growth);
1486 /* Store resize_target before read resize_initiated */
1487 cmm_smp_mb();
1488 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088
MD
1489 uatomic_inc(&ht->in_progress_resize);
1490 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1491 work = malloc(sizeof(*work));
1492 work->ht = ht;
14044b37 1493 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1494 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1495 }
ab7d5fc6 1496}
3171717f 1497
f8994aee
MD
1498#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1499
3171717f 1500static
4105056a 1501void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1502 unsigned long count)
1503{
1504 struct rcu_resize_work *work;
3171717f 1505
b8af5011
MD
1506 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1507 return;
4105056a
MD
1508 resize_target_update_count(ht, count);
1509 /* Store resize_target before read resize_initiated */
1510 cmm_smp_mb();
1511 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f
MD
1512 uatomic_inc(&ht->in_progress_resize);
1513 cmm_smp_mb(); /* increment resize count before calling it */
1514 work = malloc(sizeof(*work));
1515 work->ht = ht;
1516 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1517 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1518 }
1519}
f8994aee
MD
1520
1521#endif
This page took 0.167433 seconds and 4 git commands to generate.