rculfhash: API rename: remove -> del
[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.
196 * The dummy flag does not require to be updated atomically with the
197 * pointer, but it is added as a pointer low bit flag to save space.
198 */
d37166c6 199#define REMOVED_FLAG (1UL << 0)
f5596c94
MD
200#define DUMMY_FLAG (1UL << 1)
201#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 202
bb7b2f26 203/* Value of the end pointer. Should not interact with flags. */
f9c80341 204#define END_VALUE NULL
bb7b2f26 205
df44348d 206struct ht_items_count {
860d07e8 207 unsigned long add, del;
df44348d
MD
208} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
209
1475579c
MD
210struct rcu_level {
211 struct rcu_head head;
212 struct _cds_lfht_node nodes[0];
213};
214
395270b6 215struct rcu_table {
4105056a 216 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 217 unsigned long resize_target;
11519af6 218 int resize_initiated;
4105056a 219 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
220};
221
14044b37 222struct cds_lfht {
4105056a 223 struct rcu_table t;
14044b37
MD
224 cds_lfht_hash_fct hash_fct;
225 cds_lfht_compare_fct compare_fct;
732ad076 226 unsigned long hash_seed;
b8af5011 227 int flags;
5f511391
MD
228 /*
229 * We need to put the work threads offline (QSBR) when taking this
230 * mutex, because we use synchronize_rcu within this mutex critical
231 * section, which waits on read-side critical sections, and could
232 * therefore cause grace-period deadlock if we hold off RCU G.P.
233 * completion.
234 */
464a1ec9 235 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 236 unsigned int in_progress_resize, in_progress_destroy;
14044b37 237 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 238 void (*func)(struct rcu_head *head));
1475579c 239 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
240 void (*cds_lfht_rcu_read_lock)(void);
241 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
242 void (*cds_lfht_rcu_thread_offline)(void);
243 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
244 void (*cds_lfht_rcu_register_thread)(void);
245 void (*cds_lfht_rcu_unregister_thread)(void);
246 pthread_attr_t *resize_attr; /* Resize threads attributes */
df44348d
MD
247 unsigned long count; /* global approximate item count */
248 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
249};
250
abc490a1
MD
251struct rcu_resize_work {
252 struct rcu_head head;
14044b37 253 struct cds_lfht *ht;
abc490a1 254};
2ed95849 255
b7d619b0
MD
256struct partition_resize_work {
257 struct rcu_head head;
258 struct cds_lfht *ht;
259 unsigned long i, start, len;
260 void (*fct)(struct cds_lfht *ht, unsigned long i,
261 unsigned long start, unsigned long len);
262};
263
76a73da8
MD
264static
265struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
266 unsigned long size,
267 struct cds_lfht_node *node,
268 int unique, int dummy);
269
abc490a1
MD
270/*
271 * Algorithm to reverse bits in a word by lookup table, extended to
272 * 64-bit words.
f9830efd 273 * Source:
abc490a1 274 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 275 * Originally from Public Domain.
abc490a1
MD
276 */
277
278static const uint8_t BitReverseTable256[256] =
2ed95849 279{
abc490a1
MD
280#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
281#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
282#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
283 R6(0), R6(2), R6(1), R6(3)
284};
285#undef R2
286#undef R4
287#undef R6
2ed95849 288
abc490a1
MD
289static
290uint8_t bit_reverse_u8(uint8_t v)
291{
292 return BitReverseTable256[v];
293}
ab7d5fc6 294
abc490a1
MD
295static __attribute__((unused))
296uint32_t bit_reverse_u32(uint32_t v)
297{
298 return ((uint32_t) bit_reverse_u8(v) << 24) |
299 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
300 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
301 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
302}
303
abc490a1
MD
304static __attribute__((unused))
305uint64_t bit_reverse_u64(uint64_t v)
2ed95849 306{
abc490a1
MD
307 return ((uint64_t) bit_reverse_u8(v) << 56) |
308 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
309 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
310 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
311 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
312 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
313 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
314 ((uint64_t) bit_reverse_u8(v >> 56));
315}
316
317static
318unsigned long bit_reverse_ulong(unsigned long v)
319{
320#if (CAA_BITS_PER_LONG == 32)
321 return bit_reverse_u32(v);
322#else
323 return bit_reverse_u64(v);
324#endif
325}
326
f9830efd 327/*
24365af7
MD
328 * fls: returns the position of the most significant bit.
329 * Returns 0 if no bit is set, else returns the position of the most
330 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 331 */
24365af7
MD
332#if defined(__i386) || defined(__x86_64)
333static inline
334unsigned int fls_u32(uint32_t x)
f9830efd 335{
24365af7
MD
336 int r;
337
338 asm("bsrl %1,%0\n\t"
339 "jnz 1f\n\t"
340 "movl $-1,%0\n\t"
341 "1:\n\t"
342 : "=r" (r) : "rm" (x));
343 return r + 1;
344}
345#define HAS_FLS_U32
346#endif
347
348#if defined(__x86_64)
349static inline
350unsigned int fls_u64(uint64_t x)
351{
352 long r;
353
354 asm("bsrq %1,%0\n\t"
355 "jnz 1f\n\t"
356 "movq $-1,%0\n\t"
357 "1:\n\t"
358 : "=r" (r) : "rm" (x));
359 return r + 1;
360}
361#define HAS_FLS_U64
362#endif
363
364#ifndef HAS_FLS_U64
365static __attribute__((unused))
366unsigned int fls_u64(uint64_t x)
367{
368 unsigned int r = 64;
369
370 if (!x)
371 return 0;
372
373 if (!(x & 0xFFFFFFFF00000000ULL)) {
374 x <<= 32;
375 r -= 32;
376 }
377 if (!(x & 0xFFFF000000000000ULL)) {
378 x <<= 16;
379 r -= 16;
380 }
381 if (!(x & 0xFF00000000000000ULL)) {
382 x <<= 8;
383 r -= 8;
384 }
385 if (!(x & 0xF000000000000000ULL)) {
386 x <<= 4;
387 r -= 4;
388 }
389 if (!(x & 0xC000000000000000ULL)) {
390 x <<= 2;
391 r -= 2;
392 }
393 if (!(x & 0x8000000000000000ULL)) {
394 x <<= 1;
395 r -= 1;
396 }
397 return r;
398}
399#endif
400
401#ifndef HAS_FLS_U32
402static __attribute__((unused))
403unsigned int fls_u32(uint32_t x)
404{
405 unsigned int r = 32;
f9830efd 406
24365af7
MD
407 if (!x)
408 return 0;
409 if (!(x & 0xFFFF0000U)) {
410 x <<= 16;
411 r -= 16;
412 }
413 if (!(x & 0xFF000000U)) {
414 x <<= 8;
415 r -= 8;
416 }
417 if (!(x & 0xF0000000U)) {
418 x <<= 4;
419 r -= 4;
420 }
421 if (!(x & 0xC0000000U)) {
422 x <<= 2;
423 r -= 2;
424 }
425 if (!(x & 0x80000000U)) {
426 x <<= 1;
427 r -= 1;
428 }
429 return r;
430}
431#endif
432
433unsigned int fls_ulong(unsigned long x)
f9830efd 434{
24365af7
MD
435#if (CAA_BITS_PER_lONG == 32)
436 return fls_u32(x);
437#else
438 return fls_u64(x);
439#endif
440}
f9830efd 441
24365af7
MD
442int get_count_order_u32(uint32_t x)
443{
444 int order;
445
446 order = fls_u32(x) - 1;
447 if (x & (x - 1))
448 order++;
449 return order;
450}
451
452int get_count_order_ulong(unsigned long x)
453{
454 int order;
455
456 order = fls_ulong(x) - 1;
457 if (x & (x - 1))
458 order++;
459 return order;
f9830efd
MD
460}
461
98808fb1
MD
462#ifdef POISON_FREE
463#define poison_free(ptr) \
464 do { \
465 memset(ptr, 0x42, sizeof(*(ptr))); \
466 free(ptr); \
467 } while (0)
468#else
469#define poison_free(ptr) free(ptr)
470#endif
471
f9830efd 472static
4105056a 473void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 474
df44348d
MD
475/*
476 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
477 * available, then we support hash table item accounting.
478 * In the unfortunate event the number of CPUs reported would be
479 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
480 */
df44348d
MD
481#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
482
f8994aee 483static
4105056a 484void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
485 unsigned long count);
486
df44348d
MD
487static long nr_cpus_mask = -1;
488
489static
490struct ht_items_count *alloc_per_cpu_items_count(void)
491{
492 struct ht_items_count *count;
493
494 switch (nr_cpus_mask) {
495 case -2:
496 return NULL;
497 case -1:
498 {
499 long maxcpus;
500
501 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
502 if (maxcpus <= 0) {
503 nr_cpus_mask = -2;
504 return NULL;
505 }
506 /*
507 * round up number of CPUs to next power of two, so we
508 * can use & for modulo.
509 */
510 maxcpus = 1UL << get_count_order_ulong(maxcpus);
511 nr_cpus_mask = maxcpus - 1;
512 }
513 /* Fall-through */
514 default:
515 return calloc(nr_cpus_mask + 1, sizeof(*count));
516 }
517}
518
519static
520void free_per_cpu_items_count(struct ht_items_count *count)
521{
98808fb1 522 poison_free(count);
df44348d
MD
523}
524
525static
526int ht_get_cpu(void)
527{
528 int cpu;
529
530 assert(nr_cpus_mask >= 0);
531 cpu = sched_getcpu();
532 if (unlikely(cpu < 0))
533 return cpu;
534 else
535 return cpu & nr_cpus_mask;
536}
537
538static
4105056a 539void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 540{
3171717f 541 unsigned long percpu_count;
df44348d
MD
542 int cpu;
543
544 if (unlikely(!ht->percpu_count))
3171717f 545 return;
df44348d
MD
546 cpu = ht_get_cpu();
547 if (unlikely(cpu < 0))
3171717f
MD
548 return;
549 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d
MD
550 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
551 unsigned long count;
552
553 dbg_printf("add percpu %lu\n", percpu_count);
554 count = uatomic_add_return(&ht->count,
555 1UL << COUNT_COMMIT_ORDER);
556 /* If power of 2 */
557 if (!(count & (count - 1))) {
4105056a 558 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee
MD
559 return;
560 dbg_printf("add set global %lu\n", count);
4105056a 561 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 562 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
563 }
564 }
565}
566
567static
860d07e8 568void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
569{
570 unsigned long percpu_count;
3171717f 571 int cpu;
df44348d 572
3171717f
MD
573 if (unlikely(!ht->percpu_count))
574 return;
575 cpu = ht_get_cpu();
576 if (unlikely(cpu < 0))
577 return;
860d07e8 578 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, -1);
df44348d
MD
579 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
580 unsigned long count;
581
860d07e8 582 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 583 count = uatomic_add_return(&ht->count,
3171717f 584 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
585 /* If power of 2 */
586 if (!(count & (count - 1))) {
4105056a 587 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 588 return;
860d07e8 589 dbg_printf("del set global %lu\n", count);
4105056a 590 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 591 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
592 }
593 }
594}
595
596#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
597
598static const long nr_cpus_mask = -1;
599
600static
601struct ht_items_count *alloc_per_cpu_items_count(void)
602{
603 return NULL;
604}
605
606static
607void free_per_cpu_items_count(struct ht_items_count *count)
608{
609}
610
611static
4105056a 612void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
613{
614}
615
616static
860d07e8 617void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
618{
619}
620
621#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
622
623
f9830efd 624static
4105056a 625void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 626{
f8994aee
MD
627 unsigned long count;
628
b8af5011
MD
629 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
630 return;
f8994aee
MD
631 count = uatomic_read(&ht->count);
632 /*
633 * Use bucket-local length for small table expand and for
634 * environments lacking per-cpu data support.
635 */
636 if (count >= (1UL << COUNT_COMMIT_ORDER))
637 return;
24365af7 638 if (chain_len > 100)
f0c29ed7 639 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 640 chain_len);
3390d470 641 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 642 cds_lfht_resize_lazy(ht, size,
01370f0b 643 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
644}
645
abc490a1 646static
14044b37 647struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 648{
14044b37 649 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
650}
651
652static
14044b37 653int is_removed(struct cds_lfht_node *node)
abc490a1 654{
d37166c6 655 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
656}
657
658static
14044b37 659struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 660{
14044b37 661 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
662}
663
f5596c94 664static
14044b37 665int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
666{
667 return ((unsigned long) node) & DUMMY_FLAG;
668}
669
670static
14044b37 671struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 672{
14044b37 673 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 674}
bb7b2f26
MD
675
676static
677struct cds_lfht_node *get_end(void)
678{
679 return (struct cds_lfht_node *) END_VALUE;
680}
681
682static
683int is_end(struct cds_lfht_node *node)
684{
685 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
686}
687
abc490a1 688static
f9830efd 689unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
690{
691 unsigned long old1, old2;
692
693 old1 = uatomic_read(ptr);
694 do {
695 old2 = old1;
696 if (old2 >= v)
f9830efd 697 return old2;
abc490a1 698 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 699 return v;
abc490a1
MD
700}
701
1475579c
MD
702static
703void cds_lfht_free_level(struct rcu_head *head)
704{
705 struct rcu_level *l =
706 caa_container_of(head, struct rcu_level, head);
98808fb1 707 poison_free(l);
1475579c
MD
708}
709
273399de
MD
710/*
711 * Remove all logically deleted nodes from a bucket up to a certain node key.
712 */
713static
f9c80341 714void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 715{
14044b37 716 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 717
c90201ac
MD
718 assert(!is_dummy(dummy));
719 assert(!is_removed(dummy));
720 assert(!is_dummy(node));
721 assert(!is_removed(node));
273399de
MD
722 for (;;) {
723 iter_prev = dummy;
724 /* We can always skip the dummy node initially */
cc4fcb10
MD
725 iter = rcu_dereference(iter_prev->p.next);
726 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
727 /*
728 * We should never be called with dummy (start of chain)
729 * and logically removed node (end of path compression
730 * marker) being the actual same node. This would be a
731 * bug in the algorithm implementation.
732 */
733 assert(dummy != node);
273399de 734 for (;;) {
bb7b2f26 735 if (unlikely(is_end(iter)))
f9c80341 736 return;
76412f24 737 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 738 return;
cc4fcb10 739 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 740 if (likely(is_removed(next)))
273399de 741 break;
b453eae1 742 iter_prev = clear_flag(iter);
273399de
MD
743 iter = next;
744 }
745 assert(!is_removed(iter));
f5596c94
MD
746 if (is_dummy(iter))
747 new_next = flag_dummy(clear_flag(next));
748 else
749 new_next = clear_flag(next);
750 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 751 }
f9c80341 752 return;
273399de
MD
753}
754
abc490a1 755static
4105056a
MD
756struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
757 unsigned long size,
758 struct cds_lfht_node *node,
759 int unique, int dummy)
abc490a1 760{
14044b37 761 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
f5596c94 762 *dummy_node;
14044b37 763 struct _cds_lfht_node *lookup;
24365af7 764 unsigned long hash, index, order;
abc490a1 765
c90201ac
MD
766 assert(!is_dummy(node));
767 assert(!is_removed(node));
4105056a 768 if (!size) {
f5596c94 769 assert(dummy);
bb7b2f26 770 node->p.next = flag_dummy(get_end());
18117871
MD
771 return node; /* Initial first add (head) */
772 }
cc4fcb10 773 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 774 for (;;) {
f9830efd 775 uint32_t chain_len = 0;
abc490a1 776
11519af6
MD
777 /*
778 * iter_prev points to the non-removed node prior to the
779 * insert location.
11519af6 780 */
4105056a 781 index = hash & (size - 1);
24365af7 782 order = get_count_order_ulong(index + 1);
4105056a 783 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 784 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 785 /* We can always skip the dummy node initially */
cc4fcb10
MD
786 iter = rcu_dereference(iter_prev->p.next);
787 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 788 for (;;) {
bb7b2f26 789 if (unlikely(is_end(iter)))
273399de 790 goto insert;
76412f24 791 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 792 goto insert;
cc4fcb10 793 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 794 if (unlikely(is_removed(next)))
9dba85be 795 goto gc_node;
e43f23f8 796 if (unique
1b81fe1a 797 && !is_dummy(next)
e43f23f8
MD
798 && !ht->compare_fct(node->key, node->key_len,
799 clear_flag(iter)->key,
800 clear_flag(iter)->key_len))
18117871 801 return clear_flag(iter);
11519af6 802 /* Only account for identical reverse hash once */
24365af7
MD
803 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
804 && !is_dummy(next))
4105056a 805 check_resize(ht, size, ++chain_len);
11519af6 806 iter_prev = clear_flag(iter);
273399de 807 iter = next;
abc490a1 808 }
273399de 809 insert:
7ec59d3b 810 assert(node != clear_flag(iter));
11519af6 811 assert(!is_removed(iter_prev));
c90201ac 812 assert(!is_removed(iter));
f000907d 813 assert(iter_prev != node);
f9c80341 814 if (!dummy)
1b81fe1a 815 node->p.next = clear_flag(iter);
f9c80341
MD
816 else
817 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
818 if (is_dummy(iter))
819 new_node = flag_dummy(node);
820 else
821 new_node = node;
cc4fcb10 822 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
f5596c94 823 new_node) != iter)
273399de 824 continue; /* retry */
11519af6 825 else
273399de 826 goto gc_end;
9dba85be
MD
827 gc_node:
828 assert(!is_removed(iter));
f5596c94
MD
829 if (is_dummy(iter))
830 new_next = flag_dummy(clear_flag(next));
831 else
832 new_next = clear_flag(next);
833 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 834 /* retry */
464a1ec9 835 }
273399de
MD
836gc_end:
837 /* Garbage collect logically removed nodes in the bucket */
4105056a 838 index = hash & (size - 1);
24365af7 839 order = get_count_order_ulong(index + 1);
4105056a 840 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 841 dummy_node = (struct cds_lfht_node *) lookup;
f9c80341 842 _cds_lfht_gc_bucket(dummy_node, node);
18117871 843 return node;
abc490a1 844}
464a1ec9 845
abc490a1 846static
860d07e8 847int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a
MD
848 struct cds_lfht_node *node,
849 int dummy_removal)
abc490a1 850{
14044b37
MD
851 struct cds_lfht_node *dummy, *next, *old;
852 struct _cds_lfht_node *lookup;
abc490a1 853 int flagged = 0;
24365af7 854 unsigned long hash, index, order;
5e28c532 855
7ec59d3b 856 /* logically delete the node */
c90201ac
MD
857 assert(!is_dummy(node));
858 assert(!is_removed(node));
cc4fcb10 859 old = rcu_dereference(node->p.next);
7ec59d3b
MD
860 do {
861 next = old;
76412f24 862 if (unlikely(is_removed(next)))
7ec59d3b 863 goto end;
1475579c
MD
864 if (dummy_removal)
865 assert(is_dummy(next));
866 else
867 assert(!is_dummy(next));
cc4fcb10 868 old = uatomic_cmpxchg(&node->p.next, next,
7ec59d3b
MD
869 flag_removed(next));
870 } while (old != next);
871
872 /* We performed the (logical) deletion. */
873 flagged = 1;
874
875 /*
876 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
877 * the node, and remove it (along with any other logically removed node)
878 * if found.
11519af6 879 */
cc4fcb10 880 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
881 assert(size > 0);
882 index = hash & (size - 1);
24365af7 883 order = get_count_order_ulong(index + 1);
4105056a 884 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 885 dummy = (struct cds_lfht_node *) lookup;
f9c80341 886 _cds_lfht_gc_bucket(dummy, node);
2ed95849 887end:
11519af6
MD
888 /*
889 * Only the flagging action indicated that we (and no other)
890 * removed the node from the hash.
891 */
7ec59d3b 892 if (flagged) {
cc4fcb10 893 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 894 return 0;
7ec59d3b 895 } else
11519af6 896 return -ENOENT;
abc490a1 897}
2ed95849 898
b7d619b0
MD
899static
900void *partition_resize_thread(void *arg)
901{
902 struct partition_resize_work *work = arg;
903
904 work->ht->cds_lfht_rcu_register_thread();
905 work->fct(work->ht, work->i, work->start, work->len);
906 work->ht->cds_lfht_rcu_unregister_thread();
907 return NULL;
908}
909
910static
911void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
912 unsigned long len,
913 void (*fct)(struct cds_lfht *ht, unsigned long i,
914 unsigned long start, unsigned long len))
915{
916 unsigned long partition_len;
917 struct partition_resize_work *work;
6083a889
MD
918 int thread, ret;
919 unsigned long nr_threads;
b7d619b0
MD
920 pthread_t *thread_id;
921
6083a889
MD
922 /*
923 * Note: nr_cpus_mask + 1 is always power of 2.
924 * We spawn just the number of threads we need to satisfy the minimum
925 * partition size, up to the number of CPUs in the system.
926 */
927 nr_threads = min(nr_cpus_mask + 1,
928 len >> MIN_PARTITION_PER_THREAD_ORDER);
929 partition_len = len >> get_count_order_ulong(nr_threads);
930 work = calloc(nr_threads, sizeof(*work));
931 thread_id = calloc(nr_threads, sizeof(*thread_id));
b7d619b0 932 assert(work);
6083a889
MD
933 for (thread = 0; thread < nr_threads; thread++) {
934 work[thread].ht = ht;
935 work[thread].i = i;
936 work[thread].len = partition_len;
937 work[thread].start = thread * partition_len;
938 work[thread].fct = fct;
939 ret = pthread_create(&thread_id[thread], ht->resize_attr,
940 partition_resize_thread, &work[thread]);
b7d619b0
MD
941 assert(!ret);
942 }
6083a889
MD
943 for (thread = 0; thread < nr_threads; thread++) {
944 ret = pthread_join(thread_id[thread], NULL);
b7d619b0
MD
945 assert(!ret);
946 }
947 free(work);
948 free(thread_id);
949}
950
e8de508e
MD
951/*
952 * Holding RCU read lock to protect _cds_lfht_add against memory
953 * reclaim that could be performed by other call_rcu worker threads (ABA
954 * problem).
9ee0fc9a 955 *
b7d619b0 956 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
957 * many worker threads, based on the number of CPUs available in the system.
958 * This should therefore take care of not having the expand lagging behind too
959 * many concurrent insertion threads by using the scheduler's ability to
960 * schedule dummy node population fairly with insertions.
e8de508e 961 */
4105056a 962static
b7d619b0
MD
963void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
964 unsigned long start, unsigned long len)
4105056a
MD
965{
966 unsigned long j;
967
968 ht->cds_lfht_rcu_read_lock();
b7d619b0 969 for (j = start; j < start + len; j++) {
4105056a
MD
970 struct cds_lfht_node *new_node =
971 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
972
dc1da8f6 973 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 974 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
975 new_node->p.reverse_hash =
976 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a
MD
977 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
978 new_node, 0, 1);
979 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
980 break;
981 }
982 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
983}
984
985static
986void init_table_populate(struct cds_lfht *ht, unsigned long i,
987 unsigned long len)
988{
989 assert(nr_cpus_mask != -1);
6083a889 990 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
991 ht->cds_lfht_rcu_thread_online();
992 init_table_populate_partition(ht, i, 0, len);
993 ht->cds_lfht_rcu_thread_offline();
994 return;
995 }
996 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
997}
998
abc490a1 999static
4105056a 1000void init_table(struct cds_lfht *ht,
24365af7
MD
1001 unsigned long first_order, unsigned long len_order)
1002{
1003 unsigned long i, end_order;
1004
f0c29ed7 1005 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1006 first_order, first_order + len_order);
1007 end_order = first_order + len_order;
24365af7 1008 for (i = first_order; i < end_order; i++) {
4105056a 1009 unsigned long len;
24365af7
MD
1010
1011 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1012 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1013
1014 /* Stop expand if the resize target changes under us */
1015 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1016 break;
1017
4105056a 1018 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1019 + (len * sizeof(struct _cds_lfht_node)));
b7d619b0 1020 assert(ht->t.tbl[i]);
4105056a 1021
4105056a 1022 /*
dc1da8f6
MD
1023 * Set all dummy nodes reverse hash values for a level and
1024 * link all dummy nodes into the table.
4105056a 1025 */
dc1da8f6 1026 init_table_populate(ht, i, len);
4105056a 1027
f9c80341
MD
1028 /*
1029 * Update table size.
1030 */
1031 cmm_smp_wmb(); /* populate data before RCU size */
1032 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1033
4105056a
MD
1034 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1035 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1036 break;
1037 }
1038}
1039
e8de508e
MD
1040/*
1041 * Holding RCU read lock to protect _cds_lfht_remove against memory
1042 * reclaim that could be performed by other call_rcu worker threads (ABA
1043 * problem).
1044 * For a single level, we logically remove and garbage collect each node.
1045 *
1046 * As a design choice, we perform logical removal and garbage collection on a
1047 * node-per-node basis to simplify this algorithm. We also assume keeping good
1048 * cache locality of the operation would overweight possible performance gain
1049 * that could be achieved by batching garbage collection for multiple levels.
1050 * However, this would have to be justified by benchmarks.
1051 *
1052 * Concurrent removal and add operations are helping us perform garbage
1053 * collection of logically removed nodes. We guarantee that all logically
1054 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1055 * invoked to free a hole level of dummy nodes (after a grace period).
1056 *
1057 * Logical removal and garbage collection can therefore be done in batch or on a
1058 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1059 *
b7d619b0
MD
1060 * When we reach a certain length, we can split this removal over many worker
1061 * threads, based on the number of CPUs available in the system. This should
1062 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1063 * updater threads actively inserting into the hash table.
e8de508e 1064 */
4105056a 1065static
b7d619b0
MD
1066void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1067 unsigned long start, unsigned long len)
4105056a
MD
1068{
1069 unsigned long j;
1070
1071 ht->cds_lfht_rcu_read_lock();
b7d619b0 1072 for (j = start; j < start + len; j++) {
4105056a
MD
1073 struct cds_lfht_node *fini_node =
1074 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1075
1076 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1077 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1078 fini_node->p.reverse_hash =
1079 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1080 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
4105056a 1081 fini_node, 1);
33c7c748
MD
1082 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1083 break;
abc490a1 1084 }
4105056a 1085 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1086}
1087
1088static
1089void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1090{
1091
1092 assert(nr_cpus_mask != -1);
6083a889 1093 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1094 ht->cds_lfht_rcu_thread_online();
1095 remove_table_partition(ht, i, 0, len);
1096 ht->cds_lfht_rcu_thread_offline();
1097 return;
1098 }
1099 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1100}
1101
1475579c 1102static
4105056a 1103void fini_table(struct cds_lfht *ht,
1475579c
MD
1104 unsigned long first_order, unsigned long len_order)
1105{
1106 long i, end_order;
1107
1108 dbg_printf("fini table: first_order %lu end_order %lu\n",
1109 first_order, first_order + len_order);
1110 end_order = first_order + len_order;
1111 assert(first_order > 0);
1475579c 1112 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1113 unsigned long len;
1475579c
MD
1114
1115 len = !i ? 1 : 1UL << (i - 1);
1116 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1117
4d676753
MD
1118 /* Stop shrink if the resize target changes under us */
1119 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1120 break;
1121
1122 cmm_smp_wmb(); /* populate data before RCU size */
1123 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1124
1125 /*
1126 * We need to wait for all add operations to reach Q.S. (and
1127 * thus use the new table for lookups) before we can start
1128 * releasing the old dummy nodes. Otherwise their lookup will
1129 * return a logically removed node as insert position.
1130 */
1131 ht->cds_lfht_synchronize_rcu();
1132
21263e21 1133 /*
4105056a
MD
1134 * Set "removed" flag in dummy nodes about to be removed.
1135 * Unlink all now-logically-removed dummy node pointers.
1136 * Concurrent add/remove operation are helping us doing
1137 * the gc.
21263e21 1138 */
4105056a
MD
1139 remove_table(ht, i, len);
1140
1141 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1142
1143 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1144 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1145 break;
1146 }
1475579c
MD
1147}
1148
7a9dcf9b 1149struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1150 cds_lfht_compare_fct compare_fct,
1151 unsigned long hash_seed,
1152 unsigned long init_size,
b8af5011 1153 int flags,
14044b37 1154 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1155 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1156 void (*cds_lfht_synchronize_rcu)(void),
1157 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1158 void (*cds_lfht_rcu_read_unlock)(void),
1159 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1160 void (*cds_lfht_rcu_thread_online)(void),
1161 void (*cds_lfht_rcu_register_thread)(void),
1162 void (*cds_lfht_rcu_unregister_thread)(void),
1163 pthread_attr_t *attr)
abc490a1 1164{
14044b37 1165 struct cds_lfht *ht;
24365af7 1166 unsigned long order;
abc490a1 1167
8129be4e 1168 /* init_size must be power of two */
49619ea0 1169 if (init_size && (init_size & (init_size - 1)))
8129be4e 1170 return NULL;
14044b37 1171 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1172 assert(ht);
abc490a1 1173 ht->hash_fct = hash_fct;
732ad076
MD
1174 ht->compare_fct = compare_fct;
1175 ht->hash_seed = hash_seed;
14044b37 1176 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1177 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1178 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1179 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1180 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1181 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1182 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1183 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1184 ht->resize_attr = attr;
df44348d 1185 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1186 /* this mutex should not nest in read-side C.S. */
1187 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1188 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1189 ht->flags = flags;
5f511391 1190 ht->cds_lfht_rcu_thread_offline();
f000907d 1191 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1192 ht->t.resize_target = 1UL << (order - 1);
4105056a 1193 init_table(ht, 0, order);
f000907d 1194 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1195 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1196 return ht;
1197}
1198
14044b37 1199struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
2ed95849 1200{
bb7b2f26 1201 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1202 struct _cds_lfht_node *lookup;
4105056a 1203 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1204
732ad076 1205 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1206 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1207
4105056a
MD
1208 size = rcu_dereference(ht->t.size);
1209 index = hash & (size - 1);
24365af7 1210 order = get_count_order_ulong(index + 1);
4105056a 1211 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1212 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1213 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1214 dummy_node = (struct cds_lfht_node *) lookup;
1215 /* We can always skip the dummy node initially */
1216 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1217 node = clear_flag(node);
2ed95849 1218 for (;;) {
bb7b2f26
MD
1219 if (unlikely(is_end(node))) {
1220 node = NULL;
abc490a1 1221 break;
bb7b2f26 1222 }
cc4fcb10 1223 if (unlikely(node->p.reverse_hash > reverse_hash)) {
abc490a1
MD
1224 node = NULL;
1225 break;
2ed95849 1226 }
1b81fe1a
MD
1227 next = rcu_dereference(node->p.next);
1228 if (likely(!is_removed(next))
1229 && !is_dummy(next)
49c2e2d6 1230 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1231 break;
2ed95849 1232 }
1b81fe1a 1233 node = clear_flag(next);
2ed95849 1234 }
1b81fe1a 1235 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
abc490a1
MD
1236 return node;
1237}
e0ba718a 1238
a481e5ff
MD
1239struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
1240 struct cds_lfht_node *node)
1241{
1242 struct cds_lfht_node *next;
1243 unsigned long reverse_hash;
1244 void *key;
1245 size_t key_len;
1246
1247 reverse_hash = node->p.reverse_hash;
1248 key = node->key;
1249 key_len = node->key_len;
1250 next = rcu_dereference(node->p.next);
1251 node = clear_flag(next);
1252
1253 for (;;) {
bb7b2f26
MD
1254 if (unlikely(is_end(node))) {
1255 node = NULL;
a481e5ff 1256 break;
bb7b2f26 1257 }
a481e5ff
MD
1258 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1259 node = NULL;
1260 break;
1261 }
1262 next = rcu_dereference(node->p.next);
1263 if (likely(!is_removed(next))
1264 && !is_dummy(next)
1265 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1266 break;
1267 }
1268 node = clear_flag(next);
1269 }
1270 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1271 return node;
1272}
1273
14044b37 1274void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1275{
4105056a 1276 unsigned long hash, size;
ab7d5fc6 1277
49c2e2d6 1278 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1279 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1280
4105056a
MD
1281 size = rcu_dereference(ht->t.size);
1282 (void) _cds_lfht_add(ht, size, node, 0, 0);
1283 ht_count_add(ht, size);
3eca1b8c
MD
1284}
1285
14044b37
MD
1286struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
1287 struct cds_lfht_node *node)
3eca1b8c 1288{
4105056a 1289 unsigned long hash, size;
df44348d 1290 struct cds_lfht_node *ret;
3eca1b8c 1291
49c2e2d6 1292 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1293 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1294
4105056a
MD
1295 size = rcu_dereference(ht->t.size);
1296 ret = _cds_lfht_add(ht, size, node, 1, 0);
17f31d1b 1297 if (ret == node)
4105056a 1298 ht_count_add(ht, size);
df44348d 1299 return ret;
2ed95849
MD
1300}
1301
860d07e8 1302int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1303{
4105056a 1304 unsigned long size;
df44348d 1305 int ret;
abc490a1 1306
4105056a 1307 size = rcu_dereference(ht->t.size);
860d07e8 1308 ret = _cds_lfht_del(ht, size, node, 0);
df44348d 1309 if (!ret)
860d07e8 1310 ht_count_del(ht, size);
df44348d 1311 return ret;
2ed95849 1312}
ab7d5fc6 1313
abc490a1 1314static
14044b37 1315int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1316{
14044b37
MD
1317 struct cds_lfht_node *node;
1318 struct _cds_lfht_node *lookup;
4105056a 1319 unsigned long order, i, size;
674f7a69 1320
abc490a1 1321 /* Check that the table is empty */
4105056a 1322 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1323 node = (struct cds_lfht_node *) lookup;
abc490a1 1324 do {
1b81fe1a
MD
1325 node = clear_flag(node)->p.next;
1326 if (!is_dummy(node))
abc490a1 1327 return -EPERM;
273399de 1328 assert(!is_removed(node));
bb7b2f26 1329 } while (!is_end(node));
4105056a
MD
1330 /*
1331 * size accessed without rcu_dereference because hash table is
1332 * being destroyed.
1333 */
1334 size = ht->t.size;
abc490a1 1335 /* Internal sanity check: all nodes left should be dummy */
4105056a 1336 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1337 unsigned long len;
1338
1339 len = !order ? 1 : 1UL << (order - 1);
1340 for (i = 0; i < len; i++) {
f0c29ed7 1341 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1342 order, i,
4105056a
MD
1343 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1344 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1345 }
4105056a 1346 poison_free(ht->t.tbl[order]);
674f7a69 1347 }
abc490a1 1348 return 0;
674f7a69
MD
1349}
1350
1351/*
1352 * Should only be called when no more concurrent readers nor writers can
1353 * possibly access the table.
1354 */
b7d619b0 1355int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1356{
5e28c532
MD
1357 int ret;
1358
848d4088 1359 /* Wait for in-flight resize operations to complete */
33c7c748 1360 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
1361 while (uatomic_read(&ht->in_progress_resize))
1362 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1363 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1364 if (ret)
1365 return ret;
df44348d 1366 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1367 if (attr)
1368 *attr = ht->resize_attr;
98808fb1 1369 poison_free(ht);
5e28c532 1370 return ret;
674f7a69
MD
1371}
1372
14044b37 1373void cds_lfht_count_nodes(struct cds_lfht *ht,
273399de
MD
1374 unsigned long *count,
1375 unsigned long *removed)
1376{
14044b37
MD
1377 struct cds_lfht_node *node, *next;
1378 struct _cds_lfht_node *lookup;
24365af7 1379 unsigned long nr_dummy = 0;
273399de
MD
1380
1381 *count = 0;
1382 *removed = 0;
1383
24365af7 1384 /* Count non-dummy nodes in the table */
4105056a 1385 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1386 node = (struct cds_lfht_node *) lookup;
273399de 1387 do {
cc4fcb10 1388 next = rcu_dereference(node->p.next);
273399de 1389 if (is_removed(next)) {
1b81fe1a 1390 assert(!is_dummy(next));
273399de 1391 (*removed)++;
1b81fe1a 1392 } else if (!is_dummy(next))
273399de 1393 (*count)++;
24365af7
MD
1394 else
1395 (nr_dummy)++;
273399de 1396 node = clear_flag(next);
bb7b2f26 1397 } while (!is_end(node));
f0c29ed7 1398 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
273399de
MD
1399}
1400
1475579c 1401/* called with resize mutex held */
abc490a1 1402static
4105056a 1403void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1404 unsigned long old_size, unsigned long new_size)
abc490a1 1405{
1475579c 1406 unsigned long old_order, new_order;
1475579c
MD
1407
1408 old_order = get_count_order_ulong(old_size) + 1;
1409 new_order = get_count_order_ulong(new_size) + 1;
1410 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1411 old_size, old_order, new_size, new_order);
1475579c 1412 assert(new_size > old_size);
4105056a 1413 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1414}
1415
1416/* called with resize mutex held */
1417static
4105056a 1418void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1419 unsigned long old_size, unsigned long new_size)
464a1ec9 1420{
1475579c 1421 unsigned long old_order, new_order;
464a1ec9 1422
cd95516d 1423 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1424 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1425 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1426 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1427 old_size, old_order, new_size, new_order);
1475579c 1428 assert(new_size < old_size);
1475579c 1429
4105056a
MD
1430 /* Remove and unlink all dummy nodes to remove. */
1431 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1432}
1433
1475579c
MD
1434
1435/* called with resize mutex held */
1436static
1437void _do_cds_lfht_resize(struct cds_lfht *ht)
1438{
1439 unsigned long new_size, old_size;
4105056a
MD
1440
1441 /*
1442 * Resize table, re-do if the target size has changed under us.
1443 */
1444 do {
1445 ht->t.resize_initiated = 1;
1446 old_size = ht->t.size;
1447 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1448 if (old_size < new_size)
1449 _do_cds_lfht_grow(ht, old_size, new_size);
1450 else if (old_size > new_size)
1451 _do_cds_lfht_shrink(ht, old_size, new_size);
1452 ht->t.resize_initiated = 0;
1453 /* write resize_initiated before read resize_target */
1454 cmm_smp_mb();
4d676753 1455 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1456}
1457
abc490a1 1458static
4105056a 1459unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1460 int growth_order)
464a1ec9 1461{
4105056a
MD
1462 return _uatomic_max(&ht->t.resize_target,
1463 size << growth_order);
464a1ec9
MD
1464}
1465
1475579c 1466static
4105056a 1467void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1468 unsigned long count)
1475579c 1469{
cd95516d 1470 count = max(count, MIN_TABLE_SIZE);
4105056a 1471 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1472}
1473
1474void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1475{
4105056a
MD
1476 resize_target_update_count(ht, new_size);
1477 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1478 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1479 pthread_mutex_lock(&ht->resize_mutex);
1480 _do_cds_lfht_resize(ht);
1481 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1482 ht->cds_lfht_rcu_thread_online();
abc490a1 1483}
464a1ec9 1484
abc490a1
MD
1485static
1486void do_resize_cb(struct rcu_head *head)
1487{
1488 struct rcu_resize_work *work =
1489 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1490 struct cds_lfht *ht = work->ht;
abc490a1 1491
5f511391 1492 ht->cds_lfht_rcu_thread_offline();
abc490a1 1493 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1494 _do_cds_lfht_resize(ht);
abc490a1 1495 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1496 ht->cds_lfht_rcu_thread_online();
98808fb1 1497 poison_free(work);
848d4088
MD
1498 cmm_smp_mb(); /* finish resize before decrement */
1499 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1500}
1501
abc490a1 1502static
4105056a 1503void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1504{
abc490a1 1505 struct rcu_resize_work *work;
f9830efd 1506 unsigned long target_size;
abc490a1 1507
4105056a
MD
1508 target_size = resize_target_update(ht, size, growth);
1509 /* Store resize_target before read resize_initiated */
1510 cmm_smp_mb();
1511 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088
MD
1512 uatomic_inc(&ht->in_progress_resize);
1513 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1514 work = malloc(sizeof(*work));
1515 work->ht = ht;
14044b37 1516 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1517 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1518 }
ab7d5fc6 1519}
3171717f 1520
f8994aee
MD
1521#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1522
3171717f 1523static
4105056a 1524void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1525 unsigned long count)
1526{
1527 struct rcu_resize_work *work;
3171717f 1528
b8af5011
MD
1529 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1530 return;
4105056a
MD
1531 resize_target_update_count(ht, count);
1532 /* Store resize_target before read resize_initiated */
1533 cmm_smp_mb();
1534 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f
MD
1535 uatomic_inc(&ht->in_progress_resize);
1536 cmm_smp_mb(); /* increment resize count before calling it */
1537 work = malloc(sizeof(*work));
1538 work->ht = ht;
1539 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1540 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1541 }
1542}
f8994aee
MD
1543
1544#endif
This page took 0.170776 seconds and 4 git commands to generate.