rculfhash: support replacement operation
[urcu.git] / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
1475579c 4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
abc490a1
MD
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
5e28c532
MD
21 */
22
e753ff5a
MD
23/*
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
30 * (2002), 73-82.
31 *
1475579c 32 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
33 * implementation:
34 *
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
44 * duplicata exists.
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
1475579c
MD
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
53 * operation.
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
83 * list concurrently.
29e669f6
MD
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
88 * for it do to so.
e753ff5a
MD
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
94 * this order.
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
1475579c
MD
99 *
100 * A bit of ascii art explanation:
101 *
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
104 *
105 * This shows the nodes for a small table ordered by reversed bits:
106 *
107 * bits reverse
108 * 0 000 000
109 * 4 100 001
110 * 2 010 010
111 * 6 110 011
112 * 1 001 100
113 * 5 101 101
114 * 3 011 110
115 * 7 111 111
116 *
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
119 *
120 * order bits reverse
121 * 0 0 000 000
122 * |
f6fdd688
MD
123 * 1 | 1 001 100 <- <-
124 * | | | |
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
127 * | | | | | | |
1475579c
MD
128 * 3 -> | | | 4 100 001 | |
129 * -> | | 5 101 101 |
130 * -> | 6 110 011
131 * -> 7 111 111
e753ff5a
MD
132 */
133
2ed95849
MD
134#define _LGPL_SOURCE
135#include <stdlib.h>
e0ba718a
MD
136#include <errno.h>
137#include <assert.h>
138#include <stdio.h>
abc490a1 139#include <stdint.h>
f000907d 140#include <string.h>
e0ba718a 141
df44348d 142#include "config.h"
2ed95849 143#include <urcu.h>
abc490a1 144#include <urcu-call-rcu.h>
a42cc659
MD
145#include <urcu/arch.h>
146#include <urcu/uatomic.h>
674f7a69 147#include <urcu/jhash.h>
a42cc659 148#include <urcu/compiler.h>
abc490a1 149#include <urcu/rculfhash.h>
5e28c532 150#include <stdio.h>
464a1ec9 151#include <pthread.h>
44395fb7 152
f9830efd 153#ifdef DEBUG
f0c29ed7 154#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 155#else
e753ff5a 156#define dbg_printf(fmt, args...)
f9830efd
MD
157#endif
158
f8994aee
MD
159/*
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
164 */
165#define COUNT_COMMIT_ORDER 10
6ea6bc67
MD
166#define CHAIN_LEN_TARGET 1
167#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 168
cd95516d 169/*
76a73da8 170 * Define the minimum table size.
cd95516d 171 */
c9edd44a 172#define MIN_TABLE_SIZE 1
cd95516d 173
4105056a
MD
174#if (CAA_BITS_PER_LONG == 32)
175#define MAX_TABLE_ORDER 32
176#else
177#define MAX_TABLE_ORDER 64
178#endif
179
b7d619b0
MD
180/*
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 */
6083a889
MD
183#define MIN_PARTITION_PER_THREAD_ORDER 12
184#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 185
4105056a
MD
186#ifndef min
187#define min(a, b) ((a) < (b) ? (a) : (b))
188#endif
189
abc490a1
MD
190#ifndef max
191#define max(a, b) ((a) > (b) ? (a) : (b))
192#endif
2ed95849 193
d95bd160
MD
194/*
195 * The removed flag needs to be updated atomically with the pointer.
48ed1c18
MD
196 * It indicates that no node must attach to the node scheduled for
197 * removal. The gc flag also needs to be updated atomically with the
198 * pointer. It indicates that node garbage collection must be performed.
199 * "removed" and "gc" flags are separate for the benefit of replacement
200 * operation.
d95bd160
MD
201 * The dummy flag does not require to be updated atomically with the
202 * pointer, but it is added as a pointer low bit flag to save space.
203 */
d37166c6 204#define REMOVED_FLAG (1UL << 0)
48ed1c18
MD
205#define GC_FLAG (1UL << 1)
206#define DUMMY_FLAG (1UL << 2)
207#define FLAGS_MASK ((1UL << 3) - 1)
d37166c6 208
bb7b2f26 209/* Value of the end pointer. Should not interact with flags. */
f9c80341 210#define END_VALUE NULL
bb7b2f26 211
df44348d 212struct ht_items_count {
860d07e8 213 unsigned long add, del;
df44348d
MD
214} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
215
1475579c
MD
216struct rcu_level {
217 struct rcu_head head;
218 struct _cds_lfht_node nodes[0];
219};
220
395270b6 221struct rcu_table {
4105056a 222 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 223 unsigned long resize_target;
11519af6 224 int resize_initiated;
4105056a 225 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
226};
227
14044b37 228struct cds_lfht {
4105056a 229 struct rcu_table t;
14044b37
MD
230 cds_lfht_hash_fct hash_fct;
231 cds_lfht_compare_fct compare_fct;
732ad076 232 unsigned long hash_seed;
b8af5011 233 int flags;
5f511391
MD
234 /*
235 * We need to put the work threads offline (QSBR) when taking this
236 * mutex, because we use synchronize_rcu within this mutex critical
237 * section, which waits on read-side critical sections, and could
238 * therefore cause grace-period deadlock if we hold off RCU G.P.
239 * completion.
240 */
464a1ec9 241 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 242 unsigned int in_progress_resize, in_progress_destroy;
14044b37 243 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 244 void (*func)(struct rcu_head *head));
1475579c 245 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
246 void (*cds_lfht_rcu_read_lock)(void);
247 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
248 void (*cds_lfht_rcu_thread_offline)(void);
249 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
250 void (*cds_lfht_rcu_register_thread)(void);
251 void (*cds_lfht_rcu_unregister_thread)(void);
252 pthread_attr_t *resize_attr; /* Resize threads attributes */
df44348d
MD
253 unsigned long count; /* global approximate item count */
254 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
255};
256
abc490a1
MD
257struct rcu_resize_work {
258 struct rcu_head head;
14044b37 259 struct cds_lfht *ht;
abc490a1 260};
2ed95849 261
b7d619b0
MD
262struct partition_resize_work {
263 struct rcu_head head;
264 struct cds_lfht *ht;
265 unsigned long i, start, len;
266 void (*fct)(struct cds_lfht *ht, unsigned long i,
267 unsigned long start, unsigned long len);
268};
269
48ed1c18
MD
270enum add_mode {
271 ADD_DEFAULT = 0,
272 ADD_UNIQUE = 1,
273 ADD_REPLACE = 2,
274};
275
76a73da8
MD
276static
277struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
278 unsigned long size,
279 struct cds_lfht_node *node,
48ed1c18
MD
280 enum add_mode mode, int dummy);
281
282static
283int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
284 struct cds_lfht_node *node,
285 int dummy_removal, int do_gc);
76a73da8 286
abc490a1
MD
287/*
288 * Algorithm to reverse bits in a word by lookup table, extended to
289 * 64-bit words.
f9830efd 290 * Source:
abc490a1 291 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 292 * Originally from Public Domain.
abc490a1
MD
293 */
294
295static const uint8_t BitReverseTable256[256] =
2ed95849 296{
abc490a1
MD
297#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
298#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
299#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
300 R6(0), R6(2), R6(1), R6(3)
301};
302#undef R2
303#undef R4
304#undef R6
2ed95849 305
abc490a1
MD
306static
307uint8_t bit_reverse_u8(uint8_t v)
308{
309 return BitReverseTable256[v];
310}
ab7d5fc6 311
abc490a1
MD
312static __attribute__((unused))
313uint32_t bit_reverse_u32(uint32_t v)
314{
315 return ((uint32_t) bit_reverse_u8(v) << 24) |
316 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
317 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
318 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
319}
320
abc490a1
MD
321static __attribute__((unused))
322uint64_t bit_reverse_u64(uint64_t v)
2ed95849 323{
abc490a1
MD
324 return ((uint64_t) bit_reverse_u8(v) << 56) |
325 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
326 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
327 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
328 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
329 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
330 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
331 ((uint64_t) bit_reverse_u8(v >> 56));
332}
333
334static
335unsigned long bit_reverse_ulong(unsigned long v)
336{
337#if (CAA_BITS_PER_LONG == 32)
338 return bit_reverse_u32(v);
339#else
340 return bit_reverse_u64(v);
341#endif
342}
343
f9830efd 344/*
24365af7
MD
345 * fls: returns the position of the most significant bit.
346 * Returns 0 if no bit is set, else returns the position of the most
347 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 348 */
24365af7
MD
349#if defined(__i386) || defined(__x86_64)
350static inline
351unsigned int fls_u32(uint32_t x)
f9830efd 352{
24365af7
MD
353 int r;
354
355 asm("bsrl %1,%0\n\t"
356 "jnz 1f\n\t"
357 "movl $-1,%0\n\t"
358 "1:\n\t"
359 : "=r" (r) : "rm" (x));
360 return r + 1;
361}
362#define HAS_FLS_U32
363#endif
364
365#if defined(__x86_64)
366static inline
367unsigned int fls_u64(uint64_t x)
368{
369 long r;
370
371 asm("bsrq %1,%0\n\t"
372 "jnz 1f\n\t"
373 "movq $-1,%0\n\t"
374 "1:\n\t"
375 : "=r" (r) : "rm" (x));
376 return r + 1;
377}
378#define HAS_FLS_U64
379#endif
380
381#ifndef HAS_FLS_U64
382static __attribute__((unused))
383unsigned int fls_u64(uint64_t x)
384{
385 unsigned int r = 64;
386
387 if (!x)
388 return 0;
389
390 if (!(x & 0xFFFFFFFF00000000ULL)) {
391 x <<= 32;
392 r -= 32;
393 }
394 if (!(x & 0xFFFF000000000000ULL)) {
395 x <<= 16;
396 r -= 16;
397 }
398 if (!(x & 0xFF00000000000000ULL)) {
399 x <<= 8;
400 r -= 8;
401 }
402 if (!(x & 0xF000000000000000ULL)) {
403 x <<= 4;
404 r -= 4;
405 }
406 if (!(x & 0xC000000000000000ULL)) {
407 x <<= 2;
408 r -= 2;
409 }
410 if (!(x & 0x8000000000000000ULL)) {
411 x <<= 1;
412 r -= 1;
413 }
414 return r;
415}
416#endif
417
418#ifndef HAS_FLS_U32
419static __attribute__((unused))
420unsigned int fls_u32(uint32_t x)
421{
422 unsigned int r = 32;
f9830efd 423
24365af7
MD
424 if (!x)
425 return 0;
426 if (!(x & 0xFFFF0000U)) {
427 x <<= 16;
428 r -= 16;
429 }
430 if (!(x & 0xFF000000U)) {
431 x <<= 8;
432 r -= 8;
433 }
434 if (!(x & 0xF0000000U)) {
435 x <<= 4;
436 r -= 4;
437 }
438 if (!(x & 0xC0000000U)) {
439 x <<= 2;
440 r -= 2;
441 }
442 if (!(x & 0x80000000U)) {
443 x <<= 1;
444 r -= 1;
445 }
446 return r;
447}
448#endif
449
450unsigned int fls_ulong(unsigned long x)
f9830efd 451{
24365af7
MD
452#if (CAA_BITS_PER_lONG == 32)
453 return fls_u32(x);
454#else
455 return fls_u64(x);
456#endif
457}
f9830efd 458
24365af7
MD
459int get_count_order_u32(uint32_t x)
460{
461 int order;
462
463 order = fls_u32(x) - 1;
464 if (x & (x - 1))
465 order++;
466 return order;
467}
468
469int get_count_order_ulong(unsigned long x)
470{
471 int order;
472
473 order = fls_ulong(x) - 1;
474 if (x & (x - 1))
475 order++;
476 return order;
f9830efd
MD
477}
478
98808fb1
MD
479#ifdef POISON_FREE
480#define poison_free(ptr) \
481 do { \
482 memset(ptr, 0x42, sizeof(*(ptr))); \
483 free(ptr); \
484 } while (0)
485#else
486#define poison_free(ptr) free(ptr)
487#endif
488
f9830efd 489static
4105056a 490void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 491
df44348d
MD
492/*
493 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
494 * available, then we support hash table item accounting.
495 * In the unfortunate event the number of CPUs reported would be
496 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
497 */
df44348d
MD
498#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
499
f8994aee 500static
4105056a 501void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
502 unsigned long count);
503
df44348d
MD
504static long nr_cpus_mask = -1;
505
506static
507struct ht_items_count *alloc_per_cpu_items_count(void)
508{
509 struct ht_items_count *count;
510
511 switch (nr_cpus_mask) {
512 case -2:
513 return NULL;
514 case -1:
515 {
516 long maxcpus;
517
518 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
519 if (maxcpus <= 0) {
520 nr_cpus_mask = -2;
521 return NULL;
522 }
523 /*
524 * round up number of CPUs to next power of two, so we
525 * can use & for modulo.
526 */
527 maxcpus = 1UL << get_count_order_ulong(maxcpus);
528 nr_cpus_mask = maxcpus - 1;
529 }
530 /* Fall-through */
531 default:
532 return calloc(nr_cpus_mask + 1, sizeof(*count));
533 }
534}
535
536static
537void free_per_cpu_items_count(struct ht_items_count *count)
538{
98808fb1 539 poison_free(count);
df44348d
MD
540}
541
542static
543int ht_get_cpu(void)
544{
545 int cpu;
546
547 assert(nr_cpus_mask >= 0);
548 cpu = sched_getcpu();
549 if (unlikely(cpu < 0))
550 return cpu;
551 else
552 return cpu & nr_cpus_mask;
553}
554
555static
4105056a 556void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 557{
3171717f 558 unsigned long percpu_count;
df44348d
MD
559 int cpu;
560
561 if (unlikely(!ht->percpu_count))
3171717f 562 return;
df44348d
MD
563 cpu = ht_get_cpu();
564 if (unlikely(cpu < 0))
3171717f
MD
565 return;
566 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d
MD
567 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
568 unsigned long count;
569
570 dbg_printf("add percpu %lu\n", percpu_count);
571 count = uatomic_add_return(&ht->count,
572 1UL << COUNT_COMMIT_ORDER);
573 /* If power of 2 */
574 if (!(count & (count - 1))) {
4105056a 575 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee
MD
576 return;
577 dbg_printf("add set global %lu\n", count);
4105056a 578 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 579 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
580 }
581 }
582}
583
584static
860d07e8 585void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
586{
587 unsigned long percpu_count;
3171717f 588 int cpu;
df44348d 589
3171717f
MD
590 if (unlikely(!ht->percpu_count))
591 return;
592 cpu = ht_get_cpu();
593 if (unlikely(cpu < 0))
594 return;
860d07e8 595 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, -1);
df44348d
MD
596 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
597 unsigned long count;
598
860d07e8 599 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 600 count = uatomic_add_return(&ht->count,
3171717f 601 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
602 /* If power of 2 */
603 if (!(count & (count - 1))) {
4105056a 604 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 605 return;
860d07e8 606 dbg_printf("del set global %lu\n", count);
4105056a 607 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 608 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
609 }
610 }
611}
612
613#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
614
615static const long nr_cpus_mask = -1;
616
617static
618struct ht_items_count *alloc_per_cpu_items_count(void)
619{
620 return NULL;
621}
622
623static
624void free_per_cpu_items_count(struct ht_items_count *count)
625{
626}
627
628static
4105056a 629void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
630{
631}
632
633static
860d07e8 634void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
635{
636}
637
638#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
639
640
f9830efd 641static
4105056a 642void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 643{
f8994aee
MD
644 unsigned long count;
645
b8af5011
MD
646 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
647 return;
f8994aee
MD
648 count = uatomic_read(&ht->count);
649 /*
650 * Use bucket-local length for small table expand and for
651 * environments lacking per-cpu data support.
652 */
653 if (count >= (1UL << COUNT_COMMIT_ORDER))
654 return;
24365af7 655 if (chain_len > 100)
f0c29ed7 656 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 657 chain_len);
3390d470 658 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 659 cds_lfht_resize_lazy(ht, size,
01370f0b 660 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
661}
662
abc490a1 663static
14044b37 664struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 665{
14044b37 666 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
667}
668
669static
14044b37 670int is_removed(struct cds_lfht_node *node)
abc490a1 671{
d37166c6 672 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
673}
674
675static
14044b37 676struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 677{
14044b37 678 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
679}
680
48ed1c18
MD
681static
682int is_gc(struct cds_lfht_node *node)
683{
684 return ((unsigned long) node) & GC_FLAG;
685}
686
687static
688struct cds_lfht_node *flag_gc(struct cds_lfht_node *node)
689{
690 return (struct cds_lfht_node *) (((unsigned long) node) | GC_FLAG);
691}
692
f5596c94 693static
14044b37 694int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
695{
696 return ((unsigned long) node) & DUMMY_FLAG;
697}
698
699static
14044b37 700struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 701{
14044b37 702 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 703}
bb7b2f26
MD
704
705static
706struct cds_lfht_node *get_end(void)
707{
708 return (struct cds_lfht_node *) END_VALUE;
709}
710
711static
712int is_end(struct cds_lfht_node *node)
713{
714 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
715}
716
abc490a1 717static
f9830efd 718unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
719{
720 unsigned long old1, old2;
721
722 old1 = uatomic_read(ptr);
723 do {
724 old2 = old1;
725 if (old2 >= v)
f9830efd 726 return old2;
abc490a1 727 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 728 return v;
abc490a1
MD
729}
730
1475579c
MD
731static
732void cds_lfht_free_level(struct rcu_head *head)
733{
734 struct rcu_level *l =
735 caa_container_of(head, struct rcu_level, head);
98808fb1 736 poison_free(l);
1475579c
MD
737}
738
273399de
MD
739/*
740 * Remove all logically deleted nodes from a bucket up to a certain node key.
741 */
742static
f9c80341 743void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 744{
14044b37 745 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 746
c90201ac 747 assert(!is_dummy(dummy));
48ed1c18 748 assert(!is_gc(dummy));
c90201ac
MD
749 assert(!is_removed(dummy));
750 assert(!is_dummy(node));
48ed1c18 751 assert(!is_gc(node));
c90201ac 752 assert(!is_removed(node));
273399de
MD
753 for (;;) {
754 iter_prev = dummy;
755 /* We can always skip the dummy node initially */
cc4fcb10
MD
756 iter = rcu_dereference(iter_prev->p.next);
757 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
758 /*
759 * We should never be called with dummy (start of chain)
760 * and logically removed node (end of path compression
761 * marker) being the actual same node. This would be a
762 * bug in the algorithm implementation.
763 */
764 assert(dummy != node);
273399de 765 for (;;) {
bb7b2f26 766 if (unlikely(is_end(iter)))
f9c80341 767 return;
76412f24 768 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 769 return;
cc4fcb10 770 next = rcu_dereference(clear_flag(iter)->p.next);
48ed1c18 771 if (likely(is_gc(next)))
273399de 772 break;
b453eae1 773 iter_prev = clear_flag(iter);
273399de
MD
774 iter = next;
775 }
48ed1c18 776 assert(!is_gc(iter));
f5596c94
MD
777 if (is_dummy(iter))
778 new_next = flag_dummy(clear_flag(next));
779 else
780 new_next = clear_flag(next);
48ed1c18
MD
781 if (is_removed(iter))
782 new_next = flag_removed(new_next);
f5596c94 783 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 784 }
f9c80341 785 return;
273399de
MD
786}
787
abc490a1 788static
4105056a
MD
789struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
790 unsigned long size,
791 struct cds_lfht_node *node,
48ed1c18 792 enum add_mode mode, int dummy)
abc490a1 793{
14044b37 794 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
48ed1c18 795 *dummy_node, *return_node, *replace_pinned = NULL;
14044b37 796 struct _cds_lfht_node *lookup;
24365af7 797 unsigned long hash, index, order;
abc490a1 798
c90201ac 799 assert(!is_dummy(node));
48ed1c18 800 assert(!is_gc(node));
c90201ac 801 assert(!is_removed(node));
4105056a 802 if (!size) {
f5596c94 803 assert(dummy);
bb7b2f26 804 node->p.next = flag_dummy(get_end());
18117871
MD
805 return node; /* Initial first add (head) */
806 }
cc4fcb10 807 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 808 for (;;) {
48ed1c18 809 uint32_t chain_len;
abc490a1 810
48ed1c18
MD
811 retry:
812 chain_len = 0;
11519af6
MD
813 /*
814 * iter_prev points to the non-removed node prior to the
815 * insert location.
11519af6 816 */
4105056a 817 index = hash & (size - 1);
24365af7 818 order = get_count_order_ulong(index + 1);
4105056a 819 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 820 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 821 /* We can always skip the dummy node initially */
cc4fcb10
MD
822 iter = rcu_dereference(iter_prev->p.next);
823 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 824 for (;;) {
bb7b2f26 825 if (unlikely(is_end(iter)))
273399de 826 goto insert;
76412f24 827 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 828 goto insert;
cc4fcb10 829 next = rcu_dereference(clear_flag(iter)->p.next);
48ed1c18 830 if (unlikely(is_gc(next)))
9dba85be 831 goto gc_node;
48ed1c18
MD
832 if (unlikely(replace_pinned)) {
833 /*
834 * We're in the retry of a node
835 * replacement. Only get exact iter
836 * pointer match. We own it, so it
837 * _needs_ to be there at some point.
838 */
839 if (clear_flag(iter) == replace_pinned)
840 goto replace;
841 }
842 /*
843 * Next is removed but not gc'd. We need to
844 * busy-loop, because a concurrent replacement
845 * is keeping it temporarily pinned there but we
846 * cannot attach to it. The easiest solution is
847 * to retry.
848 */
849 if (unlikely(is_removed(next)))
850 goto retry;
851 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
1b81fe1a 852 && !is_dummy(next)
e43f23f8
MD
853 && !ht->compare_fct(node->key, node->key_len,
854 clear_flag(iter)->key,
48ed1c18
MD
855 clear_flag(iter)->key_len)) {
856 if (mode == ADD_UNIQUE)
857 return clear_flag(iter);
858 else /* mode == ADD_REPLACE */
859 goto replace;
860 }
11519af6 861 /* Only account for identical reverse hash once */
24365af7
MD
862 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
863 && !is_dummy(next))
4105056a 864 check_resize(ht, size, ++chain_len);
11519af6 865 iter_prev = clear_flag(iter);
273399de 866 iter = next;
abc490a1 867 }
48ed1c18 868
273399de 869 insert:
7ec59d3b 870 assert(node != clear_flag(iter));
11519af6 871 assert(!is_removed(iter_prev));
c90201ac 872 assert(!is_removed(iter));
48ed1c18
MD
873 assert(!is_gc(iter_prev));
874 assert(!is_gc(iter));
f000907d 875 assert(iter_prev != node);
48ed1c18 876 assert(!replace_pinned);
f9c80341 877 if (!dummy)
1b81fe1a 878 node->p.next = clear_flag(iter);
f9c80341
MD
879 else
880 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
881 if (is_dummy(iter))
882 new_node = flag_dummy(node);
883 else
884 new_node = node;
cc4fcb10 885 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 886 new_node) != iter) {
273399de 887 continue; /* retry */
48ed1c18
MD
888 } else {
889 if (mode == ADD_REPLACE)
890 return_node = NULL;
891 else /* ADD_DEFAULT and ADD_UNIQUE */
892 return_node = node;
273399de 893 goto gc_end;
48ed1c18
MD
894 }
895
896 replace:
897 assert(node != clear_flag(iter));
898 assert(!is_removed(iter_prev));
899 assert(!is_removed(iter));
900 assert(!is_gc(iter_prev));
901 assert(!is_gc(iter));
902 assert(iter_prev != node);
903 assert(!dummy);
904 node->p.next = clear_flag(next);
905 if (is_dummy(iter))
906 new_node = flag_dummy(node);
907 else
908 new_node = node;
909 /*
910 * Try to delete to-be-replaced node. Don't gc yet. Not
911 * performing gc here is important, because this lets
912 * concurrent lookups see the old node until we
913 * atomically swap the new node into its place.
914 *
915 * This algorithm is _not_ strictly lock-free between
916 * _cds_lfht_del and the uatomic_cmpxchg of the
917 * replacement operation, so a replacement should _not_
918 * crash here (which means: don't do replacements if you
919 * need strict lock-free guarantees).
920 */
921 if (!replace_pinned) {
922 if (_cds_lfht_del(ht, size, clear_flag(iter), 0, 0))
923 continue; /* concurrently removed. retry. */
924 }
925 /*
926 * After _cds_lfht_del succeeds, we have pinned the
927 * to-be-removed node in place by setting its removed
928 * flag, but not its gc flag. If we fail to cmpxchg our
929 * new node with this node, we need to retry everything
930 * from the initial lookup, and only stop when we reach
931 * the node we pinned into place.
932 */
933 return_node = uatomic_cmpxchg(&iter_prev->p.next,
934 iter, new_node);
935 if (return_node != iter) {
936 /*
937 * If cmpxchg fails, we need to do path
938 * compression, but end it by placing our own
939 * node into place.
940 */
941 replace_pinned = clear_flag(iter);
942 continue; /* retry */
943 } else {
944 /*
945 * cmpxchg succeeded. gc unnecessary, because we
946 * unlinked the return_node ourself with the
947 * cmpxchg.
948 */
949 return_node = clear_flag(return_node);
950 goto end;
951 }
952
9dba85be
MD
953 gc_node:
954 assert(!is_removed(iter));
48ed1c18 955 assert(!is_gc(iter));
f5596c94
MD
956 if (is_dummy(iter))
957 new_next = flag_dummy(clear_flag(next));
958 else
959 new_next = clear_flag(next);
960 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 961 /* retry */
464a1ec9 962 }
273399de
MD
963gc_end:
964 /* Garbage collect logically removed nodes in the bucket */
4105056a 965 index = hash & (size - 1);
24365af7 966 order = get_count_order_ulong(index + 1);
4105056a 967 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 968 dummy_node = (struct cds_lfht_node *) lookup;
f9c80341 969 _cds_lfht_gc_bucket(dummy_node, node);
48ed1c18
MD
970end:
971 return return_node;
abc490a1 972}
464a1ec9 973
abc490a1 974static
860d07e8 975int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 976 struct cds_lfht_node *node,
48ed1c18 977 int dummy_removal, int do_gc)
abc490a1 978{
14044b37
MD
979 struct cds_lfht_node *dummy, *next, *old;
980 struct _cds_lfht_node *lookup;
abc490a1 981 int flagged = 0;
24365af7 982 unsigned long hash, index, order;
5e28c532 983
7ec59d3b 984 /* logically delete the node */
c90201ac 985 assert(!is_dummy(node));
48ed1c18 986 assert(!is_gc(node));
c90201ac 987 assert(!is_removed(node));
cc4fcb10 988 old = rcu_dereference(node->p.next);
7ec59d3b 989 do {
48ed1c18
MD
990 struct cds_lfht_node *new_next;
991
7ec59d3b 992 next = old;
76412f24 993 if (unlikely(is_removed(next)))
7ec59d3b 994 goto end;
1475579c
MD
995 if (dummy_removal)
996 assert(is_dummy(next));
997 else
998 assert(!is_dummy(next));
48ed1c18
MD
999 new_next = flag_removed(next);
1000 if (do_gc)
1001 new_next = flag_gc(new_next);
1002 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
1003 } while (old != next);
1004
1005 /* We performed the (logical) deletion. */
1006 flagged = 1;
1007
48ed1c18
MD
1008 if (!do_gc)
1009 goto end;
1010
7ec59d3b
MD
1011 /*
1012 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
1013 * the node, and remove it (along with any other logically removed node)
1014 * if found.
11519af6 1015 */
cc4fcb10 1016 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
1017 assert(size > 0);
1018 index = hash & (size - 1);
24365af7 1019 order = get_count_order_ulong(index + 1);
4105056a 1020 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 1021 dummy = (struct cds_lfht_node *) lookup;
f9c80341 1022 _cds_lfht_gc_bucket(dummy, node);
2ed95849 1023end:
11519af6
MD
1024 /*
1025 * Only the flagging action indicated that we (and no other)
1026 * removed the node from the hash.
1027 */
7ec59d3b 1028 if (flagged) {
cc4fcb10 1029 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 1030 return 0;
7ec59d3b 1031 } else
11519af6 1032 return -ENOENT;
abc490a1 1033}
2ed95849 1034
b7d619b0
MD
1035static
1036void *partition_resize_thread(void *arg)
1037{
1038 struct partition_resize_work *work = arg;
1039
1040 work->ht->cds_lfht_rcu_register_thread();
1041 work->fct(work->ht, work->i, work->start, work->len);
1042 work->ht->cds_lfht_rcu_unregister_thread();
1043 return NULL;
1044}
1045
1046static
1047void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
1048 unsigned long len,
1049 void (*fct)(struct cds_lfht *ht, unsigned long i,
1050 unsigned long start, unsigned long len))
1051{
1052 unsigned long partition_len;
1053 struct partition_resize_work *work;
6083a889
MD
1054 int thread, ret;
1055 unsigned long nr_threads;
b7d619b0
MD
1056 pthread_t *thread_id;
1057
6083a889
MD
1058 /*
1059 * Note: nr_cpus_mask + 1 is always power of 2.
1060 * We spawn just the number of threads we need to satisfy the minimum
1061 * partition size, up to the number of CPUs in the system.
1062 */
1063 nr_threads = min(nr_cpus_mask + 1,
1064 len >> MIN_PARTITION_PER_THREAD_ORDER);
1065 partition_len = len >> get_count_order_ulong(nr_threads);
1066 work = calloc(nr_threads, sizeof(*work));
1067 thread_id = calloc(nr_threads, sizeof(*thread_id));
b7d619b0 1068 assert(work);
6083a889
MD
1069 for (thread = 0; thread < nr_threads; thread++) {
1070 work[thread].ht = ht;
1071 work[thread].i = i;
1072 work[thread].len = partition_len;
1073 work[thread].start = thread * partition_len;
1074 work[thread].fct = fct;
1075 ret = pthread_create(&thread_id[thread], ht->resize_attr,
1076 partition_resize_thread, &work[thread]);
b7d619b0
MD
1077 assert(!ret);
1078 }
6083a889
MD
1079 for (thread = 0; thread < nr_threads; thread++) {
1080 ret = pthread_join(thread_id[thread], NULL);
b7d619b0
MD
1081 assert(!ret);
1082 }
1083 free(work);
1084 free(thread_id);
1085}
1086
e8de508e
MD
1087/*
1088 * Holding RCU read lock to protect _cds_lfht_add against memory
1089 * reclaim that could be performed by other call_rcu worker threads (ABA
1090 * problem).
9ee0fc9a 1091 *
b7d619b0 1092 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1093 * many worker threads, based on the number of CPUs available in the system.
1094 * This should therefore take care of not having the expand lagging behind too
1095 * many concurrent insertion threads by using the scheduler's ability to
1096 * schedule dummy node population fairly with insertions.
e8de508e 1097 */
4105056a 1098static
b7d619b0
MD
1099void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1100 unsigned long start, unsigned long len)
4105056a
MD
1101{
1102 unsigned long j;
1103
1104 ht->cds_lfht_rcu_read_lock();
b7d619b0 1105 for (j = start; j < start + len; j++) {
4105056a
MD
1106 struct cds_lfht_node *new_node =
1107 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1108
dc1da8f6 1109 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1110 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1111 new_node->p.reverse_hash =
1112 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a 1113 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1114 new_node, ADD_DEFAULT, 1);
4105056a
MD
1115 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1116 break;
1117 }
1118 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1119}
1120
1121static
1122void init_table_populate(struct cds_lfht *ht, unsigned long i,
1123 unsigned long len)
1124{
1125 assert(nr_cpus_mask != -1);
6083a889 1126 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1127 ht->cds_lfht_rcu_thread_online();
1128 init_table_populate_partition(ht, i, 0, len);
1129 ht->cds_lfht_rcu_thread_offline();
1130 return;
1131 }
1132 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1133}
1134
abc490a1 1135static
4105056a 1136void init_table(struct cds_lfht *ht,
24365af7
MD
1137 unsigned long first_order, unsigned long len_order)
1138{
1139 unsigned long i, end_order;
1140
f0c29ed7 1141 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1142 first_order, first_order + len_order);
1143 end_order = first_order + len_order;
24365af7 1144 for (i = first_order; i < end_order; i++) {
4105056a 1145 unsigned long len;
24365af7
MD
1146
1147 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1148 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1149
1150 /* Stop expand if the resize target changes under us */
1151 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1152 break;
1153
4105056a 1154 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1155 + (len * sizeof(struct _cds_lfht_node)));
b7d619b0 1156 assert(ht->t.tbl[i]);
4105056a 1157
4105056a 1158 /*
dc1da8f6
MD
1159 * Set all dummy nodes reverse hash values for a level and
1160 * link all dummy nodes into the table.
4105056a 1161 */
dc1da8f6 1162 init_table_populate(ht, i, len);
4105056a 1163
f9c80341
MD
1164 /*
1165 * Update table size.
1166 */
1167 cmm_smp_wmb(); /* populate data before RCU size */
1168 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1169
4105056a
MD
1170 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1171 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1172 break;
1173 }
1174}
1175
e8de508e
MD
1176/*
1177 * Holding RCU read lock to protect _cds_lfht_remove against memory
1178 * reclaim that could be performed by other call_rcu worker threads (ABA
1179 * problem).
1180 * For a single level, we logically remove and garbage collect each node.
1181 *
1182 * As a design choice, we perform logical removal and garbage collection on a
1183 * node-per-node basis to simplify this algorithm. We also assume keeping good
1184 * cache locality of the operation would overweight possible performance gain
1185 * that could be achieved by batching garbage collection for multiple levels.
1186 * However, this would have to be justified by benchmarks.
1187 *
1188 * Concurrent removal and add operations are helping us perform garbage
1189 * collection of logically removed nodes. We guarantee that all logically
1190 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1191 * invoked to free a hole level of dummy nodes (after a grace period).
1192 *
1193 * Logical removal and garbage collection can therefore be done in batch or on a
1194 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1195 *
b7d619b0
MD
1196 * When we reach a certain length, we can split this removal over many worker
1197 * threads, based on the number of CPUs available in the system. This should
1198 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1199 * updater threads actively inserting into the hash table.
e8de508e 1200 */
4105056a 1201static
b7d619b0
MD
1202void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1203 unsigned long start, unsigned long len)
4105056a
MD
1204{
1205 unsigned long j;
1206
1207 ht->cds_lfht_rcu_read_lock();
b7d619b0 1208 for (j = start; j < start + len; j++) {
4105056a
MD
1209 struct cds_lfht_node *fini_node =
1210 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1211
1212 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1213 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1214 fini_node->p.reverse_hash =
1215 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1216 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1217 fini_node, 1, 1);
33c7c748
MD
1218 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1219 break;
abc490a1 1220 }
4105056a 1221 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1222}
1223
1224static
1225void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1226{
1227
1228 assert(nr_cpus_mask != -1);
6083a889 1229 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1230 ht->cds_lfht_rcu_thread_online();
1231 remove_table_partition(ht, i, 0, len);
1232 ht->cds_lfht_rcu_thread_offline();
1233 return;
1234 }
1235 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1236}
1237
1475579c 1238static
4105056a 1239void fini_table(struct cds_lfht *ht,
1475579c
MD
1240 unsigned long first_order, unsigned long len_order)
1241{
1242 long i, end_order;
1243
1244 dbg_printf("fini table: first_order %lu end_order %lu\n",
1245 first_order, first_order + len_order);
1246 end_order = first_order + len_order;
1247 assert(first_order > 0);
1475579c 1248 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1249 unsigned long len;
1475579c
MD
1250
1251 len = !i ? 1 : 1UL << (i - 1);
1252 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1253
4d676753
MD
1254 /* Stop shrink if the resize target changes under us */
1255 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1256 break;
1257
1258 cmm_smp_wmb(); /* populate data before RCU size */
1259 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1260
1261 /*
1262 * We need to wait for all add operations to reach Q.S. (and
1263 * thus use the new table for lookups) before we can start
1264 * releasing the old dummy nodes. Otherwise their lookup will
1265 * return a logically removed node as insert position.
1266 */
1267 ht->cds_lfht_synchronize_rcu();
1268
21263e21 1269 /*
4105056a
MD
1270 * Set "removed" flag in dummy nodes about to be removed.
1271 * Unlink all now-logically-removed dummy node pointers.
1272 * Concurrent add/remove operation are helping us doing
1273 * the gc.
21263e21 1274 */
4105056a
MD
1275 remove_table(ht, i, len);
1276
1277 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1278
1279 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1280 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1281 break;
1282 }
1475579c
MD
1283}
1284
7a9dcf9b 1285struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1286 cds_lfht_compare_fct compare_fct,
1287 unsigned long hash_seed,
1288 unsigned long init_size,
b8af5011 1289 int flags,
14044b37 1290 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1291 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1292 void (*cds_lfht_synchronize_rcu)(void),
1293 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1294 void (*cds_lfht_rcu_read_unlock)(void),
1295 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1296 void (*cds_lfht_rcu_thread_online)(void),
1297 void (*cds_lfht_rcu_register_thread)(void),
1298 void (*cds_lfht_rcu_unregister_thread)(void),
1299 pthread_attr_t *attr)
abc490a1 1300{
14044b37 1301 struct cds_lfht *ht;
24365af7 1302 unsigned long order;
abc490a1 1303
8129be4e 1304 /* init_size must be power of two */
49619ea0 1305 if (init_size && (init_size & (init_size - 1)))
8129be4e 1306 return NULL;
14044b37 1307 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1308 assert(ht);
abc490a1 1309 ht->hash_fct = hash_fct;
732ad076
MD
1310 ht->compare_fct = compare_fct;
1311 ht->hash_seed = hash_seed;
14044b37 1312 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1313 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1314 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1315 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1316 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1317 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1318 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1319 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1320 ht->resize_attr = attr;
df44348d 1321 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1322 /* this mutex should not nest in read-side C.S. */
1323 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1324 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1325 ht->flags = flags;
5f511391 1326 ht->cds_lfht_rcu_thread_offline();
f000907d 1327 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1328 ht->t.resize_target = 1UL << (order - 1);
4105056a 1329 init_table(ht, 0, order);
f000907d 1330 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1331 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1332 return ht;
1333}
1334
14044b37 1335struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
2ed95849 1336{
bb7b2f26 1337 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1338 struct _cds_lfht_node *lookup;
4105056a 1339 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1340
732ad076 1341 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1342 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1343
4105056a
MD
1344 size = rcu_dereference(ht->t.size);
1345 index = hash & (size - 1);
24365af7 1346 order = get_count_order_ulong(index + 1);
4105056a 1347 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1348 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1349 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1350 dummy_node = (struct cds_lfht_node *) lookup;
1351 /* We can always skip the dummy node initially */
1352 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1353 node = clear_flag(node);
2ed95849 1354 for (;;) {
bb7b2f26
MD
1355 if (unlikely(is_end(node))) {
1356 node = NULL;
abc490a1 1357 break;
bb7b2f26 1358 }
cc4fcb10 1359 if (unlikely(node->p.reverse_hash > reverse_hash)) {
abc490a1
MD
1360 node = NULL;
1361 break;
2ed95849 1362 }
1b81fe1a 1363 next = rcu_dereference(node->p.next);
48ed1c18
MD
1364 /*
1365 * We consider return nodes marked removed but not gc as
1366 * hits for lookup vs replacement consistency.
1367 */
1368 if (likely(!is_gc(next))
1b81fe1a 1369 && !is_dummy(next)
49c2e2d6 1370 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1371 break;
2ed95849 1372 }
1b81fe1a 1373 node = clear_flag(next);
2ed95849 1374 }
1b81fe1a 1375 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
abc490a1
MD
1376 return node;
1377}
e0ba718a 1378
a481e5ff
MD
1379struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
1380 struct cds_lfht_node *node)
1381{
1382 struct cds_lfht_node *next;
1383 unsigned long reverse_hash;
1384 void *key;
1385 size_t key_len;
1386
1387 reverse_hash = node->p.reverse_hash;
1388 key = node->key;
1389 key_len = node->key_len;
1390 next = rcu_dereference(node->p.next);
1391 node = clear_flag(next);
1392
1393 for (;;) {
bb7b2f26
MD
1394 if (unlikely(is_end(node))) {
1395 node = NULL;
a481e5ff 1396 break;
bb7b2f26 1397 }
a481e5ff
MD
1398 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1399 node = NULL;
1400 break;
1401 }
1402 next = rcu_dereference(node->p.next);
48ed1c18
MD
1403 /*
1404 * We consider return nodes marked removed but not gc as
1405 * hits for lookup vs replacement consistency.
1406 */
1407 if (likely(!is_gc(next))
a481e5ff
MD
1408 && !is_dummy(next)
1409 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1410 break;
1411 }
1412 node = clear_flag(next);
1413 }
1414 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
1415 return node;
1416}
1417
14044b37 1418void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1419{
4105056a 1420 unsigned long hash, size;
ab7d5fc6 1421
49c2e2d6 1422 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1423 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1424
4105056a 1425 size = rcu_dereference(ht->t.size);
48ed1c18 1426 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
4105056a 1427 ht_count_add(ht, size);
3eca1b8c
MD
1428}
1429
14044b37 1430struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1431 struct cds_lfht_node *node)
3eca1b8c 1432{
4105056a 1433 unsigned long hash, size;
df44348d 1434 struct cds_lfht_node *ret;
3eca1b8c 1435
49c2e2d6 1436 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1437 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1438
4105056a 1439 size = rcu_dereference(ht->t.size);
48ed1c18 1440 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
17f31d1b 1441 if (ret == node)
4105056a 1442 ht_count_add(ht, size);
df44348d 1443 return ret;
2ed95849
MD
1444}
1445
48ed1c18
MD
1446struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
1447 struct cds_lfht_node *node)
1448{
1449 unsigned long hash, size;
1450 struct cds_lfht_node *ret;
1451
1452 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1453 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1454
1455 size = rcu_dereference(ht->t.size);
1456 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1457 if (ret == NULL)
1458 ht_count_add(ht, size);
1459 return ret;
1460}
1461
860d07e8 1462int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1463{
4105056a 1464 unsigned long size;
df44348d 1465 int ret;
abc490a1 1466
4105056a 1467 size = rcu_dereference(ht->t.size);
48ed1c18 1468 ret = _cds_lfht_del(ht, size, node, 0, 1);
df44348d 1469 if (!ret)
860d07e8 1470 ht_count_del(ht, size);
df44348d 1471 return ret;
2ed95849 1472}
ab7d5fc6 1473
abc490a1 1474static
14044b37 1475int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1476{
14044b37
MD
1477 struct cds_lfht_node *node;
1478 struct _cds_lfht_node *lookup;
4105056a 1479 unsigned long order, i, size;
674f7a69 1480
abc490a1 1481 /* Check that the table is empty */
4105056a 1482 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1483 node = (struct cds_lfht_node *) lookup;
abc490a1 1484 do {
1b81fe1a
MD
1485 node = clear_flag(node)->p.next;
1486 if (!is_dummy(node))
abc490a1 1487 return -EPERM;
273399de 1488 assert(!is_removed(node));
48ed1c18 1489 assert(!is_gc(node));
bb7b2f26 1490 } while (!is_end(node));
4105056a
MD
1491 /*
1492 * size accessed without rcu_dereference because hash table is
1493 * being destroyed.
1494 */
1495 size = ht->t.size;
abc490a1 1496 /* Internal sanity check: all nodes left should be dummy */
4105056a 1497 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1498 unsigned long len;
1499
1500 len = !order ? 1 : 1UL << (order - 1);
1501 for (i = 0; i < len; i++) {
f0c29ed7 1502 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1503 order, i,
4105056a
MD
1504 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1505 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1506 }
4105056a 1507 poison_free(ht->t.tbl[order]);
674f7a69 1508 }
abc490a1 1509 return 0;
674f7a69
MD
1510}
1511
1512/*
1513 * Should only be called when no more concurrent readers nor writers can
1514 * possibly access the table.
1515 */
b7d619b0 1516int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1517{
5e28c532
MD
1518 int ret;
1519
848d4088 1520 /* Wait for in-flight resize operations to complete */
33c7c748 1521 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
1522 while (uatomic_read(&ht->in_progress_resize))
1523 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1524 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1525 if (ret)
1526 return ret;
df44348d 1527 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1528 if (attr)
1529 *attr = ht->resize_attr;
98808fb1 1530 poison_free(ht);
5e28c532 1531 return ret;
674f7a69
MD
1532}
1533
14044b37 1534void cds_lfht_count_nodes(struct cds_lfht *ht,
273399de
MD
1535 unsigned long *count,
1536 unsigned long *removed)
1537{
14044b37
MD
1538 struct cds_lfht_node *node, *next;
1539 struct _cds_lfht_node *lookup;
24365af7 1540 unsigned long nr_dummy = 0;
273399de
MD
1541
1542 *count = 0;
1543 *removed = 0;
1544
24365af7 1545 /* Count non-dummy nodes in the table */
4105056a 1546 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1547 node = (struct cds_lfht_node *) lookup;
273399de 1548 do {
cc4fcb10 1549 next = rcu_dereference(node->p.next);
48ed1c18 1550 if (is_removed(next) || is_gc(next)) {
1b81fe1a 1551 assert(!is_dummy(next));
273399de 1552 (*removed)++;
1b81fe1a 1553 } else if (!is_dummy(next))
273399de 1554 (*count)++;
24365af7
MD
1555 else
1556 (nr_dummy)++;
273399de 1557 node = clear_flag(next);
bb7b2f26 1558 } while (!is_end(node));
f0c29ed7 1559 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
273399de
MD
1560}
1561
1475579c 1562/* called with resize mutex held */
abc490a1 1563static
4105056a 1564void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1565 unsigned long old_size, unsigned long new_size)
abc490a1 1566{
1475579c 1567 unsigned long old_order, new_order;
1475579c
MD
1568
1569 old_order = get_count_order_ulong(old_size) + 1;
1570 new_order = get_count_order_ulong(new_size) + 1;
1571 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1572 old_size, old_order, new_size, new_order);
1475579c 1573 assert(new_size > old_size);
4105056a 1574 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1575}
1576
1577/* called with resize mutex held */
1578static
4105056a 1579void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1580 unsigned long old_size, unsigned long new_size)
464a1ec9 1581{
1475579c 1582 unsigned long old_order, new_order;
464a1ec9 1583
cd95516d 1584 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1585 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1586 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1587 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1588 old_size, old_order, new_size, new_order);
1475579c 1589 assert(new_size < old_size);
1475579c 1590
4105056a
MD
1591 /* Remove and unlink all dummy nodes to remove. */
1592 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1593}
1594
1475579c
MD
1595
1596/* called with resize mutex held */
1597static
1598void _do_cds_lfht_resize(struct cds_lfht *ht)
1599{
1600 unsigned long new_size, old_size;
4105056a
MD
1601
1602 /*
1603 * Resize table, re-do if the target size has changed under us.
1604 */
1605 do {
1606 ht->t.resize_initiated = 1;
1607 old_size = ht->t.size;
1608 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1609 if (old_size < new_size)
1610 _do_cds_lfht_grow(ht, old_size, new_size);
1611 else if (old_size > new_size)
1612 _do_cds_lfht_shrink(ht, old_size, new_size);
1613 ht->t.resize_initiated = 0;
1614 /* write resize_initiated before read resize_target */
1615 cmm_smp_mb();
4d676753 1616 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1617}
1618
abc490a1 1619static
4105056a 1620unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1621 int growth_order)
464a1ec9 1622{
4105056a
MD
1623 return _uatomic_max(&ht->t.resize_target,
1624 size << growth_order);
464a1ec9
MD
1625}
1626
1475579c 1627static
4105056a 1628void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1629 unsigned long count)
1475579c 1630{
cd95516d 1631 count = max(count, MIN_TABLE_SIZE);
4105056a 1632 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1633}
1634
1635void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1636{
4105056a
MD
1637 resize_target_update_count(ht, new_size);
1638 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1639 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1640 pthread_mutex_lock(&ht->resize_mutex);
1641 _do_cds_lfht_resize(ht);
1642 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1643 ht->cds_lfht_rcu_thread_online();
abc490a1 1644}
464a1ec9 1645
abc490a1
MD
1646static
1647void do_resize_cb(struct rcu_head *head)
1648{
1649 struct rcu_resize_work *work =
1650 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1651 struct cds_lfht *ht = work->ht;
abc490a1 1652
5f511391 1653 ht->cds_lfht_rcu_thread_offline();
abc490a1 1654 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1655 _do_cds_lfht_resize(ht);
abc490a1 1656 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1657 ht->cds_lfht_rcu_thread_online();
98808fb1 1658 poison_free(work);
848d4088
MD
1659 cmm_smp_mb(); /* finish resize before decrement */
1660 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1661}
1662
abc490a1 1663static
4105056a 1664void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1665{
abc490a1 1666 struct rcu_resize_work *work;
f9830efd 1667 unsigned long target_size;
abc490a1 1668
4105056a
MD
1669 target_size = resize_target_update(ht, size, growth);
1670 /* Store resize_target before read resize_initiated */
1671 cmm_smp_mb();
1672 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088
MD
1673 uatomic_inc(&ht->in_progress_resize);
1674 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1675 work = malloc(sizeof(*work));
1676 work->ht = ht;
14044b37 1677 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1678 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1679 }
ab7d5fc6 1680}
3171717f 1681
f8994aee
MD
1682#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1683
3171717f 1684static
4105056a 1685void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1686 unsigned long count)
1687{
1688 struct rcu_resize_work *work;
3171717f 1689
b8af5011
MD
1690 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1691 return;
4105056a
MD
1692 resize_target_update_count(ht, count);
1693 /* Store resize_target before read resize_initiated */
1694 cmm_smp_mb();
1695 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f
MD
1696 uatomic_inc(&ht->in_progress_resize);
1697 cmm_smp_mb(); /* increment resize count before calling it */
1698 work = malloc(sizeof(*work));
1699 work->ht = ht;
1700 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1701 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1702 }
1703}
f8994aee
MD
1704
1705#endif
This page took 0.185685 seconds and 4 git commands to generate.