rculfhash cleanup: count percpu deletes in the positive range
[urcu.git] / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
1475579c 4 * Userspace RCU library - Lock-Free Resizable RCU Hash Table
abc490a1
MD
5 *
6 * Copyright 2010-2011 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
7 *
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
5e28c532
MD
21 */
22
e753ff5a
MD
23/*
24 * Based on the following articles:
25 * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free
26 * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405.
27 * - Michael, M. M. High performance dynamic lock-free hash tables
28 * and list-based sets. In Proceedings of the fourteenth annual ACM
29 * symposium on Parallel algorithms and architectures, ACM Press,
30 * (2002), 73-82.
31 *
1475579c 32 * Some specificities of this Lock-Free Resizable RCU Hash Table
e753ff5a
MD
33 * implementation:
34 *
35 * - RCU read-side critical section allows readers to perform hash
36 * table lookups and use the returned objects safely by delaying
37 * memory reclaim of a grace period.
38 * - Add and remove operations are lock-free, and do not need to
39 * allocate memory. They need to be executed within RCU read-side
40 * critical section to ensure the objects they read are valid and to
41 * deal with the cmpxchg ABA problem.
42 * - add and add_unique operations are supported. add_unique checks if
43 * the node key already exists in the hash table. It ensures no key
44 * duplicata exists.
45 * - The resize operation executes concurrently with add/remove/lookup.
46 * - Hash table nodes are contained within a split-ordered list. This
47 * list is ordered by incrementing reversed-bits-hash value.
48 * - An index of dummy nodes is kept. These dummy nodes are the hash
49 * table "buckets", and they are also chained together in the
50 * split-ordered list, which allows recursive expansion.
1475579c
MD
51 * - The resize operation for small tables only allows expanding the hash table.
52 * It is triggered automatically by detecting long chains in the add
53 * operation.
54 * - The resize operation for larger tables (and available through an
55 * API) allows both expanding and shrinking the hash table.
56 * - Per-CPU Split-counters are used to keep track of the number of
57 * nodes within the hash table for automatic resize triggering.
e753ff5a
MD
58 * - Resize operation initiated by long chain detection is executed by a
59 * call_rcu thread, which keeps lock-freedom of add and remove.
60 * - Resize operations are protected by a mutex.
61 * - The removal operation is split in two parts: first, a "removed"
62 * flag is set in the next pointer within the node to remove. Then,
63 * a "garbage collection" is performed in the bucket containing the
64 * removed node (from the start of the bucket up to the removed node).
65 * All encountered nodes with "removed" flag set in their next
66 * pointers are removed from the linked-list. If the cmpxchg used for
67 * removal fails (due to concurrent garbage-collection or concurrent
68 * add), we retry from the beginning of the bucket. This ensures that
69 * the node with "removed" flag set is removed from the hash table
70 * (not visible to lookups anymore) before the RCU read-side critical
71 * section held across removal ends. Furthermore, this ensures that
72 * the node with "removed" flag set is removed from the linked-list
73 * before its memory is reclaimed. Only the thread which removal
74 * successfully set the "removed" flag (with a cmpxchg) into a node's
75 * next pointer is considered to have succeeded its removal (and thus
76 * owns the node to reclaim). Because we garbage-collect starting from
77 * an invariant node (the start-of-bucket dummy node) up to the
78 * "removed" node (or find a reverse-hash that is higher), we are sure
79 * that a successful traversal of the chain leads to a chain that is
80 * present in the linked-list (the start node is never removed) and
81 * that is does not contain the "removed" node anymore, even if
82 * concurrent delete/add operations are changing the structure of the
83 * list concurrently.
29e669f6
MD
84 * - The add operation performs gargage collection of buckets if it
85 * encounters nodes with removed flag set in the bucket where it wants
86 * to add its new node. This ensures lock-freedom of add operation by
87 * helping the remover unlink nodes from the list rather than to wait
88 * for it do to so.
e753ff5a
MD
89 * - A RCU "order table" indexed by log2(hash index) is copied and
90 * expanded by the resize operation. This order table allows finding
91 * the "dummy node" tables.
92 * - There is one dummy node table per hash index order. The size of
93 * each dummy node table is half the number of hashes contained in
94 * this order.
95 * - call_rcu is used to garbage-collect the old order table.
96 * - The per-order dummy node tables contain a compact version of the
97 * hash table nodes. These tables are invariant after they are
98 * populated into the hash table.
1475579c
MD
99 *
100 * A bit of ascii art explanation:
101 *
102 * Order index is the off-by-one compare to the actual power of 2 because
103 * we use index 0 to deal with the 0 special-case.
104 *
105 * This shows the nodes for a small table ordered by reversed bits:
106 *
107 * bits reverse
108 * 0 000 000
109 * 4 100 001
110 * 2 010 010
111 * 6 110 011
112 * 1 001 100
113 * 5 101 101
114 * 3 011 110
115 * 7 111 111
116 *
117 * This shows the nodes in order of non-reversed bits, linked by
118 * reversed-bit order.
119 *
120 * order bits reverse
121 * 0 0 000 000
122 * |
f6fdd688
MD
123 * 1 | 1 001 100 <- <-
124 * | | | |
125 * 2 | | 2 010 010 | |
126 * | | | 3 011 110 | <- |
127 * | | | | | | |
1475579c
MD
128 * 3 -> | | | 4 100 001 | |
129 * -> | | 5 101 101 |
130 * -> | 6 110 011
131 * -> 7 111 111
e753ff5a
MD
132 */
133
2ed95849
MD
134#define _LGPL_SOURCE
135#include <stdlib.h>
e0ba718a
MD
136#include <errno.h>
137#include <assert.h>
138#include <stdio.h>
abc490a1 139#include <stdint.h>
f000907d 140#include <string.h>
e0ba718a 141
df44348d 142#include "config.h"
2ed95849 143#include <urcu.h>
abc490a1 144#include <urcu-call-rcu.h>
a42cc659
MD
145#include <urcu/arch.h>
146#include <urcu/uatomic.h>
674f7a69 147#include <urcu/jhash.h>
a42cc659 148#include <urcu/compiler.h>
abc490a1 149#include <urcu/rculfhash.h>
5e28c532 150#include <stdio.h>
464a1ec9 151#include <pthread.h>
44395fb7 152
f9830efd 153#ifdef DEBUG
f0c29ed7 154#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 155#else
e753ff5a 156#define dbg_printf(fmt, args...)
f9830efd
MD
157#endif
158
f8994aee
MD
159/*
160 * Per-CPU split-counters lazily update the global counter each 1024
161 * addition/removal. It automatically keeps track of resize required.
162 * We use the bucket length as indicator for need to expand for small
163 * tables and machines lacking per-cpu data suppport.
164 */
165#define COUNT_COMMIT_ORDER 10
6ea6bc67
MD
166#define CHAIN_LEN_TARGET 1
167#define CHAIN_LEN_RESIZE_THRESHOLD 3
2ed95849 168
cd95516d 169/*
76a73da8 170 * Define the minimum table size.
cd95516d 171 */
c9edd44a 172#define MIN_TABLE_SIZE 1
cd95516d 173
4105056a
MD
174#if (CAA_BITS_PER_LONG == 32)
175#define MAX_TABLE_ORDER 32
176#else
177#define MAX_TABLE_ORDER 64
178#endif
179
b7d619b0
MD
180/*
181 * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
182 */
6083a889
MD
183#define MIN_PARTITION_PER_THREAD_ORDER 12
184#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
b7d619b0 185
4105056a
MD
186#ifndef min
187#define min(a, b) ((a) < (b) ? (a) : (b))
188#endif
189
abc490a1
MD
190#ifndef max
191#define max(a, b) ((a) > (b) ? (a) : (b))
192#endif
2ed95849 193
d95bd160
MD
194/*
195 * The removed flag needs to be updated atomically with the pointer.
48ed1c18 196 * It indicates that no node must attach to the node scheduled for
b198f0fd 197 * removal, and that node garbage collection must be performed.
d95bd160
MD
198 * The dummy flag does not require to be updated atomically with the
199 * pointer, but it is added as a pointer low bit flag to save space.
200 */
d37166c6 201#define REMOVED_FLAG (1UL << 0)
b198f0fd
MD
202#define DUMMY_FLAG (1UL << 1)
203#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 204
bb7b2f26 205/* Value of the end pointer. Should not interact with flags. */
f9c80341 206#define END_VALUE NULL
bb7b2f26 207
df44348d 208struct ht_items_count {
860d07e8 209 unsigned long add, del;
df44348d
MD
210} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
211
1475579c
MD
212struct rcu_level {
213 struct rcu_head head;
214 struct _cds_lfht_node nodes[0];
215};
216
395270b6 217struct rcu_table {
4105056a 218 unsigned long size; /* always a power of 2, shared (RCU) */
f9830efd 219 unsigned long resize_target;
11519af6 220 int resize_initiated;
4105056a 221 struct rcu_level *tbl[MAX_TABLE_ORDER];
395270b6
MD
222};
223
14044b37 224struct cds_lfht {
4105056a 225 struct rcu_table t;
14044b37
MD
226 cds_lfht_hash_fct hash_fct;
227 cds_lfht_compare_fct compare_fct;
732ad076 228 unsigned long hash_seed;
b8af5011 229 int flags;
5f511391
MD
230 /*
231 * We need to put the work threads offline (QSBR) when taking this
232 * mutex, because we use synchronize_rcu within this mutex critical
233 * section, which waits on read-side critical sections, and could
234 * therefore cause grace-period deadlock if we hold off RCU G.P.
235 * completion.
236 */
464a1ec9 237 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 238 unsigned int in_progress_resize, in_progress_destroy;
14044b37 239 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 240 void (*func)(struct rcu_head *head));
1475579c 241 void (*cds_lfht_synchronize_rcu)(void);
01dbfa62
MD
242 void (*cds_lfht_rcu_read_lock)(void);
243 void (*cds_lfht_rcu_read_unlock)(void);
5f511391
MD
244 void (*cds_lfht_rcu_thread_offline)(void);
245 void (*cds_lfht_rcu_thread_online)(void);
b7d619b0
MD
246 void (*cds_lfht_rcu_register_thread)(void);
247 void (*cds_lfht_rcu_unregister_thread)(void);
248 pthread_attr_t *resize_attr; /* Resize threads attributes */
df44348d
MD
249 unsigned long count; /* global approximate item count */
250 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
251};
252
abc490a1
MD
253struct rcu_resize_work {
254 struct rcu_head head;
14044b37 255 struct cds_lfht *ht;
abc490a1 256};
2ed95849 257
b7d619b0
MD
258struct partition_resize_work {
259 struct rcu_head head;
260 struct cds_lfht *ht;
261 unsigned long i, start, len;
262 void (*fct)(struct cds_lfht *ht, unsigned long i,
263 unsigned long start, unsigned long len);
264};
265
48ed1c18
MD
266enum add_mode {
267 ADD_DEFAULT = 0,
268 ADD_UNIQUE = 1,
269 ADD_REPLACE = 2,
270};
271
76a73da8
MD
272static
273struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
274 unsigned long size,
275 struct cds_lfht_node *node,
48ed1c18
MD
276 enum add_mode mode, int dummy);
277
abc490a1
MD
278/*
279 * Algorithm to reverse bits in a word by lookup table, extended to
280 * 64-bit words.
f9830efd 281 * Source:
abc490a1 282 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 283 * Originally from Public Domain.
abc490a1
MD
284 */
285
286static const uint8_t BitReverseTable256[256] =
2ed95849 287{
abc490a1
MD
288#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
289#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
290#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
291 R6(0), R6(2), R6(1), R6(3)
292};
293#undef R2
294#undef R4
295#undef R6
2ed95849 296
abc490a1
MD
297static
298uint8_t bit_reverse_u8(uint8_t v)
299{
300 return BitReverseTable256[v];
301}
ab7d5fc6 302
abc490a1
MD
303static __attribute__((unused))
304uint32_t bit_reverse_u32(uint32_t v)
305{
306 return ((uint32_t) bit_reverse_u8(v) << 24) |
307 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
308 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
309 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
310}
311
abc490a1
MD
312static __attribute__((unused))
313uint64_t bit_reverse_u64(uint64_t v)
2ed95849 314{
abc490a1
MD
315 return ((uint64_t) bit_reverse_u8(v) << 56) |
316 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
317 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
318 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
319 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
320 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
321 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
322 ((uint64_t) bit_reverse_u8(v >> 56));
323}
324
325static
326unsigned long bit_reverse_ulong(unsigned long v)
327{
328#if (CAA_BITS_PER_LONG == 32)
329 return bit_reverse_u32(v);
330#else
331 return bit_reverse_u64(v);
332#endif
333}
334
f9830efd 335/*
24365af7
MD
336 * fls: returns the position of the most significant bit.
337 * Returns 0 if no bit is set, else returns the position of the most
338 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 339 */
24365af7
MD
340#if defined(__i386) || defined(__x86_64)
341static inline
342unsigned int fls_u32(uint32_t x)
f9830efd 343{
24365af7
MD
344 int r;
345
346 asm("bsrl %1,%0\n\t"
347 "jnz 1f\n\t"
348 "movl $-1,%0\n\t"
349 "1:\n\t"
350 : "=r" (r) : "rm" (x));
351 return r + 1;
352}
353#define HAS_FLS_U32
354#endif
355
356#if defined(__x86_64)
357static inline
358unsigned int fls_u64(uint64_t x)
359{
360 long r;
361
362 asm("bsrq %1,%0\n\t"
363 "jnz 1f\n\t"
364 "movq $-1,%0\n\t"
365 "1:\n\t"
366 : "=r" (r) : "rm" (x));
367 return r + 1;
368}
369#define HAS_FLS_U64
370#endif
371
372#ifndef HAS_FLS_U64
373static __attribute__((unused))
374unsigned int fls_u64(uint64_t x)
375{
376 unsigned int r = 64;
377
378 if (!x)
379 return 0;
380
381 if (!(x & 0xFFFFFFFF00000000ULL)) {
382 x <<= 32;
383 r -= 32;
384 }
385 if (!(x & 0xFFFF000000000000ULL)) {
386 x <<= 16;
387 r -= 16;
388 }
389 if (!(x & 0xFF00000000000000ULL)) {
390 x <<= 8;
391 r -= 8;
392 }
393 if (!(x & 0xF000000000000000ULL)) {
394 x <<= 4;
395 r -= 4;
396 }
397 if (!(x & 0xC000000000000000ULL)) {
398 x <<= 2;
399 r -= 2;
400 }
401 if (!(x & 0x8000000000000000ULL)) {
402 x <<= 1;
403 r -= 1;
404 }
405 return r;
406}
407#endif
408
409#ifndef HAS_FLS_U32
410static __attribute__((unused))
411unsigned int fls_u32(uint32_t x)
412{
413 unsigned int r = 32;
f9830efd 414
24365af7
MD
415 if (!x)
416 return 0;
417 if (!(x & 0xFFFF0000U)) {
418 x <<= 16;
419 r -= 16;
420 }
421 if (!(x & 0xFF000000U)) {
422 x <<= 8;
423 r -= 8;
424 }
425 if (!(x & 0xF0000000U)) {
426 x <<= 4;
427 r -= 4;
428 }
429 if (!(x & 0xC0000000U)) {
430 x <<= 2;
431 r -= 2;
432 }
433 if (!(x & 0x80000000U)) {
434 x <<= 1;
435 r -= 1;
436 }
437 return r;
438}
439#endif
440
441unsigned int fls_ulong(unsigned long x)
f9830efd 442{
24365af7
MD
443#if (CAA_BITS_PER_lONG == 32)
444 return fls_u32(x);
445#else
446 return fls_u64(x);
447#endif
448}
f9830efd 449
24365af7
MD
450int get_count_order_u32(uint32_t x)
451{
452 int order;
453
454 order = fls_u32(x) - 1;
455 if (x & (x - 1))
456 order++;
457 return order;
458}
459
460int get_count_order_ulong(unsigned long x)
461{
462 int order;
463
464 order = fls_ulong(x) - 1;
465 if (x & (x - 1))
466 order++;
467 return order;
f9830efd
MD
468}
469
98808fb1
MD
470#ifdef POISON_FREE
471#define poison_free(ptr) \
472 do { \
473 memset(ptr, 0x42, sizeof(*(ptr))); \
474 free(ptr); \
475 } while (0)
476#else
477#define poison_free(ptr) free(ptr)
478#endif
479
f9830efd 480static
4105056a 481void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth);
f9830efd 482
df44348d
MD
483/*
484 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
485 * available, then we support hash table item accounting.
486 * In the unfortunate event the number of CPUs reported would be
487 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
488 */
df44348d
MD
489#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
490
f8994aee 491static
4105056a 492void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
f8994aee
MD
493 unsigned long count);
494
df44348d
MD
495static long nr_cpus_mask = -1;
496
497static
498struct ht_items_count *alloc_per_cpu_items_count(void)
499{
500 struct ht_items_count *count;
501
502 switch (nr_cpus_mask) {
503 case -2:
504 return NULL;
505 case -1:
506 {
507 long maxcpus;
508
509 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
510 if (maxcpus <= 0) {
511 nr_cpus_mask = -2;
512 return NULL;
513 }
514 /*
515 * round up number of CPUs to next power of two, so we
516 * can use & for modulo.
517 */
518 maxcpus = 1UL << get_count_order_ulong(maxcpus);
519 nr_cpus_mask = maxcpus - 1;
520 }
521 /* Fall-through */
522 default:
523 return calloc(nr_cpus_mask + 1, sizeof(*count));
524 }
525}
526
527static
528void free_per_cpu_items_count(struct ht_items_count *count)
529{
98808fb1 530 poison_free(count);
df44348d
MD
531}
532
533static
534int ht_get_cpu(void)
535{
536 int cpu;
537
538 assert(nr_cpus_mask >= 0);
539 cpu = sched_getcpu();
540 if (unlikely(cpu < 0))
541 return cpu;
542 else
543 return cpu & nr_cpus_mask;
544}
545
546static
4105056a 547void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d 548{
3171717f 549 unsigned long percpu_count;
df44348d
MD
550 int cpu;
551
552 if (unlikely(!ht->percpu_count))
3171717f 553 return;
df44348d
MD
554 cpu = ht_get_cpu();
555 if (unlikely(cpu < 0))
3171717f
MD
556 return;
557 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
df44348d
MD
558 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
559 unsigned long count;
560
561 dbg_printf("add percpu %lu\n", percpu_count);
562 count = uatomic_add_return(&ht->count,
563 1UL << COUNT_COMMIT_ORDER);
564 /* If power of 2 */
565 if (!(count & (count - 1))) {
4105056a 566 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
f8994aee
MD
567 return;
568 dbg_printf("add set global %lu\n", count);
4105056a 569 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 570 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
571 }
572 }
573}
574
575static
860d07e8 576void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
577{
578 unsigned long percpu_count;
3171717f 579 int cpu;
df44348d 580
3171717f
MD
581 if (unlikely(!ht->percpu_count))
582 return;
583 cpu = ht_get_cpu();
584 if (unlikely(cpu < 0))
585 return;
80d90c06 586 percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
df44348d
MD
587 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
588 unsigned long count;
589
860d07e8 590 dbg_printf("del percpu %lu\n", percpu_count);
df44348d 591 count = uatomic_add_return(&ht->count,
3171717f 592 -(1UL << COUNT_COMMIT_ORDER));
df44348d
MD
593 /* If power of 2 */
594 if (!(count & (count - 1))) {
4105056a 595 if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
f8994aee 596 return;
860d07e8 597 dbg_printf("del set global %lu\n", count);
4105056a 598 cds_lfht_resize_lazy_count(ht, size,
6ea6bc67 599 count >> (CHAIN_LEN_TARGET - 1));
df44348d
MD
600 }
601 }
602}
603
604#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
605
606static const long nr_cpus_mask = -1;
607
608static
609struct ht_items_count *alloc_per_cpu_items_count(void)
610{
611 return NULL;
612}
613
614static
615void free_per_cpu_items_count(struct ht_items_count *count)
616{
617}
618
619static
4105056a 620void ht_count_add(struct cds_lfht *ht, unsigned long size)
df44348d
MD
621{
622}
623
624static
860d07e8 625void ht_count_del(struct cds_lfht *ht, unsigned long size)
df44348d
MD
626{
627}
628
629#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
630
631
f9830efd 632static
4105056a 633void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
f9830efd 634{
f8994aee
MD
635 unsigned long count;
636
b8af5011
MD
637 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
638 return;
f8994aee
MD
639 count = uatomic_read(&ht->count);
640 /*
641 * Use bucket-local length for small table expand and for
642 * environments lacking per-cpu data support.
643 */
644 if (count >= (1UL << COUNT_COMMIT_ORDER))
645 return;
24365af7 646 if (chain_len > 100)
f0c29ed7 647 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 648 chain_len);
3390d470 649 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
4105056a 650 cds_lfht_resize_lazy(ht, size,
01370f0b 651 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
652}
653
abc490a1 654static
14044b37 655struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 656{
14044b37 657 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
658}
659
660static
14044b37 661int is_removed(struct cds_lfht_node *node)
abc490a1 662{
d37166c6 663 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
664}
665
666static
14044b37 667struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 668{
14044b37 669 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
670}
671
f5596c94 672static
14044b37 673int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
674{
675 return ((unsigned long) node) & DUMMY_FLAG;
676}
677
678static
14044b37 679struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 680{
14044b37 681 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94 682}
bb7b2f26
MD
683
684static
685struct cds_lfht_node *get_end(void)
686{
687 return (struct cds_lfht_node *) END_VALUE;
688}
689
690static
691int is_end(struct cds_lfht_node *node)
692{
693 return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
694}
695
abc490a1 696static
f9830efd 697unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
698{
699 unsigned long old1, old2;
700
701 old1 = uatomic_read(ptr);
702 do {
703 old2 = old1;
704 if (old2 >= v)
f9830efd 705 return old2;
abc490a1 706 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 707 return v;
abc490a1
MD
708}
709
1475579c
MD
710static
711void cds_lfht_free_level(struct rcu_head *head)
712{
713 struct rcu_level *l =
714 caa_container_of(head, struct rcu_level, head);
98808fb1 715 poison_free(l);
1475579c
MD
716}
717
273399de
MD
718/*
719 * Remove all logically deleted nodes from a bucket up to a certain node key.
720 */
721static
f9c80341 722void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 723{
14044b37 724 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de 725
c90201ac
MD
726 assert(!is_dummy(dummy));
727 assert(!is_removed(dummy));
728 assert(!is_dummy(node));
729 assert(!is_removed(node));
273399de
MD
730 for (;;) {
731 iter_prev = dummy;
732 /* We can always skip the dummy node initially */
cc4fcb10
MD
733 iter = rcu_dereference(iter_prev->p.next);
734 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
bd4db153
MD
735 /*
736 * We should never be called with dummy (start of chain)
737 * and logically removed node (end of path compression
738 * marker) being the actual same node. This would be a
739 * bug in the algorithm implementation.
740 */
741 assert(dummy != node);
273399de 742 for (;;) {
bb7b2f26 743 if (unlikely(is_end(iter)))
f9c80341 744 return;
76412f24 745 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
f9c80341 746 return;
cc4fcb10 747 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 748 if (likely(is_removed(next)))
273399de 749 break;
b453eae1 750 iter_prev = clear_flag(iter);
273399de
MD
751 iter = next;
752 }
b198f0fd 753 assert(!is_removed(iter));
f5596c94
MD
754 if (is_dummy(iter))
755 new_next = flag_dummy(clear_flag(next));
756 else
757 new_next = clear_flag(next);
48ed1c18
MD
758 if (is_removed(iter))
759 new_next = flag_removed(new_next);
f5596c94 760 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 761 }
f9c80341 762 return;
273399de
MD
763}
764
abc490a1 765static
4105056a
MD
766struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
767 unsigned long size,
768 struct cds_lfht_node *node,
48ed1c18 769 enum add_mode mode, int dummy)
abc490a1 770{
14044b37 771 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
adc0de68 772 *dummy_node, *return_node;
14044b37 773 struct _cds_lfht_node *lookup;
24365af7 774 unsigned long hash, index, order;
abc490a1 775
c90201ac
MD
776 assert(!is_dummy(node));
777 assert(!is_removed(node));
4105056a 778 if (!size) {
f5596c94 779 assert(dummy);
bb7b2f26 780 node->p.next = flag_dummy(get_end());
18117871
MD
781 return node; /* Initial first add (head) */
782 }
cc4fcb10 783 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 784 for (;;) {
adc0de68 785 uint32_t chain_len = 0;
abc490a1 786
11519af6
MD
787 /*
788 * iter_prev points to the non-removed node prior to the
789 * insert location.
11519af6 790 */
4105056a 791 index = hash & (size - 1);
24365af7 792 order = get_count_order_ulong(index + 1);
4105056a 793 lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
14044b37 794 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 795 /* We can always skip the dummy node initially */
cc4fcb10
MD
796 iter = rcu_dereference(iter_prev->p.next);
797 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 798 for (;;) {
bb7b2f26 799 if (unlikely(is_end(iter)))
273399de 800 goto insert;
76412f24 801 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 802 goto insert;
cc4fcb10 803 next = rcu_dereference(clear_flag(iter)->p.next);
b198f0fd 804 if (unlikely(is_removed(next)))
9dba85be 805 goto gc_node;
48ed1c18 806 if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
1b81fe1a 807 && !is_dummy(next)
e43f23f8
MD
808 && !ht->compare_fct(node->key, node->key_len,
809 clear_flag(iter)->key,
48ed1c18
MD
810 clear_flag(iter)->key_len)) {
811 if (mode == ADD_UNIQUE)
812 return clear_flag(iter);
813 else /* mode == ADD_REPLACE */
814 goto replace;
815 }
11519af6 816 /* Only account for identical reverse hash once */
24365af7
MD
817 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
818 && !is_dummy(next))
4105056a 819 check_resize(ht, size, ++chain_len);
11519af6 820 iter_prev = clear_flag(iter);
273399de 821 iter = next;
abc490a1 822 }
48ed1c18 823
273399de 824 insert:
7ec59d3b 825 assert(node != clear_flag(iter));
11519af6 826 assert(!is_removed(iter_prev));
c90201ac 827 assert(!is_removed(iter));
f000907d 828 assert(iter_prev != node);
f9c80341 829 if (!dummy)
1b81fe1a 830 node->p.next = clear_flag(iter);
f9c80341
MD
831 else
832 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
833 if (is_dummy(iter))
834 new_node = flag_dummy(node);
835 else
836 new_node = node;
cc4fcb10 837 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
48ed1c18 838 new_node) != iter) {
273399de 839 continue; /* retry */
48ed1c18
MD
840 } else {
841 if (mode == ADD_REPLACE)
842 return_node = NULL;
843 else /* ADD_DEFAULT and ADD_UNIQUE */
844 return_node = node;
273399de 845 goto gc_end;
48ed1c18
MD
846 }
847
848 replace:
adc0de68
MD
849 /* Insert after node to be replaced */
850 iter_prev = clear_flag(iter);
851 iter = next;
48ed1c18
MD
852 assert(node != clear_flag(iter));
853 assert(!is_removed(iter_prev));
854 assert(!is_removed(iter));
48ed1c18
MD
855 assert(iter_prev != node);
856 assert(!dummy);
adc0de68 857 node->p.next = clear_flag(iter);
48ed1c18
MD
858 if (is_dummy(iter))
859 new_node = flag_dummy(node);
860 else
861 new_node = node;
862 /*
adc0de68
MD
863 * Here is the whole trick for lock-free replace: we add
864 * the replacement node _after_ the node we want to
865 * replace by atomically setting its next pointer at the
b198f0fd 866 * same time we set its removal flag. Given that
adc0de68
MD
867 * the lookups/get next use an iterator aware of the
868 * next pointer, they will either skip the old node due
b198f0fd
MD
869 * to the removal flag and see the new node, or use
870 * the old node, but will not see the new one.
48ed1c18 871 */
adc0de68 872 new_node = flag_removed(new_node);
adc0de68
MD
873 if (uatomic_cmpxchg(&iter_prev->p.next,
874 iter, new_node) != iter) {
48ed1c18
MD
875 continue; /* retry */
876 } else {
adc0de68
MD
877 return_node = iter_prev;
878 goto gc_end;
48ed1c18
MD
879 }
880
9dba85be
MD
881 gc_node:
882 assert(!is_removed(iter));
f5596c94
MD
883 if (is_dummy(iter))
884 new_next = flag_dummy(clear_flag(next));
885 else
886 new_next = clear_flag(next);
887 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 888 /* retry */
464a1ec9 889 }
273399de
MD
890gc_end:
891 /* Garbage collect logically removed nodes in the bucket */
4105056a 892 index = hash & (size - 1);
24365af7 893 order = get_count_order_ulong(index + 1);
4105056a 894 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 895 dummy_node = (struct cds_lfht_node *) lookup;
f9c80341 896 _cds_lfht_gc_bucket(dummy_node, node);
48ed1c18 897 return return_node;
abc490a1 898}
464a1ec9 899
abc490a1 900static
860d07e8 901int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
4105056a 902 struct cds_lfht_node *node,
b198f0fd 903 int dummy_removal)
abc490a1 904{
14044b37
MD
905 struct cds_lfht_node *dummy, *next, *old;
906 struct _cds_lfht_node *lookup;
abc490a1 907 int flagged = 0;
24365af7 908 unsigned long hash, index, order;
5e28c532 909
7ec59d3b 910 /* logically delete the node */
c90201ac
MD
911 assert(!is_dummy(node));
912 assert(!is_removed(node));
cc4fcb10 913 old = rcu_dereference(node->p.next);
7ec59d3b 914 do {
48ed1c18
MD
915 struct cds_lfht_node *new_next;
916
7ec59d3b 917 next = old;
76412f24 918 if (unlikely(is_removed(next)))
7ec59d3b 919 goto end;
1475579c
MD
920 if (dummy_removal)
921 assert(is_dummy(next));
922 else
923 assert(!is_dummy(next));
48ed1c18 924 new_next = flag_removed(next);
48ed1c18 925 old = uatomic_cmpxchg(&node->p.next, next, new_next);
7ec59d3b
MD
926 } while (old != next);
927
928 /* We performed the (logical) deletion. */
929 flagged = 1;
930
931 /*
932 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
933 * the node, and remove it (along with any other logically removed node)
934 * if found.
11519af6 935 */
cc4fcb10 936 hash = bit_reverse_ulong(node->p.reverse_hash);
4105056a
MD
937 assert(size > 0);
938 index = hash & (size - 1);
24365af7 939 order = get_count_order_ulong(index + 1);
4105056a 940 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
14044b37 941 dummy = (struct cds_lfht_node *) lookup;
f9c80341 942 _cds_lfht_gc_bucket(dummy, node);
2ed95849 943end:
11519af6
MD
944 /*
945 * Only the flagging action indicated that we (and no other)
946 * removed the node from the hash.
947 */
7ec59d3b 948 if (flagged) {
cc4fcb10 949 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 950 return 0;
7ec59d3b 951 } else
11519af6 952 return -ENOENT;
abc490a1 953}
2ed95849 954
b7d619b0
MD
955static
956void *partition_resize_thread(void *arg)
957{
958 struct partition_resize_work *work = arg;
959
960 work->ht->cds_lfht_rcu_register_thread();
961 work->fct(work->ht, work->i, work->start, work->len);
962 work->ht->cds_lfht_rcu_unregister_thread();
963 return NULL;
964}
965
966static
967void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
968 unsigned long len,
969 void (*fct)(struct cds_lfht *ht, unsigned long i,
970 unsigned long start, unsigned long len))
971{
972 unsigned long partition_len;
973 struct partition_resize_work *work;
6083a889
MD
974 int thread, ret;
975 unsigned long nr_threads;
b7d619b0
MD
976 pthread_t *thread_id;
977
6083a889
MD
978 /*
979 * Note: nr_cpus_mask + 1 is always power of 2.
980 * We spawn just the number of threads we need to satisfy the minimum
981 * partition size, up to the number of CPUs in the system.
982 */
983 nr_threads = min(nr_cpus_mask + 1,
984 len >> MIN_PARTITION_PER_THREAD_ORDER);
985 partition_len = len >> get_count_order_ulong(nr_threads);
986 work = calloc(nr_threads, sizeof(*work));
987 thread_id = calloc(nr_threads, sizeof(*thread_id));
b7d619b0 988 assert(work);
6083a889
MD
989 for (thread = 0; thread < nr_threads; thread++) {
990 work[thread].ht = ht;
991 work[thread].i = i;
992 work[thread].len = partition_len;
993 work[thread].start = thread * partition_len;
994 work[thread].fct = fct;
995 ret = pthread_create(&thread_id[thread], ht->resize_attr,
996 partition_resize_thread, &work[thread]);
b7d619b0
MD
997 assert(!ret);
998 }
6083a889
MD
999 for (thread = 0; thread < nr_threads; thread++) {
1000 ret = pthread_join(thread_id[thread], NULL);
b7d619b0
MD
1001 assert(!ret);
1002 }
1003 free(work);
1004 free(thread_id);
1005}
1006
e8de508e
MD
1007/*
1008 * Holding RCU read lock to protect _cds_lfht_add against memory
1009 * reclaim that could be performed by other call_rcu worker threads (ABA
1010 * problem).
9ee0fc9a 1011 *
b7d619b0 1012 * When we reach a certain length, we can split this population phase over
9ee0fc9a
MD
1013 * many worker threads, based on the number of CPUs available in the system.
1014 * This should therefore take care of not having the expand lagging behind too
1015 * many concurrent insertion threads by using the scheduler's ability to
1016 * schedule dummy node population fairly with insertions.
e8de508e 1017 */
4105056a 1018static
b7d619b0
MD
1019void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
1020 unsigned long start, unsigned long len)
4105056a
MD
1021{
1022 unsigned long j;
1023
1024 ht->cds_lfht_rcu_read_lock();
b7d619b0 1025 for (j = start; j < start + len; j++) {
4105056a
MD
1026 struct cds_lfht_node *new_node =
1027 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1028
dc1da8f6 1029 dbg_printf("init populate: i %lu j %lu hash %lu\n",
4105056a 1030 i, j, !i ? 0 : (1UL << (i - 1)) + j);
dc1da8f6
MD
1031 new_node->p.reverse_hash =
1032 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
4105056a 1033 (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
48ed1c18 1034 new_node, ADD_DEFAULT, 1);
4105056a
MD
1035 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1036 break;
1037 }
1038 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1039}
1040
1041static
1042void init_table_populate(struct cds_lfht *ht, unsigned long i,
1043 unsigned long len)
1044{
1045 assert(nr_cpus_mask != -1);
6083a889 1046 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1047 ht->cds_lfht_rcu_thread_online();
1048 init_table_populate_partition(ht, i, 0, len);
1049 ht->cds_lfht_rcu_thread_offline();
1050 return;
1051 }
1052 partition_resize_helper(ht, i, len, init_table_populate_partition);
4105056a
MD
1053}
1054
abc490a1 1055static
4105056a 1056void init_table(struct cds_lfht *ht,
24365af7
MD
1057 unsigned long first_order, unsigned long len_order)
1058{
1059 unsigned long i, end_order;
1060
f0c29ed7 1061 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
1062 first_order, first_order + len_order);
1063 end_order = first_order + len_order;
24365af7 1064 for (i = first_order; i < end_order; i++) {
4105056a 1065 unsigned long len;
24365af7
MD
1066
1067 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 1068 dbg_printf("init order %lu len: %lu\n", i, len);
4d676753
MD
1069
1070 /* Stop expand if the resize target changes under us */
1071 if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
1072 break;
1073
4105056a 1074 ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
1475579c 1075 + (len * sizeof(struct _cds_lfht_node)));
b7d619b0 1076 assert(ht->t.tbl[i]);
4105056a 1077
4105056a 1078 /*
dc1da8f6
MD
1079 * Set all dummy nodes reverse hash values for a level and
1080 * link all dummy nodes into the table.
4105056a 1081 */
dc1da8f6 1082 init_table_populate(ht, i, len);
4105056a 1083
f9c80341
MD
1084 /*
1085 * Update table size.
1086 */
1087 cmm_smp_wmb(); /* populate data before RCU size */
1088 CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
1089
4105056a
MD
1090 dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
1091 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1092 break;
1093 }
1094}
1095
e8de508e
MD
1096/*
1097 * Holding RCU read lock to protect _cds_lfht_remove against memory
1098 * reclaim that could be performed by other call_rcu worker threads (ABA
1099 * problem).
1100 * For a single level, we logically remove and garbage collect each node.
1101 *
1102 * As a design choice, we perform logical removal and garbage collection on a
1103 * node-per-node basis to simplify this algorithm. We also assume keeping good
1104 * cache locality of the operation would overweight possible performance gain
1105 * that could be achieved by batching garbage collection for multiple levels.
1106 * However, this would have to be justified by benchmarks.
1107 *
1108 * Concurrent removal and add operations are helping us perform garbage
1109 * collection of logically removed nodes. We guarantee that all logically
1110 * removed nodes have been garbage-collected (unlinked) before call_rcu is
1111 * invoked to free a hole level of dummy nodes (after a grace period).
1112 *
1113 * Logical removal and garbage collection can therefore be done in batch or on a
1114 * node-per-node basis, as long as the guarantee above holds.
9ee0fc9a 1115 *
b7d619b0
MD
1116 * When we reach a certain length, we can split this removal over many worker
1117 * threads, based on the number of CPUs available in the system. This should
1118 * take care of not letting resize process lag behind too many concurrent
9ee0fc9a 1119 * updater threads actively inserting into the hash table.
e8de508e 1120 */
4105056a 1121static
b7d619b0
MD
1122void remove_table_partition(struct cds_lfht *ht, unsigned long i,
1123 unsigned long start, unsigned long len)
4105056a
MD
1124{
1125 unsigned long j;
1126
1127 ht->cds_lfht_rcu_read_lock();
b7d619b0 1128 for (j = start; j < start + len; j++) {
4105056a
MD
1129 struct cds_lfht_node *fini_node =
1130 (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
1131
1132 dbg_printf("remove entry: i %lu j %lu hash %lu\n",
1133 i, j, !i ? 0 : (1UL << (i - 1)) + j);
1134 fini_node->p.reverse_hash =
1135 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
860d07e8 1136 (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
b198f0fd 1137 fini_node, 1);
33c7c748
MD
1138 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1139 break;
abc490a1 1140 }
4105056a 1141 ht->cds_lfht_rcu_read_unlock();
b7d619b0
MD
1142}
1143
1144static
1145void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
1146{
1147
1148 assert(nr_cpus_mask != -1);
6083a889 1149 if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
b7d619b0
MD
1150 ht->cds_lfht_rcu_thread_online();
1151 remove_table_partition(ht, i, 0, len);
1152 ht->cds_lfht_rcu_thread_offline();
1153 return;
1154 }
1155 partition_resize_helper(ht, i, len, remove_table_partition);
2ed95849
MD
1156}
1157
1475579c 1158static
4105056a 1159void fini_table(struct cds_lfht *ht,
1475579c
MD
1160 unsigned long first_order, unsigned long len_order)
1161{
1162 long i, end_order;
1163
1164 dbg_printf("fini table: first_order %lu end_order %lu\n",
1165 first_order, first_order + len_order);
1166 end_order = first_order + len_order;
1167 assert(first_order > 0);
1475579c 1168 for (i = end_order - 1; i >= first_order; i--) {
4105056a 1169 unsigned long len;
1475579c
MD
1170
1171 len = !i ? 1 : 1UL << (i - 1);
1172 dbg_printf("fini order %lu len: %lu\n", i, len);
4105056a 1173
4d676753
MD
1174 /* Stop shrink if the resize target changes under us */
1175 if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
1176 break;
1177
1178 cmm_smp_wmb(); /* populate data before RCU size */
1179 CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
1180
1181 /*
1182 * We need to wait for all add operations to reach Q.S. (and
1183 * thus use the new table for lookups) before we can start
1184 * releasing the old dummy nodes. Otherwise their lookup will
1185 * return a logically removed node as insert position.
1186 */
1187 ht->cds_lfht_synchronize_rcu();
1188
21263e21 1189 /*
4105056a
MD
1190 * Set "removed" flag in dummy nodes about to be removed.
1191 * Unlink all now-logically-removed dummy node pointers.
1192 * Concurrent add/remove operation are helping us doing
1193 * the gc.
21263e21 1194 */
4105056a
MD
1195 remove_table(ht, i, len);
1196
1197 ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
1198
1199 dbg_printf("fini new size: %lu\n", 1UL << i);
1475579c
MD
1200 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
1201 break;
1202 }
1475579c
MD
1203}
1204
7a9dcf9b 1205struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
14044b37
MD
1206 cds_lfht_compare_fct compare_fct,
1207 unsigned long hash_seed,
1208 unsigned long init_size,
b8af5011 1209 int flags,
14044b37 1210 void (*cds_lfht_call_rcu)(struct rcu_head *head,
1475579c 1211 void (*func)(struct rcu_head *head)),
01dbfa62
MD
1212 void (*cds_lfht_synchronize_rcu)(void),
1213 void (*cds_lfht_rcu_read_lock)(void),
5f511391
MD
1214 void (*cds_lfht_rcu_read_unlock)(void),
1215 void (*cds_lfht_rcu_thread_offline)(void),
b7d619b0
MD
1216 void (*cds_lfht_rcu_thread_online)(void),
1217 void (*cds_lfht_rcu_register_thread)(void),
1218 void (*cds_lfht_rcu_unregister_thread)(void),
1219 pthread_attr_t *attr)
abc490a1 1220{
14044b37 1221 struct cds_lfht *ht;
24365af7 1222 unsigned long order;
abc490a1 1223
8129be4e 1224 /* init_size must be power of two */
49619ea0 1225 if (init_size && (init_size & (init_size - 1)))
8129be4e 1226 return NULL;
14044b37 1227 ht = calloc(1, sizeof(struct cds_lfht));
b7d619b0 1228 assert(ht);
abc490a1 1229 ht->hash_fct = hash_fct;
732ad076
MD
1230 ht->compare_fct = compare_fct;
1231 ht->hash_seed = hash_seed;
14044b37 1232 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
1475579c 1233 ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
01dbfa62
MD
1234 ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
1235 ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
5f511391
MD
1236 ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
1237 ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
b7d619b0
MD
1238 ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
1239 ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
1240 ht->resize_attr = attr;
df44348d 1241 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
1242 /* this mutex should not nest in read-side C.S. */
1243 pthread_mutex_init(&ht->resize_mutex, NULL);
cd95516d 1244 order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1;
b8af5011 1245 ht->flags = flags;
5f511391 1246 ht->cds_lfht_rcu_thread_offline();
f000907d 1247 pthread_mutex_lock(&ht->resize_mutex);
4d676753 1248 ht->t.resize_target = 1UL << (order - 1);
4105056a 1249 init_table(ht, 0, order);
f000907d 1250 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1251 ht->cds_lfht_rcu_thread_online();
abc490a1
MD
1252 return ht;
1253}
1254
adc0de68
MD
1255void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
1256 struct cds_lfht_iter *iter)
2ed95849 1257{
bb7b2f26 1258 struct cds_lfht_node *node, *next, *dummy_node;
14044b37 1259 struct _cds_lfht_node *lookup;
4105056a 1260 unsigned long hash, reverse_hash, index, order, size;
2ed95849 1261
732ad076 1262 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 1263 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 1264
4105056a
MD
1265 size = rcu_dereference(ht->t.size);
1266 index = hash & (size - 1);
24365af7 1267 order = get_count_order_ulong(index + 1);
4105056a 1268 lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
f0c29ed7 1269 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
554c284e 1270 hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
bb7b2f26
MD
1271 dummy_node = (struct cds_lfht_node *) lookup;
1272 /* We can always skip the dummy node initially */
1273 node = rcu_dereference(dummy_node->p.next);
bb7b2f26 1274 node = clear_flag(node);
2ed95849 1275 for (;;) {
bb7b2f26
MD
1276 if (unlikely(is_end(node))) {
1277 node = NULL;
abc490a1 1278 break;
bb7b2f26 1279 }
cc4fcb10 1280 if (unlikely(node->p.reverse_hash > reverse_hash)) {
abc490a1
MD
1281 node = NULL;
1282 break;
2ed95849 1283 }
1b81fe1a 1284 next = rcu_dereference(node->p.next);
adc0de68 1285 if (likely(!is_removed(next))
1b81fe1a 1286 && !is_dummy(next)
49c2e2d6 1287 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 1288 break;
2ed95849 1289 }
1b81fe1a 1290 node = clear_flag(next);
2ed95849 1291 }
1b81fe1a 1292 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1293 iter->node = node;
1294 iter->next = next;
abc490a1 1295}
e0ba718a 1296
adc0de68 1297void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
a481e5ff 1298{
adc0de68 1299 struct cds_lfht_node *node, *next;
a481e5ff
MD
1300 unsigned long reverse_hash;
1301 void *key;
1302 size_t key_len;
1303
adc0de68 1304 node = iter->node;
a481e5ff
MD
1305 reverse_hash = node->p.reverse_hash;
1306 key = node->key;
1307 key_len = node->key_len;
adc0de68 1308 next = iter->next;
a481e5ff
MD
1309 node = clear_flag(next);
1310
1311 for (;;) {
bb7b2f26
MD
1312 if (unlikely(is_end(node))) {
1313 node = NULL;
a481e5ff 1314 break;
bb7b2f26 1315 }
a481e5ff
MD
1316 if (unlikely(node->p.reverse_hash > reverse_hash)) {
1317 node = NULL;
1318 break;
1319 }
1320 next = rcu_dereference(node->p.next);
adc0de68 1321 if (likely(!is_removed(next))
a481e5ff
MD
1322 && !is_dummy(next)
1323 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
1324 break;
1325 }
1326 node = clear_flag(next);
1327 }
1328 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
adc0de68
MD
1329 iter->node = node;
1330 iter->next = next;
a481e5ff
MD
1331}
1332
14044b37 1333void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1 1334{
4105056a 1335 unsigned long hash, size;
ab7d5fc6 1336
49c2e2d6 1337 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1338 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 1339
4105056a 1340 size = rcu_dereference(ht->t.size);
48ed1c18 1341 (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
4105056a 1342 ht_count_add(ht, size);
3eca1b8c
MD
1343}
1344
14044b37 1345struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
48ed1c18 1346 struct cds_lfht_node *node)
3eca1b8c 1347{
4105056a 1348 unsigned long hash, size;
df44348d 1349 struct cds_lfht_node *ret;
3eca1b8c 1350
49c2e2d6 1351 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 1352 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c 1353
4105056a 1354 size = rcu_dereference(ht->t.size);
48ed1c18 1355 ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
17f31d1b 1356 if (ret == node)
4105056a 1357 ht_count_add(ht, size);
df44348d 1358 return ret;
2ed95849
MD
1359}
1360
48ed1c18
MD
1361struct cds_lfht_node *cds_lfht_replace(struct cds_lfht *ht,
1362 struct cds_lfht_node *node)
1363{
1364 unsigned long hash, size;
1365 struct cds_lfht_node *ret;
1366
1367 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
1368 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
1369
1370 size = rcu_dereference(ht->t.size);
1371 ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
1372 if (ret == NULL)
1373 ht_count_add(ht, size);
1374 return ret;
1375}
1376
860d07e8 1377int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 1378{
4105056a 1379 unsigned long size;
df44348d 1380 int ret;
abc490a1 1381
4105056a 1382 size = rcu_dereference(ht->t.size);
b198f0fd 1383 ret = _cds_lfht_del(ht, size, node, 0);
df44348d 1384 if (!ret)
860d07e8 1385 ht_count_del(ht, size);
df44348d 1386 return ret;
2ed95849 1387}
ab7d5fc6 1388
abc490a1 1389static
14044b37 1390int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 1391{
14044b37
MD
1392 struct cds_lfht_node *node;
1393 struct _cds_lfht_node *lookup;
4105056a 1394 unsigned long order, i, size;
674f7a69 1395
abc490a1 1396 /* Check that the table is empty */
4105056a 1397 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1398 node = (struct cds_lfht_node *) lookup;
abc490a1 1399 do {
1b81fe1a
MD
1400 node = clear_flag(node)->p.next;
1401 if (!is_dummy(node))
abc490a1 1402 return -EPERM;
273399de 1403 assert(!is_removed(node));
bb7b2f26 1404 } while (!is_end(node));
4105056a
MD
1405 /*
1406 * size accessed without rcu_dereference because hash table is
1407 * being destroyed.
1408 */
1409 size = ht->t.size;
abc490a1 1410 /* Internal sanity check: all nodes left should be dummy */
4105056a 1411 for (order = 0; order < get_count_order_ulong(size) + 1; order++) {
24365af7
MD
1412 unsigned long len;
1413
1414 len = !order ? 1 : 1UL << (order - 1);
1415 for (i = 0; i < len; i++) {
f0c29ed7 1416 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7 1417 order, i,
4105056a
MD
1418 bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
1419 assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
24365af7 1420 }
4105056a 1421 poison_free(ht->t.tbl[order]);
674f7a69 1422 }
abc490a1 1423 return 0;
674f7a69
MD
1424}
1425
1426/*
1427 * Should only be called when no more concurrent readers nor writers can
1428 * possibly access the table.
1429 */
b7d619b0 1430int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
674f7a69 1431{
5e28c532
MD
1432 int ret;
1433
848d4088 1434 /* Wait for in-flight resize operations to complete */
33c7c748 1435 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
1436 while (uatomic_read(&ht->in_progress_resize))
1437 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 1438 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
1439 if (ret)
1440 return ret;
df44348d 1441 free_per_cpu_items_count(ht->percpu_count);
b7d619b0
MD
1442 if (attr)
1443 *attr = ht->resize_attr;
98808fb1 1444 poison_free(ht);
5e28c532 1445 return ret;
674f7a69
MD
1446}
1447
14044b37 1448void cds_lfht_count_nodes(struct cds_lfht *ht,
273399de
MD
1449 unsigned long *count,
1450 unsigned long *removed)
1451{
14044b37
MD
1452 struct cds_lfht_node *node, *next;
1453 struct _cds_lfht_node *lookup;
24365af7 1454 unsigned long nr_dummy = 0;
273399de
MD
1455
1456 *count = 0;
1457 *removed = 0;
1458
24365af7 1459 /* Count non-dummy nodes in the table */
4105056a 1460 lookup = &ht->t.tbl[0]->nodes[0];
14044b37 1461 node = (struct cds_lfht_node *) lookup;
273399de 1462 do {
cc4fcb10 1463 next = rcu_dereference(node->p.next);
b198f0fd 1464 if (is_removed(next)) {
1b81fe1a 1465 assert(!is_dummy(next));
273399de 1466 (*removed)++;
1b81fe1a 1467 } else if (!is_dummy(next))
273399de 1468 (*count)++;
24365af7
MD
1469 else
1470 (nr_dummy)++;
273399de 1471 node = clear_flag(next);
bb7b2f26 1472 } while (!is_end(node));
f0c29ed7 1473 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
273399de
MD
1474}
1475
1475579c 1476/* called with resize mutex held */
abc490a1 1477static
4105056a 1478void _do_cds_lfht_grow(struct cds_lfht *ht,
1475579c 1479 unsigned long old_size, unsigned long new_size)
abc490a1 1480{
1475579c 1481 unsigned long old_order, new_order;
1475579c
MD
1482
1483 old_order = get_count_order_ulong(old_size) + 1;
1484 new_order = get_count_order_ulong(new_size) + 1;
1485 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
1486 old_size, old_order, new_size, new_order);
1475579c 1487 assert(new_size > old_size);
4105056a 1488 init_table(ht, old_order, new_order - old_order);
abc490a1
MD
1489}
1490
1491/* called with resize mutex held */
1492static
4105056a 1493void _do_cds_lfht_shrink(struct cds_lfht *ht,
1475579c 1494 unsigned long old_size, unsigned long new_size)
464a1ec9 1495{
1475579c 1496 unsigned long old_order, new_order;
464a1ec9 1497
cd95516d 1498 new_size = max(new_size, MIN_TABLE_SIZE);
24365af7 1499 old_order = get_count_order_ulong(old_size) + 1;
24365af7 1500 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1501 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1502 old_size, old_order, new_size, new_order);
1475579c 1503 assert(new_size < old_size);
1475579c 1504
4105056a
MD
1505 /* Remove and unlink all dummy nodes to remove. */
1506 fini_table(ht, new_order, old_order - new_order);
464a1ec9
MD
1507}
1508
1475579c
MD
1509
1510/* called with resize mutex held */
1511static
1512void _do_cds_lfht_resize(struct cds_lfht *ht)
1513{
1514 unsigned long new_size, old_size;
4105056a
MD
1515
1516 /*
1517 * Resize table, re-do if the target size has changed under us.
1518 */
1519 do {
1520 ht->t.resize_initiated = 1;
1521 old_size = ht->t.size;
1522 new_size = CMM_LOAD_SHARED(ht->t.resize_target);
1523 if (old_size < new_size)
1524 _do_cds_lfht_grow(ht, old_size, new_size);
1525 else if (old_size > new_size)
1526 _do_cds_lfht_shrink(ht, old_size, new_size);
1527 ht->t.resize_initiated = 0;
1528 /* write resize_initiated before read resize_target */
1529 cmm_smp_mb();
4d676753 1530 } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
1475579c
MD
1531}
1532
abc490a1 1533static
4105056a 1534unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
f9830efd 1535 int growth_order)
464a1ec9 1536{
4105056a
MD
1537 return _uatomic_max(&ht->t.resize_target,
1538 size << growth_order);
464a1ec9
MD
1539}
1540
1475579c 1541static
4105056a 1542void resize_target_update_count(struct cds_lfht *ht,
b8af5011 1543 unsigned long count)
1475579c 1544{
cd95516d 1545 count = max(count, MIN_TABLE_SIZE);
4105056a 1546 uatomic_set(&ht->t.resize_target, count);
1475579c
MD
1547}
1548
1549void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size)
464a1ec9 1550{
4105056a
MD
1551 resize_target_update_count(ht, new_size);
1552 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
5f511391 1553 ht->cds_lfht_rcu_thread_offline();
1475579c
MD
1554 pthread_mutex_lock(&ht->resize_mutex);
1555 _do_cds_lfht_resize(ht);
1556 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1557 ht->cds_lfht_rcu_thread_online();
abc490a1 1558}
464a1ec9 1559
abc490a1
MD
1560static
1561void do_resize_cb(struct rcu_head *head)
1562{
1563 struct rcu_resize_work *work =
1564 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1565 struct cds_lfht *ht = work->ht;
abc490a1 1566
5f511391 1567 ht->cds_lfht_rcu_thread_offline();
abc490a1 1568 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1569 _do_cds_lfht_resize(ht);
abc490a1 1570 pthread_mutex_unlock(&ht->resize_mutex);
5f511391 1571 ht->cds_lfht_rcu_thread_online();
98808fb1 1572 poison_free(work);
848d4088
MD
1573 cmm_smp_mb(); /* finish resize before decrement */
1574 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1575}
1576
abc490a1 1577static
4105056a 1578void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
ab7d5fc6 1579{
abc490a1 1580 struct rcu_resize_work *work;
f9830efd 1581 unsigned long target_size;
abc490a1 1582
4105056a
MD
1583 target_size = resize_target_update(ht, size, growth);
1584 /* Store resize_target before read resize_initiated */
1585 cmm_smp_mb();
1586 if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
848d4088
MD
1587 uatomic_inc(&ht->in_progress_resize);
1588 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1589 work = malloc(sizeof(*work));
1590 work->ht = ht;
14044b37 1591 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1592 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
f9830efd 1593 }
ab7d5fc6 1594}
3171717f 1595
f8994aee
MD
1596#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
1597
3171717f 1598static
4105056a 1599void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
3171717f
MD
1600 unsigned long count)
1601{
1602 struct rcu_resize_work *work;
3171717f 1603
b8af5011
MD
1604 if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
1605 return;
4105056a
MD
1606 resize_target_update_count(ht, count);
1607 /* Store resize_target before read resize_initiated */
1608 cmm_smp_mb();
1609 if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
3171717f
MD
1610 uatomic_inc(&ht->in_progress_resize);
1611 cmm_smp_mb(); /* increment resize count before calling it */
1612 work = malloc(sizeof(*work));
1613 work->ht = ht;
1614 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
4105056a 1615 CMM_STORE_SHARED(ht->t.resize_initiated, 1);
3171717f
MD
1616 }
1617}
f8994aee
MD
1618
1619#endif
This page took 0.183963 seconds and 4 git commands to generate.