rculfhash: count nodes (per-cpu) as expand criterion
[urcu.git] / rculfhash.c
CommitLineData
5e28c532 1/*
abc490a1
MD
2 * rculfhash.c
3 *
4 * Userspace RCU library - Lock-Free Expandable RCU Hash Table
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 *
32 * Some specificities of this Lock-Free Expandable RCU Hash Table
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.
51 * - The resize operation only allows expanding the hash table.
52 * It is triggered either through an API call or automatically by
53 * detecting long chains in the add operation.
54 * - Resize operation initiated by long chain detection is executed by a
55 * call_rcu thread, which keeps lock-freedom of add and remove.
56 * - Resize operations are protected by a mutex.
57 * - The removal operation is split in two parts: first, a "removed"
58 * flag is set in the next pointer within the node to remove. Then,
59 * a "garbage collection" is performed in the bucket containing the
60 * removed node (from the start of the bucket up to the removed node).
61 * All encountered nodes with "removed" flag set in their next
62 * pointers are removed from the linked-list. If the cmpxchg used for
63 * removal fails (due to concurrent garbage-collection or concurrent
64 * add), we retry from the beginning of the bucket. This ensures that
65 * the node with "removed" flag set is removed from the hash table
66 * (not visible to lookups anymore) before the RCU read-side critical
67 * section held across removal ends. Furthermore, this ensures that
68 * the node with "removed" flag set is removed from the linked-list
69 * before its memory is reclaimed. Only the thread which removal
70 * successfully set the "removed" flag (with a cmpxchg) into a node's
71 * next pointer is considered to have succeeded its removal (and thus
72 * owns the node to reclaim). Because we garbage-collect starting from
73 * an invariant node (the start-of-bucket dummy node) up to the
74 * "removed" node (or find a reverse-hash that is higher), we are sure
75 * that a successful traversal of the chain leads to a chain that is
76 * present in the linked-list (the start node is never removed) and
77 * that is does not contain the "removed" node anymore, even if
78 * concurrent delete/add operations are changing the structure of the
79 * list concurrently.
29e669f6
MD
80 * - The add operation performs gargage collection of buckets if it
81 * encounters nodes with removed flag set in the bucket where it wants
82 * to add its new node. This ensures lock-freedom of add operation by
83 * helping the remover unlink nodes from the list rather than to wait
84 * for it do to so.
e753ff5a
MD
85 * - A RCU "order table" indexed by log2(hash index) is copied and
86 * expanded by the resize operation. This order table allows finding
87 * the "dummy node" tables.
88 * - There is one dummy node table per hash index order. The size of
89 * each dummy node table is half the number of hashes contained in
90 * this order.
91 * - call_rcu is used to garbage-collect the old order table.
92 * - The per-order dummy node tables contain a compact version of the
93 * hash table nodes. These tables are invariant after they are
94 * populated into the hash table.
95 */
96
2ed95849
MD
97#define _LGPL_SOURCE
98#include <stdlib.h>
e0ba718a
MD
99#include <errno.h>
100#include <assert.h>
101#include <stdio.h>
abc490a1 102#include <stdint.h>
f000907d 103#include <string.h>
e0ba718a 104
df44348d 105#include "config.h"
2ed95849 106#include <urcu.h>
abc490a1 107#include <urcu-call-rcu.h>
a42cc659
MD
108#include <urcu/arch.h>
109#include <urcu/uatomic.h>
674f7a69 110#include <urcu/jhash.h>
a42cc659 111#include <urcu/compiler.h>
abc490a1 112#include <urcu/rculfhash.h>
5e28c532 113#include <stdio.h>
464a1ec9 114#include <pthread.h>
44395fb7 115
f9830efd 116#ifdef DEBUG
f0c29ed7 117#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args)
f9830efd 118#else
e753ff5a 119#define dbg_printf(fmt, args...)
f9830efd
MD
120#endif
121
01370f0b
MD
122#define CHAIN_LEN_TARGET 4
123#define CHAIN_LEN_RESIZE_THRESHOLD 8
2ed95849 124
df44348d
MD
125/* Commit counter changes to global counter each 1024 steps */
126#define COUNT_COMMIT_ORDER 10
127
abc490a1
MD
128#ifndef max
129#define max(a, b) ((a) > (b) ? (a) : (b))
130#endif
2ed95849 131
d95bd160
MD
132/*
133 * The removed flag needs to be updated atomically with the pointer.
134 * The dummy flag does not require to be updated atomically with the
135 * pointer, but it is added as a pointer low bit flag to save space.
136 */
d37166c6 137#define REMOVED_FLAG (1UL << 0)
f5596c94
MD
138#define DUMMY_FLAG (1UL << 1)
139#define FLAGS_MASK ((1UL << 2) - 1)
d37166c6 140
df44348d
MD
141struct ht_items_count {
142 unsigned long v;
143} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
144
395270b6 145struct rcu_table {
abc490a1 146 unsigned long size; /* always a power of 2 */
f9830efd 147 unsigned long resize_target;
11519af6 148 int resize_initiated;
abc490a1 149 struct rcu_head head;
14044b37 150 struct _cds_lfht_node *tbl[0];
395270b6
MD
151};
152
14044b37 153struct cds_lfht {
395270b6 154 struct rcu_table *t; /* shared */
14044b37
MD
155 cds_lfht_hash_fct hash_fct;
156 cds_lfht_compare_fct compare_fct;
732ad076 157 unsigned long hash_seed;
464a1ec9 158 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
33c7c748 159 unsigned int in_progress_resize, in_progress_destroy;
14044b37 160 void (*cds_lfht_call_rcu)(struct rcu_head *head,
abc490a1 161 void (*func)(struct rcu_head *head));
df44348d
MD
162 unsigned long count; /* global approximate item count */
163 struct ht_items_count *percpu_count; /* per-cpu item count */
2ed95849
MD
164};
165
abc490a1
MD
166struct rcu_resize_work {
167 struct rcu_head head;
14044b37 168 struct cds_lfht *ht;
abc490a1 169};
2ed95849 170
abc490a1
MD
171/*
172 * Algorithm to reverse bits in a word by lookup table, extended to
173 * 64-bit words.
f9830efd 174 * Source:
abc490a1 175 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 176 * Originally from Public Domain.
abc490a1
MD
177 */
178
179static const uint8_t BitReverseTable256[256] =
2ed95849 180{
abc490a1
MD
181#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
182#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
183#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
184 R6(0), R6(2), R6(1), R6(3)
185};
186#undef R2
187#undef R4
188#undef R6
2ed95849 189
abc490a1
MD
190static
191uint8_t bit_reverse_u8(uint8_t v)
192{
193 return BitReverseTable256[v];
194}
ab7d5fc6 195
abc490a1
MD
196static __attribute__((unused))
197uint32_t bit_reverse_u32(uint32_t v)
198{
199 return ((uint32_t) bit_reverse_u8(v) << 24) |
200 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
201 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
202 ((uint32_t) bit_reverse_u8(v >> 24));
2ed95849
MD
203}
204
abc490a1
MD
205static __attribute__((unused))
206uint64_t bit_reverse_u64(uint64_t v)
2ed95849 207{
abc490a1
MD
208 return ((uint64_t) bit_reverse_u8(v) << 56) |
209 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
210 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
211 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
212 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
213 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
214 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
215 ((uint64_t) bit_reverse_u8(v >> 56));
216}
217
218static
219unsigned long bit_reverse_ulong(unsigned long v)
220{
221#if (CAA_BITS_PER_LONG == 32)
222 return bit_reverse_u32(v);
223#else
224 return bit_reverse_u64(v);
225#endif
226}
227
f9830efd 228/*
24365af7
MD
229 * fls: returns the position of the most significant bit.
230 * Returns 0 if no bit is set, else returns the position of the most
231 * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit).
f9830efd 232 */
24365af7
MD
233#if defined(__i386) || defined(__x86_64)
234static inline
235unsigned int fls_u32(uint32_t x)
f9830efd 236{
24365af7
MD
237 int r;
238
239 asm("bsrl %1,%0\n\t"
240 "jnz 1f\n\t"
241 "movl $-1,%0\n\t"
242 "1:\n\t"
243 : "=r" (r) : "rm" (x));
244 return r + 1;
245}
246#define HAS_FLS_U32
247#endif
248
249#if defined(__x86_64)
250static inline
251unsigned int fls_u64(uint64_t x)
252{
253 long r;
254
255 asm("bsrq %1,%0\n\t"
256 "jnz 1f\n\t"
257 "movq $-1,%0\n\t"
258 "1:\n\t"
259 : "=r" (r) : "rm" (x));
260 return r + 1;
261}
262#define HAS_FLS_U64
263#endif
264
265#ifndef HAS_FLS_U64
266static __attribute__((unused))
267unsigned int fls_u64(uint64_t x)
268{
269 unsigned int r = 64;
270
271 if (!x)
272 return 0;
273
274 if (!(x & 0xFFFFFFFF00000000ULL)) {
275 x <<= 32;
276 r -= 32;
277 }
278 if (!(x & 0xFFFF000000000000ULL)) {
279 x <<= 16;
280 r -= 16;
281 }
282 if (!(x & 0xFF00000000000000ULL)) {
283 x <<= 8;
284 r -= 8;
285 }
286 if (!(x & 0xF000000000000000ULL)) {
287 x <<= 4;
288 r -= 4;
289 }
290 if (!(x & 0xC000000000000000ULL)) {
291 x <<= 2;
292 r -= 2;
293 }
294 if (!(x & 0x8000000000000000ULL)) {
295 x <<= 1;
296 r -= 1;
297 }
298 return r;
299}
300#endif
301
302#ifndef HAS_FLS_U32
303static __attribute__((unused))
304unsigned int fls_u32(uint32_t x)
305{
306 unsigned int r = 32;
f9830efd 307
24365af7
MD
308 if (!x)
309 return 0;
310 if (!(x & 0xFFFF0000U)) {
311 x <<= 16;
312 r -= 16;
313 }
314 if (!(x & 0xFF000000U)) {
315 x <<= 8;
316 r -= 8;
317 }
318 if (!(x & 0xF0000000U)) {
319 x <<= 4;
320 r -= 4;
321 }
322 if (!(x & 0xC0000000U)) {
323 x <<= 2;
324 r -= 2;
325 }
326 if (!(x & 0x80000000U)) {
327 x <<= 1;
328 r -= 1;
329 }
330 return r;
331}
332#endif
333
334unsigned int fls_ulong(unsigned long x)
f9830efd 335{
24365af7
MD
336#if (CAA_BITS_PER_lONG == 32)
337 return fls_u32(x);
338#else
339 return fls_u64(x);
340#endif
341}
f9830efd 342
24365af7
MD
343int get_count_order_u32(uint32_t x)
344{
345 int order;
346
347 order = fls_u32(x) - 1;
348 if (x & (x - 1))
349 order++;
350 return order;
351}
352
353int get_count_order_ulong(unsigned long x)
354{
355 int order;
356
357 order = fls_ulong(x) - 1;
358 if (x & (x - 1))
359 order++;
360 return order;
f9830efd
MD
361}
362
363static
14044b37 364void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth);
f9830efd 365
df44348d
MD
366/*
367 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
368 * available, then we support hash table item accounting.
369 * In the unfortunate event the number of CPUs reported would be
370 * inaccurate, we use modulo arithmetic on the number of CPUs we got.
371 */
372
373#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
374
375static long nr_cpus_mask = -1;
376
377static
378struct ht_items_count *alloc_per_cpu_items_count(void)
379{
380 struct ht_items_count *count;
381
382 switch (nr_cpus_mask) {
383 case -2:
384 return NULL;
385 case -1:
386 {
387 long maxcpus;
388
389 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
390 if (maxcpus <= 0) {
391 nr_cpus_mask = -2;
392 return NULL;
393 }
394 /*
395 * round up number of CPUs to next power of two, so we
396 * can use & for modulo.
397 */
398 maxcpus = 1UL << get_count_order_ulong(maxcpus);
399 nr_cpus_mask = maxcpus - 1;
400 }
401 /* Fall-through */
402 default:
403 return calloc(nr_cpus_mask + 1, sizeof(*count));
404 }
405}
406
407static
408void free_per_cpu_items_count(struct ht_items_count *count)
409{
410 free(count);
411}
412
413static
414int ht_get_cpu(void)
415{
416 int cpu;
417
418 assert(nr_cpus_mask >= 0);
419 cpu = sched_getcpu();
420 if (unlikely(cpu < 0))
421 return cpu;
422 else
423 return cpu & nr_cpus_mask;
424}
425
426static
427unsigned long ht_count_update(struct cds_lfht *ht, unsigned int value)
428{
429 int cpu;
430
431 if (unlikely(!ht->percpu_count))
432 return 0;
433 cpu = ht_get_cpu();
434 if (unlikely(cpu < 0))
435 return 0;
436 return uatomic_add_return(&ht->percpu_count[cpu].v, 1);
437}
438
439static
440void ht_count_add(struct cds_lfht *ht, struct rcu_table *t)
441{
442 unsigned long percpu_count;
443
444 percpu_count = ht_count_update(ht, 1);
445 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
446 unsigned long count;
447
448 dbg_printf("add percpu %lu\n", percpu_count);
449 count = uatomic_add_return(&ht->count,
450 1UL << COUNT_COMMIT_ORDER);
451 /* If power of 2 */
452 if (!(count & (count - 1))) {
453 dbg_printf("add global %lu\n", count);
454 cds_lfht_resize_lazy(ht, t, 1);
455 }
456 }
457}
458
459static
460void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t)
461{
462 unsigned long percpu_count;
463
464 percpu_count = ht_count_update(ht, -1);
465 if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
466 unsigned long count;
467
468 dbg_printf("remove percpu %lu\n", percpu_count);
469 count = uatomic_add_return(&ht->count,
470 1UL << COUNT_COMMIT_ORDER);
471 /* If power of 2 */
472 if (!(count & (count - 1))) {
473 dbg_printf("remove global %lu\n", count);
474 cds_lfht_resize_lazy(ht, t, -1);
475 }
476 }
477}
478
479#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
480
481static const long nr_cpus_mask = -1;
482
483static
484struct ht_items_count *alloc_per_cpu_items_count(void)
485{
486 return NULL;
487}
488
489static
490void free_per_cpu_items_count(struct ht_items_count *count)
491{
492}
493
494static
495void ht_count_add(struct cds_lfht *ht)
496{
497}
498
499static
500void ht_count_remove(struct cds_lfht *ht)
501{
502}
503
504#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
505
506
f9830efd 507static
14044b37 508void check_resize(struct cds_lfht *ht, struct rcu_table *t,
f9830efd
MD
509 uint32_t chain_len)
510{
df44348d 511 return; //TEST
24365af7 512 if (chain_len > 100)
f0c29ed7 513 dbg_printf("WARNING: large chain length: %u.\n",
24365af7 514 chain_len);
3390d470 515 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
14044b37 516 cds_lfht_resize_lazy(ht, t,
01370f0b 517 get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
f9830efd
MD
518}
519
abc490a1 520static
14044b37 521struct cds_lfht_node *clear_flag(struct cds_lfht_node *node)
abc490a1 522{
14044b37 523 return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK);
abc490a1
MD
524}
525
526static
14044b37 527int is_removed(struct cds_lfht_node *node)
abc490a1 528{
d37166c6 529 return ((unsigned long) node) & REMOVED_FLAG;
abc490a1
MD
530}
531
532static
14044b37 533struct cds_lfht_node *flag_removed(struct cds_lfht_node *node)
abc490a1 534{
14044b37 535 return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG);
abc490a1
MD
536}
537
f5596c94 538static
14044b37 539int is_dummy(struct cds_lfht_node *node)
f5596c94
MD
540{
541 return ((unsigned long) node) & DUMMY_FLAG;
542}
543
544static
14044b37 545struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node)
f5596c94 546{
14044b37 547 return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
f5596c94
MD
548}
549
abc490a1 550static
f9830efd 551unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
552{
553 unsigned long old1, old2;
554
555 old1 = uatomic_read(ptr);
556 do {
557 old2 = old1;
558 if (old2 >= v)
f9830efd 559 return old2;
abc490a1 560 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 561 return v;
abc490a1
MD
562}
563
273399de
MD
564/*
565 * Remove all logically deleted nodes from a bucket up to a certain node key.
566 */
567static
14044b37 568void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node)
273399de 569{
14044b37 570 struct cds_lfht_node *iter_prev, *iter, *next, *new_next;
273399de
MD
571
572 for (;;) {
573 iter_prev = dummy;
574 /* We can always skip the dummy node initially */
cc4fcb10
MD
575 iter = rcu_dereference(iter_prev->p.next);
576 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
273399de 577 for (;;) {
a2974903 578 if (unlikely(!clear_flag(iter)))
479c8a32 579 return;
76412f24 580 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 581 return;
cc4fcb10 582 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 583 if (likely(is_removed(next)))
273399de 584 break;
b453eae1 585 iter_prev = clear_flag(iter);
273399de
MD
586 iter = next;
587 }
588 assert(!is_removed(iter));
f5596c94
MD
589 if (is_dummy(iter))
590 new_next = flag_dummy(clear_flag(next));
591 else
592 new_next = clear_flag(next);
593 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de
MD
594 }
595}
596
abc490a1 597static
14044b37
MD
598struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, struct rcu_table *t,
599 struct cds_lfht_node *node, int unique, int dummy)
abc490a1 600{
14044b37 601 struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
f5596c94 602 *dummy_node;
14044b37 603 struct _cds_lfht_node *lookup;
24365af7 604 unsigned long hash, index, order;
abc490a1 605
18117871 606 if (!t->size) {
f5596c94
MD
607 assert(dummy);
608 node->p.next = flag_dummy(NULL);
18117871
MD
609 return node; /* Initial first add (head) */
610 }
cc4fcb10 611 hash = bit_reverse_ulong(node->p.reverse_hash);
abc490a1 612 for (;;) {
f9830efd 613 uint32_t chain_len = 0;
abc490a1 614
11519af6
MD
615 /*
616 * iter_prev points to the non-removed node prior to the
617 * insert location.
11519af6 618 */
24365af7
MD
619 index = hash & (t->size - 1);
620 order = get_count_order_ulong(index + 1);
621 lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
14044b37 622 iter_prev = (struct cds_lfht_node *) lookup;
11519af6 623 /* We can always skip the dummy node initially */
cc4fcb10
MD
624 iter = rcu_dereference(iter_prev->p.next);
625 assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
abc490a1 626 for (;;) {
a2974903 627 if (unlikely(!clear_flag(iter)))
273399de 628 goto insert;
76412f24 629 if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
273399de 630 goto insert;
cc4fcb10 631 next = rcu_dereference(clear_flag(iter)->p.next);
76412f24 632 if (unlikely(is_removed(next)))
9dba85be 633 goto gc_node;
e43f23f8 634 if (unique
1b81fe1a 635 && !is_dummy(next)
e43f23f8
MD
636 && !ht->compare_fct(node->key, node->key_len,
637 clear_flag(iter)->key,
638 clear_flag(iter)->key_len))
18117871 639 return clear_flag(iter);
11519af6 640 /* Only account for identical reverse hash once */
24365af7
MD
641 if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
642 && !is_dummy(next))
11519af6
MD
643 check_resize(ht, t, ++chain_len);
644 iter_prev = clear_flag(iter);
273399de 645 iter = next;
abc490a1 646 }
273399de 647 insert:
7ec59d3b 648 assert(node != clear_flag(iter));
11519af6 649 assert(!is_removed(iter_prev));
f000907d 650 assert(iter_prev != node);
f5596c94 651 if (!dummy)
1b81fe1a 652 node->p.next = clear_flag(iter);
f5596c94 653 else
1b81fe1a 654 node->p.next = flag_dummy(clear_flag(iter));
f5596c94
MD
655 if (is_dummy(iter))
656 new_node = flag_dummy(node);
657 else
658 new_node = node;
cc4fcb10 659 if (uatomic_cmpxchg(&iter_prev->p.next, iter,
f5596c94 660 new_node) != iter)
273399de 661 continue; /* retry */
11519af6 662 else
273399de 663 goto gc_end;
9dba85be
MD
664 gc_node:
665 assert(!is_removed(iter));
f5596c94
MD
666 if (is_dummy(iter))
667 new_next = flag_dummy(clear_flag(next));
668 else
669 new_next = clear_flag(next);
670 (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
273399de 671 /* retry */
464a1ec9 672 }
273399de
MD
673gc_end:
674 /* Garbage collect logically removed nodes in the bucket */
24365af7
MD
675 index = hash & (t->size - 1);
676 order = get_count_order_ulong(index + 1);
677 lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
14044b37
MD
678 dummy_node = (struct cds_lfht_node *) lookup;
679 _cds_lfht_gc_bucket(dummy_node, node);
18117871 680 return node;
abc490a1 681}
464a1ec9 682
abc490a1 683static
14044b37
MD
684int _cds_lfht_remove(struct cds_lfht *ht, struct rcu_table *t,
685 struct cds_lfht_node *node)
abc490a1 686{
14044b37
MD
687 struct cds_lfht_node *dummy, *next, *old;
688 struct _cds_lfht_node *lookup;
abc490a1 689 int flagged = 0;
24365af7 690 unsigned long hash, index, order;
5e28c532 691
7ec59d3b 692 /* logically delete the node */
cc4fcb10 693 old = rcu_dereference(node->p.next);
7ec59d3b
MD
694 do {
695 next = old;
76412f24 696 if (unlikely(is_removed(next)))
7ec59d3b 697 goto end;
1b81fe1a 698 assert(!is_dummy(next));
cc4fcb10 699 old = uatomic_cmpxchg(&node->p.next, next,
7ec59d3b
MD
700 flag_removed(next));
701 } while (old != next);
702
703 /* We performed the (logical) deletion. */
704 flagged = 1;
705
706 /*
707 * Ensure that the node is not visible to readers anymore: lookup for
273399de
MD
708 * the node, and remove it (along with any other logically removed node)
709 * if found.
11519af6 710 */
cc4fcb10 711 hash = bit_reverse_ulong(node->p.reverse_hash);
24365af7
MD
712 index = hash & (t->size - 1);
713 order = get_count_order_ulong(index + 1);
714 lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
14044b37
MD
715 dummy = (struct cds_lfht_node *) lookup;
716 _cds_lfht_gc_bucket(dummy, node);
2ed95849 717end:
11519af6
MD
718 /*
719 * Only the flagging action indicated that we (and no other)
720 * removed the node from the hash.
721 */
7ec59d3b 722 if (flagged) {
cc4fcb10 723 assert(is_removed(rcu_dereference(node->p.next)));
11519af6 724 return 0;
7ec59d3b 725 } else
11519af6 726 return -ENOENT;
abc490a1 727}
2ed95849 728
abc490a1 729static
14044b37 730void init_table(struct cds_lfht *ht, struct rcu_table *t,
24365af7
MD
731 unsigned long first_order, unsigned long len_order)
732{
733 unsigned long i, end_order;
734
f0c29ed7 735 dbg_printf("init table: first_order %lu end_order %lu\n",
24365af7
MD
736 first_order, first_order + len_order);
737 end_order = first_order + len_order;
738 t->size = !first_order ? 0 : (1UL << (first_order - 1));
739 for (i = first_order; i < end_order; i++) {
740 unsigned long j, len;
741
742 len = !i ? 1 : 1UL << (i - 1);
f0c29ed7 743 dbg_printf("init order %lu len: %lu\n", i, len);
14044b37 744 t->tbl[i] = calloc(len, sizeof(struct _cds_lfht_node));
24365af7 745 for (j = 0; j < len; j++) {
f0c29ed7 746 dbg_printf("init entry: i %lu j %lu hash %lu\n",
24365af7 747 i, j, !i ? 0 : (1UL << (i - 1)) + j);
14044b37
MD
748 struct cds_lfht_node *new_node =
749 (struct cds_lfht_node *) &t->tbl[i][j];
24365af7
MD
750 new_node->p.reverse_hash =
751 bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
14044b37 752 (void) _cds_lfht_add(ht, t, new_node, 0, 1);
33c7c748
MD
753 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
754 break;
24365af7
MD
755 }
756 /* Update table size */
757 t->size = !i ? 1 : (1UL << i);
f0c29ed7 758 dbg_printf("init new size: %lu\n", t->size);
33c7c748
MD
759 if (CMM_LOAD_SHARED(ht->in_progress_destroy))
760 break;
abc490a1 761 }
24365af7 762 t->resize_target = t->size;
11519af6 763 t->resize_initiated = 0;
2ed95849
MD
764}
765
14044b37
MD
766struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
767 cds_lfht_compare_fct compare_fct,
768 unsigned long hash_seed,
769 unsigned long init_size,
770 void (*cds_lfht_call_rcu)(struct rcu_head *head,
771 void (*func)(struct rcu_head *head)))
abc490a1 772{
14044b37 773 struct cds_lfht *ht;
24365af7 774 unsigned long order;
abc490a1 775
8129be4e 776 /* init_size must be power of two */
49619ea0 777 if (init_size && (init_size & (init_size - 1)))
8129be4e 778 return NULL;
14044b37 779 ht = calloc(1, sizeof(struct cds_lfht));
abc490a1 780 ht->hash_fct = hash_fct;
732ad076
MD
781 ht->compare_fct = compare_fct;
782 ht->hash_seed = hash_seed;
14044b37 783 ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
848d4088 784 ht->in_progress_resize = 0;
df44348d 785 ht->percpu_count = alloc_per_cpu_items_count();
abc490a1
MD
786 /* this mutex should not nest in read-side C.S. */
787 pthread_mutex_init(&ht->resize_mutex, NULL);
24365af7 788 order = get_count_order_ulong(max(init_size, 1)) + 1;
14044b37
MD
789 ht->t = calloc(1, sizeof(struct cds_lfht)
790 + (order * sizeof(struct _cds_lfht_node *)));
abc490a1 791 ht->t->size = 0;
f000907d 792 pthread_mutex_lock(&ht->resize_mutex);
24365af7 793 init_table(ht, ht->t, 0, order);
f000907d 794 pthread_mutex_unlock(&ht->resize_mutex);
abc490a1
MD
795 return ht;
796}
797
14044b37 798struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
2ed95849 799{
395270b6 800 struct rcu_table *t;
14044b37
MD
801 struct cds_lfht_node *node, *next;
802 struct _cds_lfht_node *lookup;
24365af7 803 unsigned long hash, reverse_hash, index, order;
2ed95849 804
732ad076 805 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 806 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 807
395270b6 808 t = rcu_dereference(ht->t);
24365af7
MD
809 index = hash & (t->size - 1);
810 order = get_count_order_ulong(index + 1);
811 lookup = &t->tbl[order][index & ((1UL << (order - 1)) - 1)];
f0c29ed7 812 dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
24365af7 813 hash, index, order, index & ((1UL << (order - 1)) - 1));
14044b37 814 node = (struct cds_lfht_node *) lookup;
2ed95849 815 for (;;) {
abc490a1
MD
816 if (unlikely(!node))
817 break;
cc4fcb10 818 if (unlikely(node->p.reverse_hash > reverse_hash)) {
abc490a1
MD
819 node = NULL;
820 break;
2ed95849 821 }
1b81fe1a
MD
822 next = rcu_dereference(node->p.next);
823 if (likely(!is_removed(next))
824 && !is_dummy(next)
49c2e2d6 825 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
273399de 826 break;
2ed95849 827 }
1b81fe1a 828 node = clear_flag(next);
2ed95849 829 }
1b81fe1a 830 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
abc490a1
MD
831 return node;
832}
e0ba718a 833
a481e5ff
MD
834struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
835 struct cds_lfht_node *node)
836{
837 struct cds_lfht_node *next;
838 unsigned long reverse_hash;
839 void *key;
840 size_t key_len;
841
842 reverse_hash = node->p.reverse_hash;
843 key = node->key;
844 key_len = node->key_len;
845 next = rcu_dereference(node->p.next);
846 node = clear_flag(next);
847
848 for (;;) {
849 if (unlikely(!node))
850 break;
851 if (unlikely(node->p.reverse_hash > reverse_hash)) {
852 node = NULL;
853 break;
854 }
855 next = rcu_dereference(node->p.next);
856 if (likely(!is_removed(next))
857 && !is_dummy(next)
858 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
859 break;
860 }
861 node = clear_flag(next);
862 }
863 assert(!node || !is_dummy(rcu_dereference(node->p.next)));
864 return node;
865}
866
14044b37 867void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
abc490a1
MD
868{
869 struct rcu_table *t;
49c2e2d6 870 unsigned long hash;
ab7d5fc6 871
49c2e2d6 872 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 873 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
2ed95849 874
abc490a1 875 t = rcu_dereference(ht->t);
14044b37 876 (void) _cds_lfht_add(ht, t, node, 0, 0);
df44348d 877 ht_count_add(ht, t);
3eca1b8c
MD
878}
879
14044b37
MD
880struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
881 struct cds_lfht_node *node)
3eca1b8c
MD
882{
883 struct rcu_table *t;
49c2e2d6 884 unsigned long hash;
df44348d 885 struct cds_lfht_node *ret;
3eca1b8c 886
49c2e2d6 887 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
cc4fcb10 888 node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
3eca1b8c
MD
889
890 t = rcu_dereference(ht->t);
df44348d
MD
891 ret = _cds_lfht_add(ht, t, node, 1, 0);
892 if (ret != node)
893 ht_count_add(ht, t);
894 return ret;
2ed95849
MD
895}
896
14044b37 897int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
2ed95849 898{
abc490a1 899 struct rcu_table *t;
df44348d 900 int ret;
abc490a1
MD
901
902 t = rcu_dereference(ht->t);
df44348d
MD
903 ret = _cds_lfht_remove(ht, t, node);
904 if (!ret)
905 ht_count_remove(ht, t);
906 return ret;
2ed95849 907}
ab7d5fc6 908
abc490a1 909static
14044b37 910int cds_lfht_delete_dummy(struct cds_lfht *ht)
674f7a69 911{
395270b6 912 struct rcu_table *t;
14044b37
MD
913 struct cds_lfht_node *node;
914 struct _cds_lfht_node *lookup;
24365af7 915 unsigned long order, i;
674f7a69 916
abc490a1
MD
917 t = ht->t;
918 /* Check that the table is empty */
24365af7 919 lookup = &t->tbl[0][0];
14044b37 920 node = (struct cds_lfht_node *) lookup;
abc490a1 921 do {
1b81fe1a
MD
922 node = clear_flag(node)->p.next;
923 if (!is_dummy(node))
abc490a1 924 return -EPERM;
273399de 925 assert(!is_removed(node));
a2974903 926 } while (clear_flag(node));
abc490a1 927 /* Internal sanity check: all nodes left should be dummy */
24365af7
MD
928 for (order = 0; order < get_count_order_ulong(t->size) + 1; order++) {
929 unsigned long len;
930
931 len = !order ? 1 : 1UL << (order - 1);
932 for (i = 0; i < len; i++) {
f0c29ed7 933 dbg_printf("delete order %lu i %lu hash %lu\n",
24365af7
MD
934 order, i,
935 bit_reverse_ulong(t->tbl[order][i].reverse_hash));
936 assert(is_dummy(t->tbl[order][i].next));
937 }
938 free(t->tbl[order]);
674f7a69 939 }
abc490a1 940 return 0;
674f7a69
MD
941}
942
943/*
944 * Should only be called when no more concurrent readers nor writers can
945 * possibly access the table.
946 */
14044b37 947int cds_lfht_destroy(struct cds_lfht *ht)
674f7a69 948{
5e28c532
MD
949 int ret;
950
848d4088 951 /* Wait for in-flight resize operations to complete */
33c7c748 952 CMM_STORE_SHARED(ht->in_progress_destroy, 1);
848d4088
MD
953 while (uatomic_read(&ht->in_progress_resize))
954 poll(NULL, 0, 100); /* wait for 100ms */
14044b37 955 ret = cds_lfht_delete_dummy(ht);
abc490a1
MD
956 if (ret)
957 return ret;
395270b6 958 free(ht->t);
df44348d 959 free_per_cpu_items_count(ht->percpu_count);
674f7a69 960 free(ht);
5e28c532 961 return ret;
674f7a69
MD
962}
963
14044b37 964void cds_lfht_count_nodes(struct cds_lfht *ht,
273399de
MD
965 unsigned long *count,
966 unsigned long *removed)
967{
968 struct rcu_table *t;
14044b37
MD
969 struct cds_lfht_node *node, *next;
970 struct _cds_lfht_node *lookup;
24365af7 971 unsigned long nr_dummy = 0;
273399de
MD
972
973 *count = 0;
974 *removed = 0;
975
976 t = rcu_dereference(ht->t);
24365af7
MD
977 /* Count non-dummy nodes in the table */
978 lookup = &t->tbl[0][0];
14044b37 979 node = (struct cds_lfht_node *) lookup;
273399de 980 do {
cc4fcb10 981 next = rcu_dereference(node->p.next);
273399de 982 if (is_removed(next)) {
1b81fe1a 983 assert(!is_dummy(next));
273399de 984 (*removed)++;
1b81fe1a 985 } else if (!is_dummy(next))
273399de 986 (*count)++;
24365af7
MD
987 else
988 (nr_dummy)++;
273399de
MD
989 node = clear_flag(next);
990 } while (node);
f0c29ed7 991 dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
273399de
MD
992}
993
abc490a1 994static
14044b37 995void cds_lfht_free_table_cb(struct rcu_head *head)
abc490a1
MD
996{
997 struct rcu_table *t =
998 caa_container_of(head, struct rcu_table, head);
999 free(t);
1000}
1001
1002/* called with resize mutex held */
1003static
14044b37 1004void _do_cds_lfht_resize(struct cds_lfht *ht)
464a1ec9 1005{
24365af7 1006 unsigned long new_size, old_size, old_order, new_order;
395270b6 1007 struct rcu_table *new_t, *old_t;
464a1ec9 1008
395270b6
MD
1009 old_t = ht->t;
1010 old_size = old_t->size;
24365af7 1011 old_order = get_count_order_ulong(old_size) + 1;
464a1ec9 1012
f9830efd 1013 new_size = CMM_LOAD_SHARED(old_t->resize_target);
abc490a1 1014 if (old_size == new_size)
464a1ec9 1015 return;
24365af7 1016 new_order = get_count_order_ulong(new_size) + 1;
df44348d 1017 printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
df03fab8 1018 old_size, old_order, new_size, new_order);
14044b37
MD
1019 new_t = malloc(sizeof(struct cds_lfht)
1020 + (new_order * sizeof(struct _cds_lfht_node *)));
f000907d
MD
1021 assert(new_size > old_size);
1022 memcpy(&new_t->tbl, &old_t->tbl,
14044b37 1023 old_order * sizeof(struct _cds_lfht_node *));
24365af7 1024 init_table(ht, new_t, old_order, new_order - old_order);
f000907d
MD
1025 /* Changing table and size atomically wrt lookups */
1026 rcu_assign_pointer(ht->t, new_t);
14044b37 1027 ht->cds_lfht_call_rcu(&old_t->head, cds_lfht_free_table_cb);
464a1ec9
MD
1028}
1029
abc490a1 1030static
f9830efd
MD
1031unsigned long resize_target_update(struct rcu_table *t,
1032 int growth_order)
464a1ec9 1033{
f9830efd
MD
1034 return _uatomic_max(&t->resize_target,
1035 t->size << growth_order);
464a1ec9
MD
1036}
1037
14044b37 1038void cds_lfht_resize(struct cds_lfht *ht, int growth)
464a1ec9 1039{
f9830efd
MD
1040 struct rcu_table *t = rcu_dereference(ht->t);
1041 unsigned long target_size;
1042
f0c29ed7
MD
1043 if (growth < 0) {
1044 /*
1045 * Silently refuse to shrink hash table. (not supported)
1046 */
1047 dbg_printf("shrinking hash table not supported.\n");
1048 return;
1049 }
1050
f9830efd
MD
1051 target_size = resize_target_update(t, growth);
1052 if (t->size < target_size) {
11519af6 1053 CMM_STORE_SHARED(t->resize_initiated, 1);
f9830efd 1054 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1055 _do_cds_lfht_resize(ht);
f9830efd
MD
1056 pthread_mutex_unlock(&ht->resize_mutex);
1057 }
abc490a1 1058}
464a1ec9 1059
abc490a1
MD
1060static
1061void do_resize_cb(struct rcu_head *head)
1062{
1063 struct rcu_resize_work *work =
1064 caa_container_of(head, struct rcu_resize_work, head);
14044b37 1065 struct cds_lfht *ht = work->ht;
abc490a1
MD
1066
1067 pthread_mutex_lock(&ht->resize_mutex);
14044b37 1068 _do_cds_lfht_resize(ht);
abc490a1
MD
1069 pthread_mutex_unlock(&ht->resize_mutex);
1070 free(work);
848d4088
MD
1071 cmm_smp_mb(); /* finish resize before decrement */
1072 uatomic_dec(&ht->in_progress_resize);
464a1ec9
MD
1073}
1074
abc490a1 1075static
14044b37 1076void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth)
ab7d5fc6 1077{
abc490a1 1078 struct rcu_resize_work *work;
f9830efd 1079 unsigned long target_size;
abc490a1 1080
f9830efd 1081 target_size = resize_target_update(t, growth);
11519af6 1082 if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
848d4088
MD
1083 uatomic_inc(&ht->in_progress_resize);
1084 cmm_smp_mb(); /* increment resize count before calling it */
f9830efd
MD
1085 work = malloc(sizeof(*work));
1086 work->ht = ht;
14044b37 1087 ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
11519af6 1088 CMM_STORE_SHARED(t->resize_initiated, 1);
f9830efd 1089 }
ab7d5fc6 1090}
This page took 0.129289 seconds and 4 git commands to generate.