rculfhash: fix add unique
[urcu.git] / rculfhash.c
... / ...
CommitLineData
1/*
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
21 */
22
23#define _LGPL_SOURCE
24#include <stdlib.h>
25#include <errno.h>
26#include <assert.h>
27#include <stdio.h>
28#include <stdint.h>
29#include <string.h>
30
31#include <urcu.h>
32#include <urcu-call-rcu.h>
33#include <urcu/arch.h>
34#include <urcu/uatomic.h>
35#include <urcu/jhash.h>
36#include <urcu/compiler.h>
37#include <urcu/rculfhash.h>
38#include <stdio.h>
39#include <pthread.h>
40
41#define DEBUG /* Test */
42
43#ifdef DEBUG
44#define dbg_printf(args...) printf(args)
45#else
46#define dbg_printf(args...)
47#endif
48
49#define CHAIN_LEN_TARGET 1
50#define CHAIN_LEN_RESIZE_THRESHOLD 2
51
52#ifndef max
53#define max(a, b) ((a) > (b) ? (a) : (b))
54#endif
55
56struct rcu_table {
57 unsigned long size; /* always a power of 2 */
58 unsigned long resize_target;
59 int resize_initiated;
60 struct rcu_head head;
61 struct rcu_ht_node *tbl[0];
62};
63
64struct rcu_ht {
65 struct rcu_table *t; /* shared */
66 ht_hash_fct hash_fct;
67 ht_compare_fct compare_fct;
68 unsigned long hash_seed;
69 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
70 void (*ht_call_rcu)(struct rcu_head *head,
71 void (*func)(struct rcu_head *head));
72};
73
74struct rcu_resize_work {
75 struct rcu_head head;
76 struct rcu_ht *ht;
77};
78
79/*
80 * Algorithm to reverse bits in a word by lookup table, extended to
81 * 64-bit words.
82 * Source:
83 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
84 * Originally from Public Domain.
85 */
86
87static const uint8_t BitReverseTable256[256] =
88{
89#define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
90#define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
91#define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
92 R6(0), R6(2), R6(1), R6(3)
93};
94#undef R2
95#undef R4
96#undef R6
97
98static
99uint8_t bit_reverse_u8(uint8_t v)
100{
101 return BitReverseTable256[v];
102}
103
104static __attribute__((unused))
105uint32_t bit_reverse_u32(uint32_t v)
106{
107 return ((uint32_t) bit_reverse_u8(v) << 24) |
108 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
109 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
110 ((uint32_t) bit_reverse_u8(v >> 24));
111}
112
113static __attribute__((unused))
114uint64_t bit_reverse_u64(uint64_t v)
115{
116 return ((uint64_t) bit_reverse_u8(v) << 56) |
117 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
118 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
119 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
120 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
121 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
122 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
123 ((uint64_t) bit_reverse_u8(v >> 56));
124}
125
126static
127unsigned long bit_reverse_ulong(unsigned long v)
128{
129#if (CAA_BITS_PER_LONG == 32)
130 return bit_reverse_u32(v);
131#else
132 return bit_reverse_u64(v);
133#endif
134}
135
136/*
137 * Algorithm to find the log2 of a 32-bit unsigned integer.
138 * source: http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogLookup
139 * Originally from Public Domain.
140 */
141static const char LogTable256[256] =
142{
143#define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n
144 -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
145 LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6),
146 LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7)
147};
148
149uint32_t log2_u32(uint32_t v)
150{
151 uint32_t t, tt;
152
153 if ((tt = (v >> 16)))
154 return (t = (tt >> 8))
155 ? 24 + LogTable256[t]
156 : 16 + LogTable256[tt];
157 else
158 return (t = (v >> 8))
159 ? 8 + LogTable256[t]
160 : LogTable256[v];
161}
162
163static
164void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth);
165
166static
167void check_resize(struct rcu_ht *ht, struct rcu_table *t,
168 uint32_t chain_len)
169{
170 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
171 ht_resize_lazy(ht, t,
172 log2_u32(chain_len - CHAIN_LEN_TARGET - 1));
173}
174
175static
176struct rcu_ht_node *clear_flag(struct rcu_ht_node *node)
177{
178 return (struct rcu_ht_node *) (((unsigned long) node) & ~0x1);
179}
180
181static
182int is_removed(struct rcu_ht_node *node)
183{
184 return ((unsigned long) node) & 0x1;
185}
186
187static
188struct rcu_ht_node *flag_removed(struct rcu_ht_node *node)
189{
190 return (struct rcu_ht_node *) (((unsigned long) node) | 0x1);
191}
192
193static
194unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
195{
196 unsigned long old1, old2;
197
198 old1 = uatomic_read(ptr);
199 do {
200 old2 = old1;
201 if (old2 >= v)
202 return old2;
203 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
204 return v;
205}
206
207/*
208 * Remove all logically deleted nodes from a bucket up to a certain node key.
209 */
210static
211void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node)
212{
213 struct rcu_ht_node *iter_prev, *iter, *next;
214
215 for (;;) {
216 iter_prev = dummy;
217 /* We can always skip the dummy node initially */
218 iter = rcu_dereference(iter_prev->next);
219 assert(iter_prev->reverse_hash <= node->reverse_hash);
220 for (;;) {
221 if (unlikely(!iter))
222 return;
223 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
224 return;
225 next = rcu_dereference(clear_flag(iter)->next);
226 if (is_removed(next))
227 break;
228 iter_prev = iter;
229 iter = next;
230 }
231 assert(!is_removed(iter));
232 (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next));
233 }
234}
235
236static
237int _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node,
238 int unique)
239{
240 struct rcu_ht_node *iter_prev, *dummy, *iter, *next;
241 unsigned long hash;
242
243 if (!t->size)
244 return 0;
245 hash = bit_reverse_ulong(node->reverse_hash);
246 for (;;) {
247 uint32_t chain_len = 0;
248
249 /*
250 * iter_prev points to the non-removed node prior to the
251 * insert location.
252 */
253 iter_prev = rcu_dereference(t->tbl[hash & (t->size - 1)]);
254 /* We can always skip the dummy node initially */
255 iter = rcu_dereference(iter_prev->next);
256 assert(iter_prev->reverse_hash <= node->reverse_hash);
257 for (;;) {
258 if (unlikely(!iter))
259 goto insert;
260 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
261 goto insert;
262 next = rcu_dereference(clear_flag(iter)->next);
263 if (is_removed(next))
264 goto gc_node;
265 if (unique
266 && !clear_flag(iter)->dummy
267 && !ht->compare_fct(node->key, node->key_len,
268 clear_flag(iter)->key,
269 clear_flag(iter)->key_len))
270 return -EEXIST;
271 /* Only account for identical reverse hash once */
272 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash)
273 check_resize(ht, t, ++chain_len);
274 iter_prev = clear_flag(iter);
275 iter = next;
276 }
277 insert:
278 assert(node != clear_flag(iter));
279 assert(!is_removed(iter_prev));
280 assert(iter_prev != node);
281 node->next = iter;
282 if (uatomic_cmpxchg(&iter_prev->next, iter,
283 node) != iter)
284 continue; /* retry */
285 else
286 goto gc_end;
287 gc_node:
288 assert(!is_removed(iter));
289 (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next));
290 /* retry */
291 }
292gc_end:
293 /* Garbage collect logically removed nodes in the bucket */
294 dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]);
295 _ht_gc_bucket(dummy, node);
296 return 0;
297}
298
299static
300int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
301{
302 struct rcu_ht_node *dummy, *next, *old;
303 int flagged = 0;
304 unsigned long hash;
305
306 /* logically delete the node */
307 old = rcu_dereference(node->next);
308 do {
309 next = old;
310 if (is_removed(next))
311 goto end;
312 assert(!node->dummy);
313 old = uatomic_cmpxchg(&node->next, next,
314 flag_removed(next));
315 } while (old != next);
316
317 /* We performed the (logical) deletion. */
318 flagged = 1;
319
320 /*
321 * Ensure that the node is not visible to readers anymore: lookup for
322 * the node, and remove it (along with any other logically removed node)
323 * if found.
324 */
325 hash = bit_reverse_ulong(node->reverse_hash);
326 dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]);
327 _ht_gc_bucket(dummy, node);
328end:
329 /*
330 * Only the flagging action indicated that we (and no other)
331 * removed the node from the hash.
332 */
333 if (flagged) {
334 assert(is_removed(rcu_dereference(node->next)));
335 return 0;
336 } else
337 return -ENOENT;
338}
339
340static
341void init_table(struct rcu_ht *ht, struct rcu_table *t,
342 unsigned long first, unsigned long len)
343{
344 unsigned long i, end;
345
346 end = first + len;
347 for (i = first; i < end; i++) {
348 /* Update table size when power of two */
349 if (i != 0 && !(i & (i - 1)))
350 t->size = i;
351 t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node));
352 t->tbl[i]->dummy = 1;
353 t->tbl[i]->reverse_hash = bit_reverse_ulong(i);
354 (void) _ht_add(ht, t, t->tbl[i], 0);
355 }
356 t->resize_target = t->size = end;
357 t->resize_initiated = 0;
358}
359
360struct rcu_ht *ht_new(ht_hash_fct hash_fct,
361 ht_compare_fct compare_fct,
362 unsigned long hash_seed,
363 unsigned long init_size,
364 void (*ht_call_rcu)(struct rcu_head *head,
365 void (*func)(struct rcu_head *head)))
366{
367 struct rcu_ht *ht;
368
369 ht = calloc(1, sizeof(struct rcu_ht));
370 ht->hash_fct = hash_fct;
371 ht->compare_fct = compare_fct;
372 ht->hash_seed = hash_seed;
373 ht->ht_call_rcu = ht_call_rcu;
374 /* this mutex should not nest in read-side C.S. */
375 pthread_mutex_init(&ht->resize_mutex, NULL);
376 ht->t = calloc(1, sizeof(struct rcu_table)
377 + (max(init_size, 1) * sizeof(struct rcu_ht_node *)));
378 ht->t->size = 0;
379 pthread_mutex_lock(&ht->resize_mutex);
380 init_table(ht, ht->t, 0, max(init_size, 1));
381 pthread_mutex_unlock(&ht->resize_mutex);
382 return ht;
383}
384
385struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len)
386{
387 struct rcu_table *t;
388 struct rcu_ht_node *node;
389 unsigned long hash, reverse_hash;
390
391 hash = ht->hash_fct(key, key_len, ht->hash_seed);
392 reverse_hash = bit_reverse_ulong(hash);
393
394 t = rcu_dereference(ht->t);
395 node = rcu_dereference(t->tbl[hash & (t->size - 1)]);
396 for (;;) {
397 if (unlikely(!node))
398 break;
399 if (unlikely(node->reverse_hash > reverse_hash)) {
400 node = NULL;
401 break;
402 }
403 if (likely(!is_removed(rcu_dereference(node->next)))
404 && !node->dummy
405 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
406 break;
407 }
408 node = clear_flag(rcu_dereference(node->next));
409 }
410 assert(!node || !node->dummy);
411 return node;
412}
413
414void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node)
415{
416 struct rcu_table *t;
417 unsigned long hash;
418
419 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
420 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
421
422 t = rcu_dereference(ht->t);
423 (void) _ht_add(ht, t, node, 0);
424}
425
426int ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node)
427{
428 struct rcu_table *t;
429 unsigned long hash;
430
431 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
432 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
433
434 t = rcu_dereference(ht->t);
435 return _ht_add(ht, t, node, 1);
436}
437
438int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node)
439{
440 struct rcu_table *t;
441
442 t = rcu_dereference(ht->t);
443 return _ht_remove(ht, t, node);
444}
445
446static
447int ht_delete_dummy(struct rcu_ht *ht)
448{
449 struct rcu_table *t;
450 struct rcu_ht_node *node;
451 unsigned long i;
452
453 t = ht->t;
454 /* Check that the table is empty */
455 node = t->tbl[0];
456 do {
457 if (!node->dummy)
458 return -EPERM;
459 node = node->next;
460 assert(!is_removed(node));
461 } while (node);
462 /* Internal sanity check: all nodes left should be dummy */
463 for (i = 0; i < t->size; i++) {
464 assert(t->tbl[i]->dummy);
465 free(t->tbl[i]);
466 }
467 return 0;
468}
469
470/*
471 * Should only be called when no more concurrent readers nor writers can
472 * possibly access the table.
473 */
474int ht_destroy(struct rcu_ht *ht)
475{
476 int ret;
477
478 ret = ht_delete_dummy(ht);
479 if (ret)
480 return ret;
481 free(ht->t);
482 free(ht);
483 return ret;
484}
485
486void ht_count_nodes(struct rcu_ht *ht,
487 unsigned long *count,
488 unsigned long *removed)
489{
490 struct rcu_table *t;
491 struct rcu_ht_node *node, *next;
492
493 *count = 0;
494 *removed = 0;
495
496 t = rcu_dereference(ht->t);
497 /* Check that the table is empty */
498 node = rcu_dereference(t->tbl[0]);
499 do {
500 next = rcu_dereference(node->next);
501 if (is_removed(next)) {
502 assert(!node->dummy);
503 (*removed)++;
504 } else if (!node->dummy)
505 (*count)++;
506 node = clear_flag(next);
507 } while (node);
508}
509
510static
511void ht_free_table_cb(struct rcu_head *head)
512{
513 struct rcu_table *t =
514 caa_container_of(head, struct rcu_table, head);
515 free(t);
516}
517
518/* called with resize mutex held */
519static
520void _do_ht_resize(struct rcu_ht *ht)
521{
522 unsigned long new_size, old_size;
523 struct rcu_table *new_t, *old_t;
524
525 old_t = ht->t;
526 old_size = old_t->size;
527
528 new_size = CMM_LOAD_SHARED(old_t->resize_target);
529 dbg_printf("rculfhash: resize from %lu to %lu buckets\n",
530 old_size, new_size);
531 if (old_size == new_size)
532 return;
533 new_t = malloc(sizeof(struct rcu_table)
534 + (new_size * sizeof(struct rcu_ht_node *)));
535 assert(new_size > old_size);
536 memcpy(&new_t->tbl, &old_t->tbl,
537 old_size * sizeof(struct rcu_ht_node *));
538 init_table(ht, new_t, old_size, new_size - old_size);
539 /* Changing table and size atomically wrt lookups */
540 rcu_assign_pointer(ht->t, new_t);
541 ht->ht_call_rcu(&old_t->head, ht_free_table_cb);
542}
543
544static
545unsigned long resize_target_update(struct rcu_table *t,
546 int growth_order)
547{
548 return _uatomic_max(&t->resize_target,
549 t->size << growth_order);
550}
551
552void ht_resize(struct rcu_ht *ht, int growth)
553{
554 struct rcu_table *t = rcu_dereference(ht->t);
555 unsigned long target_size;
556
557 target_size = resize_target_update(t, growth);
558 if (t->size < target_size) {
559 CMM_STORE_SHARED(t->resize_initiated, 1);
560 pthread_mutex_lock(&ht->resize_mutex);
561 _do_ht_resize(ht);
562 pthread_mutex_unlock(&ht->resize_mutex);
563 }
564}
565
566static
567void do_resize_cb(struct rcu_head *head)
568{
569 struct rcu_resize_work *work =
570 caa_container_of(head, struct rcu_resize_work, head);
571 struct rcu_ht *ht = work->ht;
572
573 pthread_mutex_lock(&ht->resize_mutex);
574 _do_ht_resize(ht);
575 pthread_mutex_unlock(&ht->resize_mutex);
576 free(work);
577}
578
579static
580void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth)
581{
582 struct rcu_resize_work *work;
583 unsigned long target_size;
584
585 target_size = resize_target_update(t, growth);
586 if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
587 work = malloc(sizeof(*work));
588 work->ht = ht;
589 ht->ht_call_rcu(&work->head, do_resize_cb);
590 CMM_STORE_SHARED(t->resize_initiated, 1);
591 }
592}
This page took 0.039212 seconds and 4 git commands to generate.