rculfhash: add unique
[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
2ed95849
MD
23#define _LGPL_SOURCE
24#include <stdlib.h>
e0ba718a
MD
25#include <errno.h>
26#include <assert.h>
27#include <stdio.h>
abc490a1 28#include <stdint.h>
f000907d 29#include <string.h>
e0ba718a 30
2ed95849 31#include <urcu.h>
abc490a1 32#include <urcu-call-rcu.h>
a42cc659
MD
33#include <urcu/arch.h>
34#include <urcu/uatomic.h>
674f7a69 35#include <urcu/jhash.h>
a42cc659 36#include <urcu/compiler.h>
abc490a1 37#include <urcu/rculfhash.h>
5e28c532 38#include <stdio.h>
464a1ec9 39#include <pthread.h>
44395fb7 40
f9830efd
MD
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
65e8e729
MD
49#define CHAIN_LEN_TARGET 1
50#define CHAIN_LEN_RESIZE_THRESHOLD 2
2ed95849 51
abc490a1
MD
52#ifndef max
53#define max(a, b) ((a) > (b) ? (a) : (b))
54#endif
2ed95849 55
395270b6 56struct rcu_table {
abc490a1 57 unsigned long size; /* always a power of 2 */
f9830efd 58 unsigned long resize_target;
11519af6 59 int resize_initiated;
abc490a1 60 struct rcu_head head;
395270b6
MD
61 struct rcu_ht_node *tbl[0];
62};
63
2ed95849 64struct rcu_ht {
395270b6 65 struct rcu_table *t; /* shared */
2ed95849 66 ht_hash_fct hash_fct;
732ad076
MD
67 ht_compare_fct compare_fct;
68 unsigned long hash_seed;
464a1ec9 69 pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */
abc490a1
MD
70 void (*ht_call_rcu)(struct rcu_head *head,
71 void (*func)(struct rcu_head *head));
2ed95849
MD
72};
73
abc490a1
MD
74struct rcu_resize_work {
75 struct rcu_head head;
2ed95849 76 struct rcu_ht *ht;
abc490a1 77};
2ed95849 78
abc490a1
MD
79/*
80 * Algorithm to reverse bits in a word by lookup table, extended to
81 * 64-bit words.
f9830efd 82 * Source:
abc490a1 83 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
f9830efd 84 * Originally from Public Domain.
abc490a1
MD
85 */
86
87static const uint8_t BitReverseTable256[256] =
2ed95849 88{
abc490a1
MD
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
2ed95849 97
abc490a1
MD
98static
99uint8_t bit_reverse_u8(uint8_t v)
100{
101 return BitReverseTable256[v];
102}
ab7d5fc6 103
abc490a1
MD
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));
2ed95849
MD
111}
112
abc490a1
MD
113static __attribute__((unused))
114uint64_t bit_reverse_u64(uint64_t v)
2ed95849 115{
abc490a1
MD
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
f9830efd
MD
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{
3390d470
MD
170 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
171 ht_resize_lazy(ht, t,
65e8e729 172 log2_u32(chain_len - CHAIN_LEN_TARGET - 1));
f9830efd
MD
173}
174
abc490a1
MD
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
f9830efd 194unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
abc490a1
MD
195{
196 unsigned long old1, old2;
197
198 old1 = uatomic_read(ptr);
199 do {
200 old2 = old1;
201 if (old2 >= v)
f9830efd 202 return old2;
abc490a1 203 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
f9830efd 204 return v;
abc490a1
MD
205}
206
207static
3eca1b8c
MD
208int _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node,
209 int unique)
abc490a1 210{
11519af6 211 struct rcu_ht_node *iter_prev, *iter, *iter_prev_next, *next;
abc490a1 212
f000907d 213 if (!t->size)
3eca1b8c 214 return 0;
abc490a1 215 for (;;) {
f9830efd 216 uint32_t chain_len = 0;
abc490a1 217
11519af6
MD
218 /*
219 * iter_prev points to the non-removed node prior to the
220 * insert location.
221 * iter iterates until it finds the next non-removed
222 * node.
223 */
abc490a1 224 iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
11519af6
MD
225 /* We can always skip the dummy node initially */
226 iter_prev_next = next = rcu_dereference(iter_prev->next);
abc490a1
MD
227 assert(iter_prev);
228 assert(iter_prev->reverse_hash <= node->reverse_hash);
229 for (;;) {
11519af6
MD
230 iter = next;
231 if (unlikely(!clear_flag(iter)))
abc490a1 232 break;
11519af6
MD
233 next = rcu_dereference(clear_flag(iter)->next);
234 if (unlikely(is_removed(next)))
235 continue;
3eca1b8c
MD
236 if (unique
237 && !clear_flag(iter)->dummy
238 && !ht->compare_fct(node->key, node->key_len,
239 clear_flag(iter)->key,
240 clear_flag(iter)->key_len))
241 return -EEXIST;
11519af6 242 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
abc490a1 243 break;
11519af6
MD
244 /* Only account for identical reverse hash once */
245 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash)
246 check_resize(ht, t, ++chain_len);
247 iter_prev = clear_flag(iter);
248 iter_prev_next = next;
abc490a1 249 }
f000907d 250 assert(node != iter);
11519af6 251 assert(!is_removed(iter_prev));
f000907d 252 assert(iter_prev != node);
11519af6
MD
253 node->next = iter;
254 if (uatomic_cmpxchg(&iter_prev->next, iter_prev_next,
255 node) != iter_prev_next)
abc490a1 256 continue;
11519af6
MD
257 else
258 break;
464a1ec9 259 }
3eca1b8c 260 return 0;
abc490a1 261}
464a1ec9 262
abc490a1
MD
263static
264int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
265{
11519af6 266 struct rcu_ht_node *iter_prev, *iter, *iter_prev_next, *next, *old;
abc490a1 267 unsigned long chain_len;
11519af6 268 int found;
abc490a1 269 int flagged = 0;
5e28c532 270
abc490a1
MD
271retry:
272 chain_len = 0;
273 found = 0;
11519af6
MD
274 /*
275 * iter_prev points to the non-removed node prior to the remove
276 * location.
277 * node is the node to remove.
278 */
abc490a1 279 iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
11519af6
MD
280 /* We can always skip the dummy node initially */
281 iter_prev_next = next = rcu_dereference(iter_prev->next);
abc490a1
MD
282 assert(iter_prev);
283 assert(iter_prev->reverse_hash <= node->reverse_hash);
2ed95849 284 for (;;) {
11519af6
MD
285 iter = next;
286 if (unlikely(!clear_flag(iter)))
abc490a1 287 break;
11519af6 288 next = rcu_dereference(clear_flag(iter)->next);
abc490a1
MD
289 if (iter == node) {
290 found = 1;
2ed95849
MD
291 break;
292 }
11519af6
MD
293 if (unlikely(is_removed(next)))
294 continue;
295 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
296 break;
297 iter_prev = clear_flag(iter);
298 iter_prev_next = next;
abc490a1 299 }
11519af6 300 if (!found)
abc490a1 301 goto end;
abc490a1 302 if (!flagged) {
11519af6 303 if (is_removed(next))
2ed95849 304 goto end;
abc490a1 305 /* set deletion flag */
f000907d
MD
306 if ((old = uatomic_cmpxchg(&iter->next, next,
307 flag_removed(next))) != next) {
11519af6 308 if (old == flag_removed(next))
abc490a1 309 goto end;
11519af6 310 else
abc490a1 311 goto retry;
abc490a1
MD
312 }
313 flagged = 1;
2ed95849 314 }
abc490a1 315 /*
98f969fc 316 * Remove the element from the list.
11519af6
MD
317 * - Retry if there has been a concurrent add before us.
318 * - Retry if the prev node has been deleted (its next removed
319 * flag would be set).
320 * - There cannot be a concurrent delete for our position, because
321 * we won the deletion flag cmpxchg.
322 * - If there is a concurrent add or remove after us while our
323 * removed flag is set, it will skip us and link directly after
324 * the prior non-removed node before us. In this case, the
325 * retry will not find the node in the list anymore.
abc490a1 326 */
11519af6
MD
327 if (uatomic_cmpxchg(&iter_prev->next, iter_prev_next,
328 clear_flag(next)) != iter_prev_next)
abc490a1 329 goto retry;
2ed95849 330end:
11519af6
MD
331 /*
332 * Only the flagging action indicated that we (and no other)
333 * removed the node from the hash.
334 */
335 if (flagged)
336 return 0;
337 else
338 return -ENOENT;
abc490a1 339}
2ed95849 340
abc490a1
MD
341static
342void init_table(struct rcu_ht *ht, struct rcu_table *t,
343 unsigned long first, unsigned long len)
344{
345 unsigned long i, end;
346
347 end = first + len;
348 for (i = first; i < end; i++) {
349 /* Update table size when power of two */
350 if (i != 0 && !(i & (i - 1)))
351 t->size = i;
352 t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node));
353 t->tbl[i]->dummy = 1;
354 t->tbl[i]->hash = i;
355 t->tbl[i]->reverse_hash = bit_reverse_ulong(i);
3eca1b8c 356 (void) _ht_add(ht, t, t->tbl[i], 0);
abc490a1 357 }
f9830efd 358 t->resize_target = t->size = end;
11519af6 359 t->resize_initiated = 0;
2ed95849
MD
360}
361
abc490a1 362struct rcu_ht *ht_new(ht_hash_fct hash_fct,
732ad076
MD
363 ht_compare_fct compare_fct,
364 unsigned long hash_seed,
abc490a1
MD
365 unsigned long init_size,
366 void (*ht_call_rcu)(struct rcu_head *head,
367 void (*func)(struct rcu_head *head)))
368{
369 struct rcu_ht *ht;
370
371 ht = calloc(1, sizeof(struct rcu_ht));
372 ht->hash_fct = hash_fct;
732ad076
MD
373 ht->compare_fct = compare_fct;
374 ht->hash_seed = hash_seed;
f000907d 375 ht->ht_call_rcu = ht_call_rcu;
abc490a1
MD
376 /* this mutex should not nest in read-side C.S. */
377 pthread_mutex_init(&ht->resize_mutex, NULL);
378 ht->t = calloc(1, sizeof(struct rcu_table)
379 + (max(init_size, 1) * sizeof(struct rcu_ht_node *)));
380 ht->t->size = 0;
f000907d 381 pthread_mutex_lock(&ht->resize_mutex);
abc490a1 382 init_table(ht, ht->t, 0, max(init_size, 1));
f000907d 383 pthread_mutex_unlock(&ht->resize_mutex);
abc490a1
MD
384 return ht;
385}
386
732ad076 387struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len)
2ed95849 388{
395270b6 389 struct rcu_table *t;
abc490a1
MD
390 struct rcu_ht_node *node;
391 unsigned long hash, reverse_hash;
2ed95849 392
732ad076 393 hash = ht->hash_fct(key, key_len, ht->hash_seed);
abc490a1 394 reverse_hash = bit_reverse_ulong(hash);
464a1ec9 395
395270b6 396 t = rcu_dereference(ht->t);
abc490a1 397 node = rcu_dereference(t->tbl[hash & (t->size - 1)]);
2ed95849 398 for (;;) {
abc490a1
MD
399 if (unlikely(!node))
400 break;
dd4505e0 401 if (unlikely(node->reverse_hash > reverse_hash)) {
abc490a1
MD
402 node = NULL;
403 break;
2ed95849 404 }
732ad076 405 if (!ht->compare_fct(node->key, node->key_len, key, key_len)) {
dd4505e0 406 if (unlikely(is_removed(rcu_dereference(node->next))))
abc490a1 407 node = NULL;
2ed95849
MD
408 break;
409 }
abc490a1 410 node = clear_flag(rcu_dereference(node->next));
2ed95849 411 }
abc490a1
MD
412 return node;
413}
e0ba718a 414
f000907d 415void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node)
abc490a1
MD
416{
417 struct rcu_table *t;
ab7d5fc6 418
732ad076 419 node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
abc490a1 420 node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash);
2ed95849 421
abc490a1 422 t = rcu_dereference(ht->t);
3eca1b8c
MD
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
430 node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
431 node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash);
432
433 t = rcu_dereference(ht->t);
434 return _ht_add(ht, t, node, 1);
2ed95849
MD
435}
436
abc490a1 437int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node)
2ed95849 438{
abc490a1
MD
439 struct rcu_table *t;
440
441 t = rcu_dereference(ht->t);
abc490a1 442 return _ht_remove(ht, t, node);
2ed95849 443}
ab7d5fc6 444
abc490a1
MD
445static
446int ht_delete_dummy(struct rcu_ht *ht)
674f7a69 447{
395270b6 448 struct rcu_table *t;
abc490a1
MD
449 struct rcu_ht_node *node;
450 unsigned long i;
674f7a69 451
abc490a1
MD
452 t = ht->t;
453 /* Check that the table is empty */
454 node = t->tbl[0];
455 do {
456 if (!node->dummy)
457 return -EPERM;
458 node = node->next;
459 } while (node);
460 /* Internal sanity check: all nodes left should be dummy */
395270b6 461 for (i = 0; i < t->size; i++) {
abc490a1
MD
462 assert(t->tbl[i]->dummy);
463 free(t->tbl[i]);
674f7a69 464 }
abc490a1 465 return 0;
674f7a69
MD
466}
467
468/*
469 * Should only be called when no more concurrent readers nor writers can
470 * possibly access the table.
471 */
5e28c532 472int ht_destroy(struct rcu_ht *ht)
674f7a69 473{
5e28c532
MD
474 int ret;
475
abc490a1
MD
476 ret = ht_delete_dummy(ht);
477 if (ret)
478 return ret;
395270b6 479 free(ht->t);
674f7a69 480 free(ht);
5e28c532 481 return ret;
674f7a69
MD
482}
483
abc490a1
MD
484static
485void ht_free_table_cb(struct rcu_head *head)
486{
487 struct rcu_table *t =
488 caa_container_of(head, struct rcu_table, head);
489 free(t);
490}
491
492/* called with resize mutex held */
493static
494void _do_ht_resize(struct rcu_ht *ht)
464a1ec9 495{
abc490a1 496 unsigned long new_size, old_size;
395270b6 497 struct rcu_table *new_t, *old_t;
464a1ec9 498
395270b6
MD
499 old_t = ht->t;
500 old_size = old_t->size;
464a1ec9 501
f9830efd
MD
502 new_size = CMM_LOAD_SHARED(old_t->resize_target);
503 dbg_printf("rculfhash: resize from %lu to %lu buckets\n",
504 old_size, new_size);
abc490a1 505 if (old_size == new_size)
464a1ec9 506 return;
f000907d 507 new_t = malloc(sizeof(struct rcu_table)
abc490a1 508 + (new_size * sizeof(struct rcu_ht_node *)));
f000907d
MD
509 assert(new_size > old_size);
510 memcpy(&new_t->tbl, &old_t->tbl,
511 old_size * sizeof(struct rcu_ht_node *));
512 init_table(ht, new_t, old_size, new_size - old_size);
f000907d
MD
513 /* Changing table and size atomically wrt lookups */
514 rcu_assign_pointer(ht->t, new_t);
515 ht->ht_call_rcu(&old_t->head, ht_free_table_cb);
464a1ec9
MD
516}
517
abc490a1 518static
f9830efd
MD
519unsigned long resize_target_update(struct rcu_table *t,
520 int growth_order)
464a1ec9 521{
f9830efd
MD
522 return _uatomic_max(&t->resize_target,
523 t->size << growth_order);
464a1ec9
MD
524}
525
464a1ec9
MD
526void ht_resize(struct rcu_ht *ht, int growth)
527{
f9830efd
MD
528 struct rcu_table *t = rcu_dereference(ht->t);
529 unsigned long target_size;
530
531 target_size = resize_target_update(t, growth);
532 if (t->size < target_size) {
11519af6 533 CMM_STORE_SHARED(t->resize_initiated, 1);
f9830efd
MD
534 pthread_mutex_lock(&ht->resize_mutex);
535 _do_ht_resize(ht);
536 pthread_mutex_unlock(&ht->resize_mutex);
537 }
abc490a1 538}
464a1ec9 539
abc490a1
MD
540static
541void do_resize_cb(struct rcu_head *head)
542{
543 struct rcu_resize_work *work =
544 caa_container_of(head, struct rcu_resize_work, head);
545 struct rcu_ht *ht = work->ht;
546
547 pthread_mutex_lock(&ht->resize_mutex);
548 _do_ht_resize(ht);
549 pthread_mutex_unlock(&ht->resize_mutex);
550 free(work);
464a1ec9
MD
551}
552
abc490a1 553static
f000907d 554void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth)
ab7d5fc6 555{
abc490a1 556 struct rcu_resize_work *work;
f9830efd 557 unsigned long target_size;
abc490a1 558
f9830efd 559 target_size = resize_target_update(t, growth);
11519af6 560 if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
f9830efd
MD
561 work = malloc(sizeof(*work));
562 work->ht = ht;
563 ht->ht_call_rcu(&work->head, do_resize_cb);
11519af6 564 CMM_STORE_SHARED(t->resize_initiated, 1);
f9830efd 565 }
ab7d5fc6 566}
This page took 0.07651 seconds and 4 git commands to generate.