rculfhash: merge node gc into add loop
[urcu.git] / rculfhash.c
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
56 struct 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
64 struct 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
74 struct 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
87 static 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
98 static
99 uint8_t bit_reverse_u8(uint8_t v)
100 {
101 return BitReverseTable256[v];
102 }
103
104 static __attribute__((unused))
105 uint32_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
113 static __attribute__((unused))
114 uint64_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
126 static
127 unsigned 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 */
141 static 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
149 uint32_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
163 static
164 void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth);
165
166 static
167 void 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
175 static
176 struct rcu_ht_node *clear_flag(struct rcu_ht_node *node)
177 {
178 return (struct rcu_ht_node *) (((unsigned long) node) & ~0x1);
179 }
180
181 static
182 int is_removed(struct rcu_ht_node *node)
183 {
184 return ((unsigned long) node) & 0x1;
185 }
186
187 static
188 struct rcu_ht_node *flag_removed(struct rcu_ht_node *node)
189 {
190 return (struct rcu_ht_node *) (((unsigned long) node) | 0x1);
191 }
192
193 static
194 unsigned 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 */
210 static
211 void _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
236 static
237 int _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
242 if (!t->size)
243 return 0;
244 for (;;) {
245 uint32_t chain_len = 0;
246
247 /*
248 * iter_prev points to the non-removed node prior to the
249 * insert location.
250 */
251 iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
252 /* We can always skip the dummy node initially */
253 iter = rcu_dereference(iter_prev->next);
254 assert(iter_prev->reverse_hash <= node->reverse_hash);
255 for (;;) {
256 if (unlikely(!iter))
257 goto insert;
258 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
259 goto insert;
260 next = rcu_dereference(clear_flag(iter)->next);
261 if (is_removed(next))
262 goto gc_node;
263 /* Only account for identical reverse hash once */
264 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash)
265 check_resize(ht, t, ++chain_len);
266 iter_prev = clear_flag(iter);
267 iter = next;
268 }
269 insert:
270 assert(node != clear_flag(iter));
271 assert(!is_removed(iter_prev));
272 assert(iter_prev != node);
273 node->next = iter;
274 if (uatomic_cmpxchg(&iter_prev->next, iter,
275 node) != iter)
276 continue; /* retry */
277 else
278 goto gc_end;
279 gc_node:
280 assert(!is_removed(iter));
281 (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next));
282 /* retry */
283 }
284 gc_end:
285 /* Garbage collect logically removed nodes in the bucket */
286 dummy = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
287 _ht_gc_bucket(dummy, node);
288 return 0;
289 }
290
291 static
292 int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
293 {
294 struct rcu_ht_node *dummy, *next, *old;
295 int flagged = 0;
296
297 /* logically delete the node */
298 old = rcu_dereference(node->next);
299 do {
300 next = old;
301 if (is_removed(next))
302 goto end;
303 assert(!node->dummy);
304 old = uatomic_cmpxchg(&node->next, next,
305 flag_removed(next));
306 } while (old != next);
307
308 /* We performed the (logical) deletion. */
309 flagged = 1;
310
311 /*
312 * Ensure that the node is not visible to readers anymore: lookup for
313 * the node, and remove it (along with any other logically removed node)
314 * if found.
315 */
316 dummy = rcu_dereference(t->tbl[node->hash & (t->size - 1)]);
317 _ht_gc_bucket(dummy, node);
318 end:
319 /*
320 * Only the flagging action indicated that we (and no other)
321 * removed the node from the hash.
322 */
323 if (flagged) {
324 assert(is_removed(rcu_dereference(node->next)));
325 return 0;
326 } else
327 return -ENOENT;
328 }
329
330 static
331 void init_table(struct rcu_ht *ht, struct rcu_table *t,
332 unsigned long first, unsigned long len)
333 {
334 unsigned long i, end;
335
336 end = first + len;
337 for (i = first; i < end; i++) {
338 /* Update table size when power of two */
339 if (i != 0 && !(i & (i - 1)))
340 t->size = i;
341 t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node));
342 t->tbl[i]->dummy = 1;
343 t->tbl[i]->hash = i;
344 t->tbl[i]->reverse_hash = bit_reverse_ulong(i);
345 (void) _ht_add(ht, t, t->tbl[i], 0);
346 }
347 t->resize_target = t->size = end;
348 t->resize_initiated = 0;
349 }
350
351 struct rcu_ht *ht_new(ht_hash_fct hash_fct,
352 ht_compare_fct compare_fct,
353 unsigned long hash_seed,
354 unsigned long init_size,
355 void (*ht_call_rcu)(struct rcu_head *head,
356 void (*func)(struct rcu_head *head)))
357 {
358 struct rcu_ht *ht;
359
360 ht = calloc(1, sizeof(struct rcu_ht));
361 ht->hash_fct = hash_fct;
362 ht->compare_fct = compare_fct;
363 ht->hash_seed = hash_seed;
364 ht->ht_call_rcu = ht_call_rcu;
365 /* this mutex should not nest in read-side C.S. */
366 pthread_mutex_init(&ht->resize_mutex, NULL);
367 ht->t = calloc(1, sizeof(struct rcu_table)
368 + (max(init_size, 1) * sizeof(struct rcu_ht_node *)));
369 ht->t->size = 0;
370 pthread_mutex_lock(&ht->resize_mutex);
371 init_table(ht, ht->t, 0, max(init_size, 1));
372 pthread_mutex_unlock(&ht->resize_mutex);
373 return ht;
374 }
375
376 struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len)
377 {
378 struct rcu_table *t;
379 struct rcu_ht_node *node;
380 unsigned long hash, reverse_hash;
381
382 hash = ht->hash_fct(key, key_len, ht->hash_seed);
383 reverse_hash = bit_reverse_ulong(hash);
384
385 t = rcu_dereference(ht->t);
386 node = rcu_dereference(t->tbl[hash & (t->size - 1)]);
387 for (;;) {
388 if (unlikely(!node))
389 break;
390 if (unlikely(node->reverse_hash > reverse_hash)) {
391 node = NULL;
392 break;
393 }
394 if (!ht->compare_fct(node->key, node->key_len, key, key_len)) {
395 if (likely(!is_removed(rcu_dereference(node->next)))
396 && likely(!node->dummy))
397 break;
398 }
399 node = clear_flag(rcu_dereference(node->next));
400 }
401 assert(!node || !node->dummy);
402 return node;
403 }
404
405 void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node)
406 {
407 struct rcu_table *t;
408
409 node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
410 node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash);
411
412 t = rcu_dereference(ht->t);
413 (void) _ht_add(ht, t, node, 0);
414 }
415
416 int ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node)
417 {
418 struct rcu_table *t;
419
420 node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
421 node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash);
422
423 t = rcu_dereference(ht->t);
424 return _ht_add(ht, t, node, 1);
425 }
426
427 int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node)
428 {
429 struct rcu_table *t;
430
431 t = rcu_dereference(ht->t);
432 return _ht_remove(ht, t, node);
433 }
434
435 static
436 int ht_delete_dummy(struct rcu_ht *ht)
437 {
438 struct rcu_table *t;
439 struct rcu_ht_node *node;
440 unsigned long i;
441
442 t = ht->t;
443 /* Check that the table is empty */
444 node = t->tbl[0];
445 do {
446 if (!node->dummy)
447 return -EPERM;
448 node = node->next;
449 assert(!is_removed(node));
450 } while (node);
451 /* Internal sanity check: all nodes left should be dummy */
452 for (i = 0; i < t->size; i++) {
453 assert(t->tbl[i]->dummy);
454 free(t->tbl[i]);
455 }
456 return 0;
457 }
458
459 /*
460 * Should only be called when no more concurrent readers nor writers can
461 * possibly access the table.
462 */
463 int ht_destroy(struct rcu_ht *ht)
464 {
465 int ret;
466
467 ret = ht_delete_dummy(ht);
468 if (ret)
469 return ret;
470 free(ht->t);
471 free(ht);
472 return ret;
473 }
474
475 void ht_count_nodes(struct rcu_ht *ht,
476 unsigned long *count,
477 unsigned long *removed)
478 {
479 struct rcu_table *t;
480 struct rcu_ht_node *node, *next;
481
482 *count = 0;
483 *removed = 0;
484
485 t = rcu_dereference(ht->t);
486 /* Check that the table is empty */
487 node = rcu_dereference(t->tbl[0]);
488 do {
489 next = rcu_dereference(node->next);
490 if (is_removed(next)) {
491 assert(!node->dummy);
492 (*removed)++;
493 } else if (!node->dummy)
494 (*count)++;
495 node = clear_flag(next);
496 } while (node);
497 }
498
499 static
500 void ht_free_table_cb(struct rcu_head *head)
501 {
502 struct rcu_table *t =
503 caa_container_of(head, struct rcu_table, head);
504 free(t);
505 }
506
507 /* called with resize mutex held */
508 static
509 void _do_ht_resize(struct rcu_ht *ht)
510 {
511 unsigned long new_size, old_size;
512 struct rcu_table *new_t, *old_t;
513
514 old_t = ht->t;
515 old_size = old_t->size;
516
517 new_size = CMM_LOAD_SHARED(old_t->resize_target);
518 dbg_printf("rculfhash: resize from %lu to %lu buckets\n",
519 old_size, new_size);
520 if (old_size == new_size)
521 return;
522 new_t = malloc(sizeof(struct rcu_table)
523 + (new_size * sizeof(struct rcu_ht_node *)));
524 assert(new_size > old_size);
525 memcpy(&new_t->tbl, &old_t->tbl,
526 old_size * sizeof(struct rcu_ht_node *));
527 init_table(ht, new_t, old_size, new_size - old_size);
528 /* Changing table and size atomically wrt lookups */
529 rcu_assign_pointer(ht->t, new_t);
530 ht->ht_call_rcu(&old_t->head, ht_free_table_cb);
531 }
532
533 static
534 unsigned long resize_target_update(struct rcu_table *t,
535 int growth_order)
536 {
537 return _uatomic_max(&t->resize_target,
538 t->size << growth_order);
539 }
540
541 void ht_resize(struct rcu_ht *ht, int growth)
542 {
543 struct rcu_table *t = rcu_dereference(ht->t);
544 unsigned long target_size;
545
546 target_size = resize_target_update(t, growth);
547 if (t->size < target_size) {
548 CMM_STORE_SHARED(t->resize_initiated, 1);
549 pthread_mutex_lock(&ht->resize_mutex);
550 _do_ht_resize(ht);
551 pthread_mutex_unlock(&ht->resize_mutex);
552 }
553 }
554
555 static
556 void do_resize_cb(struct rcu_head *head)
557 {
558 struct rcu_resize_work *work =
559 caa_container_of(head, struct rcu_resize_work, head);
560 struct rcu_ht *ht = work->ht;
561
562 pthread_mutex_lock(&ht->resize_mutex);
563 _do_ht_resize(ht);
564 pthread_mutex_unlock(&ht->resize_mutex);
565 free(work);
566 }
567
568 static
569 void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth)
570 {
571 struct rcu_resize_work *work;
572 unsigned long target_size;
573
574 target_size = resize_target_update(t, growth);
575 if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
576 work = malloc(sizeof(*work));
577 work->ht = ht;
578 ht->ht_call_rcu(&work->head, do_resize_cb);
579 CMM_STORE_SHARED(t->resize_initiated, 1);
580 }
581 }
This page took 0.040533 seconds and 5 git commands to generate.