rculfhash: wait for call_rcu quiescence before destroy
[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 unsigned int in_progress_resize;
71 void (*ht_call_rcu)(struct rcu_head *head,
72 void (*func)(struct rcu_head *head));
73 };
74
75 struct rcu_resize_work {
76 struct rcu_head head;
77 struct rcu_ht *ht;
78 };
79
80 /*
81 * Algorithm to reverse bits in a word by lookup table, extended to
82 * 64-bit words.
83 * Source:
84 * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
85 * Originally from Public Domain.
86 */
87
88 static const uint8_t BitReverseTable256[256] =
89 {
90 #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64
91 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16)
92 #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 )
93 R6(0), R6(2), R6(1), R6(3)
94 };
95 #undef R2
96 #undef R4
97 #undef R6
98
99 static
100 uint8_t bit_reverse_u8(uint8_t v)
101 {
102 return BitReverseTable256[v];
103 }
104
105 static __attribute__((unused))
106 uint32_t bit_reverse_u32(uint32_t v)
107 {
108 return ((uint32_t) bit_reverse_u8(v) << 24) |
109 ((uint32_t) bit_reverse_u8(v >> 8) << 16) |
110 ((uint32_t) bit_reverse_u8(v >> 16) << 8) |
111 ((uint32_t) bit_reverse_u8(v >> 24));
112 }
113
114 static __attribute__((unused))
115 uint64_t bit_reverse_u64(uint64_t v)
116 {
117 return ((uint64_t) bit_reverse_u8(v) << 56) |
118 ((uint64_t) bit_reverse_u8(v >> 8) << 48) |
119 ((uint64_t) bit_reverse_u8(v >> 16) << 40) |
120 ((uint64_t) bit_reverse_u8(v >> 24) << 32) |
121 ((uint64_t) bit_reverse_u8(v >> 32) << 24) |
122 ((uint64_t) bit_reverse_u8(v >> 40) << 16) |
123 ((uint64_t) bit_reverse_u8(v >> 48) << 8) |
124 ((uint64_t) bit_reverse_u8(v >> 56));
125 }
126
127 static
128 unsigned long bit_reverse_ulong(unsigned long v)
129 {
130 #if (CAA_BITS_PER_LONG == 32)
131 return bit_reverse_u32(v);
132 #else
133 return bit_reverse_u64(v);
134 #endif
135 }
136
137 /*
138 * Algorithm to find the log2 of a 32-bit unsigned integer.
139 * source: http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogLookup
140 * Originally from Public Domain.
141 */
142 static const char LogTable256[256] =
143 {
144 #define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n
145 -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
146 LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6),
147 LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7)
148 };
149
150 uint32_t log2_u32(uint32_t v)
151 {
152 uint32_t t, tt;
153
154 if ((tt = (v >> 16)))
155 return (t = (tt >> 8))
156 ? 24 + LogTable256[t]
157 : 16 + LogTable256[tt];
158 else
159 return (t = (v >> 8))
160 ? 8 + LogTable256[t]
161 : LogTable256[v];
162 }
163
164 static
165 void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth);
166
167 static
168 void check_resize(struct rcu_ht *ht, struct rcu_table *t,
169 uint32_t chain_len)
170 {
171 if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
172 ht_resize_lazy(ht, t,
173 log2_u32(chain_len - CHAIN_LEN_TARGET - 1));
174 }
175
176 static
177 struct rcu_ht_node *clear_flag(struct rcu_ht_node *node)
178 {
179 return (struct rcu_ht_node *) (((unsigned long) node) & ~0x1);
180 }
181
182 static
183 int is_removed(struct rcu_ht_node *node)
184 {
185 return ((unsigned long) node) & 0x1;
186 }
187
188 static
189 struct rcu_ht_node *flag_removed(struct rcu_ht_node *node)
190 {
191 return (struct rcu_ht_node *) (((unsigned long) node) | 0x1);
192 }
193
194 static
195 unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
196 {
197 unsigned long old1, old2;
198
199 old1 = uatomic_read(ptr);
200 do {
201 old2 = old1;
202 if (old2 >= v)
203 return old2;
204 } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
205 return v;
206 }
207
208 /*
209 * Remove all logically deleted nodes from a bucket up to a certain node key.
210 */
211 static
212 void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node)
213 {
214 struct rcu_ht_node *iter_prev, *iter, *next;
215
216 for (;;) {
217 iter_prev = dummy;
218 /* We can always skip the dummy node initially */
219 iter = rcu_dereference(iter_prev->next);
220 assert(iter_prev->reverse_hash <= node->reverse_hash);
221 for (;;) {
222 if (unlikely(!iter))
223 return;
224 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
225 return;
226 next = rcu_dereference(clear_flag(iter)->next);
227 if (is_removed(next))
228 break;
229 iter_prev = iter;
230 iter = next;
231 }
232 assert(!is_removed(iter));
233 (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next));
234 }
235 }
236
237 static
238 struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t,
239 struct rcu_ht_node *node, int unique)
240 {
241 struct rcu_ht_node *iter_prev, *dummy, *iter, *next;
242 unsigned long hash;
243
244 if (!t->size) {
245 assert(node->dummy);
246 return node; /* Initial first add (head) */
247 }
248 hash = bit_reverse_ulong(node->reverse_hash);
249 for (;;) {
250 uint32_t chain_len = 0;
251
252 /*
253 * iter_prev points to the non-removed node prior to the
254 * insert location.
255 */
256 iter_prev = rcu_dereference(t->tbl[hash & (t->size - 1)]);
257 /* We can always skip the dummy node initially */
258 iter = rcu_dereference(iter_prev->next);
259 assert(iter_prev->reverse_hash <= node->reverse_hash);
260 for (;;) {
261 if (unlikely(!iter))
262 goto insert;
263 if (clear_flag(iter)->reverse_hash > node->reverse_hash)
264 goto insert;
265 next = rcu_dereference(clear_flag(iter)->next);
266 if (is_removed(next))
267 goto gc_node;
268 if (unique
269 && !clear_flag(iter)->dummy
270 && !ht->compare_fct(node->key, node->key_len,
271 clear_flag(iter)->key,
272 clear_flag(iter)->key_len))
273 return clear_flag(iter);
274 /* Only account for identical reverse hash once */
275 if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash)
276 check_resize(ht, t, ++chain_len);
277 iter_prev = clear_flag(iter);
278 iter = next;
279 }
280 insert:
281 assert(node != clear_flag(iter));
282 assert(!is_removed(iter_prev));
283 assert(iter_prev != node);
284 node->next = iter;
285 if (uatomic_cmpxchg(&iter_prev->next, iter,
286 node) != iter)
287 continue; /* retry */
288 else
289 goto gc_end;
290 gc_node:
291 assert(!is_removed(iter));
292 (void) uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next));
293 /* retry */
294 }
295 gc_end:
296 /* Garbage collect logically removed nodes in the bucket */
297 dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]);
298 _ht_gc_bucket(dummy, node);
299 return node;
300 }
301
302 static
303 int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node)
304 {
305 struct rcu_ht_node *dummy, *next, *old;
306 int flagged = 0;
307 unsigned long hash;
308
309 /* logically delete the node */
310 old = rcu_dereference(node->next);
311 do {
312 next = old;
313 if (is_removed(next))
314 goto end;
315 assert(!node->dummy);
316 old = uatomic_cmpxchg(&node->next, next,
317 flag_removed(next));
318 } while (old != next);
319
320 /* We performed the (logical) deletion. */
321 flagged = 1;
322
323 /*
324 * Ensure that the node is not visible to readers anymore: lookup for
325 * the node, and remove it (along with any other logically removed node)
326 * if found.
327 */
328 hash = bit_reverse_ulong(node->reverse_hash);
329 dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]);
330 _ht_gc_bucket(dummy, node);
331 end:
332 /*
333 * Only the flagging action indicated that we (and no other)
334 * removed the node from the hash.
335 */
336 if (flagged) {
337 assert(is_removed(rcu_dereference(node->next)));
338 return 0;
339 } else
340 return -ENOENT;
341 }
342
343 static
344 void init_table(struct rcu_ht *ht, struct rcu_table *t,
345 unsigned long first, unsigned long len)
346 {
347 unsigned long i, end;
348
349 end = first + len;
350 for (i = first; i < end; i++) {
351 /* Update table size when power of two */
352 if (i != 0 && !(i & (i - 1)))
353 t->size = i;
354 t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node));
355 t->tbl[i]->dummy = 1;
356 t->tbl[i]->reverse_hash = bit_reverse_ulong(i);
357 (void) _ht_add(ht, t, t->tbl[i], 0);
358 }
359 t->resize_target = t->size = end;
360 t->resize_initiated = 0;
361 }
362
363 struct rcu_ht *ht_new(ht_hash_fct hash_fct,
364 ht_compare_fct compare_fct,
365 unsigned long hash_seed,
366 unsigned long init_size,
367 void (*ht_call_rcu)(struct rcu_head *head,
368 void (*func)(struct rcu_head *head)))
369 {
370 struct rcu_ht *ht;
371
372 ht = calloc(1, sizeof(struct rcu_ht));
373 ht->hash_fct = hash_fct;
374 ht->compare_fct = compare_fct;
375 ht->hash_seed = hash_seed;
376 ht->ht_call_rcu = ht_call_rcu;
377 ht->in_progress_resize = 0;
378 /* this mutex should not nest in read-side C.S. */
379 pthread_mutex_init(&ht->resize_mutex, NULL);
380 ht->t = calloc(1, sizeof(struct rcu_table)
381 + (max(init_size, 1) * sizeof(struct rcu_ht_node *)));
382 ht->t->size = 0;
383 pthread_mutex_lock(&ht->resize_mutex);
384 init_table(ht, ht->t, 0, max(init_size, 1));
385 pthread_mutex_unlock(&ht->resize_mutex);
386 return ht;
387 }
388
389 struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len)
390 {
391 struct rcu_table *t;
392 struct rcu_ht_node *node;
393 unsigned long hash, reverse_hash;
394
395 hash = ht->hash_fct(key, key_len, ht->hash_seed);
396 reverse_hash = bit_reverse_ulong(hash);
397
398 t = rcu_dereference(ht->t);
399 node = rcu_dereference(t->tbl[hash & (t->size - 1)]);
400 for (;;) {
401 if (unlikely(!node))
402 break;
403 if (unlikely(node->reverse_hash > reverse_hash)) {
404 node = NULL;
405 break;
406 }
407 if (likely(!is_removed(rcu_dereference(node->next)))
408 && !node->dummy
409 && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
410 break;
411 }
412 node = clear_flag(rcu_dereference(node->next));
413 }
414 assert(!node || !node->dummy);
415 return node;
416 }
417
418 void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node)
419 {
420 struct rcu_table *t;
421 unsigned long hash;
422
423 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
424 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
425
426 t = rcu_dereference(ht->t);
427 (void) _ht_add(ht, t, node, 0);
428 }
429
430 struct rcu_ht_node *ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node)
431 {
432 struct rcu_table *t;
433 unsigned long hash;
434
435 hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
436 node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
437
438 t = rcu_dereference(ht->t);
439 return _ht_add(ht, t, node, 1);
440 }
441
442 int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node)
443 {
444 struct rcu_table *t;
445
446 t = rcu_dereference(ht->t);
447 return _ht_remove(ht, t, node);
448 }
449
450 static
451 int ht_delete_dummy(struct rcu_ht *ht)
452 {
453 struct rcu_table *t;
454 struct rcu_ht_node *node;
455 unsigned long i;
456
457 t = ht->t;
458 /* Check that the table is empty */
459 node = t->tbl[0];
460 do {
461 if (!node->dummy)
462 return -EPERM;
463 node = node->next;
464 assert(!is_removed(node));
465 } while (node);
466 /* Internal sanity check: all nodes left should be dummy */
467 for (i = 0; i < t->size; i++) {
468 assert(t->tbl[i]->dummy);
469 free(t->tbl[i]);
470 }
471 return 0;
472 }
473
474 /*
475 * Should only be called when no more concurrent readers nor writers can
476 * possibly access the table.
477 */
478 int ht_destroy(struct rcu_ht *ht)
479 {
480 int ret;
481
482 /* Wait for in-flight resize operations to complete */
483 while (uatomic_read(&ht->in_progress_resize))
484 poll(NULL, 0, 100); /* wait for 100ms */
485 ret = ht_delete_dummy(ht);
486 if (ret)
487 return ret;
488 free(ht->t);
489 free(ht);
490 return ret;
491 }
492
493 void ht_count_nodes(struct rcu_ht *ht,
494 unsigned long *count,
495 unsigned long *removed)
496 {
497 struct rcu_table *t;
498 struct rcu_ht_node *node, *next;
499
500 *count = 0;
501 *removed = 0;
502
503 t = rcu_dereference(ht->t);
504 /* Check that the table is empty */
505 node = rcu_dereference(t->tbl[0]);
506 do {
507 next = rcu_dereference(node->next);
508 if (is_removed(next)) {
509 assert(!node->dummy);
510 (*removed)++;
511 } else if (!node->dummy)
512 (*count)++;
513 node = clear_flag(next);
514 } while (node);
515 }
516
517 static
518 void ht_free_table_cb(struct rcu_head *head)
519 {
520 struct rcu_table *t =
521 caa_container_of(head, struct rcu_table, head);
522 free(t);
523 }
524
525 /* called with resize mutex held */
526 static
527 void _do_ht_resize(struct rcu_ht *ht)
528 {
529 unsigned long new_size, old_size;
530 struct rcu_table *new_t, *old_t;
531
532 old_t = ht->t;
533 old_size = old_t->size;
534
535 new_size = CMM_LOAD_SHARED(old_t->resize_target);
536 dbg_printf("rculfhash: resize from %lu to %lu buckets\n",
537 old_size, new_size);
538 if (old_size == new_size)
539 return;
540 new_t = malloc(sizeof(struct rcu_table)
541 + (new_size * sizeof(struct rcu_ht_node *)));
542 assert(new_size > old_size);
543 memcpy(&new_t->tbl, &old_t->tbl,
544 old_size * sizeof(struct rcu_ht_node *));
545 init_table(ht, new_t, old_size, new_size - old_size);
546 /* Changing table and size atomically wrt lookups */
547 rcu_assign_pointer(ht->t, new_t);
548 ht->ht_call_rcu(&old_t->head, ht_free_table_cb);
549 }
550
551 static
552 unsigned long resize_target_update(struct rcu_table *t,
553 int growth_order)
554 {
555 return _uatomic_max(&t->resize_target,
556 t->size << growth_order);
557 }
558
559 void ht_resize(struct rcu_ht *ht, int growth)
560 {
561 struct rcu_table *t = rcu_dereference(ht->t);
562 unsigned long target_size;
563
564 target_size = resize_target_update(t, growth);
565 if (t->size < target_size) {
566 CMM_STORE_SHARED(t->resize_initiated, 1);
567 pthread_mutex_lock(&ht->resize_mutex);
568 _do_ht_resize(ht);
569 pthread_mutex_unlock(&ht->resize_mutex);
570 }
571 }
572
573 static
574 void do_resize_cb(struct rcu_head *head)
575 {
576 struct rcu_resize_work *work =
577 caa_container_of(head, struct rcu_resize_work, head);
578 struct rcu_ht *ht = work->ht;
579
580 pthread_mutex_lock(&ht->resize_mutex);
581 _do_ht_resize(ht);
582 pthread_mutex_unlock(&ht->resize_mutex);
583 free(work);
584 cmm_smp_mb(); /* finish resize before decrement */
585 uatomic_dec(&ht->in_progress_resize);
586 }
587
588 static
589 void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth)
590 {
591 struct rcu_resize_work *work;
592 unsigned long target_size;
593
594 target_size = resize_target_update(t, growth);
595 if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) {
596 uatomic_inc(&ht->in_progress_resize);
597 cmm_smp_mb(); /* increment resize count before calling it */
598 work = malloc(sizeof(*work));
599 work->ht = ht;
600 ht->ht_call_rcu(&work->head, do_resize_cb);
601 CMM_STORE_SHARED(t->resize_initiated, 1);
602 }
603 }
This page took 0.050049 seconds and 5 git commands to generate.