Commit | Line | Data |
---|---|---|
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 | |
d37166c6 MD |
56 | #define REMOVED_FLAG (1UL << 0) |
57 | #define FLAGS_MASK ((1UL << 1) - 1) | |
58 | ||
395270b6 | 59 | struct rcu_table { |
abc490a1 | 60 | unsigned long size; /* always a power of 2 */ |
f9830efd | 61 | unsigned long resize_target; |
11519af6 | 62 | int resize_initiated; |
abc490a1 | 63 | struct rcu_head head; |
395270b6 MD |
64 | struct rcu_ht_node *tbl[0]; |
65 | }; | |
66 | ||
2ed95849 | 67 | struct rcu_ht { |
395270b6 | 68 | struct rcu_table *t; /* shared */ |
2ed95849 | 69 | ht_hash_fct hash_fct; |
732ad076 MD |
70 | ht_compare_fct compare_fct; |
71 | unsigned long hash_seed; | |
464a1ec9 | 72 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
848d4088 | 73 | unsigned int in_progress_resize; |
abc490a1 MD |
74 | void (*ht_call_rcu)(struct rcu_head *head, |
75 | void (*func)(struct rcu_head *head)); | |
2ed95849 MD |
76 | }; |
77 | ||
abc490a1 MD |
78 | struct rcu_resize_work { |
79 | struct rcu_head head; | |
2ed95849 | 80 | struct rcu_ht *ht; |
abc490a1 | 81 | }; |
2ed95849 | 82 | |
abc490a1 MD |
83 | /* |
84 | * Algorithm to reverse bits in a word by lookup table, extended to | |
85 | * 64-bit words. | |
f9830efd | 86 | * Source: |
abc490a1 | 87 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
f9830efd | 88 | * Originally from Public Domain. |
abc490a1 MD |
89 | */ |
90 | ||
91 | static const uint8_t BitReverseTable256[256] = | |
2ed95849 | 92 | { |
abc490a1 MD |
93 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
94 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) | |
95 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) | |
96 | R6(0), R6(2), R6(1), R6(3) | |
97 | }; | |
98 | #undef R2 | |
99 | #undef R4 | |
100 | #undef R6 | |
2ed95849 | 101 | |
abc490a1 MD |
102 | static |
103 | uint8_t bit_reverse_u8(uint8_t v) | |
104 | { | |
105 | return BitReverseTable256[v]; | |
106 | } | |
ab7d5fc6 | 107 | |
abc490a1 MD |
108 | static __attribute__((unused)) |
109 | uint32_t bit_reverse_u32(uint32_t v) | |
110 | { | |
111 | return ((uint32_t) bit_reverse_u8(v) << 24) | | |
112 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | | |
113 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | | |
114 | ((uint32_t) bit_reverse_u8(v >> 24)); | |
2ed95849 MD |
115 | } |
116 | ||
abc490a1 MD |
117 | static __attribute__((unused)) |
118 | uint64_t bit_reverse_u64(uint64_t v) | |
2ed95849 | 119 | { |
abc490a1 MD |
120 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
121 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | | |
122 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | | |
123 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | | |
124 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | | |
125 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | | |
126 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | | |
127 | ((uint64_t) bit_reverse_u8(v >> 56)); | |
128 | } | |
129 | ||
130 | static | |
131 | unsigned long bit_reverse_ulong(unsigned long v) | |
132 | { | |
133 | #if (CAA_BITS_PER_LONG == 32) | |
134 | return bit_reverse_u32(v); | |
135 | #else | |
136 | return bit_reverse_u64(v); | |
137 | #endif | |
138 | } | |
139 | ||
f9830efd MD |
140 | /* |
141 | * Algorithm to find the log2 of a 32-bit unsigned integer. | |
142 | * source: http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogLookup | |
143 | * Originally from Public Domain. | |
144 | */ | |
145 | static const char LogTable256[256] = | |
146 | { | |
147 | #define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n | |
148 | -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, | |
149 | LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6), | |
150 | LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7) | |
151 | }; | |
152 | ||
153 | uint32_t log2_u32(uint32_t v) | |
154 | { | |
155 | uint32_t t, tt; | |
156 | ||
157 | if ((tt = (v >> 16))) | |
158 | return (t = (tt >> 8)) | |
159 | ? 24 + LogTable256[t] | |
160 | : 16 + LogTable256[tt]; | |
161 | else | |
162 | return (t = (v >> 8)) | |
163 | ? 8 + LogTable256[t] | |
164 | : LogTable256[v]; | |
165 | } | |
166 | ||
167 | static | |
168 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); | |
169 | ||
170 | static | |
171 | void check_resize(struct rcu_ht *ht, struct rcu_table *t, | |
172 | uint32_t chain_len) | |
173 | { | |
3390d470 MD |
174 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
175 | ht_resize_lazy(ht, t, | |
65e8e729 | 176 | log2_u32(chain_len - CHAIN_LEN_TARGET - 1)); |
f9830efd MD |
177 | } |
178 | ||
abc490a1 MD |
179 | static |
180 | struct rcu_ht_node *clear_flag(struct rcu_ht_node *node) | |
181 | { | |
d37166c6 | 182 | return (struct rcu_ht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
abc490a1 MD |
183 | } |
184 | ||
185 | static | |
186 | int is_removed(struct rcu_ht_node *node) | |
187 | { | |
d37166c6 | 188 | return ((unsigned long) node) & REMOVED_FLAG; |
abc490a1 MD |
189 | } |
190 | ||
191 | static | |
192 | struct rcu_ht_node *flag_removed(struct rcu_ht_node *node) | |
193 | { | |
d37166c6 | 194 | return (struct rcu_ht_node *) (((unsigned long) node) | REMOVED_FLAG); |
abc490a1 MD |
195 | } |
196 | ||
197 | static | |
f9830efd | 198 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
abc490a1 MD |
199 | { |
200 | unsigned long old1, old2; | |
201 | ||
202 | old1 = uatomic_read(ptr); | |
203 | do { | |
204 | old2 = old1; | |
205 | if (old2 >= v) | |
f9830efd | 206 | return old2; |
abc490a1 | 207 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
f9830efd | 208 | return v; |
abc490a1 MD |
209 | } |
210 | ||
273399de MD |
211 | /* |
212 | * Remove all logically deleted nodes from a bucket up to a certain node key. | |
213 | */ | |
214 | static | |
215 | void _ht_gc_bucket(struct rcu_ht_node *dummy, struct rcu_ht_node *node) | |
216 | { | |
217 | struct rcu_ht_node *iter_prev, *iter, *next; | |
218 | ||
219 | for (;;) { | |
220 | iter_prev = dummy; | |
221 | /* We can always skip the dummy node initially */ | |
cc4fcb10 MD |
222 | iter = rcu_dereference(iter_prev->p.next); |
223 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
273399de | 224 | for (;;) { |
479c8a32 MD |
225 | if (unlikely(!iter)) |
226 | return; | |
cc4fcb10 | 227 | if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) |
273399de | 228 | return; |
cc4fcb10 | 229 | next = rcu_dereference(clear_flag(iter)->p.next); |
273399de MD |
230 | if (is_removed(next)) |
231 | break; | |
273399de MD |
232 | iter_prev = iter; |
233 | iter = next; | |
234 | } | |
235 | assert(!is_removed(iter)); | |
cc4fcb10 | 236 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, clear_flag(next)); |
273399de MD |
237 | } |
238 | } | |
239 | ||
abc490a1 | 240 | static |
18117871 MD |
241 | struct rcu_ht_node *_ht_add(struct rcu_ht *ht, struct rcu_table *t, |
242 | struct rcu_ht_node *node, int unique) | |
abc490a1 | 243 | { |
273399de | 244 | struct rcu_ht_node *iter_prev, *dummy, *iter, *next; |
49c2e2d6 | 245 | unsigned long hash; |
abc490a1 | 246 | |
18117871 | 247 | if (!t->size) { |
cc4fcb10 | 248 | assert(node->p.dummy); |
18117871 MD |
249 | return node; /* Initial first add (head) */ |
250 | } | |
cc4fcb10 | 251 | hash = bit_reverse_ulong(node->p.reverse_hash); |
abc490a1 | 252 | for (;;) { |
f9830efd | 253 | uint32_t chain_len = 0; |
abc490a1 | 254 | |
11519af6 MD |
255 | /* |
256 | * iter_prev points to the non-removed node prior to the | |
257 | * insert location. | |
11519af6 | 258 | */ |
49c2e2d6 | 259 | iter_prev = rcu_dereference(t->tbl[hash & (t->size - 1)]); |
11519af6 | 260 | /* We can always skip the dummy node initially */ |
cc4fcb10 MD |
261 | iter = rcu_dereference(iter_prev->p.next); |
262 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); | |
abc490a1 | 263 | for (;;) { |
273399de MD |
264 | if (unlikely(!iter)) |
265 | goto insert; | |
cc4fcb10 | 266 | if (clear_flag(iter)->p.reverse_hash > node->p.reverse_hash) |
273399de | 267 | goto insert; |
cc4fcb10 | 268 | next = rcu_dereference(clear_flag(iter)->p.next); |
273399de | 269 | if (is_removed(next)) |
9dba85be | 270 | goto gc_node; |
e43f23f8 | 271 | if (unique |
cc4fcb10 | 272 | && !clear_flag(iter)->p.dummy |
e43f23f8 MD |
273 | && !ht->compare_fct(node->key, node->key_len, |
274 | clear_flag(iter)->key, | |
275 | clear_flag(iter)->key_len)) | |
18117871 | 276 | return clear_flag(iter); |
11519af6 | 277 | /* Only account for identical reverse hash once */ |
cc4fcb10 | 278 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash) |
11519af6 MD |
279 | check_resize(ht, t, ++chain_len); |
280 | iter_prev = clear_flag(iter); | |
273399de | 281 | iter = next; |
abc490a1 | 282 | } |
273399de | 283 | insert: |
7ec59d3b | 284 | assert(node != clear_flag(iter)); |
11519af6 | 285 | assert(!is_removed(iter_prev)); |
f000907d | 286 | assert(iter_prev != node); |
cc4fcb10 MD |
287 | node->p.next = iter; |
288 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, | |
273399de MD |
289 | node) != iter) |
290 | continue; /* retry */ | |
11519af6 | 291 | else |
273399de | 292 | goto gc_end; |
9dba85be MD |
293 | gc_node: |
294 | assert(!is_removed(iter)); | |
cc4fcb10 | 295 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, clear_flag(next)); |
273399de | 296 | /* retry */ |
464a1ec9 | 297 | } |
273399de MD |
298 | gc_end: |
299 | /* Garbage collect logically removed nodes in the bucket */ | |
49c2e2d6 | 300 | dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]); |
273399de | 301 | _ht_gc_bucket(dummy, node); |
18117871 | 302 | return node; |
abc490a1 | 303 | } |
464a1ec9 | 304 | |
abc490a1 MD |
305 | static |
306 | int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) | |
307 | { | |
273399de | 308 | struct rcu_ht_node *dummy, *next, *old; |
abc490a1 | 309 | int flagged = 0; |
49c2e2d6 | 310 | unsigned long hash; |
5e28c532 | 311 | |
7ec59d3b | 312 | /* logically delete the node */ |
cc4fcb10 | 313 | old = rcu_dereference(node->p.next); |
7ec59d3b MD |
314 | do { |
315 | next = old; | |
316 | if (is_removed(next)) | |
317 | goto end; | |
cc4fcb10 MD |
318 | assert(!node->p.dummy); |
319 | old = uatomic_cmpxchg(&node->p.next, next, | |
7ec59d3b MD |
320 | flag_removed(next)); |
321 | } while (old != next); | |
322 | ||
323 | /* We performed the (logical) deletion. */ | |
324 | flagged = 1; | |
325 | ||
326 | /* | |
327 | * Ensure that the node is not visible to readers anymore: lookup for | |
273399de MD |
328 | * the node, and remove it (along with any other logically removed node) |
329 | * if found. | |
11519af6 | 330 | */ |
cc4fcb10 | 331 | hash = bit_reverse_ulong(node->p.reverse_hash); |
49c2e2d6 | 332 | dummy = rcu_dereference(t->tbl[hash & (t->size - 1)]); |
273399de | 333 | _ht_gc_bucket(dummy, node); |
2ed95849 | 334 | end: |
11519af6 MD |
335 | /* |
336 | * Only the flagging action indicated that we (and no other) | |
337 | * removed the node from the hash. | |
338 | */ | |
7ec59d3b | 339 | if (flagged) { |
cc4fcb10 | 340 | assert(is_removed(rcu_dereference(node->p.next))); |
11519af6 | 341 | return 0; |
7ec59d3b | 342 | } else |
11519af6 | 343 | return -ENOENT; |
abc490a1 | 344 | } |
2ed95849 | 345 | |
abc490a1 MD |
346 | static |
347 | void init_table(struct rcu_ht *ht, struct rcu_table *t, | |
348 | unsigned long first, unsigned long len) | |
349 | { | |
350 | unsigned long i, end; | |
351 | ||
352 | end = first + len; | |
353 | for (i = first; i < end; i++) { | |
354 | /* Update table size when power of two */ | |
355 | if (i != 0 && !(i & (i - 1))) | |
356 | t->size = i; | |
cc4fcb10 MD |
357 | t->tbl[i] = calloc(1, sizeof(struct _rcu_ht_node)); |
358 | t->tbl[i]->p.dummy = 1; | |
359 | t->tbl[i]->p.reverse_hash = bit_reverse_ulong(i); | |
3eca1b8c | 360 | (void) _ht_add(ht, t, t->tbl[i], 0); |
abc490a1 | 361 | } |
f9830efd | 362 | t->resize_target = t->size = end; |
11519af6 | 363 | t->resize_initiated = 0; |
2ed95849 MD |
364 | } |
365 | ||
abc490a1 | 366 | struct rcu_ht *ht_new(ht_hash_fct hash_fct, |
732ad076 MD |
367 | ht_compare_fct compare_fct, |
368 | unsigned long hash_seed, | |
abc490a1 MD |
369 | unsigned long init_size, |
370 | void (*ht_call_rcu)(struct rcu_head *head, | |
371 | void (*func)(struct rcu_head *head))) | |
372 | { | |
373 | struct rcu_ht *ht; | |
374 | ||
375 | ht = calloc(1, sizeof(struct rcu_ht)); | |
376 | ht->hash_fct = hash_fct; | |
732ad076 MD |
377 | ht->compare_fct = compare_fct; |
378 | ht->hash_seed = hash_seed; | |
f000907d | 379 | ht->ht_call_rcu = ht_call_rcu; |
848d4088 | 380 | ht->in_progress_resize = 0; |
abc490a1 MD |
381 | /* this mutex should not nest in read-side C.S. */ |
382 | pthread_mutex_init(&ht->resize_mutex, NULL); | |
383 | ht->t = calloc(1, sizeof(struct rcu_table) | |
384 | + (max(init_size, 1) * sizeof(struct rcu_ht_node *))); | |
385 | ht->t->size = 0; | |
f000907d | 386 | pthread_mutex_lock(&ht->resize_mutex); |
abc490a1 | 387 | init_table(ht, ht->t, 0, max(init_size, 1)); |
f000907d | 388 | pthread_mutex_unlock(&ht->resize_mutex); |
abc490a1 MD |
389 | return ht; |
390 | } | |
391 | ||
732ad076 | 392 | struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) |
2ed95849 | 393 | { |
395270b6 | 394 | struct rcu_table *t; |
abc490a1 MD |
395 | struct rcu_ht_node *node; |
396 | unsigned long hash, reverse_hash; | |
2ed95849 | 397 | |
732ad076 | 398 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
abc490a1 | 399 | reverse_hash = bit_reverse_ulong(hash); |
464a1ec9 | 400 | |
395270b6 | 401 | t = rcu_dereference(ht->t); |
abc490a1 | 402 | node = rcu_dereference(t->tbl[hash & (t->size - 1)]); |
2ed95849 | 403 | for (;;) { |
abc490a1 MD |
404 | if (unlikely(!node)) |
405 | break; | |
cc4fcb10 | 406 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
abc490a1 MD |
407 | node = NULL; |
408 | break; | |
2ed95849 | 409 | } |
cc4fcb10 MD |
410 | if (likely(!is_removed(rcu_dereference(node->p.next))) |
411 | && !node->p.dummy | |
49c2e2d6 | 412 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
273399de | 413 | break; |
2ed95849 | 414 | } |
cc4fcb10 | 415 | node = clear_flag(rcu_dereference(node->p.next)); |
2ed95849 | 416 | } |
cc4fcb10 | 417 | assert(!node || !node->p.dummy); |
abc490a1 MD |
418 | return node; |
419 | } | |
e0ba718a | 420 | |
f000907d | 421 | void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) |
abc490a1 MD |
422 | { |
423 | struct rcu_table *t; | |
49c2e2d6 | 424 | unsigned long hash; |
ab7d5fc6 | 425 | |
49c2e2d6 | 426 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 427 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
2ed95849 | 428 | |
abc490a1 | 429 | t = rcu_dereference(ht->t); |
3eca1b8c MD |
430 | (void) _ht_add(ht, t, node, 0); |
431 | } | |
432 | ||
18117871 | 433 | struct rcu_ht_node *ht_add_unique(struct rcu_ht *ht, struct rcu_ht_node *node) |
3eca1b8c MD |
434 | { |
435 | struct rcu_table *t; | |
49c2e2d6 | 436 | unsigned long hash; |
3eca1b8c | 437 | |
49c2e2d6 | 438 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
cc4fcb10 | 439 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
3eca1b8c MD |
440 | |
441 | t = rcu_dereference(ht->t); | |
442 | return _ht_add(ht, t, node, 1); | |
2ed95849 MD |
443 | } |
444 | ||
abc490a1 | 445 | int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node) |
2ed95849 | 446 | { |
abc490a1 MD |
447 | struct rcu_table *t; |
448 | ||
449 | t = rcu_dereference(ht->t); | |
abc490a1 | 450 | return _ht_remove(ht, t, node); |
2ed95849 | 451 | } |
ab7d5fc6 | 452 | |
abc490a1 MD |
453 | static |
454 | int ht_delete_dummy(struct rcu_ht *ht) | |
674f7a69 | 455 | { |
395270b6 | 456 | struct rcu_table *t; |
abc490a1 MD |
457 | struct rcu_ht_node *node; |
458 | unsigned long i; | |
674f7a69 | 459 | |
abc490a1 MD |
460 | t = ht->t; |
461 | /* Check that the table is empty */ | |
462 | node = t->tbl[0]; | |
463 | do { | |
cc4fcb10 | 464 | if (!node->p.dummy) |
abc490a1 | 465 | return -EPERM; |
cc4fcb10 | 466 | node = node->p.next; |
273399de | 467 | assert(!is_removed(node)); |
abc490a1 MD |
468 | } while (node); |
469 | /* Internal sanity check: all nodes left should be dummy */ | |
395270b6 | 470 | for (i = 0; i < t->size; i++) { |
cc4fcb10 | 471 | assert(t->tbl[i]->p.dummy); |
abc490a1 | 472 | free(t->tbl[i]); |
674f7a69 | 473 | } |
abc490a1 | 474 | return 0; |
674f7a69 MD |
475 | } |
476 | ||
477 | /* | |
478 | * Should only be called when no more concurrent readers nor writers can | |
479 | * possibly access the table. | |
480 | */ | |
5e28c532 | 481 | int ht_destroy(struct rcu_ht *ht) |
674f7a69 | 482 | { |
5e28c532 MD |
483 | int ret; |
484 | ||
848d4088 MD |
485 | /* Wait for in-flight resize operations to complete */ |
486 | while (uatomic_read(&ht->in_progress_resize)) | |
487 | poll(NULL, 0, 100); /* wait for 100ms */ | |
abc490a1 MD |
488 | ret = ht_delete_dummy(ht); |
489 | if (ret) | |
490 | return ret; | |
395270b6 | 491 | free(ht->t); |
674f7a69 | 492 | free(ht); |
5e28c532 | 493 | return ret; |
674f7a69 MD |
494 | } |
495 | ||
273399de MD |
496 | void ht_count_nodes(struct rcu_ht *ht, |
497 | unsigned long *count, | |
498 | unsigned long *removed) | |
499 | { | |
500 | struct rcu_table *t; | |
501 | struct rcu_ht_node *node, *next; | |
502 | ||
503 | *count = 0; | |
504 | *removed = 0; | |
505 | ||
506 | t = rcu_dereference(ht->t); | |
507 | /* Check that the table is empty */ | |
508 | node = rcu_dereference(t->tbl[0]); | |
509 | do { | |
cc4fcb10 | 510 | next = rcu_dereference(node->p.next); |
273399de | 511 | if (is_removed(next)) { |
cc4fcb10 | 512 | assert(!node->p.dummy); |
273399de | 513 | (*removed)++; |
cc4fcb10 | 514 | } else if (!node->p.dummy) |
273399de MD |
515 | (*count)++; |
516 | node = clear_flag(next); | |
517 | } while (node); | |
518 | } | |
519 | ||
abc490a1 MD |
520 | static |
521 | void ht_free_table_cb(struct rcu_head *head) | |
522 | { | |
523 | struct rcu_table *t = | |
524 | caa_container_of(head, struct rcu_table, head); | |
525 | free(t); | |
526 | } | |
527 | ||
528 | /* called with resize mutex held */ | |
529 | static | |
530 | void _do_ht_resize(struct rcu_ht *ht) | |
464a1ec9 | 531 | { |
abc490a1 | 532 | unsigned long new_size, old_size; |
395270b6 | 533 | struct rcu_table *new_t, *old_t; |
464a1ec9 | 534 | |
395270b6 MD |
535 | old_t = ht->t; |
536 | old_size = old_t->size; | |
464a1ec9 | 537 | |
f9830efd MD |
538 | new_size = CMM_LOAD_SHARED(old_t->resize_target); |
539 | dbg_printf("rculfhash: resize from %lu to %lu buckets\n", | |
540 | old_size, new_size); | |
abc490a1 | 541 | if (old_size == new_size) |
464a1ec9 | 542 | return; |
f000907d | 543 | new_t = malloc(sizeof(struct rcu_table) |
abc490a1 | 544 | + (new_size * sizeof(struct rcu_ht_node *))); |
f000907d MD |
545 | assert(new_size > old_size); |
546 | memcpy(&new_t->tbl, &old_t->tbl, | |
547 | old_size * sizeof(struct rcu_ht_node *)); | |
548 | init_table(ht, new_t, old_size, new_size - old_size); | |
f000907d MD |
549 | /* Changing table and size atomically wrt lookups */ |
550 | rcu_assign_pointer(ht->t, new_t); | |
551 | ht->ht_call_rcu(&old_t->head, ht_free_table_cb); | |
464a1ec9 MD |
552 | } |
553 | ||
abc490a1 | 554 | static |
f9830efd MD |
555 | unsigned long resize_target_update(struct rcu_table *t, |
556 | int growth_order) | |
464a1ec9 | 557 | { |
f9830efd MD |
558 | return _uatomic_max(&t->resize_target, |
559 | t->size << growth_order); | |
464a1ec9 MD |
560 | } |
561 | ||
464a1ec9 MD |
562 | void ht_resize(struct rcu_ht *ht, int growth) |
563 | { | |
f9830efd MD |
564 | struct rcu_table *t = rcu_dereference(ht->t); |
565 | unsigned long target_size; | |
566 | ||
567 | target_size = resize_target_update(t, growth); | |
568 | if (t->size < target_size) { | |
11519af6 | 569 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd MD |
570 | pthread_mutex_lock(&ht->resize_mutex); |
571 | _do_ht_resize(ht); | |
572 | pthread_mutex_unlock(&ht->resize_mutex); | |
573 | } | |
abc490a1 | 574 | } |
464a1ec9 | 575 | |
abc490a1 MD |
576 | static |
577 | void do_resize_cb(struct rcu_head *head) | |
578 | { | |
579 | struct rcu_resize_work *work = | |
580 | caa_container_of(head, struct rcu_resize_work, head); | |
581 | struct rcu_ht *ht = work->ht; | |
582 | ||
583 | pthread_mutex_lock(&ht->resize_mutex); | |
584 | _do_ht_resize(ht); | |
585 | pthread_mutex_unlock(&ht->resize_mutex); | |
586 | free(work); | |
848d4088 MD |
587 | cmm_smp_mb(); /* finish resize before decrement */ |
588 | uatomic_dec(&ht->in_progress_resize); | |
464a1ec9 MD |
589 | } |
590 | ||
abc490a1 | 591 | static |
f000907d | 592 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) |
ab7d5fc6 | 593 | { |
abc490a1 | 594 | struct rcu_resize_work *work; |
f9830efd | 595 | unsigned long target_size; |
abc490a1 | 596 | |
f9830efd | 597 | target_size = resize_target_update(t, growth); |
11519af6 | 598 | if (!CMM_LOAD_SHARED(t->resize_initiated) && t->size < target_size) { |
848d4088 MD |
599 | uatomic_inc(&ht->in_progress_resize); |
600 | cmm_smp_mb(); /* increment resize count before calling it */ | |
f9830efd MD |
601 | work = malloc(sizeof(*work)); |
602 | work->ht = ht; | |
603 | ht->ht_call_rcu(&work->head, do_resize_cb); | |
11519af6 | 604 | CMM_STORE_SHARED(t->resize_initiated, 1); |
f9830efd | 605 | } |
ab7d5fc6 | 606 | } |