| 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 BUCKET_SIZE_RESIZE_THRESHOLD 4 |
| 50 | |
| 51 | #ifndef max |
| 52 | #define max(a, b) ((a) > (b) ? (a) : (b)) |
| 53 | #endif |
| 54 | |
| 55 | struct rcu_table { |
| 56 | unsigned long size; /* always a power of 2 */ |
| 57 | unsigned long resize_target; |
| 58 | struct rcu_head head; |
| 59 | struct rcu_ht_node *tbl[0]; |
| 60 | }; |
| 61 | |
| 62 | struct rcu_ht { |
| 63 | struct rcu_table *t; /* shared */ |
| 64 | ht_hash_fct hash_fct; |
| 65 | ht_compare_fct compare_fct; |
| 66 | unsigned long hash_seed; |
| 67 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
| 68 | void (*ht_call_rcu)(struct rcu_head *head, |
| 69 | void (*func)(struct rcu_head *head)); |
| 70 | }; |
| 71 | |
| 72 | struct rcu_resize_work { |
| 73 | struct rcu_head head; |
| 74 | struct rcu_ht *ht; |
| 75 | }; |
| 76 | |
| 77 | /* |
| 78 | * Algorithm to reverse bits in a word by lookup table, extended to |
| 79 | * 64-bit words. |
| 80 | * Source: |
| 81 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
| 82 | * Originally from Public Domain. |
| 83 | */ |
| 84 | |
| 85 | static const uint8_t BitReverseTable256[256] = |
| 86 | { |
| 87 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
| 88 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) |
| 89 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) |
| 90 | R6(0), R6(2), R6(1), R6(3) |
| 91 | }; |
| 92 | #undef R2 |
| 93 | #undef R4 |
| 94 | #undef R6 |
| 95 | |
| 96 | static |
| 97 | uint8_t bit_reverse_u8(uint8_t v) |
| 98 | { |
| 99 | return BitReverseTable256[v]; |
| 100 | } |
| 101 | |
| 102 | static __attribute__((unused)) |
| 103 | uint32_t bit_reverse_u32(uint32_t v) |
| 104 | { |
| 105 | return ((uint32_t) bit_reverse_u8(v) << 24) | |
| 106 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | |
| 107 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | |
| 108 | ((uint32_t) bit_reverse_u8(v >> 24)); |
| 109 | } |
| 110 | |
| 111 | static __attribute__((unused)) |
| 112 | uint64_t bit_reverse_u64(uint64_t v) |
| 113 | { |
| 114 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
| 115 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | |
| 116 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | |
| 117 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | |
| 118 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | |
| 119 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | |
| 120 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | |
| 121 | ((uint64_t) bit_reverse_u8(v >> 56)); |
| 122 | } |
| 123 | |
| 124 | static |
| 125 | unsigned long bit_reverse_ulong(unsigned long v) |
| 126 | { |
| 127 | #if (CAA_BITS_PER_LONG == 32) |
| 128 | return bit_reverse_u32(v); |
| 129 | #else |
| 130 | return bit_reverse_u64(v); |
| 131 | #endif |
| 132 | } |
| 133 | |
| 134 | /* |
| 135 | * Algorithm to find the log2 of a 32-bit unsigned integer. |
| 136 | * source: http://graphics.stanford.edu/~seander/bithacks.html#IntegerLogLookup |
| 137 | * Originally from Public Domain. |
| 138 | */ |
| 139 | static const char LogTable256[256] = |
| 140 | { |
| 141 | #define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n |
| 142 | -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, |
| 143 | LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6), |
| 144 | LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7) |
| 145 | }; |
| 146 | |
| 147 | uint32_t log2_u32(uint32_t v) |
| 148 | { |
| 149 | uint32_t t, tt; |
| 150 | |
| 151 | if ((tt = (v >> 16))) |
| 152 | return (t = (tt >> 8)) |
| 153 | ? 24 + LogTable256[t] |
| 154 | : 16 + LogTable256[tt]; |
| 155 | else |
| 156 | return (t = (v >> 8)) |
| 157 | ? 8 + LogTable256[t] |
| 158 | : LogTable256[v]; |
| 159 | } |
| 160 | |
| 161 | static |
| 162 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth); |
| 163 | |
| 164 | static |
| 165 | void check_resize(struct rcu_ht *ht, struct rcu_table *t, |
| 166 | uint32_t chain_len) |
| 167 | { |
| 168 | if (chain_len >= BUCKET_SIZE_RESIZE_THRESHOLD) |
| 169 | ht_resize_lazy(ht, t, log2_u32(chain_len)); |
| 170 | } |
| 171 | |
| 172 | static |
| 173 | struct rcu_ht_node *clear_flag(struct rcu_ht_node *node) |
| 174 | { |
| 175 | return (struct rcu_ht_node *) (((unsigned long) node) & ~0x1); |
| 176 | } |
| 177 | |
| 178 | static |
| 179 | int is_removed(struct rcu_ht_node *node) |
| 180 | { |
| 181 | return ((unsigned long) node) & 0x1; |
| 182 | } |
| 183 | |
| 184 | static |
| 185 | struct rcu_ht_node *flag_removed(struct rcu_ht_node *node) |
| 186 | { |
| 187 | return (struct rcu_ht_node *) (((unsigned long) node) | 0x1); |
| 188 | } |
| 189 | |
| 190 | static |
| 191 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
| 192 | { |
| 193 | unsigned long old1, old2; |
| 194 | |
| 195 | old1 = uatomic_read(ptr); |
| 196 | do { |
| 197 | old2 = old1; |
| 198 | if (old2 >= v) |
| 199 | return old2; |
| 200 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
| 201 | return v; |
| 202 | } |
| 203 | |
| 204 | static |
| 205 | void _ht_add(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) |
| 206 | { |
| 207 | struct rcu_ht_node *iter_prev = NULL, *iter = NULL; |
| 208 | |
| 209 | if (!t->size) |
| 210 | return; |
| 211 | for (;;) { |
| 212 | uint32_t chain_len = 0; |
| 213 | |
| 214 | iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); |
| 215 | assert(iter_prev); |
| 216 | assert(iter_prev->reverse_hash <= node->reverse_hash); |
| 217 | for (;;) { |
| 218 | iter = clear_flag(rcu_dereference(iter_prev->next)); |
| 219 | if (unlikely(!iter)) |
| 220 | break; |
| 221 | if (iter->reverse_hash < node->reverse_hash) |
| 222 | break; |
| 223 | iter_prev = iter; |
| 224 | check_resize(ht, t, ++chain_len); |
| 225 | } |
| 226 | /* add in iter_prev->next */ |
| 227 | if (is_removed(iter)) |
| 228 | continue; |
| 229 | assert(node != iter); |
| 230 | node->next = iter; |
| 231 | assert(iter_prev != node); |
| 232 | if (uatomic_cmpxchg(&iter_prev->next, iter, node) != iter) |
| 233 | continue; |
| 234 | break; |
| 235 | } |
| 236 | } |
| 237 | |
| 238 | static |
| 239 | int _ht_remove(struct rcu_ht *ht, struct rcu_table *t, struct rcu_ht_node *node) |
| 240 | { |
| 241 | struct rcu_ht_node *iter_prev, *iter, *next, *old; |
| 242 | unsigned long chain_len; |
| 243 | int found, ret = 0; |
| 244 | int flagged = 0; |
| 245 | |
| 246 | retry: |
| 247 | chain_len = 0; |
| 248 | found = 0; |
| 249 | iter_prev = rcu_dereference(t->tbl[node->hash & (t->size - 1)]); |
| 250 | assert(iter_prev); |
| 251 | assert(iter_prev->reverse_hash <= node->reverse_hash); |
| 252 | for (;;) { |
| 253 | iter = clear_flag(rcu_dereference(iter_prev->next)); |
| 254 | if (unlikely(!iter)) |
| 255 | break; |
| 256 | if (iter->reverse_hash < node->reverse_hash) |
| 257 | break; |
| 258 | if (iter == node) { |
| 259 | found = 1; |
| 260 | break; |
| 261 | } |
| 262 | iter_prev = iter; |
| 263 | } |
| 264 | if (!found) { |
| 265 | ret = -ENOENT; |
| 266 | goto end; |
| 267 | } |
| 268 | next = rcu_dereference(iter->next); |
| 269 | if (!flagged) { |
| 270 | if (is_removed(next)) { |
| 271 | ret = -ENOENT; |
| 272 | goto end; |
| 273 | } |
| 274 | /* set deletion flag */ |
| 275 | if ((old = uatomic_cmpxchg(&iter->next, next, |
| 276 | flag_removed(next))) != next) { |
| 277 | if (old == flag_removed(next)) { |
| 278 | ret = -ENOENT; |
| 279 | goto end; |
| 280 | } else { |
| 281 | goto retry; |
| 282 | } |
| 283 | } |
| 284 | flagged = 1; |
| 285 | } |
| 286 | /* |
| 287 | * Remove the element from the list. Retry if there has been a |
| 288 | * concurrent add (there cannot be a concurrent delete, because |
| 289 | * we won the deletion flag cmpxchg). |
| 290 | */ |
| 291 | if (uatomic_cmpxchg(&iter_prev->next, iter, clear_flag(next)) != iter) |
| 292 | goto retry; |
| 293 | end: |
| 294 | return ret; |
| 295 | } |
| 296 | |
| 297 | static |
| 298 | void init_table(struct rcu_ht *ht, struct rcu_table *t, |
| 299 | unsigned long first, unsigned long len) |
| 300 | { |
| 301 | unsigned long i, end; |
| 302 | |
| 303 | end = first + len; |
| 304 | for (i = first; i < end; i++) { |
| 305 | /* Update table size when power of two */ |
| 306 | if (i != 0 && !(i & (i - 1))) |
| 307 | t->size = i; |
| 308 | t->tbl[i] = calloc(1, sizeof(struct rcu_ht_node)); |
| 309 | t->tbl[i]->dummy = 1; |
| 310 | t->tbl[i]->hash = i; |
| 311 | t->tbl[i]->reverse_hash = bit_reverse_ulong(i); |
| 312 | _ht_add(ht, t, t->tbl[i]); |
| 313 | } |
| 314 | t->resize_target = t->size = end; |
| 315 | } |
| 316 | |
| 317 | struct rcu_ht *ht_new(ht_hash_fct hash_fct, |
| 318 | ht_compare_fct compare_fct, |
| 319 | unsigned long hash_seed, |
| 320 | unsigned long init_size, |
| 321 | void (*ht_call_rcu)(struct rcu_head *head, |
| 322 | void (*func)(struct rcu_head *head))) |
| 323 | { |
| 324 | struct rcu_ht *ht; |
| 325 | |
| 326 | ht = calloc(1, sizeof(struct rcu_ht)); |
| 327 | ht->hash_fct = hash_fct; |
| 328 | ht->compare_fct = compare_fct; |
| 329 | ht->hash_seed = hash_seed; |
| 330 | ht->ht_call_rcu = ht_call_rcu; |
| 331 | /* this mutex should not nest in read-side C.S. */ |
| 332 | pthread_mutex_init(&ht->resize_mutex, NULL); |
| 333 | ht->t = calloc(1, sizeof(struct rcu_table) |
| 334 | + (max(init_size, 1) * sizeof(struct rcu_ht_node *))); |
| 335 | ht->t->size = 0; |
| 336 | pthread_mutex_lock(&ht->resize_mutex); |
| 337 | init_table(ht, ht->t, 0, max(init_size, 1)); |
| 338 | pthread_mutex_unlock(&ht->resize_mutex); |
| 339 | return ht; |
| 340 | } |
| 341 | |
| 342 | struct rcu_ht_node *ht_lookup(struct rcu_ht *ht, void *key, size_t key_len) |
| 343 | { |
| 344 | struct rcu_table *t; |
| 345 | struct rcu_ht_node *node; |
| 346 | unsigned long hash, reverse_hash; |
| 347 | |
| 348 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
| 349 | reverse_hash = bit_reverse_ulong(hash); |
| 350 | |
| 351 | t = rcu_dereference(ht->t); |
| 352 | node = rcu_dereference(t->tbl[hash & (t->size - 1)]); |
| 353 | for (;;) { |
| 354 | if (unlikely(!node)) |
| 355 | break; |
| 356 | if (node->reverse_hash > reverse_hash) { |
| 357 | node = NULL; |
| 358 | break; |
| 359 | } |
| 360 | if (!ht->compare_fct(node->key, node->key_len, key, key_len)) { |
| 361 | if (is_removed(rcu_dereference(node->next))) |
| 362 | node = NULL; |
| 363 | break; |
| 364 | } |
| 365 | node = clear_flag(rcu_dereference(node->next)); |
| 366 | } |
| 367 | return node; |
| 368 | } |
| 369 | |
| 370 | void ht_add(struct rcu_ht *ht, struct rcu_ht_node *node) |
| 371 | { |
| 372 | struct rcu_table *t; |
| 373 | |
| 374 | node->hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
| 375 | node->reverse_hash = bit_reverse_ulong((unsigned long) node->hash); |
| 376 | |
| 377 | t = rcu_dereference(ht->t); |
| 378 | _ht_add(ht, t, node); |
| 379 | } |
| 380 | |
| 381 | int ht_remove(struct rcu_ht *ht, struct rcu_ht_node *node) |
| 382 | { |
| 383 | struct rcu_table *t; |
| 384 | |
| 385 | t = rcu_dereference(ht->t); |
| 386 | return _ht_remove(ht, t, node); |
| 387 | } |
| 388 | |
| 389 | static |
| 390 | int ht_delete_dummy(struct rcu_ht *ht) |
| 391 | { |
| 392 | struct rcu_table *t; |
| 393 | struct rcu_ht_node *node; |
| 394 | unsigned long i; |
| 395 | |
| 396 | t = ht->t; |
| 397 | /* Check that the table is empty */ |
| 398 | node = t->tbl[0]; |
| 399 | do { |
| 400 | if (!node->dummy) |
| 401 | return -EPERM; |
| 402 | node = node->next; |
| 403 | } while (node); |
| 404 | /* Internal sanity check: all nodes left should be dummy */ |
| 405 | for (i = 0; i < t->size; i++) { |
| 406 | assert(t->tbl[i]->dummy); |
| 407 | free(t->tbl[i]); |
| 408 | } |
| 409 | return 0; |
| 410 | } |
| 411 | |
| 412 | /* |
| 413 | * Should only be called when no more concurrent readers nor writers can |
| 414 | * possibly access the table. |
| 415 | */ |
| 416 | int ht_destroy(struct rcu_ht *ht) |
| 417 | { |
| 418 | int ret; |
| 419 | |
| 420 | ret = ht_delete_dummy(ht); |
| 421 | if (ret) |
| 422 | return ret; |
| 423 | free(ht->t); |
| 424 | free(ht); |
| 425 | return ret; |
| 426 | } |
| 427 | |
| 428 | static |
| 429 | void ht_free_table_cb(struct rcu_head *head) |
| 430 | { |
| 431 | struct rcu_table *t = |
| 432 | caa_container_of(head, struct rcu_table, head); |
| 433 | free(t); |
| 434 | } |
| 435 | |
| 436 | /* called with resize mutex held */ |
| 437 | static |
| 438 | void _do_ht_resize(struct rcu_ht *ht) |
| 439 | { |
| 440 | unsigned long new_size, old_size; |
| 441 | struct rcu_table *new_t, *old_t; |
| 442 | |
| 443 | old_t = ht->t; |
| 444 | old_size = old_t->size; |
| 445 | |
| 446 | new_size = CMM_LOAD_SHARED(old_t->resize_target); |
| 447 | dbg_printf("rculfhash: resize from %lu to %lu buckets\n", |
| 448 | old_size, new_size); |
| 449 | if (old_size == new_size) |
| 450 | return; |
| 451 | new_t = malloc(sizeof(struct rcu_table) |
| 452 | + (new_size * sizeof(struct rcu_ht_node *))); |
| 453 | assert(new_size > old_size); |
| 454 | memcpy(&new_t->tbl, &old_t->tbl, |
| 455 | old_size * sizeof(struct rcu_ht_node *)); |
| 456 | init_table(ht, new_t, old_size, new_size - old_size); |
| 457 | /* Changing table and size atomically wrt lookups */ |
| 458 | rcu_assign_pointer(ht->t, new_t); |
| 459 | ht->ht_call_rcu(&old_t->head, ht_free_table_cb); |
| 460 | } |
| 461 | |
| 462 | static |
| 463 | unsigned long resize_target_update(struct rcu_table *t, |
| 464 | int growth_order) |
| 465 | { |
| 466 | return _uatomic_max(&t->resize_target, |
| 467 | t->size << growth_order); |
| 468 | } |
| 469 | |
| 470 | void ht_resize(struct rcu_ht *ht, int growth) |
| 471 | { |
| 472 | struct rcu_table *t = rcu_dereference(ht->t); |
| 473 | unsigned long target_size; |
| 474 | |
| 475 | target_size = resize_target_update(t, growth); |
| 476 | if (t->size < target_size) { |
| 477 | pthread_mutex_lock(&ht->resize_mutex); |
| 478 | _do_ht_resize(ht); |
| 479 | pthread_mutex_unlock(&ht->resize_mutex); |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | static |
| 484 | void do_resize_cb(struct rcu_head *head) |
| 485 | { |
| 486 | struct rcu_resize_work *work = |
| 487 | caa_container_of(head, struct rcu_resize_work, head); |
| 488 | struct rcu_ht *ht = work->ht; |
| 489 | |
| 490 | pthread_mutex_lock(&ht->resize_mutex); |
| 491 | _do_ht_resize(ht); |
| 492 | pthread_mutex_unlock(&ht->resize_mutex); |
| 493 | free(work); |
| 494 | } |
| 495 | |
| 496 | static |
| 497 | void ht_resize_lazy(struct rcu_ht *ht, struct rcu_table *t, int growth) |
| 498 | { |
| 499 | struct rcu_resize_work *work; |
| 500 | unsigned long target_size; |
| 501 | |
| 502 | target_size = resize_target_update(t, growth); |
| 503 | if (t->size < target_size) { |
| 504 | work = malloc(sizeof(*work)); |
| 505 | work->ht = ht; |
| 506 | ht->ht_call_rcu(&work->head, do_resize_cb); |
| 507 | } |
| 508 | } |