| 1 | |
| 2 | /* |
| 3 | * TODO: keys are currently assumed <= sizeof(void *). Key target never freed. |
| 4 | */ |
| 5 | |
| 6 | #define _LGPL_SOURCE |
| 7 | #include <stdlib.h> |
| 8 | #include <errno.h> |
| 9 | #include <assert.h> |
| 10 | #include <stdio.h> |
| 11 | |
| 12 | #include <urcu.h> |
| 13 | #include <urcu-defer.h> |
| 14 | #include <arch.h> |
| 15 | #include <arch_atomic.h> |
| 16 | #include <compiler.h> |
| 17 | #include <urcu/jhash.h> |
| 18 | #include <stdio.h> |
| 19 | #include <pthread.h> |
| 20 | #include <urcu-ht.h> |
| 21 | |
| 22 | /* |
| 23 | * Maximum number of hash table buckets: 256M on 64-bit. |
| 24 | * Should take about 512MB max if we assume 1 node per 4 buckets. |
| 25 | */ |
| 26 | #define MAX_HT_BUCKETS ((256 << 10) / sizeof(void *)) |
| 27 | |
| 28 | /* node flags */ |
| 29 | #define NODE_STOLEN (1 << 0) |
| 30 | |
| 31 | struct rcu_ht_node; |
| 32 | |
| 33 | struct rcu_ht_node { |
| 34 | struct rcu_ht_node *next; |
| 35 | void *key; |
| 36 | void *data; |
| 37 | unsigned int flags; |
| 38 | }; |
| 39 | |
| 40 | struct rcu_table { |
| 41 | unsigned long size; |
| 42 | struct rcu_ht_node *tbl[0]; |
| 43 | }; |
| 44 | |
| 45 | struct rcu_ht { |
| 46 | struct rcu_table *t; /* shared */ |
| 47 | ht_hash_fct hash_fct; |
| 48 | void (*free_fct)(void *data); /* fct to free data */ |
| 49 | uint32_t keylen; |
| 50 | uint32_t hashseed; |
| 51 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
| 52 | int resize_ongoing; /* fast-path resize check */ |
| 53 | }; |
| 54 | |
| 55 | struct rcu_ht *ht_new(ht_hash_fct hash_fct, void (*free_fct)(void *data), |
| 56 | unsigned long init_size, uint32_t keylen, |
| 57 | uint32_t hashseed) |
| 58 | { |
| 59 | struct rcu_ht *ht; |
| 60 | |
| 61 | ht = calloc(1, sizeof(struct rcu_ht)); |
| 62 | ht->hash_fct = hash_fct; |
| 63 | ht->free_fct = free_fct; |
| 64 | ht->keylen = keylen; |
| 65 | ht->hashseed = hashseed; |
| 66 | /* this mutex should not nest in read-side C.S. */ |
| 67 | pthread_mutex_init(&ht->resize_mutex, NULL); |
| 68 | ht->resize_ongoing = 0; /* shared */ |
| 69 | ht->t = calloc(1, sizeof(struct rcu_table) |
| 70 | + (init_size * sizeof(struct rcu_ht_node *))); |
| 71 | ht->t->size = init_size; |
| 72 | return ht; |
| 73 | } |
| 74 | |
| 75 | void *ht_lookup(struct rcu_ht *ht, void *key) |
| 76 | { |
| 77 | struct rcu_table *t; |
| 78 | unsigned long hash; |
| 79 | struct rcu_ht_node *node; |
| 80 | void *ret; |
| 81 | |
| 82 | rcu_read_lock(); |
| 83 | t = rcu_dereference(ht->t); |
| 84 | smp_read_barrier_depends(); /* read t before size and table */ |
| 85 | hash = ht->hash_fct(key, ht->keylen, ht->hashseed) % t->size; |
| 86 | smp_read_barrier_depends(); /* read size before links */ |
| 87 | node = rcu_dereference(t->tbl[hash]); |
| 88 | for (;;) { |
| 89 | if (likely(!node)) { |
| 90 | ret = NULL; |
| 91 | break; |
| 92 | } |
| 93 | if (node->key == key) { |
| 94 | ret = node->data; |
| 95 | break; |
| 96 | } |
| 97 | node = rcu_dereference(node->next); |
| 98 | } |
| 99 | rcu_read_unlock(); |
| 100 | |
| 101 | return ret; |
| 102 | } |
| 103 | |
| 104 | /* |
| 105 | * Will re-try until either: |
| 106 | * - The key is already there (-EEXIST) |
| 107 | * - We successfully add the key at the head of a table bucket. |
| 108 | */ |
| 109 | int ht_add(struct rcu_ht *ht, void *key, void *data) |
| 110 | { |
| 111 | struct rcu_ht_node *node, *old_head, *new_head; |
| 112 | struct rcu_table *t; |
| 113 | unsigned long hash; |
| 114 | int ret = 0; |
| 115 | |
| 116 | new_head = calloc(1, sizeof(struct rcu_ht_node)); |
| 117 | new_head->key = key; |
| 118 | new_head->data = data; |
| 119 | new_head->flags = 0; |
| 120 | /* here comes the fun and tricky part. |
| 121 | * Add at the beginning with a cmpxchg. |
| 122 | * Hold a read lock between the moment the first element is read |
| 123 | * and the nodes traversal (to find duplicates). This ensures |
| 124 | * the head pointer has not been reclaimed when cmpxchg is done. |
| 125 | * Always adding at the head ensures that we would have to |
| 126 | * re-try if a new item has been added concurrently. So we ensure that |
| 127 | * we never add duplicates. */ |
| 128 | retry: |
| 129 | rcu_read_lock(); |
| 130 | |
| 131 | if (unlikely(LOAD_SHARED(ht->resize_ongoing))) { |
| 132 | rcu_read_unlock(); |
| 133 | /* |
| 134 | * Wait for resize to complete before continuing. |
| 135 | */ |
| 136 | ret = pthread_mutex_lock(&ht->resize_mutex); |
| 137 | assert(!ret); |
| 138 | ret = pthread_mutex_unlock(&ht->resize_mutex); |
| 139 | assert(!ret); |
| 140 | goto retry; |
| 141 | } |
| 142 | |
| 143 | t = rcu_dereference(ht->t); |
| 144 | /* no read barrier needed, because no concurrency with resize */ |
| 145 | hash = ht->hash_fct(key, ht->keylen, ht->hashseed) % t->size; |
| 146 | |
| 147 | old_head = node = rcu_dereference(t->tbl[hash]); |
| 148 | for (;;) { |
| 149 | if (likely(!node)) { |
| 150 | break; |
| 151 | } |
| 152 | if (node->key == key) { |
| 153 | ret = -EEXIST; |
| 154 | goto end; |
| 155 | } |
| 156 | node = rcu_dereference(node->next); |
| 157 | } |
| 158 | new_head->next = old_head; |
| 159 | if (rcu_cmpxchg_pointer(&t->tbl[hash], old_head, new_head) != old_head) |
| 160 | goto restart; |
| 161 | end: |
| 162 | rcu_read_unlock(); |
| 163 | return ret; |
| 164 | |
| 165 | /* restart loop, release and re-take the read lock to be kind to GP */ |
| 166 | restart: |
| 167 | rcu_read_unlock(); |
| 168 | goto retry; |
| 169 | } |
| 170 | |
| 171 | /* |
| 172 | * Restart until we successfully remove the entry, or no entry is left |
| 173 | * ((void *)(unsigned long)-ENOENT). |
| 174 | * Deal with concurrent stealers by doing an extra verification pass to check |
| 175 | * that no element in the list are still pointing to the element stolen. |
| 176 | * This could happen if two concurrent steal for consecutive objects are |
| 177 | * executed. A pointer to an object being stolen could be saved by the |
| 178 | * concurrent stealer for the previous object. |
| 179 | * Also, given that in this precise scenario, another stealer can also want to |
| 180 | * delete the doubly-referenced object; use a "stolen" flag to let only one |
| 181 | * stealer delete the object. |
| 182 | */ |
| 183 | void *ht_steal(struct rcu_ht *ht, void *key) |
| 184 | { |
| 185 | struct rcu_ht_node **prev, *node, *del_node = NULL; |
| 186 | struct rcu_table *t; |
| 187 | unsigned long hash; |
| 188 | void *data; |
| 189 | int ret; |
| 190 | |
| 191 | retry: |
| 192 | rcu_read_lock(); |
| 193 | |
| 194 | if (unlikely(LOAD_SHARED(ht->resize_ongoing))) { |
| 195 | rcu_read_unlock(); |
| 196 | /* |
| 197 | * Wait for resize to complete before continuing. |
| 198 | */ |
| 199 | ret = pthread_mutex_lock(&ht->resize_mutex); |
| 200 | assert(!ret); |
| 201 | ret = pthread_mutex_unlock(&ht->resize_mutex); |
| 202 | assert(!ret); |
| 203 | goto retry; |
| 204 | } |
| 205 | |
| 206 | t = rcu_dereference(ht->t); |
| 207 | /* no read barrier needed, because no concurrency with resize */ |
| 208 | hash = ht->hash_fct(key, ht->keylen, ht->hashseed) % t->size; |
| 209 | |
| 210 | prev = &t->tbl[hash]; |
| 211 | node = rcu_dereference(*prev); |
| 212 | for (;;) { |
| 213 | if (likely(!node)) { |
| 214 | if (del_node) { |
| 215 | goto end; |
| 216 | } else { |
| 217 | goto error; |
| 218 | } |
| 219 | } |
| 220 | if (node->key == key) { |
| 221 | break; |
| 222 | } |
| 223 | prev = &node->next; |
| 224 | node = rcu_dereference(*prev); |
| 225 | } |
| 226 | |
| 227 | if (!del_node) { |
| 228 | /* |
| 229 | * Another concurrent thread stole it ? If so, let it deal with |
| 230 | * this. Assume NODE_STOLEN is the only flag. If this changes, |
| 231 | * read flags before cmpxchg. |
| 232 | */ |
| 233 | if (cmpxchg(&node->flags, 0, NODE_STOLEN) != 0) |
| 234 | goto error; |
| 235 | } |
| 236 | |
| 237 | /* Found it ! pointer to object is in "prev" */ |
| 238 | if (rcu_cmpxchg_pointer(prev, node, node->next) == node) |
| 239 | del_node = node; |
| 240 | goto restart; |
| 241 | |
| 242 | end: |
| 243 | /* |
| 244 | * From that point, we own node. Note that there can still be concurrent |
| 245 | * RCU readers using it. We can free it outside of read lock after a GP. |
| 246 | */ |
| 247 | rcu_read_unlock(); |
| 248 | |
| 249 | data = del_node->data; |
| 250 | call_rcu(free, del_node); |
| 251 | return data; |
| 252 | |
| 253 | error: |
| 254 | data = (void *)(unsigned long)-ENOENT; |
| 255 | rcu_read_unlock(); |
| 256 | return data; |
| 257 | |
| 258 | /* restart loop, release and re-take the read lock to be kind to GP */ |
| 259 | restart: |
| 260 | rcu_read_unlock(); |
| 261 | goto retry; |
| 262 | } |
| 263 | |
| 264 | int ht_delete(struct rcu_ht *ht, void *key) |
| 265 | { |
| 266 | void *data; |
| 267 | |
| 268 | data = ht_steal(ht, key); |
| 269 | if (data && data != (void *)(unsigned long)-ENOENT) { |
| 270 | if (ht->free_fct) |
| 271 | call_rcu(ht->free_fct, data); |
| 272 | return 0; |
| 273 | } else { |
| 274 | return -ENOENT; |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | /* Delete all old elements. Allow concurrent writer accesses. */ |
| 279 | int ht_delete_all(struct rcu_ht *ht) |
| 280 | { |
| 281 | unsigned long i; |
| 282 | struct rcu_ht_node **prev, *node, *inext; |
| 283 | struct rcu_table *t; |
| 284 | int cnt = 0; |
| 285 | int ret; |
| 286 | |
| 287 | /* |
| 288 | * Mutual exclusion with resize operations, but leave add/steal execute |
| 289 | * concurrently. This is OK because we operate only on the heads. |
| 290 | */ |
| 291 | ret = pthread_mutex_lock(&ht->resize_mutex); |
| 292 | assert(!ret); |
| 293 | |
| 294 | t = rcu_dereference(ht->t); |
| 295 | /* no read barrier needed, because no concurrency with resize */ |
| 296 | for (i = 0; i < t->size; i++) { |
| 297 | rcu_read_lock(); |
| 298 | prev = &t->tbl[i]; |
| 299 | /* |
| 300 | * Cut the head. After that, we own the first element. |
| 301 | */ |
| 302 | node = rcu_xchg_pointer(prev, NULL); |
| 303 | if (!node) { |
| 304 | rcu_read_unlock(); |
| 305 | continue; |
| 306 | } |
| 307 | /* |
| 308 | * We manage a list shared with concurrent writers and readers. |
| 309 | * Note that a concurrent add may or may not be deleted by us, |
| 310 | * depending if it arrives before or after the head is cut. |
| 311 | * "node" points to our first node. Remove first elements |
| 312 | * iteratively. |
| 313 | */ |
| 314 | for (;;) { |
| 315 | inext = NULL; |
| 316 | prev = &node->next; |
| 317 | if (prev) |
| 318 | inext = rcu_xchg_pointer(prev, NULL); |
| 319 | /* |
| 320 | * "node" is the first element of the list we have cut. |
| 321 | * We therefore own it, no concurrent writer may delete |
| 322 | * it. There can only be concurrent lookups. Concurrent |
| 323 | * add can only be done on a bucket head, but we've cut |
| 324 | * it already. inext is also owned by us, because we |
| 325 | * have exchanged it for "NULL". It will therefore be |
| 326 | * safe to use it after a G.P. |
| 327 | */ |
| 328 | rcu_read_unlock(); |
| 329 | if (node->data) |
| 330 | call_rcu(ht->free_fct, node->data); |
| 331 | call_rcu(free, node); |
| 332 | cnt++; |
| 333 | if (likely(!inext)) |
| 334 | break; |
| 335 | rcu_read_lock(); |
| 336 | node = inext; |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | ret = pthread_mutex_unlock(&ht->resize_mutex); |
| 341 | assert(!ret); |
| 342 | return cnt; |
| 343 | } |
| 344 | |
| 345 | /* |
| 346 | * Should only be called when no more concurrent readers nor writers can |
| 347 | * possibly access the table. |
| 348 | */ |
| 349 | int ht_destroy(struct rcu_ht *ht) |
| 350 | { |
| 351 | int ret; |
| 352 | |
| 353 | ret = ht_delete_all(ht); |
| 354 | free(ht->t); |
| 355 | free(ht); |
| 356 | return ret; |
| 357 | } |
| 358 | |
| 359 | static void ht_resize_grow(struct rcu_ht *ht) |
| 360 | { |
| 361 | unsigned long i, new_size, old_size; |
| 362 | struct rcu_table *new_t, *old_t; |
| 363 | struct rcu_ht_node *node, *new_node, *tmp; |
| 364 | unsigned long hash; |
| 365 | |
| 366 | old_t = ht->t; |
| 367 | old_size = old_t->size; |
| 368 | |
| 369 | if (old_size == MAX_HT_BUCKETS) |
| 370 | return; |
| 371 | |
| 372 | new_size = old_size << 1; |
| 373 | new_t = calloc(1, sizeof(struct rcu_table) |
| 374 | + (new_size * sizeof(struct rcu_ht_node *))); |
| 375 | |
| 376 | for (i = 0; i < old_size; i++) { |
| 377 | /* |
| 378 | * Re-hash each entry, insert in new table. |
| 379 | * It's important that a reader looking for a key _will_ find it |
| 380 | * if it's in the table. |
| 381 | * Copy each node. (just the node, not ->data) |
| 382 | */ |
| 383 | node = old_t->tbl[i]; |
| 384 | while (node) { |
| 385 | hash = ht->hash_fct(node->key, ht->keylen, ht->hashseed) |
| 386 | % new_size; |
| 387 | new_node = malloc(sizeof(struct rcu_ht_node)); |
| 388 | new_node->key = node->key; |
| 389 | new_node->data = node->data; |
| 390 | new_node->next = new_t->tbl[hash]; /* add to head */ |
| 391 | new_t->tbl[hash] = new_node; |
| 392 | node = node->next; |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | smp_wmb(); /* write links before changing table */ |
| 397 | ht->t = new_t; /* Changing table and size atomically wrt lookups */ |
| 398 | |
| 399 | /* Ensure all concurrent lookups use new size and table */ |
| 400 | synchronize_rcu(); |
| 401 | |
| 402 | for (i = 0; i < old_size; i++) { |
| 403 | node = old_t->tbl[i]; |
| 404 | while (node) { |
| 405 | tmp = node->next; |
| 406 | free(node); |
| 407 | node = tmp; |
| 408 | } |
| 409 | } |
| 410 | free(old_t); |
| 411 | } |
| 412 | |
| 413 | static void ht_resize_shrink(struct rcu_ht *ht) |
| 414 | { |
| 415 | unsigned long i, new_size; |
| 416 | struct rcu_table *new_t, *old_t; |
| 417 | struct rcu_ht_node **prev, *node; |
| 418 | |
| 419 | old_t = ht->t; |
| 420 | if (old_t->size == 1) |
| 421 | return; |
| 422 | |
| 423 | new_size = old_t->size >> 1; |
| 424 | |
| 425 | for (i = 0; i < new_size; i++) { |
| 426 | /* Link end with first entry of i + new_size */ |
| 427 | prev = &old_t->tbl[i]; |
| 428 | node = *prev; |
| 429 | while (node) { |
| 430 | prev = &node->next; |
| 431 | node = *prev; |
| 432 | } |
| 433 | *prev = old_t->tbl[i + new_size]; |
| 434 | } |
| 435 | smp_wmb(); /* write links before changing size */ |
| 436 | STORE_SHARED(old_t->size, new_size); |
| 437 | |
| 438 | /* Ensure all concurrent lookups use new size */ |
| 439 | synchronize_rcu(); |
| 440 | |
| 441 | new_t = realloc(old_t, sizeof(struct rcu_table) |
| 442 | + (new_size * sizeof(struct rcu_ht_node *))); |
| 443 | /* shrinking, pointers should not move */ |
| 444 | assert(new_t == old_t); |
| 445 | } |
| 446 | |
| 447 | /* |
| 448 | * growth: >0: *2, <0: /2 |
| 449 | */ |
| 450 | void ht_resize(struct rcu_ht *ht, int growth) |
| 451 | { |
| 452 | int ret; |
| 453 | |
| 454 | ret = pthread_mutex_lock(&ht->resize_mutex); |
| 455 | assert(!ret); |
| 456 | STORE_SHARED(ht->resize_ongoing, 1); |
| 457 | synchronize_rcu(); |
| 458 | /* All add/remove are waiting on the mutex. */ |
| 459 | if (growth > 0) |
| 460 | ht_resize_grow(ht); |
| 461 | else if (growth < 0) |
| 462 | ht_resize_shrink(ht); |
| 463 | smp_mb(); |
| 464 | STORE_SHARED(ht->resize_ongoing, 0); |
| 465 | ret = pthread_mutex_unlock(&ht->resize_mutex); |
| 466 | assert(!ret); |
| 467 | } |
| 468 | |
| 469 | /* |
| 470 | * Expects keys <= than pointer size to be encoded in the pointer itself. |
| 471 | */ |
| 472 | uint32_t ht_jhash(void *key, uint32_t length, uint32_t initval) |
| 473 | { |
| 474 | uint32_t ret; |
| 475 | void *vkey; |
| 476 | |
| 477 | if (length <= sizeof(void *)) |
| 478 | vkey = &key; |
| 479 | else |
| 480 | vkey = key; |
| 481 | ret = jhash(vkey, length, initval); |
| 482 | return ret; |
| 483 | } |