| 1 | // SPDX-FileCopyrightText: 2010-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 2 | // |
| 3 | // SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 | |
| 5 | #ifndef _URCU_STATIC_WFSTACK_H |
| 6 | #define _URCU_STATIC_WFSTACK_H |
| 7 | |
| 8 | /* |
| 9 | * Userspace RCU library - Stack with with wait-free push, blocking traversal. |
| 10 | * |
| 11 | * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu/wfstack.h for |
| 12 | * linking dynamically with the userspace rcu library. |
| 13 | */ |
| 14 | |
| 15 | #include <pthread.h> |
| 16 | #include <poll.h> |
| 17 | #include <stdbool.h> |
| 18 | #include <urcu/assert.h> |
| 19 | #include <urcu/compiler.h> |
| 20 | #include <urcu/uatomic.h> |
| 21 | |
| 22 | #ifdef __cplusplus |
| 23 | extern "C" { |
| 24 | #endif |
| 25 | |
| 26 | #define CDS_WFS_END ((struct cds_wfs_head *) 0x1UL) |
| 27 | #define CDS_WFS_ADAPT_ATTEMPTS 10 /* Retry if being set */ |
| 28 | #define CDS_WFS_WAIT 10 /* Wait 10 ms if being set */ |
| 29 | |
| 30 | /* |
| 31 | * Stack with wait-free push, blocking traversal. |
| 32 | * |
| 33 | * Stack implementing push, pop, pop_all operations, as well as iterator |
| 34 | * on the stack head returned by pop_all. |
| 35 | * |
| 36 | * Wait-free operations: cds_wfs_push, __cds_wfs_pop_all, cds_wfs_empty, |
| 37 | * cds_wfs_first. |
| 38 | * Blocking operations: cds_wfs_pop, cds_wfs_pop_all, cds_wfs_next, |
| 39 | * iteration on stack head returned by pop_all. |
| 40 | * |
| 41 | * Synchronization table: |
| 42 | * |
| 43 | * External synchronization techniques described in the API below is |
| 44 | * required between pairs marked with "X". No external synchronization |
| 45 | * required between pairs marked with "-". |
| 46 | * |
| 47 | * cds_wfs_push __cds_wfs_pop __cds_wfs_pop_all |
| 48 | * cds_wfs_push - - - |
| 49 | * __cds_wfs_pop - X X |
| 50 | * __cds_wfs_pop_all - X - |
| 51 | * |
| 52 | * cds_wfs_pop and cds_wfs_pop_all use an internal mutex to provide |
| 53 | * synchronization. |
| 54 | */ |
| 55 | |
| 56 | /* |
| 57 | * cds_wfs_node_init: initialize wait-free stack node. |
| 58 | */ |
| 59 | static inline |
| 60 | void _cds_wfs_node_init(struct cds_wfs_node *node) |
| 61 | { |
| 62 | node->next = NULL; |
| 63 | } |
| 64 | |
| 65 | /* |
| 66 | * __cds_wfs_init: initialize wait-free stack. Don't pair with |
| 67 | * any destroy function. |
| 68 | */ |
| 69 | static inline void ___cds_wfs_init(struct __cds_wfs_stack *s) |
| 70 | { |
| 71 | s->head = CDS_WFS_END; |
| 72 | } |
| 73 | |
| 74 | /* |
| 75 | * cds_wfs_init: initialize wait-free stack. Pair with |
| 76 | * cds_wfs_destroy(). |
| 77 | */ |
| 78 | static inline |
| 79 | void _cds_wfs_init(struct cds_wfs_stack *s) |
| 80 | { |
| 81 | int ret; |
| 82 | |
| 83 | s->head = CDS_WFS_END; |
| 84 | ret = pthread_mutex_init(&s->lock, NULL); |
| 85 | urcu_posix_assert(!ret); |
| 86 | } |
| 87 | |
| 88 | /* |
| 89 | * cds_wfs_destroy: destroy wait-free stack. Pair with |
| 90 | * cds_wfs_init(). |
| 91 | */ |
| 92 | static inline |
| 93 | void _cds_wfs_destroy(struct cds_wfs_stack *s) |
| 94 | { |
| 95 | int ret = pthread_mutex_destroy(&s->lock); |
| 96 | urcu_posix_assert(!ret); |
| 97 | } |
| 98 | |
| 99 | static inline bool ___cds_wfs_end(void *node) |
| 100 | { |
| 101 | return node == CDS_WFS_END; |
| 102 | } |
| 103 | |
| 104 | /* |
| 105 | * cds_wfs_empty: return whether wait-free stack is empty. |
| 106 | * |
| 107 | * No memory barrier is issued. No mutual exclusion is required. |
| 108 | */ |
| 109 | static inline bool _cds_wfs_empty(cds_wfs_stack_ptr_t u_stack) |
| 110 | { |
| 111 | struct __cds_wfs_stack *s = u_stack._s; |
| 112 | |
| 113 | return ___cds_wfs_end(CMM_LOAD_SHARED(s->head)); |
| 114 | } |
| 115 | |
| 116 | /* |
| 117 | * cds_wfs_push: push a node into the stack. |
| 118 | * |
| 119 | * Issues a full memory barrier before push. No mutual exclusion is |
| 120 | * required. |
| 121 | * |
| 122 | * Returns 0 if the stack was empty prior to adding the node. |
| 123 | * Returns non-zero otherwise. |
| 124 | */ |
| 125 | static inline |
| 126 | int _cds_wfs_push(cds_wfs_stack_ptr_t u_stack, struct cds_wfs_node *node) |
| 127 | { |
| 128 | struct __cds_wfs_stack *s = u_stack._s; |
| 129 | struct cds_wfs_head *old_head, *new_head; |
| 130 | |
| 131 | urcu_posix_assert(node->next == NULL); |
| 132 | new_head = caa_container_of(node, struct cds_wfs_head, node); |
| 133 | /* |
| 134 | * uatomic_xchg() implicit memory barrier orders earlier stores |
| 135 | * to node (setting it to NULL) before publication. |
| 136 | */ |
| 137 | old_head = uatomic_xchg(&s->head, new_head); |
| 138 | /* |
| 139 | * At this point, dequeuers see a NULL node->next, they should |
| 140 | * busy-wait until node->next is set to old_head. |
| 141 | */ |
| 142 | CMM_STORE_SHARED(node->next, &old_head->node); |
| 143 | return !___cds_wfs_end(old_head); |
| 144 | } |
| 145 | |
| 146 | /* |
| 147 | * Waiting for push to complete enqueue and return the next node. |
| 148 | */ |
| 149 | static inline struct cds_wfs_node * |
| 150 | ___cds_wfs_node_sync_next(struct cds_wfs_node *node, int blocking) |
| 151 | { |
| 152 | struct cds_wfs_node *next; |
| 153 | int attempt = 0; |
| 154 | |
| 155 | /* |
| 156 | * Adaptative busy-looping waiting for push to complete. |
| 157 | */ |
| 158 | while ((next = CMM_LOAD_SHARED(node->next)) == NULL) { |
| 159 | if (!blocking) |
| 160 | return CDS_WFS_WOULDBLOCK; |
| 161 | if (++attempt >= CDS_WFS_ADAPT_ATTEMPTS) { |
| 162 | (void) poll(NULL, 0, CDS_WFS_WAIT); /* Wait for 10ms */ |
| 163 | attempt = 0; |
| 164 | } else { |
| 165 | caa_cpu_relax(); |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | return next; |
| 170 | } |
| 171 | |
| 172 | static inline |
| 173 | struct cds_wfs_node * |
| 174 | ___cds_wfs_pop(cds_wfs_stack_ptr_t u_stack, int *state, int blocking) |
| 175 | { |
| 176 | struct cds_wfs_head *head, *new_head; |
| 177 | struct cds_wfs_node *next; |
| 178 | struct __cds_wfs_stack *s = u_stack._s; |
| 179 | |
| 180 | if (state) |
| 181 | *state = 0; |
| 182 | for (;;) { |
| 183 | head = CMM_LOAD_SHARED(s->head); |
| 184 | if (___cds_wfs_end(head)) { |
| 185 | return NULL; |
| 186 | } |
| 187 | next = ___cds_wfs_node_sync_next(&head->node, blocking); |
| 188 | if (!blocking && next == CDS_WFS_WOULDBLOCK) { |
| 189 | return CDS_WFS_WOULDBLOCK; |
| 190 | } |
| 191 | new_head = caa_container_of(next, struct cds_wfs_head, node); |
| 192 | if (uatomic_cmpxchg(&s->head, head, new_head) == head) { |
| 193 | if (state && ___cds_wfs_end(new_head)) |
| 194 | *state |= CDS_WFS_STATE_LAST; |
| 195 | return &head->node; |
| 196 | } |
| 197 | if (!blocking) { |
| 198 | return CDS_WFS_WOULDBLOCK; |
| 199 | } |
| 200 | /* busy-loop if head changed under us */ |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | /* |
| 205 | * __cds_wfs_pop_with_state_blocking: pop a node from the stack, with state. |
| 206 | * |
| 207 | * Returns NULL if stack is empty. |
| 208 | * |
| 209 | * __cds_wfs_pop_blocking needs to be synchronized using one of the |
| 210 | * following techniques: |
| 211 | * |
| 212 | * 1) Calling __cds_wfs_pop_blocking under rcu read lock critical |
| 213 | * section. The caller must wait for a grace period to pass before |
| 214 | * freeing the returned node or modifying the cds_wfs_node structure. |
| 215 | * 2) Using mutual exclusion (e.g. mutexes) to protect |
| 216 | * __cds_wfs_pop_blocking and __cds_wfs_pop_all callers. |
| 217 | * 3) Ensuring that only ONE thread can call __cds_wfs_pop_blocking() |
| 218 | * and __cds_wfs_pop_all(). (multi-provider/single-consumer scheme). |
| 219 | * |
| 220 | * "state" saves state flags atomically sampled with pop operation. |
| 221 | */ |
| 222 | static inline |
| 223 | struct cds_wfs_node * |
| 224 | ___cds_wfs_pop_with_state_blocking(cds_wfs_stack_ptr_t u_stack, int *state) |
| 225 | { |
| 226 | return ___cds_wfs_pop(u_stack, state, 1); |
| 227 | } |
| 228 | |
| 229 | static inline |
| 230 | struct cds_wfs_node * |
| 231 | ___cds_wfs_pop_blocking(cds_wfs_stack_ptr_t u_stack) |
| 232 | { |
| 233 | return ___cds_wfs_pop_with_state_blocking(u_stack, NULL); |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * __cds_wfs_pop_with_state_nonblocking: pop a node from the stack. |
| 238 | * |
| 239 | * Same as __cds_wfs_pop_with_state_blocking, but returns |
| 240 | * CDS_WFS_WOULDBLOCK if it needs to block. |
| 241 | * |
| 242 | * "state" saves state flags atomically sampled with pop operation. |
| 243 | */ |
| 244 | static inline |
| 245 | struct cds_wfs_node * |
| 246 | ___cds_wfs_pop_with_state_nonblocking(cds_wfs_stack_ptr_t u_stack, int *state) |
| 247 | { |
| 248 | return ___cds_wfs_pop(u_stack, state, 0); |
| 249 | } |
| 250 | |
| 251 | /* |
| 252 | * __cds_wfs_pop_nonblocking: pop a node from the stack. |
| 253 | * |
| 254 | * Same as __cds_wfs_pop_blocking, but returns CDS_WFS_WOULDBLOCK if |
| 255 | * it needs to block. |
| 256 | */ |
| 257 | static inline |
| 258 | struct cds_wfs_node * |
| 259 | ___cds_wfs_pop_nonblocking(cds_wfs_stack_ptr_t u_stack) |
| 260 | { |
| 261 | return ___cds_wfs_pop_with_state_nonblocking(u_stack, NULL); |
| 262 | } |
| 263 | |
| 264 | /* |
| 265 | * __cds_wfs_pop_all: pop all nodes from a stack. |
| 266 | * |
| 267 | * __cds_wfs_pop_all does not require any synchronization with other |
| 268 | * push, nor with other __cds_wfs_pop_all, but requires synchronization |
| 269 | * matching the technique used to synchronize __cds_wfs_pop_blocking: |
| 270 | * |
| 271 | * 1) If __cds_wfs_pop_blocking is called under rcu read lock critical |
| 272 | * section, both __cds_wfs_pop_blocking and cds_wfs_pop_all callers |
| 273 | * must wait for a grace period to pass before freeing the returned |
| 274 | * node or modifying the cds_wfs_node structure. However, no RCU |
| 275 | * read-side critical section is needed around __cds_wfs_pop_all. |
| 276 | * 2) Using mutual exclusion (e.g. mutexes) to protect |
| 277 | * __cds_wfs_pop_blocking and __cds_wfs_pop_all callers. |
| 278 | * 3) Ensuring that only ONE thread can call __cds_wfs_pop_blocking() |
| 279 | * and __cds_wfs_pop_all(). (multi-provider/single-consumer scheme). |
| 280 | */ |
| 281 | static inline |
| 282 | struct cds_wfs_head * |
| 283 | ___cds_wfs_pop_all(cds_wfs_stack_ptr_t u_stack) |
| 284 | { |
| 285 | struct __cds_wfs_stack *s = u_stack._s; |
| 286 | struct cds_wfs_head *head; |
| 287 | |
| 288 | /* |
| 289 | * Implicit memory barrier after uatomic_xchg() matches implicit |
| 290 | * memory barrier before uatomic_xchg() in cds_wfs_push. It |
| 291 | * ensures that all nodes of the returned list are consistent. |
| 292 | * There is no need to issue memory barriers when iterating on |
| 293 | * the returned list, because the full memory barrier issued |
| 294 | * prior to each uatomic_cmpxchg, which each write to head, are |
| 295 | * taking care to order writes to each node prior to the full |
| 296 | * memory barrier after this uatomic_xchg(). |
| 297 | */ |
| 298 | head = uatomic_xchg(&s->head, CDS_WFS_END); |
| 299 | if (___cds_wfs_end(head)) |
| 300 | return NULL; |
| 301 | return head; |
| 302 | } |
| 303 | |
| 304 | /* |
| 305 | * cds_wfs_pop_lock: lock stack pop-protection mutex. |
| 306 | */ |
| 307 | static inline void _cds_wfs_pop_lock(struct cds_wfs_stack *s) |
| 308 | { |
| 309 | int ret; |
| 310 | |
| 311 | ret = pthread_mutex_lock(&s->lock); |
| 312 | urcu_posix_assert(!ret); |
| 313 | } |
| 314 | |
| 315 | /* |
| 316 | * cds_wfs_pop_unlock: unlock stack pop-protection mutex. |
| 317 | */ |
| 318 | static inline void _cds_wfs_pop_unlock(struct cds_wfs_stack *s) |
| 319 | { |
| 320 | int ret; |
| 321 | |
| 322 | ret = pthread_mutex_unlock(&s->lock); |
| 323 | urcu_posix_assert(!ret); |
| 324 | } |
| 325 | |
| 326 | /* |
| 327 | * Call __cds_wfs_pop_with_state_blocking with an internal pop mutex held. |
| 328 | */ |
| 329 | static inline |
| 330 | struct cds_wfs_node * |
| 331 | _cds_wfs_pop_with_state_blocking(struct cds_wfs_stack *s, int *state) |
| 332 | { |
| 333 | struct cds_wfs_node *retnode; |
| 334 | cds_wfs_stack_ptr_t stack; |
| 335 | |
| 336 | _cds_wfs_pop_lock(s); |
| 337 | stack.s = s; |
| 338 | retnode = ___cds_wfs_pop_with_state_blocking(stack, state); |
| 339 | _cds_wfs_pop_unlock(s); |
| 340 | return retnode; |
| 341 | } |
| 342 | |
| 343 | /* |
| 344 | * Call _cds_wfs_pop_with_state_blocking without saving any state. |
| 345 | */ |
| 346 | static inline |
| 347 | struct cds_wfs_node * |
| 348 | _cds_wfs_pop_blocking(struct cds_wfs_stack *s) |
| 349 | { |
| 350 | return _cds_wfs_pop_with_state_blocking(s, NULL); |
| 351 | } |
| 352 | |
| 353 | /* |
| 354 | * Call __cds_wfs_pop_all with an internal pop mutex held. |
| 355 | */ |
| 356 | static inline |
| 357 | struct cds_wfs_head * |
| 358 | _cds_wfs_pop_all_blocking(struct cds_wfs_stack *s) |
| 359 | { |
| 360 | struct cds_wfs_head *rethead; |
| 361 | cds_wfs_stack_ptr_t stack; |
| 362 | |
| 363 | _cds_wfs_pop_lock(s); |
| 364 | stack.s = s; |
| 365 | rethead = ___cds_wfs_pop_all(stack); |
| 366 | _cds_wfs_pop_unlock(s); |
| 367 | return rethead; |
| 368 | } |
| 369 | |
| 370 | /* |
| 371 | * cds_wfs_first: get first node of a popped stack. |
| 372 | * |
| 373 | * Content written into the node before enqueue is guaranteed to be |
| 374 | * consistent, but no other memory ordering is ensured. |
| 375 | * |
| 376 | * Used by for-like iteration macros in urcu/wfstack.h: |
| 377 | * cds_wfs_for_each_blocking() |
| 378 | * cds_wfs_for_each_blocking_safe() |
| 379 | * |
| 380 | * Returns NULL if popped stack is empty, top stack node otherwise. |
| 381 | */ |
| 382 | static inline struct cds_wfs_node * |
| 383 | _cds_wfs_first(struct cds_wfs_head *head) |
| 384 | { |
| 385 | if (___cds_wfs_end(head)) |
| 386 | return NULL; |
| 387 | return &head->node; |
| 388 | } |
| 389 | |
| 390 | static inline struct cds_wfs_node * |
| 391 | ___cds_wfs_next(struct cds_wfs_node *node, int blocking) |
| 392 | { |
| 393 | struct cds_wfs_node *next; |
| 394 | |
| 395 | next = ___cds_wfs_node_sync_next(node, blocking); |
| 396 | /* |
| 397 | * CDS_WFS_WOULDBLOCK != CSD_WFS_END, so we can check for end |
| 398 | * even if ___cds_wfs_node_sync_next returns CDS_WFS_WOULDBLOCK, |
| 399 | * and still return CDS_WFS_WOULDBLOCK. |
| 400 | */ |
| 401 | if (___cds_wfs_end(next)) |
| 402 | return NULL; |
| 403 | return next; |
| 404 | } |
| 405 | |
| 406 | /* |
| 407 | * cds_wfs_next_blocking: get next node of a popped stack. |
| 408 | * |
| 409 | * Content written into the node before enqueue is guaranteed to be |
| 410 | * consistent, but no other memory ordering is ensured. |
| 411 | * |
| 412 | * Used by for-like iteration macros in urcu/wfstack.h: |
| 413 | * cds_wfs_for_each_blocking() |
| 414 | * cds_wfs_for_each_blocking_safe() |
| 415 | * |
| 416 | * Returns NULL if reached end of popped stack, non-NULL next stack |
| 417 | * node otherwise. |
| 418 | */ |
| 419 | static inline struct cds_wfs_node * |
| 420 | _cds_wfs_next_blocking(struct cds_wfs_node *node) |
| 421 | { |
| 422 | return ___cds_wfs_next(node, 1); |
| 423 | } |
| 424 | |
| 425 | |
| 426 | /* |
| 427 | * cds_wfs_next_nonblocking: get next node of a popped stack. |
| 428 | * |
| 429 | * Same as cds_wfs_next_blocking, but returns CDS_WFS_WOULDBLOCK if it |
| 430 | * needs to block. |
| 431 | */ |
| 432 | static inline struct cds_wfs_node * |
| 433 | _cds_wfs_next_nonblocking(struct cds_wfs_node *node) |
| 434 | { |
| 435 | return ___cds_wfs_next(node, 0); |
| 436 | } |
| 437 | |
| 438 | #ifdef __cplusplus |
| 439 | } |
| 440 | #endif |
| 441 | |
| 442 | #endif /* _URCU_STATIC_WFSTACK_H */ |