Commit | Line | Data |
---|---|---|
d3d3857f MJ |
1 | // SPDX-FileCopyrightText: 2010-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
2 | // | |
3 | // SPDX-License-Identifier: LGPL-2.1-or-later | |
4 | ||
d3bfcb24 MD |
5 | #ifndef _URCU_STATIC_LFSTACK_H |
6 | #define _URCU_STATIC_LFSTACK_H | |
7 | ||
8 | /* | |
d3bfcb24 MD |
9 | * Userspace RCU library - Lock-Free Stack |
10 | * | |
07c2a4fd MD |
11 | * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu/lfstack.h for |
12 | * linking dynamically with the userspace rcu library. | |
d3bfcb24 MD |
13 | */ |
14 | ||
7294103b MD |
15 | #include <stdbool.h> |
16 | #include <pthread.h> | |
01477510 | 17 | #include <urcu/assert.h> |
d3bfcb24 MD |
18 | #include <urcu/uatomic.h> |
19 | #include <urcu-pointer.h> | |
20 | ||
21 | #ifdef __cplusplus | |
22 | extern "C" { | |
23 | #endif | |
24 | ||
7294103b MD |
25 | /* |
26 | * Lock-free stack. | |
27 | * | |
28 | * Stack implementing push, pop, pop_all operations, as well as iterator | |
29 | * on the stack head returned by pop_all. | |
30 | * | |
31 | * Synchronization table: | |
32 | * | |
33 | * External synchronization techniques described in the API below is | |
34 | * required between pairs marked with "X". No external synchronization | |
35 | * required between pairs marked with "-". | |
36 | * | |
37 | * cds_lfs_push __cds_lfs_pop __cds_lfs_pop_all | |
38 | * cds_lfs_push - - - | |
39 | * __cds_lfs_pop - X X | |
40 | * __cds_lfs_pop_all - X - | |
41 | * | |
42 | * cds_lfs_pop_blocking and cds_lfs_pop_all_blocking use an internal | |
43 | * mutex to provide synchronization. | |
44 | */ | |
45 | ||
46 | /* | |
47 | * cds_lfs_node_init: initialize lock-free stack node. | |
48 | */ | |
d3bfcb24 | 49 | static inline |
70469b43 | 50 | void _cds_lfs_node_init(struct cds_lfs_node *node __attribute__((unused))) |
d3bfcb24 MD |
51 | { |
52 | } | |
53 | ||
7294103b | 54 | /* |
200d100e MD |
55 | * cds_lfs_init: initialize lock-free stack (with lock). Pair with |
56 | * cds_lfs_destroy(). | |
7294103b | 57 | */ |
d3bfcb24 MD |
58 | static inline |
59 | void _cds_lfs_init(struct cds_lfs_stack *s) | |
60 | { | |
7294103b MD |
61 | int ret; |
62 | ||
d3bfcb24 | 63 | s->head = NULL; |
7294103b | 64 | ret = pthread_mutex_init(&s->lock, NULL); |
01477510 | 65 | urcu_posix_assert(!ret); |
7294103b MD |
66 | } |
67 | ||
48a8832b | 68 | /* |
200d100e MD |
69 | * cds_lfs_destroy: destroy lock-free stack (with lock). Pair with |
70 | * cds_lfs_init(). | |
71 | */ | |
72 | static inline | |
73 | void _cds_lfs_destroy(struct cds_lfs_stack *s) | |
74 | { | |
75 | int ret = pthread_mutex_destroy(&s->lock); | |
01477510 | 76 | urcu_posix_assert(!ret); |
200d100e MD |
77 | } |
78 | ||
79 | /* | |
80 | * ___cds_lfs_init: initialize lock-free stack (without lock). | |
81 | * Don't pair with any destroy function. | |
48a8832b MD |
82 | */ |
83 | static inline | |
84 | void ___cds_lfs_init(struct __cds_lfs_stack *s) | |
85 | { | |
86 | s->head = NULL; | |
87 | } | |
88 | ||
7294103b MD |
89 | static inline |
90 | bool ___cds_lfs_empty_head(struct cds_lfs_head *head) | |
91 | { | |
92 | return head == NULL; | |
93 | } | |
94 | ||
95 | /* | |
96 | * cds_lfs_empty: return whether lock-free stack is empty. | |
97 | * | |
98 | * No memory barrier is issued. No mutual exclusion is required. | |
99 | */ | |
100 | static inline | |
48a8832b | 101 | bool _cds_lfs_empty(cds_lfs_stack_ptr_t s) |
7294103b | 102 | { |
48a8832b | 103 | return ___cds_lfs_empty_head(CMM_LOAD_SHARED(s._s->head)); |
d3bfcb24 MD |
104 | } |
105 | ||
106 | /* | |
107 | * cds_lfs_push: push a node into the stack. | |
108 | * | |
109 | * Does not require any synchronization with other push nor pop. | |
110 | * | |
111 | * Lock-free stack push is not subject to ABA problem, so no need to | |
112 | * take the RCU read-side lock. Even if "head" changes between two | |
113 | * uatomic_cmpxchg() invocations here (being popped, and then pushed | |
114 | * again by one or more concurrent threads), the second | |
115 | * uatomic_cmpxchg() invocation only cares about pushing a new entry at | |
116 | * the head of the stack, ensuring consistency by making sure the new | |
117 | * node->next is the same pointer value as the value replaced as head. | |
118 | * It does not care about the content of the actual next node, so it can | |
119 | * very well be reallocated between the two uatomic_cmpxchg(). | |
120 | * | |
121 | * We take the approach of expecting the stack to be usually empty, so | |
122 | * we first try an initial uatomic_cmpxchg() on a NULL old_head, and | |
123 | * retry if the old head was non-NULL (the value read by the first | |
124 | * uatomic_cmpxchg() is used as old head for the following loop). The | |
125 | * upside of this scheme is to minimize the amount of cacheline traffic, | |
126 | * always performing an exclusive cacheline access, rather than doing | |
127 | * non-exclusive followed by exclusive cacheline access (which would be | |
128 | * required if we first read the old head value). This design decision | |
ffa11a18 | 129 | * might be revisited after more thorough benchmarking on various |
d3bfcb24 MD |
130 | * platforms. |
131 | * | |
132 | * Returns 0 if the stack was empty prior to adding the node. | |
133 | * Returns non-zero otherwise. | |
134 | */ | |
135 | static inline | |
48a8832b | 136 | bool _cds_lfs_push(cds_lfs_stack_ptr_t u_s, |
d3bfcb24 MD |
137 | struct cds_lfs_node *node) |
138 | { | |
48a8832b | 139 | struct __cds_lfs_stack *s = u_s._s; |
7294103b MD |
140 | struct cds_lfs_head *head = NULL; |
141 | struct cds_lfs_head *new_head = | |
142 | caa_container_of(node, struct cds_lfs_head, node); | |
d3bfcb24 MD |
143 | |
144 | for (;;) { | |
7294103b | 145 | struct cds_lfs_head *old_head = head; |
d3bfcb24 MD |
146 | |
147 | /* | |
148 | * node->next is still private at this point, no need to | |
149 | * perform a _CMM_STORE_SHARED(). | |
150 | */ | |
7294103b | 151 | node->next = &head->node; |
d3bfcb24 MD |
152 | /* |
153 | * uatomic_cmpxchg() implicit memory barrier orders earlier | |
154 | * stores to node before publication. | |
155 | */ | |
7294103b | 156 | head = uatomic_cmpxchg(&s->head, old_head, new_head); |
d3bfcb24 MD |
157 | if (old_head == head) |
158 | break; | |
159 | } | |
e47048cc | 160 | return !___cds_lfs_empty_head(head); |
d3bfcb24 MD |
161 | } |
162 | ||
163 | /* | |
7294103b | 164 | * __cds_lfs_pop: pop a node from the stack. |
d3bfcb24 MD |
165 | * |
166 | * Returns NULL if stack is empty. | |
167 | * | |
7294103b | 168 | * __cds_lfs_pop needs to be synchronized using one of the following |
d3bfcb24 MD |
169 | * techniques: |
170 | * | |
f9b5c2b6 MD |
171 | * 1) Calling __cds_lfs_pop under rcu read lock critical section. |
172 | * Both __cds_lfs_pop and __cds_lfs_pop_all callers must wait for a | |
173 | * grace period to pass before freeing the returned node or pushing | |
174 | * the node back into the stack. It is valid to overwrite the content | |
175 | * of cds_lfs_node immediately after __cds_lfs_pop and | |
176 | * __cds_lfs_pop_all. No RCU read-side critical section is needed | |
177 | * around __cds_lfs_pop_all. | |
7294103b MD |
178 | * 2) Using mutual exclusion (e.g. mutexes) to protect __cds_lfs_pop |
179 | * and __cds_lfs_pop_all callers. | |
180 | * 3) Ensuring that only ONE thread can call __cds_lfs_pop() and | |
181 | * __cds_lfs_pop_all(). (multi-provider/single-consumer scheme). | |
d3bfcb24 MD |
182 | */ |
183 | static inline | |
48a8832b | 184 | struct cds_lfs_node *___cds_lfs_pop(cds_lfs_stack_ptr_t u_s) |
d3bfcb24 | 185 | { |
48a8832b MD |
186 | struct __cds_lfs_stack *s = u_s._s; |
187 | ||
d3bfcb24 | 188 | for (;;) { |
7294103b MD |
189 | struct cds_lfs_head *head, *next_head; |
190 | struct cds_lfs_node *next; | |
d3bfcb24 MD |
191 | |
192 | head = _CMM_LOAD_SHARED(s->head); | |
7294103b MD |
193 | if (___cds_lfs_empty_head(head)) |
194 | return NULL; /* Empty stack */ | |
195 | ||
196 | /* | |
197 | * Read head before head->next. Matches the implicit | |
198 | * memory barrier before uatomic_cmpxchg() in | |
199 | * cds_lfs_push. | |
200 | */ | |
201 | cmm_smp_read_barrier_depends(); | |
202 | next = _CMM_LOAD_SHARED(head->node.next); | |
203 | next_head = caa_container_of(next, | |
204 | struct cds_lfs_head, node); | |
205 | if (uatomic_cmpxchg(&s->head, head, next_head) == head) | |
206 | return &head->node; | |
207 | /* busy-loop if head changed under us */ | |
d3bfcb24 MD |
208 | } |
209 | } | |
210 | ||
7294103b MD |
211 | /* |
212 | * __cds_lfs_pop_all: pop all nodes from a stack. | |
213 | * | |
214 | * __cds_lfs_pop_all does not require any synchronization with other | |
215 | * push, nor with other __cds_lfs_pop_all, but requires synchronization | |
216 | * matching the technique used to synchronize __cds_lfs_pop: | |
217 | * | |
218 | * 1) If __cds_lfs_pop is called under rcu read lock critical section, | |
f9b5c2b6 MD |
219 | * both __cds_lfs_pop and __cds_lfs_pop_all callers must wait for a |
220 | * grace period to pass before freeing the returned node or pushing | |
221 | * the node back into the stack. It is valid to overwrite the content | |
222 | * of cds_lfs_node immediately after __cds_lfs_pop and | |
223 | * __cds_lfs_pop_all. No RCU read-side critical section is needed | |
224 | * around __cds_lfs_pop_all. | |
7294103b MD |
225 | * 2) Using mutual exclusion (e.g. mutexes) to protect __cds_lfs_pop and |
226 | * __cds_lfs_pop_all callers. | |
227 | * 3) Ensuring that only ONE thread can call __cds_lfs_pop() and | |
228 | * __cds_lfs_pop_all(). (multi-provider/single-consumer scheme). | |
229 | */ | |
230 | static inline | |
48a8832b | 231 | struct cds_lfs_head *___cds_lfs_pop_all(cds_lfs_stack_ptr_t u_s) |
7294103b | 232 | { |
48a8832b MD |
233 | struct __cds_lfs_stack *s = u_s._s; |
234 | ||
7294103b MD |
235 | /* |
236 | * Implicit memory barrier after uatomic_xchg() matches implicit | |
237 | * memory barrier before uatomic_cmpxchg() in cds_lfs_push. It | |
238 | * ensures that all nodes of the returned list are consistent. | |
239 | * There is no need to issue memory barriers when iterating on | |
240 | * the returned list, because the full memory barrier issued | |
241 | * prior to each uatomic_cmpxchg, which each write to head, are | |
242 | * taking care to order writes to each node prior to the full | |
243 | * memory barrier after this uatomic_xchg(). | |
244 | */ | |
245 | return uatomic_xchg(&s->head, NULL); | |
246 | } | |
247 | ||
248 | /* | |
249 | * cds_lfs_pop_lock: lock stack pop-protection mutex. | |
250 | */ | |
251 | static inline void _cds_lfs_pop_lock(struct cds_lfs_stack *s) | |
252 | { | |
253 | int ret; | |
254 | ||
255 | ret = pthread_mutex_lock(&s->lock); | |
01477510 | 256 | urcu_posix_assert(!ret); |
7294103b MD |
257 | } |
258 | ||
259 | /* | |
260 | * cds_lfs_pop_unlock: unlock stack pop-protection mutex. | |
261 | */ | |
262 | static inline void _cds_lfs_pop_unlock(struct cds_lfs_stack *s) | |
263 | { | |
264 | int ret; | |
265 | ||
266 | ret = pthread_mutex_unlock(&s->lock); | |
01477510 | 267 | urcu_posix_assert(!ret); |
7294103b MD |
268 | } |
269 | ||
270 | /* | |
271 | * Call __cds_lfs_pop with an internal pop mutex held. | |
272 | */ | |
273 | static inline | |
274 | struct cds_lfs_node * | |
275 | _cds_lfs_pop_blocking(struct cds_lfs_stack *s) | |
276 | { | |
277 | struct cds_lfs_node *retnode; | |
28757437 | 278 | cds_lfs_stack_ptr_t stack; |
7294103b MD |
279 | |
280 | _cds_lfs_pop_lock(s); | |
28757437 SM |
281 | stack.s = s; |
282 | retnode = ___cds_lfs_pop(stack); | |
7294103b MD |
283 | _cds_lfs_pop_unlock(s); |
284 | return retnode; | |
285 | } | |
286 | ||
287 | /* | |
288 | * Call __cds_lfs_pop_all with an internal pop mutex held. | |
289 | */ | |
290 | static inline | |
291 | struct cds_lfs_head * | |
292 | _cds_lfs_pop_all_blocking(struct cds_lfs_stack *s) | |
293 | { | |
294 | struct cds_lfs_head *rethead; | |
28757437 | 295 | cds_lfs_stack_ptr_t stack; |
7294103b MD |
296 | |
297 | _cds_lfs_pop_lock(s); | |
28757437 SM |
298 | stack.s = s; |
299 | rethead = ___cds_lfs_pop_all(stack); | |
7294103b MD |
300 | _cds_lfs_pop_unlock(s); |
301 | return rethead; | |
302 | } | |
303 | ||
d3bfcb24 MD |
304 | #ifdef __cplusplus |
305 | } | |
306 | #endif | |
307 | ||
308 | #endif /* _URCU_STATIC_LFSTACK_H */ |