Commit | Line | Data |
---|---|---|
d3bfcb24 MD |
1 | #ifndef _URCU_STATIC_LFSTACK_H |
2 | #define _URCU_STATIC_LFSTACK_H | |
3 | ||
4 | /* | |
5 | * urcu/static/lfstack.h | |
6 | * | |
7 | * Userspace RCU library - Lock-Free Stack | |
8 | * | |
9 | * Copyright 2010-2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
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 | * This library is free software; you can redistribute it and/or | |
15 | * modify it under the terms of the GNU Lesser General Public | |
16 | * License as published by the Free Software Foundation; either | |
17 | * version 2.1 of the License, or (at your option) any later version. | |
18 | * | |
19 | * This library is distributed in the hope that it will be useful, | |
20 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
21 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
22 | * Lesser General Public License for more details. | |
23 | * | |
24 | * You should have received a copy of the GNU Lesser General Public | |
25 | * License along with this library; if not, write to the Free Software | |
26 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | |
27 | */ | |
28 | ||
7294103b MD |
29 | #include <stdbool.h> |
30 | #include <pthread.h> | |
31 | #include <assert.h> | |
d3bfcb24 MD |
32 | #include <urcu/uatomic.h> |
33 | #include <urcu-pointer.h> | |
34 | ||
35 | #ifdef __cplusplus | |
36 | extern "C" { | |
37 | #endif | |
38 | ||
7294103b MD |
39 | /* |
40 | * Lock-free stack. | |
41 | * | |
42 | * Stack implementing push, pop, pop_all operations, as well as iterator | |
43 | * on the stack head returned by pop_all. | |
44 | * | |
45 | * Synchronization table: | |
46 | * | |
47 | * External synchronization techniques described in the API below is | |
48 | * required between pairs marked with "X". No external synchronization | |
49 | * required between pairs marked with "-". | |
50 | * | |
51 | * cds_lfs_push __cds_lfs_pop __cds_lfs_pop_all | |
52 | * cds_lfs_push - - - | |
53 | * __cds_lfs_pop - X X | |
54 | * __cds_lfs_pop_all - X - | |
55 | * | |
56 | * cds_lfs_pop_blocking and cds_lfs_pop_all_blocking use an internal | |
57 | * mutex to provide synchronization. | |
58 | */ | |
59 | ||
60 | /* | |
61 | * cds_lfs_node_init: initialize lock-free stack node. | |
62 | */ | |
d3bfcb24 MD |
63 | static inline |
64 | void _cds_lfs_node_init(struct cds_lfs_node *node) | |
65 | { | |
66 | } | |
67 | ||
7294103b | 68 | /* |
93aee570 MD |
69 | * cds_lfs_init: initialize lock-free stack (with lock). Pair with |
70 | * cds_lfs_destroy(). | |
7294103b | 71 | */ |
d3bfcb24 MD |
72 | static inline |
73 | void _cds_lfs_init(struct cds_lfs_stack *s) | |
74 | { | |
7294103b MD |
75 | int ret; |
76 | ||
d3bfcb24 | 77 | s->head = NULL; |
7294103b MD |
78 | ret = pthread_mutex_init(&s->lock, NULL); |
79 | assert(!ret); | |
80 | } | |
81 | ||
93aee570 MD |
82 | /* |
83 | * cds_lfs_destroy: destroy lock-free stack (with lock). Pair with | |
84 | * cds_lfs_init(). | |
85 | */ | |
86 | static inline | |
87 | void _cds_lfs_destroy(struct cds_lfs_stack *s) | |
88 | { | |
89 | int ret = pthread_mutex_destroy(&s->lock); | |
90 | assert(!ret); | |
91 | } | |
92 | ||
7294103b MD |
93 | static inline |
94 | bool ___cds_lfs_empty_head(struct cds_lfs_head *head) | |
95 | { | |
96 | return head == NULL; | |
97 | } | |
98 | ||
99 | /* | |
100 | * cds_lfs_empty: return whether lock-free stack is empty. | |
101 | * | |
102 | * No memory barrier is issued. No mutual exclusion is required. | |
103 | */ | |
104 | static inline | |
105 | bool _cds_lfs_empty(struct cds_lfs_stack *s) | |
106 | { | |
107 | return ___cds_lfs_empty_head(CMM_LOAD_SHARED(s->head)); | |
d3bfcb24 MD |
108 | } |
109 | ||
110 | /* | |
111 | * cds_lfs_push: push a node into the stack. | |
112 | * | |
113 | * Does not require any synchronization with other push nor pop. | |
114 | * | |
115 | * Lock-free stack push is not subject to ABA problem, so no need to | |
116 | * take the RCU read-side lock. Even if "head" changes between two | |
117 | * uatomic_cmpxchg() invocations here (being popped, and then pushed | |
118 | * again by one or more concurrent threads), the second | |
119 | * uatomic_cmpxchg() invocation only cares about pushing a new entry at | |
120 | * the head of the stack, ensuring consistency by making sure the new | |
121 | * node->next is the same pointer value as the value replaced as head. | |
122 | * It does not care about the content of the actual next node, so it can | |
123 | * very well be reallocated between the two uatomic_cmpxchg(). | |
124 | * | |
125 | * We take the approach of expecting the stack to be usually empty, so | |
126 | * we first try an initial uatomic_cmpxchg() on a NULL old_head, and | |
127 | * retry if the old head was non-NULL (the value read by the first | |
128 | * uatomic_cmpxchg() is used as old head for the following loop). The | |
129 | * upside of this scheme is to minimize the amount of cacheline traffic, | |
130 | * always performing an exclusive cacheline access, rather than doing | |
131 | * non-exclusive followed by exclusive cacheline access (which would be | |
132 | * required if we first read the old head value). This design decision | |
ffa11a18 | 133 | * might be revisited after more thorough benchmarking on various |
d3bfcb24 MD |
134 | * platforms. |
135 | * | |
136 | * Returns 0 if the stack was empty prior to adding the node. | |
137 | * Returns non-zero otherwise. | |
138 | */ | |
139 | static inline | |
7294103b | 140 | bool _cds_lfs_push(struct cds_lfs_stack *s, |
d3bfcb24 MD |
141 | struct cds_lfs_node *node) |
142 | { | |
7294103b MD |
143 | struct cds_lfs_head *head = NULL; |
144 | struct cds_lfs_head *new_head = | |
145 | caa_container_of(node, struct cds_lfs_head, node); | |
d3bfcb24 MD |
146 | |
147 | for (;;) { | |
7294103b | 148 | struct cds_lfs_head *old_head = head; |
d3bfcb24 MD |
149 | |
150 | /* | |
151 | * node->next is still private at this point, no need to | |
152 | * perform a _CMM_STORE_SHARED(). | |
153 | */ | |
7294103b | 154 | node->next = &head->node; |
d3bfcb24 MD |
155 | /* |
156 | * uatomic_cmpxchg() implicit memory barrier orders earlier | |
157 | * stores to node before publication. | |
158 | */ | |
7294103b | 159 | head = uatomic_cmpxchg(&s->head, old_head, new_head); |
d3bfcb24 MD |
160 | if (old_head == head) |
161 | break; | |
162 | } | |
0fba1252 | 163 | return !___cds_lfs_empty_head(head); |
d3bfcb24 MD |
164 | } |
165 | ||
166 | /* | |
7294103b | 167 | * __cds_lfs_pop: pop a node from the stack. |
d3bfcb24 MD |
168 | * |
169 | * Returns NULL if stack is empty. | |
170 | * | |
7294103b | 171 | * __cds_lfs_pop needs to be synchronized using one of the following |
d3bfcb24 MD |
172 | * techniques: |
173 | * | |
7294103b | 174 | * 1) Calling __cds_lfs_pop under rcu read lock critical section. The |
d3bfcb24 MD |
175 | * caller must wait for a grace period to pass before freeing the |
176 | * returned node or modifying the cds_lfs_node structure. | |
7294103b MD |
177 | * 2) Using mutual exclusion (e.g. mutexes) to protect __cds_lfs_pop |
178 | * and __cds_lfs_pop_all callers. | |
179 | * 3) Ensuring that only ONE thread can call __cds_lfs_pop() and | |
180 | * __cds_lfs_pop_all(). (multi-provider/single-consumer scheme). | |
d3bfcb24 MD |
181 | */ |
182 | static inline | |
7294103b | 183 | struct cds_lfs_node *___cds_lfs_pop(struct cds_lfs_stack *s) |
d3bfcb24 MD |
184 | { |
185 | for (;;) { | |
7294103b MD |
186 | struct cds_lfs_head *head, *next_head; |
187 | struct cds_lfs_node *next; | |
d3bfcb24 MD |
188 | |
189 | head = _CMM_LOAD_SHARED(s->head); | |
7294103b MD |
190 | if (___cds_lfs_empty_head(head)) |
191 | return NULL; /* Empty stack */ | |
192 | ||
193 | /* | |
194 | * Read head before head->next. Matches the implicit | |
195 | * memory barrier before uatomic_cmpxchg() in | |
196 | * cds_lfs_push. | |
197 | */ | |
198 | cmm_smp_read_barrier_depends(); | |
199 | next = _CMM_LOAD_SHARED(head->node.next); | |
200 | next_head = caa_container_of(next, | |
201 | struct cds_lfs_head, node); | |
202 | if (uatomic_cmpxchg(&s->head, head, next_head) == head) | |
203 | return &head->node; | |
204 | /* busy-loop if head changed under us */ | |
d3bfcb24 MD |
205 | } |
206 | } | |
207 | ||
7294103b MD |
208 | /* |
209 | * __cds_lfs_pop_all: pop all nodes from a stack. | |
210 | * | |
211 | * __cds_lfs_pop_all does not require any synchronization with other | |
212 | * push, nor with other __cds_lfs_pop_all, but requires synchronization | |
213 | * matching the technique used to synchronize __cds_lfs_pop: | |
214 | * | |
215 | * 1) If __cds_lfs_pop is called under rcu read lock critical section, | |
216 | * both __cds_lfs_pop and cds_lfs_pop_all callers must wait for a | |
217 | * grace period to pass before freeing the returned node or modifying | |
218 | * the cds_lfs_node structure. However, no RCU read-side critical | |
219 | * section is needed around __cds_lfs_pop_all. | |
220 | * 2) Using mutual exclusion (e.g. mutexes) to protect __cds_lfs_pop and | |
221 | * __cds_lfs_pop_all callers. | |
222 | * 3) Ensuring that only ONE thread can call __cds_lfs_pop() and | |
223 | * __cds_lfs_pop_all(). (multi-provider/single-consumer scheme). | |
224 | */ | |
225 | static inline | |
226 | struct cds_lfs_head *___cds_lfs_pop_all(struct cds_lfs_stack *s) | |
227 | { | |
228 | /* | |
229 | * Implicit memory barrier after uatomic_xchg() matches implicit | |
230 | * memory barrier before uatomic_cmpxchg() in cds_lfs_push. It | |
231 | * ensures that all nodes of the returned list are consistent. | |
232 | * There is no need to issue memory barriers when iterating on | |
233 | * the returned list, because the full memory barrier issued | |
234 | * prior to each uatomic_cmpxchg, which each write to head, are | |
235 | * taking care to order writes to each node prior to the full | |
236 | * memory barrier after this uatomic_xchg(). | |
237 | */ | |
238 | return uatomic_xchg(&s->head, NULL); | |
239 | } | |
240 | ||
241 | /* | |
242 | * cds_lfs_pop_lock: lock stack pop-protection mutex. | |
243 | */ | |
244 | static inline void _cds_lfs_pop_lock(struct cds_lfs_stack *s) | |
245 | { | |
246 | int ret; | |
247 | ||
248 | ret = pthread_mutex_lock(&s->lock); | |
249 | assert(!ret); | |
250 | } | |
251 | ||
252 | /* | |
253 | * cds_lfs_pop_unlock: unlock stack pop-protection mutex. | |
254 | */ | |
255 | static inline void _cds_lfs_pop_unlock(struct cds_lfs_stack *s) | |
256 | { | |
257 | int ret; | |
258 | ||
259 | ret = pthread_mutex_unlock(&s->lock); | |
260 | assert(!ret); | |
261 | } | |
262 | ||
263 | /* | |
264 | * Call __cds_lfs_pop with an internal pop mutex held. | |
265 | */ | |
266 | static inline | |
267 | struct cds_lfs_node * | |
268 | _cds_lfs_pop_blocking(struct cds_lfs_stack *s) | |
269 | { | |
270 | struct cds_lfs_node *retnode; | |
271 | ||
272 | _cds_lfs_pop_lock(s); | |
273 | retnode = ___cds_lfs_pop(s); | |
274 | _cds_lfs_pop_unlock(s); | |
275 | return retnode; | |
276 | } | |
277 | ||
278 | /* | |
279 | * Call __cds_lfs_pop_all with an internal pop mutex held. | |
280 | */ | |
281 | static inline | |
282 | struct cds_lfs_head * | |
283 | _cds_lfs_pop_all_blocking(struct cds_lfs_stack *s) | |
284 | { | |
285 | struct cds_lfs_head *rethead; | |
286 | ||
287 | _cds_lfs_pop_lock(s); | |
288 | rethead = ___cds_lfs_pop_all(s); | |
289 | _cds_lfs_pop_unlock(s); | |
290 | return rethead; | |
291 | } | |
292 | ||
d3bfcb24 MD |
293 | #ifdef __cplusplus |
294 | } | |
295 | #endif | |
296 | ||
297 | #endif /* _URCU_STATIC_LFSTACK_H */ |