Commit | Line | Data |
---|---|---|
ee566d46 MD |
1 | /* |
2 | * rcuwfqueue.h | |
3 | * | |
4 | * Userspace RCU library - RCU Queue with Wait-Free Enqueue/Blocking Dequeue | |
5 | * | |
6 | * Copyright 2010 - 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 | #include <urcu/urcu_ref.h> | |
24 | #include <assert.h> | |
25 | ||
26 | #if (!defined(_GNU_SOURCE) && !defined(_LGPL_SOURCE)) | |
27 | #error "Dynamic loader LGPL wrappers not implemented yet" | |
28 | #endif | |
29 | ||
30 | /* | |
31 | * RCU queue with wait-free enqueue/blocking dequeue using reference counting. | |
32 | * Enqueue and dequeue operations hold a RCU read lock to deal with cmpxchg ABA | |
33 | * problem. This implementation keeps a dummy head node to ensure we can always | |
34 | * update the queue locklessly. Given that this is a queue, the dummy head node | |
35 | * must always advance as we dequeue entries. Therefore, we keep a reference | |
36 | * count on each entry we are dequeueing, so they can be kept as dummy head node | |
37 | * until the next dequeue, at which point their reference count will be | |
38 | * decremented. | |
39 | */ | |
40 | ||
41 | #define URCU_WFQ_PERMANENT_REF 128 | |
42 | ||
43 | struct rcu_wfq_node { | |
44 | struct rcu_wfq_node *next; | |
45 | struct urcu_ref ref; | |
46 | }; | |
47 | ||
48 | struct rcu_wfq_queue { | |
49 | struct rcu_wfq_node *head, *tail; | |
50 | struct rcu_wfq_node init; /* Dummy initialization node */ | |
51 | }; | |
52 | ||
53 | void rcu_wfq_node_init(struct rcu_wfq_node *node) | |
54 | { | |
55 | node->next = NULL; | |
56 | urcu_ref_init(&node->ref); | |
57 | } | |
58 | ||
59 | void rcu_wfq_init(struct rcu_wfq_queue *q) | |
60 | { | |
61 | rcu_wfq_node_init(&q->init); | |
62 | /* Make sure the initial node is never freed. */ | |
63 | urcu_ref_set(&q->init.ref, URCU_WFQ_PERMANENT_REF); | |
64 | /* Set queue end */ | |
65 | q->head = q->tail = &q->init; | |
66 | } | |
67 | ||
68 | void rcu_wfq_enqueue(struct rcu_wfq_queue *q, struct rcu_wfq_node *node) | |
69 | { | |
70 | struct rcu_wfq_node *old_tail; | |
71 | ||
72 | urcu_ref_get(&node->ref); | |
73 | /* | |
74 | * uatomic_xchg() implicit memory barrier orders earlier stores to node | |
75 | * (setting it to NULL and incrementing the refcount) before | |
76 | * publication. | |
77 | */ | |
78 | old_tail = uatomic_xchg(&q->tail, node); | |
79 | /* | |
80 | * At this point, dequeuers see a NULL old_tail->next, which indicates | |
81 | * end of queue. The following store will append "node" to the queue | |
82 | * from a dequeuer perspective. | |
83 | */ | |
84 | STORE_SHARED(old_tail->next, node); | |
85 | } | |
86 | ||
87 | /* | |
88 | * The entry returned by dequeue must be taken care of by doing a urcu_ref_put, | |
89 | * which calls the release primitive when the reference count drops to zero. A | |
90 | * grace period must be waited before performing the actual memory reclamation | |
91 | * in the release primitive. The wfq node returned by dequeue must not be | |
92 | * modified/re-used/freed until the reference count reaches zero and a grace | |
93 | * period has elapsed (after the refcount reached 0). | |
94 | * | |
95 | * TODO: implement adaptative busy-wait and wait/wakeup scheme rather than busy | |
96 | * loops. Better for UP. | |
97 | */ | |
98 | struct rcu_wfq_node * | |
99 | rcu_wfq_dequeue_blocking(struct rcu_wfq_queue *q, | |
100 | void (*release)(struct urcu_ref *)) | |
101 | { | |
102 | for (;;) { | |
103 | struct rcu_wfq_node *head, *next; | |
104 | ||
105 | rcu_read_lock(); | |
106 | head = rcu_dereference(q->head); | |
107 | next = rcu_dereference(head->next); | |
108 | if (next) { | |
109 | if (uatomic_cmpxchg(&q->head, head, next) == head) { | |
110 | rcu_read_unlock(); | |
111 | urcu_ref_put(&head->ref, release); | |
112 | return next; | |
113 | } else { | |
114 | /* Concurrently pushed, retry */ | |
115 | rcu_read_unlock(); | |
116 | continue; | |
117 | } | |
118 | } else { | |
119 | /* Empty */ | |
120 | rcu_read_unlock(); | |
121 | return NULL; | |
122 | } | |
123 | } | |
124 | } |