| 1 | // SPDX-FileCopyrightText: 2010 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 2 | // |
| 3 | // SPDX-License-Identifier: LGPL-2.1-or-later |
| 4 | |
| 5 | #ifndef _URCU_RCULFQUEUE_STATIC_H |
| 6 | #define _URCU_RCULFQUEUE_STATIC_H |
| 7 | |
| 8 | /* |
| 9 | * Userspace RCU library - Lock-Free RCU Queue |
| 10 | * |
| 11 | * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See rculfqueue.h for linking |
| 12 | * dynamically with the userspace rcu library. |
| 13 | */ |
| 14 | |
| 15 | #include <urcu-call-rcu.h> |
| 16 | #include <urcu/assert.h> |
| 17 | #include <urcu/uatomic.h> |
| 18 | #include <urcu-pointer.h> |
| 19 | #include <errno.h> |
| 20 | |
| 21 | #ifdef __cplusplus |
| 22 | extern "C" { |
| 23 | #endif |
| 24 | |
| 25 | struct cds_lfq_node_rcu_dummy { |
| 26 | struct cds_lfq_node_rcu parent; |
| 27 | struct rcu_head head; |
| 28 | struct cds_lfq_queue_rcu *q; |
| 29 | }; |
| 30 | |
| 31 | /* |
| 32 | * Lock-free RCU queue. Enqueue and dequeue operations hold a RCU read |
| 33 | * lock to deal with cmpxchg ABA problem. This queue is *not* circular: |
| 34 | * head points to the oldest node, tail points to the newest node. |
| 35 | * A dummy node is kept to ensure enqueue and dequeue can always proceed |
| 36 | * concurrently. Keeping a separate head and tail helps with large |
| 37 | * queues: enqueue and dequeue can proceed concurrently without |
| 38 | * wrestling for exclusive access to the same variables. |
| 39 | * |
| 40 | * Dequeue retry if it detects that it would be dequeueing the last node |
| 41 | * (it means a dummy node dequeue-requeue is in progress). This ensures |
| 42 | * that there is always at least one node in the queue. |
| 43 | * |
| 44 | * In the dequeue operation, we internally reallocate the dummy node |
| 45 | * upon dequeue/requeue and use call_rcu to free the old one after a |
| 46 | * grace period. |
| 47 | */ |
| 48 | |
| 49 | static inline |
| 50 | struct cds_lfq_node_rcu *make_dummy(struct cds_lfq_queue_rcu *q, |
| 51 | struct cds_lfq_node_rcu *next) |
| 52 | { |
| 53 | struct cds_lfq_node_rcu_dummy *dummy; |
| 54 | |
| 55 | dummy = (struct cds_lfq_node_rcu_dummy *) |
| 56 | malloc(sizeof(struct cds_lfq_node_rcu_dummy)); |
| 57 | urcu_posix_assert(dummy); |
| 58 | dummy->parent.next = next; |
| 59 | dummy->parent.dummy = 1; |
| 60 | dummy->q = q; |
| 61 | return &dummy->parent; |
| 62 | } |
| 63 | |
| 64 | static inline |
| 65 | void free_dummy_cb(struct rcu_head *head) |
| 66 | { |
| 67 | struct cds_lfq_node_rcu_dummy *dummy = |
| 68 | caa_container_of(head, struct cds_lfq_node_rcu_dummy, head); |
| 69 | free(dummy); |
| 70 | } |
| 71 | |
| 72 | static inline |
| 73 | void rcu_free_dummy(struct cds_lfq_node_rcu *node) |
| 74 | { |
| 75 | struct cds_lfq_node_rcu_dummy *dummy; |
| 76 | |
| 77 | urcu_posix_assert(node->dummy); |
| 78 | dummy = caa_container_of(node, struct cds_lfq_node_rcu_dummy, parent); |
| 79 | dummy->q->queue_call_rcu(&dummy->head, free_dummy_cb); |
| 80 | } |
| 81 | |
| 82 | static inline |
| 83 | void free_dummy(struct cds_lfq_node_rcu *node) |
| 84 | { |
| 85 | struct cds_lfq_node_rcu_dummy *dummy; |
| 86 | |
| 87 | urcu_posix_assert(node->dummy); |
| 88 | dummy = caa_container_of(node, struct cds_lfq_node_rcu_dummy, parent); |
| 89 | free(dummy); |
| 90 | } |
| 91 | |
| 92 | static inline |
| 93 | void _cds_lfq_node_init_rcu(struct cds_lfq_node_rcu *node) |
| 94 | { |
| 95 | node->next = NULL; |
| 96 | node->dummy = 0; |
| 97 | } |
| 98 | |
| 99 | static inline |
| 100 | void _cds_lfq_init_rcu(struct cds_lfq_queue_rcu *q, |
| 101 | void queue_call_rcu(struct rcu_head *head, |
| 102 | void (*func)(struct rcu_head *head))) |
| 103 | { |
| 104 | q->tail = make_dummy(q, NULL); |
| 105 | q->head = q->tail; |
| 106 | q->queue_call_rcu = queue_call_rcu; |
| 107 | } |
| 108 | |
| 109 | /* |
| 110 | * The queue should be emptied before calling destroy. |
| 111 | * |
| 112 | * Return 0 on success, -EPERM if queue is not empty. |
| 113 | */ |
| 114 | static inline |
| 115 | int _cds_lfq_destroy_rcu(struct cds_lfq_queue_rcu *q) |
| 116 | { |
| 117 | struct cds_lfq_node_rcu *head; |
| 118 | |
| 119 | head = rcu_dereference(q->head); |
| 120 | if (!(head->dummy && head->next == NULL)) |
| 121 | return -EPERM; /* not empty */ |
| 122 | free_dummy(head); |
| 123 | return 0; |
| 124 | } |
| 125 | |
| 126 | /* |
| 127 | * Should be called under rcu read lock critical section. |
| 128 | */ |
| 129 | static inline |
| 130 | void _cds_lfq_enqueue_rcu(struct cds_lfq_queue_rcu *q, |
| 131 | struct cds_lfq_node_rcu *node) |
| 132 | { |
| 133 | /* |
| 134 | * uatomic_cmpxchg() implicit memory barrier orders earlier stores to |
| 135 | * node before publication. |
| 136 | */ |
| 137 | for (;;) { |
| 138 | struct cds_lfq_node_rcu *tail, *next; |
| 139 | |
| 140 | tail = rcu_dereference(q->tail); |
| 141 | cmm_emit_legacy_smp_mb(); |
| 142 | next = uatomic_cmpxchg_mo(&tail->next, NULL, node, |
| 143 | CMM_SEQ_CST, CMM_SEQ_CST); |
| 144 | if (next == NULL) { |
| 145 | /* |
| 146 | * Tail was at the end of queue, we successfully |
| 147 | * appended to it. Now move tail (another |
| 148 | * enqueue might beat us to it, that's fine). |
| 149 | */ |
| 150 | (void) uatomic_cmpxchg_mo(&q->tail, tail, node, |
| 151 | CMM_SEQ_CST, CMM_SEQ_CST); |
| 152 | return; |
| 153 | } else { |
| 154 | /* |
| 155 | * Failure to append to current tail. |
| 156 | * Help moving tail further and retry. |
| 157 | */ |
| 158 | (void) uatomic_cmpxchg_mo(&q->tail, tail, next, |
| 159 | CMM_SEQ_CST, CMM_SEQ_CST); |
| 160 | continue; |
| 161 | } |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | static inline |
| 166 | void enqueue_dummy(struct cds_lfq_queue_rcu *q) |
| 167 | { |
| 168 | struct cds_lfq_node_rcu *node; |
| 169 | |
| 170 | /* We need to reallocate to protect from ABA. */ |
| 171 | node = make_dummy(q, NULL); |
| 172 | _cds_lfq_enqueue_rcu(q, node); |
| 173 | } |
| 174 | |
| 175 | /* |
| 176 | * Should be called under rcu read lock critical section. |
| 177 | * |
| 178 | * The caller must wait for a grace period to pass before freeing the returned |
| 179 | * node or modifying the cds_lfq_node_rcu structure. |
| 180 | * Returns NULL if queue is empty. |
| 181 | */ |
| 182 | static inline |
| 183 | struct cds_lfq_node_rcu *_cds_lfq_dequeue_rcu(struct cds_lfq_queue_rcu *q) |
| 184 | { |
| 185 | for (;;) { |
| 186 | struct cds_lfq_node_rcu *head, *next; |
| 187 | |
| 188 | head = rcu_dereference(q->head); |
| 189 | next = rcu_dereference(head->next); |
| 190 | if (head->dummy && next == NULL) |
| 191 | return NULL; /* empty */ |
| 192 | /* |
| 193 | * We never, ever allow dequeue to get to a state where |
| 194 | * the queue is empty (we need at least one node in the |
| 195 | * queue). This is ensured by checking if the head next |
| 196 | * is NULL, which means we need to enqueue a dummy node |
| 197 | * before we can hope dequeuing anything. |
| 198 | */ |
| 199 | if (!next) { |
| 200 | enqueue_dummy(q); |
| 201 | next = rcu_dereference(head->next); |
| 202 | } |
| 203 | if (uatomic_cmpxchg_mo(&q->head, head, next, |
| 204 | CMM_SEQ_CST, CMM_SEQ_CST) != head) |
| 205 | continue; /* Concurrently pushed. */ |
| 206 | if (head->dummy) { |
| 207 | /* Free dummy after grace period. */ |
| 208 | rcu_free_dummy(head); |
| 209 | continue; /* try again */ |
| 210 | } |
| 211 | return head; |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | #ifdef __cplusplus |
| 216 | } |
| 217 | #endif |
| 218 | |
| 219 | #endif /* _URCU_RCULFQUEUE_STATIC_H */ |