rculfstack: document "push"

[urcu.git] / urcu / static / rculfstack.h

diff --git a/urcu/static/rculfstack.h b/urcu/static/rculfstack.h
index 99d3d4ccc92f8cd22011d099417986e274e08d73..9d69fa736996c4dfc704bbffc88d9f14395344c8 100644 (file)
--- a/urcu/static/rculfstack.h
+++ b/urcu/static/rculfstack.h
@@ -27,21 +27,45 @@
  */
 
 #include <urcu/uatomic.h>
-/* A urcu implementation header should be already included. */
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+static inline
 void _cds_lfs_node_init_rcu(struct cds_lfs_node_rcu *node)
 {
 }
 
+static inline
 void _cds_lfs_init_rcu(struct cds_lfs_stack_rcu *s)
 {
        s->head = NULL;
 }
 
+/*
+ * Lock-free stack push is not subject to ABA problem, so no need to
+ * take the RCU read-side lock. Even if "head" changes between two
+ * uatomic_cmpxchg() invocations here (being popped, and then pushed
+ * again by one or more concurrent threads), the second
+ * uatomic_cmpxchg() invocation only cares about pushing a new entry at
+ * the head of the stack, ensuring consistency by making sure the new
+ * node->next is the same pointer value as the value replaced as head.
+ * It does not care about the content of the actual next node, so it can
+ * very well be reallocated between the two uatomic_cmpxchg().
+ *
+ * We take the approach of expecting the stack to be usually empty, so
+ * we first try an initial uatomic_cmpxchg() on a NULL old_head, and
+ * retry if the old head was non-NULL (the value read by the first
+ * uatomic_cmpxchg() is used as old head for the following loop). The
+ * upside of this scheme is to minimize the amount of cacheline traffic,
+ * always performing an exclusive cacheline access, rather than doing
+ * non-exclusive followed by exclusive cacheline access (which would be
+ * required if we first read the old head value). This design decision
+ * might be revisited after more throrough benchmarking on various
+ * platforms.
+ */
+static inline
 void _cds_lfs_push_rcu(struct cds_lfs_stack_rcu *s, struct cds_lfs_node_rcu *node)
 {
        struct cds_lfs_node_rcu *head = NULL;
@@ -61,30 +85,35 @@ void _cds_lfs_push_rcu(struct cds_lfs_stack_rcu *s, struct cds_lfs_node_rcu *nod
 }
 
 /*
- * Should be called under rcu read-side lock.
+ * Acts as a RCU reader.
  *
  * The caller must wait for a grace period to pass before freeing the returned
  * node or modifying the cds_lfs_node_rcu structure.
  * Returns NULL if stack is empty.
  */
+static inline
 struct cds_lfs_node_rcu *
 _cds_lfs_pop_rcu(struct cds_lfs_stack_rcu *s)
 {
        for (;;) {
                struct cds_lfs_node_rcu *head;
 
+               rcu_read_lock();
                head = rcu_dereference(s->head);
                if (head) {
                        struct cds_lfs_node_rcu *next = rcu_dereference(head->next);
 
                        if (uatomic_cmpxchg(&s->head, head, next) == head) {
+                               rcu_read_unlock();
                                return head;
                        } else {
                                /* Concurrent modification. Retry. */
+                               rcu_read_unlock();
                                continue;
                        }
                } else {
                        /* Empty stack */
+                       rcu_read_unlock();
                        return NULL;
                }
        }