Skip to content

2.0.1

Latest

Choose a tag to compare

@mjansson mjansson released this 15 Jul 14:53
beef233

Fixed a data race in concurrent lazy initialization where simultaneous first allocations from
multiple threads could race on the global init state and return a block backed by a partially
written configuration. Initialization is now serialized through an atomic state machine.

Spin waits now escalate to an OS scheduler yield after a spin threshold, avoiding throughput
collapse when a preempted lock holder is starved by spinners under high thread contention.