sys_membarrier()

In essence, yes, the read-side memory barriers are only executed when the updater requires that synchronization. With RCU, we only have to execute the barriers each time a synchronize_rcu() is called, which occurs typically far less often than variable updates.

Rephrasing what you said: "the idea is to make CPUs execute memory barriers only when synchronization is required by the updater thread, as opposed to executing them each time before and after accessing it from a reader thread".

Thanks,

Mathieu