On Thu, May 30, 2019 at 02:05:31PM +0200, Oleg Nesterov wrote: > > Anyway, looking at it, I think smp_acquire__after_ctrl_dep() doesn't > > make sense here; > > Well I still _think_ it should work, it provides the LOAD-LOAD ordering > and this is what we need.
So it hard relies on being part of a control dependency, note how the comment says that architectures that do not do load speculation can override the smp_rmb() default with barrier() (and we used have an architecture that made use of that, although it's been since removed). IOW, it is an error to use smp_acquire__after_ctrl_dep() without an (immediate) preceding branch.
