On Thu, Mar 16, 2017 at 06:20:46PM +0100, Peter Zijlstra wrote:
> On Thu, Mar 16, 2017 at 06:26:55PM +0300, Kirill A. Shutemov wrote:
> > +config HAVE_GENERIC_RCU_GUP
> > + def_bool y
> > +
>
> Nothing immediately jumped out to me; except that this option might be
> misnamed.
>
> AFAICT that code does not in fact rely on HAVE_RCU_TABLE_FREE; it will
> happily work with the (x86) broadcast IPI invalidate model, as you show
> here.
>
> Architectures that do not do that obviously need HAVE_RCU_TABLE_FREE,
> but that is not the point I feel.
>
> Also, this code hard relies on IRQ-disable delaying grace periods, which
> is mostly true I think, but has always been something Paul didn't really
> want to commit too firmly to.
That is quite true!
The only case where IRQ-disable is guaranteed to delay grace periods is
when you are using RCU-sched, in other words synchronize_sched() and
call_rcu_sched(). And even then, the CPU cannot be in the idle loop,
cannot be offline, and cannot be a nohz_full CPU on its way to/from
userspace execution.
Thanx, Paul
