Re: [PATCH 03/11] ARM: pm: some ARMv7 requires a dsb in resume to ensure correctness

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Wed, Sep 07, 2011 at 04:41:32PM +0100, Catalin Marinas wrote:
> On 1 September 2011 13:49, Russell King - ARM Linux
> <linux@arm.linux.org.uk> wrote:
> > Add a dsb after the isb to ensure that the previous writes to the
> > CP15 registers take effect before we enable the MMU.
> >
> > Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
> > ---
> >  arch/arm/mm/proc-v7.S |    1 +
> >  1 files changed, 1 insertions(+), 0 deletions(-)
> >
> > diff --git a/arch/arm/mm/proc-v7.S b/arch/arm/mm/proc-v7.S
> > index dec72ee..a773f4e 100644
> > --- a/arch/arm/mm/proc-v7.S
> > +++ b/arch/arm/mm/proc-v7.S
> > @@ -255,6 +255,7 @@ ENTRY(cpu_v7_do_resume)
> >        mcr     p15, 0, r4, c10, c2, 0  @ write PRRR
> >        mcr     p15, 0, r5, c10, c2, 1  @ write NMRR
> >        isb
> > +       dsb
> 
> Isn't an ISB enough here? We usually have the DSB for some background
> operations like cache maintenance.

That depends whether you're including the effects of the cache
maintanence instructions in this.  The ARM ARM tells me that
a DSB is required to ensure that all cache maintanence is issued
before the dsb is complete at the point that the dsb is executed.

So for architectural compliance, yes, a dsb is required.  The isb
is also required to ensure that instruction cache maintanence is
properly visible.