[PATCH 2/8] ARM: cache: remove redundant dsb instruction from v7_coherent_user_range

[Will Deacon]

On Fri, May 09, 2014 at 05:16:09PM +0100, Catalin Marinas wrote:
> On Fri, May 02, 2014 at 04:24:09PM +0100, Will Deacon wrote:
> > v7_coherent_user_range takes a virtual address range, cleans the D-side
> > to PoU and then invalidates the I-side so that subsequent instruction
> > fetches can see any new data written to the range in question.
> > 
> > Since cache maintenance by MVA is architected to execute in program
> > order with respect to other cache maintenance operations specifying
> > the same virtual address, we do not require a barrier between the
> > D-side clean and the I-side invalidation.
> > 
> > This patch removes the redundant dsb.
> > 
> > Signed-off-by: Will Deacon <will.deacon at arm.com>
> > ---
> >  arch/arm/mm/cache-v7.S | 1 -
> >  1 file changed, 1 deletion(-)
> > 
> > diff --git a/arch/arm/mm/cache-v7.S b/arch/arm/mm/cache-v7.S
> > index 615c99e38ba1..b040d3ca20ac 100644
> > --- a/arch/arm/mm/cache-v7.S
> > +++ b/arch/arm/mm/cache-v7.S
> > @@ -282,7 +282,6 @@ ENTRY(v7_coherent_user_range)
> >  	add	r12, r12, r2
> >  	cmp	r12, r1
> >  	blo	1b
> > -	dsb	ishst
> >  	icache_line_size r2, r3
> >  	sub	r3, r2, #1
> >  	bic	r12, r0, r3
> 
> The original implementation follows the ARMv7 ARM example for self
> modifying code which has a DSB. I agree with you that the section B2.2.9
> (ARMv7 ARM - Ordering of cache and branch predictor maintenance
> operations) states that ops by MVA would be ordered with each-other.

Those examples also don't make use of barrier options, so I wouldn't pay too
much attention to them :)

Still, I'll chase this up internally since the ARM ARM needs fixing.

Will