[PATCH v2 3/5] ARM: KVM: relax cache maintainance when building page tables

[Will Deacon]

On Thu, May 02, 2013 at 04:15:05PM +0100, Christoffer Dall wrote:
> On Thu, May 2, 2013 at 8:03 AM, Will Deacon <will.deacon at arm.com> wrote:
> > On Thu, May 02, 2013 at 03:39:00PM +0100, Marc Zyngier wrote:
> >> Patch 5a677ce044f1 (ARM: KVM: switch to a dual-step HYP init code)
> >> introduced code that flushes page tables to the point of coherency.
> >> This is overkill (point of unification is enough and already done),
> >> and actually not required if running on a SMP capable platform
> >> (the HW PTW can snoop other cpus' L1).
> >>
> >> Remove this code and let ae8a8b9553bd (ARM: 7691/1: mm: kill unused
> >> TLB_CAN_READ_FROM_L1_CACHE and use ALT_SMP instead) turn it into
> >> a no-op for SMP ARMv7.
> >>
> >> Reported-by: Catalin Marinas <catalin.marinas at arm.com>
> >> Cc: Will Deacon <will.deacon at arm.com>
> >> Signed-off-by: Marc Zyngier <marc.zyngier at arm.com>
> >> ---
> >>  arch/arm/kvm/mmu.c | 3 ---
> >>  1 file changed, 3 deletions(-)
> >>
> >> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> >> index 4a838db..d79b594 100644
> >> --- a/arch/arm/kvm/mmu.c
> >> +++ b/arch/arm/kvm/mmu.c
> >> @@ -234,7 +234,6 @@ static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
> >>               pte = pte_offset_kernel(pmd, addr);
> >>               kvm_set_pte(pte, pfn_pte(pfn, prot));
> >>               get_page(virt_to_page(pte));
> >> -             kvm_flush_dcache_to_poc(pte, sizeof(*pte));
> >>               pfn++;
> >>       } while (addr += PAGE_SIZE, addr != end);
> >
> > Could you remove the flushing code out of kvm_set_pte, then flush the range
> > after the loop? Then you get one flush and one barrier for the whole range.
> >
> that would be different from how it's done in the rest of the kernel
> where you're expecting set_pte to do the necessary flushing for you,
> no?

Correct, but since kvm_set_pte seems only to be called from this file, I'm
not sure you need to match thes rest of the kernel.

> instead, if this optimization is worth it, the code could just use
> pte_val(pte) = XX;

or just *pte = XX;

Will