Re: [PATCH 2/3] KVM: arm64: Don't map PUD huge page if it's not available

[Gavin Shan]

Hi Marc,

On 10/25/20 9:05 PM, Marc Zyngier wrote:
> On Sun, 25 Oct 2020 01:27:38 +0100,
> Gavin Shan <gshan@xxxxxxxxxx> wrote:
> > PUD huge page isn't available when CONFIG_ARM64_4K_PAGES is disabled.
> > In this case, we needn't try to map the memory through PUD huge pages
> > to save some CPU cycles in the hot path.
> >
> > This also corrects the code style issue, which was introduced by
> > commit <523b3999e5f6> ("KVM: arm64: Try PMD block mappings if PUD mappings
> > are not supported").
> >
> > Signed-off-by: Gavin Shan <gshan@xxxxxxxxxx>
> > ---
> >   arch/arm64/kvm/mmu.c | 4 +++-
> >   1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
> > index a816cb8e619b..0f51585adc04 100644
> > --- a/arch/arm64/kvm/mmu.c
> > +++ b/arch/arm64/kvm/mmu.c
> > @@ -787,9 +787,11 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
> >   		vma_shift = PAGE_SHIFT;
> >   	}
> >   
> > +#ifdef CONFIG_ARM64_4K_PAGES
> >   	if (vma_shift == PUD_SHIFT &&
> >   	    !fault_supports_stage2_huge_mapping(memslot, hva, PUD_SIZE))
> > -	       vma_shift = PMD_SHIFT;
> > +		vma_shift = PMD_SHIFT;
> > +#endif
> >   
> >   	if (vma_shift == PMD_SHIFT &&
> >   	    !fault_supports_stage2_huge_mapping(memslot, hva, PMD_SIZE)) {
>
>
> I really don't buy the "CPU cycles" argument here either. Can you
> actually measure any difference here?
>
> You have taken a fault, gone through a full guest exit, triaged it,
> and are about to into a page mapping operation which may result in a
> TLBI, and reenter the guest. It only happen a handful of times per
> page over the lifetime of the guest unless you start swapping. Hot
> path? I don't think so.

Thanks for the explanation. Agreed and I will drop this in v2.

Thanks,
Gavin