Re: [PATCH 3/3] KVM: PPC: e500: Implement TLB1-in-TLB0 mapping

[Scott Wood <scottwood@freescale.com>]

On 01/17/2013 04:50:41 PM, Alexander Graf wrote:
> When a host mapping fault happens in a guest TLB1 entry today, we
> map the translated guest entry into the host's TLB1.
> 
> This isn't particularly clever when the guest is mapped by normal 4k
> pages, since these would be a lot better to put into TLB0 instead.
> 
> This patch adds the required logic to map 4k TLB1 shadow maps into
> the host's TLB0.
> 
> Signed-off-by: Alexander Graf <agraf@suse.de>
> ---
>  arch/powerpc/kvm/e500.h          |    1 +
>  arch/powerpc/kvm/e500_mmu_host.c |   58  
> +++++++++++++++++++++++++++++--------
>  2 files changed, 46 insertions(+), 13 deletions(-)
> 
> diff --git a/arch/powerpc/kvm/e500.h b/arch/powerpc/kvm/e500.h
> index 00f96d8..d32e6a8 100644
> --- a/arch/powerpc/kvm/e500.h
> +++ b/arch/powerpc/kvm/e500.h
> @@ -28,6 +28,7 @@
> 
>  #define E500_TLB_VALID 1
>  #define E500_TLB_BITMAP 2
> +#define E500_TLB_TLB0		(1 << 2)
> 
>  struct tlbe_ref {
>  	pfn_t pfn;
> diff --git a/arch/powerpc/kvm/e500_mmu_host.c  
> b/arch/powerpc/kvm/e500_mmu_host.c
> index 3bb2154..cbb6cf8 100644
> --- a/arch/powerpc/kvm/e500_mmu_host.c
> +++ b/arch/powerpc/kvm/e500_mmu_host.c
> @@ -198,6 +198,11 @@ void inval_gtlbe_on_host(struct kvmppc_vcpu_e500  
> *vcpu_e500, int tlbsel,
>  		local_irq_restore(flags);
> 
>  		return;
> +	} else if (tlbsel == 1 &&
> +		   vcpu_e500->gtlb_priv[1][esel].ref.flags &  
> E500_TLB_TLB0) {
> +		/* This is a slow path, so just invalidate everything */
> +		kvmppc_e500_tlbil_all(vcpu_e500);
> +		vcpu_e500->gtlb_priv[1][esel].ref.flags &=  
> ~E500_TLB_TLB0;
>  	}

What if the guest TLB1 entry is backed by a mix of TLB0 and TLB1  
entries on the host?  I don't see checks elsewhere that would prevent  
this situation.

> @@ -529,9 +556,14 @@ void kvmppc_mmu_map(struct kvm_vcpu *vcpu, u64  
> eaddr, gpa_t gpaddr,
>  	case 1: {
>  		gfn_t gfn = gpaddr >> PAGE_SHIFT;
> 
> -		stlbsel = 1;
>  		sesel = kvmppc_e500_tlb1_map(vcpu_e500, eaddr, gfn,
>  					     gtlbe, &stlbe, esel);
> +		if (sesel < 0) {
> +			/* TLB0 mapping */
> +			sesel = 0;
> +			stlbsel = 0;
> +		} else
> +			stlbsel = 1;
>  		break;
>  	}

Maybe push the call to write_tlbe() into the tlb0/1_map functions,  
getting rid of the need to pass sesel/stlbsel/stlbe back?

-Scott