Re: [Qemu-ppc] [qemu-ppc] quesstion about kvm e500 tlb search method

[Alexander Graf]

Hi Kevin,

On 24.01.2013, at 14:15, Wangkai (Kevin,C) wrote:

> Dear,
>  
> I check e500 core reference when the e500 core lookup for L2mmu entries,
> TLB1 is preferred to TLB0.

Could you please point me to the respective part of the documentation?

> And for e500 KVM l2mmu lookup, I find that TLB0 is searched first, and it
> Was done by software.
>  
> And I suggest we can search TLB1 first, because TLB0 have more entries
> Than TLB1, and this can improve the guest kernel performance very much.

Actually TLB1 has more entries, because we only search one set at a time. For 
TLB1, one set means "all entries" which on e500mc would mean 64. For TLB0, we 
only need to check 4 entries. So looking at TLB0 is a lot faster.

Also, most misses in "normal" real world workloads should happen in user space.

> Can I do this change to the kvm code? Is there some other affect?

Let me try to understand what you're trying to fix. Are you seeing performance 
issues? What kernel are you running on? Are you using hugetlbfs? There are a 
lot of things that improve performance a lot more than this change would.

Also, let me CC address@hidden, since you're really asking a KVM question here 
:).


Alex

>  
> Thanks!
> Wangkai
>  
>  
> int kvmppc_e500_tlb_search(struct kvm_vcpu *vcpu,
>                             gva_t eaddr, unsigned int pid, int as)
> {
>        struct kvmppc_vcpu_e500 *vcpu_e500 = to_e500(vcpu);
>        int esel, tlbsel;
>  
>        for (tlbsel = 0; tlbsel < 2; tlbsel++) {  // first tlb0, and then tlb1
>               esel = kvmppc_e500_tlb_index(vcpu_e500, eaddr, tlbsel, pid, as);
>               if (esel >= 0)
>                      return index_of(tlbsel, esel);
>        }
>  
>        return -1;
> }
>  
>