On 09/23/2010 04:59 AM, Xiao Guangrong wrote:
On 09/20/2010 11:24 PM, Avi Kivity wrote:
> On 09/20/2010 04:21 PM, Xiao Guangrong wrote:
>> Gfn may have many shadow pages, when one sp need be synced, we write
>> protected sp->gfn and sync this sp but we keep other shadow pages
>> asynchronous
>>
>> So, while gfn happen page fault, let it not touches unsync page, the
>> unsync
>> page only updated at invlpg/flush TLB time
>>
>> @@ -3157,6 +3164,9 @@ void kvm_mmu_pte_write(struct kvm_vcpu *vcpu,
>> gpa_t gpa,
>>
>> mask.cr0_wp = mask.cr4_pae = mask.nxe = 1;
>> for_each_gfn_indirect_valid_sp(vcpu->kvm, sp, gfn, node) {
>> + if (sp->unsync)
>> + continue;
>> +
>>
>
> Not sure this is a win. If a gpte is updated from having p=0 to p=1 (or
> permissions upgraded), we may not have an invlpg to sync the spte, since
> the hardware doesn't require it. With this change, we may get an extra
> #PF.
>
Avi,
Thanks for your review, i think this case is not too bad since:
1: This case only impacts local vcpu since if permissions is increased, it's
no need send IPT to flush remote vcpu's tlb, so even if we update unsync
spte in kvm_mmu_pte_write() path, the #PF still occur on other vcpus.
IIRC, the cpu will re-validate the tlb entry from the page tables before
issuing a fault, so we won't see a spurious fault. Not 100% sure.
For !P -> P, there won't be a tlb entry, so 100% there won't be a
spurious fault.
2: If the unsync sp which is updated in kvm_mmu_pte_write() is not using by the
vcpu, it will sync automatically after it's loaded.
True, and this is a likely case.
3: If the sp is using, update this sp in kvm_mmu_pte_write() will avoid extra #PF,
in this case, two(or more) sps have the same gfn, there are mapped in the same
page table and with different kinds(unsync/sync), i thinks this case is infrequency.
And even we updated it, we can not sure it can be accessed latter,
If it's infrequent, the why do we optimize it?
So, i think it's better lazily update unsync sp until it's used or the flush time,
your opinion? :-)
Any performance numbers?
To me it seems saving a possible exit is worth extra computation.
--
error compiling committee.c: too many arguments to function
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
