Re: [PATCH 4/7] KVM: MMU: don't mark pte notrap if it's just sync transient

[Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>]

Marcelo Tosatti wrote:
> On Fri, Jun 11, 2010 at 09:31:38PM +0800, Xiao Guangrong wrote:
>> If the sync-sp just sync transient, don't mark its pte notrap
>>
>> Signed-off-by: Xiao Guangrong <xiaoguangrong@cn.fujitsu.com>
>> ---
>>  arch/x86/include/asm/kvm_host.h |    2 +-
>>  arch/x86/kvm/mmu.c              |   11 ++++-------
>>  arch/x86/kvm/paging_tmpl.h      |    5 +++--
>>  3 files changed, 8 insertions(+), 10 deletions(-)
> 
> Xiao,
> 
> Can you explain the reasoning for this? 
> 

Marcelo,

In the kvm_sync_page_transient() path, the sp is keep unsync and sp->gfn is not
write protected, we can't set 'spte == shadow_notrap_nonpresent_pte' if the 
'gpte.p == 0' in this case.

It's because if 'gpte.p == 0', when guest change the gpte's mapping, it's no need
to flush tlb(p == 0 means the mapping is not in tlb), if we set spte to
'shadow_notrap_nonpresent_pte', we will miss the chance to update the mapping, also
a incorrect #PF will pass to guest directly, may cause guest crash.

And, this is the reasoning we do mmu_convert_notrap() in kvm_sync_page_transient()
before, this patch just avoid this unnecessary workload. :-)