Re: KVM: MMU: drop read-only large sptes when creating lower level sptes

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Feb 25, 2014 at 11:11:19PM +0800, Xiao Guangrong wrote:
> 
> On Feb 25, 2014, at 9:13 PM, Marcelo Tosatti <mtosatti@redhat.com> wrote:
> 
> > On Tue, Feb 25, 2014 at 11:30:37AM +0800, Xiao Guangrong wrote:
> >> On 02/25/2014 12:59 AM, Marcelo Tosatti wrote:
> >>> 
> >>> Read-only large sptes can be created due to read-only faults as
> >>> follows:
> >>> 
> >>> - QEMU pagetable entry that maps guest memory is read-only 
> >>> due to COW.
> >>> - Guest read faults such memory, COW is not broken, because
> >>> it is a read-only fault.
> >>> - Enable dirty logging, large spte not nuked because it is read-only.
> >>> - Write-fault on such memory causes guest to loop endlessly
> >>> (which must go down to level 1 because dirty logging is enabled).
> >> 
> >> Hi Marcelo,
> >> 
> >> It surprised me that the large-readonly mapping was not dropped
> >> by mmu-notifer as this is write fault on readonly mapping in Qemu.
> >> Hmm... i missed something?
> > 
> > You mean COW was not broken by gup? (that is the problem, so a
> > read-only large spte is created).
> > 
> 
> I mean the final step that “Write-fault on such memory causes guest” should
> break the readonly QEMU pagetable entry and change it to writable, at that
> time, mmu-notifier should be called since the page’s permission has been
> changed.
> 
> After read the code more carefully, i figure the reason out that mmu-notifier
> is not called when the page can be reused. 

Well a transition from read-only host-pte to read-write host-pte does
not require the read-only spte to be upgraded immediatelly, right?
Permission is upgraded lazily.

> I am thinking whether it is
> reasonable:
> - anyway, the permission bits on the pre are changed, why not call
>  mmu-notifier->change_pte()?

It could, but lazy updates are better i suppose.

> - kvm and other users of mum-notifier need to check the permission
>  changing very carefully, that is really subtle, like the case fixed here.

Yes.

> - if we do proper prefetch when it happen, the addition page-fault can
>  be avoided, that is good for the performance.

Agree.