Guest memory access within hypervisor Inbox x

Guest memory access within hypervisor Inbox x

[Piyus Kedia]

Hi,

I am working on record replay in KVM. I just want to confirm that KVM
doesn't access guest memory except for DMA. The functions
kvm_read_guest_page() and kvm_write_guest_page() in kvm_main.c are
only used for mmio emulation which access memory that is always
emulated. It will be nice if somebody else also confirm this.

Thanks,
Piyus Kedia

Re: Guest memory access within hypervisor Inbox x

[Stefan Hajnoczi]

On Fri, Sep 27, 2013 at 03:19:28AM +0530, Piyus Kedia wrote:
> I am working on record replay in KVM. I just want to confirm that KVM
> doesn't access guest memory except for DMA. The functions
> kvm_read_guest_page() and kvm_write_guest_page() in kvm_main.c are
> only used for mmio emulation which access memory that is always
> emulated. It will be nice if somebody else also confirm this.

kvmclock comes to mind, I think it has a shared page between the guest
and the host kernel where clock updates are published from host ->
guest.  I think kvmclock can be disabled by choosing another clocksource
in the guest Linux kernel.

There are other exceptions like the ivshmem device which is rarely used
but can provide shared memory between guest <-> host or guest <-> guest.
I think that memory is an MMIO PCI BAR in the guest just that neither
kvm.ko nor QEMU may be catch the update if another process on the host
is writing to it.

Stefan

Re: Guest memory access within hypervisor Inbox x

[Piyus Kedia]

Thanks.

On Fri, Sep 27, 2013 at 3:40 PM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> On Fri, Sep 27, 2013 at 03:19:28AM +0530, Piyus Kedia wrote:
>> I am working on record replay in KVM. I just want to confirm that KVM
>> doesn't access guest memory except for DMA. The functions
>> kvm_read_guest_page() and kvm_write_guest_page() in kvm_main.c are
>> only used for mmio emulation which access memory that is always
>> emulated. It will be nice if somebody else also confirm this.
>
> kvmclock comes to mind, I think it has a shared page between the guest
> and the host kernel where clock updates are published from host ->
> guest.  I think kvmclock can be disabled by choosing another clocksource
> in the guest Linux kernel.
>
> There are other exceptions like the ivshmem device which is rarely used
> but can provide shared memory between guest <-> host or guest <-> guest.
> I think that memory is an MMIO PCI BAR in the guest just that neither
> kvm.ko nor QEMU may be catch the update if another process on the host
> is writing to it.
>
> Stefan