Re: [PATCH 4/5] KVM: Fix compound mmio.

[Gleb Natapov <gleb@redhat.com>]

On Tue, Apr 12, 2011 at 03:27:40PM +0300, Avi Kivity wrote:
> On 04/12/2011 03:22 PM, Gleb Natapov wrote:
> >On Tue, Apr 12, 2011 at 03:19:00PM +0300, Avi Kivity wrote:
> >>  On 04/12/2011 12:36 PM, Gleb Natapov wrote:
> >>  >mmio_index should be taken into account when copying data from
> >>  >userspace.
> >>  >
> >>  >Signed-off-by: Gleb Natapov<gleb@redhat.com>
> >>  >---
> >>  >   arch/x86/kvm/x86.c |    3 ++-
> >>  >   1 files changed, 2 insertions(+), 1 deletions(-)
> >>  >
> >>  >diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> >>  >index b568779..609c7ab 100644
> >>  >--- a/arch/x86/kvm/x86.c
> >>  >+++ b/arch/x86/kvm/x86.c
> >>  >@@ -5518,7 +5518,8 @@ static int complete_mmio(struct kvm_vcpu *vcpu)
> >>  >   	if (vcpu->mmio_needed) {
> >>  >   		vcpu->mmio_needed = 0;
> >>  >   		if (!vcpu->mmio_is_write)
> >>  >-			memcpy(vcpu->mmio_data, run->mmio.data, 8);
> >>  >+			memcpy(vcpu->mmio_data + vcpu->mmio_index,
> >>  >+			       run->mmio.data, 8);
> >>  >   		vcpu->mmio_index += 8;
> >>  >   		if (vcpu->mmio_index<   vcpu->mmio_size) {
> >>  >   			run->exit_reason = KVM_EXIT_MMIO;
> >>
> >>  Interesting, the code passed the emulator.flat sse test.  Does it now?
> >>
> >It pass now and before. Probably by chance.
> 
> I don't understand how.  I explicitly set the values so that it
> would fail in that case.
> 
> Can you patch the test to fail with the current code?
> 
If I understand correctly you've already found the explanation why
test case worked?

> >But if I change read_emulated() to do
> >
> >  int n = min(size, (unsigned)KVM_MMIO_SIZE);
> >
> >instead of
> >
> >  int n = min(size, 8u);
> >
> >emulator.flat fails to emulate far jump instruction.
> 
> Ouch, looks like we have the multi-transaction support in two
> places.  I guess this is what made sse mmio work.
> 
> Not sure what we should do (patch is fine, question is how to
> resolve the duplication).
> 
Multi-transaction works faster in complete_mmio, so read_emulated()
should use it by doing int n = min(size, (unsigned)KVM_MMIO_SIZE).
Other code in read_emulated() is still needed since it provides read
re-play for multiple mmio reads during instruction emulation.

--
			Gleb.