Re: [PATCH 2/2] KVM: Fix writeback on page boundary that propagate changes in spite of #PF

[Gleb Natapov <gleb@redhat.com>]

On Thu, Jan 12, 2012 at 12:21:00PM +0200, Avi Kivity wrote:
> On 01/12/2012 12:12 PM, Gleb Natapov wrote:
> > On Wed, Jan 11, 2012 at 06:53:31PM +0200, Nadav Amit wrote:
> > > Consider the case in which an instruction emulation writeback is performed on a page boundary.
> > > In such case, if a #PF occurs on the second page, the write to the first page already occurred and cannot be retracted.
> > > Therefore, validation of the second page access must be performed prior to writeback.
> > > 
> > > Signed-off-by: Nadav Amit <nadav.amit@gmail.com>
> > > ---
> > >  arch/x86/kvm/x86.c |   13 +++++++++++++
> > >  1 files changed, 13 insertions(+), 0 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > > index 05fd3d7..7af3d67 100644
> > > --- a/arch/x86/kvm/x86.c
> > > +++ b/arch/x86/kvm/x86.c
> > > @@ -3626,6 +3626,8 @@ struct read_write_emulator_ops {
> > >  			       int bytes, void *val);
> > >  	int (*read_write_exit_mmio)(struct kvm_vcpu *vcpu, gpa_t gpa,
> > >  				    void *val, int bytes);
> > > +	gpa_t (*read_write_validate)(struct kvm_vcpu *vcpu, gva_t gva,
> > > +				     struct x86_exception *exception);
> > >  	bool write;
> > >  };
> > >  
> > > @@ -3686,6 +3688,7 @@ static struct read_write_emulator_ops write_emultor = {
> > >  	.read_write_emulate = write_emulate,
> > >  	.read_write_mmio = write_mmio,
> > >  	.read_write_exit_mmio = write_exit_mmio,
> > > +	.read_write_validate = kvm_mmu_gva_to_gpa_write,
> > >  	.write = true,
> > >  };
> > >  
> > > @@ -3750,6 +3753,16 @@ int emulator_read_write(struct x86_emulate_ctxt *ctxt, unsigned long addr,
> > >  		int rc, now;
> > >  
> > >  		now = -addr & ~PAGE_MASK;
> > > +
> > > +		/* First check there is no page-fault on the next page */
> > > +		if (ops->read_write_validate &&
> > > +			ops->read_write_validate(vcpu, addr+now, exception) ==
> > > +			UNMAPPED_GVA) {
> > > +			/* #PF on the first page should be reported first */
> > > +			ops->read_write_validate(vcpu, addr, exception);
> > > +			return X86EMUL_PROPAGATE_FAULT;
> > > +		}
> > > +
> > This undoes optimization that vcpu_mmio_gva_to_gpa() has for handling
> > mmio. 
> 
> Right.  I suggest changing I/O to have two phases: first, translate the
> virtual address into an array of two physical addresses; check
> exceptions and report.  Then do the actual writes.
> 
> > Furthermore for common (non faulting) case we will check page
> > tables twice on each write that crosses page boundary, first time here
> > and second time in emulator_read_write_onepage().
> 
> Those should be very uncommon.
> 
Still it is better to have all the checks in one place like you suggest
above.

--
			Gleb.