Re: [PATCH 3/3] KVM: Update gfn_to_pfn_cache khva when it moves within the same page

[Sean Christopherson <seanjc@google.com>]

On Wed, Nov 23, 2022, David Woodhouse wrote:
> On Wed, 2022-11-23 at 00:20 +0000, David Woodhouse wrote:
> > From: David Woodhouse <dwmw@amazon.co.uk>
> > 
> > In the case where a GPC is refreshed to a different location within the
> > same page, we didn't bother to update it. Mostly we don't need to, but
> > since the ->khva field also includes the offset within the page, that
> > does have to be updated.
> > 
> > Fixes: 982ed0de4753 ("KVM: Reinstate gfn_to_pfn_cache with invalidation support")
> 
> Hm, wait. That commit wasn't actually broken because at that point the
> page offset was included in the uhva too, so the uhva *did* change and
> we'd (gratuitously) take the slower path through hva_to_pfn_retry()
> when the GPA moved within the same page.
> 
> So I think this should actually be:
> 
> Fixes: 3ba2c95ea180 ("KVM: Do not incorporate page offset into gfn=>pfn cache user address")

Ya.

> Which means it's only relevant back to v6.0 stable, not all the way
> back to v5.17.

Probably a moot point in the long run since that commit was tagged for stable@
too, in order to simplify the fixes that followed.

> > Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
> > Reviewed-by: Paul Durrant <paul@xen.org>
> > Cc: stable@kernel.org
> > 
> > ---

Reviewed-by: Sean Christopherson <seanjc@google.com>

> >  virt/kvm/pfncache.c | 7 ++++++-
> >  1 file changed, 6 insertions(+), 1 deletion(-)
> > 
> > diff --git a/virt/kvm/pfncache.c b/virt/kvm/pfncache.c
> > index bd4a46aee384..5f83321bfd2a 100644
> > --- a/virt/kvm/pfncache.c
> > +++ b/virt/kvm/pfncache.c
> > @@ -297,7 +297,12 @@ int kvm_gfn_to_pfn_cache_refresh(struct kvm *kvm, struct gfn_to_pfn_cache *gpc,
> >  	if (!gpc->valid || old_uhva != gpc->uhva) {
> >  		ret = hva_to_pfn_retry(kvm, gpc);
> >  	} else {
> > -		/* If the HVA→PFN mapping was already valid, don't unmap it. */
> > +		/*
> > +		 * If the HVA→PFN mapping was already valid, don't unmap it.
> > +		 * But do update gpc->khva because the offset within the page
> > +		 * may have changed.
> > +		 */
> > +		gpc->khva = old_khva + page_offset;

If/when we rework the APIs, another possible approach would be to store only the
the page aligned address, e.g. force the user to pass in offset+len by doing
something like:


	r = kvm_gpc_lock(...);
	if (r)
		return r;

	my_struct = kvm_gpc_kmap(..., offset, len);