Re: [PATCH RFC 03/39] KVM: x86/xen: register shared_info page

[Ankur Arora]

On 2020-12-02 2:44 a.m., Joao Martins wrote:
> [late response - was on holiday yesterday]
>
> On 12/2/20 12:40 AM, Ankur Arora wrote:
> > On 2020-12-01 5:07 a.m., David Woodhouse wrote:
> > > On Wed, 2019-02-20 at 20:15 +0000, Joao Martins wrote:
> > > > +static int kvm_xen_shared_info_init(struct kvm *kvm, gfn_t gfn)
> > > > +{
> > > > +       struct shared_info *shared_info;
> > > > +       struct page *page;
> > > > +
> > > > +       page = gfn_to_page(kvm, gfn);
> > > > +       if (is_error_page(page))
> > > > +               return -EINVAL;
> > > > +
> > > > +       kvm->arch.xen.shinfo_addr = gfn;
> > > > +
> > > > +       shared_info = page_to_virt(page);
> > > > +       memset(shared_info, 0, sizeof(struct shared_info));
> > > > +       kvm->arch.xen.shinfo = shared_info;
> > > > +       return 0;
> > > > +}
> > > > +
> > >
> > > Hm.
> > >
> > > How come we get to pin the page and directly dereference it every time,
> > > while kvm_setup_pvclock_page() has to use kvm_write_guest_cached()
> > > instead?
> >
> > So looking at my WIP trees from the time, this is something that
> > we went back and forth on as well with using just a pinned page or a
> > persistent kvm_vcpu_map().
> >
> > I remember distinguishing shared_info/vcpu_info from kvm_setup_pvclock_page()
> > as shared_info is created early and is not expected to change during the
> > lifetime of the guest which didn't seem true for MSR_KVM_SYSTEM_TIME (or
> > MSR_KVM_STEAL_TIME) so that would either need to do a kvm_vcpu_map()
> > kvm_vcpu_unmap() dance or do some kind of synchronization.
> >
> > That said, I don't think this code explicitly disallows any updates
> > to shared_info.
> >
> > > If that was allowed, wouldn't it have been a much simpler fix for
> > > CVE-2019-3016? What am I missing?
> >
> > Agreed.
> >
> > Perhaps, Paolo can chime in with why KVM never uses pinned page
> > and always prefers to do cached mappings instead?
> Part of the CVE fix to not use cached versions.
>
> It's not a longterm pin of the page unlike we try to do here (partly due to the nature
> of the pages we are mapping) but we still we map the gpa, RMW the steal time struct, and
> then unmap the page.
>
> See record_steal_time() -- but more specifically commit b043138246 ("x86/KVM: Make sure
> KVM_VCPU_FLUSH_TLB flag is not missed").
>
> But I am not sure it's a good idea to follow the same as record_steal_time() given that
> this is a fairly sensitive code path for event channels.
>
> > > Should I rework these to use kvm_write_guest_cached()?
> >
> > kvm_vcpu_map() would be better. The event channel logic does RMW operations
> > on shared_info->vcpu_info.
> Indeed, yes.
>
> Ankur IIRC, we saw missed event channels notifications when we were using the
> {write,read}_cached() version of the patch.
>
> But I can't remember the reason it was due to, either the evtchn_pending or the mask
> word -- which would make it not inject an upcall.

If memory serves, it was the mask. Though I don't think that we had
kvm_{write,read}_cached in use at that point -- given that they were
definitely not RMW safe.


Ankur

> 	Joao