Re: [Xen-devel] [PATCH v7 10/10] xen/common: do not implicitly permit access to mapped I/O memory

[Julien Grall <julien.grall@xxxxxxxxxx>]

On 26/05/14 14:22, Jan Beulich wrote:
> > > > On 26.05.14 at 14:15, <julien.grall@xxxxxxxxxx> wrote:
> > On 26/05/14 12:51, Jan Beulich wrote:
> > > > > > On 26.05.14 at 13:42, <julien.grall@xxxxxxxxxx> wrote:
> > > > On 26/05/14 12:37, Jan Beulich wrote:
> > > > > > > > On 26.05.14 at 13:24, <julien.grall@xxxxxxxxxx> wrote:
> > > > > > On 26/05/14 12:14, Jan Beulich wrote:
> > > > > > > > > > On 26.05.14 at 12:53, <julien.grall@xxxxxxxxxx> wrote:
> > > > > > > > On 26/05/14 11:14, Jan Beulich wrote:
> > > > > > > > > Or maybe I wasn't wrong - the patch context doesn't really make
> > > > > > > > > clear whether it's the granting or mapping operation that gets
> > > > > > > > > adjusted here (since an earlier patch moved the mapping one into
> > > > > > > > > this function).
> > > > > > > >               ret = -EPERM;
> > > > > > > > -        if ( !iomem_access_permitted(current->domain, mfn, mfn_end) )
> > > > > > > > +        if ( !iomem_access_permitted(d, mfn, mfn_end) )
> > > > > > > >                   break;
> > > > > > > >
> > > > > > > >               ret = xsm_iomem_mapping(XSM_HOOK, d, mfn, mfn_end, add);
> > > > > > > >
> > > > > > > > There is an xsm_iomem_mapping just after, so the change has been done in
> > > > > > > > XEN_DOMCTL_memory_mapping.
> > > > > > > In which case I indeed stick to my original comment - it's perhaps
> > > > > > > best to check _both_.
> > > > > > Why? We may want to map the region in the guest P2M without giving the
> > > > > > permission to the guest (I'm thinking about ARM passthrough case).
> > > > > How can you put a mapping of memory into a guest's P2M for which
> > > > > that guest has no access permission? To me this reads like you're
> > > > > intending to create a security issue here.
> > > > iomem_access_permitted is used to check if we allow the current guest to
> > > > map a region in another guest P2M.
> > > >
> > > > Once the mapping is done, at least on ARM, we don't use anymore the
> > > > permission check. This is because there is no trap involved afterwards.
> > > I don't see how absence or presence of traps is involved here. The
> > > problem I see is that by putting in such a P2M entrry you allow a
> > > guest access to memory that it wasn't granted access to.
> > In the case of an HVM guest (or ARM guest), the permission seems to be
> > used only during DOMCTL_memory_mapping hypercall. So I understand the
> > permission as "I'm allowed to map/unmap this MMIO range from a guest P2M".
> Ah, I start seeing where you're coming from: Taking current uses of
> a certain construct to imply a meaning of that construct is kind of
> backwards. You should start with the meaning, and all of
> {iomem,ioports,pirq}_access_permitted() are inquiries to find out
> whether the specified domain is permitted to access the specified
> resource.
>
> And the fact that for HVM/ARM there's currently only that one use
> in the domctl handling isn't indicative of anything. Just take the use
> in xen/common/grant_table.c - if the paging_mode_external()
> check got dropped (assuming the backing functions can handle this)
> you'd all of the sudden have another use. And the check here, from
> all I know, isn't there because such an operation makes sense only
> for PV guests, but because the same operation just can't be handled
> for HVM/PVH/ARM ones. The moment someone finds a legitimate use
> of e.g. granting pages from a passed-through frame buffer (to e.g.
> snapshot its contents directly to storage) this would need to be
> fixed.
I got your point now. Many thanks for your explanation.

I'm still not convince, it's perfectly valid to:
        1) call iomem_permission to set the permission
        2) call memory_mapping
        3) call iomem_permission to remove the permission

In this case the memory mapping will be there without any granting. 
Hence, we won't be able to remove those mappings.
IHMO, this make the requirement of the permission useless and add more 
code in the toolstack to map an MMIO region for peanuts.
> > If we request the guest to have the permission on this range, we also
> > allow the guest to map in its P2M (assuming XSM is not there) theses MMIOs.
> >
> > I don't think, at least on ARM, we want to let the guest doing what it
> > wants with the mapping MMIO region.
> And I didn't say that.
I didn't intend to blame you with this sentence.

Regards,

--
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel