Re: [Xen-devel] [RFC PATCH 04/16] x86/xen: hypercall support for xenhost_t

[Andrew Cooper]

On 14/06/2019 08:35, Juergen Gross wrote:
> On 14.06.19 09:20, Ankur Arora wrote:
>> On 2019-06-12 2:15 p.m., Andrew Cooper wrote:
>>> On 09/05/2019 18:25, Ankur Arora wrote:
>>>> Allow for different hypercall implementations for different xenhost
>>>> types.
>>>> Nested xenhost, which has two underlying xenhosts, can use both
>>>> simultaneously.
>>>>
>>>> The hypercall macros (HYPERVISOR_*) implicitly use the default
>>>> xenhost.x
>>>> A new macro (hypervisor_*) takes xenhost_t * as a parameter and
>>>> does the
>>>> right thing.
>>>>
>>>> TODO:
>>>> ÂÂ - Multicalls for now assume the default xenhost
>>>> ÂÂ - xen_hypercall_* symbols are only generated for the default
>>>> xenhost.
>>>>
>>>> Signed-off-by: Ankur Arora <ankur.a.arora@xxxxxxxxxx>
>>>
>>> Again, what is the hypervisor nesting and/or guest layout here?
>> Two hypervisors, L0 and L1, and the guest is a child of the L1
>> hypervisor but could have PV devices attached to both L0 and L1
>> hypervisors.
>>
>>>
>>> I can't think of any case where a single piece of software can
>>> legitimately have two hypercall pages, because if it has one working
>>> one, it is by definition a guest, and therefore not privileged
>>> enough to
>>> use the outer one.
>> Depending on which hypercall page is used, the hypercall would
>> (eventually) land in the corresponding hypervisor.
>>
>> Juergen elsewhere pointed out proxying hypercalls is a better approach,
>> so I'm not really considering this any more but, given this layout, and
>> assuming that the hypercall pages could be encoded differently would it
>> still not work?
>
> Hypercalls might work, but it is a bad idea and a violation of layering
> to let a L1 guest issue hypercalls to L0 hypervisor, as those hypercalls
> could influence other L1 guests and even the L1 hypervisor.
>
> Hmm, thinking more about it, I even doubt those hypercalls could work in
> all cases: when issued from a L1 PV guest the hypercalls would seem to
> be issued from user mode for the L0 hypervisor, and this is not allowed.

That is exactly the point I was trying to make.

If L2 is an HVM guest, then both its hypercall pages will be using
VMCALL/VMMCALL which will end up making hypercalls to L1, rather than
having one go to L0.

If L2 is a PV guest, then one hypercall page will be SYSCALL/INT 82
which will go to L1, and one will be VMCALL/VMMCALL which goes to L0,
but L0 will see it from ring1/ring3 and reject the hypercall.

However you nest the system, every guest only has a single occurrence of
"supervisor software", so only has a single context that will be
tolerated to make hypercalls by the next hypervisor up.

~Andrew