Re: [Xen-devel] Xenstore domains and XS_RESTRICT

[Juergen Gross <jgross@xxxxxxxx>]

On 18/01/17 19:26, Stefano Stabellini wrote:
> On Wed, 18 Jan 2017, Juergen Gross wrote:
>> On 18/01/17 12:03, Wei Liu wrote:
>>> On Mon, Jan 16, 2017 at 05:47:15PM +0100, Juergen Gross wrote:
>>>> On 07/12/16 08:44, Juergen Gross wrote:
>>>>> Hi,
>>>>>
>>>>> today the XS_RESTRICT wire command of Xenstore is supported by
>>>>> oxenstored only to drop the privilege of a connection to that of the
>>>>> domid given as a parameter to the command.
>>>>>
>>>>> Using this mechanism with Xenstore running in a stubdom will lead to
>>>>> problems as instead of only a dom0 process dropping its privileges
>>>>> the privileges of dom0 will be dropped (all dom0 Xenstore requests
>>>>> share the same connection).
>>>>>
>>>>> In order to solve the problem I suggest the following change to the
>>>>> Xenstore wire protocol:
>>>>>
>>>>>  struct xsd_sockmsg
>>>>>  {
>>>>> -    uint32_t type;  /* XS_??? */
>>>>> +    uint16_t type;  /* XS_??? */
>>>>> +    uint16_t domid; /* Use privileges of this domain */
>>>>>      uint32_t req_id;/* Request identifier, echoed in daemon's response.  
>>>>> */
>>>>>      uint32_t tx_id; /* Transaction id (0 if not related to a
>>>>> transaction). */
>>>>>      uint32_t len;   /* Length of data following this. */
>>>>>
>>>>>      /* Generally followed by nul-terminated string(s). */
>>>>>  };
>>>>>
>>>>> domid will normally be zero having the same effect as today.
>>>>>
>>>>> Using XS_RESTRICT via a socket connection will run as today by dropping
>>>>> the privileges of that connection.
>>>>>
>>>>> Using XS_RESTRICT via the kernel (Xenstore domain case) will save the
>>>>> domid given as parameter in the connection specific private kernel
>>>>> structure. All future Xenstore commands of the connection will have
>>>>> this domid set in xsd_sockmsg. The kernel will never forward the
>>>>> XS_RESTRICT command to Xenstore.
>>>>>
>>>>> A domid other than 0 in xsd_sockmsg will be handled by Xenstore to use
>>>>> the privileges of that domain. Specifying a domid in xsd_sockmsg is
>>>>> allowed for privileged domain only, of course. XS_RESTRICT via a
>>>>> non-socket connection will be rejected in all cases.
>>>>>
>>>>> The needed modifications for Xenstore and the kernel are rather small.
>>>>> As there is currently no Xenstore domain available supporting
>>>>> XS_RESTRICT there are no compatibility issues to expect.
>>>>>
>>>>> Thoughts?
>>>>
>>>> As I don't get any further constructive responses even after asking for
>>>> them: would patches removing all XS_RESTRICT support be accepted?
>>>>
>>>
>>> We don't need to actually remove it, do we? If XS_RESTRICT is not supported 
>>> by
>>> xenstored, the client would get meaningful error code. A patch to
>>> deprecate that command should be good enough, right?
>>
>> Uuh, no.
>>
>> oxenstored does support XS_RESTRICT. The longer it stays the better the
>> chances someone is using it.
>>
>>> And sorry for the late reply, I'm still mulling over your proposal, I
>>> will try to respond as soon as possible.
>>
>> I thought a little bit further: the idea of XS_RESTRICT is to avoid qemu
>> being capable to overwrite any Xenstore entries of other domains
>> including dom0.
>>
>> I fail to see how this should work with qemu-based backends (qdisk,
>> pvusb), as those rely on paths in Xenstore writable by dom0 only.
> 
> It does not work. However, QEMU based backends can be run on a separate
> QEMU. Patches were submitted by IanJ and me to run 2 QEMUs per domain,
> one to provide emulation, the other to provide the backends. Not sure
> what happen to them, but they were more then prototypes.
> 
> 
>> We already have a mechanism to de-privilege the device model of a HVM
>> domain without hurting the backends: ioemu-stubdom. So I believe we
>> should try to make qmeu upstream usable in stubdom instead of
>> introducing mechanisms limited in usability ("if you want a secure
>> device model you can't use features x, y and z.").
> 
> Yes, but ioemu-stubdoms have drawbacks that make them not viable in many
> scenarios. There are reasons why they are not enabled by default.
> XS_RESTRICT should not replace, but complement ioemu-stubdoms. If we
> remove XS_RESTRICT, what's the plan to make QEMU in Dom0 secure by
> default?

Currently none. OTOH there is no plan how to make XS_RESTRICT work in
other cases like Xenstore domain.

We need to design a solution which has no such drawbacks. We don't have
to implement them all from beginning, but we should know how to do it.
Otherwise something like XS_RESTRICT will be the result which isn't
activated as default as it isn't working for all cases.


Juergen


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