Re: [RFC PATCH 1/1] xsm: allows system domains to allocate evtchn

["Daniel P. Smith" <dpsmith@xxxxxxxxxxxxxxxxxxxx>]

On 3/29/22 03:29, Roger Pau Monné wrote:
> On Mon, Mar 28, 2022 at 04:36:22PM -0400, Daniel P. Smith wrote:
>> During domain construction under dom0less and hyperlaunch it is necessary to
>> allocate at least the event channel for xenstore and potentially the event
>> channel for the core console. When dom0less and hyperlaunch are doing their
>> construction logic they are executing under the idle domain context. The idle
>> domain is not a privileged domain, it is not the target domain, and as a 
>> result
>> under the current default XSM policy is not allowed to allocate the event
>> channel.
> 
> I've not been following the discussion around this patch, but I would
> assume such privileges are only required for init code when no other
> domains are running?

At this time, correct.

> Since it's only at that point where the idle domain context needs to
> allocate event channels would it make sense to temporary elevate it's
> privileges by setting d->is_privileged while doing the domain creation?

This is initially what I did but it seemed like there was some
reluctance. As I was looking to formalize/abstract this in XSM instead
of doing direct manipulations, I realized I could achieve it in the hook
which would allow the hyperlaunch and dom0less code work without having
to ensure priv escalation is properly handled.

> That way we wouldn't need to grant those permissions for the lifetime
> of the host when they are only needed for initialization code.

Correct, which is why I adjusted the effective default policy only on
the check instead of in xsm_default_action() as Jan has suggested.
Outside of a code fault, all other times that evtchn_alloc_unbound() is
called `current->domain` should be pointing at the caller of the hypercall.

This works as an interim solution with minimal impact as it is all
internal to XSM and can easily be evolved. My concern is that exposing a
function call to provide priv escalation for the idle domain as an
interim solution for dom0less and hyperlaunch will have more impactful
code churn in both of these when a longer term approach is adopted.

> Another option would be switching to the initial vCPU of the domain
> being created, but that's likely to be more complex, or even create a
> short lived system domain with is_privileged set just for the purpose
> of building other domains.

Longer term I would like to explore doing this in general. Some initial
thinking is the fact that hypervisor has a few contexts, relative to
external entities, under which it is executing. When it is handling
internal house keeping (e.g. scheduler and security server), when it is
interacting with guest domains, when it is interacting with hardware
(e.g. vpci), and now when it is processing boot material to construct
domains. It  has been mentioned that today in Xen if one of these
contexts acting with external entities is corrupted, it can interfere
with operations occurring in the other contexts. In the past the have
advocated and been working to split these contexts using hard L0/L1
separation. As noted in other discussions, some architectures are
gaining hardware features that can be used in hard L0/L1 partitioning
but also could be used in a more "soft" partitioning more a kin to
Nested Kernel[1] and Dune[2]. Again just some initial thoughts.

> Overall I'm not sure it's worth giving those extra privileges to the
> idle domain when they are just need for a known and bounded period of
> time.

IMHO that is a slight over simplification. Setting is_privileged to the
idle domain while it is processing domain construction data from outside
the hypervisor means that during that bounded period the idle domain is
complete unrestricted and may invoke any XSM protected call. Contrast
this with only granting the idle domain the ability to allocate event
channels between domains at any time with the only codified usage is
during init/setup. While I am unsure how, theoretically malformed
construction data could expose a logic flaw to do some very unsavory
allocations without any guards. Whereas during runtime if the idle
domain was tricked into establishing an event channel between two
domains, it would only serve to provide a covert channel between the two
domains. Neither is desirable but IMHO I find the former a little more
concerning than the latter.

With that said, I am not completely against doing the priv escalation if
overall this is the direction that is preferred. If so, I would prefer
to provide a pair of static inlines under XSM name space to provide a
consistent implementation and be able to easily locate the places where
it is applied if/when a longer term approach is implemented.

v/r,
dps

[1]
https://nathandautenhahn.com/downloads/publications/asplos200-dautenhahn.pdf
[2]
https://web.stanford.edu/group/mast/cgi-bin/drupal/system/files/2012.dune_.osdi_.pdf