[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] XSM dummy policy blocking event channel creation



On 05/02/2013 11:37 AM, Ross Philipson wrote:
On 05/02/2013 11:10 AM, Daniel De Graaf wrote:
On 05/02/2013 10:23 AM, Ross Philipson wrote:
On 05/01/2013 06:02 PM, Daniel De Graaf wrote:
On 05/01/2013 05:09 PM, Ross Philipson wrote:
On 05/01/2013 04:50 PM, Daniel De Graaf wrote:
On 05/01/2013 03:40 PM, Ross Philipson wrote:
I am working on the next set of V4V patches to post to the list. I
have pulled the very latest staging branch of Xen and I quickly ran
into a new problem. We are basically trying to create an event
channel
during the creation of dom0. We have split up the
evtchn_alloc_unbound() function into two functions but the basic
problem is the same. The call to xsm_evtchn_unbound() is returning
-EPERM from the new code in xsm/dummy.h. This patch set added this
functionality:

http://lists.xen.org/archives/html/xen-devel/2012-11/msg01920.html

Specifically we are failing this part of the test return -EPERM:

static always_inline int xsm_default_action(
xsm_default_t action, struct domain *src, struct domain *target)
{
...
case XSM_TARGET:
if ( src != target && !IS_PRIV_FOR(src, target) )
return -EPERM;

The src domain is the current->domain which is idle_domain and target
is dom0 which is in the process of being created. Neither of them is
privileged (dom0 is not set to privileged yet). And I have not gotten
past dom0 creation yet so I don't know what will happen when V4V
tries
to initialize for domU's.

I need some advice on how to proceed here. I am not terribly
conversant in the working of XSM and do not have a clear idea how to
proceed.

Thanks,
Ross Philipson

The answer for domUs is actually easier: the domain builder (dom0)
will
satisfy IS_PRIV_FOR(dom0, newdomU) and so the creation will be
allowed.
With FLASK, the remote domain will also be considered, so I presume
you
have set something valid there (it needs to have a struct domain* that
rcu_lock_domain_by_any_id can be used to fetch).

Event channels created by the hypervisor are normally allocated using
alloc_unbound_xen_event_channel instead of evtchn_alloc_unbound; the
former will not fail the creation of the event channel if the XSM hook
denies, but will instead change the remote domain ID to DOMID_INVALID.
Depending on what you intend the remote domid to be in this newly
allocated event channel, this may or may not be useful behavior; the
semantics of alloc_unbound_xen_event_channel are also different, so
you would need to change your calling code with this in mind.

If you want to use evtchn_alloc_unbound directly, you probably need to
add logic to omit the XSM check during the creation of dom0 - and this
likely belongs in the evtchn_alloc_unbound function rather than the
dummy XSM hook. The XSM changes you referenced don't actually
introduce
this, since before the call to rcu_lock_target_domain_by_id would have
failed because the idle domain does not satisfy IS_PRIV_FOR on dom0.

Thank you for getting back to me so quickly.

Perhaps I should share the new split up evtchn_alloc* functions since
the way they are done is relevant to your answer. I attached a snippet
with the split functions. We are directly calling
evtchn_alloc_unbound_domain() which does not call
rcu_lock_domain_by_any_id(). So the permissions block is related only
to the changes in dummy.c I believe.

Well, the way you have it written happens to avoid the permission check,
presumably because the hypervisor is doing the creation. I can't tell if
this is still correct in the new function since I can't see who calls
it;
if this is only called during domain creation and remote_domid is always
something special (DOMID_SELF, perhaps?) then completely removing the
XSM
check would be justified for internal callers.

The very first time the call is made to create the event channel in
V4V, the target is dom0 and the src (which is current->domain) is the
idle_domain. This is during creation of dom0 and this is the case that
fails the XSM check. Maybe this is the one we can special case and
bypass the XSM hook.

All subsequent calls that I see are either src == dom0 and target is
dom0 or some other domU which of course passes. I also see cases where
src == target for both dom0 and domU's which also pass the test.

In the case of V4V calling evtchn_alloc_unbound_domain(), remote_domid
is always the same as the domid which becomes the target domid in the
XSM calls. I am attaching the relevant v4v_init() code for clarity.

Actually, I think the call to evtchn_alloc_unbound_domain in v4v_init
should always bypass the XSM hook. The invocation of this function is
either during dom0 creation (which is always allowed) or has already
passed the access check for domain creation. Since this event channel
is always bound to the local domain, there is also no need for a FLASK
check for creating a new communication path. This would also avoid
special-casing the idle domain's creation of event channels.

Ok that seems to make sense to me. Would a boolean flag passed to 
evtchn_alloc_unbound_domain() be the way to go in case something else in the 
future wants to use that routine? Also evtchn_alloc_unbound() calls to the 
second routine and that call always used the XSM call before.

That would work; it looks like it would be easier than pulling the XSM hook
out of the function. The important part is that the user-accessible caller
still needs to execute the hook. Other callers bypassing the permission
check will need similar justification for doing that, but that discussion
should wait until those additional callers exist.
The goal of the XSM check here is to control the creation of event
channels:
we don't want a domain creating event channels for another unless it
controls
that domain (the check between current->domain and d that you have hit
here),
and with FLASK we also want to control what domains can communicate with
each
other (which is the reason remote_domid is passed to the XSM hook). V4V
will
require its own XSM hooks to satisfy the second goal, and the first does
not
apply when the hypervisor is the entity creating the event channel.


What I am confused about here is that the remote_domid that is passed
to evtchn_alloc_unbound_domain() is then passed to
xsm_evtchn_unbound() as id2 but is never used in the XSM code.

It's not used in the dummy XSM module, but it is used in FLASK; see the
flask_evtchn_unbound function in xen/xsm/flask/hooks.c, which is called
by xsm_evtchn_unbound if using FLASK. If V4V ends up calling this XSM
hook before the target domain is inserted into the domain hash, the
prototype for the function will need to be changed to take a struct
domain* instead of a domid for the target (or FLASK will always deny the
creation because it cannot determine the target).

Make sense, thanks.


I am still digging through the history of the posts and responses to
see how they settled on that configuration of the evtchn_alloc*
functions but this might explain things for you a bit better (e.g. why
it worked prior to the dummy.c changes).

Thanks,
Ross

What is this event channel being used for - is it similar to the HVM
ioreq
event channel created on a domain that someone else can bind and send
to, or
a VIRQ implemented as an unbound event channel, or something else?
What are
you passing as remote_domid here?

Yes it is like a VIRQ implemented as a unbound event channel. Each
domain has a v4v structure associated with it. The event channel is
used to signal a guess from within Xen V4V bits. It used to actually
be a VIRQ but we were asked to change it to its current form.

OK, that makes sense.

The V4V-like code that I am currently using uses a VIRQ for this
purpose, as
does the last V4V patch RFC that I checked (dated 2012-05-31).


Thanks
Ross

--
Daniel De Graaf
National Security Agency

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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.