[PATCH 1/3] qemu_passt: Don't make passt transition to svirt_t/virt_domain on start

Wed Feb 22 11:30:10 UTC 2023

On Wed, 22 Feb 2023 12:21:09 +0100
Michal Prívozník <mprivozn at redhat.com> wrote:

> On 2/22/23 11:05, Stefano Brivio wrote:
> > On Wed, 22 Feb 2023 09:46:42 +0000
> > Daniel P. Berrangé <berrange at redhat.com> wrote:
> >   
> >> On Tue, Feb 21, 2023 at 10:49:46PM +0100, Stefano Brivio wrote:  
> >>> On Tue, 21 Feb 2023 19:43:33 +0000
> >>> Daniel P. Berrangé <berrange at redhat.com> wrote:
> >>>     
> >>>> On Tue, Feb 21, 2023 at 08:19:05PM +0100, Stefano Brivio wrote:    
> >>>>> qemuSecurityCommandRun() causes an explicit domain transition of the
> >>>>> new process, but passt ships with its own SELinux policy, with
> >>>>> external interfaces for libvirtd, so we simply need to transition
> >>>>> from virtd_t to passt_t as passt is executed. The qemu type
> >>>>> enforcement rules have little to do with it.      
> >>>>
> >>>> Can you clarify the difference here ?    
> >>>
> >>> Between...?
> >>>
> >>> I mean, virCommandRun() will just keep things running under virtd_t, so
> >>> that passt later can transition to passt_t.
> >>>
> >>> With qemuSecurityCommandRun(), there would be a transition from virtd_t
> >>> to svirt_t (it's the function that's called to start qemu, but
> >>> shouldn't be called here), and not to passt_t.
> >>>
> >>> But I'm not really sure that's what you were asking for.    
> >>
> >> Yes, it is.
> >>  
> >>>     
> >>>> Runing passt under 'svirt_t' is not desirable as that allows
> >>>> many actions that are only relevant to QEMU.    
> >>>
> >>> Right, that's what this patch avoids. There are also actions, such as
> >>> starting passt or killing it, that we don't want to allow QEMU to do.
> >>>     
> >>>> Running passt under the MCS label that is associated with the
> >>>> VM is highly desirable though. Two passt instances belonging
> >>>> to separate VMs are isolated from each other if they each use
> >>>> the VM specific MCS label, than if they use the global default
> >>>> MCS label.
> >>>>
> >>>> To use the VM specific MCS label would require libvirt to
> >>>> explicitly set the desired selinux label on exec, it can't
> >>>> happen automatically via an SELinux transition rule.
> >>>>
> >>>> We do stil want to use the passt_t type though.
> >>>>
> >>>> IOW, if we have a VM running
> >>>>
> >>>>   svirt_t:s0:c710,c716
> >>>>
> >>>> IMHO we would its corresponding passt instance to be
> >>>> running
> >>>>
> >>>>   passt_t:s0:c710,c716
> >>>>
> >>>>
> >>>> not
> >>>>
> >>>>   passt_t:s0:c0.c1023    
> >>>
> >>> Practically speaking, it doesn't make a huge difference for passt
> >>> because it unshares any relevant namespace right after it starts --
> >>> that's *in theory* a strictly stronger isolation compared to what
> >>> SELinux provides (at least once we reach the main loop).    
> >>
> >> Even docker/podman will apply SELinux isolation per container,
> >> rather than only relying on namespaces.   
> > 
> > Sure, I'm not saying it's not desirable -- but still, many (most?)
> > host-facing services they rely on are not isolated in this sense. Same
> > for the current implementation of libvirt with dnsmasq, for example.
> >   
> >>> But it makes sense, and I guess we should relabel to a specific MCS
> >>> with still 'virtd_t' as a type, then later the domain would transition
> >>> to passt_t. This could probably be done as an extension of
> >>> qemuSecurityCommandRun(), I haven't looked into it yet. I will.
> >>>
> >>> Anyway, right now, I think this provides better security than
> >>> 'setenforce 0', which is the only way to run passt from libvirt at the
> >>> moment on some distributions.    
> >>
> >> If running with SELinux permissive, this patch has no effect, as it
> >> is unconfined regardless. That's not a situation we care about. If
> >> people want to run without protection they get to keep the pieces
> >> when it all goes wrong.  
> > 
> > The current implementation simply does not and cannot work with SELinux
> > in enforcing mode, so users have no other choice.
> >   
> >>> I'm not sure how you handle these cases on libvirt, but generally
> >>> speaking, this patch is a vast improvement on the current situation,
> >>> and I can follow up with an extension or a different version of
> >>> qemuSecurityCommandRun() later.    
> >>
> >> No, I don't think it is a vast improvement.
> >>
> >> The goal of sVirt is to guarantee isolation between VMs.
> >>
> >> Running passt under svirt_t:MCS is not ideal because the svirt_t
> >> policy allows things that passt should not get. It does still
> >> guarantee isolation between VMs, because the MCS label is present.  
> > 
> > It's a bit more than that: it's not ideal because libvirt simply won't
> > start passt. There's no isolation and no VMs.
> >   
> >> Switching to running passt_t:c0.c1023 will be more correct in terms
> >> of what permissions passt should be allowed, but it has disabled
> >> isolation of passt between VMs. This is a clear degradation of
> >> capabilities from the POV of sVirt's goals.  
> > 
> > It's not a degradation because VMs can't start passt with SELinux in
> > enforcing mode, without this patch. No service, no degradation.
> > 
> > I looked into options to rework
> > virSecurityManagerSetChildProcessLabel() and friends with a more
> > flexible modeling/building of labels -- just doing some trick all the
> > way down from qemuSecurityCommandRun() implies a number of layering
> > violations that I would like to avoid.
> > 
> > It all looks doable, but implementing the type of functionality/API
> > that's currently missing there isn't a small rework -- some refactoring
> > of interfaces is definitely needed. I started, but it's not quick.
> > 
> > So for the moment being I would suggest that, if passt can't be
> > relabeled as passt_t (i.e. if this patch or equivalent can't be
> > applied), the whole passt back-end should be dropped.  
> 
> I don't think we need such drastic measure. I think you can use:
> 
> qemuPasstStart()
> {
> 
> 
>   seclabel = virDomainDefGetSecurityLabelDef(vm->def, "selinux");
>   s = context_new(seclabel->label);
>   context_type_set(s, "virt_t);
>   newLabel = context_str(s);
> 
>   virCommandSetSELinuxLabel(cmd, newLabel);
> 
>   virCommandRun();
> }

Yes, I actually tried something like this and it seemed to work, but I
didn't propose it as it looks (is) gross.

On the other hand, if you think it's acceptable as a temporary measure,
let me test it (in a bit). Thanks for the snippet.

-- 
Stefano