[PATCH] qemu: Label vhostuser net device

[Maxime Coquelin]

On 8/19/21 11:10 AM, Michael S. Tsirkin wrote:
> On Wed, Aug 18, 2021 at 09:00:52AM -0400, Laine Stump wrote:
>> On 8/18/21 4:21 AM, Michal Prívozník wrote:
>>> On 8/18/21 12:31 AM, Jim Fehlig wrote:
>>>> On 8/17/21 12:11 PM, Jim Fehlig wrote:
>>>>> On 8/17/21 4:13 AM, Michal Prívozník wrote:
>>>>>> On 8/13/21 11:36 PM, Jim Fehlig wrote:
>>>>>>> Attaching a newly created vhostuser port to a VM fails due to an
>>>>>>> apparmor denial
>>>>>>>
>>>>>>> internal error: unable to execute QEMU command 'chardev-add': Failed
>>>>>>> to bind socket to /run/openvswitch/vhu838c4d29-c9: Permission denied
>>>>>>>
>>>>>>> In the case of a net device type VIR_DOMAIN_NET_TYPE_VHOSTUSER, the
>>>>>>> underlying chardev is not labeled in qemuDomainAttachNetDevice prior
>>>>>>> to calling qemuMonitorAttachCharDev. Label the chardev before calling
>>>>>>> qemuMonitorAttachCharDev, and restore the label when removing the
>>>>>>> net device.
>>>>>>>
>>>>>>> Signed-off-by: Jim Fehlig <jfehlig at suse.com>
>>>>>>> ---
>>>>>>>    src/qemu/qemu_hotplug.c | 9 +++++++++
>>>>>>>    1 file changed, 9 insertions(+)
>>>>>>>
>>>>>>> diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c
>>>>>>> index c00e8a7852..42e7997112 100644
>>>>>>> --- a/src/qemu/qemu_hotplug.c
>>>>>>> +++ b/src/qemu/qemu_hotplug.c
>>>>>>> @@ -1467,6 +1467,11 @@ qemuDomainAttachNetDevice(virQEMUDriver *driver,
>>>>>>>        }
>>>>>>>        if (actualType == VIR_DOMAIN_NET_TYPE_VHOSTUSER) {
>>>>>>> +        virDomainChrDef chr = { .source = net->data.vhostuser };
>>>>>>> +
>>>>>>> +        if (qemuSecuritySetChardevLabel(driver, vm, &chr) < 0)
>>>>>>> +            goto cleanup;
>>>>>>> +
>>>>>>>            if (qemuMonitorAttachCharDev(priv->mon, charDevAlias,
>>>>>>> net->data.vhostuser) < 0) {
>>>>>>>                ignore_value(qemuDomainObjExitMonitor(driver, vm));
>>>>>>>                virDomainAuditNet(vm, NULL, net, "attach", false);
>>>>>>> @@ -4692,6 +4697,8 @@ qemuDomainRemoveNetDevice(virQEMUDriver *driver,
>>>>>>>        }
>>>>>>>        if (actualType == VIR_DOMAIN_NET_TYPE_VHOSTUSER) {
>>>>>>> +        virDomainChrDef chr = { .source = net->data.vhostuser };
>>>>>>> +
>>>>>>>            /* vhostuser has a chardev too */
>>>>>>>            if (qemuMonitorDetachCharDev(priv->mon, charDevAlias) < 0) {
>>>>>>>                /* well, this is a messy situation. Guest visible PCI
>>>>>>> device has
>>>>>>> @@ -4699,6 +4706,8 @@ qemuDomainRemoveNetDevice(virQEMUDriver *driver,
>>>>>>>                 * to just ignore the error and carry on.
>>>>>>>                 */
>>>>>>>            }
>>>>>>> +        if (qemuSecurityRestoreChardevLabel(driver, vm, &chr) < 0)
>>>>>>> +            VIR_WARN("Unable to restore security label on vhostuser
>>>>>>> char device");
>>>>>>>        } else if (actualType == VIR_DOMAIN_NET_TYPE_VDPA) {
>>>>>>>            int vdpafdset = -1;
>>>>>>>            g_autoptr(qemuMonitorFdsets) fdsets = NULL;
>>>>>>>
>>>>>>
>>>>>> What an interesting bug. Previously we assumed that the UNIX socket is
>>>>>> created with broad enough permissions so that QEMU can connect to it. I
>>>>>> mean, when a VM is starting up nor DAC nor SELinux drivers care about
>>>>>> VHOSTUSER. It's AppArmor driver that cares. And it makes sense.
>>>>>> But, what's problematic here is that upon attach the socket perms will
>>>>>> be changed (because both DAC and SELinux drivers implement
>>>>>> domainSetSecurityChardevLabel callback). And since sockets can't have
>>>>>> XATTRs libvirt won't remember its original labels and thus subsequent
>>>>>> restore changes owner to root:root.
>>>>>
>>>>> How are existing chardevs with socket backends handled? It seems they
>>>>> would suffer the same problem when restoring the labels. The DAC and
>>>>> selinux callbacks seem to avoid labeling for "server" sockets
>>>>>
>>>>> https://gitlab.com/libvirt/libvirt/-/blob/master/src/security/security_dac.c#L1534
>>>>>
>>>>> https://gitlab.com/libvirt/libvirt/-/blob/master/src/security/security_selinux.c#L2544
>>>>>
>>>
>>> Yeah, you're right. But I view chardevs as runtime, fire and go things.
>>> I mean, you start a VM with a chardev, say an UNIX socket and its
>>> lifetime is identical to the VM lifetime. vhostuser on the other hand is
>>> handled by a third party daemon (typically OVS) and thus UNIX socket
>>> lifetime is different to VM lifetime. IOW, I worry that if we changed
>>> the UNIX socket label we might be preventing other VMs from connecting
>>> to it.
>>>
>>> And I don't even know if a single socket can be shared between two VMs.
>>> For instance, if OVS created an UNIX socket whether I can attach
>>> vhostuser interface to one domain and then to the other. Because if I
>>> could, then we should not change labels.

In theory, I think there is nothing that prevents a Vhost-user socket
created by OVS to be used by one VM and later by another one. But in
practice, the management layer likely request OVS a new socket on VM
creation.

I'd like to add that the mode where OVS is the Vhost-user socket server
is deprecated in OVS (but not yet removed) [0]:

"
Open vSwitch provides two types of vHost User ports:

    vhost-user (dpdkvhostuser)
    vhost-user-client (dpdkvhostuserclient)

vHost User uses a client-server model. The server
creates/manages/destroys the vHost User sockets, and the client connects
to the server. Depending on which port type you use, dpdkvhostuser or
dpdkvhostuserclient, a different configuration of the client-server
model is used.

For vhost-user ports, Open vSwitch acts as the server and QEMU the
client. This means if OVS dies, all VMs must be restarted. On the other
hand, for vhost-user-client ports, OVS acts as the client and QEMU the
server. This means OVS can die and be restarted without issue, and it is
also possible to restart an instance itself. For this reason,
vhost-user-client ports are the preferred type for all known use cases;
the only limitation is that vhost-user client mode ports require QEMU
version 2.7. Ports of type vhost-user are currently deprecated and will
be removed in a future release.
"

>>>
>>> Perhaps Laine can shed more light here.
>>
>> Sorry, I really know nothing about this topic. Who would be the proper
>> person at the next level down to ask? Michael, do you know?
> 
> hmm...
> Maxime?
> 
> 
>>
>>>
>>> I do understand that apparmor needs to add an entry to the VM's profile,
>>> but I worry that DAC and SELinux might screw things up.
>>>
>>>>>
>>>>>> So I think we should address this inconsistency in behavior. But I don't
>>>>>> know how :-(
>>>>>
>>>>> My first attempt at fixing this introduced
>>>>> domain{Set,Restore}SecurityNetdevLabel to the security driver, but it
>>>>> seemed like overkill after I discovered virDomainChrSourceDef embedded
>>>>> in the virDomainNetDef structure. I can revisit that approach if it
>>>>> sounds more promising.
>>>
>>> I think it does sound more promising given my assumption above is
>>> correct. This way we can have only AppArmor driver implement the
>>> callback leaving us with consistent behavior.
>>>
>>> Michal
>>>
>