When using command line passthrough users will often trip up over the
security protections like SELinux, DAC, namespaces, etc which will
deny access to files they are passing. This document explains the
various protections and how to deal with their policy, and/or how
to disable them.

Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
---
 docs/kbase.html.in                       |   4 +
 docs/kbase/qemu-passthrough-security.rst | 157 +++++++++++++++++++++++
 2 files changed, 161 insertions(+)
 create mode 100644 docs/kbase/qemu-passthrough-security.rst

diff --git a/docs/kbase.html.in b/docs/kbase.html.in
index c156414c41..db84b95b60 100644
--- a/docs/kbase.html.in
+++ b/docs/kbase.html.in
@@ -29,6 +29,10 @@
         <dt><a href="kbase/backing_chains.html">Backing chain management</a></dt>
         <dd>Explanation of how disk backing chain specification impacts libvirt's
           behaviour and basic troubleshooting steps of disk problems.</dd>
+
+        <dt><a href="kbase/qemu-passthrough-security.html">Security with QEMU passthrough</a></dt>
+        <dd>Examination of the security protections used for QEMU and how they need
+          configuring to allow use of QEMU passthrough with host files/devices.</dd>
       </dl>
     </div>
 
diff --git a/docs/kbase/qemu-passthrough-security.rst b/docs/kbase/qemu-passthrough-security.rst
new file mode 100644
index 0000000000..7fb1f6fbdd
--- /dev/null
+++ b/docs/kbase/qemu-passthrough-security.rst
@@ -0,0 +1,157 @@
+=============================
+QEMU command line passthrough
+=============================
+
+.. contents::
+
+Libvirt aims to provide explicit modelling of virtualization features in
+the domain XML document schema. QEMU has a very broad range of features
+and not all of these can be mapped to elements in the domain XML. Libvirt
+would like to reduce the gap to QEMU, however, with finite resources there
+will always be cases which aren't covered by the domain XML schema.
+
+
+XML document additions
+======================
+
+To deal with the problem, libvirt introduced support for command line
+passthrough of QEMU arguments. This is achieved by supporting a custom
+XML namespace, under which some QEMU driver specific elements are defined.
+
+The canonical place to declare the namespace is on the top level ``<domain>``
+element. At the very end of the document, arbitrary command line arguments
+can now be added, using the namespace prefix ``qemu:``
+
+::
+
+   <domain type='kvm' xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'>
+     <name>QEMUGuest1</name>
+     <uuid>c7a5fdbd-edaf-9455-926a-d65c16db1809</uuid>
+     ...
+     <qemu:commandline>
+       <qemu:arg value='-newarg'/>
+       <qemu:arg value='parameter'/>
+       <qemu:env name='ID' value='wibble'/>
+       <qemu:env name='BAR'/>
+     </qemu:commandline>
+   </domain>
+
+Note that when an argument takes a value eg ``-newarg parameter``, the argument
+and the value must be passed as separate ``<qemu:arg>`` entries.
+
+Instead of declaring the XML namespace on the top level ``<domain>`` it is also
+possible to declare it at time of use, which is more convenient for humans
+writing the XML documents manually. So the following example is functionally
+identical:
+
+::
+
+   <domain type='kvm'>
+     <name>QEMUGuest1</name>
+     <uuid>c7a5fdbd-edaf-9455-926a-d65c16db1809</uuid>
+     ...
+     <commandline xmlns="http://libvirt.org/schemas/domain/qemu/1.0">
+       <arg value='-newarg'/>
+       <arg value='parameter'/>
+       <env name='ID' value='wibble'/>
+       <env name='BAR'/>
+     </commandline>
+   </domain>
+
+Note that when querying the XML from libvirt, it will have been translated into
+the canonical syntax once more with the namespace on the top level element.
+
+Security confinement / sandboxing
+=================================
+
+When libvirt launches a QEMU process it makes use of a number of security
+technologies to confine QEMU and thus protect the host from malicious VM
+breakouts.
+
+When configuring security protection, however, libvirt generally needs to know
+exactly which host resources the VM is permitted to access. It gets this
+information from the domain XML document. This only works for elements in the
+regular schema, the arguments used with command line passthrough are completely
+opaque to libvirt.
+
+As a result, if command line passthrough is used to expose a file on the host
+to QEMU, the security protections will activate and either kill QEMU or deny it
+access.
+
+There are two strategies for dealing with this problem, either figure out what
+steps are needed to grant QEMU access to the device, or disable the security
+protections.  The former is harder, but more secure, while the latter is simple.
+
+Granting access per VM
+----------------------
+
+* SELinux - the file on the host needs an SELinux label that will grant access
+  to QEMU's ``svirt_t`` policy.
+
+  - Read only access - use the ``virt_content_t`` label
+  - Shared, write access - use the ``svirt_image_t:s0`` label (ie no MCS
+    category appended)
+  - Exclusive, write access - use the ``svirt_image_t:s0:MCS`` label for the VM.
+    The MCS is auto-generatd at boot time, so this may require re-configuring
+    the VM to have a fixed MCS label
+
+* DAC - the file on the host needs to be readable/writable to the ``qemu``
+  user or ``qemu`` group. This can be done by changing the file ownership to
+  ``qemu``, or relaxing the permissions to allow world read, or adding file
+  ACLs to allow access to ``qemu``.
+
+* Namespaces - a private ``mount`` namespace is used for QEMU by default
+  which populates a new ``/dev`` with only the device nodes needed by QEMU.
+  There is no way to augment the set of device nodes ahead of time.
+
+* Seccomp - libvirt launches QEMU with its built-in seccomp policy enabled with
+  ``obsolete=deny``, ``elevateprivileges=deny``, ``spawn=deny`` and
+  ``resourcecontrol=deny`` settings active. There is no way to change this
+  policy on a per VM basis
+
+* Cgroups - a custom cgroup is created per VM and this will either use the
+  ``devices`` controller or an ``BPF`` rule to whitelist a set of device nodes.
+  There is no way to change this policy on a per VM basis.
+
+Disabling security protection per VM
+------------------------------------
+
+Some of the security protections can be disabled per-VM:
+
+* SELinux - in the domain XML the ``<seclabel>`` model can be changed to
+  ``none`` instead of ``selinux``, which will make the VM run unconfined.
+
+* DAC - in the domain XML an ``<seclabel>`` element with the ``dac`` model can
+  be added, configured with a user / group account of ``root`` to make QEMU run
+  with full privileges
+
+* Namespaces - there is no way to disable this per VM
+
+* Seccomp - there is no way to disable this per VM
+
+* Cgroups - there is no way to disable this per VM
+
+Disabling security protection host-wide
+---------------------------------------
+
+As a last resort it is possible to disable security protection host wide which
+will affect all virtual machines. These settings are all made in
+``/etc/libvirt/qemu.conf``
+
+* SELinux - set ``security_default_confied = 0`` to make QEMU run unconfined by
+  default, while still allowing explicit opt-in to SELinux for VMs.
+
+* DAC - set ``user = root`` and ``group = root`` to make QEMU run as the root
+  account
+
+* SELinux, DAC - set ``security_driver = []`` to entirely disable both the
+  SELinux and DAC security drivers.
+
+* Namespaces - set ``namespaces = []`` to disable use of the ``mount``
+  namespaces, causing QEMU to see the normal fully popualated ``dev``
+
+* Seccomp - set ``seccomp_sandbox = 0`` to disable use of the Seccomp sandboxing
+  in QEMU
+
+* Cgroups - set ``cgroup_device_acl`` to include the desired device node, or
+  ``cgroup_controllers = [...]`` to exclude the ``devices`` controller.
-- 
2.24.1

On Thu, Feb 06, 2020 at 01:05:37PM +0000, Daniel P. Berrangé wrote:
>When using command line passthrough users will often trip up over the
>security protections like SELinux, DAC, namespaces, etc which will
>deny access to files they are passing. This document explains the
>various protections and how to deal with their policy, and/or how
>to disable them.
>
>Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
>---
> docs/kbase.html.in                       |   4 +
> docs/kbase/qemu-passthrough-security.rst | 157 +++++++++++++++++++++++
> 2 files changed, 161 insertions(+)
> create mode 100644 docs/kbase/qemu-passthrough-security.rst
>

Reviewed-by: Ján Tomko <jtomko@redhat.com>

Jano

On Thu, Feb 06, 2020 at 01:05:37PM +0000, Daniel P. Berrangé wrote:

The core content reads very well.  A couple of minor nit-picks inline.

[...]

> diff --git a/docs/kbase/qemu-passthrough-security.rst b/docs/kbase/qemu-passthrough-security.rst
> new file mode 100644
> index 0000000000..7fb1f6fbdd
> --- /dev/null
> +++ b/docs/kbase/qemu-passthrough-security.rst
> @@ -0,0 +1,157 @@

[...]

> +XML document additions
> +======================
> +
> +To deal with the problem, libvirt introduced support for command line

Nit: s/command line/command-line/g  (there are a few occurrences)

> +passthrough of QEMU arguments. This is achieved by supporting a custom
> +XML namespace, under which some QEMU driver specific elements are defined.
> +
> +The canonical place to declare the namespace is on the top level ``<domain>``
> +element. At the very end of the document, arbitrary command line arguments
> +can now be added, using the namespace prefix ``qemu:``
> +
> +::

If you can stomach the syntax chance, you can put the :: at the end of
the sentence.

> +
> +   <domain type='kvm' xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'>
> +     <name>QEMUGuest1</name>
> +     <uuid>c7a5fdbd-edaf-9455-926a-d65c16db1809</uuid>
> +     ...
> +     <qemu:commandline>
> +       <qemu:arg value='-newarg'/>
> +       <qemu:arg value='parameter'/>

I'd guess you intentionally took a generic example, rather than specific
QEMU command-line parameter to illustrate the XML, in case the example
command-line is deprecated, etc.

> +       <qemu:env name='ID' value='wibble'/>
> +       <qemu:env name='BAR'/>
> +     </qemu:commandline>
> +   </domain>

Is it worth calling out that the 'env' fragments are envirnoment
variables?  As it isn't obvious to those who don't dwell on libvirt/QEMU
daily.

> +Note that when an argument takes a value eg ``-newarg parameter``, the argument
> +and the value must be passed as separate ``<qemu:arg>`` entries.
>
> +
> +Instead of declaring the XML namespace on the top level ``<domain>`` it is also
> +possible to declare it at time of use, which is more convenient for humans
> +writing the XML documents manually. So the following example is functionally
> +identical:
> +
> +::

Here too, you can put the :: at the end of the sentence, saving one
colon :D

> +
> +   <domain type='kvm'>
> +     <name>QEMUGuest1</name>
> +     <uuid>c7a5fdbd-edaf-9455-926a-d65c16db1809</uuid>
> +     ...
> +     <commandline xmlns="http://libvirt.org/schemas/domain/qemu/1.0">
> +       <arg value='-newarg'/>
> +       <arg value='parameter'/>
> +       <env name='ID' value='wibble'/>
> +       <env name='BAR'/>
> +     </commandline>
> +   </domain>
> +
> +Note that when querying the XML from libvirt, it will have been translated into
> +the canonical syntax once more with the namespace on the top level element.

Here you might want to use the rST "note" admonition:

.. note:: When querying the XML from libvirt, it will have been
          translated into  canonical syntax once more with the namespace
          on the top level element.

    
> +
> +Security confinement / sandboxing
> +=================================
> +
> +When libvirt launches a QEMU process it makes use of a number of security
> +technologies to confine QEMU and thus protect the host from malicious VM
> +breakouts.
> +
> +When configuring security protection, however, libvirt generally needs to know
> +exactly which host resources the VM is permitted to access. It gets this
> +information from the domain XML document. This only works for elements in the
> +regular schema, the arguments used with command line passthrough are completely
> +opaque to libvirt.
> +
> +As a result, if command line passthrough is used to expose a file on the host
> +to QEMU, the security protections will activate and either kill QEMU or deny it
> +access.
> +
> +There are two strategies for dealing with this problem, either figure out what
> +steps are needed to grant QEMU access to the device, or disable the security
> +protections.  The former is harder, but more secure, while the latter is simple.
> +
> +Granting access per VM
> +----------------------
> +
> +* SELinux - the file on the host needs an SELinux label that will grant access
> +  to QEMU's ``svirt_t`` policy.
> +
> +  - Read only access - use the ``virt_content_t`` label

Nit: s/"Read only"/Read-only/

> +  - Shared, write access - use the ``svirt_image_t:s0`` label (ie no MCS
> +    category appended)
> +  - Exclusive, write access - use the ``svirt_image_t:s0:MCS`` label for the VM.
> +    The MCS is auto-generatd at boot time, so this may require re-configuring
> +    the VM to have a fixed MCS label
> +
> +* DAC - the file on the host needs to be readable/writable to the ``qemu``

Nit: let's please expand acronyms on first use: "Discretionary Access
Control (DAC)"; although DAC and ACL (below) might be common enough for
"Linux dwellers" that we don't have to be pedantic about it.  But MCS
(Multi-Category Security) is familiar only for those who are
SELinux-aware.

So, your choice, as I don't want to make you expand every acronym; but
only the obscure ones. :-)

> +  user or ``qemu`` group. This can be done by changing the file ownership to
> +  ``qemu``, or relaxing the permissions to allow world read, or adding file
> +  ACLs to allow access to ``qemu``.
> +
> +* Namespaces - a private ``mount`` namespace is used for QEMU by default
> +  which populates a new ``/dev`` with only the device nodes needed by QEMU.
> +  There is no way to augment the set of device nodes ahead of time.
> +
> +* Seccomp - libvirt launches QEMU with its built-in seccomp policy enabled with
> +  ``obsolete=deny``, ``elevateprivileges=deny``, ``spawn=deny`` and
> +  ``resourcecontrol=deny`` settings active. There is no way to change this
> +  policy on a per VM basis

Missing full stop at the end here ...

> +
> +* Cgroups - a custom cgroup is created per VM and this will either use the
> +  ``devices`` controller or an ``BPF`` rule to whitelist a set of device nodes.
> +  There is no way to change this policy on a per VM basis.
> +
> +Disabling security protection per VM
> +------------------------------------
> +
> +Some of the security protections can be disabled per-VM:
> +
> +* SELinux - in the domain XML the ``<seclabel>`` model can be changed to
> +  ``none`` instead of ``selinux``, which will make the VM run unconfined.
> +
> +* DAC - in the domain XML an ``<seclabel>`` element with the ``dac`` model can
> +  be added, configured with a user / group account of ``root`` to make QEMU run
> +  with full privileges

... here,

> +* Namespaces - there is no way to disable this per VM
> +
> +* Seccomp - there is no way to disable this per VM
> +
> +* Cgroups - there is no way to disable this per VM
> +
> +Disabling security protection host-wide
> +---------------------------------------
> +
> +As a last resort it is possible to disable security protection host wide which
> +will affect all virtual machines. These settings are all made in
> +``/etc/libvirt/qemu.conf``

... and here.

> +
> +* SELinux - set ``security_default_confied = 0`` to make QEMU run unconfined by
> +  default, while still allowing explicit opt-in to SELinux for VMs.
> +
> +* DAC - set ``user = root`` and ``group = root`` to make QEMU run as the root
> +  account
> +
> +* SELinux, DAC - set ``security_driver = []`` to entirely disable both the
> +  SELinux and DAC security drivers.
> +
> +* Namespaces - set ``namespaces = []`` to disable use of the ``mount``
> +  namespaces, causing QEMU to see the normal fully popualated ``dev``
> +
> +* Seccomp - set ``seccomp_sandbox = 0`` to disable use of the Seccomp sandboxing
> +  in QEMU
> +
> +* Cgroups - set ``cgroup_device_acl`` to include the desired device node, or
> +  ``cgroup_controllers = [...]`` to exclude the ``devices`` controller.


I'll let you pick what you want to address, as this doc is an
improvement as-is, FWIW:

Reviewed-by: Kashyap Chamarthy <kchamart@redhat.com>

-- 
/kashyap