[Virtio-fs] virtiofs mounted filesystems & SELinux

[Link Dupont]

On Mon, Jun 7 2021 at 09:01:08 AM -0400, Daniel Walsh 
<dwalsh at redhat.com> wrote:
> On 6/4/21 09:59, Daniel P. Berrangé wrote:
>> On Fri, Jun 04, 2021 at 09:44:39AM -0400, Vivek Goyal wrote:
>>> On Thu, Jun 03, 2021 at 10:14:24PM -0400, Link Dupont wrote:
>>>> On Thu, Jun 3 2021 at 08:56:46 PM -0400, Link Dupont 
>>>> <link at sub-pop.net>
>>>> wrote:
>>>>>   reproducible scenarios
>>>> Alright. I reran my tests with a CentOS 8 guest. On CentOS 8 (with 
>>>> a
>>>> virtiofs filesystem and with xattr on), the type of files in the 
>>>> mounted
>>>> hierarchy are unlabeled_t. I can work around that by switching 
>>>> SELinux in
>>>> the guest to permissive or disabled.
>>> cc Dan Walsh. I was discussing this with Dan Walsh yesterday in 
>>> general.
>>> 
>>> In general, if we want to enable SELinux both on host and guest, 
>>> then
>>> both host and guest should have same SELinux policy. Otherwise there
>>> will be lot of different kind of conflicts because both host and
>>> guest will try to work with same selinux label. I guess that in
>>> practice this will be very hard to achieve as people will run
>>> different host and guest flavors and these might have different
>>> policies.
>> Yeah, I think there's little to no chance of people keeping the
>> same SELinux policy in host/guest, except in very tightly controlled
>> narrow use cases where the host admin exerts direct control over
>> the precise guest config.
>> 
>> 
>>> So another option is to rename selinux xattr in virtiofs so that
>>> any selinux xattr coming from guest is saved as
>>> user.virtiofs.security.selinux xattr on host. That way host and 
>>> guest
>>> can have their separate labels without interfering with each other.
>>> David Gilbert already has added support for this. I can't remember
>>> the exact syntax but you can figure it out from documentation here
>>> in xattr remappig section.
>> For general purpose virt usage, I think remapping in some way is
>> likely to be needed as the default strategy.
>> 
>>> https://github.com/qemu/qemu/blob/master/docs/tools/virtiofsd.rst
>>> 
>>> But I have question with selinux xattr remapping. What will happen
>>> to initial labels when fs is exported. I mean until and unless
>>> some process in guest labels all the exported files, they all
>>> with either be unlabeled or pick some generic label for all the
>>> files.
>> I'd say you need some mechanism to force a re-label inside the
>> guest. Normally a relabel will be done in /.autorelabel file
>> is present, or in certain other scenarios like selinux policy
>> RPM updates.
>> 
>> We wouldn't want to force a relabel neccesarily for the entire
>> FS if we're just hotplugging a new virtiofs export though. So
>> perhaps there's scope for supporting usage of a per-mount
>> point relabel trigger. eg Host creates $VIRTIOFS-ROOT/.autorelabel
>> and whenever the guest sees a new virtiofs export arriving, it
>> can look for $VIRTIOFS-MOUNT-POINT/.autorelabel
>> 
>>> Another option is, can we use a single label for whole of the
>>> virtiofs (using context=<label>) option in guest. That way nothing
>>> is saved in files as such. But this means that processes in guest
>>> can't have different selinux labels on different virtiofs dir/files.
>> Forcing a single label for the entire export is passable as a
>> fallback plan. This is what people have done for years with
>> NFS v3 mounts. It has annoying usage limitations though, so
>> if at all possible remapping is a preferrable approach.
>> 
>>> Dan, what do you think?
>>> 
>>> Thanks
>>> Vivek
>>> 
>>> 
>>>> With a CentOS 7 guest, things get less usable. I digested this to a
>>>> reproducible scenario.
>>>> 
>>>> Build a disk image with `virt-builder`, configuring the CentOS 
>>>> Plus kernel
>>>> to get 9p support.
>>>> 
>>>> virt-builder centos-7.8 \
>>>> --root-password password:centos \
>>>> --output centos-7.8.qcow2 \
>>>> --install yum-utils \
>>>> --run-command 'yum-config-manager --enable centosplus' \
>>>> --run-command 'sed -ie 
>>>> "s/DEFAULTKERNEL=kernel/DEFAULTKERNEL=kernel-plus/"
>>>> /etc/sysconfig/kernel' \
>>>> --append-line 
>>>> '/etc/dracut.conf.d/virtio.conf:add_drivers+="virtio_scsi
>>>> virtio_pci virtio_console"' \
>>>> --append-line '/etc/modules-load.d/9pnet_virtio.conf:9pnet_virtio' 
>>>> \
>>>> --install kernel-plus \
>>>> --append-line '/etc/fstab:home /home 9p 
>>>> trans=virtio,version=9p2000.L 0 0'
>>>> 
>>>> Install the volume into the `default` pool.
>>>> 
>>>> sudo install -m644 centos-7.8.qcow2 /var/lib/libvirt/images
>>>> 
>>>> Next, define a domain using the disk image (using `virt-install` 
>>>> here for
>>>> "easy mode").
>>>> 
>>>> virt-install \
>>>> --import \
>>>> --os-variant centos7.0 \
>>>> --name centos \
>>>> --ram 2048 \
>>>> --disk path=/var/lib/libvirt/images/centos-7.8.qcow2 \
>>>> --memorybacking access.mode=shared \
>>>> --filesystem source=/home,target=home,accessmode=passthrough \
>>>> --autoconsole none
>>>> 
>>>> Now with SELinux enforcing, I cannot list the contents of the 
>>>> directories in
>>>> the mounted hierarchy.
>>>> 
>>>> [root at localhost ~]# ls -lZ /home/link
>>>> ls: cannot open directory /home/link: Permission denied
>>>> 
>>>> 
>>>> 
>>>> _______________________________________________
>>>> Virtio-fs mailing list
>>>> Virtio-fs at redhat.com
>>>> https://listman.redhat.com/mailman/listinfo/virtio-fs
>>>> 
>> Regards,
>> Daniel
> 
> The separate XAttr support would also allow virtiofsd to work with 
> labels inside of the container when the virtiofsd is confined on the 
> host.
> 
> 
> I would guess/hope virtiofsd on the host will be confined and only 
> able to write certain labels like container_file_t or svirt_image_t.  
> If you want to allow virt file systems within the "VM" to be able to 
> set labels, these labels will have to be something other then SELinux 
> labels on the host, or SELinux will prevent them from being set.
> 

I don't think the guest needs to be able to set labels on the virtiofs 
filesystems; as long as it can read labels that grant read and write 
access as the guest user, I'd call it a success. The way I use this 
feature, I'm treating my host filesystem as the "source", and really 
only mounting it into the guest as a zero-delay synchronization. (I 
previously did this using sync solutions like rsync or lsyncd, but 
those have a slight delay before files sync over.)

If the labels as seen by the guest allow for reading and writing files 
from the guest to the host filesystem, I'd be happy. Doing any 
relabeling from the host is totally fine.