libvirt/docs/kbase/virtiofs.rst
Peter Krempa 590a6c1b78 kbase: virtiofs: Add a note that virtiofs is not migratable
Note that certain operations will not work.

Resolves: https://gitlab.com/libvirt/libvirt/-/issues/452
Signed-off-by: Peter Krempa <pkrempa@redhat.com>
Reviewed-by: Martin Kletzander <mkletzan@redhat.com>
2023-03-01 11:57:37 +01:00

5.4 KiB

Sharing files with Virtiofs

Virtiofs

Virtiofs is a shared file system that lets virtual machines access a directory tree on the host. Unlike existing approaches, it is designed to offer local file system semantics and performance.

See https://virtio-fs.gitlab.io/

Note: virtiofs currently does not support migration so operations such as migration, save/managed-save, or snapshots with memory are not supported if a VM has a virtiofs filesystem connected.

Sharing a host directory with a guest

  1. Add the following domain XML elements to share the host directory /path with the guest

    <domain>
      ...
      <memoryBacking>
        <source type='memfd'/>
        <access mode='shared'/>
      </memoryBacking>
      ...
      <devices>
        ...
        <filesystem type='mount' accessmode='passthrough'>
          <driver type='virtiofs' queue='1024'/>
          <source dir='/path'/>
          <target dir='mount_tag'/>
        </filesystem>
        ...
      </devices>
    </domain>

    Don't forget the <memoryBacking> elements. They are necessary for the vhost-user connection with the virtiofsd daemon.

    Note that despite its name, the target dir is an arbitrary string called a mount tag that is used inside the guest to identify the shared file system to be mounted. It does not have to correspond to the desired mount point in the guest.

  2. Boot the guest and mount the filesystem

    guest# mount -t virtiofs mount_tag /mnt/mount/path

    Note: this requires virtiofs support in the guest kernel (Linux v5.4 or later)

Optional parameters

More optional elements can be specified

<filesystem type='mount' accessmode='passthrough'>
  <driver type='virtiofs' queue='1024'/>
  ...
  <binary path='/usr/libexec/virtiofsd' xattr='on'>
    <cache mode='always'/>
    <lock posix='on' flock='on'/>
  </binary>
</filesystem>

Externally-launched virtiofsd

Libvirtd can also connect the vhost-user-fs device to a virtiofsd daemon launched outside of libvirtd. In that case socket permissions, the mount tag and all the virtiofsd options are out of libvirtd's control and need to be set by the application running virtiofsd.

<filesystem type='mount'>
  <driver type='virtiofs' queue='1024'/>
  <source socket='/var/virtiofsd.sock'/>
  <target dir='tag'/>
</filesystem>

Other options for vhost-user memory setup

The following information is necessary if you are using older versions of QEMU and libvirt or have special memory backend requirements.

Almost all virtio devices (all that use virtqueues) require access to at least certain portions of guest RAM (possibly policed by DMA). In case of virtiofsd, much like in case of other vhost-user (see https://www.qemu.org/docs/master/interop/vhost-user.html) virtio devices that are realized by an userspace process, this in practice means that QEMU needs to allocate the backing memory for all the guest RAM as shared memory. As of QEMU 4.2, it is possible to explicitly specify a memory backend when specifying the NUMA topology. This method is however only viable for machine types that do support NUMA. As of QEMU 5.0.0 and libvirt 6.9.0, it is possible to specify the memory backend without NUMA (using the so called memobject interface).

  1. Set up the memory backend

    • Use memfd memory

      No host setup is required when using the Linux memfd memory backend.

    • Use file-backed memory

      Configure the directory where the files backing the memory will be stored with the memory_backing_dir option in /etc/libvirt/qemu.conf

      # This directory is used for memoryBacking source if configured as file.
      # NOTE: big files will be stored here
      memory_backing_dir = "/dev/shm/"
    • Use hugepage-backed memory

      Make sure there are enough huge pages allocated for the requested guest memory. For example, for one guest with 2 GiB of RAM backed by 2 MiB hugepages:

      # virsh allocpages 2M 1024
  2. Specify the NUMA topology (this step is only required for the NUMA case)

    in the domain XML of the guest. For the simplest one-node topology for a guest with 2GiB of RAM and 8 vCPUs:

    <domain>
      ...
      <cpu ...>
        <numa>
          <cell id='0' cpus='0-7' memory='2' unit='GiB' memAccess='shared'/>
        </numa>
      </cpu>
     ...
    </domain>

    Note that the CPU element might already be specified and only one is allowed.

  3. Specify the memory backend

    One of the following:

    • memfd memory

      <domain>
        ...
        <memoryBacking>
          <source type='memfd'/>
          <access mode='shared'/>
        </memoryBacking>
        ...
      </domain>
    • File-backed memory

      <domain>
        ...
        <memoryBacking>
          <access mode='shared'/>
        </memoryBacking>
        ...
      </domain>

      This will create a file in the directory specified in qemu.conf

    • Hugepage-backed memory

      <domain>
        ...
        <memoryBacking>
          <hugepages>
            <page size='2' unit='M'/>
          </hugepages>
          <access mode='shared'/>
        </memoryBacking>
        ...
      </domain>