On 07/26/2011 06:51 AM, Corey Bryant wrote:
There are some additional features provided by certain image types
where Qemu reopens the image file. All of these scenarios will be
unsupported for the fd: protocol, at least for this patch:
- The -snapshot command line option
- The savevm monitor command
- The snapshot_blkdev monitor command
- Use of copy-on-write image files
- The -cdrom command line option
- The -drive command line option with media=cdrom
- The change monitor command
The thought is that this support can be added in the future, but is
not required for the initial fd: support.
Libvirt will eventually need support for fd passing on savevm,
snapshot_blkdev, and change monitor commands, as well as for -cdrom,
before this feature can be used to provide the desired security
enhancements. I agree that for an incremental patch, you don't have to
solve all points at once, but until all places have been modified to
support fd usage, you aren't gaining any security, except for severely
constrained guests.
Furthermore, how do you plan to map fd: to filename? There's already
been big threads on why snapshot_blkdev needs both the new fd: and the
name of the old backing file at the same time, so that qemu can write
the correct headers into new qcow2 files. But your proposal precludes
that, since "qemu -drive file=fd:4,format=qcow2" is not letting qemu
know the file name of fd:4 that would later have to be written into a
qcow2 header. I'm afraid that we need a better solution that gets both
fd and filename mapped together, before this stands a chance of being
useful. That said, I'm strongly in favor of getting the open() burden
moved out of qemu into libvirt, because of the potential it has for
increased security.
--
Eric Blake eblake(a)redhat.com +1-801-349-2682
Libvirt virtualization library
http://libvirt.org