[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Libvir] Thoughts on remote storage support

OOps, my spam filter ate your reply, sorry for the delay.

Richard W.M. Jones wrote:
Tóth István wrote:

I had a hobby project where I needed to manipulate xen disk images on
remote systems, that used a model similar to libirt's remote support.
Based on what I learnt from it, I came up with a possible model for
libvirt's remote storage support. I present it here for discussion.

We typically store the images in volumes on LVM, or in dedicated file
system folders.
The folders and volume groups usable by libvirt can be limited in a
config file.

It is probably not neccessary to differentiate between defined and
created files, as you can not stop and start a file like a domain, you
either have it on disk, or not.


Libvirt should not store information on these files, everything should
be checked/listed on the fly, so that if you just copy an image to a
directory, libvirt can deduct all information (well, all it can) on it,
and handle it just as if the file was created by it.


The handle for the file is its path, plus its virConnect object (i.e.
the host it is on). For consistency, it may be possible to create an
object for it, but as disk images  have no persistent properties apart
from what is on the disk, and it can always be checked from there, it
provides no extra functionality.

Probably the 'handle' is its filename or device name + the virDomain object.

For example, here are the domains and their images running on my Xen host at the moment. I got this by writing a simple script which parses the domain XML:

        /var/lib/xen/images/fc6_0.img -> xvda
        /var/lib/xen/images/home.disk -> xvdb
        /var/lib/xen/images/fc6_1.img -> xvda
        /var/lib/xen/images/debian32fv.img -> hda
        /dev/Images/f764pv -> xvda
        /var/lib/xen/images/freebsd32fv.img -> hda
        [CD] -> hdc
        /var/lib/xen/images/gentoo32fv.img -> hda
Well, this is a good handle for the images that belong too an active domain.
But I can see other images laying around, backup images, snapshots, virgin installed images for provisioning of new VMs, and you need to refer to those as well. Hence, I still think that it would be better to use host+path. For example, you need to be able to say in effect: copy "/var/lib/xen/images/fc6_0.img" to "/backups/fc6_xvda_1", and you have to refer to target image somehow. You could just use the local path in this case, but I think that being able work with images on other libvirt hosts would be a bonus.

I think there is no need to support remote files explicitly, as the
domains mount local files/volumes. The file/volume may actually be
mounted from a NAS or SAN, of course, but it does not matter because we
use the local path names, and AFAIK all virtualization tools use local
files or local devices as blockdevs.

I have added compression to the mix because it is immensely useful.
I have used lzop in my project, and a full backup and restore was much
faster when using a compressed backup file, than with and uncrompessed
one. It conserves disk space, as well as cpu/bus capacity.

Zeroing out newly allocated files, helps with compressed backups, as
well as security. It also means that no holey files can be used.

The objects we are dealing with are disk images.
They have the following properties:
-Path: The unix path of the file ( /mnt/images/fc7.img or /dev/VG/fc7)
   -Compression: Mountable/compressed
   -Type: Plain file/LVM volume/ What else?
   -Filesystem: swap/ext3/xfs/....
   -Is it mounted?

This is where it gets very complicated. Files or partitions may represent simple filesystems, or partitioned block devices, or LVM PVs, or filesystems or partitions in formats that we have no chance of understanding (eg. NTFS), or snapshots, or dm_crypt, or compressed and so on and so on.

To do this in any feasible way, I'm sure we'll need a library, the obvious one being gparted (http://www.gnu.org/software/parted/). However parted has a lot of problems, and most specifically it doesn't support LVM.

I am aware of a project to make LVM accessible as a library and to include LVM library support in gparted/libparted, but I'm not sure how it is progressing.

Indeed, back when I created this specifications the supported mode for Xen was to feed it partitions instead of whole disk images, and the partitioning problem was not apparent. I checked the LVM library project about a moth ago, but it does not seem to be in a generally usable shape yet. The supported way is just to write, call and parse the command lines.

In fact, the more I thought about it, and the more scenarios popped into my mind (plus the ones you describe above), the more I think that at least an initial implementation should not try to see into the partition, exactly because of the problems you mention above. Even if partition/filesystem handling is included in libvirt, it should probably be somewhat orthogonal to the rest of the image handling functions.

i.e. the operations I detailed below (except for the growfs-related ones) for creating, moving, backing up, etc. of raw images, and a different set of operations that partitions, adds/removes paritions, creates file systems, growfs-es, etc.

This limits the complexity to just supporting simple files, block devices, and LVMs ( or the equivalent functionality on other platforms), and the parted-like functionality can be added on top of it.

We can do the following operations on the images:
[... operations ...]

I'd prefer to stick to the minimum set of operations needed now and add to them later. But yes, the mix of operations looks OK.

Much of this work is needed by virt-p2v and virt-df (http://et.redhat.com/~rjones/virt-p2v/ and virt-df to be released soon).
virt-p2v looks immensely useful.


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]