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

Re: [libvirt] Extensions to the libvirt Storage API



On 07/28/2010 01:24 PM, Daniel P. Berrange wrote:
On Wed, Jul 28, 2010 at 12:24:53PM -0500, Patrick Dignan wrote:
On 07/28/2010 05:08 AM, Daniel P. Berrange wrote:
On Tue, Jul 27, 2010 at 06:28:01PM -0500, Patrick Dignan wrote:

Hi Everyone,

I'm looking at implementing some functionality in libvirt that would
allow it to call functions in an unpublished iSCSI library.  Some of the
functionality I wish to implement is not currently part of the libvirt
storage API.  I wanted to suggest the following additions to the storage
API: grow volumes, show whether thin provisioning is enabled, enable
thin provisioning, disable thin provisioning, create snapshots, and
delete snapshots.  I've added a patch at the end of the mail showing how
I think these functions should be implemented.  Note that I have not
included details about the virStorageSnapshotDefPtr yet, that's the next
step.

Perhaps this should be in a separate mail for better threading, but it
seems a bit strange to me that the storage interface isn't pluggable in
the traditional sense.  In order to add a backend to libvirt, one has to
make modifications all over the place, for example: virt-inst, the
Makefile.am, the configure.ac, storage_backend.h, and several other
places.  It would make sense to me to make this pluggable such that
someone could just load in a library that implements the required
functions and some identifying information (eg type of storage,
description, etc).  A list of supported backends could be stored in
empty files in a directory somewhere, or some similar hack.  This way
someone could write a plugin for tgtd for example, or in my case the
library I'm working with.  I think this would also help others with
writing plugins for more storage backends.  How difficult do you think
this would be?  I'm willing to do a reasonable amount of work to get
this implemented, but I want to know what the experts think!

We explicitly don't support external driver plugins in libvirt for a
couple of reasons

  - We don't want to support use of closed source plugins
  - We don't want to guarentee stability of any aspect of
    libvirt's internal API

We would like to see support for the various vendor specific iSCSI
extensions to allow volume creation/deletion, but want that code to
be part of the libvirt codebase.


Understandable.  I was thinking that there is currently no way to
specify a vendor of a storage backend.  For example, an iSCSI vendor.
This makes it look like implementing vendor-specific extensions requires
creating a new backend, even though there's already an iSCSI backend.
It seems like a secondary field for vendor, and maybe even model, could
help this.
Yes, we'd want to add some kind of vendor and/or model tag to the
storage pool XML description, to indicate what variant of iSCSI
is to be used.

Makes sense to me!  maybe something like:

<device vendor="vendorA" model="modelT" path="demo-target" />

In the pool source element would make sense.
  /* File creation/cloning functions used for cloning between backends */
  int virStorageBackendCreateRaw(virConnectPtr conn,
@@ -76,6 +83,12 @@
      virStorageBackendCreateVol createVol;
      virStorageBackendRefreshVol refreshVol;
      virStorageBackendDeleteVol deleteVol;
+    virStorageBackendGrowVol growVol;

I'd call this 'resizeVol' since there's no reason we can't also support
shrinking.

Do all backends support shrinking?  I was under the impression shrinking
is not quite a universal feature, so it made sense to me to break this
out.  If most backends support shrinking, it makes sense to use
resizeVol instead then.
Plain files, LVM, and disk partitions can all be shrunk, so I
think this is fine. If a particular storage type does not support
shrinking then it can raise  VIR_ERR_NO_SUPPORT if the users tries
to shrink.

Alright sounds good, I'll change the patch to reflect that.
+    virStorageBackendThinProvisionShow thinProvisionShow;
+    virStorageBackendThinProvisionEnable thinProvisionEnable;
+    virStorageBackendThinProvisionDisable thinProvisionDisable;

I'm not really liking this as a concept. The other storage drivers, and
indeed my iSCSI server, deal with thin provisioning on a per-volume basis
when creating the volume. The libvirt model is that if in the XML, then
<allocation>   value is zero then the user is requesting thin provisioning
of that volume. ie no storage allocated for it upfront. If<allocation>
matches<capacity>   then the volume should be fully allocated upfront.


Sorry, that was a bit of a misnomer on my part, these functions are
intended to be used at the volume level.  However, what if they want to
enable thin provisioning or disable it after the fact?  Is that not
going to be supported?  An example use case of enabling after the fact
is if they want to enable it, and then grow the volume to a larger
size.  Disabling, of course, could be done at any time (eg for a speed
increase).  Do you object to the thinProvisionShow function, in either case?
I think this can all still be done using a combination of allocation
and capacity.  Some examples

  * create new volume, thin provisioning

       <allocation>0</allocation>
       <capacity>10000000</capacity>

  * create new volume, fat provisioning

       <allocation>10000000</allocation>
       <capacity>10000000</capacity>

The virStorageVolResize() API would probably want to take the same
XML format. So to extend a volume with thin provisioning

       <allocation>10000000</allocation>
       <capacity>50000000</capacity>

Or to extend with fat provisioning

       <allocation>10000000</allocation>
       <capacity>50000000</capacity>


Finally if you want to query current volume status the XML
config will show this, as will virStorageVolGetInfo.

So basically the virStorageVolResize() should handle thin-provisioning changes? What about situations where the storage is already allocated, but there is free space, and the user wants to de-allocate that in favor of thin provisioning (or vice-versa)? It seems like a bit of an edge case, but I can see it happening.
+    virStorageBackendCreateSnapshot createSnapshot;
+    virStorageBackendDeleteSnapshot deleteSnapshot;

There's no need for snapshot APIs. This functionality is already supported
via the normal volume creation API, just specify the original volume to be
snapshotted in the XML as the backing store.

I wasn't aware of this functionality.  It looks like it's implemented on
a per-hypervisor basis.  It'd be really cool to get snapshotting
integrated into storage backends with snapshotting support, so that
snapshots would show up in both libvirt and the storage backend's UI,
but I can see how this would be nearly impossible.
Namespace clash ! The virDomainSnapshot APIs are per-hypervisor. They
do snapshotting of the guest VM (including its storage).

I was actually just talking about the storage backends though which
can do snapshots independently of any hypervisor. See the<backingstorage>
element here:

   http://libvirt.org/formatstorage.html#StorageVolBacking

This is already implemented with the LVM pool doing LVM snapshots. We
also use it for external qcow2 backing files.

I'm not quite sure how this works, specifically with regards to multiple snapshots. If a VM is running with one backing store, and the user wants to do a snapshot of the current state, what happens? Is the backing store swapped to the current state?

Thanks for the help!

Best,

Patrick Dignan


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