On 03/06/2012 10:56 AM, Paolo Bonzini wrote: > With it comes a new image creation mode, "no-backing-file", that can > be used to stream an image so that the destination does not need the > original image's backing file(s). > > Both bdrv_append and blkmirror will set the backing_hd on the target, > even if the image is created without one, so that both streaming and > copy-on-write work properly (at least with qcow2 or qed, not raw). > > Streaming mode works with the following gotchas: > > - streaming will rewrite every bit of the source image; > > - zero writes are not supported by the blkmirror driver, hence both > the source and the destination image will grow to full size. >
> +++ b/qapi-schema.json
> @@ -1133,10 +1133,12 @@
> # @absolute-paths: QEMU should create a new image with absolute paths
> # for the backing file.
> #
> +# @no-backing-file: QEMU should create a new image with no backing file.
> +#
> # Since: 1.1
> ##
> { 'enum': 'NewImageMode'
> - 'data': [ 'existing', 'absolute-paths' ] }
> + 'data': [ 'existing', 'absolute-paths', 'no-backing-file' ] }
Offline, Paolo and I were discussing this. It may make more sense to
_not_ expose a 'no-backing-file' mode, but instead...
> +# @BlockdevMirror
> +#
> +# @device: the name of the device to start mirroring.
> +#
> +# @target: the image that will start receiving writes for @device. A new
> +# file will be created unless @mode is "existing".
> +#
> +# @format: #optional the format of the target image, default is 'qcow2'.
> +#
> +# @mode: #optional whether and how QEMU should create a new image, default is
> +# 'absolute-paths'.
> +##
> +{ 'type': 'BlockdevMirror',
> + 'data': { 'device': 'str', 'target': 'str', '*format': 'str',
> + '*mode': 'NewImageMode' } }
have either a mandatory 'full':'bool' or an optional '*base':'str' for
choosing how much of the backing chain to be mirrored, as is the case
with the block_stream command. That is, the ability to supply an
external file ('mode':'existing') should not imply a shallow copy.
Having an optional base (where omitting the base gives a full copy) is
the most flexible, because then I can start with:
base <- snap1 <- snap2 <- snap3
as well as use qemu-img to create an empty file backed by snap1, then issue:
{ 'execute': 'drive-mirror', 'arguments':
{ 'device': 'ide0-hd0', 'target': 'copy', 'base': 'snap1', 'mode':
'existing' } }
and get the following setup after a drive-reopen:
base <- snap1 <- copy
without having to do an extra block_stream.
--
Eric Blake ebl...@redhat.com +1-919-301-3266
Libvirt virtualization library http://libvirt.org
signature.asc
Description: OpenPGP digital signature
