Re: [PATCH 1/3] qemu: implementation of transient option for qcow2 file

On Thu, Jul 16, 2020 at 20:55:29 -0400, Masayoshi Mizuma wrote: > Thank you for your review. > > On Tue, Jul 07, 2020 at 06:36:23AM -0500, Eric Blake wrote: > > On 7/7/20 2:12 AM, Peter Krempa wrote: > > > > > You can install a qcow2 overlay on top of a raw file too. IMO the > > > implications of using <transient/> allow that. > > > > > > As said above I'd strongly prefer if the overlay is created in qemu > > > using the blockdev-create blockjob (there is already infrastructure in > > > libvirt to achieve that). > > > > Agreed. At this point, any time we call out to qemu-img as a separate > > process, we are probably doing it wrong. > > Got it. I'm thinking about the procedure such as followings. > Does that make sense? > > 1) Open the monitor with qemuProcessQMPNew()/qemuProcessQMPStart(), > and connect it. Starting a new qemu process just to format an image is extreme overkill and definitely not what we want to do.

> 2) Setup the transient disk with qemuDomainPrepareStorageSourceBlockdev(), > qemuBlockStorageSourceAttachApplyStorage(), qemuBlockStorageSourceCreateGetFormatProps() > and something... > > 3) Run blockdev-create command with qemuMonitorBlockdevCreate(), then > close the monitor. These two steps should be exectued in the qemu process which already will run the VM prior to starting the guest CPUs. > 4) Switch the original disk to the transient disk. > > 5) Build the blockdev argument for qemu. And instead of this step, you use the external snapshot infrastructure to install the overlays via 'blockdev-snapshot' QMP command

> > 6) Run qemu And instead of this the VM cpus will be started.

The above steps require factoring out snapshot code a bit. I have a few patches in that direction so I'll try posting them next week hopefully.