Re: [PATCH] qemu: blockjob: Avoid spurious log errors when cancelling a shallow copy with reused images
On Wed, Feb 23, 2022 at 14:17:40 +0100, Kashyap Chamarthy wrote:
On Tue, Feb 22, 2022 at 05:55:38PM +0100, Peter Krempa wrote:
> In case when a user starts a block copy operation with
> VIR_DOMAIN_BLOCK_COPY_SHALLOW and VIR_DOMAIN_BLOCK_COPY_REUSE_EXT and
> both the reused image and the original disk have a backing image libvirt
> specifically does not insert the backing image until after the job is
> asked to be completed via virBlockJobAbort with
> VIR_DOMAIN_BLOCK_JOB_ABORT_PIVOT.
Thanks for the quick patch!
First, to refresh my memory (and for others reading the thread), these
are the two main uses of combining '--reuse-external' and '--shallow'
flags with "blockcopy":
(a) You can control what the backing file is, as long as the top of that
backing file chain has identical guest-visible contents to what the
original chain had in its backing file.
(b) The new backing file can have a _different_ backing chain depth or
even different format.
* * *
Now to construct a reproducer with `virsh`. Peter, tell me what I got
wrong :-)
(1) Let's start with this image chain (overlays are always of qcow2
format):
base.raw <-- overlay1 <-- overlay2 (live QEMU).
With the goal of copying the above into the below chain (note, here
we're flattening both base.raw and overlay1 into a single file,
"flat-o-b1")
flat-o-b1.raw <-- copy (live QEMU)
(2) Make a *raw* variant of "base <-- overlay1", call it
"flat-b-o1.raw"
(i.e. flattened version of combined base and overlay1):
$ qemu-img convert -f qcow2 -O raw overlay1.qcow2 \
flat-b-o1.raw
(3) Then, create an empty QCOW2 file to create "flat-b-o1 <-- copy
(empty)":
$ qemu-img create -f qcow2 \
-o backing_file=flat-b-o1.raw,backing_fmt=raw copy.qcow2
(4) *Then* perform the `blockcopy --reuse-external --shallow`:
$ virsh blockcopy \
--domain vm1 vda ./copy.qcow2 \
--wait --verbose --reuse-external --shallow \
--finish
If you are working on a persistent domain/vm use --transient-job to
avoid the need to undefine it.
Also note that '--finish' is equivalent to aborting the job after it
reaches synchronized phase ...
(5) And then pivot the job, so that live QEMU now points to copy
$ virsh blockjob --pivot
... so this will actually say that there is no blockjob, because the
previous step actually cancelled it.
And the final result, we get the "goal" chain in step (1).
src: base <-- overlay1 <-- overlay2
== ==
dst: flat-o-b1 <-- copy (live QEMU)
Am I missing anything else?
Now the bug is that when you _cancel_ the job we attempt to blockdev-del
some images which were not blockdev-added, and thus create some spurious
log entries, but the rest of the code behaves correctly.
So you'll only see the failure this is fixing in the logs.