The trouble I would expect is when you have a front end that allows
external users to create snapshots as long as they're not exceeding a
quota, at which point they would have to delete some in order to create
more. Then 1000 snapshots of 10MB each (10GB total) might be well within
the quota, meanwhile deletion is permitted when over quota because it's
the expected means to free up space. Which makes it a concern if
deletion can turn each pair of 10MB existing snapshots into 1TB of
zeroes. But I'm after getting it fixed rather than classifying it.
There might also be two bugs here. I put them together because they're
related and could potentially be fixed together, but at least the lesser
one might still be in libvirt.
The main one is in the interaction between libvirt and qemu. In trixie,
qemu doesn't do detect_zeroes during blockcommit unless libvirt tells it
to and libvirt frequently doesn't tell it to even when
detect_zeroes='unmap' is set in the XML. That could be fixed in either
place by either having libvirt tell it to or having qemu do it without
needing to be told. I don't know of any reason not to fix it in qemu
instead and if that's what others did then fair enough. I'll see if I
can find what's different in qemu in sid.
Here's the other one:
# virsh qemu-monitor-command trogdor --pretty
'{"execute":"query-block"}'|grep detect_zeroes
"detect_zeroes": "unmap",
# virsh snapshot-create-as --domain trogdor snap42 --diskspec
vda,file=/var/lib/libvirt/images/trogdor-snap42.qcow2 --disk-only
Domain snapshot snap42 created
# virsh qemu-monitor-command trogdor --pretty
'{"execute":"query-block"}'|grep detect_zeroes
"detect_zeroes": "off",
This one isn't an issue for snapshot deletion or blockcommit. But if you
actually wanted detect_zeroes='unmap' in the active top image, creating
a snapshot turns it off until you shut down the VM because
qemuDomainPrepareDiskSourceData() isn't properly copying the
detect_zeroes property to the new top image during snapshot creation. In
theory qemu could also deduce that one without being told if it gets a
new top image when detect_zeroes was enabled for the previous one, but
it isn't doing that even in sid and even if it was that would still
leave the "detect_zeroes" field in "struct _virStorageSource" set to the
wrong value for the new top image in libvirt.
I suppose that one wants to be a separate bug at this point, doesn't it?
