[pve-devel] [RFC guest-common 09/13] vzdump: schema: add fleecing property string

[Fiona Ebner]

Am 01.02.24 um 13:39 schrieb DERUMIER, Alexandre:
>>> LVM and non-sparse ZFS need enough space for a copy for the full disk
>>> up-front, so are not suitable as fleecing storages in many cases.
> 
> can't we force sparse for theses fleecing volumes, even if the storage
> don't have sparse enabled ? (I can understand that it could make sense
> for user to have non sparse for production for performance or
> allocation reservation, but for fleecing image, it should be
> exceptionnal to rewrite a full image)
> 

For ZFS, we could always allocate fleecing images sparsely, but would
require a change to the storage API as you can't tell vdisk_alloc() to
do that right now. There could also be a new helper altogether,
allocate_fleecing_image() then the storage plugin itself could decide
what the best settings are.

>>> Should the setting rather be VM-specific than backup job-specific?
>>> These issues
>>> mostly defeat the purpose of the default here.
> 
> can't we forbidden it in storage plugin features ? { fleecing => 1} ?
> 

There is no feature list for storage plugins right now, just
volume_has_feature() and that doesn't help if don't already have a volume.

There is storage_can_replicate() and we could either switch to a common
helper for storage features and deprecate the old or simply add a
storage_supports_fleecing() helper.

But the question remains if the setting should be VM-specific or
job-wide. Most flexible would be both, but I'd rather not overcomplicate
things. Maybe my idea for the default with "use same storage for
fleecing" is not actually a good one and having a dedicated storage for
fleecing is better. Then it needs to be a conscious decision.

>>> IIRC older version of NFS lack the ability to discard. While not
>>> quite
>>> as bad as the above, it's still far from ideal. Might also be worth
>>> trying to detect? Will add something to the docs in any case.
> 
> I never have seen working discard with nfs, I think (never tested) it's
> possible with 4.2, but 4.2 is really new on nas appliance (netapp,...).
> So I think than 90% of user don't have working discard with nfs.
> 

With NFS 4.2 discard works nicely even with raw format. But you might be
right about most users not having new enough versions. We discussed this
off-list too and an improvement would be to use qcow2, so the discards
could happen at least internally. The qcow2 could not free the allocated
blocks, but re-use already allocated ones.

> Is it a problem if the vm main storage support discard , but not
> fleecing storage ? (I don't have looked yet how exactly fleecing is
> working)
> 

It doesn't matter if the main storage does or not. It only depends on
the fleecing storage.

> If it's a problem, I think we should forbind to use a fleecing storage
> not supporting discard, if the vm have discard on 1 disk.
> 

The problem is that the space usage can be very high. It's not a
fundamental problem, you can still use fleecing on such storages if you
have enough space.

There are already ideas to have a limit setting, monitor the space usage
and abort when the limit is hit, but nothing concrete yet.