[pve-devel] [RFC storage] work-around #6543: do not use preallocation for qcow2 on top of LVM

[Fiona Ebner]

Hi Alexandre,

Am 23.07.25 um 2:15 PM schrieb DERUMIER, Alexandre via pve-devel:
> Hi Fiona, I'm on holiday, can't verify, but before using qemu-img
> measure I had implemented compute of metadatas size (It was not 100%
> perfect).
> 
> This is strange, because I thinked that "qemu-img measure" was working
> correctly (we need to pass it blocksize && l2_extended option too). I
> had tried with a 1TB qcow2 volume.
> 
> 
> Note that I'm almost pretty sure that l2_extended=on need
> preallocation. (If I remember it's failing without it, and performance
> of backed volumes are pretty bad with l2_extended).

allocation seems to work fine:

Formatting '/dev/sharedlvm/vm-101-disk-0.qcow2', fmt=qcow2
cluster_size=131072 extended_l2=on compression_type=zlib size=1073741824
backing_file=snap_vm-101-disk-0_foobar.qcow2 backing_fmt=qcow2
lazy_refcounts=off refcount_bits=16

Do you mean for the performance benefits?

> it seem to be normal that preallocation option don't make a difference
> here. qemu-img measure always compute size of all metadatas.  (only
> cluster_size && l2_extended change the result)
> 
> 
> I don't understand why disabling preallocation could help here (until
> it's bugged), because the same amount of metadatas need to be allocated
> later on the fly  with disabled preallocation.

Yes, intuitively and telling from qemu-img measure, preallocation
shouldn't make a difference, but it seems to. The bug has not been fully
investigated, but I can clearly see qcow2 trying to write to an offset
beyond the LV size (which is even slightly larger than what qemu-img
measure reported, because it rounded up to full extent).

In practice, something is wrong, see the bug report. Friedrich and I
were both able to reproduce the issue, but only with
preallocation=metadata for now. I used a 4GiB disk on a ZFS storage that
first was filled with zeroes from within the VM (like in the bug
report). Then offline move it to qcow2-on-LVM and again fill it with
zeroes from within the VM. The error doesn't happen every time though,
still need to investigate further.

Best Regards,
Fiona