[Qemu-devel] qemu-img convert corruption (raw to raw)

[Dominik Csapak <d.csapak@proxmox.com>]

Hi,

since i got no answer on the qemu-discuss mailing list, i post it here
maybe someone can help me or point me in the right direction
(copy-pasted from my previous mail)

Hi,

I have a strange issue with qemu-img convert which corrupts my images, 
and found nothing relevant online.
Maybe someone here can help me.

My problem:

I have a Windows 10 VM with a hd image file of type RAW with cache mode 
writeback.
I boot it up, write a bunch of data (a few gigabytes with crystal disk 
mark) and shutdown.
As soon as the qemu/kvm process is gone i do a
"qemu-img convert -p -f raw -O raw vmdisk.raw vmdisk2.raw"
When it finishes i do a compare:
"cmp vmdisk.raw vmidsk2.raw"
The output is:
"vmdisk.raw vmdisk2.raw differ: byte yyyyy, line zzzzz"
this it tells me that the files are different.
this manifests by the fact that sometimes in windows on the new disk i 
get filesystem errors or damaged files.
(depending on where the writes were)

now the strange part:
it does not happen with linux guests
it does not happen when i do a "echo 3 > /proc/sys/vm/drop_caches" 
before the qemu-img command
it does not happen when i have the "-t none" or "-t writethrough" flags 
on my qemu-img command
it DOES happen when i do a sync before the qemu-img command
the "-T" parameter makes no difference
it does also not happen when i cp the file
also when i wait a few minutes before the qemu-img command it works 
correctly

now the real questions:
is writeback with windows not recommended (i would assume after a guest 
shutdown it would be safe)
what do the -t and -T flags on qemu-img convert and why does it bypass 
the host page cache (apparently?)
why does it work when i specify no caching on the target
(this baffles me, since it would seem that the source cache is the 
culprit, not the target)

i apologize for the long message
and hope someone can explain this to me

with kind regards
Dominik Csapak