[pve-devel] scsi-hd vs scsi-generic with iSCSI

[Thomas Lamprecht]

On 10/8/19 12:54 PM, Daniel Berteaud wrote:
> ----- Le 8 Oct 19, à 12:28, Thomas Lamprecht t.lamprecht at proxmox.com a écrit :
> 
>>
>> Thanks for the nice write up and clear reproducer!
>>
>> It seems that if we cannot use the same backend for all disks we need to
>> die when a disk move to a storage backend is request, and that move would
>> need to change the scsi "backend".
>> As I'd not like to die it would be better to see if there's still the need
>> for different backends.
> 
> Dying wouldn't be very nice indeed (I need to be able to move disks between NFS and ZFS over iSCSI on a regular basis) 
> 
> If scsi-hd was always selected, there would be no issue. I've patched my QemuServer.pm to do that for now.
> Not sure if scsi-generic/scsi-block has any advantages, but I couldn't measure performance diff in my case. unmap is also passed correctly with scsi-hd. IMHO, unless there are strong values with them (which I'am unaware off, but I couldn't find any documentation about all those backends), we should always use scsi-hd, as it's working with all storage types and allows live disk move from any storage type to any other, including the issue I have specific to ZFS over iSCSI (guest I/O error during live move from ZFS over iSCSI to something else)

I would actually really like to change this to scsi-hd, but we need to be sure
it's OK for all possible supported setups..

So I tried to investigate a bit how it came to the use of scsi-generic, I came
to a commit[0] from Alexandre (CCd) which adds support for access with libiscsi.

Maybe he knows why the -generic was used and not the -hd one?

[0]: https://git.proxmox.com/?p=qemu-server.git;a=commitdiff;h=d454d040338a6216c8d3e5cc9623d6223476cb5a

cheers,
Thomas