Low overhead low-level method of sharing filesystems

[QLP24]

Hi,
Wondering what methods there are of sharing a filesystem from a host to a guest in Proxmox. I looked into 9p and found a forum thread from 2012 where dietmar says "[...]We do not want to implement that now."

Long story short, I have a 10TB XFS filesystem that I need to access read-only from a host, but read-write from within a guest.
NFS, CIFS and the like are not an option dude to the overhead of going through a network stack from host to guest.

 

[dietmar]

You can use containers with bind mounts.

 

[QLP24]

I need this in a KVM machine

 

[tom]

did you tested the "overhead"?

I assume a virtio net connection from the host to the guest is blazing fast.

 

[QLP24]

@tom:
Yeah, I have a virtio network card on the guest, and I'm currently using NFS. It's unneccesary overhead, and is stripping my POSIX ACL's.

So far my idea is to make the logical volume visible to the host and guest, then mount the XFS filesystem in the guest and host at the same time, but use "ro,norecovery" on the host. This would mount it read-only and skip running log recovery. There's no documentation on whether this is safe or not, but here's some discussion: http://oss.sgi.com/archives/xfs/2011-01/msg00038.html.

 

[mir]

Another approach could be to use iscsi on the promox host and make an either block based or a file based Lun which is exposed to the VM.

 

[John Radley]

@mir
I have this question too. Can you say more about using a "file based LUN"
Open-iSCSI appears to be running on my Proxmox 4.3, but I'm sure that is the Initiator (client).
What should I use to make the Proxmox server provide a iSCSI target?

 

[mir]

It seems iscsi target is changing in Debian. Current status for Debian Jessie, which proxmox is based on, is iscsitarget but the problem is that this will not and will never compile with the current kernel in proxmox since the package has officially been removed from Debian as of Debian stretch. Your best choice will therefore be to use the official in-kernel linux iscsitarget which is LIO (http://www.linux-iscsi.org/wiki/Main_Page) bad thing about this is that usertools are only available in Debian stretch and onward so you will have to backport these tools yourself.

 

[QLP24]

@dietmar what's preventing you guys from implementing 9p virtio? Would be great to have this feature.

 

[mir]

@dietmar what's preventing you guys from implementing 9p virtio? Would be great to have this feature.

Could be interesting but remember using this will prevent you from migrating the VM.

 

[dietmar]

@dietmar what's preventing you guys from implementing 9p virtio?

No spare time. But yes, this would be interesting. I would also accept patches...

 

[QLP24]

No spare time. But yes, this would be interesting. I would also accept patches...

I see. Anyway, consider it for a future release! Would be great

 

[John Radley]

@mir, thank you.
I was able to install "targetcli_2.1.fb43-1_all.deb" and its dependencies, and create a iSCSI target.

 

[mir]

For I good intro watch this from Redhat kernel developer Andy Grover on youtube: