Re: [Iscsitarget-devel] Consistent snapshots on IET

[Pasi K. <pa...@ik...>]

On Sun, Apr 19, 2009 at 03:39:10PM +0300, Pasi Kärkkäinen wrote:
> On Wed, Apr 15, 2009 at 05:53:18PM -0400, Ross S. W. Walker wrote:
> > Pasi Kärkkäinen wrote:
> > > On Tue, Apr 14, 2009 at 06:55:49PM -0400, Ross S. W. Walker wrote:
> > > > On Apr 14, 2009, at 9:46 AM, "Mark Hunting" <ma...@ne...> wrote:
> > > > 
> > > > >Hi all,
> > > > >
> > > > >I use IET to export LVM (well, LVM2) volumes as iSCSI devices. On the
> > > > >initiator I use the iSCSI devices as disks for KVM (Kernel-based  
> > > > >Virtual
> > > > >Machine). I can create a snapshot of a LVM volume on the target just
> > > > >fine. But can I be 100% sure that the resulting snapshot contains
> > > > >consistent data? For example, when a database on the KVM is just  
> > > > >writing
> > > > >data when I make the snapshot, is there a change that the snapshot  
> > > > >will be broken/non-consistent?
> > > > >
> > > > >When using for example a local ext3 filesystem on a LVM volume the
> > > > >filesystem halts for a moment when the snapshot is created. Does IET
> > > > >also do this?
> > > > >
> > > > Actually this really has nothing to do with IET as IET simply acts as  
> > > > a disk drive and snapshots aren't done at the disk drive level.
> > > > 
> > > > You would need an LVM process to communicate somehow with the  
> > > > initiators to tell them, ok, it's snapshot time and that can be very  
> > > > tricky.
> > > > 
> > > > CLVM still doesn't support cluster snapshots to this day for this  
> > > > reason.
> > > > 
> > > > Another idea is another target or local disk on the initiator for  
> > > > snapshots and doing all snapshots on the initiator where applications  
> > > > can get notification to quiesce disk activity. Local disks for  
> > > > snapshots would reduce network utilization and probably increase  
> > > > throughput. Having these on the target side would slow operations for  
> > > > all targets in the volume group due to the overhead of the extra COW  
> > > > operations.
> > > > 
> > > 
> > > Hello!
> > > 
> > > Ross: Are you aware of any Windows VSS style projects for Linux? 
> > 
> > No unfortunately not. I don't think Linux would actually use VSS
> > specification because I don't think it is GPL'd. Somebody could
> > write a userland daemon though that would interface to Linux's
> > "whatever" library for LVM events and send those to Windows
> > hosts using Microsoft's VDS specification. Then the VSS writers
> > on those hosts could quiesce there services (File, SQL, Exchange).
> > 
> 
> OK. Yeah, I didn't meant to implement exactly VSS, but something like that..
> 
> > Those individuals who write a VDS daemon could also conceivably
> > have it listen for RPC/IPC commands from the Windows hosts per
> > the VDS specification and act on those with backend scripts.
> > 
> > Unfortunately you need both Linux daemons and Windows COM objects
> > for this to work.
> > 
> 
> Hmm.. is VDS requires for VSS? 
> 
> If I understood you correctly I was thinking of something different..
> I was mostly thinking of the following scenario:
> 
> - Backup agent on the Linux server starts a backup, as requested by the backup server
> - Backup agent on the Linux server calls VSS-style library/API to prepare for snapshot
> - VSS-style library signals all the 'registered' applications on the Linux server to prepare for snapshot.. flush their buffers etc. 
> - VSS-style library flushes/freezes the filesystem
> - VSS-style library calls the ietd-specific SAN-plugin to take a snapshot on the SAN (target)
> - when the snapshot is taken, VSS-style library signals the applications they can continue running normally
> - Backup agent (or server) continues the backup from the snapshot
> 

Of course it would be nice to have VSS/VDS support for ietd target in 
Windows initiators aswell :)

But yeah, I was mostly thinking of Linux initiators in this discussion.. 

-- Pasi