Re: [linux-lvm] Snapshots

[Mike Snitzer <snitzer@redhat.com>]

On Thu, Feb 16 2012 at 11:46pm -0500,
Stuart D. Gathman <stuart@bmsi.com> wrote:

> Long ago, Nostradamus foresaw that on Feb 16, Mike Snitzer would write:
> 
> >>An elegant part of the LVM system is that the device mapper kernel support is
> >>very general, and new data structures can be experimented with entirely in user
> >>code - with a script language even.  Metadata for experimental structures does
> >>not have to stored with the main metadata.
> >
> >Please note that the dm-thinp code has metadata in the kernel (on-disk
> >format for btrees, etc) much like a filesystem would have.  So there is
> >both kernel and userspace (lvm2) metadata for dm-thinp.
> 
> Yes, but isn't this loaded into the kernel via userland tools like
> device-mapper?

The 'thin-pool' and 'thin' DM device tables are loaded from userspace
via DM interfaces.

The kernel metadata isn't loaded from userspace.  It is created and/or
changed by certain actions taken from userspace (via DM messages).

> So while a kernel feature would be required for a new
> type of kernel metadata, experimental uses of existing formats can
> be done in userland.

The kernel manages the kernel's metadata.  But the LVM metadata that
userspace uses to coordinate and manage the thin devices can be
changed independently.

The thin-provisioning-tools know the kernel's metadata format and can
check it (and in the future repair it).