Device-mapper filesystem interface

Device-mapper filesystem interface

[Joe Thornber]

I thought I'd kick off a thread concerning the filesystem interface
for device-mapper after it came up on last nights 'must-fix' meeting.

To recap:

Alasdair Kergon and I spent a lot of time thinking last autumn about
how to best map the dm semantics onto an fs.  The end result was this
very rough and ready patchset:

http://people.sistina.com/~thornber/patches/2.5-unstable/2.5.51/2.5.51-dmfs-1.tar.bz2

The reception was not favourable.  People didn't like the way creating
a directory was analagous to creating a device, or the fact that these
device directories were pre-populated with table, status and
dependency files.  Gregkh was the only person who put forward
alternatives ideas (sysfs), and I don't think even he had thought
through how all of the dm functionality was going to be mapped.  eg,
with dmfs as it stands the 'wait for event' ioctl has translated into
a poll on the status file, ie wait until the status file changes - I
think this is neat.

I must admit I rather let the issue die; having played with these fs
ideas I do not see any particular advantage over the ioctl interface.
It will involve more code on the kernel side (which I will need help
with since I know v. little about the vfs), and more code on the
userland side.  I can't even argue that the fs interface makes it
easier for scripting languages to use dm, since the simple dmsetup
tool will always be far simpler to use than poking about in the fs.

I've always been careful to keep core dm seperate from the interface,
interfaces can be seperate modules, and multiple interfaces can be
present at the same time - they are just clients of the core dm code.
So let's treat dmfs as a seperate project.  I'm happy to work on it,
but I don't have the enthusiasm to drive it, especially after the luke
warm response to my initial attempts.

- Joe

Re: Device-mapper filesystem interface

[Greg KH]

On Thu, May 22, 2003 at 09:50:36AM +0100, Joe Thornber wrote:
> I thought I'd kick off a thread concerning the filesystem interface
> for device-mapper after it came up on last nights 'must-fix' meeting.
> 
> To recap:
> 
> Alasdair Kergon and I spent a lot of time thinking last autumn about
> how to best map the dm semantics onto an fs.  The end result was this
> very rough and ready patchset:
> 
> http://people.sistina.com/~thornber/patches/2.5-unstable/2.5.51/2.5.51-dmfs-1.tar.bz2
> 
> The reception was not favourable.  People didn't like the way creating
> a directory was analagous to creating a device, or the fact that these
> device directories were pre-populated with table, status and
> dependency files.  Gregkh was the only person who put forward
> alternatives ideas (sysfs), and I don't think even he had thought
> through how all of the dm functionality was going to be mapped.  eg,
> with dmfs as it stands the 'wait for event' ioctl has translated into
> a poll on the status file, ie wait until the status file changes - I
> think this is neat.

Yeah, I went down the sysfs path for a while, then got distracted by
other issues (driver core, etc.)  If you need this feature, then yes, a
dmfs does make sense to have, and not use sysfs.

I'm not opposed to your implementation, it was just a bit strange at
first glance.  Care to update your old patch for this?

thanks,

greg k-h