Re: [PATCH] Device-mapper submission 6/7

[Jeff Garzik <jgarzik@pobox.com>]

Joe Thornber wrote:
> On Tue, Oct 15, 2002 at 02:15:35PM -0400, Jeff Garzik wrote:
> 
>>>[Device mapper]
>>>Provide a traditional ioctl based interface to control device-mapper
>>
>>>from userland.
>>
>>
>>If you're adding a new interface, there should be no need to add new
>>ioctls and all that they entail.  Just control via a ramfs-based fs...
> 
> 
> We originally did have a fs based interface written by Steve
> Whitehouse.  However at the time (about a year ago) it wasn't obvious
> that everyone would think it a good idea.  Also the code was
> significantly larger than the ioctl interface.  I would be more than
> happy to do away with the ioctl stuff if people are now in full
> agreement that an fs interface is the way to go.


Which people didn't like it?  ;-)

AFAIK Linus and Al Viro (and myself <g>) have always considered ioctls 
an ugly -ism that should have never made it into Unix.  Over and above 
the Unix/VFS design problems with ioctl(2), ioctl(2) is a pain for 
people like David Miller who must maintain 32<->64 bit ioctl translation 
layers for their architecture.  ia64 and x64-64 must do this too.  Each 
ioctl you add is an additional headache for them.

We now have libfs.c in 2.5.x that makes ramfs-based filesystems even 
more tiny, too.  With the added flexibility of an fs -- it makes the 
userland tools much more simple and sane -- and the pain of ioctls, it 
seems a clear choice for new interfaces.

	Jeff