On Thu, Jan 13, 2011 at 16:59, Karel Zak [off-list ref] wrote:

On Thu, Jan 13, 2011 at 03:43:38PM +0100, Kay Sievers wrote:

quoted

On Thu, Jan 13, 2011 at 15:30, Tejun Heo [off-list ref] wrote:

quoted

On Thu, Jan 13, 2011 at 3:25 PM, Milan Broz [off-list ref] wrote:

quoted

Maybe, but this was not invented in DM/MD camp:-)
Probably Kay or Greg can answer why it was done this way?

It's not from Greg or Kay. It just appeared some day in the context of dm. :)

And yes, symlinks *look* nice and simple for the outside, but they are
not, and have all sorts of problems like non-atomic updates, make it

 Sounds like sysfs implementation problem, right?

It's a normal multi-file problem. It can by-definition not be atomic
without doing really weird locking things.

 If there is noway to fix sysfs then we can add a generic ioctl or
 /sys/block/<device>/{slave,holder}_list files with list of
 holders/slaves.

Yeah, we've been there with the btrfs problem. For btrfs it woud
probably need to be something statfs()-like.

 But please, don't force userspace to use *claimer-specific*
 methods to answer *generic questions* like slave/holder dependencies
 between devices.

The links exist only for dm and md so far, I think. It's the classical
multiple-parents-in-a-tree problem. We have that for bonded network
devices and some IO buses too. There is no nice representation for
these reversed-trees-in-the-tree so far.

Kay