[dm-devel] multipath affinity to SCSI

[goggin, edward]

On Wed, 27 Apr 2005 19:11:33 +0200, Lars Marowsky-Bree wrote:
 
> > Has anyone tried multipathing to non-SCSI target devices, 
> e.g., LOOP,
> > NBD, USB?  While the architecture looks fully capable of supporting
> > such, it looks like both the kernel and user implementation 
> would have
> > some problems due to some SCSI affinity.
> 
> Well, I can see the user-space not auto-detecting the paths right now
> (probably because simply noone has added those features yet 
> ;-), but the
> kernel shouldn't have any issues with it. I've played with it 
> over loop
> devices for testing.
> 
> Which SCSI affinity are you referring to?

Kernel components look fine.  Mapped and target device concepts seem
abstract enough to encompass multiple underlying block device session
protocols.  I was wrong to implicate them.

As reported yesterday by Benjamin Marzinski, sysfs_devinfo()
definitely shows this affinity.  Look in libmultipath/structs.h
for more examples.  

I think the definition of a path may be too focused on SCSI.  Path
attributes and path discovery are clearly SCSI centric.  Possibly
the path data structure could use unions for the different
definitions for a path (address attributes for a path are surely
different across these protocols) in order to accommodate
different underlying protocols.  Another idea is to have per
protocol callouts from sysfs_devinfo to perform path discovery.

This isn't all that different (although its at the session layer
not the SCSI transport layer) than the need for transport specific
handlers if/when the path testing starts making use of transport
level attributes as pointed out by Mike Anderson yesterday.