FW: [openib-general] Minutes from DAPL BOF at OpenIB Workshop

[Christoph Hellwig]

On Sun, Feb 13, 2005 at 03:25:24PM +0100, Christoph Hellwig wrote:
> On Sun, Feb 13, 2005 at 10:37:12AM +0200, Dan Bar Dov wrote:
> > > I don't know who "we" here is, but we as the kernel developer 
> > > community certainly disagree.
> > 
> > So how do you describe scsi mid layer?
> > Scsi mid layer is exaclty a unifying API that hides the specific hardware device drivers from the upper layer abstractions.
> > The kdapl model is identical to the scsi model - a single in-kernel API that hides different hardware specific implementations of RDMA.
> 
> It's not really analogue.  In SCSI the hardware (or at least the drivers)
> all speak the same protocol (in a few different revisions), and the scsi
> midlayer and upper level drivers implement most of that protocol.  The
> LLDDs program the different hardware implementing this protocol.
> 
> See the SAM-2 or SAM-3 documents for details.

The right analogy for kDAPL in scsi land would be the CAM standard from
T10.org that tried to specifiy in-kernel interfaces for SCSI.  While a
few Operating Systems tried to model their stack afer CAM (Digital Unix,
FreeBSD), it soon became clear that the standard isn't implementable 1:1
in a real life system, was a hidrance in implementing clean and lean
driver and had no chance following the development of new SAM
substandards.