On Fri, Feb 17, 2017 at 3:27 AM, Christoph Hellwig <h...@lst.de> wrote: > > On Fri, Feb 17, 2017 at 09:06:14AM +0100, Hannes Reinecke wrote: > > We could, but why? > > ATM we're only having SCSI devices able to use device handler; adding > > another layer of indirection doesn't solve anything here. > > Moving the infrastructure one level up will only make sense if we're > > getting non-SCSI device handler (ANA?), but until then I'd think it's > > just overengineering. > > Agreed. Independent of what does the balancing between queues hardware > handler should be attached by the low-level driver for any future > transport without any control from DM.
But doesn't Keith's abstraction makes a lot of sense given that you're providing a device handler interface for NVMe? The most important part of scsi_dh that DM uses is its calls to scsi_dh_activate. Attachment isn't interesting or the issue (DM's call to scsi_dh_attach is purely legacy now that SCSI attaches the proper scsi_dh during SCSI's device scan).
