On Mon, 2008-06-02 at 14:11 -0400, Jeff Garzik wrote: > James Bottomley wrote: > > The quickest way seems to be to break the scsi_host <-> ata_port link by > > mapping scsi_host <-> ata_host instead. > > Correct for newer controllers... not desirable for older master/slave > > > > Legacy controllers with only a > > single port can keep the apparent 1:1 mapping (we can even keep the > > hostdata stuff). > > Not sure what you mean by this I think it's possible to do the shift while still making it look to single port hosts that nothing has changed > > Unfortunately, the standard way of doing this is via > > the transport classes, but as long as you have a pointer from the port > > to the host and from the device to the port (which you do) it should be > > possible. > > I've always been open to an ATA transport class and thought that was the > best way to go long term. But I confess to not knowing the best way to > implement that goal in concrete terms. Perhaps it is best to approach it from what libsas needs, that way the structures can be made to map to each other. > > The downside I can see is that qc_defer handling changes non trivially > > because of this, but there don't seem to be many other issues. > > There are a lot of little details that are easily fixable. > > The main detail is not breaking queueing for master/slave. > Master/slave, if you recall, _requires_ one-shost-per-port because it > relies on scsi queueing to handle the balancing between master/slave. > We set can_queue to 1, and for the 2-device case -- master and slave > present -- SCSI does the heavy lifting to ensure proper arbitration. Well, we can be elastic on this. The common case is a two port master/slave, isn't it. This particular one could just be treated as two separate interfaces and hence two hosts (it's reasonably analagous to the 53c896---which is a single chip that presents two SCSI channels as two PCI functions). > Adding support for NCQ and qc_defer made it a bit easier to change the > master/slave setup, though. > > The other reason why we use one-short-per-port on master/slave is that, > in many respects, each legacy IDE port really can behave like a > completely separate controller, each port with its own interrupts and > independent reset logic. > > Thus ATA really has two models: 1-port-1-host and N-ports-1-host. > > [meta: I'm not disagreeing with you here, just explaining how all this > came about...] OK ... so how bad would it be to maintain two attachment models for the two cases? The architect in me says yuk because special casing like this always leads to complexity, but I can't see a simpler way of doing it. James -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
