Re: [RFC/PARTIAL PATCH 0/3] dma: passing "attributes" to dma_map_* routines

[Roland Dreier <rdreier@cisco.com>]

 > > I think the case before us that Arthur is dealing with is a
 > > counterexample for this: there's nothing bus-specific about it all.
 > > The issue is related to reordering of DMAs within the Altix system
 > > fabric, after they've left the PCI world.  This issue would be present
 > > no matter what kind of host bridge you hook up to the system fabric,
 > > be it PCI-X, PCIe, ISA, MCA or whatever.

 > But it is: for performance reasons, the Altix boxes have a rather non
 > standard PCI bridge implementation that gives relaxed ordering on the
 > PCI bus.

I don't think this is accurate. As I understand things, the reordering
happens within the Altix system interconnect -- nothing to do with the
PCI bridge hanging off this fabric.  It is "platform" behavior and I
think is properly handled within the dma_ API, which exists to
abstract platforms.

 > This behaviour was later standardised to some degree in PCIe,
 > so you could argue they actually have an altix specific PCI bus (PCIa
 > anyone?).  Regardless, other manufacturers are probably going to demand
 > something equivalent to this based on the PCIe standard, so we should be
 > ready for it, hence the desire for the bus specific attributes.

But:
 a) The Altix has PCI-X, not PCIe, so having something PCIe-specific
    is not a solution for this case; and
 b) the PCIe behavior is opt-in, in the sense that you have to
    specifically ask for looser ordering, while the Altix is loosely
    ordered unless you ask for this "flush" property.  So I don't
    think the same attribute will work for both cases.

 - R.