Re: Proposed changes to io.h

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Sat, 2004-04-03 at 01:41, John Whitney wrote:
> > No, you are not supposed to pass an address coming from
> > pci_alloc_consistent() to pci_map_single(), period.
> >
>
> Why take that stance, when a simple change to virt_to_phys() makes all
> virtual addresses coming from any routine work correctly?  Would that
> be far simpler system under which to implement code, with less prone to
> error (no need to worry about making sure you have the "correct"
> virtual memory type for a given routine - all virtual memory addresses
> would work).  This just sounds entirely political, and very little
> practical.

No. virt_to_phys() operates on the linear mapping only, that has
always been the case, doing otherwise would be non portable anyway.

I think we should deprecate virt_to_phys() anyway and use __pa in
the arch code, that would avoid people trying to use it for what it
is not.

Regarding pci_map_single(), it's semantics are clear regarding the
kind of memory it can map. Doing a ppc only hack to make the nonsense
work (passing it memory obtained by pci_alloc_consistent) would be
letting drivers writers do the wrong thing on ppc and break whenever
they port the driver to a platform that enforce that limitation.

We could add a runtime check to pci_map_single() to verify it gets
passed an address within the linear mapping though. That would catch
the bugs you are afraid of.

(Oh, and let's keep those discussions on-list).

Ben.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/