Re: [RFC v2 25/39] pcmcia: add HAS_IOPORT dependencies

[Niklas Schnelle]

On Wed, 2022-05-04 at 12:33 +0200, Arnd Bergmann wrote:
> On Wed, May 4, 2022 at 1:38 AM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote:
> > On Fri, Apr 29, 2022 at 03:50:41PM +0200, Niklas Schnelle wrote:
> > > In a future patch HAS_IOPORT=n will result in inb()/outb() and friends
> > > not being declared. PCMCIA devices are either LEGACY_PCI devices
> > > which implies HAS_IOPORT or require HAS_IOPORT.
> > > 
> > > Acked-by: Dominik Brodowski <linux@xxxxxxxxxxxxxxxxxxxx>
> > > Co-developed-by: Arnd Bergmann <arnd@xxxxxxxxxx>
> > > Signed-off-by: Niklas Schnelle <schnelle@xxxxxxxxxxxxx>
> > > ---
> > >  drivers/pcmcia/Kconfig | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/drivers/pcmcia/Kconfig b/drivers/pcmcia/Kconfig
> > > index 2ce261cfff8e..32b5cd324c58 100644
> > > --- a/drivers/pcmcia/Kconfig
> > > +++ b/drivers/pcmcia/Kconfig
> > > @@ -5,7 +5,7 @@
> > > 
> > >  menuconfig PCCARD
> > >       tristate "PCCard (PCMCIA/CardBus) support"
> > > -     depends on !UML
> > > +     depends on HAS_IOPORT
> > 
> > I don't know much about PC Card.  Is there a requirement that these
> > devices must use I/O port space?  If so, can you include a spec
> > reference in the commit log?
> 
> I think for PCMCIA devices, the dependency makes sense because
> all device drivers for PCMCIA devices need I/O ports.
> 
> For cardbus, we can go either way, I don't see any reference to
> I/O ports in yenta_socket.c or the pccard core, so it would build
> fine with or without I/O ports.
> 
> > I do see the PC Card spec, r8.1, sec 5.5.4.2.2 says:
> > 
> >   All CardBus PC Card adapters must support either memory-mapped I/O
> >   or both memory-mapped I/O and I/O space. The selection will depend
> >   largely on the system architecture the adapter is intended to be
> >   used in. The requirement to also support memory-mapped I/O, if I/O
> >   space is supported, is driven by the potential emergence of
> >   memory-mapped I/O only cards. Supporting both modes may also
> >   position the adapter to be sold into multiple system architectures.
> > 
> > which sounds like I/O space is optional.
> 
> An earlier version of the patch series had a separate
> CONFIG_LEGACY_PCI that required CONFIG_HAS_IOPORT
> here, which I think made this clearer:
> 
> Almost all architectures that support CONFIG_PCI also provide
> HAS_IOPORT today (at least at compile time, if not at runtime),
> with s390 as a notable exception. Any machines that have legacy
> PCI device support will also have I/O ports because a lot of
> legacy PCI cards used it, and any machine with a pc-card slot
> should also support legacy PCI devices.
> 
> If we get new architectures without I/O space in the future, they
> would certainly not care about supporting old cardbus devices.
> 
>         Arnd

I tested removing the HAS_IOPORT dependency on PCCARD thus ungating
PCMCIA and CARDBUS. This then requires about 30 additional HAS_IOPORT
dependencies to build my s390 allyesconfig.

The only exceptions I found that depends on PCMCIA and isn't otherwise
dependend on ISA or a specific platform are USB_SL811_CS and
PCMCIA_RAYCS (Aviator/Raytheon 2.4GHz wireless).

As for CARDBUS the only "depends on CARDBUS" is in
drivers/net/ethernet/dec/tulip/Kconfig. Now I tested with allyesconfig
which also sets CONFIG_TULIP_MMIO and with that the drivers there did
compile. I guess one should then have "select TULIP_MMIO if
!HAS_IOPORT" in the config NET_TULIP.

So not sure, it seems unlikely we're going to see any new
PCMCIA/CardBus device drivers added and that's a lot of churn for
compile testing so few drivers but in theory it looks like it is
possible to have non-I/O port PCMCIA/CardBus.

Thanks,
Niklas