Re: [PATCH v4] Kconfig: introduce HAS_IOPORT option and select it as necessary

[Geert Uytterhoeven <geert@xxxxxxxxxxxxxx>]

Hi David,

On Wed, Apr 5, 2023 at 11:37 PM David Laight <David.Laight@xxxxxxxxxx> wrote:
> From: Linuxppc-dev Arnd Bergmann
> > Sent: 05 April 2023 21:32
> >
> > On Wed, Apr 5, 2023, at 22:00, H. Peter Anvin wrote:
> > > On April 5, 2023 8:12:38 AM PDT, Niklas Schnelle <schnelle@xxxxxxxxxxxxx> wrote:
> > >>On Thu, 2023-03-23 at 17:33 +0100, Niklas Schnelle wrote:
> > >>> We introduce a new HAS_IOPORT Kconfig option to indicate support for I/O
> > >>> Port access. In a future patch HAS_IOPORT=n will disable compilation of
> > >>> the I/O accessor functions inb()/outb() and friends on architectures
> > >>> which can not meaningfully support legacy I/O spaces such as s390.
> > >>> >>
> > >>Gentle ping. As far as I can tell this hasn't been picked to any tree
> > >>sp far but also hasn't seen complains so I'm wondering if I should send
> > >>a new version of the combined series of this patch plus the added
> > >>HAS_IOPORT dependencies per subsystem or wait until this is picked up.
> > >
> > > You need this on a system supporting not just ISA but also PCI.
> > >
> > > Typically on non-x86 architectures this is simply mapped into a memory window.
> >
> > I'm pretty confident that the list is correct here, as the HAS_IOPORT
> > symbol is enabled exactly for the architectures that have a way to
> > map the I/O space. PCIe generally works fine without I/O space, the
> > only exception are drivers for devices that were around as early PCI.
>
> Isn't there a difference between cpu that have inb()/outb() (probably
> only x86?) and architectures (well computer designs) that can generate
> PCI 'I/O' cycles by some means.
> It isn't even just PCI I/O cycles, I've used an ARM cpu (SA1100)
> that mapped a chuck of physical address space onto PCMCIA I/O cycles.
>
> If the hardware can map a PCI 'IO' bar into normal kernel address
> space then the bar and accesses can be treated exactly like a memory bar.
> This probably leaves x86 as the outlier where you need (IIRC) io_readl()
> and friends that can generate in/out instructions for those accesses.
>
> There are also all the x86 ISA devices which need in/out instructions.
> But (with the likely exception of the UART) they are pretty much
> platform specific.
>
> So, to my mind at least, HAS_IOPORT is just the wrong question.

Not all PCI controllers support mapping the I/O bar in MMIO space, so
in general you cannot say that CONFIG_PCI=y means CONFIG_HAS_IOPORT=y.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds