Re: [PATCH 4/4] lib/iomap.c: improve comment about pci anomaly

From: Arnd Bergmann
Date: Wed Nov 29 2023 - 12:38:00 EST


On Wed, Nov 29, 2023, at 11:16, Philipp Stanner wrote:
> On Fri, 2023-11-24 at 20:08 +0100, Danilo Krummrich wrote:
>>
>> lib/pci_iomap.c contains another definition of pci_iounmap() which is
>> guarded by ARCH_WANTS_GENERIC_PCI_IOUNMAP to prevent multiple
>> definitions
>> in case either GENERIC_IOMAP is set or the architecture already
>> defined
>> pci_iounmap().
>
> To clarify that, here's the relevant excerpt from include/asm-
> generic/io.h:
>
> #ifndef CONFIG_GENERIC_IOMAP
> #ifndef pci_iounmap
> #define ARCH_WANTS_GENERIC_PCI_IOUNMAP
> #endif
> #endif

Right, this was added fairly recently in an effort to
unify the architectures that can share a simple implementation
based on the way that modern PCI host bridges on non-x86
work.

>> What's the purpose of not having set ARCH_HAS_GENERIC_IOPORT_MAP
>> producing
>> an empty definition of pci_iounmap() though [1]?
>>
>> And more generally, is there any other (subtle) logic behind this?
>
> That's indeed also very hard to understand for me, because you'd expect
> that if pci_iomap() exists (and does something), pci_iounmap() should
> also exist and, of course, unmapp the memory again.

Right, I think that was a leak introduced in 316e8d79a095
("pci_iounmap'2: Electric Boogaloo: try to make sense of
it all") that should be fixed like

--- a/lib/pci_iomap.c
+++ b/lib/pci_iomap.c
@@ -170,8 +170,8 @@ void pci_iounmap(struct pci_dev *dev, void __iomem *p)

if (addr >= start && addr < start + IO_SPACE_LIMIT)
return;
- iounmap(p);
#endif
+ iounmap(p);
}
EXPORT_SYMBOL(pci_iounmap);

i.e. architectures without port I/O just call iounmap() but those
that support the normal ioport_map() have to skip iounmap()
for ports in that special PIO range.

> Regarding the last point, a number of architectures define their own
> ioport_map():
>
> arch/alpha/kernel/io.c, line 684 (as a function)
> arch/arc/include/asm/io.h, line 27 (as a function)
> arch/arm/mm/iomap.c, line 19 (as a function)
> arch/m68k/include/asm/kmap.h, line 60 (as a function)
> arch/parisc/lib/iomap.c, line 504 (as a function)
> arch/powerpc/kernel/iomap.c, line 14 (as a function)
> arch/s390/include/asm/io.h, line 38 (as a function)
> arch/sh/kernel/ioport.c, line 24 (as a function)
> arch/sparc/lib/iomap.c, line 10 (as a function)
>
> I grepped through those archs and as I see it, none of those specify an
> empty pci_iomap() that could be a counterpart to the potentially empty
> pci_iounmap() in lib/pci_iomap.c

I'm trying to unwind what you are saying here, and there are
two separate issues:

- an empty unmap() function still makes sense if the map() function
just returns a usable pointer like the asm-generic version
of ioport_map(), it only has to be non-empty if the map function
allocates a resource that has to be freed later, like the
page table entries for most ioremap() implementations.

- pci_iounmap() in lib/pci_iomap.c being empty is probably
just a bug

>> From what I can tell looking at the header, I think we can
>> just remove the "#elif defined(CONFIG_GENERIC_PCI_IOMAP)"
>> bit entirely, as it no longer serves the purpose it originally
>> had.
>
> So it seems that the empty unmap-function in pci_iomap.c is the left-
> over counterpart of those mapping functions always returning NULL.

no

> @Arnd:
> Your code draft
>
> void pci_iounmap(struct pci_dev *dev, void __iomem * addr)
> {
> #ifdef CONFIG_HAS_IOPORT
> if (iomem_is_ioport(addr)) {
> ioport_unmap(addr);
> return;
> }
> #endif
> iounmap(addr)
> }
>
> seems to agree with that: There will never be the need for an empty
> function that does nothing. Correct?

Agreed, while arch/sparc/ currently has an empty pci_iounmap(),
that is just because the normal iounmap() on that architecture
is also empty, given that all MMIO memory is always mapped.

>> > {
>> > #ifdef CONFIG_HAS_IOPORT
>> >         if (iomem_is_ioport(addr)) {
>> >                ioport_unmap(addr);
>> >                return;
>> >         }
>> > #endif
>> >        iounmap(addr)
>> > }
>> >
>> > and then define iomem_is_ioport() in lib/iomap.c for x86,
>> > while defining it in asm-generic/io.h for the rest,
>> > with an override in asm/io.h for those architectures
>> > that need a custom inb().
>>
>> So, that would be similar to IO_COND(), right? What would we need
>> inb() for in this context?

In general, any architecture that has a custom inb() also
needs a custom ioport_map() and iomem_is_ioport() in this
scheme, while the "normal" architectures like arm/arm64 and
riscv should be able to just use the asm-generic version.

IO_COND() is really specific to those architectures that
rely on the rather misnamed GENERIC_IOMAP for implementing
ioport_map().

Arnd