On Mon, Sep 20, 2021 at 12:14 PM John Paul Adrian Glaubitz
<glaubitz@xxxxxxxxxxxxxxxxxxx> wrote:
>
> If you want to get feedback whether the kernel actually boots, let me know.
So having looked around more sparc64 actually looks to be ok as-is,
because it doesn't do any ioremap at all, and the PIO accesses are
done at physical address zero.
Sparc uses a special IO memory address space and can basically map all
of PCI that way, and it looks like the hardware does all the required
special things for the PIO range at address 0-0xffff.
So it turns out that the "missing iounmap()" is actually ok on sparc,
because it's a no-op anyway - because the ioremap() was just a pointer
cast with no actual remapping necessary.
And the generic IOMAP thing does assume that PIO is special, in ways
that sparc doesn't need. On x86, PIO is not remapped, but also uses
different instructions, so it's not just pointer games that could be
done at iomap/unmap case.
(And on many other architectures you need to do different
synchronization, even if you could perhaps otherwise make the
PIO-vs-MMIO be only about the pointer mapping - so "writeb()" and
"outb()" aren't just different in the addressing).
End result: the only downside of sparc not using the generic iomap is
likely that sparc will happily use a NULL __iomap pointer (error) and
basically use it as a PIO access. But since other architectures like
x86-64 would warn for that case (see 'bad_io_access()' in
lib/iomap.c), even that isn't actually a big deal - any such bugs
would have been found elsewhere.
And having looked at this, I'm starting to suspect that sparc oddity
is _why_ the fallback version in <asm-generic/io.h> was so broken. It
did the right thing on sparc, but leaks iomap remappings almost
anywhere else. But maybe sparc ended up being the only user of it?
Linus
