Re: [PATCH] x86/calgary: fix bitcast type warnings

[Jann Horn]

On Fri, Mar 29, 2019 at 10:32 PM Logan Gunthorpe <logang@xxxxxxxxxxxx> wrote:
> On 2019-03-29 3:19 p.m., Jann Horn wrote:
> >>> Can the existing api's not be used here like iowrite64be/ioread64be/ or
> >>> similar variant in "include/asm-generic/io.h"
> >>
> >> Oooh! I didn't realize that those exist. I'll change that and send a v2.
> Yes, they are very new! It took me a while to get those patches in as it
> is a bit more complicated than it seems.
>
> > Actually, that doesn't work at the moment on x86-64:
> >
> > include/asm-generic/io.h only defines these things if
> > CONFIG_GENERIC_IOMAP isn't defined; and X86 unconditionally defines
> > it. With CONFIG_GENERIC_IOMAP set, these functions are provided by
> > include/asm-generic/iomap.h.
>
> For CONFIG_GENERIC_IOMAP, lib/iomap.c provides _lo_hi and _hi_lo
> implementations seeing the pio regions must be emulated with 32bit
> operations and we have to define the order.
>
> > The definitions for these are in lib/iomap.c, except that there are no
> > definitions for ioread64be() and iowrite64be(); if you try to use
> > them, you get linker errors.
>
> Some platforms implement these but most do not. If I recall correctly
> only powerpc does.
>
> If you want to use 64 bit operations in a portable fashion, you should
> include "linux/io-64-nonatomic-hi-lo.h" or
> "linux/io-64-nonatomic-lo-hi.h", depending on weather you want the lower
> bits or the higher bits to be written or read first in cases where an
> atomic operation is not available.

So what is the right thing to do in the context of
arch/x86/kernel/pci-calgary_64.c? That code wants to perform MMIO with
endianness conversion, and these accesses are always performed as
MMIO. Using the non-atomic 64-bit I/O helpers for this would be kind
of awkward, since the accesses would never actually be split...