On Wed, Jul 16, 2014 at 12:01:22PM +0100, Thierry Reding wrote:

From: Thierry Reding <redacted>

Currently driver writers need to use io{read,write}{8,16,32}_rep() when
accessing FIFO registers portably. This is bad for two reasons: it is
inconsistent with how other registers are accessed using the standard
{read,write}{b,w,l}() functions, which can lead to confusion. On some
architectures the io{read,write}*() functions also need to perform some
extra checks to determine whether an address is memory-mapped or refers
to I/O space. Drivers which can be expected to never use I/O can safely
use the {read,write}s{b,w,l,q}(), just like they use their non-string
variants and there's no need for these extra checks.

This patch implements generic versions of readsb(), readsw(), readsl(),
readsq(), writesb(), writesw(), writesl() and writesq(). Variants of
these string functions for I/O accesses (ins*() and outs*() as well as
ioread*_rep() and iowrite*_rep()) are now implemented in terms of the
new functions.

Going forward, {read,write}{,s}{b,w,l,q}() should be used consistently
by drivers for devices that will only ever be memory-mapped and hence
don't need to access I/O space, whereas io{read,write}{8,16,32}_rep()
should be used by drivers for devices that can be either memory-mapped
or I/O-mapped.

While at it, also make sure that any of the functions provided as
fallback for architectures that don't override them can't be overridden
subsequently.

This is compile- and runtime-tested on 32-bit and 64-bit ARM and compile
tested on Microblaze, s390, SPARC and Xtensa. For ARC, Blackfin, Metag,
OpenRISC, Score and Unicore32 which also use asm-generic/io.h I couldn't
find or build a cross-compiler that would run on my system. But by code
inspection they shouldn't break with this patch.

Signed-off-by: Thierry Reding <redacted>

Acked-by: Catalin Marinas <catalin.marinas@arm.com>