openbmc-20160222-1

Make and use new routines to access the fifo when in user mode.
Instead of using memcpy_fromio, base the new routines on the io
port string accessors.

When the smc controllers are configured to transfer in user mode,
the exact address within the chip select region is ignored and
all valid data bytes are sent to the device fifo.  It has been
determined the reason _memcpy_fromio, an arm specific byte-by-byte
variant of memcpy_fromio was required in read_reg, is that the
later will stutter when coping from memory, discarding some
bytes read, when switching from words to bytes for the final
unaligned length.

This change includes a hack to make inw/outw work.  The default
generic IO_SPACE_LIMIT is 0, and twenty bits when PCI is selected.
While there is a PCMCIA SOC config to select the full 4GB space it
would bring in the PCMCIA core for no use to the ast2400.  We want
to be part of the generic multi-platform arm kernel and therefore
mach/io.h is not going to work.  However, the limit is under an
ifndef, so define it in this file an check it was not redefine
at probe time.  To avoid redefinition errors or warnings, guard
the setting with CONFIG_ARM for cross-arch compile testing and
with a Kconfig dependency for arm cross-compile tests.  While some
other architecture may not accept unsigned long addresses as ioport
cookies, that can be addressed when, if ever, the hardware appears
on another arch.

While the change to use the designed-for-memory optimised copy
template, even reusing the bytes of memcpy, was a recent change
and only for little endian arm, designing and proposing a patch
will take some time and there could be debate if the routine
needs to work on changing memory.

Signed-off-by: Milton Miller <miltonm@us.ibm.com>
Reviewed-by: Andrew Jeffery <andrew@aj.id.au>
Signed-off-by: Joel Stanley <joel@jms.id.au>