On Thu, 2016-06-02 at 11:01 +0200, Arnd Bergmann wrote:

On Wednesday, June 1, 2016 8:24:20 PM CEST Scott Wood wrote:

quoted

On Mon, 2016-05-30 at 15:18 +0200, Arnd Bergmann wrote:

quoted

All users of this driver are PowerPC specific and the header file
has no business in the global include/linux/ hierarchy, so move
it back before anyone starts using it on ARM.

This reverts commit 948486544713492f00ac8a9572909101ea892cb0.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
This part of the series is not required for the eSDHC quirk,
but it restores the asm/fsl_guts.h header so it doesn't accidentally
get abused for this in the future. I found two drivers outside of
arch/powerpc that already accessed the registers directly, but the
functions look fairly contained, and can be easily hidden in an
#ifdef CONFIG_PPC

NACK

Besides adding ifdef pollution for no good reason, this register block is
used
on some ARM chips as well.  Why is it a problem if "anyone starts using it
on
ARM"?

It's just not a good interface when it's defined as "this is the layout of
a register area that any driver can ioremap() if they can figure out the
device node".

That's why I want to move accesses into one guts driver.

 It's not uncommon to have register areas like that, but
normally you have at the minimum a 'syscon' device to handle locking
between drivers accessing the same registers and to avoid having to map
the same area multiple times.

syscon requires device tree changes.

I don't see read-modify-write operations in regmap -- how does locking around
an individual, inherently-atomic load or store help?

If we need to use 'guts' registers on ARM, we can find a way to abstract
them properly for the given use cases, using a syscon or a driver with
exported functions, but just making a PowerPC platform specific header
global to all Linux drivers by putting it into include/linux doesn't seem
right.

Again, it's not PowerPC-specific!  It started that way but then the same
register block got put onto some ARM chips.

It's not global to "all Linux drivers", just the ones that choose to include
an fsl-specific header.

If and when all uses of guts are moved into the guts driver, the header can be
moved into drivers/soc/fsl.

Note that the header file uses a structure definition rather than the more
common macros with register offsets, which is fine for a driver that has
its own registers and abstracts them, but it doesn't really work with
the regmap interface, so if we want to use it with syscon, it also needs to
be rewritten.

We don't want to use it with syscon.  If we did, the solution wouldn't be to
move the header back to arch/powerpc, but to convert the struct into offsets.

-Scott