Re: [PATCH v3 2/2] i2c: xiic: Use 32bit accesses only

[Arnd Bergmann <arnd@arndb.de>]

On Wednesday 23 February 2011 10:37:31 Michal Simek wrote:
> Arnd Bergmann wrote:
> > On Wednesday 23 February 2011 09:08:47 Michal Simek wrote:
> >> Grant Likely wrote:
> >>> On Tue, Feb 22, 2011 at 10:49 AM, Michal Simek <monstr@monstr.eu> wrote:
> >>>> i2c driver is used for LE/BE that's why is useful to use
> >>>> 32bit accesses. Then it is not necessary to solve any
> >>>> endian issues.
> >>> Are you sure?  I would expect the BE version needs to use
> >>> io{read,write}32be variants of the accessors.  What platforms have you
> >>> tested on?
> >> iowrite32 is the same with iowrite32be for Microblaze.
> >> I have no problem to change it to iowrite32be if you like.
> >>
> >> I have tested it on microblaze big and little endian platforms.
> > 
> > I think what Grant was saying is that iowrite32 being the same as
> > iowrite32be is a bug, because iowrite32 is documented to be little-endian.
> 
> Can you pointed me to that documentation?

Documented was maybe a term that is bit too strong in this case ;-)

Documentation/memory-barriers.txt documents that iowrite32 is the same
as either writel or outl, depending on the source of the mapping.

Both include/asm-generic/io.h and include/asm-generic/iomap.h define
variants of iowrite32 and iowrite32be, but the definition in io.h is
actually broken.

> > This is probably fine as long as you don't have any PCI devices,
> > but if you ever get PCI support, it won't work.
> 
> Microblaze have PCI support but I don't have the board for testing.

Ok, I see.
 
> > Also, it heavily confuses other developers such as Grant and me
> > if one architecture defines things to mean something completely
> > different from the other architectures.
> 
> As I see will be good to review microblaze io.h and use generic one.

The generic io.h is lacking in a number of ways and should really
be improved. Guan Xuetao also fell into the same trap with the
unicore32 port.

The problem is that I copied asm/io.h from mn10300, which seemed
generic enough, but it really doesn't work correctly with PCI.
There is one patch for this that I did in the unicore32 tree,
I should probably go through it and fix everything else that sticks
out.

	Arnd