writel(), readl() in <asm-ppc/io.h>

writel(), readl() in <asm-ppc/io.h>

[Shawn Jin]

Hi,

In <asm-ppc/io.h>, writew(), readw(), writel(), and readl() are
defined to little endian access for all platforms unless either
CONFIG_APUS or CONFIG_8260_PCI9 is defined.

Why? Aren't they correct in big endian systems, are they? Maybe I miss
something here.

Regards,
-Shawn.

Re: writel(), readl() in <asm-ppc/io.h>

[Matt Porter]

On Fri, Aug 12, 2005 at 06:11:14PM -0700, Shawn Jin wrote:
> Hi,
> 
> In <asm-ppc/io.h>, writew(), readw(), writel(), and readl() are
> defined to little endian access for all platforms unless either
> CONFIG_APUS or CONFIG_8260_PCI9 is defined.
> 
> Why? Aren't they correct in big endian systems, are they? Maybe I miss
> something here.

I'm not sure what your qustion is but I'll take a stab at an answer. :)

read*()/write*() are accessors for PCI and PCI only.  PCI is little
endian. PPC is big endian.  All platforms must byte swap on access
to PCI memory space except in special cases.  The two exceptions
must byte swap in hardware. Looks like it is due to errata in the
8260 PCI bridge case. APUS is probably byte swapping in hardware
because APUS is simply odd.

-Matt

Re: writel(), readl() in <asm-ppc/io.h>

[Shawn Jin]

> read*()/write*() are accessors for PCI and PCI only.  PCI is little

If read*()/write*() are designed for PCI access only as you claimed,
that explains why they call in/out_leXX() funcitons.

The problem is that read*()/write*() are misused in some places, e.g.,
serial drivers such as serial8250. The serial_in() and serial_out()
call read*() and write*() respectively. So what's your recommendation
in such a case?

Regards,
-Shawn.

Re: writel(), readl() in <asm-ppc/io.h>

[Eugene Surovegin]

On Sat, Aug 13, 2005 at 08:56:50PM -0700, Shawn Jin wrote:
> > read*()/write*() are accessors for PCI and PCI only.  PCI is little
> 
> If read*()/write*() are designed for PCI access only as you claimed,
> that explains why they call in/out_leXX() funcitons.
> 
> The problem is that read*()/write*() are misused in some places, e.g.,
> serial drivers such as serial8250. The serial_in() and serial_out()
> call read*() and write*() respectively. So what's your recommendation
> in such a case?

Why do you think this is a problem? Why do you think they are 
misused? Please explain. 

FYI, there _are_ serial PCI cards.

-- 
Eugene

Re: writel(), readl() in <asm-ppc/io.h>

[Matt Porter]

On Sat, Aug 13, 2005 at 08:56:50PM -0700, Shawn Jin wrote:
> > read*()/write*() are accessors for PCI and PCI only.  PCI is little
> 
> If read*()/write*() are designed for PCI access only as you claimed,
> that explains why they call in/out_leXX() funcitons.
> 
> The problem is that read*()/write*() are misused in some places, e.g.,
> serial drivers such as serial8250. The serial_in() and serial_out()
> call read*() and write*() respectively. So what's your recommendation
> in such a case?

Keep misusing them. There's no generic accessors for memory mapped
I/O.  People just started using them in generic drivers because they
are convenient. I use the the __raw_read*() for non byte swapped access
(or in/outbe32() depending on my memory barrier requirements.

Also, portions of the 8250 serial driver using readl/writel are
assuming that the serial port in on PCI or other little endian bus.
readb/writeb usage doesn't matter in that driver.

-Matt

Re: writel(), readl() in <asm-ppc/io.h>

[Arthur Othieno]

On Sat, Aug 13, 2005 at 08:56:50PM -0700, Shawn Jin wrote:
> > read*()/write*() are accessors for PCI and PCI only.  PCI is little
> 
> If read*()/write*() are designed for PCI access only as you claimed,
> that explains why they call in/out_leXX() funcitons.
> 
> The problem is that read*()/write*() are misused in some places, e.g.,
> serial drivers such as serial8250. The serial_in() and serial_out()
> call read*() and write*() respectively. So what's your recommendation
> in such a case?
 
See here for the gory details:

  http://lwn.net/Articles/102232/
  http://lwn.net/Articles/132804/

	Arthur

Re: writel(), readl() in <asm-ppc/io.h>

[Shawn Jin]

> > The problem is that read*()/write*() are misused in some places, e.g.,
> > serial drivers such as serial8250. The serial_in() and serial_out()
> > call read*() and write*() respectively. So what's your recommendation
> > in such a case?
>=20
> Why do you think this is a problem? Why do you think they are
> misused? Please explain.

IMHO, serial_in() and serial_out() should be able to handle all the
possilbilities. But the current implementation cannot achieve this by
simply calling read*() and write*() in the case that UART control
registers are 32-bit or 16-bit wide and can only be acessed in 32/16
bits in some big endian systems. As Matt pointed out before, read*()
and write*() are designed for PCI access only.

Regards,
-Shawn.

Re: writel(), readl() in <asm-ppc/io.h>

[Shawn Jin]

>   http://lwn.net/Articles/132804/

This is what exactly I want. Thanks for the URL.

-Shawn.