Hi Arnd,

> > ioread64_rep is then only defined if CONFIG_64BIT. As it is based
> > on readsq/writesq() and those must be defined (as you said), I don't get
> > why the *64_rep() helpers are not defined in all cases. Maybe because no
> > 32-bit system _should_ need them? But then compile testing gets more
> > difficult.  
> 
> Both readsq/writesq and ioread64_rep/iowrite64_rep must be defined
> for 64-bit architectures and cannot be defined for 32-bit ones.

Yeah, ok.

> >> - For a FIFO, you cannot use readq() but have to use __raw_readq()
> >>   to get the correct endianness. You cannot use this for an
> >>   MMIO register with side-effects though, as this needs the byteswap
> >>   and the barrier in readsl().  
> >
> > I'm not sure about the true definition of "FIFO" as you say. I guess
> > you just mean reading from a BE device?
> >
> > In this case I guess we need the barrier+byteswap helpers.  
> 
> The difference is that a register has a fixed length, and gets
> accessed with a device specific endianness, which may have to
> be swapped if the device and the CPU disagree.
> 
> A FIFO register is what you use for transferring a stream of
> bytes, such as reading a file system block from disk. The
> first byte in the register corresponds to the first byte in
> memory later, so there must not be any byteswap while copying
> to/from memory. If the data itself is structured (i.e. an
> on-disk inode or a network packet), then the byteswap will
> happen if necessary while interpreting the data.

Ok, I fully get what you mean, I was just not used to the word FIFO for
this definition as I use it as a more generic term, but it completely
makes sense. Thanks for taking the time to explain.

Thanks,
Miquèl