Re: [patch 2.6.28-rc9] spi: spi_write_then_read() regression fix

[David Brownell <david-b@pacbell.net>]

On Monday 22 December 2008, Ingo Oeser wrote:
> Hi David,
> 
> On Monday 22 December 2008, you wrote:
> > 	  The write_then_read() helper is
> > simplifying a common half-duplex idiom for short operations,
> > but the hardware still does full duplex.  Buffer layout is:
> > 
> >   Before:	WWWWW0000000
> >   After:	xxxxxRRRRRRR
> > 
> > That is, for every bit shifted out (W, 0) another one gets
> > shifted in (x, R).  The I/O primitive essentially swaps
> > contents of a one-word shift register between master and
> > slave; or, sequences of such words.  Words don't need to
> > be byte-size, though that's a common option.
> 
> ...
> 
> Wow, what interesting hardware logic and a nice explanation.

It's standard SPI ... yes, it's kind of interesting, and I've
not come across anything else that synchronizes RX and TX in
quite the same way.

I suspect that hardware designers find the ability to get
quick transfers (tens of megabits/second) from just a shift
register is fairly attractive.  Less work than high speed
I2C, for one example.


> Could you put that into a comment somewhere close to those helpers?

Fair enough.  That file does lack any mention of the I/O model
for SPI ... and it's certainly a bit of a surprise to anyone
who's not had to look at it before!

 
> You can safely assume, that any code which Linus doesn't understand
> is non-trivial and needs a comment :-)

I'm not sure I'd agree with that -- while he's many things,
"expert in everything" is not one of them! -- but I do agree
that comment would be worth adding.


However, with regards to $PATCH ... considering that the '28
release is just around the corner, I changed my mind and would
suggest just reverting f9b90e39cbc5c4d6ef60022fd1f25d541df0aad1
much though I very much like that fix in general.

Reason being:  it needs a bit more work, I forgot that it will
break for example drivers using drivers/spi/omap_uwire.c since
that is built on top of truly half-duplex hardware.  (Used with
OMAP1 SOCs.)  The kerneldoc says it's for providing MicroWire
style transactions, so it's kind of significant that it work
on MicroWire hardware too.  ;)

So an updated version would need to both zero the TX bits when
using normal full-duplex SPI hardware ... and use more or less
the current code when running on half-duplex variants.

- Dave