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

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

On Sunday 21 December 2008, Linus Torvalds wrote:
> 
> On Sat, 20 Dec 2008, David Brownell wrote:
> > 
> > All SPI transfers are full duplex, and are packaged as half duplex
> > by either discarding the data that's read ("write only"), or else
> > by writing zeroes ("read only").  That patch wasn't ensuring that
> > zeroes were getting written out during the "half duplex read" part
> > of the transaction; instead, old RX bits were getting sent.
> 
> Hmm. In addition, isn't this broken (in that same function):

No -- this is full duplex.  The write_then_read() helper is
simplifying a common half-duplex idiom for short operations,
but the harware 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.

 
>         memcpy(local_buf, txbuf, n_tx);
>         x.tx_buf = local_buf;
>         x.rx_buf = local_buf;
> 
>         /* do the i/o */
>         status = spi_sync(spi, &message);
>         if (status == 0)
>                 memcpy(rxbuf, x.rx_buf + n_tx, n_rx);
> 
> shouldn't that 'rx_buf' setup be
> 
> 	x.rx_buf = local_buf + n_tx;
> 
> since the whole point was that we allocated a buffer that can hold _both_ 
> the rx and tx parts? Especially as that final copy into the resulting 
> "rxbuf" thing uses that "+ n_tx" addition?

See above.  We only want the "R" bits which were shifted in
right *after* the n_tx "W" bits.  If we offset rx_buf before
the I/O, we'd start with the "x" don't-care bits and need to
do something else to discard them.  (Plus, allocate more
space at the end of the buffer.)

- Dave


> 			Linus
> 
>