On Wed, 5 Jul 2023 12:36:40 +0300
Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> On Wed, Jul 5, 2023 at 12:28 PM Jonathan Cameron
> <Jonathan.Cameron@xxxxxxxxxx> wrote:
> > On Wed, 5 Jul 2023 11:53:17 +0300
> > Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > > On Wed, Jul 5, 2023 at 10:55 AM Jonathan Cameron
> > > <Jonathan.Cameron@xxxxxxxxxx> wrote:
> > > > > > From: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
> > > > > > Sent: Sunday, July 2, 2023 6:04 PM
> > > > > > On Thu, 22 Jun 2023 22:32:27 +0800
> > > > > > Kim Seer Paller <kimseer.paller@xxxxxxxxxx> wrote:
>
> ...
>
> > > > > > > + /*
> > > > > > > + * Convert transmit buffer to big-endian format and reverse transmit
> > > > > > > + * buffer to align with the LSB-first input on SDI port.
> > > > > > > + */
> > > > > > > + st->spi_tx_buffer =
> > > > > > cpu_to_be16(bitrev16(FIELD_PREP(MAX14001_ADDR_MASK,
> > > > > > > + reg_addr)));
> > > > > > > +
> > > > > > > + ret = spi_sync_transfer(st->spi, xfers, ARRAY_SIZE(xfers));
> > > > > > > + if (ret)
> > > > > > > + return ret;
> > > > > > > +
> > > > > > > + /*
> > > > > > > + * Align received data from the receive buffer, reversing and reordering
> > > > > > > + * it to match the expected MSB-first format.
> > > > > > > + */
> > > > > > > + *data = (__force u16)(be16_to_cpu(bitrev16(st->spi_rx_buffer))) &
> > > > > > > +
> > > > > > MAX14001_DATA_MASK;
> > > > > > > +
> > > > > > These sequences still confuse me a lot because I'd expect the values in tx
> > > > > > to have the opposite operations applied to those for rx and that's not the
> > > > > > case.
> > > > > >
> > > > > > Let's take a le system.
> > > > > > tx = cpu_to_be16(bitrev16(x))
> > > > > > = cpu_to_be16((__bitrev8(x & 0xff) << 8) | __bitrev8(x >> 8));
> > > > > > = __bitrev8(x & 0xff) | (__bitrev8(x >> 8) << 8)
> > > > > > or swap all the bits in each byte, but don't swap the bytes.
> > > > > >
> > > > > > rx = cpu_to_be16(bitrev16(x))
> > > > > > = be16_to_cpu(((__bitrev8(x & 0xff) << 8) | __bitrev8(x >> 8)_
> > > > > > = __bitrev8(x & 0xff) | __bitrev(x >> 8)
> > > > > >
> > > > > > also swap all the bits in each byte, but don't swap the bytes.
> > > > > >
> > > > > > So it is the reverse because the bytes swaps unwind themselves somewhat.
> > > > > > For a be system cpu_to_be16 etc are noop.
> > > > > > tx = (__bitrev8(x & 0xff) << 8) | __bitrev8(x >> 8)
> > > > > > rx = (__bitrev8(x & 0xff) << 8) | __bitrev8(x >> 8)
> > > > > >
> > > > > > So in this case swap all 16 bits.
> > > > > >
> > > > > > Now, given I'd expected them to be reversed for the tx vs rx case.
> > > > > > E.g.
> > > > > > tx = cpu_to_be16(bitrev16(x))
> > > > > > As above.
> > > > > > For rx, le host
> > > > > > rx = bitrev16(be16_to_cpu(x))
> > > > > > = __bitrev8((x >> 8) & 0xff) << 8) | __bitrev8((((x & 0xff) << 8) >> 8)
> > > > > > same as above (if you swap the two terms I think.
> > > > > >
> > > > > > For be the be16_to_cpu is a noop again, so it's just bitrev16(x) as expected.
> > > > > >
> > > > > > Hence if I've understood this correctly you could reverse the terms so that
> > > > > > it was 'obvious' you were doing the opposite for the tx term vs the rx one
> > > > > > without making the slightest bit of difference....
> > > > > >
> > > > > > hmm. Might be worth doing simply to avoid questions.
> > > > >
> > > > > Thank you for your feedback. I have tested the modifications based on your
> > > > > suggestions, taking the le system into account, and it appears that the code is
> > > > > functioning correctly. Before sending the new patch version, I would like to
> > > > > confirm if this aligns with your comments.
> > >
> > > > Yes. This looks good to me.
> > >
> > > I think the implementation is still incorrect. See below.
> > >
> > > > > static int max14001_read(void *context, unsigned int reg_addr, unsigned int *data)
> > > > > {
> > > > > struct max14001_state *st = context;
> > > > > int ret;
> > > > >
> > > > > struct spi_transfer xfers[] = {
> > > > > {
> > > > > .tx_buf = &st->spi_tx_buffer,
> > > > > .len = sizeof(st->spi_tx_buffer),
> > > > > .cs_change = 1,
> > > > > }, {
> > > > > .rx_buf = &st->spi_rx_buffer,
> > > > > .len = sizeof(st->spi_rx_buffer),
> > > > > },
> > > > > };
> > >
> > > > > st->spi_tx_buffer = cpu_to_be16(bitrev16(FIELD_PREP(MAX14001_ADDR_MASK, reg_addr)));
> > >
> > > Here we got bits in CPU order, reversed them and converted to BE16.
> > >
> > > > > ret = spi_sync_transfer(st->spi, xfers, ARRAY_SIZE(xfers));
> > > > > if (ret)
> > > > > return ret;
> > >
> > > > > *data = cpu_to_be16(bitrev16(st->spi_rx_buffer));
> > >
> > > Here we take __be16 response, reverse them and convert to BE16?!
> > > This is weird. You should have be16_to_cpu() somewhere, not the opposite.
> > Good point - though functionally they end up the same (and the bitrev
> > is making mess of type markings anyway). It is more logical
> > to ensure the direction is reversed as you suggest.
>
> Also a question why we don't do that in reversed order.
> Logically it sounds like bitrev16(be16_to_cpu()) should be.
> Will it give the wrong results?
Shouldn't make any difference as the two operations commute.
I'd missed this. You are right that the other order makes more
sense.
Jonathan
>
> All in all this algo should be described in the comment in the code
> (if not yet).
>
> > > > > return 0;
> > > > > }
>
