Re: [PATCH v5 1/2] spi: dw: Add 32 bpw support to DW DMA Controller

[Joy Chakraborty <joychakr@google.com>]

On Tue, Apr 11, 2023 at 9:52 PM Andy Shevchenko
<andriy.shevchenko@intel.com> wrote:
>
> On Tue, Apr 11, 2023 at 08:48:52PM +0530, Joy Chakraborty wrote:
> > On Tue, Apr 11, 2023 at 8:47 PM Andy Shevchenko
> > <andriy.shevchenko@intel.com> wrote:
> > > On Tue, Apr 11, 2023 at 06:09:16PM +0300, Serge Semin wrote:
> > > > On Tue, Apr 11, 2023 at 05:46:34PM +0300, Andy Shevchenko wrote:
> > > > > On Tue, Apr 11, 2023 at 05:11:15PM +0300, Serge Semin wrote:
> > > > > > On Tue, Apr 11, 2023 at 03:13:49PM +0300, Andy Shevchenko wrote:
> > > > > > > On Thu, Mar 30, 2023 at 06:34:49AM +0000, Joy Chakraborty wrote:
>
> ...
>
> > > > > > > > -     if (n_bytes == 1)
> > > > > > > > +     switch (n_bytes) {
> > > > > > > > +     case 1:
> > > > > > > >               return DMA_SLAVE_BUSWIDTH_1_BYTE;
> > > > > > > > -     else if (n_bytes == 2)
> > > > > > > > +     case 2:
> > > > > > > >               return DMA_SLAVE_BUSWIDTH_2_BYTES;
> > > > > > > > -
> > > > > > > > -     return DMA_SLAVE_BUSWIDTH_UNDEFINED;
> > > > > > >
> > > > > > > > +     case 3:
> > > > > > >
> > > > > > > I'm not sure about this.
> > > > > >
> > > > > > This actually makes sense seeing the function argument can have values
> > > > > > 1, 2, _3_ and 4:
> > > > > > dws->n_bytes = DIV_ROUND_UP(transfer->bits_per_word, BITS_PER_BYTE);
> > > > > > transfer->bits_per_word = __F__(master->bits_per_word_mask = SPI_BPW_RANGE_MASK(4, 32));
> > > > > > ...
> > > > > > dw_spi_dma_convert_width(dws->n_bytes)
> > > > > >
> > > > > > The spi_transfer.bits_per_word field value depends on the
> > > > > > SPI peripheral device communication protocol requirements which may
> > > > > > imply the 3-bytes word xfers (even though it's indeed unluckily).
> > > > > >
> > > > > > This semantic will also match to what we currently have in the
> > > > > > IRQ-based SPI-transfer implementation (see dw_writer() and
> > > > > > dw_reader()).
> > > >
> > > > > Nice, but we have DMA_SLAVE_BUSWIDTH_3_BYTES definition for that. Why we don't
> > > > > use it?
> > > >
> > > > We could but there are two more-or-less firm reasons not to do
> > > > that:
> > > > 1. There aren't that much DMA-engines with the
> > > > DMA_SLAVE_BUSWIDTH_3_BYTES capability meanwhile the DW APB SSI just
> > > > ignores the upper bits if CTRLR0.DFS is less than the value actual
> > > > written to the DR registers. Note DW DMAC engine isn't one of such
> > > > controllers. So if we get to meet a peripheral SPI-device with 3-bytes
> > > > word protocol transfers and the DMA-engine doesn't support it the
> > > > DMA-based transfers may fail (depending on the DMA-engine driver
> > > > implementation).
> > > > 2. The DW APB SSIs (3.x and 4.x) can be synthesized with the APB Data
> > > > Bus Width of 8, 16 and 32. So no matter whether DMA-engine supports
> > > > the 3-bytes bus width the system bus most likely will either convert
> > > > the transfers to the proper sized bus-transactions or fail.
> > > >
> > > > So taking all of the above into account not using the
> > > > DMA_SLAVE_BUSWIDTH_3_BYTES macro here seems better than using it with
> > > > a risk to fail some of the platform setups especially seeing the DW
> > > > APB SSI ignores the upper bits anyway.
> > >
> > > But this is not about SPI host hardware, it's about the consumers.
> > > They should know about supported sizes. Either we add the corresponding support
> > > to the driver or remove 3 case as I suggested. I don't think it's correct to
> > > use 3 as 4.
> >
> > Another thing to add is that as per spi.h even if bits per word is a
> > not a power of 2 the buffer should be formatted in memory as a power
> > of 2
> > ...
> >  * @bits_per_word: Data transfers involve one or more words; word sizes
> >  * like eight or 12 bits are common.  In-memory wordsizes are
> >  * powers of two bytes (e.g. 20 bit samples use 32 bits).
> >  * This may be changed by the device's driver, or left at the
> >  * default (0) indicating protocol words are eight bit bytes.
> >  * The spi_transfer.bits_per_word can override this for each transfer.
> > ...
> > Hence for n_bytes = 3 or 24 bits/per word we expect the client SW to
> > format it to 4 byte buffers hence the transaction generated should
> > also be of size 4 from the DMA.
>
> You didn't get my point. The consumer wants to know if the 3 bytes is supported
> or not, that's should be part of the DMA related thing, right?
>
> It's incorrectly to say 4 for 3 if the backend DMA controller behaves differently
> on this. How do you know that (any) DMA controller integrated with this hardware
> has no side effects for this change.
>
> So, I don't think the case 3 is correct in this patch.

I see, I am of the opposite opinion in this case.

Other then Serge(y)'s points,
I was trying to say that irrespective of what the underlying DMA
controller supports we should use DMA_SLAVE_BUSWIDTH_4_BYTES when
n_bytes = 3 from SPI perspective as we get n_bytes from bits per word
passed by the client / spi framework " dws->n_bytes =
DIV_ROUND_UP(transfer->bits_per_word, BITS_PER_BYTE) ".
Based on the spi header what I perceive is that for bits/word between
17 and 32 the data has to be stored in 32bit words in memory as per
the example in the header " (e.g. 20 bit samples use 32 bits) ".

Hence, taking an example to transfer 6 bytes (say 0xAA 0xBB 0xCC 0xDD
0xEE 0xFF) with bits per word as 24 (n_bytes = 3) i.e. a total of 2
words I expect the buffer to look as follows which is coming from the
client:
_ _____address|__________0________4________8________C
    SD:00000000|>00CCBBAA 00FFEEDD 00000000 00000000
Hence to transfer this successfully the DMA controller would need to
copy 4 bytes per word .

Please correct me if my understanding of this is incorrect.

>
> > > > > > > > +     case 4:
> > > > > > > > +             return DMA_SLAVE_BUSWIDTH_4_BYTES;
> > > > > > > > +     default:
> > > > > > > > +             return DMA_SLAVE_BUSWIDTH_UNDEFINED;
> > > > > > > > +     }
>
> --
> With Best Regards,
> Andy Shevchenko
>
>
Thanks
Joy