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

[Andy Shevchenko <andriy.shevchenko@xxxxxxxxx>]

On Wed, Apr 12, 2023 at 02:51:44PM +0530, Joy Chakraborty wrote:
> On Wed, Apr 12, 2023 at 1:11 AM Joy Chakraborty <joychakr@xxxxxxxxxx> wrote:
> > On Tue, Apr 11, 2023 at 9:52 PM Andy Shevchenko
> > <andriy.shevchenko@xxxxxxxxx> 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@xxxxxxxxx> 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.

Thank you for finding the answer for me by yourself!

> On the other hand I do see that in the fifo driver dw_writer() /
> dw_reader() increments the pointer with 3 incase n_bytes = 3 even
> though it copies 4 bytes.
> ...
>    if (dws->n_bytes == 1)
>       txw = *(u8 *)(dws->tx);
>    else if (dws->n_bytes == 2)
>       txw = *(u16 *)(dws->tx);
>    else
>       txw = *(u32 *)(dws->tx);
> 
> dws->tx += dws->n_bytes;
> ...
> This will not behave as using DMA_SLAVE_BUSWIDTH_4_BYTES in the DMA so
> maybe I am not correct in interpretting the spi.h header file.
> Can CPU's in general access u32 from unaligned odd addresses ?

Generally speaking the above code must check number of bytes for being 4.

> From Serge(y)'s comment regarding this, the APB interface writing data
> to the FIFO register definitely cannot handle
> DMA_SLAVE_BUSWIDTH_3_BYTES since it handles a power of 2 only.
> Hence we can possibly remove "case 3:" completely and also mask out
> DMA_SLAVE_BUSWIDTH_3_BYTES from dma_addr_width capabilities so that
> can_dma api does not allow n_bytes = 3 to use DMA.
> 
> Would that be correct ?

We have to fix the above and the DIV_ROUND_UP(transfer->bits_per_word,
BITS_PER_BYTE) one to be something like

roundup_pow_of_two(round_up(..., BITS_PER_BYTE))

> > > > > > > > > > +     case 4:
> > > > > > > > > > +             return DMA_SLAVE_BUSWIDTH_4_BYTES;
> > > > > > > > > > +     default:
> > > > > > > > > > +             return DMA_SLAVE_BUSWIDTH_UNDEFINED;
> > > > > > > > > > +     }

-- 
With Best Regards,
Andy Shevchenko