spi: Provide core support for full duplex devices

[Geert Uytterhoeven]

Hi Mark,

On Thu, Feb 6, 2014 at 1:56 PM, Geert Uytterhoeven <geert at linux-m68k.org> wrote:
> I found this in your for-next branch:
>
> commit 3a2eba9bd0a6447dfbc01635e4cd0689f5f2bdad
> Author: Mark Brown <broonie at linaro.org>
> Date:   Tue Jan 28 20:17:03 2014 +0000
>
>     spi: Provide core support for full duplex devices
>
> If I'm not mistaken, you forgot to email this out to linux-spi?
>
> I'll see whether I can make use of this in the RSPI driver.

I gave it a quick try, with RSPI single transfers.

For Dual/Quad transfers, which are half-duplex, the buffer pointer for the
unused direction must be NULL, so the driver knows in which direction to
transfer. Hence SPI_MASTER_MUST_TX cannot be set for Dual/Quad
capable SPI master.

I'm also worried about the buffer allocation. With m25p80 and mtchar, the
transfer buffer can be quite large (up to 4 MiB in my case), and this can be
triggered from userspace.

While mtdchar uses mtd_kmalloc_up_to(), and reduces the temporary
buffer size under memory pressure, the SPI core must always allocate a
buffer of the matching size. This may fail.

It's also a little bit slower (O(20 ms) for 1 MiB read), due to the overhead of
calling krealloc(), and zeroing the dummy buffer.

Alternatively, to avoid the memory allocation, you could zero rx_buf,
and use rx_buf for both tx and rx? That may break some DMA setups,
though.

Concluding, for PIO mode, I'm still in favor of using

    data = tx_buf ? *tx_buf++ : DUMMY_DATA;

in the SPI master driver itself.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert at linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds