Re: [PATCH 01/13] spi: add core support for controllers with offload capabilities

[David Lechner <dlechner@xxxxxxxxxxxx>]

On Thu, Jan 11, 2024 at 2:54 PM David Lechner <dlechner@xxxxxxxxxxxx> wrote:
>
> On Wed, Jan 10, 2024 at 3:36 PM Mark Brown <broonie@xxxxxxxxxx> wrote:
> >
> > On Wed, Jan 10, 2024 at 01:49:42PM -0600, David Lechner wrote:
> > > This adds a feature for specialized SPI controllers that can record
> > > a series of SPI transfers, including tx data, cs assertions, delays,
> > > etc. and then play them back using a hardware trigger without CPU
> > > intervention.
> >
> > > The intended use case for this is with the AXI SPI Engine to capture
> > > data from ADCs at high rates (MSPS) with a stable sample period.
> >
> > > Most of the implementation is controller-specific and will be handled by
> > > drivers that implement the offload_ops callbacks. The API follows a
> > > prepare/enable pattern that should be familiar to users of the clk
> > > subsystem.
> >
> > This is a lot to do in one go, and I think it's a bit too off on the
> > side and unintegrated with the core.  There's two very high level bits
> > here, there's the pre-cooking a message for offloading to be executed by
> > a hardware engine and there's the bit where that's triggered by some
> > hardwar event rather than by software.
> >
> > There was a bunch of discussion of the former case with David Jander
>
> I found [1] which appears to be the conversation you are referring to.
> Is that all or is there more that I missed?
>
> [1]: https://lore.kernel.org/linux-spi/20220512163445.6dcca126@erd992/
>
> > (CCed) a while back when he was doing all the work he did on optimising
> > the core for uncontended uses, the thinking there was to have a
> > spi_prepare_message() (or similar) API that drivers could call and then
> > reuse the same transfer repeatedly, and even without any interface for
> > client drivers it's likely that we'd be able to take advantage of it in
> > the core for multi-transfer messages.  I'd be surprised if there weren't
> > wins when the message goes over the DMA copybreak size.  A much wider
> > range of hardware would be able to do this bit, for example David's case
> > was a Raspberry Pi using the DMA controller to write into the SPI

For those, following along, it looks like the RPi business was
actually a 2013 discussion with Martin Sperl [2]. Both this and [1]
discuss proposed spi_prepare_message() APIs.

[2]: https://lore.kernel.org/linux-spi/CACRpkdb4mn_Hxg=3tuBu89n6eyJ082EETkwtNbzZDFZYTHbVVg@xxxxxxxxxxxxxx/T/#u