Re: [PATCH 3/9] regmap: spi: SPI_CONTROLLER_CS_PER_TRANSFER affects max read/write

From: Mark Brown <broonie@kernel.org>
To: Lucas tanure <tanureal@opensource.cirrus.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
	"Rafael J . Wysocki" <rafael@kernel.org>,
	Sanjay R Mehta <sanju.mehta@amd.com>,
	Nehal Bakulchandra Shah <Nehal-Bakulchandra.shah@amd.com>,
	linux-kernel@vger.kernel.org, linux-spi@vger.kernel.org,
	patches@opensource.cirrus.com
Subject: Re: [PATCH 3/9] regmap: spi: SPI_CONTROLLER_CS_PER_TRANSFER affects max read/write
Date: Wed, 25 Aug 2021 18:21:25 +0100	[thread overview]
Message-ID: <20210825172125.GN5186@sirena.org.uk> (raw)
In-Reply-To: <4c604d13-f177-ff75-d21f-27613e1b763f@opensource.cirrus.com>

[-- Attachment #1: Type: text/plain, Size: 1897 bytes --]

On Wed, Aug 25, 2021 at 06:13:01PM +0100, Lucas tanure wrote:
> On 8/24/21 5:37 PM, Mark Brown wrote:

> > This should be handled by the SPI core, it's already relying on being
> > able to do multiple transfers to handle message size limits and in any
> > case this is a super standard thing to do so many clients would require

> For a message with N transfers how can spi core decide what to merge or what
> not merge. If mergers everything and is less than max_transfer_size success,

In the same way it does for transfers that are too long.  If the
controller has a property saying that it can't handle more than one
transfer then the core needs to either combine multiple transfers in a
single message into a single transfer or return an error to the caller
(modulo handling of cs_change).  If the controller can handle the
message it should just get passed straight through.

> but if bigger will need to stop merging and add an address in front of the
> next not merged transfer, but spi core is not aware of addresses
> And in the case of multiple addresses and data transfers, how it will know
> doesn't need to be merged?

The spi_message says what the message should look like on the bus.  The
semantics of what's in the message don't matter.  

> For me seems more reasonable for the regmap-spi stop splitting address
> and data. Or at least if the controller has some flag change the bus for
> one where it uses different functions for gather_write, async_write etc

This would force us to marshall the data in memory prior to sending
which adds overhead.

> Can you point which way you think the code should go? Investigate more spi
> core to coalesce transfers or change regmap-spi to not split address and
> data anymore?

Like I said in reply to your driver patch it looks like this
fundamentally doesn't do what you want in the first place.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]