Date:   Tue, 17 Dec 2019 09:35:39 +0100
From:   Luca Ceresoli <luca@...aceresoli.net>
To:     Daniel Mack <daniel@...que.org>, Wolfram Sang <wsa@...-dreams.de>
Cc:     linux-kernel@...r.kernel.org, linux-gpio@...r.kernel.org,
        linux-i2c@...r.kernel.org, alsa-devel@...a-project.org,
        devicetree@...r.kernel.org, linux-clk@...r.kernel.org,
        mturquette@...libre.com, sboyd@...nel.org, robh+dt@...nel.org,
        broonie@...nel.org, lee.jones@...aro.org, lars@...afoo.de,
        pascal.huerst@...il.com
Subject: Re: [PATCH 07/10] i2c: Add driver for AD242x bus controller

Hi Daniel,

On 15/12/19 21:27, Daniel Mack wrote:
> Hi,
> 
> Thanks for the review!
> 
> On 12/12/2019 5:33 pm, Wolfram Sang wrote:
>> Hi Luca,
>>
>> thanks for the review!
>>
>>> good, but I think there's a problem in this function. A "normal"
>>> master_xfer function issues a repeated start between one msg and the
>>> next one, at least in the typical case where all msgs have the same
>>> slave address. Your implementation breaks repeated start. At first sight
>>> we might need more complex code here to coalesce all consecutive msgs
>>> with the same address into a single i2c_transfer() call.
>>
>> Note that it is by far the standard case that all messages in a transfer
>> have the same client address (99,999%?). But technically, this is not a
>> requirement and the repeated start on the bus is totally independent of
>> the addresses used. It is just a master wanting to send without being
>> interrupted by another master.
> 
> I'm not quite sure I understand.
> 
> Let's assume the following setup. An i2c client (some driver code) is
> sending a list of messages to the a2b xfer function, which in turn is
> logically connected to a 'real' i2c bus master that'll put the data on
> the wire.
> 
> The a2b code has to tell the 'master node' the final destination of the
> payload by programming registers on its primary i2c address, and then
> forwards the messages to its secondary i2c address. The layout of the
> messages don't change, and neither do the flags; i2c messages are being
> sent as i2c messages, except their addresses are changed, a bit like NAT
> in networking. That procedure is described on page 3-4 of the TRM,
> "Remote Peripheral I2C Accesses".
> 
> The 'real' i2c master that handles the hardware bus is responsible for
> adding start conditions, and as the messages as such are untouched, I
> believe it should do the right thing. The code in my xfer functions
> merely suppresses reprogramming remote addresses by remembering the last
> one that was used, but that is independent of the start conditions on
> the wire.

My concern is not about the start condition, it's about the *repeated*
start condition.

The first question is whether the A2B chips can do it. What if the host
processor sets a slave chip address and then issues two messages
separated by a repeated start condition? Will the slave transceiver emit
a repeated start condition too?

If the answer is "yes", then the issue moves to the driver code. A
master xfer function receives a set of messages that are normally
emitted with a repeated start between each other. But ad242x_i2c_xfer()
splits the msgs and calls i2c_transfer_buffer_flags() with one msg at a
time. i2c_transfer_buffer_flags() then will emit a stop condition.

This is not necessarily a problem, unless multi-master is used, but if
there are limitations or deviations from the standard they should at
least be well known and documented.

-- 
Luca

Powered by blists - more mailing lists