[lkml]   [2013]   [Sep]   [23]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RESEND] spi/tegra114: Correct support for cs_change
On 9/23/2013 3:58 PM, Stephen Warren wrote:
> On 09/23/2013 01:48 PM, Rhyland Klein wrote:
>> On 9/23/2013 2:51 PM, Stephen Warren wrote:
>>> On 09/18/2013 12:17 PM, Rhyland Klein wrote:
>>>> The tegra114 driver wasn't currently handling the cs_change functionality.
>>>> It is meant to invert normal behavior, and we were only using it to possibly
>>>> delay at the end of a transfer.
> ...
>>>> diff --git a/drivers/spi/spi-tegra114.c b/drivers/spi/spi-tegra114.c
> ...
>>>> @@ -717,7 +718,12 @@ static int tegra_spi_start_transfer_one(struct spi_device *spi,
>>>> else if (req_mode == SPI_MODE_3)
>>>> command1 |= SPI_CONTROL_MODE_3;
>>>> - tegra_spi_writel(tspi, command1, SPI_COMMAND1);
>>>> + if (tspi->cs_control) {
>>>> + if (tspi->cs_control != spi)
>>>> + tegra_spi_writel(tspi, command1, SPI_COMMAND1);
>>> Do you really need a separate write call there? The value of command1
>>> isn't fully set up there (the CS bits are all set up immediately after),
>>> so won't that glitch the CS lines in some cases?
>> On our hardware (as far as I've seen), the CS line is normally low. We
> I assume you meant "normally *active* low", not "normally low"?

I mean that when I look at CS, before a transaction starts, the line is
low. Then we do a write like this (default SPI_COMMAND1) you see CS rise
and fall very soon after. This is purely how we generate the edge
triggers (as far as I can tell on all Tegra hw, though Laxman can
correct me if I am wrong).

>> need to generate a falling-edge to trigger the beginning of a SPI
>> transaction. Doing this write with the default value of SPI_COMMAND1
>> causes a brief rise and fall of CS, giving us our falling-edge.
> That sounds like exactly the glitch I was talking about.
> Assuming CS isn't held constantly asserted (low), isn't CS de-asserted
> (rises) at the end of transaction n-1, and re-asserted (falls) at the
> start of transaction n? If so, I'm not sure why the setup for
> transaction n needs to both de-assert and then re-assert it? It seems
> like cs_control should be handled at the end of a transaction, not at
> the start of the next one.

cs_change has to maintain state over spi_message boundries, not just
between spi_transfers within a spi_message.

In this specific case, this is a safe guard.

>>>> + if (tspi->cs_control) {
This sees that the previous transfer stored its spi_device pointer,
meaning it had cs_change set on the last part of the last spi_message
(I.E. CS hasn't been deasserted, so the SPI transaction is still on-going).

>>>> + if (tspi->cs_control != spi)
This however finds that the device trying to send a SPI message isn't
the same one that was in the middle of its transaction. This is a
collision between 2 clients on 1 bus. This simply ends the previous
transaction (the ongoing one) in favor of starting a new transaction.

Otherwise, this logic allows us to skip the spi of COMMAND1 which would
normally be used to create the falling edge to start a new transaction,
leaving the previous one open for more transfers.



 \ /
  Last update: 2013-09-23 23:21    [W:0.089 / U:12.724 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site