Re: [PATCH] spi: Avoid setting the chip select if we don't need to

["Ardelean, Alexandru" <alexandru.Ardelean@xxxxxxxxxx>]

On Mon, 2020-06-29 at 16:41 -0700, Douglas Anderson wrote:
> [External]
> 
> On some SPI controllers (like spi-geni-qcom) setting the chip select
> is a heavy operation.  For instance on spi-geni-qcom, with the current
> code, is was measured as taking upwards of 20 us.  Even on SPI
> controllers that aren't as heavy, setting the chip select is at least
> something like a MMIO operation over some peripheral bus which isn't
> as fast as a RAM access.
> 
> While it would be good to find ways to mitigate problems like this in
> the drivers for those SPI controllers, it can also be noted that the
> SPI framework could also help out.  Specifically, in some situations,
> we can see the SPI framework calling the driver's set_cs() with the
> same parameter several times in a row.  This is specifically observed
> when looking at the way the Chrome OS EC SPI driver (cros_ec_spi)
> works but other drivers likely trip it to some extent.
> 
> Let's solve this by caching the chip select state in the core and only
> calling into the controller if there was a change.  We check not only
> the "enable" state but also the chip select mode (active high or
> active low) since controllers may care about both the mode and the
> enable flag in their callback.

I think checkpatch suggested I be added here, since I touched some parts of
the delay/timings code.

> Signed-off-by: Douglas Anderson <dianders@xxxxxxxxxxxx>
> ---
> 
>  drivers/spi/spi.c       | 11 +++++++++++
>  include/linux/spi/spi.h |  4 ++++
>  2 files changed, 15 insertions(+)
> 
> diff --git a/drivers/spi/spi.c b/drivers/spi/spi.c
> index 6fa56590bba2..d4ba723a30da 100644
> --- a/drivers/spi/spi.c
> +++ b/drivers/spi/spi.c
> @@ -778,6 +778,17 @@ static void spi_set_cs(struct spi_device *spi, bool
> enable)
>  {
>  	bool enable1 = enable;
>  
> +	/*
> +	 * Avoid calling into the driver (or doing delays) if the chip
> select
> +	 * isn't actually changing from the last time this was called.
> +	 */
> +	if ((spi->controller->last_cs_enable == enable) &&
> +	    (spi->controller->last_cs_mode_high == (spi->mode &
> SPI_CS_HIGH)))
> +		return;
> +
> +	spi->controller->last_cs_enable = enable;
> +	spi->controller->last_cs_mode_high = spi->mode & SPI_CS_HIGH;
> +

I don't feel like this is the best approach for the SPI CS handling,
because it's pretty difficult to guess the last CS state, and whether this
return would cause other weirder issues [like not toggling CS when it
should].

Maybe a question is: when should this CS be toggled [or not]?
Is it between 2 calls of spi_transfer_one_message() or between 2
spi_transfers?
Or, is "xfer->cs_change == 1" where it shouldn't be?

I think, there are some ways to not toggle CS between some of these, or if
there aren't, some controls could be added/proposed to avoid toggling CS,
vs doing caching.

>  	if (!spi->controller->set_cs_timing) {
>  		if (enable1)
>  			spi_delay_exec(&spi->controller->cs_setup, NULL);
> diff --git a/include/linux/spi/spi.h b/include/linux/spi/spi.h
> index b4917df79637..0e67a9a3a1d3 100644
> --- a/include/linux/spi/spi.h
> +++ b/include/linux/spi/spi.h
> @@ -368,6 +368,8 @@ static inline void spi_unregister_driver(struct
> spi_driver *sdrv)
>   * @cur_msg_prepared: spi_prepare_message was called for the currently
>   *                    in-flight message
>   * @cur_msg_mapped: message has been mapped for DMA
> + * @last_cs_enable: was enable true on the last call to set_cs.
> + * @last_cs_mode_high: was (mode & SPI_CS_HIGH) true on the last call to
> set_cs.
>   * @xfer_completion: used by core transfer_one_message()
>   * @busy: message pump is busy
>   * @running: message pump is running
> @@ -604,6 +606,8 @@ struct spi_controller {
>  	bool				auto_runtime_pm;
>  	bool                            cur_msg_prepared;
>  	bool				cur_msg_mapped;
> +	bool				last_cs_enable;
> +	bool				last_cs_mode_high;
>  	bool                            fallback;
>  	struct completion               xfer_completion;
>  	size_t				max_dma_len;