Hi doug,
Thanx for your comments.
On 06/23/2017 05:41 AM, Doug Anderson wrote:
Hi,
On Tue, Jun 13, 2017 at 8:38 PM, Jeffy Chen <jeffy.c...@rock-chips.com> wrote:
Support using "cs-gpios" property to specify cs gpios.
Signed-off-by: Jeffy Chen <jeffy.c...@rock-chips.com>
---
Changes in v3:
include linux/gpio/consumer.h for compile errors on ARCH_X86
(reported by kbuild test robot <l...@intel.com>)
Changes in v2:
1/ request cs gpios in probe for better error handling
2/ use gpiod* function
(suggested by Heiko Stuebner)
3/ split dt-binding changes to new patch
(suggested by Shawn Lin & Heiko Stuebner)
drivers/spi/spi-rockchip.c | 31 ++++++++++++++++++++++++++++++-
1 file changed, 30 insertions(+), 1 deletion(-)
diff --git a/drivers/spi/spi-rockchip.c b/drivers/spi/spi-rockchip.c
index bab9b13..4bcf251 100644
--- a/drivers/spi/spi-rockchip.c
+++ b/drivers/spi/spi-rockchip.c
@@ -16,7 +16,8 @@
#include <linux/clk.h>
#include <linux/dmaengine.h>
#include <linux/module.h>
-#include <linux/of.h>
+#include <linux/gpio/consumer.h>
+#include <linux/of_gpio.h>
#include <linux/pinctrl/consumer.h>
#include <linux/platform_device.h>
#include <linux/spi/spi.h>
@@ -663,6 +664,27 @@ static bool rockchip_spi_can_dma(struct spi_master *master,
return (xfer->len > rs->fifo_len);
}
+static int rockchip_spi_setup_cs_gpios(struct device *dev)
+{
+ struct device_node *np = dev->of_node;
+ struct gpio_desc *cs_gpio;
+ int i, nb;
+
+ if (!np)
+ return 0;
Not sure you really to check for NULL "np". Do we really run properly
without device tree? We already call of_property_read_u32()
unconditionally...
hmm, right
+
+ nb = of_gpio_named_count(np, "cs-gpios");
+ for (i = 0; i < nb; i++) {
Implicitly if there is any error getting "cs-gpios" (AKA if it doesn't
exist) you'll return a negative value here, then return "0" for the
function. AKA cs-gpios is optional... The behavior is correct, but
it's a bit non-obvious. Personally I would have at least put a
comment even if you didn't put an explicit check.
ok
+ /* We support both GPIO CS and HW CS */
+ cs_gpio = devm_gpiod_get_index_optional(dev, "cs",
+ i, GPIOD_ASIS);
+ if (IS_ERR(cs_gpio))
+ return PTR_ERR(cs_gpio);
As per your discussion with Brian, your whole reason for having this
function is that:
1. Core SPI framework treats errors getting the GPIO as non-fatal (SPI
framework falls back on using the HW chip select).
Mark is the expert, but IMHO that seems like a bug in the core SPI
framework and you should fix it there rather than hacking around in
the driver. _In theory_ you could break backward compatibility
(someone could have been relying on the old behavior that an error
caused you to fallback to the HW chip select), but I think that's not
likely as long as you handle things like:
cs-gpios = <&gpio1 0 0>, <0>, <&gpio1 1 0>, <&gpio1 2 0>;
AKA if someone has explicitly specified <0> for the GPIO then _that_
shouldn't be an error and we should do the fallback to HW chip select.
If we really expect old buggy DTS files that get broken by the old
behavior then we'd have to ask for advice from Mark and/or device tree
experts...
right, it would be good to be handled in the spi core.
and i think devm_gpiod_get_index_optional would take care of the <0>
fallback case
As evidence that the current SPI core is broken in the way it is and
could use a patch, it would be easy to "fall back" to a chip select
that's greater than "master->num_chipselect", which seems to me like a
clear bug.
--
2. The SPI framework doesn't end up calling gpiod_request(). It seems
like it ought to. Requiring the sub driver to do this seems wrong.
+ }
+
+ return 0;
+}
+
static int rockchip_spi_probe(struct platform_device *pdev)
{
int ret = 0;
@@ -749,6 +771,7 @@ static int rockchip_spi_probe(struct platform_device *pdev)
master->transfer_one = rockchip_spi_transfer_one;
master->max_transfer_size = rockchip_spi_max_transfer_size;
master->handle_err = rockchip_spi_handle_err;
+ master->flags = SPI_MASTER_GPIO_SS;
IMHO this one line in your patch makes total sense and it seems like
you could post it by itself and it could land. All the error check
and gpiod_request() bikeshedding could be deferred to a separate
patch.
ok, that make sense, will do in next version.
