Re: [PATCH 1/2 v9] serial: mctrl_gpio: Check if GPIO property exisits before requesting it

From: Stefan Roese
Date: Tue Jun 25 2019 - 01:27:51 EST

Hi Geert,

On 24.06.19 17:35, Geert Uytterhoeven wrote:
On Mon, Jun 24, 2019 at 5:29 PM Stefan Roese <sr@xxxxxxx> wrote:
On 24.06.19 10:42, Geert Uytterhoeven wrote:
On Thu, Jun 20, 2019 at 8:24 AM Stefan Roese <sr@xxxxxxx> wrote:
This patch adds a check for the GPIOs property existence, before the
GPIO is requested. This fixes an issue seen when the 8250 mctrl_gpio
support is added (2nd patch in this patch series) on x86 platforms using

Here Mika's comments from 2016-08-09:

I noticed that with v4.8-rc1 serial console of some of our Broxton
systems does not work properly anymore. I'm able to see output but input
does not work.

I bisected it down to commit 4ef03d328769eddbfeca1f1c958fdb181a69c341
("tty/serial/8250: use mctrl_gpio helpers").

The reason why it fails is that in ACPI we do not have names for GPIOs
(except when _DSD is used) so we use the "idx" to index into _CRS GPIO
resources. Now mctrl_gpio_init_noauto() goes through a list of GPIOs
calling devm_gpiod_get_index_optional() passing "idx" of 0 for each. The
UART device in Broxton has following (simplified) ACPI description:

Device (URT4)
Name (_CRS, ResourceTemplate () {
GpioIo (Exclusive, PullDefault, 0x0000, 0x0000, IoRestrictionOutputOnly,
"\\_SB.GPO0", 0x00, ResourceConsumer)
GpioIo (Exclusive, PullDefault, 0x0000, 0x0000, IoRestrictionOutputOnly,
"\\_SB.GPO0", 0x00, ResourceConsumer)

In this case it finds the first GPIO (0x003A which happens to be RX pin
for that UART), turns it into GPIO which then breaks input for the UART
device. This also breaks systems with bluetooth connected to UART (those
typically have some GPIOs in their _CRS).

Any ideas how to fix this?

We cannot just drop the _CRS index lookup fallback because that would
break many existing machines out there so maybe we can limit this to
only DT enabled machines. Or alternatively probe if the property first
exists before trying to acquire the GPIOs (using

This patch implements the fix suggested by Mika in his statement above.

Signed-off-by: Stefan Roese <sr@xxxxxxx>

--- a/drivers/tty/serial/serial_mctrl_gpio.c
+++ b/drivers/tty/serial/serial_mctrl_gpio.c

@@ -116,6 +117,19 @@ struct mctrl_gpios *mctrl_gpio_init_noauto(struct device *dev, unsigned int idx)

for (i = 0; i < UART_GPIO_MAX; i++) {
enum gpiod_flags flags;
+ char *gpio_str;
+ bool present;
+ /* Check if GPIO property exists and continue if not */
+ gpio_str = kasprintf(GFP_KERNEL, "%s-gpios",
+ mctrl_gpios_desc[i].name);

This will silently break DTBs using "(cts|dsr|dcd|rng|rts|dtr)-gpio" instead
of "(cts|dsr|dcd|rng|rts|dtr)-gpios".

Should both options be supported ("cts-gpio" vs "cts-gpios")?
Documentation/devicetree/bindings/serial/serial.txt only mentions
the "-gpios" variant.

Well, the "-gpio" variant is deprecated, but still supported by
devm_gpiod_get_index_optional(), and there are active users in upstream
DTS files.

My main objection is (trying to) replicate the matching logic inside
gpiolib.c, causing subtle semantic differences. And keeping it consistent,
of course.

It would be nice if this could be fixed inside acpi_find_gpio(), so
users don't need to be updated. There may be other subsystems where
the difference between DT and ACPI may cause issues, unbeknownst.

Sure, I can fix this. I would prefer to do this in a follow-up patch
though, if nobody objects.