[PATCH 2/2] gpio: mvebu: fix gpio bank registration when pwm is used

2017-05-30 17:14 GMT+02:00 Ralph Sennhauser [off-list ref]:

Hi Richard

On Tue, 30 May 2017 16:45:24 +0200
Richard Genoud [off-list ref] wrote:

quoted

2017-05-30 15:16 GMT+02:00 Gregory CLEMENT
[off-list ref]:

quoted

Hi Richard,

Hi Greg !

quoted

 On mar., mai 30 2017, Richard Genoud [off-list ref]
wrote:

quoted

If more than one gpio bank has the "pwm" property, only one will be
registered successfully, all the others will fail with:
mvebu-gpio: probe of f1018140.gpio failed with error -17

That's because in alloc_pwms(), the chip->base (aka "int pwm"),
was not set (thus, ==0) ; and 0 is a meaningful start value in
alloc_pwm(). What was intended is chip->base = -1.
Like that, the numbering will be done auto-magically

Tested on clearfog-pro (Marvell 88F6828)

Signed-off-by: Richard Genoud <redacted>
---
 drivers/gpio/gpio-mvebu.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/gpio/gpio-mvebu.c b/drivers/gpio/gpio-mvebu.c
index cdef2c78cb3b..4734923e11fd 100644
--- a/drivers/gpio/gpio-mvebu.c
+++ b/drivers/gpio/gpio-mvebu.c

@@ -768,6 +768,7 @@ static int mvebu_pwm_probe(struct

platform_device *pdev, mvpwm->chip.dev = dev;
      mvpwm->chip.ops = &mvebu_pwm_ops;
      mvpwm->chip.npwm = mvchip->chip.ngpio;
+     mvpwm->chip.base = -1;

Why not using
mvpwm->chip.base = id * MVEBU_MAX_GPIO_PER_BANK;
as it is done in the mvebu_gpio_probe() function?

Yes, that was my first move:
mvpwm->chip.base = mvchip->chip.base;

But after some reflexion, mvpwm->chip.base is not the GPIO base, it's
the PWM base,
(mvpwm->chip is a struct pwm_chip), so it would we weird to have
"holes" in the declared PWMs.
I'm not clear, so here's an example:
If, in the DTS, we have:
            gpio0: gpio at 18100 {
                compatible = "marvell,armada-370-xp-gpio",
                         "marvell,orion-gpio";
                reg = <0x18100 0x40>, <0x181c0 0x08>;
                reg-names = "gpio";  /* "pwm" missing */
[...]
            gpio1: gpio at 18140 {
                compatible = "marvell,armada-370-xp-gpio",
                         "marvell,orion-gpio";
                reg = <0x18140 0x40>, <0x181c8 0x08>;
                reg-names = "gpio", "pwm";
In this case, if gpio0 is not declared as PWM capable, the PWM
numbering will start at 32 if we have
mvpwm->chip.base = mvchip->chip.base;
but it will start at 0 if we have mvpwm->chip.base = -1;

The pros for having mvpwm->chip.base = mvchip->chip.base; is mainly
the stable numbering:
if we add the "pwm" feature to gpio0 afterwards, the pwm numbering in
sysfs will stay the same.
And if we have mvpwm->chip.base = -1; the pwm numbering will be
shifted.

Looking back at the V5 of this patch
https://www.spinics.net/lists/kernel/msg2484889.html
There was the line:
mvpwm->chip.base = mvchip->chip.base;
I guess it got lost in the v6 rebase.

So I could change it back, but I'm not sure which one is better.

Thierry Redding pointed out that the region might already be occupied
by another PWM chip, unlikely but not impossible. That's why it got
changed for v6.

See https://www.spinics.net/lists/devicetree/msg173138.html

Hum, I see.

But still, if mvpwm->chip.base is 0, we will have an error in alloc_pwms():

in pmvebu_pwm_probe(), mvpwm->chip.base is set to 0 (kzalloc), then
pwmchip_add(&mvpwm->chip) is called,
which will call alloc_pwms(chip->base/* == 0 */, chip->npwm);
The 1st time (1st gpio bank), it will be ok because:
start = bitmap_find_next_zero_area(allocated_pwms, MAX_PWMS, from, count, 0);
if (pwm >= 0 && start != pwm)
     return -EEXIST;
start is 0, pwm is 0, everything is good.
BUT, on the second time (for gpio1), we will have start = 32 and pwm=0
=> -EEXIST !

So there's something to fix.
And It seems that setting mvpwm->chip.base to -1 is a good choice
since it doesn't
force the numbering to something already taken.


Regards;
Richard.