On 11/15/2016 03:32 AM, Vladimir Zapolskiy wrote:

On 11/15/2016 01:10 PM, Vladimir Zapolskiy wrote:

quoted

Hello Heiko,

On 11/15/2016 12:20 PM, Heiko Schocher wrote:

quoted

Hello,

commit e188cbf7564f: "gpio: mxc: shift gpio_mxc_init() to subsys_initcall level"
moves the gpio initialization of the mxc gpio driver
from the arch_initcall level into subsys_initcall level.

This leads now on mxc boards, which use a gpio wdt driver
and the CONFIG_GPIO_WATCHDOG_ARCH_INITCALL option enabled,
to unwanted driver probe deferrals during kernel boot.

I see this currently on an imx6 based board (which has unfortunately
3 WDT: imx6 internal (disabled), gpio wdt and da9063 WDT ...).

Also a side effect from above commit is, that the da9063 WDT driver
is now probed before the gpio WDT driver ... so /dev/watchdog now
does not point to the gpio_wdt, instead it points to the da9063 WDT.

So there are 2 solutions possible:

- add a CONFIG_GPIO_MCX_ARCH_INITCALL option
  in drivers/gpio/gpio-mxc.c like for the gpio_wdt.c driver?

in my opinion this is overly heavy solution and it might be
better to avoid it if possible.

I would rather prefer to reconsider GPIO_WATCHDOG_ARCH_INITCALL
usage in the watchdog driver.

Moreover adding this proposed GPIO_MCX_ARCH_INITCALL to call
the driver on arch level will result in deferring the GPIO driver.

quoted

  But how can we guarantee, that first the gpio driver and then
  the gpio_wdt driver gets probed?

- move the arch_initcall in gpio_wdt.c into a subsys_initcall
  (Tested this, and the probe dereferral messages are gone ...)

  But this may results in problems on boards, which needs an early
  trigger on an gpio wdt ...

The level of "earliness" can not be defined in absolute time value
in any case, why decreasing the init level of the watchdog driver
to subsys level can cause problems? For that there should exist
some kind of a dependency on IC or PCB hardware level, can you
name it please?

Also please note that more than a half of all GPIO drivers settle
on subsys or later initcall level, this means that there is
an expected GPIO watchdog driver deferral for all of them.

Please find two more late notes though.

quoted

I propose to send two patches for review:

1. remove GPIO_WATCHDOG_ARCH_INITCALL option completely and decouple
   module_platform_driver() into arch_initcall() and module_exit()
   unconditionally.

2. change arch_initcall() in the watchdog driver to subsys_initcall().
   This change removes probe deferrals on boot, when the driver is
   used with the most of the GPIO controllers.

Alternatively commit 5e53c8ed813d ("watchdog: gpio_wdt: Add option for
early registration") can be reverted and then module_platform_driver()
is decoupled into subsys_initcall() and module_exit() as its replacement.

Sure, only the reason for that was that there are situations where
subsys_initcall() was too late. Also, when using arch_initcall() only,
we get deferrals again, which is apparently hated by many and a reason
for all those "avoid probe deferrals" patches.

And also please note that since quite many GPIO controller drivers
live on initcall levels after subsys_initcall(), the solution won't
let to avoid watchdog driver deferrals totally, this should be accepted.

... except for others it isn't, and we are back to square one.

GPIO_WATCHDOG_ARCH_INITCALL was intended to be only used in situations
where needed. Why is it used here in the first place if that is not
the case ?

Guenter