2014-10-28 17:37 GMT+03:00 Linus Walleij <linus.walleij@xxxxxxxxxx>:
> On Wed, Oct 15, 2014 at 1:12 PM, Sergey Ryazanov <ryazanov.s.a@xxxxxxxxx> wrote:
>> 2014-10-15 12:58 GMT+04:00, Linus Walleij <linus.walleij@xxxxxxxxxx>:
>>> On Sun, Sep 28, 2014 at 8:33 PM, Sergey Ryazanov <ryazanov.s.a@xxxxxxxxx>
>
>> I have one more generic question: could you merge driver without
>> GPIOLIB_IRQCHIP usage?
>
> No.
>
Ok.
>> Currently no one driver for the AR231x SoCs
>> uses irq_domain and I do not like to enable IRQ_DOMAIN just for one
>> driver. I plan to convert drivers to make them irq_domain aware a bit
>> later.
>
> I don't believe any such promises. It's nothing personal, just I've
> been burned too many times by people promising to "fix later".
>
Now I drop the driver from the series and return to the development a
bit later, when finished the basic code for the MIPS architecture. In
that case I will have a time to write the driver that does not require
further fixes.
>>>> +static u32 ar2315_gpio_intmask;
>>>> +static u32 ar2315_gpio_intval;
>>>> +static unsigned ar2315_gpio_irq_base;
>>>> +static void __iomem *ar2315_mem;
>>>
>>> No static locals. Allocate and use a state container, see
>>> Documentation/driver-model/design-patterns.txt
>>>
>> Is that rule mandatory for drivers, which serve only one device?
>
> There is no central authority which decides what is mandatory
> or not. It is mandatory to get a driver past the GPIO maintainer.
>
Nice point :)
>>>> +static inline u32 ar2315_gpio_reg_read(unsigned reg)
>>>> +{
>>>> + return __raw_readl(ar2315_mem + reg);
>>>> +}
>>>
>>> Use readl_relaxed() instead.
>>>
>> readl_relaxed() converts the bit ordering and seems inapplicable in this case.
>
> It assumes the peripherals IO memory is little endian.
>
> If the IO memory for this device is little endian, please stay with
> [readl|writel]_relaxed so it looks familiar.
>
> Or is this machine really using big endian hardware registers?
> In that case I understand your comment...
>
Yes, AR5312 and AR2315 SoCs are big endian machines with big endian registers.
>>> When you use the state container, you need to do a
>>> state dereference like that:
>>>
>>> mystate->base + reg
>>>
>>> So I don't think these inlines buy you anything. Just use
>>> readl/writel_relaxed directly in the code.
>>>
>> These helpers make code shorter and clearer. I can use macros if you
>> do preferred.
>
> No big deal. Keep it if you like it this way.
>
>>> So why is .set_type() not implemented and instead hard-coded into
>>> the unmask function? Please fix this. It will be called by the
>>> core eventually no matter what.
>>>
>> The interrupt configuration is a bit complex. This controller could be
>> configured to generate interrupts only for two lines at once. Or in
>> other words: user could select any two lines to generate interrupt.
>
> Oh well, better just handle it I guess...
>
Will do in v2.
>>>> +static int __init ar2315_gpio_init(void)
>>>> +{
>>>> + return platform_driver_register(&ar2315_gpio_driver);
>>>> +}
>>>> +subsys_initcall(ar2315_gpio_init);
>>>
>>> Why are you using subsys_initcall()?
>>>
>>> This should not be necessary.
>>>
>> I have users of GPIO in arch code, what called earlier than the
>> devices initcall.
>
> OK? Why are there such users that early and what do they
> use the GPIOs for? Any reason they cannot be device_initcall()s?
>
One GPIO line is used in reset handler to be able to reliably reset
the chip. This is a workaround from vendor's reference design to
eliminate a hw bug in the reset circuit of the AR2315 SoC. So I prefer
to have GPIO controller in ready state as soon as possible.
> Yours,
> Linus Walleij
--
BR,
Sergey
