[v2,3/5] gpio: msc313: MStar MSC313 GPIO driver

This adds a driver that supports the GPIO block found in
MStar/SigmaStar ARMv7 SoCs.

The controller seems to support 128 lines but where they
are wired up differs between chips and no currently known
chip uses anywhere near 128 lines so there needs to be some
per-chip data to collect together what lines actually have
physical pins attached and map the right names to them.

The core peripherals seem to use the same lines on the
currently known chips but the lines used for the sensor
interface, lcd controller etc pins seem to be totally
different between the infinity and mercury chips

The code tries to collect all of the re-usable names,
offsets etc together so that it's easy to build the extra
per-chip data for other chips in the future.

So far this only supports the MSC313 and MSC313E chips.

Support for the SSC8336N (mercury5) is trivial to add once
all of the lines have been mapped out.

Signed-off-by: Daniel Palmer <daniel@0x0f.com>
---
 MAINTAINERS                |   1 +
 drivers/gpio/Kconfig       |   9 +
 drivers/gpio/Makefile      |   1 +
 drivers/gpio/gpio-msc313.c | 406 +++++++++++++++++++++++++++++++++++++
 4 files changed, 417 insertions(+)
 create mode 100644 drivers/gpio/gpio-msc313.c

On Mon, Oct 19, 2020 at 5:11 PM Daniel Palmer <daniel@0x0f.com> wrote:
>
> This adds a driver that supports the GPIO block found in
> MStar/SigmaStar ARMv7 SoCs.
>
> The controller seems to support 128 lines but where they
> are wired up differs between chips and no currently known
> chip uses anywhere near 128 lines so there needs to be some
> per-chip data to collect together what lines actually have
> physical pins attached and map the right names to them.
>
> The core peripherals seem to use the same lines on the
> currently known chips but the lines used for the sensor
> interface, lcd controller etc pins seem to be totally
> different between the infinity and mercury chips
>
> The code tries to collect all of the re-usable names,
> offsets etc together so that it's easy to build the extra
> per-chip data for other chips in the future.
>
> So far this only supports the MSC313 and MSC313E chips.
>
> Support for the SSC8336N (mercury5) is trivial to add once
> all of the lines have been mapped out.

...

> +config GPIO_MSC313
> +       bool "MStar MSC313 GPIO support"

Why boolean?

> +       default y if ARCH_MSTARV7

Simply
       default ARCH_MSTARV7
should work as well.

Are you planning to extend this to other boards?

> +       depends on ARCH_MSTARV7
> +       select GPIOLIB_IRQCHIP
> +       help
> +         Say Y here to support GPIO on MStar MSC313 and later SoCs.

Please, be more specific. Also it's recommended to have a module name
to be included (but let's understand first why it's not a module)

...

> +/*
> + * Copyright (C) 2020 Daniel Palmer<daniel@thingy.jp>
> + */

One line.

...

> +#include <linux/of_device.h>
> +#include <linux/of_irq.h>
> +#include <linux/gpio/driver.h>

I believe this should be reworked.
For example, it misses mod_devicetable.h, bits.h, io.h, types.h, etc, but has

...

> +/* These bits need to be saved to correctly restore the
> + * gpio state when resuming from suspend to memory.
> + */

/*
 * For this subsystem the comment style for multi-line
 * like this.
 */

...

> +#ifdef CONFIG_MACH_INFINITY

Does it make any sense?

> +#endif

...

> +       return readb_relaxed(gpio->base + gpio->gpio_data->offsets[offset])
> +                       & MSC313_GPIO_IN;

Usual pattern is to leave operators on the previous line.

...

> +static struct irq_chip msc313_gpio_irqchip = {
> +       .name = "GPIO",

Is this name good enough?

> +       .irq_eoi = irq_chip_eoi_parent,
> +       .irq_mask = irq_chip_mask_parent,
> +       .irq_unmask = irq_chip_unmask_parent,
> +       .irq_set_type = irq_chip_set_type_parent,
> +};

...

> +static int msc313e_gpio_child_to_parent_hwirq(struct gpio_chip *chip,
> +                                            unsigned int child,
> +                                            unsigned int child_type,
> +                                            unsigned int *parent,
> +                                            unsigned int *parent_type)
> +{
> +       struct msc313_gpio *priv = gpiochip_get_data(chip);
> +       unsigned int offset = priv->gpio_data->offsets[child];
> +       int ret = -EINVAL;
> +
> +       /* only the spi0 pins have interrupts on the parent
> +        * on all of the known chips and so far they are all
> +        * mapped to the same place
> +        */

Comment style!

> +       if (offset >= OFF_SPI0_CZ && offset <= OFF_SPI0_DO) {
> +               *parent_type = child_type;
> +               *parent = ((offset - OFF_SPI0_CZ) >> 2) + 28;
> +               ret = 0;
> +       }
> +
> +       return ret;

Oh, can, for a starter, we use a regular (not-so-twisted) pattern

  if (...)
    return -EINVAL;
  ...
  return 0;

?

> +}

...

> +       gpio->saved = devm_kzalloc(&pdev->dev, gpio->gpio_data->num * sizeof(*gpio->saved), GFP_KERNEL);

devm_kcalloc()

> +       if (!gpio->saved)
> +               return -ENOMEM;

...

> +       res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> +       gpio->base = devm_ioremap_resource(&pdev->dev, res);

devm_platform_ioremap_resource()

> +       if (IS_ERR(gpio->base))
> +               return PTR_ERR(gpio->base);

...

> +       gpiochip->label = DRIVER_NAME;

Not good. When you use user space how do you distinguish if more than
one chip appears in the system?

...

> +       ret = gpiochip_add_data(gpiochip, gpio);
> +       return ret;

return ...(...);

Why not devm_gpiochip_add_data() ?

...

> +static const struct of_device_id msc313_gpio_of_match[] = {

> +#ifdef CONFIG_MACH_INFINITY

To me this makes no sense.

> +       {
> +               .compatible = "mstar,msc313-gpio",
> +               .data = &msc313_data,
> +       },
> +#endif
> +       { }
> +};

...

> +/* The GPIO controller loses the state of the registers when the
> + * SoC goes into suspend to memory mode so we need to save some
> + * of the register bits before suspending and put it back when resuming
> + */

Comment style!

> +

Redundant blank line.

...

> +static int __maybe_unused msc313_gpio_resume(struct device *dev)
> +{

> +}

> +

Redundant blank line.

> +static SIMPLE_DEV_PM_OPS(msc313_gpio_ops, msc313_gpio_suspend, msc313_gpio_resume);

...

> +static struct platform_driver msc313_gpio_driver = {
> +       .driver = {
> +               .name = DRIVER_NAME,
> +               .of_match_table = msc313_gpio_of_match,
> +               .pm = &msc313_gpio_ops,

You still allow to unbind.

> +       },
> +       .probe = msc313_gpio_probe,
> +};

> +

Redundant blank line.

> +builtin_platform_driver(msc313_gpio_driver);

Why?

Hi Andy,

On Tue, 20 Oct 2020 at 20:59, Andy Shevchenko <andy.shevchenko@gmail.com> wrote:
> > +config GPIO_MSC313
> > +       bool "MStar MSC313 GPIO support"
>
> Why boolean?

Because it's a built in driver. I could change it to tristate/a module
but I didn't think it needed to be one.
The machines this is used on generally only have 64 or 128MB of RAM so
the kernel is usually built without modules and only the totally
required stuff built in.

> > +       default y if ARCH_MSTARV7
>
> Simply
>        default ARCH_MSTARV7
> should work as well.
>
> Are you planning to extend this to other boards?

I think I copy/pasted the block above there. I'll fix this up.

As for other boards. I think this GPIO controller is only present in
MStar's SoCs and some MediaTek SoCs that inherited parts from MStar
after MediaTek bought them. Like the MStar interrupt controller is
present in some MediaTek TV chips.

>
> > +       depends on ARCH_MSTARV7
> > +       select GPIOLIB_IRQCHIP
> > +       help
> > +         Say Y here to support GPIO on MStar MSC313 and later SoCs.
>
> Please, be more specific. Also it's recommended to have a module name
> to be included (but let's understand first why it's not a module)

Ok. I'll rework that. As for it not being a module. I can make it
possible to build it
as a module. I just didn't really think it needed to be one.

> > +#include <linux/of_device.h>
> > +#include <linux/of_irq.h>
> > +#include <linux/gpio/driver.h>
>
> I believe this should be reworked.
> For example, it misses mod_devicetable.h, bits.h, io.h, types.h, etc, but has

I'll look at this again.

> > +/* These bits need to be saved to correctly restore the
> > + * gpio state when resuming from suspend to memory.
> > + */
>
> /*
>  * For this subsystem the comment style for multi-line
>  * like this.
>  */

Sorry, I'll fix these up.

> > +#ifdef CONFIG_MACH_INFINITY
>
> Does it make any sense?
>
> > +#endif

It doesn't make a lot of sense right now but it makes a bit more sense
when the support for the mercury chips is added. They have their own
set of offsets
for the used pins. I cut that out of this series because I haven't
fully reverse engineered all of the pins for those chips yet so the
support for them has a lot of guesses and notes in the code.

Anyhow, with only 64MB of RAM I thought it made sense to not compile
in the mercury tables when support for those machines isn't enabled.
I'll drop the if/defs for the next version.

> > +static struct irq_chip msc313_gpio_irqchip = {
> > +       .name = "GPIO",
>
> Is this name good enough?
>

There is only one GPIO block in the chips this is for so I think it's
unique at least.

> > +       gpiochip->label = DRIVER_NAME;
>
> Not good. When you use user space how do you distinguish if more than
> one chip appears in the system?

So far there is only ever one of these GPIO controller for the whole system.
There is another GPIO block in the system but it uses another driver
as the register layout is totally different.

Thank you for all of the comments. Sorry about the multiple style issues.
I thought checkpatch had my back on that.

Thanks,

Daniel

On Tue, Oct 20, 2020 at 1:59 PM Andy Shevchenko
<andy.shevchenko@gmail.com> wrote:
> On Mon, Oct 19, 2020 at 5:11 PM Daniel Palmer <daniel@0x0f.com> wrote:

> > +config GPIO_MSC313
> > +       bool "MStar MSC313 GPIO support"

> > +       default y if ARCH_MSTARV7

I think it is possible to write:

default ARCH_MSTARV7

For this.

> Why boolean?

I answered this question in some other mail but there is usually somewhat
a good reason for this but let's discuss it a bit.

The following usecase:

- You have a generic distribution such as Android

- The generic distribution does not use initramfs to support basic
  drivers (does Android? Nowadays?)

- So all modules would need to load from actual storage media.

- The storage media is an SD card.

- The SD card reader has a card detect line connected to one
  of the GPIO lines.

- Now we have catch-22 because there is no way to load the
  GPIO driver modules from the storage media because the
  driver is needed to mount the storage media.

I do not know if this "never happens" because every generic
distribution "should" be using initramfs for their drivers. But it
provides a convenient way for users to shoot themselves in the
foot and be frustrated about that their root filesystem is not
mounting.

I do not think this is limited to GPIO card detects but it is a very
immediate example.

What is the consensus here?

Yours,
Linus Walleij

On Mon, Oct 19, 2020 at 4:10 PM Daniel Palmer <daniel@0x0f.com> wrote:

> This adds a driver that supports the GPIO block found in
> MStar/SigmaStar ARMv7 SoCs.
>
> The controller seems to support 128 lines but where they
> are wired up differs between chips and no currently known
> chip uses anywhere near 128 lines so there needs to be some
> per-chip data to collect together what lines actually have
> physical pins attached and map the right names to them.
>
> The core peripherals seem to use the same lines on the
> currently known chips but the lines used for the sensor
> interface, lcd controller etc pins seem to be totally
> different between the infinity and mercury chips
>
> The code tries to collect all of the re-usable names,
> offsets etc together so that it's easy to build the extra
> per-chip data for other chips in the future.
>
> So far this only supports the MSC313 and MSC313E chips.
>
> Support for the SSC8336N (mercury5) is trivial to add once
> all of the lines have been mapped out.
>
> Signed-off-by: Daniel Palmer <daniel@0x0f.com>

This looks really nice, the generic hierarchical IRQchip does
its job very well here.

I saw you would send another version but this already looks
like merge material.

Certainly any remaining issues can be ironed out in-tree.

> +config GPIO_MSC313
> +       bool "MStar MSC313 GPIO support"
> +       default y if ARCH_MSTARV7
> +       depends on ARCH_MSTARV7
> +       select GPIOLIB_IRQCHIP

select IRQ_DOMAIN_HIERARCHY

since you are dependent on this.

(Else people will soon report problems with randconfig...)

> +/* The parent interrupt controller needs the GIC interrupt type set to GIC_SPI
> + * so we need to provide the fwspec. Essentially gpiochip_populate_parent_fwspec_twocell
> + * that puts GIC_SPI into the first cell.
> + */
> +static void *msc313_gpio_populate_parent_fwspec(struct gpio_chip *gc,
> +                                            unsigned int parent_hwirq,
> +                                            unsigned int parent_type)
> +{
> +       struct irq_fwspec *fwspec;
> +
> +       fwspec = kmalloc(sizeof(*fwspec), GFP_KERNEL);
> +       if (!fwspec)
> +               return NULL;
> +
> +       fwspec->fwnode = gc->irq.parent_domain->fwnode;
> +       fwspec->param_count = 3;
> +       fwspec->param[0] = GIC_SPI;
> +       fwspec->param[1] = parent_hwirq;
> +       fwspec->param[2] = parent_type;
> +
> +       return fwspec;
> +}

Clever. Looping in Marc Z so he can say if this looks allright to him.

> +static int msc313e_gpio_child_to_parent_hwirq(struct gpio_chip *chip,
> +                                            unsigned int child,
> +                                            unsigned int child_type,
> +                                            unsigned int *parent,
> +                                            unsigned int *parent_type)
> +{
> +       struct msc313_gpio *priv = gpiochip_get_data(chip);
> +       unsigned int offset = priv->gpio_data->offsets[child];
> +       int ret = -EINVAL;
> +
> +       /* only the spi0 pins have interrupts on the parent
> +        * on all of the known chips and so far they are all
> +        * mapped to the same place
> +        */
> +       if (offset >= OFF_SPI0_CZ && offset <= OFF_SPI0_DO) {
> +               *parent_type = child_type;
> +               *parent = ((offset - OFF_SPI0_CZ) >> 2) + 28;
> +               ret = 0;
> +       }
> +
> +       return ret;
> +}

Neat!

> +static int msc313_gpio_probe(struct platform_device *pdev)
> +{
> +       const struct msc313_gpio_data *match_data;
> +       struct msc313_gpio *gpio;
> +       struct gpio_chip *gpiochip;
> +       struct gpio_irq_chip *gpioirqchip;
> +       struct resource *res;
> +       struct irq_domain *parent_domain;
> +       struct device_node *parent_node;
> +       int ret;
> +
> +       match_data = of_device_get_match_data(&pdev->dev);

There is a lot of referencing &pdev->dev.

I would add a local variable like this:

struct device *dev = &pdev->dev

and replace all &pdev->dev with that. It will make the code more
compact and easier to read.

Yours,
Linus Walleij

On 2020-11-05 09:40, Linus Walleij wrote:
> On Mon, Oct 19, 2020 at 4:10 PM Daniel Palmer <daniel@0x0f.com> wrote:

[...]

>> +/* The parent interrupt controller needs the GIC interrupt type set 
>> to GIC_SPI
>> + * so we need to provide the fwspec. Essentially 
>> gpiochip_populate_parent_fwspec_twocell
>> + * that puts GIC_SPI into the first cell.
>> + */

nit: comment style.

>> +static void *msc313_gpio_populate_parent_fwspec(struct gpio_chip *gc,
>> +                                            unsigned int 
>> parent_hwirq,
>> +                                            unsigned int parent_type)
>> +{
>> +       struct irq_fwspec *fwspec;
>> +
>> +       fwspec = kmalloc(sizeof(*fwspec), GFP_KERNEL);
>> +       if (!fwspec)
>> +               return NULL;
>> +
>> +       fwspec->fwnode = gc->irq.parent_domain->fwnode;
>> +       fwspec->param_count = 3;
>> +       fwspec->param[0] = GIC_SPI;
>> +       fwspec->param[1] = parent_hwirq;
>> +       fwspec->param[2] = parent_type;
>> +
>> +       return fwspec;
>> +}
> 
> Clever. Looping in Marc Z so he can say if this looks allright to him.

Yup, this looks correct. However, looking at the bit of the patch that
isn't quoted here, I see that msc313_gpio_irqchip doesn't have a
.irq_set_affinity callback. Is this system UP only?

Thanks,

         M.

Hi Marc,

On Thu, 5 Nov 2020 at 21:08, Marc Zyngier <maz@kernel.org> wrote:
>

> On 2020-11-05 09:40, Linus Walleij wrote:

> > On Mon, Oct 19, 2020 at 4:10 PM Daniel Palmer <daniel@0x0f.com> wrote:

>

> [...]

>

> >> +/* The parent interrupt controller needs the GIC interrupt type set

> >> to GIC_SPI

> >> + * so we need to provide the fwspec. Essentially

> >> gpiochip_populate_parent_fwspec_twocell

> >> + * that puts GIC_SPI into the first cell.

> >> + */

>

> nit: comment style.


I've fixed these and some other bits for the v3.
I've held off on pushing that until the rest of it seemed right.

> >> +static void *msc313_gpio_populate_parent_fwspec(struct gpio_chip *gc,

> >> +                                            unsigned int

> >> parent_hwirq,

> >> +                                            unsigned int parent_type)

> >> +{

> >> +       struct irq_fwspec *fwspec;

> >> +

> >> +       fwspec = kmalloc(sizeof(*fwspec), GFP_KERNEL);

> >> +       if (!fwspec)

> >> +               return NULL;

> >> +

> >> +       fwspec->fwnode = gc->irq.parent_domain->fwnode;

> >> +       fwspec->param_count = 3;

> >> +       fwspec->param[0] = GIC_SPI;

> >> +       fwspec->param[1] = parent_hwirq;

> >> +       fwspec->param[2] = parent_type;

> >> +

> >> +       return fwspec;

> >> +}

> >

> > Clever. Looping in Marc Z so he can say if this looks allright to him.

>

> Yup, this looks correct. However, looking at the bit of the patch that

> isn't quoted here, I see that msc313_gpio_irqchip doesn't have a

> .irq_set_affinity callback. Is this system UP only?


What is in mainline right now is UP only but there are chips with a
second cortex A7 that I have working in my tree.
So I will add that in for v3 if I can work out what I should actually
do there. :)

Thanks,

Daniel

On 2020-11-05 15:23, Daniel Palmer wrote:
> Hi Marc,

> 

> On Thu, 5 Nov 2020 at 21:08, Marc Zyngier <maz@kernel.org> wrote:

>> 

>> On 2020-11-05 09:40, Linus Walleij wrote:

>> > On Mon, Oct 19, 2020 at 4:10 PM Daniel Palmer <daniel@0x0f.com> wrote:

>> 

>> [...]

>> 

>> >> +/* The parent interrupt controller needs the GIC interrupt type set

>> >> to GIC_SPI

>> >> + * so we need to provide the fwspec. Essentially

>> >> gpiochip_populate_parent_fwspec_twocell

>> >> + * that puts GIC_SPI into the first cell.

>> >> + */

>> 

>> nit: comment style.

> 

> I've fixed these and some other bits for the v3.

> I've held off on pushing that until the rest of it seemed right.

> 

>> >> +static void *msc313_gpio_populate_parent_fwspec(struct gpio_chip *gc,

>> >> +                                            unsigned int

>> >> parent_hwirq,

>> >> +                                            unsigned int parent_type)

>> >> +{

>> >> +       struct irq_fwspec *fwspec;

>> >> +

>> >> +       fwspec = kmalloc(sizeof(*fwspec), GFP_KERNEL);

>> >> +       if (!fwspec)

>> >> +               return NULL;

>> >> +

>> >> +       fwspec->fwnode = gc->irq.parent_domain->fwnode;

>> >> +       fwspec->param_count = 3;

>> >> +       fwspec->param[0] = GIC_SPI;

>> >> +       fwspec->param[1] = parent_hwirq;

>> >> +       fwspec->param[2] = parent_type;

>> >> +

>> >> +       return fwspec;

>> >> +}

>> >

>> > Clever. Looping in Marc Z so he can say if this looks allright to him.

>> 

>> Yup, this looks correct. However, looking at the bit of the patch that

>> isn't quoted here, I see that msc313_gpio_irqchip doesn't have a

>> .irq_set_affinity callback. Is this system UP only?

> 

> What is in mainline right now is UP only but there are chips with a

> second cortex A7 that I have working in my tree.

> So I will add that in for v3 if I can work out what I should actually

> do there. :)


Probably nothing more than setting the callback to 
irq_chip_set_affinity_parent,
I'd expect.

         M.
-- 
Jazz is not dead. It just smells funny...

On Thu, Nov 5, 2020 at 4:43 PM Marc Zyngier <maz@kernel.org> wrote:
> On 2020-11-05 15:23, Daniel Palmer wrote:

> > On Thu, 5 Nov 2020 at 21:08, Marc Zyngier <maz@kernel.org> wrote:


> > >  I see that msc313_gpio_irqchip doesn't have a

> >> .irq_set_affinity callback. Is this system UP only?

> >

> > What is in mainline right now is UP only but there are chips with a

> > second cortex A7 that I have working in my tree.

> > So I will add that in for v3 if I can work out what I should actually

> > do there. :)

>

> Probably nothing more than setting the callback to

> irq_chip_set_affinity_parent,


Hm, is this something all GPIO irqchips used on SMP systems
should be doing? Or just hierarchical ones?

Yours,
Linus Walleij

On 2020-11-10 14:02, Linus Walleij wrote:
> On Thu, Nov 5, 2020 at 4:43 PM Marc Zyngier <maz@kernel.org> wrote:

>> On 2020-11-05 15:23, Daniel Palmer wrote:

>> > On Thu, 5 Nov 2020 at 21:08, Marc Zyngier <maz@kernel.org> wrote:

> 

>> > >  I see that msc313_gpio_irqchip doesn't have a

>> >> .irq_set_affinity callback. Is this system UP only?

>> >

>> > What is in mainline right now is UP only but there are chips with a

>> > second cortex A7 that I have working in my tree.

>> > So I will add that in for v3 if I can work out what I should actually

>> > do there. :)

>> 

>> Probably nothing more than setting the callback to

>> irq_chip_set_affinity_parent,

> 

> Hm, is this something all GPIO irqchips used on SMP systems

> should be doing? Or just hierarchical ones?


Probably only the hierarchical ones. I'd expect the non-hierarchical
GPIOs to be muxed behind a single interrupt, which makes it impossible
to move a single GPIO around, and moving the mux interrupt would break
userspace's expectations that interrupts move independently of each 
others.

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...

On 2020-11-11 14:09, Linus Walleij wrote:
> On Tue, Nov 10, 2020 at 3:19 PM Marc Zyngier <maz@kernel.org> wrote:

>> On 2020-11-10 14:02, Linus Walleij wrote:

> 

>> >> Probably nothing more than setting the callback to

>> >> irq_chip_set_affinity_parent,

>> >

>> > Hm, is this something all GPIO irqchips used on SMP systems

>> > should be doing? Or just hierarchical ones?

>> 

>> Probably only the hierarchical ones. I'd expect the non-hierarchical

>> GPIOs to be muxed behind a single interrupt, which makes it impossible

>> to move a single GPIO around, and moving the mux interrupt would break

>> userspace's expectations that interrupts move independently of each

>> others.

> 

> I found two suspects and sent patches. I think I might have some

> more candidates down in pinctrl. I do have some hierarchical IRQ

> that is on UP systems, I suppose these are not affected.


Yup, they look good. Feel free to add my Ack to them.
And yes, UP systems are naturally oblivious of interrupt affinity.

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...