Re: [PATCH v2] gpio: winbond: add driver

[Maciej S. Szmigiero]

On 29.12.2017 11:27, Andy Shevchenko wrote:
> On Fri, 2017-12-29 at 00:44 +0100, Maciej S. Szmigiero wrote:
>> On 28.12.2017 16:12, Andy Shevchenko wrote:
>>> On Fri, 2017-12-22 at 19:58 +0100, Maciej S. Szmigiero wrote:
>>>>
> 
>>>> +	  You will need to provide a module parameter "gpios", or
>>>> a
>>>> +	  boot-time parameter "gpio_winbond.gpios" with a bitmask
>>>> of
>>>> GPIO
>>>> +	  ports to enable (bit 0 is GPIO1, bit 1 is GPIO2, etc.).
>>>
>>> 1. Why it's required?
>>
>> It is required bacause "[o]ne should be careful which ports one
>> tinkers
>> with since some might be managed by the firmware (for functions like
>> powering on and off, sleeping, BIOS recovery, etc.) and some of GPIO
>> port
>> pins are physically shared with other devices included in the Super
>> I/O
>> chip".
> 
> I would like to be clear, I was asking about module parameters. Nowadays
> we won't have new parameters to the kernel.
> 
> Is there any strong argument to do it so? Is it one above? Can we detect
> as much as possible run time?

These GPIO devices are not described anywhere by the platform AFAIK.
It is also very likely that these particular devices and their pins that
aren't used internally by the firmware aren't even configured correctly.

That's why we don't really have a general ability to detect which GPIO
device(s) should be configured and how.

Other issue is that random hacking and reverse engineering of
motherboard signals (SIO GPIO lines are often used internally to
implement switching of various motherboard functions, like LED control,
main / backup BIOS swapping, etc.) will probably represent the most
common use case for this driver so its settings should be changeable
without needing to modify the kernel.

Don't think about this driver like a ready driver for some platform,
like a driver for a Foo motherboard lights or a Bar laptop extra keys.
In such cases a dedicated driver should be written (which would then
interact with LED and input subsystems, respectively).

>>> 2. GPIO1 -> GPIO0, GPIO2 -> GPIO1, etc ?
>>
>> The chip datasheet calls GPIO devices and their pins "GPIO1", "GPIO2",
>> etc., however the driver uses a more common zero-based indexing (so,
>> for example, we don't waste the zeroth bit).
> 
> I dunno what makes less confusion here.
> 
> One way, is to provide labels according to data sheet (not so flexible,
> though), another waste 0th bit for good.
> 
> Linus, what's your opinion?
>
(..)
>>>> +#define WB_SIO_DEV_WDGPIO56 8
>>>> +#define WB_SIO_WDGPIO56_REG_ENABLE 0x30
>>>
>>> Why do we have duplication here?
>>
>> Registers with offset >= 0x30 are
>> specific for a particular device.
>>
>> That's a register in a different device
>> (which happen to have similar function as
>> register 0x30 in, for example, UARTC but
>> nothing in the datasheet guarantees that
>> such mapping will be universal).
> 
> OK, then can you put a comment like this one before this very definition
> block (with offsets >= 0x30) and put a one line comment before each
> *different* device.

Ok.

>>>> +static u8 gpios;
>>>> +static u8 ppgpios;
>>>> +static u8 odgpios;
>>>> +static bool pledgpio;
>>>> +static bool beepgpio;
>>>> +static bool i2cgpio;
>>>
>>> Hmm... Too many global variables.
>>
>> All of them, besides i2cgpio, are module parameters, which require
>> global variables.
> 
> Same question as on top of this mail.
> Why do we need all of them?

I've responded above to this module parameters issue.

>>>> +static void winbond_sio_reg_write(u16 base, u8 reg, u8 data)
>>>> +static u8 winbond_sio_reg_read(u16 base, u8 reg)
>>>> +static void winbond_sio_reg_bset(u16 base, u8 reg, u8 bit)
>>>> +static void winbond_sio_reg_bclear(u16 base, u8 reg, u8 bit)
>>>> +static bool winbond_sio_reg_btest(u16 base, u8 reg, u8 bit)
>>>
>>> regmap API?
>>
>> Looking at the regmap API:
>> * It does not support a I/O port space,
> 
> Not an issue, you may provide your own IO accessors.
I meant here that there is no ready regmap support for the I/O port
space like there is for AC'97, I2C, MMIO, SPI, SPMI and W1, sorry if I
wasn't clear enough.

>> * It does not support a shared (muxed) register space, where one needs
>> to first request the space for a particular driver, then can perform
>> required sequence of operations, then should release the space,
> 
> If it's an MFD, it would be done at MFD level and shared across devices.
> Also, there is no problem to remap same IO space as many times as
> driver(s) want(s) to.
> 
>> * It does not support multiple logical devices sharing a register
>> address space where one needs first to select the logical device, then
>> perform the required sequence of operations.
> 
> We have MFD for precisely that kind of devices.

Looks like MFD infrastructure isn't generally used for Super I/Os:
it seems that currently no driver for this type of hardware uses it,
despite there being a lot of Super I/O hwmon and watchdog drivers in
the kernel tree (and even two existing GPIO ones).

Also, w83627ehf hwmon and w83627hf_wdt watchdog drivers (which access
the same hardware as this driver) would need to be rewritten to use
this infrastructure, too.

Overall, I think it is just an overkill for such a small driver.

>>>> +	for (i = 0; i < ARRAY_SIZE(winbond_gpio_infos); i++) {
>>>> +		if (!(gpios & BIT(i)))
>>>> +			continue;
>>>
>>> for_each_set_bit()
>>
>> Do we really want to replace a simple 6-iteration "for" loop with a
>> one that (possibly) does a function call at setup time then an another
>> per iteration?
> 
> Yes. Rationale is:
> - makes less LOCs
> - makes code more readable and understandble at a glance
> - unloads optimization stuff to compiler's shoulders.

Ok.

>>
>>>> +
>>>> +		if (gpio_num < 8)
>>>> +			break;
>>>> +
>>>> +		gpio_num -= 8;
>>>
>>> Yeah, consider to use % and / paired, see below.
>>
>> Here it is only a simple subtraction of 8 lines per each set bit in
>> 'gpios', how do you suggest to replace it with a faster
>> division-with-reminder operation (I guess it isn't supposed to be 
>> "if (gpio_num / 8 == 0) break;")?
> 
> Looking above and below, I actually missed that what you need is
> hweight().

Here hweight() felt like a poor fit, but it was used in the code below.

>>>> +	if (WARN_ON(i >= ARRAY_SIZE(winbond_gpio_infos)))
>>>> +		i = 0;
>>>
>>> Something is even more seriously wrong if you are going to mess with
>>> GPIO 0.
>>>
>>> You have to return an error here.
>>>
>>> Although it should not happen at all, AFAIU.
>>
>> Yes, this condition should never happen in principle (it's only a
>> defensive programming check) so I think that printing a warning and
>> letting it read the first GPIO that is present while disallowing write
>> access would be fine in this very unlikely situation.
> 
> I think it's just a waste of effort. If this happens it happens to all
> GPIO drivers, otherwise it might be some HW issue with memory, for
> example, that again makes entire system unstable.
> 
> % git grep -n WARN -- drivers/gpio/ | cut -f1 -d: | sort -u
> drivers/gpio/devres.c
> drivers/gpio/gpio-aspeed.c
> drivers/gpio/gpio-brcmstb.c
> drivers/gpio/gpio-bt8xx.c
> drivers/gpio/gpio-davinci.c
> drivers/gpio/gpio-grgpio.c
> drivers/gpio/gpiolib-acpi.c
> drivers/gpio/gpiolib.c
> drivers/gpio/gpiolib-of.c
> drivers/gpio/gpio-omap.c
> drivers/gpio/gpio-tegra186.c
> drivers/gpio/gpio-thunderx.c
> drivers/gpio/gpio-twl4030.c
> drivers/gpio/gpio-tz1090.c
> drivers/gpio/gpio-uniphier.c
> drivers/gpio/gpio-zynq.c
> 
> Basically above list minus 4 library files (and perhaps even less to
> care about similar "should never happen" cases).
> 
> $ ls -1 drivers/gpio/ | wc -l
> 147

I've removed this check now but I cannot say I agree that checks like
this are just a waste of effort: in fact only yesterday I've hit a
userspace-triggerable NULL pointer dereference in the pktcdvd driver
caused by bdev_get_queue() dereferencing an intermediate pointer blindly.

Of course, the actual cause of this bug probably lies somewhere else, but
a "soft" failure of returning an error (or printing a warning, or not
performing the requested operation, etc.) is no doubt preferred over a
"hard" failure of a kernel NULL pointer dereference.

Although in case of bdev_get_queue() function it can be argued that
this check wasn't done for performance reasons, since it is a core
blkdev API, one really cannot say so about winbond_gpio_get_info() in
this driver.

(..)
>>>> +	if (winbond_sio_reg_btest(*base, info->invreg, gpio_num))
>>>> +		val = !val;
>>>> +
>>>> +	if (val)
>>>
>>> if (val ^ winbond_sio_reg_btest()) ?
>>
>> Ok.
> 
> Just to be clear, I put "question mark" at the end of my proposals in
> which I completely leave choice on author(s) / maintainer(s).

All right, so I left this one as-is, since I think it is clearer this
way.

>>>> +	/*
>>>> +	 * Add 8 gpios for every GPIO port that was enabled in
>>>> gpios
>>>> +	 * module parameter (that wasn't disabled earlier in
>>>> +	 * winbond_gpio_configure() & co. due to, for example, a
>>>> pin
>>>> conflict).
>>>> +	 */
>>>> +	winbond_gpio_chip.ngpio = 0;
>>>> +	for (i = 0; i < 5; i++)
>>>
>>> 5 is a magic.
>>
>> Ok, will replace with ARRAY_SIZE as it is done in similar places.
>>
>>>> +		if (gpios & BIT(i))
>>>> +			winbond_gpio_chip.ngpio += 8;
>>>
>>> for_each_set_bit()
>>
>> The same situation as with the for_each_set_bit() remark above.
> 
> I believe hweight() is suitable here.
> 

I've adapted the code to use hweight() here.

Maciej