[U-Boot] [PATCH 3/4] gpio: xilinx: Not read output values via regs

[Stefan Herbrechtsmeier]

Am 30.07.2018 um 16:10 schrieb Michal Simek:
> On 30.7.2018 15:32, Stefan Herbrechtsmeier wrote:
>> Am 30.07.2018 um 14:40 schrieb Michal Simek:
>>> On 27.7.2018 10:41, Stefan Herbrechtsmeier wrote:
>>>> Am 27.07.2018 um 09:05 schrieb Michal Simek:
>>>>> On 26.7.2018 21:46, Stefan Herbrechtsmeier wrote:
>>>>>> Am 26.07.2018 um 10:41 schrieb Michal Simek:
>>>>>>> On 25.7.2018 20:21, Stefan Herbrechtsmeier wrote:
>>>>>>>> Am 25.07.2018 um 08:39 schrieb Michal Simek:
>>>>>>>>> On 24.7.2018 21:56, Stefan Herbrechtsmeier wrote:
>>>>>>>>>> Am 24.07.2018 um 12:31 schrieb Michal Simek:
>>>>>>>>>>> On 23.7.2018 20:42, Stefan Herbrechtsmeier wrote:
>>>>>>>>>>>> Am 23.07.2018 um 13:43 schrieb Michal Simek:
>>>>>>>>>>>>> Reading registers for finding out output value is not working
>>>>>>>>>>>>> because
>>>>>>>>>>>>> input value is read instead in case of tristate.
>>>>>>>>>>>>>
>>>>>>>>>>>>> Reported-by: Stefan Herbrechtsmeier
>>>>>>>>>>>>> <stefan at herbrechtsmeier.net>
>>>>>>>>>>>>> Signed-off-by: Michal Simek <michal.simek at xilinx.com>
>>>>>>>>>>>>> ---
>>>>>>>>>>>>>
>>>>>>>>>>>>>         drivers/gpio/xilinx_gpio.c | 38
>>>>>>>>>>>>> +++++++++++++++++++++++++++++++++-----
>>>>>>>>>>>>>         1 file changed, 33 insertions(+), 5 deletions(-)
>>>>>>>>>>>>>
>>>>>>>>>>>>> diff --git a/drivers/gpio/xilinx_gpio.c
>>>>>>>>>>>>> b/drivers/gpio/xilinx_gpio.c
>>>>>>>>>>>>> index 4da9ae114d87..9d3e9379d0e5 100644
>>>>>>>>>>>>> --- a/drivers/gpio/xilinx_gpio.c
>>>>>>>>>>>>> +++ b/drivers/gpio/xilinx_gpio.c
>>>>>>>>>> [snip]
>>>>>>>>>>
>>>>>>>>>>>>>         +    priv->output_val[bank] = val;
>>>>>>>>>>>>> +
>>>>>>>>>>>>>             return val;
>>>>>>>>>>>>>         };
>>>>>>>>>>>>>         @@ -441,6 +449,7 @@ static int
>>>>>>>>>>>>> xilinx_gpio_get_function(struct
>>>>>>>>>>>>> udevice *dev, unsigned offset)
>>>>>>>>>>>>>         static int xilinx_gpio_get_value(struct udevice *dev,
>>>>>>>>>>>>> unsigned
>>>>>>>>>>>>> offset)
>>>>>>>>>>>>>         {
>>>>>>>>>>>>>             struct xilinx_gpio_platdata *platdata =
>>>>>>>>>>>>> dev_get_platdata(dev);
>>>>>>>>>>>>> +    struct xilinx_gpio_privdata *priv = dev_get_priv(dev);
>>>>>>>>>>>>>             int val, ret;
>>>>>>>>>>>>>             u32 bank, pin;
>>>>>>>>>>>>>         @@ -451,7 +460,14 @@ static int
>>>>>>>>>>>>> xilinx_gpio_get_value(struct
>>>>>>>>>>>>> udevice
>>>>>>>>>>>>> *dev, unsigned offset)
>>>>>>>>>>>>>             debug("%s: regs: %lx, gpio: %x, bank %x, pin %x\n",
>>>>>>>>>>>>> __func__,
>>>>>>>>>>>>>                   (ulong)platdata->regs, offset, bank, pin);
>>>>>>>>>>>>>         -    val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>>>>>>>>>> +    if (xilinx_gpio_get_function(dev, offset) ==
>>>>>>>>>>>>> GPIOF_INPUT) {
>>>>>>>>>>>>> +        debug("%s: Read input value from reg\n", __func__);
>>>>>>>>>>>>> +        val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>>>>>>>>>> +    } else {
>>>>>>>>>>>>> +        debug("%s: Read saved output value\n", __func__);
>>>>>>>>>>>>> +        val = priv->output_val[bank];
>>>>>>>>>>>>> +    }
>>>>>>>>>>>> Why you don't always read the data register? This doesn't
>>>>>>>>>>>> work for
>>>>>>>>>>>> three
>>>>>>>>>>>> state outputs.
>>>>>>>>>>> In three state register every bit/pin is 0 - output, 1 input.
>>>>>>>>>>> It means else part is output and I read saved value in
>>>>>>>>>>> priv->output_val.
>>>>>>>>>>> If pin is setup as INPUT then I need read data reg to find out
>>>>>>>>>>> input
>>>>>>>>>>> value.
>>>>>>>>>>> Maybe you are commenting something else but please let me know if
>>>>>>>>>>> there
>>>>>>>>>>> is any other bug.
>>>>>>>>>> What happen if I have an open drain output. Even if the gpio
>>>>>>>>>> output
>>>>>>>>>> is 1
>>>>>>>>>> the input could read a 0. You driver will always return the output
>>>>>>>>>> value
>>>>>>>>>> and not the real input value. According to the picture in
>>>>>>>>>> documentation
>>>>>>>>>> and my tests a data register write writes the output registers
>>>>>>>>>> and a
>>>>>>>>>> data register read reads the input registers.
>>>>>>>>>>
>>>>>>>>>> Why should the driver return the desired state (output register)
>>>>>>>>>> and not
>>>>>>>>>> the real state (input register)?
>>>>>>>>> First of all thanks for description.
>>>>>>>>>
>>>>>>>>> I have another example where you have output only and you can't
>>>>>>>>> read
>>>>>>>>> input because there is no wire.
>>>>>>>> Does you mean the all outputs configuration? Does this removes the
>>>>>>>> "gpio_io_i" signal from the IP?
>>>>>>> I am not sure what synthesis is doing with that unused IP pins but I
>>>>>>> would consider as a bug if this is automatically connected together.
>>>>>> I mean does the IP generator removes the gpio_io_i signal because it
>>>>>> isn't needed? If the IP generator creates the gpio_io_i signal I would
>>>>>> expect that you can't leave it unconnected as this would lead to
>>>>>> undefined values.
>>>>> Normally when you know that you have output only there is no no
>>>>> gpio_io_i or tristate signal. The same for input only.
>>>> And in this case the device tree flags "xlnx,all-inputs" or
>>>> "xlnx,all-outputs" should be set.
>>> yes.
>>>
>>>>>>>      And
>>>>>>> also wasting a logic if there is unused part.
>>>>>>> But in Vivado you should be able to setup output pins to and input
>>>>>>> pins
>>>>>>> separately. There are In/Out/Tristate.
>>>>>>> If you don't want to deal with external pin you can connect them
>>>>>>> inside PL.
>>>>>> This isn't my point. I mean that if you have an gpio_io_i signal you
>>>>>> have to connected it to a signal. You could connect it to the
>>>>>> output of
>>>>>> an IO, to the gpio_io_o signal or to fixed value (0 or 1). If you
>>>>>> connect it to a fixed value (0 or 1) you get this value if you read
>>>>>> the
>>>>>> status of this GPIO.
>>>>> Not a HW to tell you what's vivado is going to do with that. But you
>>>>> can
>>>>> select that ouput/input only where these signals are out.
>>>> You mean the all-inputs or all-outputs case.
>>> yes.
>>>
>>>>>>>>> Regarding open drain output.
>>>>>>>>>
>>>>>>>>> Also this is what it is written in manual:
>>>>>>>>> "AXI GPIO Data Register Description"
>>>>>>>>> "For each I/O bit programmed as output:
>>>>>>>>> • R: Reads to these bits always return zeros"
>>>>>>>> This must be a mistake in the documentation. I could read the input
>>>>>>>> value.
>>>>>>>>
>>>>>>>> The old driver and the Linux driver always read the register.
>>>>>>> In old u-boot driver I see
>>>>>>>      179         if (gpio_get_direction(gpio) == GPIO_DIRECTION_OUT)
>>>>>>>      180                 val = gpio_get_output_value(gpio);
>>>>>>>      181         else
>>>>>>>      182                 val = gpio_get_input_value(gpio);
>>>>>>>
>>>>>>> gpio_get_output_value() reads it from gpiodata_store for output
>>>>>>> gpio_get_input_value() reads the reg.
>>>>>> Sorry you are right. I mixed it with the zynq driver.
>>>>> ok.
>>>>>
>>>>>
>>>>>>> In Linux you are right it is read from reg in both cases.
>>>>>>>
>>>>>>>> Additionally the documentation states that a write to an input
>>>>>>>> doesn't
>>>>>>>> work:
>>>>>>>> AXI GPIO Data Register.
>>>>>>>> For each I/O bit programmed as input:
>>>>>>>> •  R: Reads value on the input pin.
>>>>>>>> •  W: No effect.
>>>>>>>>
>>>>>>>> However the Linux driver use the common sequence and sets first the
>>>>>>>> data
>>>>>>>> register and afterwards the direction register.
>>>>>>> Possible that driver has pretty long history and I don't think it was
>>>>>>> tested for all cases.
>>>>>> I test the IP again and it works like expected and the
>>>>>> documentation is
>>>>>> wrong. You could always write the output register and read the input
>>>>>> register. This means you could write the output first and afterwards
>>>>>> enable the output via the tri-state register.
>>>>> ok.
>>>> Maybe the set_direction function should be adapted to avoid glitches
>>>> because it first enabled the output with the old value.
>>> No problem with this change at all.
>>>
>>>>>>>>> If you want to support this configuration then you would have to
>>>>>>>>> change
>>>>>>>>> pin to input and read it and revert it back. And all of this
>>>>>>>>> should be
>>>>>>>>> done based on flag.
>>>>>>>>> There is macro GPIO_OPEN_DRAIN which could be specified via DT
>>>>>>>>> binding.
>>>>>>>>> Unfortunately this is for consumer not for generic listing. I can't
>>>>>>>>> also
>>>>>>>>> see it in gpio_desc flags to be able to handle it in the driver.
>>>>>>>>> That's why I have doubts if this is supported by any u-boot gpio
>>>>>>>>> driver
>>>>>>>>> but I can be wrong.
>>>>>>>> I think the GPIO_OPEN_DRAIN is used to not set the output to 1.
>>>>>>>> In our
>>>>>>>> case the open drain is transparent. The output has only two states
>>>>>>>> High-Z and 0. The input value of the FPGA output block is always
>>>>>>>> 0 and
>>>>>>>> the tri-state input is connected to the output register of the GPIO
>>>>>>>> controller. The output of the FPGA input block is connected to the
>>>>>>>> input
>>>>>>>> register of the GPIO controller.
>>>>>>> I just found out in connection to second thread we have together
>>>>>>> that in
>>>>>>> gpio-uclass there are dm_gpio_set/get_open_drain functions but
>>>>>>> they are
>>>>>>> not wired anywhere.
>>>>>>> I think in your case when these two functions are implemented you
>>>>>>> will
>>>>>>> get this functionality which you are asking about.
>>>>>> No. This function is to control a special register.
>>>>> I am not quite sure if this is really about special register or can be
>>>>> also used in our case.
>>>> In my case the open drain is transparent for the controller. The output
>>>> replace a 1 by a High-Z. This only means that you don't know the real
>>>> value on the wire.
>>> please look below.
>>>
>>>>>> Beside the open drain. What happens if the output is stuck to a fixed
>>>>>> signal. This could be detected if you read the input register.
>>>>> And again if that wires are there. Not covering the case where you have
>>>>> output only without any input signal at all.
>>>> This can be detect by the all-outputs flag.
>>> please look below.
>>>
>>>>>>> My guess is that gpio command should be extended to call these two
>>>>>>> functions. When set_open_drain is called special flag for pin is
>>>>>>> setup
>>>>>>> and for this pin get_status will always read data reg in our case and
>>>>>>> you get your values.
>>>>>> The main quest is which value should the get function return for an
>>>>>> output. The value from the input register which represents the real
>>>>>> value or the value from the output register which represents the
>>>>>> desired
>>>>>> value. Because of the "Warning: value of pin is still %d" inside the
>>>>>> gpio cmd I would expect the real value. This is also implemented by
>>>>>> most
>>>>>> drivers. Only tegra, kw and rcar distinguish between input and output.
>>>>>> Thereby the rcar explicitly mention a bug:
>>>>>>
>>>>>> Testing on r8a7790 shows that INDT does not show correct pin state
>>>>>> when
>>>>>> configured as output, so use OUTDT in case of output pins.
>>>>>>
>>>>>> A test of the Xilinx GPIO controller shows that a read always returns
>>>>>> the value of the gpio_io_i signal.
>>>>> What about to use this logic?
>>>>>
>>>>> if (platdata->bank_input[bank]) // input only
>>>>>       val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>> else if (platdata->bank_output[bank])) // output only
>>>>>       val = priv->output_val[bank];
>>>> Okay
>>>>
>>>>> else { // tristate because it is not output or input only.
>>>>>       state = xilinx_gpio_get_function(dev, offset)
>>>>>       if (state == GPIOF_INPUT) // if input now - just read it
>>>>>           val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>>       else { // if output - change it to input
>>>>>               tmp1 = readl(&platdata->regs->gpiodir + bank * 2);
>>>>>               tmp = tmp1 | (1 << pin);
>>>>>               writel(tmp, &platdata->regs->gpiodir + bank * 2);
>>>>>
>>>>>           // read value
>>>>>           val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>>
>>>>>           // revert it back
>>>>>                    writel(tmp1, &platdata->regs->gpiodir + bank * 2);
>>>>>       }
>>>>> }
>>>> This will generate a glitch on the wire.
>>>>
>>>> Why don't we simple read the value? The hardware support this and we
>>>> have already handle the all-output case:
>>>>
>>>> if (platdata->bank_output[bank]))
>>>>        val = priv->output_val[bank];
>>>> else
>>>>        val = readl(&platdata->regs->gpiodata + bank * 2);
>>>>
>>>> If you don't trust my test could you please ask your hardware co works
>>>> why the data register behavior should depend on the tri-state register.
>>> I am looking for IP owner but as far as I see it was created 9 years ago
>>> that's why it will take some time.
>>> Anyway I have created hw design on zcu102 which has output on led and
>>> input on dip and output value is kept in data reg that's why your
>>> snippet above can be used.
>> Do you use a signal port for both or two separated ports?
> I expect single - yes the same pin 0 for output led and input dip.

The same pin or port?

>> Have you connect the gpio_io_i and gpio_io_o signals of the outputs?
> It is really pin 0 io_o to one fpga pin and pin0 io_i to another fpga pin.

To make sure I understand you correct: The gpio_io_o[0] is connected to 
a LED and gpio_io_i[0] is connected to a switch?

Have you connect the gpio_io_t[0] signal?

>> What happens if you connect the gpio_io_i signals to zero for the output
>> to save resources?
>> What happens if you disable the power supply for the LED bank?
> didn't try that.
>
>>>    Documentation is wrong in this
>>> "For each I/O bit programmed as output:
>>> •  R: Reads to these bits always return zeros"
>>>
>>> I have sent 4 patches which we discussed that's why please reviewed them.
>>> I have one more patch in queue which is keeping output_val for certain
>>> bank but not sure if make sense to apply it because as above when
>>> direction is setup to output you can read desired output value from
>>> gpiodata reg.
>> This is only true if the gpio_io_i and gpio_io_o signals are connected
>> together or the input and output signals have always the same value.
>> Otherwise you read the input signal and write it into the output value.
> That's what I see on zcu102 without any connection between io_i io_o.
> There is incorrect behavior when gpio toggle command is called and input
> value is 1. Because the first gpio toggle cmd reads input value first
> (because code is doing that) and then setting up output with opposite value.
>
>
>> In our use case we have two gpio controllers (for two processors) which
>> are connected to one io port. Thereby only the output is multiplexed
>> between the two controllers and the input is always connected to the io.
>> This means both controller always read the input value of the io but
>> only one of them controls the output of the io. Your set function reads
>> the input value of the io (maybe controlled by the other controller) and
>> thereby overwrite the old output value of the controller.
>>
>>>    I have tested it on gpio output only led case and gpiodata
>>> reg contain correct value.
>> Have you test outputs with differences between output value and input
>> value?
> I believe so.
>
>>> Anyway if I miss any patch in connection to this please send a patch and
>>> let's look at that use case.
>> Okay
> TBH it is starting to be a little bit confusing.

Me too.

It looks like we have observe different behavior. I will redo my test in 
the next days.

It would be so much easier if we could check the IP's HDL.

The hardware have three physical registers (gpio_io_o, gpio_io_i, 
gpio_io_t) per pin and two bus registers (GPIOx_DATA, GPIOx_TRI) per port.

-> GPIOx_TRI
   -> Read
     -> gpio_io_t
   -> Write
     -> gpio_io_t

-> GPIOx_DATA
      -> Read
           -> gpio_io_i (my observation)
           -> (gpio_io_t == 1) ? gpio_io_i : gpio_io_o (your observation ?)
      -> Write
           -> gpio_io_o

> What we should really do is to use standard boards and connection and
> describe that cases and expected behavior and changes in code which
> should happen to support this.
>
> We are talking about two functions.
> xilinx_gpio_set_value and xilinx_gpio_get_value.
>
> And let's take mainline driver with that 4 patches I have sent.

Okay

> Let's start with xilinx_gpio_get_value()
> Mainline is all the time reading value from gpiodata reg no matter if
> this is output/input only or io.
>
> Above is this sequence which you have written that it is right
> if (platdata->bank_output[bank]))
>         val = priv->output_val[bank];
> else
>        val = readl(&platdata->regs->gpiodata + bank * 2);

Based on your observations the driver should work without this change 
because the controller returns the output register value if the tristate 
register is cleared. Hopefully this also happens if the all_outputs flag 
is set.

> I have took at look at gpio output only case
>
> ZynqMP> gpio toggle gpio at a00010000
> gpio: pin gpio at a00010000 (gpio 187) value is 1
> ZynqMP> gpio status -a gpio at a0001000
> Bank gpio at a0001000:
> gpio at a00010000: output: 1 [ ]
> gpio at a00010001: output: 0 [ ]
> gpio at a00010002: output: 0 [ ]
> gpio at a00010003: output: 0 [ ]
> gpio at a00010004: output: 0 [ ]
> gpio at a00010005: output: 0 [ ]
> gpio at a00010006: output: 0 [ ]
> gpio at a00010007: output: 0 [ ]
> ZynqMP> gpio toggle gpio at a00010004
> gpio: pin gpio at a00010004 (gpio 191) value is 1
> ZynqMP> gpio status -a gpio at a0001000
> Bank gpio at a0001000:
> gpio at a00010000: output: 1 [ ]
> gpio at a00010001: output: 0 [ ]
> gpio at a00010002: output: 0 [ ]
> gpio at a00010003: output: 0 [ ]
> gpio at a00010004: output: 1 [ ]
> gpio at a00010005: output: 0 [ ]
> gpio at a00010006: output: 0 [ ]
> gpio at a00010007: output: 0 [ ]
> ZynqMP> gpio toggle gpio at a00010006
> gpio: pin gpio at a00010006 (gpio 193) value is 1
> ZynqMP> gpio status -a gpio at a0001000
> Bank gpio at a0001000:
> gpio at a00010000: output: 1 [ ]
> gpio at a00010001: output: 0 [ ]
> gpio at a00010002: output: 0 [ ]
> gpio at a00010003: output: 0 [ ]
> gpio at a00010004: output: 1 [ ]
> gpio at a00010005: output: 0 [ ]
> gpio at a00010006: output: 1 [ ]
> gpio at a00010007: output: 0 [ ]
> ZynqMP> gpio toggle gpio at a00010004
> gpio: pin gpio at a00010004 (gpio 191) value is 0
> ZynqMP> gpio status -a gpio at a0001000
> Bank gpio at a0001000:
> gpio at a00010000: output: 1 [ ]
> gpio at a00010001: output: 0 [ ]
> gpio at a00010002: output: 0 [ ]
> gpio at a00010003: output: 0 [ ]
> gpio at a00010004: output: 0 [ ]
> gpio at a00010005: output: 0 [ ]
> gpio at a00010006: output: 1 [ ]
> gpio at a00010007: output: 0 [ ]
> ZynqMP> gpio toggle gpio at a00010007
> gpio: pin gpio at a00010007 (gpio 194) value is 1
> ZynqMP> gpio status -a gpio at a0001000
> Bank gpio at a0001000:
> gpio at a00010000: output: 1 [ ]
> gpio at a00010001: output: 0 [ ]
> gpio at a00010002: output: 0 [ ]
> gpio at a00010003: output: 0 [ ]
> gpio at a00010004: output: 0 [ ]
> gpio at a00010005: output: 0 [ ]
> gpio at a00010006: output: 1 [ ]
> gpio at a00010007: output: 1 [ ]
>
> And for this case there is no need to take that value from any saved
> location.
> It means having
>        val = readl(&platdata->regs->gpiodata + bank * 2);
> is enough based on my test.
> And for the rest as your example above this should be also fine.

Yes. If your observations are correct it is impossible to read the input 
state if the GPIO is configured as output and if my observations are 
correct the read will return the input state independently of the mode.

> When we have agreement on this we can look at xilinx_gpio_set_value.

The direct read is fine for me.

Can you read the two switch states if you run the following commands:

/* Enable LED */
ZynqMP> gpio set gpio at a00010000
/* Read switch state and keep LED on */
ZynqMP> gpio input gpio at a00010000
ZynqMP> gpio status -a gpio at a0001000
/* Manually change switch state and read again */
ZynqMP> gpio status -a gpio at a0001000
/* Disable LED */
ZynqMP> gpio clear gpio at a00010000
ZynqMP> gpio status -a gpio at a0001000
/* Manually change switch state and read again */
ZynqMP> gpio status -a gpio at a0001000

Best regards
   Stefan