Skip to contentSkip to conversation

gpiolib: Fix gpio_direction_* for single direction GPIOs

PATCH In review Ricardo Ribalda Delgado 5 participants 22 messages Sep 21, 2018 – Oct 2, 2018 1 message not in the archive GPIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 gpiolib: Fix gpio_direction_* for single direction GPIOs

GPIOs with no programmable direction are not required to implement direction_output nor direction_input.

If we try to set an output direction on an output-only GPIO or input direction on an input-only GPIO simply return 0.

This allows this single direction GPIO to be used by libgpiod.

---
 drivers/gpio/gpiolib.c | 36 ++++++++++++++++++++++++++++--------
 1 file changed, 28 insertions(+), 8 deletions(-)
drivers/gpio/gpiolib.c +28 -8
index a57300c1d649..4b45de883ada 100644
@@ -2512,19 +2512,27 @@ EXPORT_SYMBOL_GPL(gpiochip_free_own_desc);
 int gpiod_direction_input(struct gpio_desc *desc)
 {
 	struct gpio_chip	*chip;
-	int			status = -EINVAL;
+	int			status = 0;
 
 	VALIDATE_DESC(desc);
 	chip = desc->gdev->chip;
 
-	if (!chip->get || !chip->direction_input) {
+	if (!chip->get && chip->direction_input) {
 		gpiod_warn(desc,
-			"%s: missing get() or direction_input() operations\n",
+			"%s: missing get() and direction_input() operations\n",
 			__func__);
 		return -EIO;
 	}
 
-	status = chip->direction_input(chip, gpio_chip_hwgpio(desc));
+	if (chip->direction_input) {
+		status = chip->direction_input(chip, gpio_chip_hwgpio(desc));
+	} else if (chip->get_direction &&
+		  (chip->get_direction(chip, gpio_chip_hwgpio(desc)) != 1)) {
+		gpiod_warn(desc,
+			"%s: missing direction_input() operation\n",
+			__func__);
+		return -EIO;
+	}
 	if (status == 0)
 		clear_bit(FLAG_IS_OUT, &desc->flags);
 
@@ -2546,16 +2554,28 @@ static int gpiod_direction_output_raw_commit(struct gpio_desc *desc, int value)
 {
 	struct gpio_chip *gc = desc->gdev->chip;
 	int val = !!value;
-	int ret;
+	int ret = 0;
 
-	if (!gc->set || !gc->direction_output) {
+	if (!gc->set && !gc->direction_output) {
 		gpiod_warn(desc,
-		       "%s: missing set() or direction_output() operations\n",
+		       "%s: missing set() and direction_output() operations\n",
 		       __func__);
 		return -EIO;
 	}
 
-	ret = gc->direction_output(gc, gpio_chip_hwgpio(desc), val);
+	if (gc->direction_output) {
+		ret = gc->direction_output(gc, gpio_chip_hwgpio(desc), val);
+	} else {
+		if (gc->get_direction &&
+		    gc->get_direction(gc, gpio_chip_hwgpio(desc))) {
+			gpiod_warn(desc,
+				"%s: missing direction_output() operation\n",
+				__func__);
+			return -EIO;
+		}
+		gc->set(gc, gpio_chip_hwgpio(desc), val);
+	}
+
 	if (!ret)
 		set_bit(FLAG_IS_OUT, &desc->flags);
 	trace_gpio_value(desc_to_gpio(desc), 0, val);
Signature · 1 line
2.18.0
#1 lore raw

PATCH v2 1/1 gpiolib: Show correct direction from the beginning

Current code assumes that the direction is input if direction_input function is set.
This might not be the case on GPIOs with programmable direction.

---
 drivers/gpio/gpiolib.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpio/gpiolib.c +6 -1
index 4b45de883ada..00c17f64d9ff 100644
@@ -1352,7 +1352,12 @@ int gpiochip_add_data_with_key(struct gpio_chip *chip, void *data,
 		 * it does, and in case chip->get_direction is not set, we may
 		 * expose the wrong direction in sysfs.
 		 */
-		desc->flags = !chip->direction_input ? (1 << FLAG_IS_OUT) : 0;
+		if (chip->get_direction && gpiochip_line_is_valid(chip, i))
+			desc->flags = !chip->get_direction(chip, i) ?
+					(1 << FLAG_IS_OUT) : 0;
+		else
+			desc->flags = !chip->direction_input ?
+					(1 << FLAG_IS_OUT) : 0;
 	}
 
 #ifdef CONFIG_PINCTRL
Signature · 1 line
2.18.0
#2 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Jeff, can you test these two patches on Amberwing to make sure that they don't cause an XPU violation on boot?

The call to gpiochip_line_is_valid() should return false on any GPIOs that aren't listed in the ACPI table.

My concern is that this patch might be calling gpiochip_line_is_valid() too early, before all the arrays have been set up.

Thanks.

Quoted 28 lines · Ricardo Ribalda Delgado

Current code assumes that the direction is input if direction_input function is set.
This might not be the case on GPIOs with programmable direction.

---
  drivers/gpio/gpiolib.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpio/gpiolib.c +6 -1
index 4b45de883ada..00c17f64d9ff 100644
@@ -1352,7 +1352,12 @@ int gpiochip_add_data_with_key(struct gpio_chip *chip, void *data,
  		 * it does, and in case chip->get_direction is not set, we may
  		 * expose the wrong direction in sysfs.
  		 */
-		desc->flags = !chip->direction_input ? (1 << FLAG_IS_OUT) : 0;
+		if (chip->get_direction && gpiochip_line_is_valid(chip, i))
+			desc->flags = !chip->get_direction(chip, i) ?
+					(1 << FLAG_IS_OUT) : 0;
+		else
+			desc->flags = !chip->direction_input ?
+					(1 << FLAG_IS_OUT) : 0;
  	}
 
  #ifdef CONFIG_PINCTRL
Reviewed-by on v2 1/1
#3 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

In reply to Ricardo Ribalda Delgado

Quoting Ricardo Ribalda Delgado (2018-09-21 03:36:04)

Quoted 6 lines

Current code assumes that the direction is input if direction_input function is set.
This might not be the case on GPIOs with programmable direction.

---

Looks OK to me visually. I haven't tested it because I don't have access to the locked down hardware anymore.

#4 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 2 lines · Stephen Boyd

Looks OK to me visually. I haven't tested it because I don't have access to the locked down hardware anymore.

Same here. Please wait for Jeff Hugo to test it before applying. Thanks.

#5 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 5 lines · Timur Tabi
Quoted 2 lines · Stephen Boyd

Looks OK to me visually. I haven't tested it because I don't have access to the locked down hardware anymore.

Same here.  Please wait for Jeff Hugo to test it before applying.  Thanks.

I guess its lucky I saw this then.

I should be able to test within a week.

Signature · 4 lines
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
#7 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 5 lines · Timur Tabi
Quoted 2 lines · Jeffrey Hugo

I guess its lucky I saw this then.

Did you not get this email: https://lore.kernel.org/patchwork/patch/989545/#1173771

Apparently I did. I found it in my deleted items. I must have accidentally done that when sorting through my backlog after Linaro Connect. Sorry about that.

Signature · 4 lines
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
#8 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

In reply to Jeffrey Hugo
Quoted 11 lines · Jeffrey Hugo
Quoted 5 lines · Timur Tabi
Quoted 2 lines · Stephen Boyd

Looks OK to me visually. I haven't tested it because I don't have access to the locked down hardware anymore.

Same here.  Please wait for Jeff Hugo to test it before applying. Thanks.

I guess its lucky I saw this then.

I should be able to test within a week.

Nack. Causes a XPU violation to the GPIO config registers.

Signature · 4 lines
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
#9 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 1 line · Jeffrey Hugo

Nack. Causes a XPU violation to the GPIO config registers.

That doesn't surprise me at all.

I believe that the problem is that gpiochip_line_is_valid() is being called before gpiochip_irqchip_init_valid_mask() is called, which means that gpiochip_line_is_valid() always returns true.

#10 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Hi Timur

In fact gpiochip_init_valid_mask is called some lines after in the same function. We could reorder the function. Would that work for you?

The driver breaking is upstream? Is it possible to access the hardware?

Thanks

[Sorry for the two html mails in a row, I did not try to work with my phone before :) ]

Quoted 9 lines · Timur Tabi
Quoted 1 line · Jeffrey Hugo

Nack. Causes a XPU violation to the GPIO config registers.

That doesn't surprise me at all.

I believe that the problem is that gpiochip_line_is_valid() is being called before gpiochip_irqchip_init_valid_mask() is called, which means that gpiochip_line_is_valid() always returns true.

Signature · 1 line
Ricardo Ribalda
#11 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 2 lines · Ricardo Ribalda Delgado

In fact gpiochip_init_valid_mask is called some lines after in the same function. We could reorder the function. Would that work for you?

It might. It might break something else, though.

Quoted 1 line

The driver breaking is upstream?

Yes.

Quoted 1 line

Is it possible to access the hardware?

Linaro and some Linux OSVs should still have systems, but I usually just ask Jeff to test code for me.

#12 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 4 lines · Timur Tabi
Quoted 2 lines

Is it possible to access the hardware?

Linaro and some Linux OSVs should still have systems, but I usually just ask Jeff to test code for me.

Alternatively, you can just add valid_mask support to your driver, and add a check to your get_direction() function to see if it ever is asked to access an invalid GPIO.

A message that is not in the archive (sent off-list or not yet received): [email protected]

#14 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 2 lines · Ricardo Ribalda Delgado

How do we proceed from here? Can you fix your driver somehow to init the valid mask before enabling the gpio?

Just include a hunk to the qcom driver reordering this call at the same time. No need to make it separate patches, it need to be tested together anyways.

I guess just switch the order of these two:

        ret = gpiochip_add_data(&pctrl->chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed register gpiochip\n");
                return ret;
        }

        ret = msm_gpio_init_valid_mask(chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed to setup irq valid bits\n");
                gpiochip_remove(&pctrl->chip);
                return ret;
        }
Quoted 1 line

Do we need to make more severe changes on the core?

Don't think so.

Yours,
Linus Walleij

#15 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Hi Linus

Quoted 13 lines · Linus Walleij
Quoted 2 lines · Ricardo Ribalda Delgado

How do we proceed from here? Can you fix your driver somehow to init the valid mask before enabling the gpio?

Just include a hunk to the qcom driver reordering this call at the same time. No need to make it separate patches, it need to be tested together anyways.

I guess just switch the order of these two:

Quoted 13 lines
        ret = gpiochip_add_data(&pctrl->chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed register gpiochip\n");
                return ret;
        }

        ret = msm_gpio_init_valid_mask(chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed to setup irq valid bits\n");
                gpiochip_remove(&pctrl->chip);
                return ret;
        }

the problem is that valid_mask is not a long/integer, is a struct that needs to be malloced, and is malloc at gpiochip_add_data :(

Maybe we need a callback from the driver to init that mask just after the allocation?

A fast grep shows that the only driver using need_valid_mask (not for irq) is msm:

ricardo@neopili:~/curro/kernel-upstream$ git grep "need_valid_mask ="
| grep -v irq
drivers/gpio/gpiolib.c: gpiochip->need_valid_mask = true;
drivers/pinctrl/intel/pinctrl-cherryview.c: bool need_valid_mask =
!dmi_check_system(chv_no_valid_mask);
drivers/pinctrl/qcom/pinctrl-msm.c: chip->need_valid_mask =
msm_gpio_needs_valid_mask(pctrl);

so hacking something in the driver might not be a terrible idea.

Also

Quoted 6 lines
Quoted 1 line

Do we need to make more severe changes on the core?

Don't think so.

Yours,
Linus Walleij

Signature · 1 line
Ricardo Ribalda
#16 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 46 lines · Ricardo Ribalda Delgado

Hi Linus

Quoted 13 lines · Linus Walleij
Quoted 2 lines · Ricardo Ribalda Delgado

How do we proceed from here? Can you fix your driver somehow to init the valid mask before enabling the gpio?

Just include a hunk to the qcom driver reordering this call at the same time. No need to make it separate patches, it need to be tested together anyways.

I guess just switch the order of these two:

Quoted 13 lines
         ret = gpiochip_add_data(&pctrl->chip, pctrl);
         if (ret) {
                 dev_err(pctrl->dev, "Failed register gpiochip\n");
                 return ret;
         }

         ret = msm_gpio_init_valid_mask(chip, pctrl);
         if (ret) {
                 dev_err(pctrl->dev, "Failed to setup irq valid bits\n");
                 gpiochip_remove(&pctrl->chip);
                 return ret;
         }

the problem is that valid_mask is not a long/integer, is a struct that needs to be malloced, and is malloc at gpiochip_add_data :(

Maybe we need a callback from the driver to init that mask just after the allocation?

A fast grep shows that the only driver using need_valid_mask (not for irq) is msm:

ricardo@neopili:~/curro/kernel-upstream$ git grep "need_valid_mask ="
| grep -v irq
drivers/gpio/gpiolib.c: gpiochip->need_valid_mask = true;
drivers/pinctrl/intel/pinctrl-cherryview.c: bool need_valid_mask =
!dmi_check_system(chv_no_valid_mask);
drivers/pinctrl/qcom/pinctrl-msm.c: chip->need_valid_mask =
msm_gpio_needs_valid_mask(pctrl);
Quoted 2 lines

so hacking something in the driver might not be a terrible idea.

IMO, hack up the driver and I'll test it. We can figure out what is needed to work, then determine what is the proper solution for that.

Signature · 4 lines
Jeffrey Hugo
Qualcomm Datacenter Technologies as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the
Code Aurora Forum, a Linux Foundation Collaborative Project.
#17 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

In reply to Ricardo Ribalda Delgado
Quoted 27 lines · Ricardo Ribalda Delgado
Quoted 12 lines · Linus Walleij
Quoted 2 lines · Ricardo Ribalda Delgado

How do we proceed from here? Can you fix your driver somehow to init the valid mask before enabling the gpio?

Just include a hunk to the qcom driver reordering this call at the same time. No need to make it separate patches, it need to be tested together anyways.

I guess just switch the order of these two:

Quoted 13 lines
        ret = gpiochip_add_data(&pctrl->chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed register gpiochip\n");
                return ret;
        }

        ret = msm_gpio_init_valid_mask(chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed to setup irq valid bits\n");
                gpiochip_remove(&pctrl->chip);
                return ret;
        }
Quoted 2 lines

the problem is that valid_mask is not a long/integer, is a struct that needs to be malloced, and is malloc at gpiochip_add_data :(

I don't get it, but maybe I'm not smart enough.

gpiochip_add_data() doesn't allocate anything, it just adds a already allocated (or static!) gpio_chip to the gpiolib subsystem.

In fact I think it is wrong to set up the mask after calling gpiolob_add_data(), because of exactly the type of problem you're seeing.

Don't get confused by the &pctrl->chip
vs just chip variables, it's just some sloppiness.

Yours,
Linus Walleij

#18 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Hi

Quoted 30 lines · Linus Walleij
Quoted 27 lines · Ricardo Ribalda Delgado
Quoted 25 lines · Linus Walleij
Quoted 2 lines · Ricardo Ribalda Delgado

How do we proceed from here? Can you fix your driver somehow to init the valid mask before enabling the gpio?

Just include a hunk to the qcom driver reordering this call at the same time. No need to make it separate patches, it need to be tested together anyways.

I guess just switch the order of these two:

        ret = gpiochip_add_data(&pctrl->chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed register gpiochip\n");
                return ret;
        }

        ret = msm_gpio_init_valid_mask(chip, pctrl);
        if (ret) {
                dev_err(pctrl->dev, "Failed to setup irq valid bits\n");
                gpiochip_remove(&pctrl->chip);
                return ret;
        }
Quoted 4 lines
Quoted 2 lines

the problem is that valid_mask is not a long/integer, is a struct that needs to be malloced, and is malloc at gpiochip_add_data :(

I don't get it, but maybe I'm not smart enough.

Dont take my job, I am the not smart of the conversation :P

Quoted 5 lines

gpiochip_add_data() doesn't allocate anything, it just adds a already allocated (or static!) gpio_chip to the gpiolib subsystem.

Take a look to gpiochip_add_data_with_key()
->gpiochip_init_valid_mask()

       -> gpiochip->valid_mask = gpiochip_allocate_mask(gpiochip);
Quoted 3 lines

In fact I think it is wrong to set up the mask after calling gpiolob_add_data(), because of exactly the type of problem you're seeing.

I agree, and I believe that the cleaner way would be to add a function pointer that replaces:

gpiochip_allocate_mask()
bitmap_fill(p, chip->ngpio);

with a proper initalization from the driver

But as today the only driver that seems to be using valid_mask is msm, so perhaps a hack is something better and then when we have a second driver that requires it we figure out the real requirements. But it is definately your decision ;)

Quoted 6 lines

Don't get confused by the &pctrl->chip
vs just chip variables, it's just some sloppiness.

Yours,
Linus Walleij

Thanks!

Signature · 1 line
Ricardo Ribalda
#19 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 9 lines · Ricardo Ribalda Delgado
Quoted 3 lines · Linus Walleij

gpiochip_add_data() doesn't allocate anything, it just adds a already allocated (or static!) gpio_chip to the gpiolib subsystem.

Take a look to gpiochip_add_data_with_key()
->gpiochip_init_valid_mask()

       -> gpiochip->valid_mask = gpiochip_allocate_mask(gpiochip);

Aha I see...

Quoted 11 lines
Quoted 3 lines

In fact I think it is wrong to set up the mask after calling gpiolob_add_data(), because of exactly the type of problem you're seeing.

I agree, and I believe that the cleaner way would be to add a function pointer that replaces:

gpiochip_allocate_mask()
bitmap_fill(p, chip->ngpio);

with a proper initalization from the driver

OK

Quoted 4 lines

But as today the only driver that seems to be using valid_mask is msm, so perhaps a hack is something better and then when we have a second driver that requires it we figure out the real requirements. But it is definately your decision ;)

I would just add some exported function to gpiolib to do what you need so you can set up the valid_mask before calling gpiochip_add*.

Yours,
Linus Walleij

#20 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 4 lines · Linus Walleij
Quoted 4 lines

But as today the only driver that seems to be using valid_mask is msm, so perhaps a hack is something better and then when we have a second driver that requires it we figure out the real requirements. But it is definately your decision;)

Please note that MSM is supposed to be the *first* driver, not the only, driver that needs valid_mask. So let's not make any code changes that limit this feature to the MSM driver.

Quoted 3 lines

I would just add some exported function to gpiolib to do what you need so you can set up the valid_mask before calling gpiochip_add*.

I think that should be okay. Drivers should know pretty early whether they need valid_mask or not.

#21 lore raw

Re: [PATCH v2] gpiolib: Show correct direction from the beginning

Quoted 9 lines · Timur Tabi
Quoted 4 lines · Linus Walleij
Quoted 4 lines

But as today the only driver that seems to be using valid_mask is msm, so perhaps a hack is something better and then when we have a second driver that requires it we figure out the real requirements. But it is definately your decision;)

Please note that MSM is supposed to be the *first* driver, not the only, driver that needs valid_mask. So let's not make any code changes that limit this feature to the MSM driver.

I just recently encouraged Thierry to reuse this for the Tegra186 driver:
https://marc.info/?l=linux-gpio&m=153787164826821&w=2

Yours,
Linus Walleij

#22 lore raw
In reply to Ricardo Ribalda Delgado

Hi Ricardo,

thanks for the patch and sorry for taking time before responding.

Quoted 9 lines · Ricardo Ribalda Delgado

GPIOs with no programmable direction are not required to implement direction_output nor direction_input.

If we try to set an output direction on an output-only GPIO or input direction on an input-only GPIO simply return 0.

This allows this single direction GPIO to be used by libgpiod.

It makes perfect sense, patch applied.

I'll go in and add some comments to the code so I understand it right as well in the future.

Yours,
Linus Walleij