Re: [PATCH v2 2/3] gpio: Add gpio driver support for ThunderX and OCTEON-TX

From: David Daney
Date: Mon Jan 09 2017 - 15:02:56 EST


On 01/09/2017 11:35 AM, Linus Walleij wrote:
On Sat, Jan 7, 2017 at 12:22 AM, David Daney <ddaney.cavm@xxxxxxxxx> wrote:

From: David Daney <david.daney@xxxxxxxxxx>

Cavium ThunderX and OCTEON-TX are arm64 based SoCs. Add driver for
the on-chip GPIO pins.

Signed-off-by: David Daney <david.daney@xxxxxxxxxx>

(...)
+config GPIO_THUNDERX
+ tristate "Cavium ThunderX/OCTEON-TX GPIO"

Do you really load this as module? OK then...

The driver does work when loaded as a module. I have tested loading and unloading it many times.


+#include <linux/gpio.h>

No.
#include <linux/gpio/driver.h>

Nothing else.

Right, I will fix that.


+#define GLITCH_FILTER_400NS ((4ull << GPIO_BIT_CFG_FIL_SEL_SHIFT) | \
+ (9ull << GPIO_BIT_CFG_FIL_CNT_SHIFT))
+
+static unsigned int bit_cfg_reg(unsigned int line)
+{
+ return 8 * line + GPIO_BIT_CFG;
+}
+
+static unsigned int intr_reg(unsigned int line)
+{
+ return 8 * line + GPIO_INTR;
+}

This looks a bit overzealous, but OK.

+struct thunderx_gpio;
+
+struct thunderx_irqdev {
+ struct thunderx_gpio *gpio;
+ char *name;
+ unsigned int line;
+};
+
+struct thunderx_gpio {
+ struct gpio_chip chip;
+ u8 __iomem *register_base;
+ struct msix_entry *msix_entries;
+ struct thunderx_irqdev *irqdev_entries;
+ raw_spinlock_t lock;
+ unsigned long invert_mask[2];
+ unsigned long od_mask[2];
+ int base_msi;
+};

Why can't you just move the thunderx_irqdev fields
into thunderx_gpio?

There is a variable length array of them. It is not a single instance of the structure. Same for the msix_entries.

It will save very little memory for
non-irq systems, I do not think footprint warrants it.

+
+static bool thunderx_gpio_is_gpio(struct thunderx_gpio *gpio,
+ unsigned int line)
+{
+ u64 bit_cfg = readq(gpio->register_base + bit_cfg_reg(line));
+ bool rv = (bit_cfg & GPIO_BIT_CFG_PIN_SEL_MASK) == 0;
+
+ WARN_RATELIMIT(!rv, "Pin %d not available for GPIO\n", line);
+
+ return rv;
+}

Nifty. So this actually has a pin controller back-end?

The hardware does function like a "pin controller"

I haven't seen the code for that yet, I think.

This seems like a cheap version of

/* External interface to pin control */
extern int pinctrl_request_gpio(unsigned gpio);
extern void pinctrl_free_gpio(unsigned gpio);
extern int pinctrl_gpio_direction_input(unsigned gpio);
extern int pinctrl_gpio_direction_output(unsigned gpio);

From <linux/pinctrl/consumer.h>

So are you planning pin control support separately in drivers/pinctrl/*
in the future?

Not at this time. Currently early boot firmware is programming the pin functions, and we don't see a need to move the complexity of pinctrl into the kernel, especially as there is not really any ACPI support for pinctrl at this point, and we must also support ACPI.

Maybe some comment to replace this with proper pin control
in the future is warranted?


Good idea. I will add a comment about this.


+static int thunderx_gpio_dir_in(struct gpio_chip *chip, unsigned int line)
+{
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);

1. Please use gpiochip_get_data() instead of the container_of() construction,
utilized devm_gpiochip_add_data() in your probe() call.

2. Do not call this local variable "gpio" that is superconfusing, at least call
it txgpio or something.

1 & 2 applies EVERYWHERE in thid driver.

OK, I will make that change for the next revision of the patch set.


+static void thunderx_gpio_set(struct gpio_chip *chip, unsigned int line,
+ int value)
+{
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);
+ int bank = line / 64;
+ int bank_bit = line % 64;
+
+ void __iomem *reg = gpio->register_base +
+ (bank * GPIO_2ND_BANK) + (value ? GPIO_TX_SET : GPIO_TX_CLR);
+
+ writeq(1ull << bank_bit, reg);

Use this:

#include <linus/bitops.h>

writeq(BIT(bank_bit), reg);

Applies EVERYWHERE you use (1ULL << n)

OK.


+static int thunderx_gpio_set_single_ended(struct gpio_chip *chip,
+ unsigned int line,
+ enum single_ended_mode mode)

Thanks for implementing this properly.

+ read_bits >>= bank_bit;
+
+ if (test_bit(line, gpio->invert_mask))
+ return !(read_bits & 1);
+ else
+ return read_bits & 1;

This looks superconvoluted.

Can't you just:

if (test_bit(line, gpio->invert_mask))
return !(read_bits & BIT(bank_bit));
else
return !!(read_bits & BIT(bank_bit));

OK maybe not much clearer but seems clearer to me.

As I really dislike the "!!" idiom, would you settle for:

if (test_bit(line, gpio->invert_mask))
return (read_bits & BIT(bank_bit)) == 0;
else
return (read_bits & BIT(bank_bit)) != 0;


+static int thunderx_gpio_irq_request_resources(struct irq_data *data)
+{
+ struct gpio_chip *chip = irq_data_get_irq_chip_data(data);
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);
+ unsigned int line = data->hwirq;
+ struct thunderx_irqdev *irqdev;
+ int err;
+
+ if (!thunderx_gpio_is_gpio(gpio, line))
+ return -EIO;
+
+ irqdev = gpio->irqdev_entries + line;
+
+ irqdev->gpio = gpio;
+ irqdev->line = line;
+ irqdev->name = devm_kasprintf(chip->parent, GFP_KERNEL,
+ "gpio-%d", line + chip->base);
+
+ writeq(GPIO_INTR_ENA_W1C, gpio->register_base + intr_reg(line));
+
+ err = devm_request_irq(chip->parent, gpio->msix_entries[line].vector,
+ thunderx_gpio_chain_handler, IRQF_NO_THREAD, irqdev->name, irqdev);
+ return err;
+}
+
+static void thunderx_gpio_irq_release_resources(struct irq_data *data)
+{
+ struct gpio_chip *chip = irq_data_get_irq_chip_data(data);
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);
+ unsigned int line = data->hwirq;
+ struct thunderx_irqdev *irqdev;
+
+ irqdev = gpio->irqdev_entries + line;
+
+ /*
+ * The request_resources/release_resources functions may be
+ * called multiple times in the lifitime of the driver, so we
+ * need to clean up the devm_* things to avoid a resource
+ * leak.
+ */
+ devm_free_irq(chip->parent, gpio->msix_entries[line].vector, irqdev);
+
+ writeq(GPIO_INTR_ENA_W1C, gpio->register_base + intr_reg(line));
+
+ devm_kfree(chip->parent, irqdev->name);
+}

Then just do not use the devm* variants. Explicitly allocate and free instead.

These callbacks should be called on all resources anyways.

Rithe.


+static void thunderx_gpio_irq_unmask(struct irq_data *data)
+{
+ struct gpio_chip *chip = irq_data_get_irq_chip_data(data);
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);
+ unsigned int line = data->hwirq;
+
+ writeq(GPIO_INTR_ENA_W1S, gpio->register_base + intr_reg(line));
+}
+
+static int thunderx_gpio_irq_set_type(struct irq_data *data,
+ unsigned int flow_type)
+{
+ struct gpio_chip *chip = irq_data_get_irq_chip_data(data);
+ struct thunderx_gpio *gpio = container_of(chip, struct thunderx_gpio, chip);
+ unsigned int line = data->hwirq;
+ u64 bit_cfg;
+
+ irqd_set_trigger_type(data, flow_type);
+
+ bit_cfg = GLITCH_FILTER_400NS | GPIO_BIT_CFG_INT_EN;
+
+ if (flow_type & IRQ_TYPE_EDGE_BOTH) {
+ irq_set_handler_locked(data, handle_edge_irq);
+ bit_cfg |= GPIO_BIT_CFG_INT_TYPE;
+ } else {
+ irq_set_handler_locked(data, handle_level_irq);
+ }
+
+ raw_spin_lock(&gpio->lock);
+ if (flow_type & (IRQ_TYPE_EDGE_FALLING | IRQ_TYPE_LEVEL_LOW)) {
+ bit_cfg |= GPIO_BIT_CFG_PIN_XOR;
+ set_bit(line, gpio->invert_mask);
+ } else {
+ clear_bit(line, gpio->invert_mask);
+ }
+ clear_bit(line, gpio->od_mask);
+ writeq(bit_cfg, gpio->register_base + bit_cfg_reg(line));
+ raw_spin_unlock(&gpio->lock);
+
+ return IRQ_SET_MASK_OK;
+}
+
+/*
+ * Interrupts are chained from underlying MSI-X vectors. We have
+ * these irq_chip functions to be able to handle level triggering
+ * semantics and other acknowledgment tasks associated with the GPIO
+ * mechanism.
+ */
+static struct irq_chip thunderx_gpio_irq_chip = {
+ .name = "GPIO",
+ .irq_enable = thunderx_gpio_irq_unmask,
+ .irq_disable = thunderx_gpio_irq_mask,
+ .irq_ack = thunderx_gpio_irq_ack,
+ .irq_mask = thunderx_gpio_irq_mask,
+ .irq_mask_ack = thunderx_gpio_irq_mask_ack,
+ .irq_unmask = thunderx_gpio_irq_unmask,
+ .irq_set_type = thunderx_gpio_irq_set_type,
+ .irq_request_resources = thunderx_gpio_irq_request_resources,
+ .irq_release_resources = thunderx_gpio_irq_release_resources,
+ .flags = IRQCHIP_SET_TYPE_MASKED
+};

This looks wrong.

If you're calling devm_request_irq() on *every* *single* *irq* like you
do here, you are dealing with a hieararchical irqdomain, not a linear one,
and GPIOLIB_IRQCHIP may not be used.

Look at drivers/gpio/gpio-xgene-sb.c for inspiration. That is the only
hierarchical GPIO irqdomain I have right now.

Consult Marc Zyngier's IRQ domain lecture if you have time:
https://www.youtube.com/watch?v=YE8cRHVIM4E

If you have ideas how to combine hierarchical irqdomain and GPIOLIB_IRQCHIP
pls help out.

+ gpio->irqdev_entries = devm_kzalloc(dev,
+ sizeof(struct thunderx_irqdev) * ngpio,
+ GFP_KERNEL);
+ if (!gpio->irqdev_entries) {
+ err = -ENOMEM;
+ goto out;
+ }

I think this is overkill. Use hierarchical irqdomain.

I will look into it. I suspect it will require more lines of driver code to implement it than what I have here (that does actually work).

Thanks for looking at this,
David