On Tue, Sep 13, 2016 at 02:22:25PM +0200, Linus Walleij wrote:

On Tue, Sep 13, 2016 at 11:33 AM, Mika Westerberg
[off-list ref] wrote:

quoted

On Tue, Sep 13, 2016 at 11:18:49AM +0200, Linus Walleij wrote:

quoted

On Mon, Sep 12, 2016 at 3:11 PM, Mika Westerberg
[off-list ref] wrote:

quoted

On Mon, Sep 12, 2016 at 09:04:44PM +0800, Phidias Chiang wrote:

quoted

On Mon, Sep 12, 2016 at 12:04:01PM +0300, Mika Westerberg wrote:

quoted

OK, I see what is going on now. When I changed handle_simple_irq to
handle_bad_irq, the IRQ core in __irq_do_set_handler() thinks the
handler is uninstalled and masks the line.

If you change handle_bad_irq to handle_simple_irq, in call to
gpiochip_irqchip_add(), does it work then?

Yes it does :), thank you for the support!

Thanks for testing.

So we need to use handle_simple_irq here instead.

Linus, do you see any problems with that?

I need to see the patch in its context with a commit message,
I can't figure it out from the thread.

handle_simple_irq() is for something generic not level- or
edge-triggered. If you support specific triggers only, it
should not be used.

Nominally assigning handle_bad_irq() until a specific
edge or level is requested is the right thing to do, since
the IRQ is really not configured for anything at all and
hence has undefined behaviour.

For Cherryview/Braswell some interrupts are actually configured by the
BIOS but they are routed directly to the I/O-APIC and are supposed to be
handled without involvement of the GPIO driver (an example of this is
the ACPI SCI interrupt). However, INTMASK GPIO register can still be
used to mask the interrupt in question.

A-ha! But why are you registering a irqdomain entry for an interrupt
that cannot be used, hm?

Unfortunately there is no way to figure out from the hardware (or
firmware) whether the interrupt is supposed to be used by the GPIO
driver or something else.

quoted

So when we specify handle_bad_irq as handler the IRQ core thinks the
handler is being uninstalled and masks the interrupt.

You should not register any handle for it.

In fact, IMO the irqdomain should reject it being mapped, as it is not
for the kernel to use.

Check:
drivers/irqchip/irq-vic.c

We supply a u32 to the driver from the device tree named "valid-mask".
This has bits set to 1 for the valid (to be mapped) IRQs and zero
for those we may not touch.

So in our vic_irqdomain_map() call we have:

/* Skip invalid IRQs, only register handlers for the real ones */
if (!(v->valid_sources & (1 << hwirq)))
        return -EPERM;

So it can never be mapped.

Also notice:

/* create an IRQ mapping for each valid IRQ */
for (i = 0; i < fls(valid_sources); i++)
        if (valid_sources & (1 << i))
                irq_create_mapping(v->domain, i);

So we only create mappings where there are valid IRQs,
skipping over any "holes" in the mapping.

We should do something like this.

To make this play nicely with gpiolib_irqchip_add() I suggest
adding a bitmap valid_mask to struct gpio_chip() in
include/linux/gpio/driver.h
inside the CONFIG_GPIOLIB_IRQCHIP

I guess

u32 bitmap[MAX_IRQS_FOR_A_GPIO_CHIP];

Then augment the generic GPIO IRQCHIP helpers in
drivers/gpio/gpiolib.c per above so that the invalid
IRQs can't be mapped.

The driver would just:

/* Mark line N as invalid: used by BIOS */
set_bit(&chip->valid_mask, N);

Otherwise this probably works but it forces us to hardcode these
"special" lines in the driver and I would like to avoid that if
possible.