Re: [PATCH] irq: Request and release resources for chained IRQs

On Fri, 7 Dec 2018, Linus Walleij wrote:

Sorry for answering late.

On Thu, Nov 29, 2018 at 2:31 PM Linus Walleij [off-list ref] wrote:

quoted

This irqchip is in drivers/pinctrl/qcom/pinctrl-msm.c and known
as TLMM.

TLMM == Totally Lost Manufacturer Mess?

quoted

However, the irqchip core helpers in GPIO assumes that
the .irq_request_resources() callback is always called before
.irq_enable(), and the latter is where module refcount and
also gpiochip_lock_as_irq() is normally called.

When .irq_enable() is called without .irq_request_resources()
having been called first, it seems like we are enabling
an IRQ on a GPIO line that has not first been locked to be
used as IRQ and we get the above warning. This happens since
as of
commit 461c1a7d4733d ("gpiolib: override irq_enable/disable")
this is strictly assumed for all GPIO-based IRQs.

As it seems reasonable to assume that .irq_request_resources()
is always strictly called before .irq_enable(), we add the
irq_[request|release]_resources() functions to the internal
API and call them also when adding a chained irqchip to
an IRQ.

I am a bit uncertain about the call site for
irq_released_resources() inside chip.c, but it appears this
path is for uninstalling a chained handler.

You cannot invoke those callbacks from __irq_do_set_handler() as that is
holding the irq descriptor lock with interrupts disabled.

The calling convention for irq_released_resources() is that it's called
with the bus lock held (if the chip provides the bus lock callbacks), but
definitely not with desc->lock held.

Needs more thought. Btw, the uninstall path does not invoke irq_deactive()
either.... I so hate that chained handler mess....

Thanks,

	tglx