Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20230512042806.3438373-1-chris.packham@alliedtelesis.co.nz>
In-Reply-To: <20230512042806.3438373-1-chris.packham@alliedtelesis.co.nz>
From:   Linus Walleij <linus.walleij@linaro.org>
Date:   Fri, 12 May 2023 09:56:17 +0200
Message-ID: <CACRpkdYz9ipNTo2ORXKWy5Q4uCpKL=9Gd+kK76pestX7Onuz-Q@mail.gmail.com>
Subject: Re: [PATCH] gpiolib: Avoid side effects in gpio_is_visible()
To:     Chris Packham <chris.packham@alliedtelesis.co.nz>
Cc:     brgl@bgdev.pl, johan@kernel.org, andy.shevchenko@gmail.com,
        maz@kernel.org, Ben Brown <ben.brown@alliedtelesis.co.nz>,
        linux-gpio@vger.kernel.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Fri, May 12, 2023 at 6:28=E2=80=AFAM Chris Packham
<chris.packham@alliedtelesis.co.nz> wrote:

> Calling gpiod_to_irq() creates an irq_desc for the GPIO.

Normally gpiod_to_irq() should not have side effects, it's just
a helper function that is there to translate a descriptor to the
corresponding IRQ number.

> This is not
> something that should happen when just exporting the GPIO via sysfs. The
> effect of this was observed as triggering a warning in
> gpiochip_disable_irq() when kexec()ing after exporting a GPIO.
>
> Remove the call to gpiod_to_irq() from gpio_is_visible(). The actual
> intended creation of the irq_desc comes via edge_store() when requested
> by the user.
>
> Fixes: ebbeba120ab2 ("gpio: sysfs: fix gpio attribute-creation race")
> Signed-off-by: Chris Packham <chris.packham@alliedtelesis.co.nz>

I have a hard time understanding this fix.

The problem is rather this see gpiolib.c:

int gpiod_to_irq(const struct gpio_desc *desc)
{
        struct gpio_chip *gc;
        int offset;

        /*
         * Cannot VALIDATE_DESC() here as gpiod_to_irq() consumer semantics
         * requires this function to not return zero on an invalid descript=
or
         * but rather a negative error number.
         */
        if (!desc || IS_ERR(desc) || !desc->gdev || !desc->gdev->chip)
                return -EINVAL;

        gc =3D desc->gdev->chip;
        offset =3D gpio_chip_hwgpio(desc);
        if (gc->to_irq) {
                int retirq =3D gc->to_irq(gc, offset);

Here gc->to_irq() is called unconditionally.

Since this is using gpiolib_irqchip this ->to_irq() will be
gpiochip_to_irq() and that finally ends in the call:

return irq_create_mapping(domain, offset);

which seems to be the problem here. Why is this a problem?
The IRQ mappings are dynamic, meaning they are created
on-demand, but for this hardware it should be fine to essentially
just call irq_create_mapping() on all of them as the device
is created, we just don't do it in order to save space.

I don't understand why calling irq_create_mapping(domain, offset);
creates a problem? It's just a mapping between a GPIO and the
corresponding IRQ. What am I missing here?

Yours,
Linus Walleij