[PATCH 01/23] gpio: sysfs: fix memory leaks and device hotplug

[PATCH 01/23] gpio: sysfs: fix memory leaks and device hotplug

[Johan Hovold]

Unregister GPIOs requested through sysfs at chip remove to avoid leaking
the associated memory and sysfs entries.

The stale sysfs entries prevented the gpio numbers from being exported
when the gpio range was later reused (e.g. at device reconnect).

This also fixes the related module-reference leak.

Note that kernfs makes sure that any on-going sysfs operations finish
before the class devices are unregistered and that further accesses
fail.

The chip exported flag is used to prevent gpiod exports during removal.
This also makes it harder to trigger, but does not fix, the related race
between gpiochip_remove and export_store, which is really a race with
gpiod_request that needs to be addressed separately.

Also note that this would prevent the crashes (e.g. NULL-dereferences)
at reconnect that affects pre-3.18 kernels, as well as use-after-free on
operations on open attribute files on pre-3.14 kernels (prior to
kernfs).

Fixes: d8f388d8dc8d ("gpio: sysfs interface")
Cc: stable <stable@vger.kernel.org>	# v2.6.27: 01cca93a9491
Signed-off-by: Johan Hovold <johan@kernel.org>
---
 drivers/gpio/gpiolib-sysfs.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/drivers/gpio/gpiolib-sysfs.c b/drivers/gpio/gpiolib-sysfs.c
index 7722ed53bd65..af3bc7a8033b 100644
--- a/drivers/gpio/gpiolib-sysfs.c
+++ b/drivers/gpio/gpiolib-sysfs.c
@@ -551,6 +551,7 @@ static struct class gpio_class = {
  */
 int gpiod_export(struct gpio_desc *desc, bool direction_may_change)
 {
+	struct gpio_chip	*chip;
 	unsigned long		flags;
 	int			status;
 	const char		*ioname = NULL;
@@ -568,8 +569,16 @@ int gpiod_export(struct gpio_desc *desc, bool direction_may_change)
 		return -EINVAL;
 	}
 
+	chip = desc->chip;
+
 	mutex_lock(&sysfs_lock);
 
+	/* check if chip is being removed */
+	if (!chip || !chip->exported) {
+		status = -ENODEV;
+		goto fail_unlock;
+	}
+
 	spin_lock_irqsave(&gpio_lock, flags);
 	if (!test_bit(FLAG_REQUESTED, &desc->flags) ||
 	     test_bit(FLAG_EXPORT, &desc->flags)) {
@@ -783,12 +792,15 @@ void gpiochip_unexport(struct gpio_chip *chip)
 {
 	int			status;
 	struct device		*dev;
+	struct gpio_desc *desc;
+	unsigned int i;
 
 	mutex_lock(&sysfs_lock);
 	dev = class_find_device(&gpio_class, NULL, chip, match_export);
 	if (dev) {
 		put_device(dev);
 		device_unregister(dev);
+		/* prevent further gpiod exports */
 		chip->exported = false;
 		status = 0;
 	} else
@@ -797,6 +809,13 @@ void gpiochip_unexport(struct gpio_chip *chip)
 
 	if (status)
 		chip_dbg(chip, "%s: status %d\n", __func__, status);
+
+	/* unregister gpiod class devices owned by sysfs */
+	for (i = 0; i < chip->ngpio; i++) {
+		desc = &chip->desc[i];
+		if (test_and_clear_bit(FLAG_SYSFS, &desc->flags))
+			gpiod_free(desc);
+	}
 }
 
 static int __init gpiolib_sysfs_init(void)
-- 
2.0.5

Re: [PATCH 01/23] gpio: sysfs: fix memory leaks and device hotplug

[Linus Walleij]

On Tue, Apr 21, 2015 at 5:42 PM, Johan Hovold <johan@kernel.org> wrote:

> Unregister GPIOs requested through sysfs at chip remove to avoid leaking
> the associated memory and sysfs entries.
>
> The stale sysfs entries prevented the gpio numbers from being exported
> when the gpio range was later reused (e.g. at device reconnect).
>
> This also fixes the related module-reference leak.
>
> Note that kernfs makes sure that any on-going sysfs operations finish
> before the class devices are unregistered and that further accesses
> fail.
>
> The chip exported flag is used to prevent gpiod exports during removal.
> This also makes it harder to trigger, but does not fix, the related race
> between gpiochip_remove and export_store, which is really a race with
> gpiod_request that needs to be addressed separately.
>
> Also note that this would prevent the crashes (e.g. NULL-dereferences)
> at reconnect that affects pre-3.18 kernels, as well as use-after-free on
> operations on open attribute files on pre-3.14 kernels (prior to
> kernfs).
>
> Fixes: d8f388d8dc8d ("gpio: sysfs interface")
> Cc: stable <stable@vger.kernel.org>     # v2.6.27: 01cca93a9491
> Signed-off-by: Johan Hovold <johan@kernel.org>

Patch applied for fixes.

I worry a bit about what userspaces do out there, but they
cannot reasonably have behaviours tied to in-flight removal
of GPIO chips, that would be bizarre.

Yours,
Linus Walleij

Re: [PATCH 01/23] gpio: sysfs: fix memory leaks and device hotplug

[Johan Hovold]

On Wed, Apr 29, 2015 at 11:44:18PM +0200, Linus Walleij wrote:
> On Tue, Apr 21, 2015 at 5:42 PM, Johan Hovold <johan@kernel.org> wrote:
> 
> > Unregister GPIOs requested through sysfs at chip remove to avoid leaking
> > the associated memory and sysfs entries.
> >
> > The stale sysfs entries prevented the gpio numbers from being exported
> > when the gpio range was later reused (e.g. at device reconnect).
> >
> > This also fixes the related module-reference leak.
> >
> > Note that kernfs makes sure that any on-going sysfs operations finish
> > before the class devices are unregistered and that further accesses
> > fail.
> >
> > The chip exported flag is used to prevent gpiod exports during removal.
> > This also makes it harder to trigger, but does not fix, the related race
> > between gpiochip_remove and export_store, which is really a race with
> > gpiod_request that needs to be addressed separately.
> >
> > Also note that this would prevent the crashes (e.g. NULL-dereferences)
> > at reconnect that affects pre-3.18 kernels, as well as use-after-free on
> > operations on open attribute files on pre-3.14 kernels (prior to
> > kernfs).
> >
> > Fixes: d8f388d8dc8d ("gpio: sysfs interface")
> > Cc: stable <stable@vger.kernel.org>     # v2.6.27: 01cca93a9491
> > Signed-off-by: Johan Hovold <johan@kernel.org>
> 
> Patch applied for fixes.
> 
> I worry a bit about what userspaces do out there, but they
> cannot reasonably have behaviours tied to in-flight removal
> of GPIO chips, that would be bizarre.

You shouldn't worry too much; even before this patch userspace would see
an -ENODEV when accessing an open sysfs attribute file of a disconnected
device as kernfs would orphan the file -- only now without the associated
leaks and crashes. ;)

Johan