Re: [PATCH] Fix sysfs_notify and kobject_uevent in fan_alarm_notify

[Guenter Roeck <linux@roeck-us.net>]

On Wed, Jun 19, 2019 at 12:00:37PM +0200, Christian Schneider wrote:
> Ok, seems that moving that block, where fan_alarm_init is called down, isn't
> a problem on master.
> I need to use 4.9, cause we have an embedded device that needs patches, and
> they are only available for older kernels.
> 
> In 4.9, fan_alarm_init() does fan_data->alarm = alarm;
> When fan_alarm_init() is called after
> devm_hwmon_device_register_with_groups(), then fan_data->alarm is not yet
> set, so the sysfs attribute is not created.
> 
> when I move the fan_data->alarm = alarm; out of fan_alarm_init() before
> devm_hwmon_device_register_with_groups() and the fan_alarm_init() after,
> then it works, and doesn't have a race condition.
> 
> On master, populating fan_data->alarm_gpio is done in
> gpio_fan_get_of_data(fan_data); as far as I understand, so moving the
> fan_alarm_init() block further down shouldn't be a problem.
> 
> I don't know, how I can test this with master kernel. Unfortunately, the
> device, I'm using doesn't have a recent kernel. And on my normal (desktop)
> pc, I don't have a gpio fan.
> How should I proceed now?
> 

Essentially, we'll need two patches: One to move fan_alarm_init() after
devm_hwmon_device_register_with_groups(), and then the notification patch.
For the backport to v4.9, the first patch will have to be slightly modified:
The first part will a simple
	fan_data->alarm = pdata->alarm;
at the location where fan_alarm_init() is called now, the second part is to move
the call to fan_alarm_init() after devm_hwmon_device_register_with_groups().

It is ok to test the result with your 4.9 based kernel; we'll just have to be
careful when reviewing the mainline patch.

Thanks,
Guenter

> 
> BR, Christian
> 
> Am 19.06.2019 um 09:47 schrieb Christian Schneider:
> >Hi,
> >unfrotunately I found a problem with my patch. It didn't appear on the
> >devices, where I tested, only when distributing the patch to more devices,
> >it occurs sometimes.
> >
> >There is a possible race condition. In gpio_fan_probe, the IRQ handler for
> >the fan alarm is registered before the hwmon_dev is initialised.
> >So it is possible, that the handler executes, while fandata->hwmon_dev is
> >still unitialised and the handler runs in an invalid pointer dereference.
> >
> >The obvious fix, moving the
> >/* Configure alarm GPIO if available. */
> >if (fan_data->alarm_gpio) {
> >     err = fan_alarm_init(fan_data);
> >     if (err)
> >         return err;
> >}
> >block after
> >
> >/* Make this driver part of hwmon class. */
> >fan_data->hwmon_dev =
> >     devm_hwmon_device_register_with_groups(dev,
> >                            "gpio_fan", fan_data,
> >                            gpio_fan_groups);
> >
> >doesn't work. Sysfs doesn't have a fan1_alarm attribute anymore then.
> >
> >It is not yet clear to me, why creation of the attribute fails in this
> >case. I need to investigate further.
> >In the meanwhile, please revert my patch.
> >
> >I'm sorry for the disturbance I created....
> >
> >BR, Christian
> >
> >
> >Am 17.06.2019 um 18:04 schrieb Guenter Roeck:
> >>On Mon, Jun 17, 2019 at 11:33:43AM +0200, cschneider@radiodata.biz wrote:
> >>>From: Christian Schneider <christian@ch-sc.de>
> >>>
> >>>sysfs_notify and kobject_uevent are passed the wrong kobject.
> >>>that why notifications can't be received and uevents have the wrong
> >>>path.
> >>>this patch fixes this.
> >>>
> >>>Signed-off-by: Christian Schneider <cschneider@radiodata.biz>
> >>
> >>Applied.
> >>
> >>In the future, please version your patches, add a change log,
> >>and reflect the subsystem and driver in the subject line.
> >>
> >>Thanks,
> >>Guenter
> >>
> >>>---
> >>>  drivers/hwmon/gpio-fan.c | 4 ++--
> >>>  1 file changed, 2 insertions(+), 2 deletions(-)
> >>>
> >>>diff --git a/drivers/hwmon/gpio-fan.c b/drivers/hwmon/gpio-fan.c
> >>>index 84753680a4e8..76377791ff0e 100644
> >>>--- a/drivers/hwmon/gpio-fan.c
> >>>+++ b/drivers/hwmon/gpio-fan.c
> >>>@@ -54,8 +54,8 @@ static void fan_alarm_notify(struct work_struct *ws)
> >>>      struct gpio_fan_data *fan_data =
> >>>          container_of(ws, struct gpio_fan_data, alarm_work);
> >>>-    sysfs_notify(&fan_data->dev->kobj, NULL, "fan1_alarm");
> >>>-    kobject_uevent(&fan_data->dev->kobj, KOBJ_CHANGE);
> >>>+    sysfs_notify(&fan_data->hwmon_dev->kobj, NULL, "fan1_alarm");
> >>>+    kobject_uevent(&fan_data->hwmon_dev->kobj, KOBJ_CHANGE);
> >>>  }
> >>>  static irqreturn_t fan_alarm_irq_handler(int irq, void *dev_id)