Re: [PATCH 29/31] media: track media device unregister in progress

[Shuah Khan]

On 01/28/2016 01:42 PM, Shuah Khan wrote:
> On 01/28/2016 10:28 AM, Mauro Carvalho Chehab wrote:
>> Em Thu, 28 Jan 2016 10:04:24 -0700
>> Shuah Khan <shuahkh@xxxxxxxxxxxxxxx> escreveu:
>>
>>> On 01/28/2016 10:01 AM, Mauro Carvalho Chehab wrote:
>>>> Em Wed,  6 Jan 2016 13:27:18 -0700
>>>> Shuah Khan <shuahkh@xxxxxxxxxxxxxxx> escreveu:
>>>>   
>>>>> Add support to track media device unregister in progress
>>>>> state to prevent more than one driver entering unregister.
>>>>> This enables fixing the general protection faults while
>>>>> snd-usb-audio was cleaning up media resources for pcm
>>>>> streams and mixers. In this patch a new interface is added
>>>>> to return the unregister in progress state. Subsequent
>>>>> patches to snd-usb-audio and au0828-core use this interface
>>>>> to avoid entering unregister and attempting to unregister
>>>>> entities and remove devnodes while unregister is in progress.
>>>>> Media device unregister removes entities and interface nodes.  
>>>>
>>>> Hmm... isn't the spinlock enough to serialize it? It seems weird the
>>>> need of an extra bool here to warrant that this is really serialized.
>>>>   
>>>
>>> The spinlock and check for media_devnode_is_registered(&mdev->devnode)
>>> aren't enough to ensure only one driver enters the unregister. 
>>>
>>> Please
>>> note that the devnode isn't marked unregistered until the end in
>>> media_device_unregister().
>>
>> I guess the call to:
>> 	device_remove_file(&mdev->devnode.dev, &dev_attr_model);
>>
>> IMO, This should be, instead, at media_devnode_unregister().
>>
>> Then, we can change the logic at media_devnode_unregister() to:
>>
>> void media_devnode_unregister(struct media_devnode *mdev)
>> {
>> 	mutex_lock(&media_devnode_lock);
>>
>> 	/* Check if mdev was ever registered at all */
>> 	if (!media_devnode_is_registered(mdev)) {
>> 		mutex_unlock(&media_devnode_lock);
>> 		return;
>> 	}
>>
>> 	clear_bit(MEDIA_FLAG_REGISTERED, &mdev->flags);
>> 	mutex_unlock(&media_devnode_lock);
>> 	device_remove_file(&mdev->devnode.dev, &dev_attr_model);
>> 	device_unregister(&mdev->dev);
>> }
>>
>> This sounds enough to avoid device_unregister() or device_remove_file()
>> to be called twice.
>>
> 
> I can give it a try. There might other problems that could
> result from media device being a devres in this case. The
> last put_device on the usbdev parent device (media device
> is created as devres for this), all device resources get
> released. That might have to be solved in a different way.
> 
> For now I will see if your solution works.

Hi Mauro,

Moving device_remove_file() won't be easy without
making more changes. The file is created in
media_device_regsiter() and all the attributes are
handled in media-device.c

One solution I can think of is clearing the
MEDIA_FLAG_REGISTERED bit very early in
media_device_unregister()

--- a/drivers/media/media-device.c
+++ b/drivers/media/media-device.c
@@ -759,6 +759,9 @@ void media_device_unregister(struct media_device *mdev)
                return;
        }
 
+       /* Protect unregister path - clear MEDIA_FLAG_REGISTERED */
+       clear_bit(MEDIA_FLAG_REGISTERED, &mdev->devnode.flags);
+
        /* Remove all entities from the media device */
        list_for_each_entry_safe(entity, next, &mdev->entities, graph_obj.list)
                __media_device_unregister_entity(entity);

and changing media_devnode_unregister() to simply call
device_unregister(&mdev->dev);

Again clearing MEDIA_FLAG_REGISTERED bit in
media_device_unregister() some problems.
For one thing clearing this bit should be
done holding media_devnode_lock. It can be
done cleanly if we do the following:

How about if we split media_devnode_unregister()
into twp ohases:

media_devnode_start_unregister()
to clear this bit. It can do:

media_devnode_start_unregister()
{
	mutex_lock(&media_devnode_lock);
	if (!media_devnode_is_registered(mdev)) {
		mutex_unlock(&media_devnode_lock);
		return;
	}
	clear_bit(MEDIA_FLAG_REGISTERED, &mdev->flags);
	mutex_unlock(&media_devnode_lock);
}

then:media_device_unregister(struct media_device *mdev)
will call this first thing and then hold mdev->lock
do the rest and the call media_devnode_unregister()
and which will be changed to as follows:

--- a/drivers/media/media-devnode.c
+++ b/drivers/media/media-devnode.c
@@ -274,13 +274,6 @@ error:
 
 void media_devnode_unregister(struct media_devnode *mdev)
 {
-       /* Check if mdev was ever registered at all */
-       if (!media_devnode_is_registered(mdev))
-               return;
-
-       mutex_lock(&media_devnode_lock);
-       clear_bit(MEDIA_FLAG_REGISTERED, &mdev->flags);
-       mutex_unlock(&media_devnode_lock);
        device_unregister(&mdev->dev);
 }

thanks,
-- Shuah

-- 
Shuah Khan
Sr. Linux Kernel Developer
Open Source Innovation Group
Samsung Research America (Silicon Valley)
shuahkh@xxxxxxxxxxxxxxx | (970) 217-8978