Re: [PATCH v8 20/22] counter: Implement events_queue_size sysfs attribute

[David Lechner <david@lechnology.com>]

On 2/25/21 6:03 PM, William Breathitt Gray wrote:
> On Sun, Feb 21, 2021 at 03:51:40PM +0000, Jonathan Cameron wrote:
>> On Thu, 18 Feb 2021 19:32:16 +0900
>> William Breathitt Gray <vilhelm.gray@gmail.com> wrote:
>>
>>> On Sun, Feb 14, 2021 at 06:11:46PM +0000, Jonathan Cameron wrote:
>>>> On Fri, 12 Feb 2021 21:13:44 +0900
>>>> William Breathitt Gray <vilhelm.gray@gmail.com> wrote:
>>>>    
>>>>> The events_queue_size sysfs attribute provides a way for users to
>>>>> dynamically configure the Counter events queue size for the Counter
>>>>> character device interface. The size is in number of struct
>>>>> counter_event data structures. The number of elements will be rounded-up
>>>>> to a power of 2 due to a requirement of the kfifo_alloc function called
>>>>> during reallocation of the queue.
>>>>>
>>>>> Cc: Oleksij Rempel <o.rempel@pengutronix.de>
>>>>> Signed-off-by: William Breathitt Gray <vilhelm.gray@gmail.com>
>>>>> ---
>>>>>   Documentation/ABI/testing/sysfs-bus-counter |  8 +++++++
>>>>>   drivers/counter/counter-chrdev.c            | 23 +++++++++++++++++++
>>>>>   drivers/counter/counter-chrdev.h            |  2 ++
>>>>>   drivers/counter/counter-sysfs.c             | 25 +++++++++++++++++++++
>>>>>   4 files changed, 58 insertions(+)
>>>>>
>>>>> diff --git a/Documentation/ABI/testing/sysfs-bus-counter b/Documentation/ABI/testing/sysfs-bus-counter
>>>>> index 847e96f19d19..f6cb2a8b08a7 100644
>>>>> --- a/Documentation/ABI/testing/sysfs-bus-counter
>>>>> +++ b/Documentation/ABI/testing/sysfs-bus-counter
>>>>> @@ -212,6 +212,14 @@ Description:
>>>>>   		both edges:
>>>>>   			Any state transition.
>>>>>   
>>>>> +What:		/sys/bus/counter/devices/counterX/events_queue_size
>>>>> +KernelVersion:	5.13
>>>>> +Contact:	linux-iio@vger.kernel.org
>>>>> +Description:
>>>>> +		Size of the Counter events queue in number of struct
>>>>> +		counter_event data structures. The number of elements will be
>>>>> +		rounded-up to a power of 2.
>>>>> +
>>>>>   What:		/sys/bus/counter/devices/counterX/name
>>>>>   KernelVersion:	5.2
>>>>>   Contact:	linux-iio@vger.kernel.org
>>>>> diff --git a/drivers/counter/counter-chrdev.c b/drivers/counter/counter-chrdev.c
>>>>> index 16f02df7f73d..53eea894e13f 100644
>>>>> --- a/drivers/counter/counter-chrdev.c
>>>>> +++ b/drivers/counter/counter-chrdev.c
>>>>> @@ -375,6 +375,29 @@ void counter_chrdev_remove(struct counter_device *const counter)
>>>>>   	cdev_del(&counter->chrdev);
>>>>>   }
>>>>>   
>>>>> +int counter_chrdev_realloc_queue(struct counter_device *const counter,
>>>>> +				 size_t queue_size)
>>>>> +{
>>>>> +	int err;
>>>>> +	DECLARE_KFIFO_PTR(events, struct counter_event);
>>>>> +	unsigned long flags;
>>>>> +
>>>>> +	/* Allocate new events queue */
>>>>> +	err = kfifo_alloc(&events, queue_size, GFP_ATOMIC);
>>>>
>>>> Is there any potential for losing events?
>>>
>>> We take the events_list_lock down below so we're safe against missing an
>>> event, but past events currently unread in the queue will be lost.
>>>
>>> Shortening the size of the queue is inherently a destructive process if
>>> we have more events in the current queue than can fit in the new queue.
>>> Because we a liable to lose some events in such a case, I think it's
>>> best to keep the behavior of this reallocation consistent and have it
>>> provide a fresh empty queue every time, as opposed to sometimes dropping
>>> events and sometimes not.
>>>
>>> I also suspect an actual user would be setting the size of their queue
>>> to the required amount before they begin watching events, rather than
>>> adjusting it sporadically during a live operation.
>>>
>>
>> Absolutely agree.   As such I wonder if you are better off enforcing this
>> behaviour?  If the cdev is open for reading, don't allow the fifo to be
>> resized.
>>
>> Jonathan
> 
> I can't really think of a good reason not to, so let's enforce it: if
> the cdev is open, then we'll return an EINVAL if the user attempts to
> resize the queue.
> 
> What is a good way to check for this condition? Should I just call
> kref_read() and see if it's greater than 1? For example, in
> counter_chrdev_realloc_queue():
> 
> 	if (kref_read(&counter->dev.kobj.kref) > 1)
> 		return -EINVAL;
> 
> William Breathitt Gray
> 

Wouldn't EBUSY make more sense?