Re: [PATCH 06/10] Input: sparcspkr - use cleanup facility for device_node

[Javier Carrasco <javier.carrasco.cruz@gmail.com>]

On 11/10/2024 00:01, Javier Carrasco wrote:
> On 10/10/2024 23:43, Al Viro wrote:
>> On Thu, Oct 10, 2024 at 11:25:56PM +0200, Javier Carrasco wrote:
>>>
>>> diff --git a/drivers/input/misc/sparcspkr.c b/drivers/input/misc/sparcspkr.c
>>> index 20020cbc0752..bb1c732c8f95 100644
>>> --- a/drivers/input/misc/sparcspkr.c
>>> +++ b/drivers/input/misc/sparcspkr.c
>>> @@ -188,7 +188,6 @@ static int bbc_beep_probe(struct platform_device *op)
>>>  {
>>>  	struct sparcspkr_state *state;
>>>  	struct bbc_beep_info *info;
>>> -	struct device_node *dp;
>>>  	int err = -ENOMEM;
>>>  
>>>  	state = kzalloc(sizeof(*state), GFP_KERNEL);
>>> @@ -199,14 +198,13 @@ static int bbc_beep_probe(struct platform_device *op)
>>>  	state->event = bbc_spkr_event;
>>>  	spin_lock_init(&state->lock);
>>>  
>>> -	dp = of_find_node_by_path("/");
>>>  	err = -ENODEV;
>>> +	struct device_node *dp __free(device_node) = of_find_node_by_path("/");
>>>  	if (!dp)
>>>  		goto out_free;
>>
>> Sigh...  See that
>>         state = kzalloc(sizeof(*state), GFP_KERNEL);
>> 	if (!state)
>> 		goto out_err;
>> above?
>>
>> IOW, this will quietly generate broken code if built with gcc (and refuse to
>> compile with clang).  Yeah, this one is trivially fixed (return -ENOMEM instead
>> of a goto), but...
>>
>> __cleanup() can be useful, but it's really *not* safe for blind use; you
>> need to watch out for changed scopes (harmless in case of device_node)
>> and for gotos (broken here).
> 
> Hi Al Viro,
> 
> sorry, but I think I don't get you. First, I don't see sparc64 as a
> supported architecture for clang in the Linux documentation. Maybe the
> documentation is not up-to-date, but I tried to compile with clang and
> it seems to be true that it is not supported. Anyway, that is not the
> issue here.
> 
> Second, I might be missing something about the scopes you are
> mentioning. 'state' gets allocated before the device_node is declared,
> and when the device_node is declared and its initialization fails, it
> should jump to 'out_free' to free 'state', shouldn't it? Sorry if I have
> overlooked something here.
> 
> Thank your for your feedback and best regards,
> Javier Carrasco
> 


I think that the issue you are talking about is that the goto will
trigger the cleanup function of the device_node, which will not be
initialized at that point.

Yes, the goto before the device_node declaration hurts, and a return as
you said would be better.

Thanks and best regards,
Javier Carrasco