Re: [RFC PATCH] cleanup: make scoped_guard() to be return-friendly

[Przemek Kitszel <przemyslaw.kitszel@intel.com>]

On 9/27/24 17:04, Dan Carpenter wrote:
> On Fri, Sep 27, 2024 at 04:08:30PM +0200, Przemek Kitszel wrote:
>> On 9/27/24 09:31, Dan Carpenter wrote:
>>> On Thu, Sep 26, 2024 at 03:41:38PM +0200, Przemek Kitszel wrote:
>>>> diff --git a/include/linux/cleanup.h b/include/linux/cleanup.h
>>>> index d9e613803df1..6b568a8a7f9c 100644
>>>> --- a/include/linux/cleanup.h
>>>> +++ b/include/linux/cleanup.h
>>>> @@ -168,9 +168,16 @@ static inline class_##_name##_t class_##_name##ext##_constructor(_init_args) \
>>>>    #define __guard_ptr(_name) class_##_name##_lock_ptr
>>>> -#define scoped_guard(_name, args...)					\
>>>> -	for (CLASS(_name, scope)(args),					\
>>>> -	     *done = NULL; __guard_ptr(_name)(&scope) && !done; done = (void *)1)
>>>> +#define scoped_guard(_name, args...)	\
>>>> +	__scoped_guard_labeled(__UNIQUE_ID(label), _name, args)
>>>> +
>>>> +#define __scoped_guard_labeled(_label, _name, args...)	\
>>>> +	if (0)						\
>>>> +		_label: ;				\
>>>> +	else						\
>>>> +		for (CLASS(_name, scope)(args);		\
>>>> +		     __guard_ptr(_name)(&scope), 1;	\
>>>                                                  ^^^
>>>> +		     ({ goto _label; }))
>>>
>>> Remove the ", 1".  The point of the __guard_ptr() condition is for try_locks
>>> but the ", 1" means they always succeed.  The only try lock I can find in
>>
>> You are right that the __guard_ptr() is conditional for the benefit of
>> try_locks. But here we have unconditional lock. And removing ", 1" part
>> makes compiler complaining with the very same message:
>> error: control reaches end of non-void function [-Werror=return-type]
>>
>> so ", 1" part is on purpose and must stay there to aid compiler.
>>
>>> the current tree is tsc200x_esd_work().
> 
> Obviously, we can't break stuff and also checking __guard_ptr(_name)(&scope) is
> pointless if we're going to ignore the return value.
> 
> But, sure, I get that we want to the compiler to know that regular spin_lock()
> is going to succeed and spin_trylock() might not.  As a static checker
> developer, I want that as well.  Currently, whenever someone creates a new class
> of locks, I have to add a couple lines to Smatch to add this information.  It's
> not a huge deal, but it would be nice to avoid this.
> 
> I did a `git grep scoped_guard | grep try` and I think tsc200x_esd_work() is the
> only place which actually uses try locks with scoped_guard().  If it's just the
> one, then why don't we create a scoped_guard_trylock() macro?

Most of the time it is just easier to bend your driver than change or
extend the core of the kernel.

There is actually scoped_cond_guard() which is a trylock variant.

scoped_guard(mutex_try, &ts->mutex) you have found is semantically
wrong and must be fixed.

---
I have received also a bot message about "if (x) scoped_guard(y, z)"
usage (without braces), so will need to adjust it too.