[PATCH 2/3] arm: include: asm: atomic.h: use 'unsigned int' and 'atomic' instead of 'unsigned long' for atomic_clear_mask()

[Chen Gang]

On 10/10/2013 05:58 PM, Will Deacon wrote:
> On Thu, Oct 10, 2013 at 03:34:02AM +0100, Chen Gang wrote:
>> In current kernel wide source, for arm, only s390 scsi drivers use
>> atomic_clear_mask(), now, s390 itself need use 'unsigned int' and
>> 'atomic_t', so need match s390's atomic_clear_mask().
>>
>>
>> Signed-off-by: Chen Gang <gang.chen at asianux.com>
>> ---
>>  arch/arm/include/asm/atomic.h |   13 +++++++------
>>  1 files changed, 7 insertions(+), 6 deletions(-)
>>
>> diff --git a/arch/arm/include/asm/atomic.h b/arch/arm/include/asm/atomic.h
>> index da1c77d..0832a7f 100644
>> --- a/arch/arm/include/asm/atomic.h
>> +++ b/arch/arm/include/asm/atomic.h
>> @@ -134,9 +134,10 @@ static inline int atomic_cmpxchg(atomic_t *ptr, int old, int new)
>>  	return oldval;
>>  }
>>  
>> -static inline void atomic_clear_mask(unsigned long mask, unsigned long *addr)
>> +static inline void atomic_clear_mask(unsigned int mask, atomic_t *ptr)
>>  {
>> -	unsigned long tmp, tmp2;
>> +	unsigned int tmp;
> 
> I reckon this should be int (the mask parameter is unsigned, but
> atomic_t.counter is signed).

For 'ldrex' and 'strex' (loading/storing instruction), it is really
better to match 'atomic_t.counter', but for 'bic' (operating
instruction), it is better to match 'mask'.

In my opinion, for signed/unsigned, 'operating' has higher priority than
'loading/storing' (especially for *mask functions, by default, suggest
using unsigned).

Commonly, for loading/storing (e.g. 'ldrex', 'strex'), must be sure of
bits wide (signed/unsigned will not cause real issues), but for
operating, signed/unsigned may cause real issues.


> 
>> +	unsigned long tmp2;
>>  
>>  	__asm__ __volatile__("@ atomic_clear_mask\n"
>>  "1:	ldrex	%0, [%3]\n"
>> @@ -144,8 +145,8 @@ static inline void atomic_clear_mask(unsigned long mask, unsigned long *addr)
>>  "	strex	%1, %0, [%3]\n"
>>  "	teq	%1, #0\n"
>>  "	bne	1b"
>> -	: "=&r" (tmp), "=&r" (tmp2), "+Qo" (*addr)
>> -	: "r" (addr), "Ir" (mask)
>> +	: "=&r" (tmp), "=&r" (tmp2), "+Qo" (ptr->counter)
>> +	: "r" (&ptr->counter), "Ir" (mask)
>>  	: "cc");
>>  }
>>  
>> @@ -197,12 +198,12 @@ static inline int atomic_cmpxchg(atomic_t *v, int old, int new)
>>  	return ret;
>>  }
>>  
>> -static inline void atomic_clear_mask(unsigned long mask, unsigned long *addr)
>> +static inline void atomic_clear_mask(unsigned int mask, atomic_t *v)
>>  {
>>  	unsigned long flags;
>>  
>>  	raw_local_irq_save(flags);
>> -	*addr &= ~mask;
>> +	v->counter &= ~mask;
>>  	raw_local_irq_restore(flags);
>>  }
> 
> This is now identical to asm-generic/atomic.h. I wonder whether we could
> just #include that file for the ARMv6 case? You'd need to check the
> differences carefully.
> 

If most of functions for ARMv6 case can use "asm-generic/atomic.h", your
idea sounds good to me, although we don't need 'atomic_set_mask' (it is
inconsistent with 'atomic_clear_mask' in "asm-generic/atomic.h").

> Finally, I still question the need for the clear_mask function anyway. We
> don't implement set_mask, and these functions are only used by either other
> arch code or in drivers that don't work on ARM anyway.
> 

Hmm... can we remove atomic_*_mask() for both arm and arm64?

It seems before get a conclusion, it is necessary to let arm and arm64
pass 'allmodconfig' firstly (and then try to remove these functions to
see the compiling result).

I will/should try 'allmodconfig' for them, but excuse me, it needs
waiting (I am just trying 'arc' architecture with 'allmodconfig', and
already delayed, because I have no enough time resources on it :-( ).

> Will
> 
> 

Thanks.
-- 
Chen Gang