Re: [kvm-unit-tests PATCH 1/8] s390x: lib: Extend bitops

[Thomas Huth <thuth@redhat.com>]

On 18/08/2021 10.39, Janosch Frank wrote:
> On 8/18/21 10:20 AM, Thomas Huth wrote:
>> On 13/08/2021 13.31, Janosch Frank wrote:
>>> On 8/13/21 10:32 AM, Claudio Imbrenda wrote:
>>>> On Fri, 13 Aug 2021 07:36:08 +0000
>>>> Janosch Frank <frankja@linux.ibm.com> wrote:
>>>>
>>>>> Bit setting and clearing is never bad to have.
>>>>>
>>>>> Signed-off-by: Janosch Frank <frankja@linux.ibm.com>
>>>>> ---
>>>>>    lib/s390x/asm/bitops.h | 102
>>>>> +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 102
>>>>> insertions(+)
>>>>>
>>>>> diff --git a/lib/s390x/asm/bitops.h b/lib/s390x/asm/bitops.h
>>>>> index 792881ec..f5612855 100644
>>>>> --- a/lib/s390x/asm/bitops.h
>>>>> +++ b/lib/s390x/asm/bitops.h
>>>>> @@ -17,6 +17,78 @@
>>>>>    
>>>>>    #define BITS_PER_LONG	64
>>>>>    
>>>>> +static inline unsigned long *bitops_word(unsigned long nr,
>>>>> +					 const volatile unsigned
>>>>> long *ptr) +{
>>>>> +	unsigned long addr;
>>>>> +
>>>>> +	addr = (unsigned long)ptr + ((nr ^ (nr & (BITS_PER_LONG -
>>>>> 1))) >> 3);
>>>>> +	return (unsigned long *)addr;
>>>>
>>>> why not just
>>>>
>>>> return ptr + (nr / BITS_PER_LONG);
>>>>
>>>>> +}
>>>>> +
>>>>> +static inline unsigned long bitops_mask(unsigned long nr)
>>>>> +{
>>>>> +	return 1UL << (nr & (BITS_PER_LONG - 1));
>>>>> +}
>>>>> +
>>>>> +static inline uint64_t laog(volatile unsigned long *ptr, uint64_t
>>>>> mask) +{
>>>>> +	uint64_t old;
>>>>> +
>>>>> +	/* load and or 64bit concurrent and interlocked */
>>>>> +	asm volatile(
>>>>> +		"	laog	%[old],%[mask],%[ptr]\n"
>>>>> +		: [old] "=d" (old), [ptr] "+Q" (*ptr)
>>>>> +		: [mask] "d" (mask)
>>>>> +		: "memory", "cc" );
>>>>> +	return old;
>>>>> +}
>>>>
>>>> do we really need the artillery (asm) here?
>>>> is there a reason why we can't do this in C?
>>>
>>> Those are the interlocked/atomic instructions and even though we don't
>>> exactly need them right now I wanted to add them for completeness.
>>
>> I think I agree with Claudio - unless we really need them, we should not
>> clog the sources with arbitrary inline assembly functions.
> 
> Alright I can trim it down
> 
>>
>>> We might be able to achieve the same via compiler functionality but this
>>> is not my expertise. Maybe Thomas or David have a few pointers for me?
>>
>> I'm not an expert with atomic builtins either, but what's the point of this
>> at all? Loading a value and OR-ing something into the value in one go?
>> What's that good for?
> 
> Well it's a block-concurrent interlocked-update load, or and store.
> I.e. it loads the data from the ptr and copies it into [old] then ors
> the mask and stores it back to the ptr address.
> 
> The instruction name "load and or" does not represent the full actions
> of the instruction.

Ok, thanks, that makes more sense now, but you could at least have mentioned 
this in the comment that you added in front of it :-)

Anyway, I guess it's easier to use the builtin atomic functions like 
__atomic_or_fetch() for stuff like this in case we ever need it.

  Thomas