[PATCH] locking/atomic: Make test_and_*_bit() ordered on failure

[Hector Martin]

On 16/08/2022 23.04, Will Deacon wrote:
>> diff --git a/Documentation/atomic_bitops.txt b/Documentation/atomic_bitops.txt
>> index 093cdaefdb37..d8b101c97031 100644
>> --- a/Documentation/atomic_bitops.txt
>> +++ b/Documentation/atomic_bitops.txt
>> @@ -59,7 +59,7 @@ Like with atomic_t, the rule of thumb is:
>>   - RMW operations that have a return value are fully ordered.
>>  
>>   - RMW operations that are conditional are unordered on FAILURE,
>> -   otherwise the above rules apply. In the case of test_and_{}_bit() operations,
>> +   otherwise the above rules apply. In the case of test_and_set_bit_lock(),
>>     if the bit in memory is unchanged by the operation then it is deemed to have
>>     failed.
> 
> The next sentence is:
> 
>   | Except for a successful test_and_set_bit_lock() which has ACQUIRE
>   | semantics and clear_bit_unlock() which has RELEASE semantics.
> 
> so I think it reads a bit strangely now. How about something like:
> 
> 
> diff --git a/Documentation/atomic_bitops.txt b/Documentation/atomic_bitops.txt
> index 093cdaefdb37..3b516729ec81 100644
> --- a/Documentation/atomic_bitops.txt
> +++ b/Documentation/atomic_bitops.txt
> @@ -59,12 +59,15 @@ Like with atomic_t, the rule of thumb is:
>   - RMW operations that have a return value are fully ordered.
>  
>   - RMW operations that are conditional are unordered on FAILURE,
> -   otherwise the above rules apply. In the case of test_and_{}_bit() operations,
> -   if the bit in memory is unchanged by the operation then it is deemed to have
> -   failed.
> +   otherwise the above rules apply. For the purposes of ordering, the
> +   test_and_{}_bit() operations are treated as unconditional.
>  
> -Except for a successful test_and_set_bit_lock() which has ACQUIRE semantics and
> -clear_bit_unlock() which has RELEASE semantics.
> +Except for:
> +
> + - test_and_set_bit_lock() which has ACQUIRE semantics on success and is
> +   unordered on failure;
> +
> + - clear_bit_unlock() which has RELEASE semantics.
>  
>  Since a platform only has a single means of achieving atomic operations
>  the same barriers as for atomic_t are used, see atomic_t.txt.

Makes sense! I'll send a v2 with that in a couple of days if nothing
else comes up.

>> diff --git a/include/asm-generic/bitops/atomic.h b/include/asm-generic/bitops/atomic.h
>> index 3096f086b5a3..71ab4ba9c25d 100644
>> --- a/include/asm-generic/bitops/atomic.h
>> +++ b/include/asm-generic/bitops/atomic.h
>> @@ -39,9 +39,6 @@ arch_test_and_set_bit(unsigned int nr, volatile unsigned long *p)
>>  	unsigned long mask = BIT_MASK(nr);
>>  
>>  	p += BIT_WORD(nr);
>> -	if (READ_ONCE(*p) & mask)
>> -		return 1;
>> -
>>  	old = arch_atomic_long_fetch_or(mask, (atomic_long_t *)p);
>>  	return !!(old & mask);
>>  }
>> @@ -53,9 +50,6 @@ arch_test_and_clear_bit(unsigned int nr, volatile unsigned long *p)
>>  	unsigned long mask = BIT_MASK(nr);
>>  
>>  	p += BIT_WORD(nr);
>> -	if (!(READ_ONCE(*p) & mask))
>> -		return 0;
>> -
>>  	old = arch_atomic_long_fetch_andnot(mask, (atomic_long_t *)p);
>>  	return !!(old & mask);
> 
> I suppose one sad thing about this is that, on arm64, we could reasonably
> keep the READ_ONCE() path with a DMB LD (R->RW) barrier before the return
> but I don't think we can express that in the Linux memory model so we
> end up in RmW territory every time.

You'd need a barrier *before* the READ_ONCE(), since what we're trying
to prevent is a consumer from writing to the value without being able to
observe the writes that happened prior, while this side read the old
value. A barrier after the READ_ONCE() doesn't do anything, as that read
is the last memory operation in this thread (of the problematic sequence).

At that point, I'm not sure DMB LD / early read / LSE atomic would be
any faster than just always doing the LSE atomic?

- Hector