Re: [PATCH 1/2] asm-generic: rwsem: ensure sem->cnt is only accessed via atomic_long_*

[Peter Hurley <peter@hurleysoftware.com>]

On 02/28/2014 07:13 AM, Will Deacon wrote:
> On Thu, Feb 27, 2014 at 05:28:24AM +0000, Davidlohr Bueso wrote:
>> On Fri, 2014-02-21 at 17:22 +0000, Will Deacon wrote:
>>> The asm-generic rwsem implementation directly acceses sem->cnt when
>>> performing a __down_read_trylock operation. Whilst this is probably safe
>>> on all architectures, we should stick to the atomic_long_* API and use
>>> atomic_long_read instead.
>>>
>>> Signed-off-by: Will Deacon <will.deacon@arm.com>
>>> ---
>>>   include/asm-generic/rwsem.h | 2 +-
>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/include/asm-generic/rwsem.h b/include/asm-generic/rwsem.h
>>> index bb1e2cdeb9bf..75af612f54f8 100644
>>> --- a/include/asm-generic/rwsem.h
>>> +++ b/include/asm-generic/rwsem.h
>>> @@ -41,7 +41,7 @@ static inline int __down_read_trylock(struct rw_semaphore *sem)
>>>   {
>>>   	long tmp;
>>>
>>> -	while ((tmp = sem->count) >= 0) {
>>> +	while ((tmp = atomic_long_read((atomic_long_t *)&sem->count)) >= 0) {
>>
>> That's pretty ugly, how about having infinite look and just do the tmp
>> assign separately from the conditional?
>>
>> It also looks like a cpu_relax() could help here between iterations.

This is the read trylock so no cpu_relax().

>> Other than that:
>>
>> Reviewed-by: Davidlohr Bueso <davidlohr@hp.com>
>
> Actually, we should make that cmpxchg an atomic_long_cmpxchg, so the extra
> diff ends up looking like below. It's ugly adding a cpu_relax(), since you
> only want it when the cmpxchg fails (and we don't have such logic in the
> asm-generic __atomic_add_unless, for example).
>
> Will
>
> --->8
>
> diff --git a/include/asm-generic/rwsem.h b/include/asm-generic/rwsem.h
> index 603a0a11e592..2b6401f9e428 100644
> --- a/include/asm-generic/rwsem.h
> +++ b/include/asm-generic/rwsem.h
> @@ -40,14 +40,16 @@ static inline void __down_read(struct rw_semaphore *sem)
>   static inline int __down_read_trylock(struct rw_semaphore *sem)
>   {
>          long tmp;
> +       atomic_long_t *cnt = (atomic_long_t *)&sem->count;

The shared rwsem failure paths (kernel/locking/rwsem_xadd.c) peek at
sem->count as long type, so this isn't really necessary.

>
> -       while ((tmp = atomic_long_read((atomic_long_t *)&sem->count)) >= 0) {
> -               if (tmp == cmpxchg(&sem->count, tmp,
> -                                  tmp + RWSEM_ACTIVE_READ_BIAS)) {
> -                       return 1;
> -               }
> -       }
> -       return 0;
> +       do {
> +               tmp = atomic_long_read(cnt);
> +               if (tmp < 0)
> +                       return 0;
> +       } while (tmp != atomic_long_cmpxchg(cnt, tmp,
> +                                           tmp + RWSEM_ACTIVE_READ_BIAS));
> +
> +       return 1;
>   }

Regards,
Peter Hurley