Re: [PATCH 2/2] locking/rtmutex: optimize rt_mutex_cmpxchgs

[Thomas Gleixner]

Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx> writes:

> Checking l->owner first to skip time cost cmpxchgs.

I don't see what that buys.

It actually adds an extra conditional in the non-contended case, which
is the case we are optimizing for.

In the contended case, i.e. when l->owner != c the cmpxchg cost is
completely irrelevant compared to the slowpath costs.

>  #ifndef CONFIG_DEBUG_RT_MUTEXES
> -# define rt_mutex_cmpxchg_acquire(l,c,n) (cmpxchg_acquire(&l->owner, c, n) == c)
> -# define rt_mutex_cmpxchg_release(l,c,n) (cmpxchg_release(&l->owner, c, n) == c)
> +# define rt_mutex_cmpxchg_acquire(l, c, n)	\
> +			(l->owner == c && cmpxchg_acquire(&l->owner, c, n) == c)
> +# define rt_mutex_cmpxchg_release(l, c, n)	\
> +			(l->owner == c && cmpxchg_release(&l->owner, c, n) == c)

This kind of micro-optimizing is more than dubious especially w/o
numbers backing up the benefit.

Thanks,

        tglx