[PATCH v3] nptl: Optimize trylock for high cache contention workloads (BZ #33704)

[Yann Droneaud]

Hi,

Le 16/12/2025 à 03:21, Sunil K Pandey a écrit :
> Check lock availability before acquisition to reduce cache line
> bouncing.  Significantly improves trylock throughput on multi-core
> systems under heavy contention.
>
> mutex trylock throughput benchmark runtime
>
> Core  baseline         patch
> ====  ===============  =============
> 80    66Min 32Sec      45Sec
> 24     8Min 4Sec       14Sec
> 12     5Min 14Sec      34Sec
> 8      3Min 40Sec      20Sec
>
> Throughput data(Core count: 80)
>
> tcount baseline	pthread_mutex_trylock  Improvement
> ====== ======== =====================  ===========
> 80     232112   3119301                13.44x
> 64     247737   937424                 3.78x
> 32     441441   888738                 2.01x
> 16     863201   1529987                1.77x
> 8      1587950  2775892                1.75x
> 4      3208012  6487683                2.02x
> 2      7729917  11811184               1.53x
>
> Tested on x86_64.

Which one ?

Any results for other architectures ?


> Fixes BZ #33704.
>
> Add recursive.
> ---
>   nptl/pthread_mutex_trylock.c | 8 ++++++--
>   1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/nptl/pthread_mutex_trylock.c b/nptl/pthread_mutex_trylock.c
> index 94621cc254..fe9ba8da67 100644
> --- a/nptl/pthread_mutex_trylock.c
> +++ b/nptl/pthread_mutex_trylock.c
> @@ -47,7 +47,8 @@ ___pthread_mutex_trylock (pthread_mutex_t *mutex)
>   	  return 0;
>   	}
>   
> -      if (lll_trylock (mutex->__data.__lock) == 0)
> +      if (atomic_load_relaxed (&(mutex->__data.__lock)) == 0
> +	  && lll_trylock (mutex->__data.__lock) == 0)
>   	{
>   	  /* Record the ownership.  */
>   	  mutex->__data.__owner = id;
> @@ -60,7 +61,10 @@ ___pthread_mutex_trylock (pthread_mutex_t *mutex)
>       case PTHREAD_MUTEX_TIMED_NP:
>       case PTHREAD_MUTEX_ADAPTIVE_NP:
>       case PTHREAD_MUTEX_ERRORCHECK_NP:
> -      if (lll_trylock (mutex->__data.__lock) != 0)
> +      /* Mutex type is already loaded, lock check overhead should
> +         be minimal.  */
> +      if (atomic_load_relaxed (&(mutex->__data.__lock)) != 0
> +	  || lll_trylock (mutex->__data.__lock) != 0)
>   	break;
>   
>         /* Record the ownership.  */

as lll_trylock() is a macro, could the relaxed load be inserted into it, 
providing the improvement to other user of lll_trylock() ?

Regards.