Re: [PATCH 1/2] locking/qspinlock: Store lock holder cpu in lock if feasible

[Peter Zijlstra]

On Sat, Jul 11, 2020 at 02:21:27PM -0400, Waiman Long wrote:

> +static DEFINE_PER_CPU_READ_MOSTLY(u8, pcpu_lockval) = _Q_LOCKED_VAL;
>  
>  /*
>   * We must be able to distinguish between no-tail and the tail at 0:0,
> @@ -138,6 +139,19 @@ struct mcs_spinlock *grab_mcs_node(struct mcs_spinlock *base, int idx)
>  
>  #define _Q_LOCKED_PENDING_MASK (_Q_LOCKED_MASK | _Q_PENDING_MASK)
>  
> +static __init int __init_pcpu_lockval(void)
> +{
> +	int cpu;
> +
> +	for_each_possible_cpu(cpu) {
> +		u8 lockval = (cpu + 2 < _Q_LOCKED_MASK - 1) ? cpu + 2
> +							    : _Q_LOCKED_VAL;
> +		per_cpu(pcpu_lockval, cpu) = lockval;
> +	}
> +	return 0;
> +}
> +early_initcall(__init_pcpu_lockval);

> +	u8 lockval = this_cpu_read(pcpu_lockval);

Urgh... so you'd rather read a guaranteed cold line than to use
smp_processor_id(), which we already use anyway?

I'm skeptical this helps anything, and it certainly makes the code more
horrible :-(