[patch] mm, pcp: allow restoring percpu_pagelist_fraction default

Oleg Drokin green at linuxhacker.ru
Wed Jun 4 02:19:12 UTC 2014


Hello!

On Jun 3, 2014, at 9:22 PM, David Rientjes wrote:
> 
> @@ -5849,21 +5850,32 @@ int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
> int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
> 	void __user *buffer, size_t *length, loff_t *ppos)
> {
> +	const int old_percpu_pagelist_fraction = percpu_pagelist_fraction;
> 	struct zone *zone;
> -	unsigned int cpu;
> 	int ret;
> 
> 	ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
> -	if (!write || (ret < 0))
> +	if (!write || ret < 0)
> 		return ret;
> 
> +	/* Sanity checking to avoid pcp imbalance */
> +	if (percpu_pagelist_fraction &&
> +	    percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
> +		percpu_pagelist_fraction = old_percpu_pagelist_fraction;
> +		return -EINVAL;
> +	}
> +
> +	/* No change? */
> +	if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
> +		return 0;
> +
> 	mutex_lock(&pcp_batch_high_lock);
> 	for_each_populated_zone(zone) {
> -		unsigned long  high;
> -		high = zone->managed_pages / percpu_pagelist_fraction;
> +		unsigned int cpu;
> +
> 		for_each_possible_cpu(cpu)
> -			pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
> -					 high);
> +			pageset_set_high_and_batch(zone,
> +					per_cpu_ptr(zone->pageset, cpu));

BTW, I just realized this version is racy (as was the previous one). 
A parallel writer could write a value of 0 while we are in the middle of pageset_set_high_and_batch
and it's possible that'll result in division by zero still.
Also it's possible an incorrect value might set for some of the zones.

I imagine we might want to expand the lock area all the way up to before the proc_dointvec_minmax call jsut to be extra safe.

> 	}
> 	mutex_unlock(&pcp_batch_high_lock);
> 	return 0;

Bye,
    Oleg


More information about the devel mailing list