On Sun, 23 Oct 2011, Gilad Ben-Yossef wrote:

> +/* Which CPUs have per cpu pages  */
> +cpumask_var_t cpus_with_pcp;
> +static DEFINE_PER_CPU(unsigned long, total_cpu_pcp_count);

This increases the cache footprint of a hot vm path. Is it possible to do
the same than what you did for slub? Run a loop over all zones when
draining to check for remaining pcp pages and build the set of cpus
needing IPIs temporarily while draining?