On Tue, Feb 19, 2013 at 04:21:28PM -0800, Tejun Heo wrote:
On Tue, Feb 19, 2013 at 4:17 PM, Minchan Kim minchan@kernel.org wrote:
Should we really enable memcg for just pressure notificaion in embedded side? I didn't check the size(cgroup + memcg) and performance penalty but I don't want to add unnecessary overhead if it is possible. Do you have a plan to support it via global knob(ie, /proc/mempressure), NOT memcg?
That should be handled by mempressure at the root cgroup. If that adds significant amount of overhead code or memory-wise, we just need to fix root cgroup handling in memcg. No reason to further complicate the interface which already is pretty complex.
For what it worth, I agree here. Even if we decide to make another interface to vmpressure (which, say, would not require memcg), then it is better to keep the API the same: eventfd + control file. That way, API/ABI-wise there will be no differnce between memcg and non-memcg kernels, which is cool.
Thanks, Anton