On Mon, 23 Jul 2018, David Rientjes wrote: > > > The huge zero page can be reclaimed under memory pressure and, if it is, > > > it is attempted to be allocted again with gfp flags that attempt memory > > > compaction that can become expensive. If we are constantly under memory > > > pressure, it gets freed and reallocated millions of times always trying to > > > compact memory both directly and by kicking kcompactd in the background. > > > > > > It likely should also be per node. > > > > Have you benchmarked making the non-huge zero page per-node? > > > > Not since we disable it :) I will, though. The more concerning issue for > us, modulo CVE-2017-1000405, is the cpu cost of constantly directly > compacting memory for allocating the hzp in real time after it has been > reclaimed. We've observed this happening tens or hundreds of thousands > of times on some systems. It will be 2MB per node on x86 if the data > suggests we should make it NUMA aware, I don't think the cost is too high > to leave it persistently available even under memory pressure if > use_zero_page is enabled. > Measuring access latency to 4GB of memory on Naples I observe ~6.7% slower access latency intrasocket and ~14% slower intersocket. use_zero_page is currently a simple thp flag, meaning it rejects writes where val != !!val, so perhaps it would be best to overload it with additional options? I can imagine 0x2 defining persistent allocation so that the hzp is not freed when the refcount goes to 0 and 0x4 defining if the hzp should be per node. Implementing persistent allocation fixes our concern with it, so I'd like to start there. Comments?
