On Fri 17-04-20 09:41:04, Shakeel Butt wrote: > On Fri, Apr 17, 2020 at 9:03 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > > > On Fri 17-04-20 22:43:43, Alex Shi wrote: > > > This patch fold MEMCG_SWAP feature into kernel as default function. That > > > required a short size memcg id for each of page. As Johannes mentioned > > > > > > "the overhead of tracking is tiny - 512k per G of swap (0.04%).' > > > > > > So all swapout page could be tracked for its memcg id. > > > > I am perfectly OK with dropping the CONFIG_MEMCG_SWAP. The code that is > > guarded by it is negligible and the resulting code is much easier to > > read so no objection on that front. I just do not really see any real > > reason to flip the default for cgroup v1. Why do we want/need that? > > > > Yes, the changelog is lacking the motivation of this change. This is > proposed by Johannes and I was actually expecting the patch from him. > The motivation is to make the things simpler for per-memcg LRU locking > and workingset for anon memory (Johannes has described these really > well, lemme find the email). If we keep the differentiation between > cgroup v1 and v2, then there is actually no point of this cleanup as > per-memcg LRU locking and anon workingset still has to handle the > !do_swap_account case. All those details really have to go into the changelog. I have to say that I still do not understand why the actual accounting swap or not makes any difference for per per-memcg LRU. Especially when your patch keeps the kernel command line parameter still in place. Anyway, it would be much more simpler to have a patch that drops the CONFIG_MEMCG_SWAP and a separate one which switches the default beahvior. I am not saying I am ok with the later but if the justification is convincing then I might change my mind. -- Michal Hocko SUSE Labs
