On Tue, Mar 02, 2021 at 12:24:41PM -0800, Hugh Dickins wrote: > On Tue, 2 Mar 2021, Michal Hocko wrote: > > [Cc Johannes for awareness and fixup Nick's email] > > > > On Tue 02-03-21 01:34:51, Zhou Guanghui wrote: > > > When split page, the memory cgroup info recorded in first page is > > > not copied to tail pages. In this case, when the tail pages are > > > freed, the uncharge operation is not performed. As a result, the > > > usage of this memcg keeps increasing, and the OOM may occur. > > > > > > So, the copying of first page's memory cgroup info to tail pages > > > is needed when split page. > > > > I was not aware that alloc_pages_exact is used for accounted allocations > > but git grep told me otherwise so this is not a theoretical one. Both > > users (arm64 and s390 kvm) are quite recent AFAICS. split_page is also > > used in dma allocator but I got lost in indirection so I have no idea > > whether there are any users there. > > Yes, it's a bit worrying that such a low-level thing as split_page() > can now get caught up in memcg accounting, but I suppose that's okay. > > I feel rather strongly that whichever way it is done, THP splitting > and split_page() should use the same interface to memcg. > > And a look at mem_cgroup_split_huge_fixup() suggests that nowadays > there need to be css_get()s too - or better, a css_get_many(). > > Its #ifdef CONFIG_TRANSPARENT_HUGEPAGE should be removed, rename > it mem_cgroup_split_page_fixup(), and take order from caller.
+1 There is already a split_page_owner() in both these places as well which does a similar thing. Mabye we can match that by calling it split_page_memcg() and having it take a nr of pages? > Though I've never much liked that separate pass: would it be > better page by page, like this copy_page_memcg() does? Though > mem_cgroup_disabled() and css_getting make that less appealing. Agreed on both counts. mem_cgroup_disabled() is a jump label and would be okay, IMO, but the refcounting - though it is (usually) per-cpu - adds at least two branches and rcu read locking.