On Thu, Oct 01, 2020 at 07:00:36PM +0200, Michal Koutný wrote: > Hi. > > On Wed, Sep 30, 2020 at 05:27:10PM -0700, Roman Gushchin <[email protected]> wrote: > > @@ -369,8 +371,12 @@ enum page_memcg_data_flags { > > */ > > static inline struct mem_cgroup *page_memcg(struct page *page) > > { > > + unsigned long memcg_data = page->memcg_data; > > + > > VM_BUG_ON_PAGE(PageSlab(page), page); > > - return (struct mem_cgroup *)page->memcg_data; > > + VM_BUG_ON_PAGE(memcg_data & MEMCG_DATA_OBJCGS, page); > > + > > + return (struct mem_cgroup *)(memcg_data & ~MEMCG_DATA_FLAGS_MASK); > > } > Shouldn't this change go also into page_memcg_rcu()? (I don't think the > current single user (workingset_activation() would pass a non-slab > kernel page but for consistency sake.)
+1 > Alternatively, I'm thinking why (in its single use) is there > page_memcg_rcu() a separate function to page_memcg() (cross memcg page > migration?). It goes back to commit 55779ec759ccc3c12b917b3712a7716e1140c652. The activation code is the only path where page migration is not excluded. Because unlike with page state statistics, we don't really mind a race when counting an activation event. I do think there is a bug, though: mem_cgroup_move_account() should use WRITE_ONCE() on page->mem_cgroup.
