On Mon, 8 Mar 2021 20:47:31 +0000 Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> On Mon, Mar 08, 2021 at 12:42:27PM -0800, Andrew Morton wrote:
> > On Mon, 8 Mar 2021 09:41:38 +0100 Michal Hocko <mhocko@xxxxxxxx> wrote:
> >
> > > On Fri 05-03-21 15:58:40, Andrew Morton wrote:
> > > > On Fri, 5 Mar 2021 12:52:52 +0100 Michal Hocko <mhocko@xxxxxxxx> wrote:
> > > >
> > > > > On Thu 04-03-21 07:40:53, Zhou Guanghui wrote:
> > > > > > As described in the split_page function comment, for the non-compound
> > > > > > high order page, the sub-pages must be freed individually. If the
> > > > > > memcg of the fisrt page is valid, the tail pages cannot be uncharged
> > > > > > when be freed.
> > > > > >
> > > > > > For example, when alloc_pages_exact is used to allocate 1MB continuous
> > > > > > physical memory, 2MB is charged(kmemcg is enabled and __GFP_ACCOUNT is
> > > > > > set). When make_alloc_exact free the unused 1MB and free_pages_exact
> > > > > > free the applied 1MB, actually, only 4KB(one page) is uncharged.
> > > > > >
> > > > > > Therefore, the memcg of the tail page needs to be set when split page.
> > > > > >
> > > > >
> > > > > As already mentioned there are at least two explicit users of
> > > > > __GFP_ACCOUNT with alloc_exact_pages added recently. It would be good to
> > > > > mention that explicitly and maybe even mention 7efe8ef274024 resp.
> > > > > c419621873713 so that it is clear this is not just a theoretical issue.
> > > >
> > > > I added
> > > >
> > > > : Michel:
> > > > :
> > > > : There are at least two explicit users of __GFP_ACCOUNT with
> > > > : alloc_exact_pages added recently. See 7efe8ef274024 ("KVM: arm64:
> > > > : Allocate stage-2 pgd pages with GFP_KERNEL_ACCOUNT") and c419621873713
> > > > : ("KVM: s390: Add memcg accounting to KVM allocations"), so this is not
> > > > : just a theoretical issue.
> > > >
> > > > And should we cc:stable on this one?
> > >
> > > Somebody more familiar with iommu dma allocation layer should have a
> > > look as well (__iommu_dma_alloc_pages) so that we know whether there are
> > > kernels outside of the above two ones mentioned above that need a fix.
> > > But in general this sounds like a good fit for the stable tree.
> >
> > OK. I reversed the order of these two patches so we don't need to
> > burden -stable with a cosmetic rename.
>
> Eek, no.
>
> The alloc_pages_exact() is done to pages that _aren't_ compound.
> So you have to pass the number of pages to the memcg split function,
> because a non-compound page doesn't know the size of its allocation.
Ah, OK, the patch title fooled me.
It should have been a three-patch series, really
1: add nr_pages arg to mem_cgroup_split_huge_fixup()
2: call mem_cgroup_split_huge_fixup() when splitting
3: rename mem_cgroup_split_huge_fixup() to split_page_memcg()
That way, the third cosmetic patch could be deferred so we don't feed
the cosmetic renaming into -stable.
But whatever, the rename isn't a big deal so I'll go with the 2-patch
series as sent for -stable.
