Re: [External] Re: [PATCH v2 3/6] mm: hugetlb: fix a race between freeing and dissolving the page

[Muchun Song <songmuchun@bytedance.com>]

On Fri, Jan 8, 2021 at 4:43 PM Michal Hocko <mhocko@suse.com> wrote:
>
> On Thu 07-01-21 23:11:22, Muchun Song wrote:
> > On Thu, Jan 7, 2021 at 10:11 PM Michal Hocko <mhocko@suse.com> wrote:
> > >
> > > On Thu 07-01-21 20:59:33, Muchun Song wrote:
> > > > On Thu, Jan 7, 2021 at 8:38 PM Michal Hocko <mhocko@suse.com> wrote:
> > > [...]
> > > > > Right. Can we simply back off in the dissolving path when ref count is
> > > > > 0 && PageHuge() if list_empty(page->lru)? Is there any other scenario
> > > > > when the all above is true and the page is not being freed?
> > > >
> > > > The list_empty(&page->lru) may always return false.
> > > > The page before freeing is on the active list
> > > > (hstate->hugepage_activelist).Then it is on the free list
> > > > after freeing. So list_empty(&page->lru) is always false.
> > >
> > > The point I was trying to make is that the page has to be enqueued when
> > > it is dissolved and freed. If the page is not enqueued then something
> > > racing. But then I have realized that this is not a great check to
> > > detect the race because pages are going to be released to buddy
> > > allocator and that will reuse page->lru again. So scratch that and sorry
> > > for the detour.
> > >
> > > But that made me think some more and one way to reliably detect the race
> > > should be PageHuge() check in the freeing path. This is what dissolve
> > > path does already. PageHuge becomes false during update_and_free_page()
> > > while holding the hugetlb_lock. So can we use that?
> >
> > It may make the thing complex. Apart from freeing it to the
> > buddy allocator, free_huge_page also does something else for
> > us. If we detect the race in the freeing path, if it is not a HugeTLB
> > page, the freeing path just returns. We also should move those
> > things to the dissolve path. Right?
>
> Not sure what you mean. Dissolving is a subset of the freeing path. It
> effectivelly only implements the update_and_free_page branch (aka free
> to buddy). It skips some of the existing steps because it believes it
> sees a freed page. But as you have pointed out this is racy and I
> strongly suspect it is simply wrong in those assumptions. E.g. hugetlb
> cgroup accounting can get wrong right?

OK. I know what you mean. The update_and_free_page should
do the freeing which is similar to __free_huge_page().

>
> The more I think about it the more I think that dissolving path should
> simply have a common helper with  __free_huge_page that would release
> the huge page to the allocator. The only thing that should be specific
> to the dissolving path is HWpoison handling. It shouldn't be playing
> with accounting and what not. Btw the HWpoison handling is suspicious as
> well, a lost race would mean this doesn't happen. But maybe there is
> some fixup handled later on...
>
> > But I find a tricky problem to solve. See free_huge_page().
> > If we are in non-task context, we should schedule a work
> > to free the page. We reuse the page->mapping. If the page
> > is already freed by the dissolve path. We should not touch
> > the page->mapping. So we need to check PageHuge().
> > The check and llist_add() should be protected by
> > hugetlb_lock. But we cannot do that. Right? If dissolve
> > happens after it is linked to the list. We also should
> > remove it from the list (hpage_freelist). It seems to make
> > the thing more complex.
>
> I am not sure I follow you here but yes PageHuge under hugetlb_lock
> should be the reliable way to check for the race. I am not sure why we
> really need to care about mapping or other state.

CPU0:                               CPU1:
free_huge_page(page)
  if (PageHuge(page))
                                    dissolve_free_huge_page(page)
                                      spin_lock(&hugetlb_lock)
                                      update_and_free_page(page)
                                      spin_unlock(&hugetlb_lock)
    llist_add(page->mapping)
    // the mapping is corrupted

The PageHuge(page) and llist_add() should be protected by
hugetlb_lock. Right? If so, we cannot hold hugetlb_lock
in free_huge_page() path.

> --
> Michal Hocko
> SUSE Labs