Re: [PATCH v3 2/4] mm/hwpoison: check if a subpage of a hugetlb folio is raw HWPOISON

[Jiaqi Yan <jiaqiyan@google.com>]

On Mon, Jul 10, 2023 at 8:16 AM Jiaqi Yan <jiaqiyan@google.com> wrote:
>
> On Fri, Jul 7, 2023 at 7:57 PM Miaohe Lin <linmiaohe@huawei.com> wrote:
> >
> > On 2023/7/8 4:19, Jiaqi Yan wrote:
> > > Add the functionality, is_raw_hwp_subpage, to tell if a subpage of a
> > > hugetlb folio is a raw HWPOISON page. This functionality relies on
> > > RawHwpUnreliable to be not set; otherwise hugepage's raw HWPOISON list
> > > becomes meaningless.
> > >
> > > is_raw_hwp_subpage needs to hold hugetlb_lock in order to synchronize
> > > with __get_huge_page_for_hwpoison, who iterates and inserts an entry to
> > > raw_hwp_list. llist itself doesn't ensure insertion is synchornized with
> > > the iterating used by __is_raw_hwp_list. Caller can minimize the
> > > overhead of lock cycles by first checking if folio / head page's
> > > HWPOISON flag is set.
> > >
> > > Exports this functionality to be immediately used in the read operation
> > > for hugetlbfs.
> > >
> > > Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>
> > > Reviewed-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
> > > Signed-off-by: Jiaqi Yan <jiaqiyan@google.com>
> > > ---
> > >  include/linux/hugetlb.h | 19 +++++++++++++++++++
> > >  include/linux/mm.h      |  7 +++++++
> > >  mm/hugetlb.c            | 10 ++++++++++
> > >  mm/memory-failure.c     | 34 ++++++++++++++++++++++++----------
> > >  4 files changed, 60 insertions(+), 10 deletions(-)
> > > ...
> > > -static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
> > > +bool __is_raw_hwp_subpage(struct folio *folio, struct page *subpage)
> > >  {
> > > -     return (struct llist_head *)&folio->_hugetlb_hwpoison;
> > > +     struct llist_head *raw_hwp_head;
> > > +     struct raw_hwp_page *p, *tmp;
> > > +     bool ret = false;
> > > +
> > > +     if (!folio_test_hwpoison(folio))
> > > +             return false;
> > > +
> > > +     /*
> > > +      * When RawHwpUnreliable is set, kernel lost track of which subpages
> > > +      * are HWPOISON. So return as if ALL subpages are HWPOISONed.
> > > +      */
> > > +     if (folio_test_hugetlb_raw_hwp_unreliable(folio))
> > > +             return true;
> > > +
> > > +     raw_hwp_head = raw_hwp_list_head(folio);
> > > +     llist_for_each_entry_safe(p, tmp, raw_hwp_head->first, node) {
> >
> > Since we don't free the raw_hwp_list, does llist_for_each_entry works same as llist_for_each_entry_safe?

Sorry I missed this comment. Yes they are the same but
llist_for_each_entry doesn't need "tmp". I will switch to
llist_for_each_entry in v4.

>
> >
> > > +             if (subpage == p->page) {
> > > +                     ret = true;
> > > +                     break;
> > > +             }
> > > +     }
> > > +
> > > +     return ret;
> > >  }
> >
> > It seems there's a race between __is_raw_hwp_subpage and unpoison_memory:
> >   unpoison_memory               __is_raw_hwp_subpage
> >                                   if (!folio_test_hwpoison(folio)) -- hwpoison is set
> >     folio_free_raw_hwp            llist_for_each_entry_safe raw_hwp_list
> >       llist_del_all                 ..
> >     folio_test_clear_hwpoison
> >
>
> Thanks Miaohe for raising this concern.
>
> > But __is_raw_hwp_subpage is used in hugetlbfs, unpoison_memory couldn't reach here because there's a
> > folio_mapping == NULL check before folio_free_raw_hwp.
>
> I agree. But in near future I do want to make __is_raw_hwp_subpage
> work for shared-mapping hugetlb, so it would be nice to work with
> unpoison_memory. It doesn't seem to me that holding mf_mutex in
> __is_raw_hwp_subpage is nice or even absolutely correct. Let me think
> if I can come up with something in v4.

At my 2nd thought, if __is_raw_hwp_subpage simply takes mf_mutex
before llist_for_each_entry, it will introduce a deadlock:

unpoison_memory                       __is_raw_hwp_subpage
  held mf_mutex                         held hugetlb_lock
  get_hwpoison_hugetlb_folio            attempts mf_mutex
    attempts hugetlb lock

Not for this patch series, but for future, is it a good idea to make
mf_mutex available to hugetlb code? Then enforce the order of locking
to be mf_mutex first, hugetlb_lock second? I believe this is the
current locking pattern / order for try_memory_failure_hugetlb.


>
> >
> > Anyway, this patch looks good to me.
> >
> > Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
> > Thanks.
> >