Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
MIME-Version: 1.0
References: <20220228140245.24552-1-linmiaohe@huawei.com> <20220228140245.24552-3-linmiaohe@huawei.com>
 <20220304082714.GB3778609@hori.linux.bs1.fc.nec.co.jp> <227af111-9264-02fd-9e46-744d39ecfed0@huawei.com>
 <CAHbLzkom__59_RCpJCZDA+ray-t5qAWatujXWha8BX2-x8GiMA@mail.gmail.com> <8a35d8d8-3078-89ad-4061-358a500c5d61@huawei.com>
In-Reply-To: <8a35d8d8-3078-89ad-4061-358a500c5d61@huawei.com>
From:   Yang Shi <shy828301@gmail.com>
Date:   Tue, 8 Mar 2022 10:51:42 -0800
Message-ID: <CAHbLzkqjoQgXZDUnuLkVABLbPx-=yfgZ-Akt26rgALyQP07dbA@mail.gmail.com>
Subject: Re: [PATCH 2/4] mm/memory-failure.c: fix wrong user reference report
To:     Miaohe Lin <linmiaohe@huawei.com>
Cc:     =?UTF-8?B?SE9SSUdVQ0hJIE5BT1lBKOWggOWPoyDnm7TkuZ8p?= 
        <naoya.horiguchi@nec.com>,
        "akpm@linux-foundation.org" <akpm@linux-foundation.org>,
        "linux-mm@kvack.org" <linux-mm@kvack.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Tue, Mar 8, 2022 at 5:11 AM Miaohe Lin <linmiaohe@huawei.com> wrote:
>
> On 2022/3/8 4:14, Yang Shi wrote:
> > On Mon, Mar 7, 2022 at 3:26 AM Miaohe Lin <linmiaohe@huawei.com> wrote:
> >>
> >> On 2022/3/4 16:27, HORIGUCHI NAOYA(=E5=A0=80=E5=8F=A3 =E7=9B=B4=E4=B9=
=9F) wrote:
> >>> On Mon, Feb 28, 2022 at 10:02:43PM +0800, Miaohe Lin wrote:
> >>>> The dirty swapcache page is still residing in the swap cache after i=
t's
> >>>> hwpoisoned. So there is always one extra refcount for swap cache.
> >>>
> >>> The diff seems fine at a glance, but let me have a few question to
> >>> understand the issue more.
> >>>
> >>> - Is the behavior described above the effect of recent change on shme=
m where
> >>>   dirty pagecache is pinned on hwpoison (commit a76054266661 ("mm: sh=
mem:
> >>>   don't truncate page if memory failure happens"). Or the older kerne=
ls
> >>>   behave as the same?
> >>>
> >>> - Is the behavior true for normal anonymous pages (not shmem pages)?
> >>>
> >>
> >> The behavior described above is aimed at swapcache not pagecache. So i=
t should be
> >> irrelevant with the recent change on shmem.
> >>
> >> What I try to fix here is that me_swapcache_dirty holds one extra pin =
via SwapCache
> >> regardless of the return value of delete_from_lru_cache. We should try=
 to report more
> >> accurate extra refcount for debugging purpose.
> >
> > I think you misunderstood the code. The delete_from_lru_cache()
> > returning 0 means the page was on LRU and isolated from LRU
> > successfully now. Returning -EIO means the page was not on LRU, so it
> > should have at least an extra pin on it.
> >
> > So MF_DELAYED means there is no other pin other than hwpoison and
> > swapcache which is expected, MF_FAILED means there might be extra
> > pins.
> >
> > The has_extra_refcount() raised error then there is *unexpected* refcou=
nt.
>
> Many thanks for your explanation. It seems you're right. If page is held =
on
> the lru_pvecs when we try to do delete_from_lru_cache, and after that it'=
s
> drained to the lru list( so its refcnt might be 2 now). Then we might hav=
e
> the following complain if extra_pins is always true:
>         "Memory failure: ... still referenced by 0 users\n"
>
> But it seems the origin code can not report the correct reason too becaus=
e
> if we retry, page can be delete_from_lru_cache and we can succeed now.

Retry is ok, but it seems overkilling to me IMHO.

>
> Anyway, many thanks for pointing this out.
>
> >
> >>
> >>> I'm trying to test hwpoison hitting the dirty swapcache, but it seems=
 that
> >>> in my testing memory_faliure() fails with "hwpoison: unhandlable page=
"
> >>
> >> Maybe memory_faliure is racing with page reclaim where page is isolate=
d?
> >>
> >>> warning at get_any_page().  So I'm still not sure that me_pagecache_d=
irty()
> >>> fixes any visible problem.
> >>
> >> IIUC, me_pagecache_dirty can't do much except for the corresponding ad=
dress_space supporting
> >> error_remove_page which can truncate the dirty pagecache page. But thi=
s may cause silent data
> >> loss. It's better to keep the page stay in the pagecache until the fil=
e is truncated, hole
> >> punched or removed as commit a76054266661 pointed out.
> >>
> >> Thanks.
> >>
> >>>> Thanks,
> >>> Naoya Horiguchi
> >>>
> >>>>
> >>>> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
> >>>> ---
> >>>>  mm/memory-failure.c | 6 +-----
> >>>>  1 file changed, 1 insertion(+), 5 deletions(-)
> >>>>
> >>>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> >>>> index 0d7c58340a98..5f9503573263 100644
> >>>> --- a/mm/memory-failure.c
> >>>> +++ b/mm/memory-failure.c
> >>>> @@ -984,7 +984,6 @@ static int me_pagecache_dirty(struct page_state =
*ps, struct page *p)
> >>>>  static int me_swapcache_dirty(struct page_state *ps, struct page *p=
)
> >>>>  {
> >>>>      int ret;
> >>>> -    bool extra_pins =3D false;
> >>>>
> >>>>      ClearPageDirty(p);
> >>>>      /* Trigger EIO in shmem: */
> >>>> @@ -993,10 +992,7 @@ static int me_swapcache_dirty(struct page_state=
 *ps, struct page *p)
> >>>>      ret =3D delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
> >>>>      unlock_page(p);
> >>>>
> >>>> -    if (ret =3D=3D MF_DELAYED)
> >>>> -            extra_pins =3D true;
> >>>> -
> >>>> -    if (has_extra_refcount(ps, p, extra_pins))
> >>>> +    if (has_extra_refcount(ps, p, true))
> >>>>              ret =3D MF_FAILED;
> >>>>
> >>>>      return ret;
> >>>> --
> >>>> 2.23.0
> >>
> >>
> > .
> >
>