[v1] mm, hwpoison: add is_free_buddy_page() in HWPoisonHandlable()

Commit Message

Naoya Horiguchi Sept. 9, 2021, 12:41 a.m. UTC

From: Naoya Horiguchi <[e-mail]>

commit fcc00621d88b ("mm/hwpoison: retry with shake_page() for
unhandlable pages") changes the return value of __get_hwpoison_page() to
retry for transiently unhandlable cases.  However, __get_hwpoison_page()
currently fails to properly judge buddy pages as handlable, so hard/soft
offline for buddy pages always fail as "unhandlable page".  This is
totally regrettable.

So let's add is_free_buddy_page() in HWPoisonHandlable(), so that
__get_hwpoison_page() returns different return values between buddy
pages and unhandlable pages as intended.

Fixes: fcc00621d88b ("mm/hwpoison: retry with shake_page() for unhandlable pages")
Cc: <[e-mail]>
Signed-off-by: Naoya Horiguchi <[e-mail]>
---
 mm/memory-failure.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Yang Shi Sept. 9, 2021, 3:22 a.m. UTC | #1

On Wed, Sep 8, 2021 at 5:41 PM Naoya Horiguchi
<[e-mail]> wrote:
>
> From: Naoya Horiguchi <[e-mail]>
>
> commit fcc00621d88b ("mm/hwpoison: retry with shake_page() for
> unhandlable pages") changes the return value of __get_hwpoison_page() to
> retry for transiently unhandlable cases.  However, __get_hwpoison_page()
> currently fails to properly judge buddy pages as handlable, so hard/soft
> offline for buddy pages always fail as "unhandlable page".  This is
> totally regrettable.
>
> So let's add is_free_buddy_page() in HWPoisonHandlable(), so that
> __get_hwpoison_page() returns different return values between buddy
> pages and unhandlable pages as intended.
>
> Fixes: fcc00621d88b ("mm/hwpoison: retry with shake_page() for unhandlable pages")
> Cc: <[e-mail]>
> Signed-off-by: Naoya Horiguchi <[e-mail]>
> ---
>  mm/memory-failure.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> index 60df8fcd0444..3416c55be810 100644
> --- v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c
> +++ v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> @@ -1126,7 +1126,7 @@ static int page_action(struct page_state *ps, struct page *p,
>   */
>  static inline bool HWPoisonHandlable(struct page *page)
>  {
> -       return PageLRU(page) || __PageMovable(page);
> +       return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);

It seems to work. Although this may change the return value of
get_any_page() to 1 when MF_COUNT_INCREASED is set. This may cause
soft offline to mishandle free buddy page, but MF_COUNT_INCREASED is
only set when madvise is used, and madvise definitely can't soft
offline free buddy page. It did take me a while to figure out this
trick. Maybe need some refactor?

Anyway this patch looks fine to me. Reviewed-by: Yang Shi <[e-mail]>

>  }
>
>  static int __get_hwpoison_page(struct page *page)
> --
> 2.25.1
>

HORIGUCHI NAOYA(堀口 直也) Sept. 9, 2021, 3:58 a.m. UTC | #2

On Wed, Sep 08, 2021 at 08:22:14PM -0700, Yang Shi wrote:
> On Wed, Sep 8, 2021 at 5:41 PM Naoya Horiguchi
> <[e-mail]> wrote:
> >
> > From: Naoya Horiguchi <[e-mail]>
> >
> > commit fcc00621d88b ("mm/hwpoison: retry with shake_page() for
> > unhandlable pages") changes the return value of __get_hwpoison_page() to
> > retry for transiently unhandlable cases.  However, __get_hwpoison_page()
> > currently fails to properly judge buddy pages as handlable, so hard/soft
> > offline for buddy pages always fail as "unhandlable page".  This is
> > totally regrettable.
> >
> > So let's add is_free_buddy_page() in HWPoisonHandlable(), so that
> > __get_hwpoison_page() returns different return values between buddy
> > pages and unhandlable pages as intended.
> >
> > Fixes: fcc00621d88b ("mm/hwpoison: retry with shake_page() for unhandlable pages")
> > Cc: <[e-mail]>
> > Signed-off-by: Naoya Horiguchi <[e-mail]>
> > ---
> >  mm/memory-failure.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> > index 60df8fcd0444..3416c55be810 100644
> > --- v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c
> > +++ v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> > @@ -1126,7 +1126,7 @@ static int page_action(struct page_state *ps, struct page *p,
> >   */
> >  static inline bool HWPoisonHandlable(struct page *page)
> >  {
> > -       return PageLRU(page) || __PageMovable(page);
> > +       return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
> 
> It seems to work. Although this may change the return value of
> get_any_page() to 1 when MF_COUNT_INCREASED is set. This may cause
> soft offline to mishandle free buddy page, but MF_COUNT_INCREASED is
> only set when madvise is used, and madvise definitely can't soft
> offline free buddy page. 

Yes, madvise(MADV_SOFT_OFFLINE) should not be called for free buddy pages,
because the interface covers only user-mapped pages.

> It did take me a while to figure out this
> trick. Maybe need some refactor?

I think so, the current code looks to me fragile for future change, but
I'm not sure how we can improve this.

One (maybe not related to MF_COUNT_INCREASED) room for refactoring is
"if (PageTransHuge(head))" block in __get_hwpoison_page(), which might
be old and useless code for now.

> 
> Anyway this patch looks fine to me. Reviewed-by: Yang Shi <[e-mail]>

Thank you for the review.

- Naoya

David Hildenbrand Sept. 10, 2021, 11:16 a.m. UTC | #3

On 09.09.21 02:41, Naoya Horiguchi wrote:
> From: Naoya Horiguchi <[e-mail]>
> 
> commit fcc00621d88b ("mm/hwpoison: retry with shake_page() for
> unhandlable pages") changes the return value of __get_hwpoison_page() to
> retry for transiently unhandlable cases.  However, __get_hwpoison_page()
> currently fails to properly judge buddy pages as handlable, so hard/soft
> offline for buddy pages always fail as "unhandlable page".  This is
> totally regrettable.
> 
> So let's add is_free_buddy_page() in HWPoisonHandlable(), so that
> __get_hwpoison_page() returns different return values between buddy
> pages and unhandlable pages as intended.
> 
> Fixes: fcc00621d88b ("mm/hwpoison: retry with shake_page() for unhandlable pages")
> Cc: <[e-mail]>
> Signed-off-by: Naoya Horiguchi <[e-mail]>
> ---
>   mm/memory-failure.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> index 60df8fcd0444..3416c55be810 100644
> --- v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c
> +++ v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
> @@ -1126,7 +1126,7 @@ static int page_action(struct page_state *ps, struct page *p,
>    */
>   static inline bool HWPoisonHandlable(struct page *page)
>   {
> -	return PageLRU(page) || __PageMovable(page);
> +	return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
>   }
>   
>   static int __get_hwpoison_page(struct page *page)
> 

LGTM and I agree that code might desrve some cleanups.

Acked-by: David Hildenbrand <[e-mail]>

Patch

diff --git v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
index 60df8fcd0444..3416c55be810 100644
--- v5.14-rc7-mmotm-2021-08-23-16-42/mm/memory-failure.c
+++ v5.14-rc7-mmotm-2021-08-23-16-42_patched/mm/memory-failure.c
@@ -1126,7 +1126,7 @@  static int page_action(struct page_state *ps, struct page *p,
  */
 static inline bool HWPoisonHandlable(struct page *page)
 {
-	return PageLRU(page) || __PageMovable(page);
+	return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
 }
 
 static int __get_hwpoison_page(struct page *page)