mm/memory_failure: reject unsupported non-folio compound page

[PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Zi Yan 2 days, 2 hours ago

When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
a userspace mapping via either vm_insert_*() functions or
vm_operatios_struct->fault(). They are not folios, thus should not be
considered for folio operations like split. Change memory_failure() and
soft_offline_page() to reject these non-folio compound pages as
EOPNOTSUPP.

Add PageNonFolioCompound() helper function. This function is functionally
equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
is supposed to be used on struct page. So open code it instead.

Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
Reported-by: 是参差 <shicenci@gmail.com>
Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
Signed-off-by: Zi Yan <ziy@nvidia.com>
---
 include/linux/page-flags.h | 16 ++++++++++++++++
 mm/memory-failure.c        |  9 ++++++---
 2 files changed, 22 insertions(+), 3 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index f7a0e4af0c73..2fe8047f42a3 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -1102,6 +1102,22 @@ static inline bool folio_contain_hwpoisoned_page(struct folio *folio)
 
 bool is_free_buddy_page(const struct page *page);
 
+static inline bool PageNonFolioCompound(const struct page *page)
+{
+	if (PageCompound(page)) {
+		const struct page *head = compound_head(page);
+
+		/*
+		 * Without CONFIG_TRANSPARENT_HUGEPAGE, PG_large_rmappable
+		 * should not be set/used.
+		 */
+		return !IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ||
+		       !test_bit(PG_large_rmappable, &head[1].flags.f);
+	}
+
+	return false;
+}
+
 #ifdef CONFIG_MIGRATION
 /*
  * This page is migratable through movable_ops (for selected typed pages
diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index cf0d526e6d41..8b6b5950bb66 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -2440,9 +2440,12 @@ int memory_failure(unsigned long pfn, int flags)
 
 	folio = page_folio(p);
 
-	/* filter pages that are protected from hwpoison test by users */
+	/*
+	 * filter pages that are protected from hwpoison test by users or
+	 * unsupported non folio compound ones
+	 */
 	folio_lock(folio);
-	if (hwpoison_filter(p)) {
+	if (hwpoison_filter(p) || PageNonFolioCompound(p)) {
 		ClearPageHWPoison(p);
 		folio_unlock(folio);
 		folio_put(folio);
@@ -2945,7 +2948,7 @@ int soft_offline_page(unsigned long pfn, int flags)
 	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
 	put_online_mems();
 
-	if (hwpoison_filter(page)) {
+	if (hwpoison_filter(page) || PageNonFolioCompound(page)) {
 		if (ret > 0)
 			put_page(page);
 
-- 
2.51.0

Re: [PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Zi Yan 1 day, 23 hours ago

Hi Andrew,

Miaohe has a more appropriate solution[1]. I think it is better to drop this patch
and wait for Miaohe's proper patch fix.

Thanks.


[1] https://lore.kernel.org/all/c5206f67-1e9e-d487-3514-8867f00cc626@huawei.com/


On 4 Feb 2026, at 19:56, Zi Yan wrote:

> When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
> a userspace mapping via either vm_insert_*() functions or
> vm_operatios_struct->fault(). They are not folios, thus should not be
> considered for folio operations like split. Change memory_failure() and
> soft_offline_page() to reject these non-folio compound pages as
> EOPNOTSUPP.
>
> Add PageNonFolioCompound() helper function. This function is functionally
> equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
> is supposed to be used on struct page. So open code it instead.
>
> Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
> Reported-by: 是参差 <shicenci@gmail.com>
> Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---
>  include/linux/page-flags.h | 16 ++++++++++++++++
>  mm/memory-failure.c        |  9 ++++++---
>  2 files changed, 22 insertions(+), 3 deletions(-)
>
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index f7a0e4af0c73..2fe8047f42a3 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -1102,6 +1102,22 @@ static inline bool folio_contain_hwpoisoned_page(struct folio *folio)
>
>  bool is_free_buddy_page(const struct page *page);
>
> +static inline bool PageNonFolioCompound(const struct page *page)
> +{
> +	if (PageCompound(page)) {
> +		const struct page *head = compound_head(page);
> +
> +		/*
> +		 * Without CONFIG_TRANSPARENT_HUGEPAGE, PG_large_rmappable
> +		 * should not be set/used.
> +		 */
> +		return !IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ||
> +		       !test_bit(PG_large_rmappable, &head[1].flags.f);
> +	}
> +
> +	return false;
> +}
> +
>  #ifdef CONFIG_MIGRATION
>  /*
>   * This page is migratable through movable_ops (for selected typed pages
> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> index cf0d526e6d41..8b6b5950bb66 100644
> --- a/mm/memory-failure.c
> +++ b/mm/memory-failure.c
> @@ -2440,9 +2440,12 @@ int memory_failure(unsigned long pfn, int flags)
>
>  	folio = page_folio(p);
>
> -	/* filter pages that are protected from hwpoison test by users */
> +	/*
> +	 * filter pages that are protected from hwpoison test by users or
> +	 * unsupported non folio compound ones
> +	 */
>  	folio_lock(folio);
> -	if (hwpoison_filter(p)) {
> +	if (hwpoison_filter(p) || PageNonFolioCompound(p)) {
>  		ClearPageHWPoison(p);
>  		folio_unlock(folio);
>  		folio_put(folio);
> @@ -2945,7 +2948,7 @@ int soft_offline_page(unsigned long pfn, int flags)
>  	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
>  	put_online_mems();
>
> -	if (hwpoison_filter(page)) {
> +	if (hwpoison_filter(page) || PageNonFolioCompound(page)) {
>  		if (ret > 0)
>  			put_page(page);
>
> -- 
> 2.51.0


--
Best Regards,
Yan, Zi

Re: [PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Harry Yoo 2 days ago

On Wed, Feb 04, 2026 at 07:56:03PM -0500, Zi Yan wrote:
> When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
> a userspace mapping via either vm_insert_*() functions or
> vm_operatios_struct->fault(). They are not folios, thus should not be
> considered for folio operations like split. Change memory_failure() and
> soft_offline_page() to reject these non-folio compound pages as
> EOPNOTSUPP.
> 
> Add PageNonFolioCompound() helper function. This function is functionally
> equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
> is supposed to be used on struct page. So open code it instead.
> 
> Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
> Reported-by: 是参差 <shicenci@gmail.com>
> Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---

Looks good to me,
Acked-by: Harry Yoo <harry.yoo@oracle.com>

-- 
Cheers,
Harry / Hyeonggon

Re: [PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Miaohe Lin 2 days ago

On 2026/2/5 8:56, Zi Yan wrote:
> When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
> a userspace mapping via either vm_insert_*() functions or
> vm_operatios_struct->fault(). They are not folios, thus should not be
> considered for folio operations like split. Change memory_failure() and
> soft_offline_page() to reject these non-folio compound pages as
> EOPNOTSUPP.
> 
> Add PageNonFolioCompound() helper function. This function is functionally
> equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
> is supposed to be used on struct page. So open code it instead.
> 
> Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
> Reported-by: 是参差 <shicenci@gmail.com>
> Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---
>  include/linux/page-flags.h | 16 ++++++++++++++++
>  mm/memory-failure.c        |  9 ++++++---
>  2 files changed, 22 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> index f7a0e4af0c73..2fe8047f42a3 100644
> --- a/include/linux/page-flags.h
> +++ b/include/linux/page-flags.h
> @@ -1102,6 +1102,22 @@ static inline bool folio_contain_hwpoisoned_page(struct folio *folio)
>  
>  bool is_free_buddy_page(const struct page *page);
>  
> +static inline bool PageNonFolioCompound(const struct page *page)
> +{
> +	if (PageCompound(page)) {
> +		const struct page *head = compound_head(page);
> +
> +		/*
> +		 * Without CONFIG_TRANSPARENT_HUGEPAGE, PG_large_rmappable
> +		 * should not be set/used.
> +		 */
> +		return !IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ||
> +		       !test_bit(PG_large_rmappable, &head[1].flags.f);
> +	}
> +
> +	return false;
> +}
> +
>  #ifdef CONFIG_MIGRATION
>  /*
>   * This page is migratable through movable_ops (for selected typed pages
> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> index cf0d526e6d41..8b6b5950bb66 100644
> --- a/mm/memory-failure.c
> +++ b/mm/memory-failure.c
> @@ -2440,9 +2440,12 @@ int memory_failure(unsigned long pfn, int flags)
>  
>  	folio = page_folio(p);
>  
> -	/* filter pages that are protected from hwpoison test by users */
> +	/*
> +	 * filter pages that are protected from hwpoison test by users or
> +	 * unsupported non folio compound ones
> +	 */
>  	folio_lock(folio);
> -	if (hwpoison_filter(p)) {
> +	if (hwpoison_filter(p) || PageNonFolioCompound(p)) {
>  		ClearPageHWPoison(p);
>  		folio_unlock(folio);
>  		folio_put(folio);
> @@ -2945,7 +2948,7 @@ int soft_offline_page(unsigned long pfn, int flags)
>  	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
>  	put_online_mems();
>  
> -	if (hwpoison_filter(page)) {
> +	if (hwpoison_filter(page) || PageNonFolioCompound(page)) {

There should be no problem in soft_offline_page(). HWPoisonHandlable() check will be used
by get_hwpoison_page() to reject PageNonFolioCompound folios. Or am I miss something?

Thanks.
.

Re: [PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Zi Yan 2 days ago

On 4 Feb 2026, at 22:25, Miaohe Lin wrote:

> On 2026/2/5 8:56, Zi Yan wrote:
>> When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
>> a userspace mapping via either vm_insert_*() functions or
>> vm_operatios_struct->fault(). They are not folios, thus should not be
>> considered for folio operations like split. Change memory_failure() and
>> soft_offline_page() to reject these non-folio compound pages as
>> EOPNOTSUPP.
>>
>> Add PageNonFolioCompound() helper function. This function is functionally
>> equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
>> is supposed to be used on struct page. So open code it instead.
>>
>> Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
>> Reported-by: 是参差 <shicenci@gmail.com>
>> Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>> ---
>>  include/linux/page-flags.h | 16 ++++++++++++++++
>>  mm/memory-failure.c        |  9 ++++++---
>>  2 files changed, 22 insertions(+), 3 deletions(-)
>>
>> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
>> index f7a0e4af0c73..2fe8047f42a3 100644
>> --- a/include/linux/page-flags.h
>> +++ b/include/linux/page-flags.h
>> @@ -1102,6 +1102,22 @@ static inline bool folio_contain_hwpoisoned_page(struct folio *folio)
>>
>>  bool is_free_buddy_page(const struct page *page);
>>
>> +static inline bool PageNonFolioCompound(const struct page *page)
>> +{
>> +	if (PageCompound(page)) {
>> +		const struct page *head = compound_head(page);
>> +
>> +		/*
>> +		 * Without CONFIG_TRANSPARENT_HUGEPAGE, PG_large_rmappable
>> +		 * should not be set/used.
>> +		 */
>> +		return !IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ||
>> +		       !test_bit(PG_large_rmappable, &head[1].flags.f);
>> +	}
>> +
>> +	return false;
>> +}
>> +
>>  #ifdef CONFIG_MIGRATION
>>  /*
>>   * This page is migratable through movable_ops (for selected typed pages
>> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
>> index cf0d526e6d41..8b6b5950bb66 100644
>> --- a/mm/memory-failure.c
>> +++ b/mm/memory-failure.c
>> @@ -2440,9 +2440,12 @@ int memory_failure(unsigned long pfn, int flags)
>>
>>  	folio = page_folio(p);
>>
>> -	/* filter pages that are protected from hwpoison test by users */
>> +	/*
>> +	 * filter pages that are protected from hwpoison test by users or
>> +	 * unsupported non folio compound ones
>> +	 */
>>  	folio_lock(folio);
>> -	if (hwpoison_filter(p)) {
>> +	if (hwpoison_filter(p) || PageNonFolioCompound(p)) {
>>  		ClearPageHWPoison(p);
>>  		folio_unlock(folio);
>>  		folio_put(folio);
>> @@ -2945,7 +2948,7 @@ int soft_offline_page(unsigned long pfn, int flags)
>>  	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
>>  	put_online_mems();
>>
>> -	if (hwpoison_filter(page)) {
>> +	if (hwpoison_filter(page) || PageNonFolioCompound(page)) {
>
> There should be no problem in soft_offline_page(). HWPoisonHandlable() check will be used
> by get_hwpoison_page() to reject PageNonFolioCompound folios. Or am I miss something?

I did not know that. Why does memory_failure() not call HWPosonHandlable() to check the input
page? It looks to me that HWPosonHandlable() is more appropriate than PageNonFolioCompound()
here.

--
Best Regards,
Yan, Zi

Re: [PATCH] mm/memory_failure: reject unsupported non-folio compound page

Posted by Harry Yoo 1 day, 23 hours ago

On Wed, Feb 04, 2026 at 10:40:22PM -0500, Zi Yan wrote:
> On 4 Feb 2026, at 22:25, Miaohe Lin wrote:
> 
> > On 2026/2/5 8:56, Zi Yan wrote:
> >> When !CONFIG_TRANSPARENT_HUGEPAGE, a non-folio compound page can appear in
> >> a userspace mapping via either vm_insert_*() functions or
> >> vm_operatios_struct->fault(). They are not folios, thus should not be
> >> considered for folio operations like split. Change memory_failure() and
> >> soft_offline_page() to reject these non-folio compound pages as
> >> EOPNOTSUPP.
> >>
> >> Add PageNonFolioCompound() helper function. This function is functionally
> >> equivalent to folio_test_large() && !folio_test_large_rmappable(), but it
> >> is supposed to be used on struct page. So open code it instead.
> >>
> >> Fixes: 689b8986776c ("mm/memory-failure: improve large block size folio handling")
> >> Reported-by: 是参差 <shicenci@gmail.com>
> >> Closes: https://lore.kernel.org/all/PS1PPF7E1D7501F1E4F4441E7ECD056DEADAB98A@PS1PPF7E1D7501F.apcprd02.prod.outlook.com/
> >> Signed-off-by: Zi Yan <ziy@nvidia.com>
> >> ---
> >>  include/linux/page-flags.h | 16 ++++++++++++++++
> >>  mm/memory-failure.c        |  9 ++++++---
> >>  2 files changed, 22 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
> >> index f7a0e4af0c73..2fe8047f42a3 100644
> >> --- a/include/linux/page-flags.h
> >> +++ b/include/linux/page-flags.h
> >> @@ -1102,6 +1102,22 @@ static inline bool folio_contain_hwpoisoned_page(struct folio *folio)
> >>
> >>  bool is_free_buddy_page(const struct page *page);
> >>
> >> +static inline bool PageNonFolioCompound(const struct page *page)
> >> +{
> >> +	if (PageCompound(page)) {
> >> +		const struct page *head = compound_head(page);
> >> +
> >> +		/*
> >> +		 * Without CONFIG_TRANSPARENT_HUGEPAGE, PG_large_rmappable
> >> +		 * should not be set/used.
> >> +		 */
> >> +		return !IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ||
> >> +		       !test_bit(PG_large_rmappable, &head[1].flags.f);
> >> +	}
> >> +
> >> +	return false;
> >> +}
> >> +
> >>  #ifdef CONFIG_MIGRATION
> >>  /*
> >>   * This page is migratable through movable_ops (for selected typed pages
> >> diff --git a/mm/memory-failure.c b/mm/memory-failure.c
> >> index cf0d526e6d41..8b6b5950bb66 100644
> >> --- a/mm/memory-failure.c
> >> +++ b/mm/memory-failure.c
> >> @@ -2440,9 +2440,12 @@ int memory_failure(unsigned long pfn, int flags)
> >>
> >>  	folio = page_folio(p);
> >>
> >> -	/* filter pages that are protected from hwpoison test by users */
> >> +	/*
> >> +	 * filter pages that are protected from hwpoison test by users or
> >> +	 * unsupported non folio compound ones
> >> +	 */
> >>  	folio_lock(folio);
> >> -	if (hwpoison_filter(p)) {
> >> +	if (hwpoison_filter(p) || PageNonFolioCompound(p)) {
> >>  		ClearPageHWPoison(p);
> >>  		folio_unlock(folio);
> >>  		folio_put(folio);
> >> @@ -2945,7 +2948,7 @@ int soft_offline_page(unsigned long pfn, int flags)
> >>  	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
> >>  	put_online_mems();
> >>
> >> -	if (hwpoison_filter(page)) {
> >> +	if (hwpoison_filter(page) || PageNonFolioCompound(page)) {
> >
> > There should be no problem in soft_offline_page(). HWPoisonHandlable() check will be used
> > by get_hwpoison_page() to reject PageNonFolioCompound folios. Or am I miss something?

Oops, I missed Miohe's email.
So it's rejected because it's not on LRU.

> I did not know that. Why does memory_failure() not call HWPosonHandlable() to check the input
> page? It looks to me that HWPosonHandlable() is more appropriate than PageNonFolioCompound()
> here.

soft_offline_page() unconditionally calls get_hwpoison_page(), but
memory_failure() doesn't call it if MF_COUNT_INCREASED is set.

(MF_COUNT_INCREASED set by madvise_inject_error(), in this case)

-- 
Cheers,
Harry / Hyeonggon