If large folio is in the range of VM_LOCKED VMA, it should be
mlocked to avoid being picked by page reclaim. Which may split
the large folio and then mlock each pages again.
Mlock this kind of large folio to prevent them being picked by
page reclaim.
For the large folio which cross the boundary of VM_LOCKED VMA,
we'd better not to mlock it. So if the system is under memory
pressure, this kind of large folio will be split and the pages
ouf of VM_LOCKED VMA can be reclaimed.
Signed-off-by: Yin Fengwei <[email protected]>
---
mm/internal.h | 11 ++++++++---
mm/rmap.c | 3 ++-
2 files changed, 10 insertions(+), 4 deletions(-)
diff --git a/mm/internal.h b/mm/internal.h
index 66117523d7d71..c7b8f0b008d81 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -637,7 +637,8 @@ static inline void mlock_vma_folio(struct folio *folio,
* still be set while VM_SPECIAL bits are added: so ignore it then.
*/
if (unlikely((vma->vm_flags & (VM_LOCKED|VM_SPECIAL)) == VM_LOCKED) &&
- (compound || !folio_test_large(folio)))
+ (compound || !folio_test_large(folio) ||
+ folio_in_range(folio, vma, vma->vm_start, vma->vm_end)))
mlock_folio(folio);
}
@@ -645,8 +646,12 @@ void munlock_folio(struct folio *folio);
static inline void munlock_vma_folio(struct folio *folio,
struct vm_area_struct *vma, bool compound)
{
- if (unlikely(vma->vm_flags & VM_LOCKED) &&
- (compound || !folio_test_large(folio)))
+ /*
+ * To handle the case that a mlocked large folio is unmapped from VMA
+ * piece by piece, allow munlock the large folio which is partially
+ * mapped to VMA.
+ */
+ if (unlikely(vma->vm_flags & VM_LOCKED))
munlock_folio(folio);
}
diff --git a/mm/rmap.c b/mm/rmap.c
index 2668f5ea35342..7d6547d1bd096 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -817,7 +817,8 @@ static bool folio_referenced_one(struct folio *folio,
address = pvmw.address;
if ((vma->vm_flags & VM_LOCKED) &&
- (!folio_test_large(folio) || !pvmw.pte)) {
+ (!folio_test_large(folio) || !pvmw.pte ||
+ folio_in_range(folio, vma, vma->vm_start, vma->vm_end))) {
/* Restore the mlock which got missed */
mlock_vma_folio(folio, vma, !pvmw.pte);
page_vma_mapped_walk_done(&pvmw);
--
2.39.2
On Fri, Jul 7, 2023 at 10:52 AM Yin Fengwei <[email protected]> wrote:
>
> If large folio is in the range of VM_LOCKED VMA, it should be
> mlocked to avoid being picked by page reclaim. Which may split
> the large folio and then mlock each pages again.
>
> Mlock this kind of large folio to prevent them being picked by
> page reclaim.
>
> For the large folio which cross the boundary of VM_LOCKED VMA,
> we'd better not to mlock it. So if the system is under memory
> pressure, this kind of large folio will be split and the pages
> ouf of VM_LOCKED VMA can be reclaimed.
>
> Signed-off-by: Yin Fengwei <[email protected]>
> ---
> mm/internal.h | 11 ++++++++---
> mm/rmap.c | 3 ++-
> 2 files changed, 10 insertions(+), 4 deletions(-)
>
> diff --git a/mm/internal.h b/mm/internal.h
> index 66117523d7d71..c7b8f0b008d81 100644
> --- a/mm/internal.h
> +++ b/mm/internal.h
> @@ -637,7 +637,8 @@ static inline void mlock_vma_folio(struct folio *folio,
> * still be set while VM_SPECIAL bits are added: so ignore it then.
> */
> if (unlikely((vma->vm_flags & (VM_LOCKED|VM_SPECIAL)) == VM_LOCKED) &&
> - (compound || !folio_test_large(folio)))
> + (compound || !folio_test_large(folio) ||
> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end)))
> mlock_folio(folio);
> }
>
> @@ -645,8 +646,12 @@ void munlock_folio(struct folio *folio);
> static inline void munlock_vma_folio(struct folio *folio,
> struct vm_area_struct *vma, bool compound)
> {
> - if (unlikely(vma->vm_flags & VM_LOCKED) &&
> - (compound || !folio_test_large(folio)))
> + /*
> + * To handle the case that a mlocked large folio is unmapped from VMA
> + * piece by piece, allow munlock the large folio which is partially
> + * mapped to VMA.
> + */
> + if (unlikely(vma->vm_flags & VM_LOCKED))
> munlock_folio(folio);
> }
>
> diff --git a/mm/rmap.c b/mm/rmap.c
> index 2668f5ea35342..7d6547d1bd096 100644
> --- a/mm/rmap.c
> +++ b/mm/rmap.c
> @@ -817,7 +817,8 @@ static bool folio_referenced_one(struct folio *folio,
> address = pvmw.address;
>
> if ((vma->vm_flags & VM_LOCKED) &&
> - (!folio_test_large(folio) || !pvmw.pte)) {
> + (!folio_test_large(folio) || !pvmw.pte ||
> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end))) {
> /* Restore the mlock which got missed */
> mlock_vma_folio(folio, vma, !pvmw.pte);
> page_vma_mapped_walk_done(&pvmw);
It needs to bail out if large but not within range so that the
references within the locked VMA can be ignored. Otherwise, a hot
locked portion can prevent a cold unlocked portion from getting
reclaimed.
On 7/8/2023 1:11 PM, Yu Zhao wrote:
> On Fri, Jul 7, 2023 at 10:52 AM Yin Fengwei <[email protected]> wrote:
>>
>> If large folio is in the range of VM_LOCKED VMA, it should be
>> mlocked to avoid being picked by page reclaim. Which may split
>> the large folio and then mlock each pages again.
>>
>> Mlock this kind of large folio to prevent them being picked by
>> page reclaim.
>>
>> For the large folio which cross the boundary of VM_LOCKED VMA,
>> we'd better not to mlock it. So if the system is under memory
>> pressure, this kind of large folio will be split and the pages
>> ouf of VM_LOCKED VMA can be reclaimed.
>>
>> Signed-off-by: Yin Fengwei <[email protected]>
>> ---
>> mm/internal.h | 11 ++++++++---
>> mm/rmap.c | 3 ++-
>> 2 files changed, 10 insertions(+), 4 deletions(-)
>>
>> diff --git a/mm/internal.h b/mm/internal.h
>> index 66117523d7d71..c7b8f0b008d81 100644
>> --- a/mm/internal.h
>> +++ b/mm/internal.h
>> @@ -637,7 +637,8 @@ static inline void mlock_vma_folio(struct folio *folio,
>> * still be set while VM_SPECIAL bits are added: so ignore it then.
>> */
>> if (unlikely((vma->vm_flags & (VM_LOCKED|VM_SPECIAL)) == VM_LOCKED) &&
>> - (compound || !folio_test_large(folio)))
>> + (compound || !folio_test_large(folio) ||
>> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end)))
>> mlock_folio(folio);
>> }
>>
>> @@ -645,8 +646,12 @@ void munlock_folio(struct folio *folio);
>> static inline void munlock_vma_folio(struct folio *folio,
>> struct vm_area_struct *vma, bool compound)
>> {
>> - if (unlikely(vma->vm_flags & VM_LOCKED) &&
>> - (compound || !folio_test_large(folio)))
>> + /*
>> + * To handle the case that a mlocked large folio is unmapped from VMA
>> + * piece by piece, allow munlock the large folio which is partially
>> + * mapped to VMA.
>> + */
>> + if (unlikely(vma->vm_flags & VM_LOCKED))
>> munlock_folio(folio);
>> }
>>
>> diff --git a/mm/rmap.c b/mm/rmap.c
>> index 2668f5ea35342..7d6547d1bd096 100644
>> --- a/mm/rmap.c
>> +++ b/mm/rmap.c
>> @@ -817,7 +817,8 @@ static bool folio_referenced_one(struct folio *folio,
>> address = pvmw.address;
>>
>> if ((vma->vm_flags & VM_LOCKED) &&
>> - (!folio_test_large(folio) || !pvmw.pte)) {
>> + (!folio_test_large(folio) || !pvmw.pte ||
>> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end))) {
>> /* Restore the mlock which got missed */
>> mlock_vma_folio(folio, vma, !pvmw.pte);
>> page_vma_mapped_walk_done(&pvmw);
>
> It needs to bail out if large but not within range so that the
> references within the locked VMA can be ignored. Otherwise, a hot
> locked portion can prevent a cold unlocked portion from getting
> reclaimed.
Good point. We can't bail out here as return here means folio should
not be reclaimed. My understanding is that we should skip the entries
which is in the range of VM_LOCKED VMA. Will address this in coming
version. Thanks.
Regards
Yin, Fengwei
On Fri, Jul 7, 2023 at 11:34 PM Yin, Fengwei <[email protected]> wrote:
>
>
>
> On 7/8/2023 1:11 PM, Yu Zhao wrote:
> > On Fri, Jul 7, 2023 at 10:52 AM Yin Fengwei <[email protected]> wrote:
> >>
> >> If large folio is in the range of VM_LOCKED VMA, it should be
> >> mlocked to avoid being picked by page reclaim. Which may split
> >> the large folio and then mlock each pages again.
> >>
> >> Mlock this kind of large folio to prevent them being picked by
> >> page reclaim.
> >>
> >> For the large folio which cross the boundary of VM_LOCKED VMA,
> >> we'd better not to mlock it. So if the system is under memory
> >> pressure, this kind of large folio will be split and the pages
> >> ouf of VM_LOCKED VMA can be reclaimed.
> >>
> >> Signed-off-by: Yin Fengwei <[email protected]>
> >> ---
> >> mm/internal.h | 11 ++++++++---
> >> mm/rmap.c | 3 ++-
> >> 2 files changed, 10 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/mm/internal.h b/mm/internal.h
> >> index 66117523d7d71..c7b8f0b008d81 100644
> >> --- a/mm/internal.h
> >> +++ b/mm/internal.h
> >> @@ -637,7 +637,8 @@ static inline void mlock_vma_folio(struct folio *folio,
> >> * still be set while VM_SPECIAL bits are added: so ignore it then.
> >> */
> >> if (unlikely((vma->vm_flags & (VM_LOCKED|VM_SPECIAL)) == VM_LOCKED) &&
> >> - (compound || !folio_test_large(folio)))
> >> + (compound || !folio_test_large(folio) ||
> >> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end)))
> >> mlock_folio(folio);
> >> }
> >>
> >> @@ -645,8 +646,12 @@ void munlock_folio(struct folio *folio);
> >> static inline void munlock_vma_folio(struct folio *folio,
> >> struct vm_area_struct *vma, bool compound)
> >> {
> >> - if (unlikely(vma->vm_flags & VM_LOCKED) &&
> >> - (compound || !folio_test_large(folio)))
> >> + /*
> >> + * To handle the case that a mlocked large folio is unmapped from VMA
> >> + * piece by piece, allow munlock the large folio which is partially
> >> + * mapped to VMA.
> >> + */
> >> + if (unlikely(vma->vm_flags & VM_LOCKED))
> >> munlock_folio(folio);
> >> }
> >>
> >> diff --git a/mm/rmap.c b/mm/rmap.c
> >> index 2668f5ea35342..7d6547d1bd096 100644
> >> --- a/mm/rmap.c
> >> +++ b/mm/rmap.c
> >> @@ -817,7 +817,8 @@ static bool folio_referenced_one(struct folio *folio,
> >> address = pvmw.address;
> >>
> >> if ((vma->vm_flags & VM_LOCKED) &&
> >> - (!folio_test_large(folio) || !pvmw.pte)) {
> >> + (!folio_test_large(folio) || !pvmw.pte ||
> >> + folio_in_range(folio, vma, vma->vm_start, vma->vm_end))) {
> >> /* Restore the mlock which got missed */
> >> mlock_vma_folio(folio, vma, !pvmw.pte);
> >> page_vma_mapped_walk_done(&pvmw);
> >
> > It needs to bail out if large but not within range so that the
> > references within the locked VMA can be ignored. Otherwise, a hot
> > locked portion can prevent a cold unlocked portion from getting
> > reclaimed.
> Good point. We can't bail out here as return here means folio should
> not be reclaimed. My understanding is that we should skip the entries
> which is in the range of VM_LOCKED VMA. Will address this in coming
> version. Thanks.
Yes, that's what I mean. A wrapper would be cleaner:
while () {
...
if (vma->vm_flags & VM_LOCKED) {
if (cant_mlock())
goto next;
...
return false;
}
...
next:
pra->mapcount--;
}