2023-05-18 11:24:08

by Ryan Roberts

[permalink] [raw]
Subject: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

It is racy to non-atomically read a pte, then clear the young bit, then
write it back as this could discard dirty information. Further, it is
bad practice to directly set a pte entry within a table. Instead
clearing young must go through the arch-provided helper,
ptep_test_and_clear_young() to ensure it is modified atomically and to
give the arch code visibility and allow it to check (and potentially
modify) the operation.

Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
Signed-off-by: Ryan Roberts <[email protected]>
Reviewed-by: Zi Yan <[email protected]>
---
mm/damon/ops-common.c | 16 ++++++----------
mm/damon/ops-common.h | 4 ++--
mm/damon/paddr.c | 4 ++--
mm/damon/vaddr.c | 4 ++--
4 files changed, 12 insertions(+), 16 deletions(-)

diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
index cc63cf953636..acc264b97903 100644
--- a/mm/damon/ops-common.c
+++ b/mm/damon/ops-common.c
@@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
return folio;
}

-void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
+void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
{
bool referenced = false;
struct folio *folio = damon_get_folio(pte_pfn(*pte));
@@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
if (!folio)
return;

- if (pte_young(*pte)) {
+ if (ptep_test_and_clear_young(vma, addr, pte))
referenced = true;
- *pte = pte_mkold(*pte);
- }

#ifdef CONFIG_MMU_NOTIFIER
- if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
+ if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
referenced = true;
#endif /* CONFIG_MMU_NOTIFIER */

@@ -62,7 +60,7 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
folio_put(folio);
}

-void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
+void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr)
{
#ifdef CONFIG_TRANSPARENT_HUGEPAGE
bool referenced = false;
@@ -71,13 +69,11 @@ void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
if (!folio)
return;

- if (pmd_young(*pmd)) {
+ if (pmdp_test_and_clear_young(vma, addr, pmd))
referenced = true;
- *pmd = pmd_mkold(*pmd);
- }

#ifdef CONFIG_MMU_NOTIFIER
- if (mmu_notifier_clear_young(mm, addr, addr + HPAGE_PMD_SIZE))
+ if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + HPAGE_PMD_SIZE))
referenced = true;
#endif /* CONFIG_MMU_NOTIFIER */

diff --git a/mm/damon/ops-common.h b/mm/damon/ops-common.h
index 14f4bc69f29b..18d837d11bce 100644
--- a/mm/damon/ops-common.h
+++ b/mm/damon/ops-common.h
@@ -9,8 +9,8 @@

struct folio *damon_get_folio(unsigned long pfn);

-void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr);
-void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr);
+void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr);
+void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr);

int damon_cold_score(struct damon_ctx *c, struct damon_region *r,
struct damos *s);
diff --git a/mm/damon/paddr.c b/mm/damon/paddr.c
index 467b99166b43..5b3a3463d078 100644
--- a/mm/damon/paddr.c
+++ b/mm/damon/paddr.c
@@ -24,9 +24,9 @@ static bool __damon_pa_mkold(struct folio *folio, struct vm_area_struct *vma,
while (page_vma_mapped_walk(&pvmw)) {
addr = pvmw.address;
if (pvmw.pte)
- damon_ptep_mkold(pvmw.pte, vma->vm_mm, addr);
+ damon_ptep_mkold(pvmw.pte, vma, addr);
else
- damon_pmdp_mkold(pvmw.pmd, vma->vm_mm, addr);
+ damon_pmdp_mkold(pvmw.pmd, vma, addr);
}
return true;
}
diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
index 1fec16d7263e..37994fb6120c 100644
--- a/mm/damon/vaddr.c
+++ b/mm/damon/vaddr.c
@@ -311,7 +311,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
}

if (pmd_trans_huge(*pmd)) {
- damon_pmdp_mkold(pmd, walk->mm, addr);
+ damon_pmdp_mkold(pmd, walk->vma, addr);
spin_unlock(ptl);
return 0;
}
@@ -323,7 +323,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
if (!pte_present(*pte))
goto out;
- damon_ptep_mkold(pte, walk->mm, addr);
+ damon_ptep_mkold(pte, walk->vma, addr);
out:
pte_unmap_unlock(pte, ptl);
return 0;
--
2.25.1



2023-05-18 17:31:25

by SeongJae Park

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

Hi Ryan,


Nit. For consistency of DAMON commits subjects, I'd prefer something like
"mm/damon/ops-common: atomically test and clear young on ptes and pmds".

On Thu, 18 May 2023 12:07:24 +0100 Ryan Roberts <[email protected]> wrote:

> It is racy to non-atomically read a pte, then clear the young bit, then
> write it back as this could discard dirty information. Further, it is
> bad practice to directly set a pte entry within a table. Instead
> clearing young must go through the arch-provided helper,
> ptep_test_and_clear_young() to ensure it is modified atomically and to
> give the arch code visibility and allow it to check (and potentially
> modify) the operation.
>
> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")

Nit. The commit is only for refactoring. The original code camde from commit
3f49584b262c ("mm/damon: implement primitives for the virtual memory address
spaces").

> Signed-off-by: Ryan Roberts <[email protected]>
> Reviewed-by: Zi Yan <[email protected]>

Other than above nits,

Reviewed-by: SeongJae Park <[email protected]>


Thanks,
SJ

> ---
> mm/damon/ops-common.c | 16 ++++++----------
> mm/damon/ops-common.h | 4 ++--
> mm/damon/paddr.c | 4 ++--
> mm/damon/vaddr.c | 4 ++--
> 4 files changed, 12 insertions(+), 16 deletions(-)
>
> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
> index cc63cf953636..acc264b97903 100644
> --- a/mm/damon/ops-common.c
> +++ b/mm/damon/ops-common.c
> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
> return folio;
> }
>
> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
> {
> bool referenced = false;
> struct folio *folio = damon_get_folio(pte_pfn(*pte));
> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> if (!folio)
> return;
>
> - if (pte_young(*pte)) {
> + if (ptep_test_and_clear_young(vma, addr, pte))
> referenced = true;
> - *pte = pte_mkold(*pte);
> - }
>
> #ifdef CONFIG_MMU_NOTIFIER
> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
> referenced = true;
> #endif /* CONFIG_MMU_NOTIFIER */
>
> @@ -62,7 +60,7 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> folio_put(folio);
> }
>
> -void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
> +void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr)
> {
> #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> bool referenced = false;
> @@ -71,13 +69,11 @@ void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
> if (!folio)
> return;
>
> - if (pmd_young(*pmd)) {
> + if (pmdp_test_and_clear_young(vma, addr, pmd))
> referenced = true;
> - *pmd = pmd_mkold(*pmd);
> - }
>
> #ifdef CONFIG_MMU_NOTIFIER
> - if (mmu_notifier_clear_young(mm, addr, addr + HPAGE_PMD_SIZE))
> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + HPAGE_PMD_SIZE))
> referenced = true;
> #endif /* CONFIG_MMU_NOTIFIER */
>
> diff --git a/mm/damon/ops-common.h b/mm/damon/ops-common.h
> index 14f4bc69f29b..18d837d11bce 100644
> --- a/mm/damon/ops-common.h
> +++ b/mm/damon/ops-common.h
> @@ -9,8 +9,8 @@
>
> struct folio *damon_get_folio(unsigned long pfn);
>
> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr);
> -void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr);
> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr);
> +void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr);
>
> int damon_cold_score(struct damon_ctx *c, struct damon_region *r,
> struct damos *s);
> diff --git a/mm/damon/paddr.c b/mm/damon/paddr.c
> index 467b99166b43..5b3a3463d078 100644
> --- a/mm/damon/paddr.c
> +++ b/mm/damon/paddr.c
> @@ -24,9 +24,9 @@ static bool __damon_pa_mkold(struct folio *folio, struct vm_area_struct *vma,
> while (page_vma_mapped_walk(&pvmw)) {
> addr = pvmw.address;
> if (pvmw.pte)
> - damon_ptep_mkold(pvmw.pte, vma->vm_mm, addr);
> + damon_ptep_mkold(pvmw.pte, vma, addr);
> else
> - damon_pmdp_mkold(pvmw.pmd, vma->vm_mm, addr);
> + damon_pmdp_mkold(pvmw.pmd, vma, addr);
> }
> return true;
> }
> diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
> index 1fec16d7263e..37994fb6120c 100644
> --- a/mm/damon/vaddr.c
> +++ b/mm/damon/vaddr.c
> @@ -311,7 +311,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
> }
>
> if (pmd_trans_huge(*pmd)) {
> - damon_pmdp_mkold(pmd, walk->mm, addr);
> + damon_pmdp_mkold(pmd, walk->vma, addr);
> spin_unlock(ptl);
> return 0;
> }
> @@ -323,7 +323,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
> pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
> if (!pte_present(*pte))
> goto out;
> - damon_ptep_mkold(pte, walk->mm, addr);
> + damon_ptep_mkold(pte, walk->vma, addr);
> out:
> pte_unmap_unlock(pte, ptl);
> return 0;
> --
> 2.25.1
>
>

2023-05-18 23:35:32

by Yu Zhao

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On Thu, May 18, 2023 at 5:07 AM Ryan Roberts <[email protected]> wrote:
>
> It is racy to non-atomically read a pte, then clear the young bit, then
> write it back as this could discard dirty information. Further, it is
> bad practice to directly set a pte entry within a table. Instead
> clearing young must go through the arch-provided helper,
> ptep_test_and_clear_young() to ensure it is modified atomically and to
> give the arch code visibility and allow it to check (and potentially
> modify) the operation.
>
> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")

This should be a separate patch, since it's independent from what the
series tries to do.

And that patch should cc stable, since it fixes user data corruption.

> Signed-off-by: Ryan Roberts <[email protected]>
> Reviewed-by: Zi Yan <[email protected]>
> ---
> mm/damon/ops-common.c | 16 ++++++----------
> mm/damon/ops-common.h | 4 ++--
> mm/damon/paddr.c | 4 ++--
> mm/damon/vaddr.c | 4 ++--
> 4 files changed, 12 insertions(+), 16 deletions(-)
>
> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
> index cc63cf953636..acc264b97903 100644
> --- a/mm/damon/ops-common.c
> +++ b/mm/damon/ops-common.c
> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
> return folio;
> }
>
> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
> {
> bool referenced = false;
> struct folio *folio = damon_get_folio(pte_pfn(*pte));
> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> if (!folio)
> return;
>
> - if (pte_young(*pte)) {
> + if (ptep_test_and_clear_young(vma, addr, pte))
> referenced = true;
> - *pte = pte_mkold(*pte);
> - }
>
> #ifdef CONFIG_MMU_NOTIFIER
> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
> referenced = true;
> #endif /* CONFIG_MMU_NOTIFIER */

Use ptep_clear_young_notify(). Similar below.

2023-05-19 08:55:28

by Ryan Roberts

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On 18/05/2023 18:13, SeongJae Park wrote:
> Hi Ryan,
>
>
> Nit. For consistency of DAMON commits subjects, I'd prefer something like
> "mm/damon/ops-common: atomically test and clear young on ptes and pmds".
>
> On Thu, 18 May 2023 12:07:24 +0100 Ryan Roberts <[email protected]> wrote:
>
>> It is racy to non-atomically read a pte, then clear the young bit, then
>> write it back as this could discard dirty information. Further, it is
>> bad practice to directly set a pte entry within a table. Instead
>> clearing young must go through the arch-provided helper,
>> ptep_test_and_clear_young() to ensure it is modified atomically and to
>> give the arch code visibility and allow it to check (and potentially
>> modify) the operation.
>>
>> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
>
> Nit. The commit is only for refactoring. The original code camde from commit
> 3f49584b262c ("mm/damon: implement primitives for the virtual memory address
> spaces").
>
>> Signed-off-by: Ryan Roberts <[email protected]>
>> Reviewed-by: Zi Yan <[email protected]>
>
> Other than above nits,
>
> Reviewed-by: SeongJae Park <[email protected]>

Thanks for the reviews! I'll fix the nits you raised for v3.

>
>
> Thanks,
> SJ
>

2023-05-19 09:04:25

by Ryan Roberts

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On 19/05/2023 00:19, Yu Zhao wrote:
> On Thu, May 18, 2023 at 5:07 AM Ryan Roberts <[email protected]> wrote:
>>
>> It is racy to non-atomically read a pte, then clear the young bit, then
>> write it back as this could discard dirty information. Further, it is
>> bad practice to directly set a pte entry within a table. Instead
>> clearing young must go through the arch-provided helper,
>> ptep_test_and_clear_young() to ensure it is modified atomically and to
>> give the arch code visibility and allow it to check (and potentially
>> modify) the operation.
>>
>> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
>
> This should be a separate patch, since it's independent from what the
> series tries to do.
>
> And that patch should cc stable, since it fixes user data corruption.

Fair point. The first 3 patches are fixes for issues I found during the
conversion. So if you're ok with it, I'll split into 2 series; the first 3
patches in the first, and the conversion to ptep_deref() (or ptep_get(), as per
your comments in the other thread) in the second. I guess the whole first series
should go to stable.

Let me know if you disagree.

>
>> Signed-off-by: Ryan Roberts <[email protected]>
>> Reviewed-by: Zi Yan <[email protected]>
>> ---
>> mm/damon/ops-common.c | 16 ++++++----------
>> mm/damon/ops-common.h | 4 ++--
>> mm/damon/paddr.c | 4 ++--
>> mm/damon/vaddr.c | 4 ++--
>> 4 files changed, 12 insertions(+), 16 deletions(-)
>>
>> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
>> index cc63cf953636..acc264b97903 100644
>> --- a/mm/damon/ops-common.c
>> +++ b/mm/damon/ops-common.c
>> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
>> return folio;
>> }
>>
>> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
>> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
>> {
>> bool referenced = false;
>> struct folio *folio = damon_get_folio(pte_pfn(*pte));
>> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
>> if (!folio)
>> return;
>>
>> - if (pte_young(*pte)) {
>> + if (ptep_test_and_clear_young(vma, addr, pte))
>> referenced = true;
>> - *pte = pte_mkold(*pte);
>> - }
>>
>> #ifdef CONFIG_MMU_NOTIFIER
>> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
>> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
>> referenced = true;
>> #endif /* CONFIG_MMU_NOTIFIER */
>
> Use ptep_clear_young_notify(). Similar below.

This looks sensible but I'd like clarification from SeongJae: Are you happy for
me to do this refactoring as part of the patch?

Thanks,
Ryan


2023-05-19 20:08:45

by SeongJae Park

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On Fri, 19 May 2023 10:02:48 +0100 Ryan Roberts <[email protected]> wrote:

> On 19/05/2023 00:19, Yu Zhao wrote:
> > On Thu, May 18, 2023 at 5:07 AM Ryan Roberts <[email protected]> wrote:
> >>
> >> It is racy to non-atomically read a pte, then clear the young bit, then
> >> write it back as this could discard dirty information. Further, it is
> >> bad practice to directly set a pte entry within a table. Instead
> >> clearing young must go through the arch-provided helper,
> >> ptep_test_and_clear_young() to ensure it is modified atomically and to
> >> give the arch code visibility and allow it to check (and potentially
> >> modify) the operation.
> >>
> >> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
> >
> > This should be a separate patch, since it's independent from what the
> > series tries to do.
> >
> > And that patch should cc stable, since it fixes user data corruption.
>
> Fair point. The first 3 patches are fixes for issues I found during the
> conversion. So if you're ok with it, I'll split into 2 series; the first 3
> patches in the first, and the conversion to ptep_deref() (or ptep_get(), as per
> your comments in the other thread) in the second. I guess the whole first series
> should go to stable.
>
> Let me know if you disagree.
>
> >
> >> Signed-off-by: Ryan Roberts <[email protected]>
> >> Reviewed-by: Zi Yan <[email protected]>
> >> ---
> >> mm/damon/ops-common.c | 16 ++++++----------
> >> mm/damon/ops-common.h | 4 ++--
> >> mm/damon/paddr.c | 4 ++--
> >> mm/damon/vaddr.c | 4 ++--
> >> 4 files changed, 12 insertions(+), 16 deletions(-)
> >>
> >> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
> >> index cc63cf953636..acc264b97903 100644
> >> --- a/mm/damon/ops-common.c
> >> +++ b/mm/damon/ops-common.c
> >> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
> >> return folio;
> >> }
> >>
> >> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> >> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
> >> {
> >> bool referenced = false;
> >> struct folio *folio = damon_get_folio(pte_pfn(*pte));
> >> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> >> if (!folio)
> >> return;
> >>
> >> - if (pte_young(*pte)) {
> >> + if (ptep_test_and_clear_young(vma, addr, pte))
> >> referenced = true;
> >> - *pte = pte_mkold(*pte);
> >> - }
> >>
> >> #ifdef CONFIG_MMU_NOTIFIER
> >> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
> >> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
> >> referenced = true;
> >> #endif /* CONFIG_MMU_NOTIFIER */
> >
> > Use ptep_clear_young_notify(). Similar below.
>
> This looks sensible but I'd like clarification from SeongJae: Are you happy for
> me to do this refactoring as part of the patch?

Yes, I would be happy for that :) Nevertheless, because that's a refactoring
rather than a fix, I'd be happier if you could do that with a separate patch.


Thanks,
SJ

>
> Thanks,
> Ryan
>

2023-05-22 09:01:09

by Ryan Roberts

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On 19/05/2023 20:54, SeongJae Park wrote:
> On Fri, 19 May 2023 10:02:48 +0100 Ryan Roberts <[email protected]> wrote:
>
>> On 19/05/2023 00:19, Yu Zhao wrote:
>>> On Thu, May 18, 2023 at 5:07 AM Ryan Roberts <[email protected]> wrote:
>>>>
>>>> It is racy to non-atomically read a pte, then clear the young bit, then
>>>> write it back as this could discard dirty information. Further, it is
>>>> bad practice to directly set a pte entry within a table. Instead
>>>> clearing young must go through the arch-provided helper,
>>>> ptep_test_and_clear_young() to ensure it is modified atomically and to
>>>> give the arch code visibility and allow it to check (and potentially
>>>> modify) the operation.
>>>>
>>>> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
>>>
>>> This should be a separate patch, since it's independent from what the
>>> series tries to do.
>>>
>>> And that patch should cc stable, since it fixes user data corruption.
>>
>> Fair point. The first 3 patches are fixes for issues I found during the
>> conversion. So if you're ok with it, I'll split into 2 series; the first 3
>> patches in the first, and the conversion to ptep_deref() (or ptep_get(), as per
>> your comments in the other thread) in the second. I guess the whole first series
>> should go to stable.
>>
>> Let me know if you disagree.
>>
>>>
>>>> Signed-off-by: Ryan Roberts <[email protected]>
>>>> Reviewed-by: Zi Yan <[email protected]>
>>>> ---
>>>> mm/damon/ops-common.c | 16 ++++++----------
>>>> mm/damon/ops-common.h | 4 ++--
>>>> mm/damon/paddr.c | 4 ++--
>>>> mm/damon/vaddr.c | 4 ++--
>>>> 4 files changed, 12 insertions(+), 16 deletions(-)
>>>>
>>>> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
>>>> index cc63cf953636..acc264b97903 100644
>>>> --- a/mm/damon/ops-common.c
>>>> +++ b/mm/damon/ops-common.c
>>>> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
>>>> return folio;
>>>> }
>>>>
>>>> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
>>>> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
>>>> {
>>>> bool referenced = false;
>>>> struct folio *folio = damon_get_folio(pte_pfn(*pte));
>>>> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
>>>> if (!folio)
>>>> return;
>>>>
>>>> - if (pte_young(*pte)) {
>>>> + if (ptep_test_and_clear_young(vma, addr, pte))
>>>> referenced = true;
>>>> - *pte = pte_mkold(*pte);
>>>> - }
>>>>
>>>> #ifdef CONFIG_MMU_NOTIFIER
>>>> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
>>>> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
>>>> referenced = true;
>>>> #endif /* CONFIG_MMU_NOTIFIER */
>>>
>>> Use ptep_clear_young_notify(). Similar below.
>>
>> This looks sensible but I'd like clarification from SeongJae: Are you happy for
>> me to do this refactoring as part of the patch?
>
> Yes, I would be happy for that :) Nevertheless, because that's a refactoring
> rather than a fix, I'd be happier if you could do that with a separate patch.

Great - I'll do it like this for v3.

>
>
> Thanks,
> SJ
>
>>
>> Thanks,
>> Ryan
>>


2023-05-24 19:10:36

by Mike Rapoport

[permalink] [raw]
Subject: Re: [PATCH v2 2/5] mm: damon must atomically clear young on ptes and pmds

On Thu, May 18, 2023 at 12:07:24PM +0100, Ryan Roberts wrote:
> It is racy to non-atomically read a pte, then clear the young bit, then
> write it back as this could discard dirty information. Further, it is
> bad practice to directly set a pte entry within a table. Instead
> clearing young must go through the arch-provided helper,
> ptep_test_and_clear_young() to ensure it is modified atomically and to
> give the arch code visibility and allow it to check (and potentially
> modify) the operation.
>
> Fixes: 46c3a0accdc4 ("mm/damon/vaddr: separate commonly usable functions")
> Signed-off-by: Ryan Roberts <[email protected]>
> Reviewed-by: Zi Yan <[email protected]>

Reviewed-by: Mike Rapoport (IBM) <[email protected]>

> ---
> mm/damon/ops-common.c | 16 ++++++----------
> mm/damon/ops-common.h | 4 ++--
> mm/damon/paddr.c | 4 ++--
> mm/damon/vaddr.c | 4 ++--
> 4 files changed, 12 insertions(+), 16 deletions(-)
>
> diff --git a/mm/damon/ops-common.c b/mm/damon/ops-common.c
> index cc63cf953636..acc264b97903 100644
> --- a/mm/damon/ops-common.c
> +++ b/mm/damon/ops-common.c
> @@ -37,7 +37,7 @@ struct folio *damon_get_folio(unsigned long pfn)
> return folio;
> }
>
> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr)
> {
> bool referenced = false;
> struct folio *folio = damon_get_folio(pte_pfn(*pte));
> @@ -45,13 +45,11 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> if (!folio)
> return;
>
> - if (pte_young(*pte)) {
> + if (ptep_test_and_clear_young(vma, addr, pte))
> referenced = true;
> - *pte = pte_mkold(*pte);
> - }
>
> #ifdef CONFIG_MMU_NOTIFIER
> - if (mmu_notifier_clear_young(mm, addr, addr + PAGE_SIZE))
> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + PAGE_SIZE))
> referenced = true;
> #endif /* CONFIG_MMU_NOTIFIER */
>
> @@ -62,7 +60,7 @@ void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr)
> folio_put(folio);
> }
>
> -void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
> +void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr)
> {
> #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> bool referenced = false;
> @@ -71,13 +69,11 @@ void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr)
> if (!folio)
> return;
>
> - if (pmd_young(*pmd)) {
> + if (pmdp_test_and_clear_young(vma, addr, pmd))
> referenced = true;
> - *pmd = pmd_mkold(*pmd);
> - }
>
> #ifdef CONFIG_MMU_NOTIFIER
> - if (mmu_notifier_clear_young(mm, addr, addr + HPAGE_PMD_SIZE))
> + if (mmu_notifier_clear_young(vma->vm_mm, addr, addr + HPAGE_PMD_SIZE))
> referenced = true;
> #endif /* CONFIG_MMU_NOTIFIER */
>
> diff --git a/mm/damon/ops-common.h b/mm/damon/ops-common.h
> index 14f4bc69f29b..18d837d11bce 100644
> --- a/mm/damon/ops-common.h
> +++ b/mm/damon/ops-common.h
> @@ -9,8 +9,8 @@
>
> struct folio *damon_get_folio(unsigned long pfn);
>
> -void damon_ptep_mkold(pte_t *pte, struct mm_struct *mm, unsigned long addr);
> -void damon_pmdp_mkold(pmd_t *pmd, struct mm_struct *mm, unsigned long addr);
> +void damon_ptep_mkold(pte_t *pte, struct vm_area_struct *vma, unsigned long addr);
> +void damon_pmdp_mkold(pmd_t *pmd, struct vm_area_struct *vma, unsigned long addr);
>
> int damon_cold_score(struct damon_ctx *c, struct damon_region *r,
> struct damos *s);
> diff --git a/mm/damon/paddr.c b/mm/damon/paddr.c
> index 467b99166b43..5b3a3463d078 100644
> --- a/mm/damon/paddr.c
> +++ b/mm/damon/paddr.c
> @@ -24,9 +24,9 @@ static bool __damon_pa_mkold(struct folio *folio, struct vm_area_struct *vma,
> while (page_vma_mapped_walk(&pvmw)) {
> addr = pvmw.address;
> if (pvmw.pte)
> - damon_ptep_mkold(pvmw.pte, vma->vm_mm, addr);
> + damon_ptep_mkold(pvmw.pte, vma, addr);
> else
> - damon_pmdp_mkold(pvmw.pmd, vma->vm_mm, addr);
> + damon_pmdp_mkold(pvmw.pmd, vma, addr);
> }
> return true;
> }
> diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
> index 1fec16d7263e..37994fb6120c 100644
> --- a/mm/damon/vaddr.c
> +++ b/mm/damon/vaddr.c
> @@ -311,7 +311,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
> }
>
> if (pmd_trans_huge(*pmd)) {
> - damon_pmdp_mkold(pmd, walk->mm, addr);
> + damon_pmdp_mkold(pmd, walk->vma, addr);
> spin_unlock(ptl);
> return 0;
> }
> @@ -323,7 +323,7 @@ static int damon_mkold_pmd_entry(pmd_t *pmd, unsigned long addr,
> pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
> if (!pte_present(*pte))
> goto out;
> - damon_ptep_mkold(pte, walk->mm, addr);
> + damon_ptep_mkold(pte, walk->vma, addr);
> out:
> pte_unmap_unlock(pte, ptl);
> return 0;
> --
> 2.25.1
>
>

--
Sincerely yours,
Mike.