mm: pgtable: Ensure pml spinlock gets unlock

[PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by I Hsin Cheng 12 months ago

When !start_pte is true, the "pml" spinlock is still being holded and
the branch "out_pte" is taken. If "ptl" is equal to "pml", the lock
"pml" will still be locked when the function returns.

It'll be better to set a new branch "out_pte" and jump to it when
!start_pte is true at the first place, therefore no additional check for
"start_pte" or "ptl != pml" is needed, simply unlock "pml" and return.

Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
---
 mm/pt_reclaim.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
index 7e9455a18aae..163e38f1728d 100644
--- a/mm/pt_reclaim.c
+++ b/mm/pt_reclaim.c
@@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
 	pml = pmd_lock(mm, pmd);
 	start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, &ptl);
 	if (!start_pte)
-		goto out_ptl;
+		goto out_pte;
 	if (ptl != pml)
 		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
 
@@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
 		pte_unmap_unlock(start_pte, ptl);
 	if (ptl != pml)
 		spin_unlock(pml);
+	return;
+
+out_pte:
+	spin_unlock(pml);
 }
-- 
2.43.0

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by Qi Zheng 12 months ago


On 2025/2/9 02:49, I Hsin Cheng wrote:
> When !start_pte is true, the "pml" spinlock is still being holded and
> the branch "out_pte" is taken. If "ptl" is equal to "pml", the lock
> "pml" will still be locked when the function returns.

No. When start_pte is NULL, the ptl must also be NULL, so the ptl and
pml will not be equal.

> 
> It'll be better to set a new branch "out_pte" and jump to it when
> !start_pte is true at the first place, therefore no additional check for
> "start_pte" or "ptl != pml" is needed, simply unlock "pml" and return.
> 
> Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
> ---
>   mm/pt_reclaim.c | 6 +++++-
>   1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
> index 7e9455a18aae..163e38f1728d 100644
> --- a/mm/pt_reclaim.c
> +++ b/mm/pt_reclaim.c
> @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
>   	pml = pmd_lock(mm, pmd);
>   	start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, &ptl);
>   	if (!start_pte)
> -		goto out_ptl;
> +		goto out_pte;
>   	if (ptl != pml)
>   		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>   
> @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
>   		pte_unmap_unlock(start_pte, ptl);
>   	if (ptl != pml)
>   		spin_unlock(pml);
> +	return;
> +
> +out_pte:
> +	spin_unlock(pml);
>   }

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by I Hsin Cheng 12 months ago

On Mon, Feb 10, 2025 at 12:05:05PM +0800, Qi Zheng wrote:
> 
> 
> On 2025/2/9 02:49, I Hsin Cheng wrote:
> > When !start_pte is true, the "pml" spinlock is still being holded and
> > the branch "out_pte" is taken. If "ptl" is equal to "pml", the lock
> > "pml" will still be locked when the function returns.
> 
> No. When start_pte is NULL, the ptl must also be NULL, so the ptl and
> pml will not be equal.
> 
> > 
> > It'll be better to set a new branch "out_pte" and jump to it when
> > !start_pte is true at the first place, therefore no additional check for
> > "start_pte" or "ptl != pml" is needed, simply unlock "pml" and return.
> > 
> > Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
> > ---
> >   mm/pt_reclaim.c | 6 +++++-
> >   1 file changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
> > index 7e9455a18aae..163e38f1728d 100644
> > --- a/mm/pt_reclaim.c
> > +++ b/mm/pt_reclaim.c
> > @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
> >   	pml = pmd_lock(mm, pmd);
> >   	start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, &ptl);
> >   	if (!start_pte)
> > -		goto out_ptl;
> > +		goto out_pte;
> >   	if (ptl != pml)
> >   		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
> > @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
> >   		pte_unmap_unlock(start_pte, ptl);
> >   	if (ptl != pml)
> >   		spin_unlock(pml);
> > +	return;
> > +
> > +out_pte:
> > +	spin_unlock(pml);
> >   }

Hi Qi,

Thanks for your kindly review!

> No. When start_pte is NULL, the ptl must also be NULL, so the ptl and
> pml will not be equal.

Since this is the case, we don't have to do any addtional check for
"start_pte" and "ptl != pml", we can be sure that only the second check
will be true so we can unlock "pml" without the redundant check, that's
my understanding, what do you think?

Best regards,
I Hsin Cheng

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by Qi Zheng 12 months ago


On 2025/2/10 16:20, I Hsin Cheng wrote:
> On Mon, Feb 10, 2025 at 12:05:05PM +0800, Qi Zheng wrote:
>>
>>
>> On 2025/2/9 02:49, I Hsin Cheng wrote:
>>> When !start_pte is true, the "pml" spinlock is still being holded and
>>> the branch "out_pte" is taken. If "ptl" is equal to "pml", the lock
>>> "pml" will still be locked when the function returns.
>>
>> No. When start_pte is NULL, the ptl must also be NULL, so the ptl and
>> pml will not be equal.
>>
>>>
>>> It'll be better to set a new branch "out_pte" and jump to it when
>>> !start_pte is true at the first place, therefore no additional check for
>>> "start_pte" or "ptl != pml" is needed, simply unlock "pml" and return.
>>>
>>> Signed-off-by: I Hsin Cheng <richard120310@gmail.com>
>>> ---
>>>    mm/pt_reclaim.c | 6 +++++-
>>>    1 file changed, 5 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
>>> index 7e9455a18aae..163e38f1728d 100644
>>> --- a/mm/pt_reclaim.c
>>> +++ b/mm/pt_reclaim.c
>>> @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
>>>    	pml = pmd_lock(mm, pmd);
>>>    	start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, &ptl);
>>>    	if (!start_pte)
>>> -		goto out_ptl;
>>> +		goto out_pte;
>>>    	if (ptl != pml)
>>>    		spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>>> @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t *pmd, unsigned long addr,
>>>    		pte_unmap_unlock(start_pte, ptl);
>>>    	if (ptl != pml)
>>>    		spin_unlock(pml);
>>> +	return;
>>> +
>>> +out_pte:
>>> +	spin_unlock(pml);
>>>    }
> 
> Hi Qi,
> 
> Thanks for your kindly review!
> 
>> No. When start_pte is NULL, the ptl must also be NULL, so the ptl and
>> pml will not be equal.
> 
> Since this is the case, we don't have to do any addtional check for
> "start_pte" and "ptl != pml", we can be sure that only the second check
> will be true so we can unlock "pml" without the redundant check, that's
> my understanding, what do you think?
Adding a label

vs

Redundant check in rare cases

Not sure if this is worth it. ;)

> 
> Best regards,
> I Hsin Cheng
> 
>

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by Qi Zheng 12 months ago


On 2025/2/10 16:31, Qi Zheng wrote:
> 

[...]

>>>>
>>>> diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
>>>> index 7e9455a18aae..163e38f1728d 100644
>>>> --- a/mm/pt_reclaim.c
>>>> +++ b/mm/pt_reclaim.c
>>>> @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t 
>>>> *pmd, unsigned long addr,
>>>>        pml = pmd_lock(mm, pmd);
>>>>        start_pte = pte_offset_map_rw_nolock(mm, pmd, addr, &pmdval, 
>>>> &ptl);
>>>>        if (!start_pte)

Maybe we can return directly here:

             if (!start_pte) {
		spin_unlock(pml);
                 return;
             }

>>>> -        goto out_ptl;
>>>> +        goto out_pte;
>>>>        if (ptl != pml)
>>>>            spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
>>>> @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm, pmd_t 
>>>> *pmd, unsigned long addr,
>>>>            pte_unmap_unlock(start_pte, ptl);
>>>>        if (ptl != pml)
>>>>            spin_unlock(pml);
>>>> +    return;
>>>> +
>>>> +out_pte:
>>>> +    spin_unlock(pml);
>>>>    }
>>

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by I Hsin Cheng 12 months ago

On Mon, Feb 10, 2025 at 04:42:13PM +0800, Qi Zheng wrote:
> 
> 
> On 2025/2/10 16:31, Qi Zheng wrote:
> > 
> 
> [...]
> 
> > > > > 
> > > > > diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
> > > > > index 7e9455a18aae..163e38f1728d 100644
> > > > > --- a/mm/pt_reclaim.c
> > > > > +++ b/mm/pt_reclaim.c
> > > > > @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm,
> > > > > pmd_t *pmd, unsigned long addr,
> > > > >        pml = pmd_lock(mm, pmd);
> > > > >        start_pte = pte_offset_map_rw_nolock(mm, pmd, addr,
> > > > > &pmdval, &ptl);
> > > > >        if (!start_pte)
> 
> Maybe we can return directly here:
> 
>             if (!start_pte) {
> 		spin_unlock(pml);
>                 return;
>             }
> 
> > > > > -        goto out_ptl;
> > > > > +        goto out_pte;
> > > > >        if (ptl != pml)
> > > > >            spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
> > > > > @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm,
> > > > > pmd_t *pmd, unsigned long addr,
> > > > >            pte_unmap_unlock(start_pte, ptl);
> > > > >        if (ptl != pml)
> > > > >            spin_unlock(pml);
> > > > > +    return;
> > > > > +
> > > > > +out_pte:
> > > > > +    spin_unlock(pml);
> > > > >    }
> > >

I've send a new patch stating for this change and change the title,
because pml will sure get unlocked, we just prevent the redundant
branches.

Best regards,
I Hsin Cheng

Re: [PATCH] mm: pgtable: Ensure pml spinlock gets unlock

Posted by I Hsin Cheng 12 months ago

On Mon, Feb 10, 2025 at 04:42:13PM +0800, Qi Zheng wrote:
> 
> 
> On 2025/2/10 16:31, Qi Zheng wrote:
> > 
> 
> [...]
> 
> > > > > 
> > > > > diff --git a/mm/pt_reclaim.c b/mm/pt_reclaim.c
> > > > > index 7e9455a18aae..163e38f1728d 100644
> > > > > --- a/mm/pt_reclaim.c
> > > > > +++ b/mm/pt_reclaim.c
> > > > > @@ -43,7 +43,7 @@ void try_to_free_pte(struct mm_struct *mm,
> > > > > pmd_t *pmd, unsigned long addr,
> > > > >        pml = pmd_lock(mm, pmd);
> > > > >        start_pte = pte_offset_map_rw_nolock(mm, pmd, addr,
> > > > > &pmdval, &ptl);
> > > > >        if (!start_pte)
> 
> Maybe we can return directly here:
> 
>             if (!start_pte) {
> 		spin_unlock(pml);
>                 return;
>             }
> 
> > > > > -        goto out_ptl;
> > > > > +        goto out_pte;
> > > > >        if (ptl != pml)
> > > > >            spin_lock_nested(ptl, SINGLE_DEPTH_NESTING);
> > > > > @@ -68,4 +68,8 @@ void try_to_free_pte(struct mm_struct *mm,
> > > > > pmd_t *pmd, unsigned long addr,
> > > > >            pte_unmap_unlock(start_pte, ptl);
> > > > >        if (ptl != pml)
> > > > >            spin_unlock(pml);
> > > > > +    return;
> > > > > +
> > > > > +out_pte:
> > > > > +    spin_unlock(pml);
> > > > >    }
> > >

Hi Qi,

> Maybe we can return directly here:
>
>             if (!start_pte) {
>               spin_unlock(pml);
>                 return;
>             }

Ahh that's right I think it's better than adding another label,
it doesn't even need to jump.

Should I send patch v2 for it or wait for the maintainer's review
first?

Best regards,
I Hsin Cheng