mm: move mask update out of the atomic context

[PATCH] mm: move mask update out of the atomic context

Posted by Alexander Gordeev 7 months, 2 weeks ago

There is not need to modify page table synchronization mask
while apply_to_pte_range() holds user page tables spinlock.

Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
---
 mm/memory.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/memory.c b/mm/memory.c
index 8eba595056fe..6849ab4e44bf 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -3035,12 +3035,13 @@ static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
 			}
 		} while (pte++, addr += PAGE_SIZE, addr != end);
 	}
-	*mask |= PGTBL_PTE_MODIFIED;
 
 	arch_leave_lazy_mmu_mode();
 
 	if (mm != &init_mm)
 		pte_unmap_unlock(mapped_pte, ptl);
+	*mask |= PGTBL_PTE_MODIFIED;
+
 	return err;
 }
 
-- 
2.48.1

Re: [PATCH] mm: move mask update out of the atomic context

Posted by Dev Jain 7 months, 2 weeks ago

On 23/06/25 1:34 pm, Alexander Gordeev wrote:
> There is not need to modify page table synchronization mask
> while apply_to_pte_range() holds user page tables spinlock.

I don't get you, what is the problem with the current code?
Are you just concerned about the duration of holding the
lock?

>
> Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
> ---
>   mm/memory.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 8eba595056fe..6849ab4e44bf 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -3035,12 +3035,13 @@ static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
>   			}
>   		} while (pte++, addr += PAGE_SIZE, addr != end);
>   	}
> -	*mask |= PGTBL_PTE_MODIFIED;
>   
>   	arch_leave_lazy_mmu_mode();
>   
>   	if (mm != &init_mm)
>   		pte_unmap_unlock(mapped_pte, ptl);
> +	*mask |= PGTBL_PTE_MODIFIED;
> +
>   	return err;
>   }
>

Re: [PATCH] mm: move mask update out of the atomic context

Posted by Alexander Gordeev 7 months, 2 weeks ago

On Mon, Jun 23, 2025 at 02:26:29PM +0530, Dev Jain wrote:
> 
> On 23/06/25 1:34 pm, Alexander Gordeev wrote:
> > There is not need to modify page table synchronization mask
> > while apply_to_pte_range() holds user page tables spinlock.
> 
> I don't get you, what is the problem with the current code?
> Are you just concerned about the duration of holding the
> lock?

Yes.

> > Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
> > ---
> >   mm/memory.c | 3 ++-
> >   1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 8eba595056fe..6849ab4e44bf 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -3035,12 +3035,13 @@ static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
> >   			}
> >   		} while (pte++, addr += PAGE_SIZE, addr != end);
> >   	}
> > -	*mask |= PGTBL_PTE_MODIFIED;
> >   	arch_leave_lazy_mmu_mode();
> >   	if (mm != &init_mm)
> >   		pte_unmap_unlock(mapped_pte, ptl);
> > +	*mask |= PGTBL_PTE_MODIFIED;
> > +
> >   	return err;
> >   }

Re: [PATCH] mm: move mask update out of the atomic context

Posted by Dev Jain 7 months, 2 weeks ago

On 23/06/25 3:07 pm, Alexander Gordeev wrote:
> On Mon, Jun 23, 2025 at 02:26:29PM +0530, Dev Jain wrote:
>> On 23/06/25 1:34 pm, Alexander Gordeev wrote:
>>> There is not need to modify page table synchronization mask
>>> while apply_to_pte_range() holds user page tables spinlock.
>> I don't get you, what is the problem with the current code?
>> Are you just concerned about the duration of holding the
>> lock?
> Yes.

Doesn't really matter but still a correct change:

Reviewed-by: Dev Jain <dev.jain@arm.com>

>
>>> Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
>>> ---
>>>    mm/memory.c | 3 ++-
>>>    1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/mm/memory.c b/mm/memory.c
>>> index 8eba595056fe..6849ab4e44bf 100644
>>> --- a/mm/memory.c
>>> +++ b/mm/memory.c
>>> @@ -3035,12 +3035,13 @@ static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
>>>    			}
>>>    		} while (pte++, addr += PAGE_SIZE, addr != end);
>>>    	}
>>> -	*mask |= PGTBL_PTE_MODIFIED;
>>>    	arch_leave_lazy_mmu_mode();
>>>    	if (mm != &init_mm)
>>>    		pte_unmap_unlock(mapped_pte, ptl);
>>> +	*mask |= PGTBL_PTE_MODIFIED;
>>> +
>>>    	return err;
>>>    }

Re: [PATCH] mm: move mask update out of the atomic context

Posted by David Hildenbrand 7 months, 2 weeks ago

On 23.06.25 12:09, Dev Jain wrote:
> 
> On 23/06/25 3:07 pm, Alexander Gordeev wrote:
>> On Mon, Jun 23, 2025 at 02:26:29PM +0530, Dev Jain wrote:
>>> On 23/06/25 1:34 pm, Alexander Gordeev wrote:
>>>> There is not need to modify page table synchronization mask
>>>> while apply_to_pte_range() holds user page tables spinlock.
>>> I don't get you, what is the problem with the current code?
>>> Are you just concerned about the duration of holding the
>>> lock?
>> Yes.
> 
> Doesn't really matter but still a correct change:

Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?

I see

if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
	arch_sync_kernel_mappings(start, start + size);

And then

arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED


Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?

(same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)

-- 
Cheers,

David / dhildenb

Re: [PATCH] mm: move mask update out of the atomic context

Posted by Alexander Gordeev 7 months, 2 weeks ago

On Mon, Jun 23, 2025 at 09:45:34PM +0200, David Hildenbrand wrote:
...
> Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?
> 
> I see
> 
> if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
> 	arch_sync_kernel_mappings(start, start + size);
> 
> And then
> 
> arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> 
> 
> Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?
> 
> (same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)

AFAICT it was thought as architecture-specific:

/*                                                                              
 * Architectures can set this mask to a combination of PGTBL_P?D_MODIFIED values
 * and let generic vmalloc and ioremap code know when arch_sync_kernel_mappings()
 * needs to be called.                                                          
 */                                                                             
#ifndef ARCH_PAGE_TABLE_SYNC_MASK
#define ARCH_PAGE_TABLE_SYNC_MASK 0
#endif

Not sure if that needs to be addressed at all.

> -- 
> Cheers,
> 
> David / dhildenb

Thanks!

Re: [PATCH] mm: move mask update out of the atomic context

Posted by David Hildenbrand 7 months, 2 weeks ago

On 24.06.25 11:37, Alexander Gordeev wrote:
> On Mon, Jun 23, 2025 at 09:45:34PM +0200, David Hildenbrand wrote:
> ...
>> Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?
>>
>> I see
>>
>> if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
>> 	arch_sync_kernel_mappings(start, start + size);
>>
>> And then
>>
>> arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>> arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>> arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>>
>>
>> Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?
>>
>> (same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)
> 
> AFAICT it was thought as architecture-specific:
> 
> /*
>   * Architectures can set this mask to a combination of PGTBL_P?D_MODIFIED values
>   * and let generic vmalloc and ioremap code know when arch_sync_kernel_mappings()
>   * needs to be called.
>   */
> #ifndef ARCH_PAGE_TABLE_SYNC_MASK
> #define ARCH_PAGE_TABLE_SYNC_MASK 0
> #endif
> 
> Not sure if that needs to be addressed at all.

Okay, if there are no users of PGTBL_PTE_MODIFIED we could just ... 
remove it. Dead code.

-- 
Cheers,

David / dhildenb

Re: [PATCH] mm: move mask update out of the atomic context

Posted by Alexander Gordeev 7 months, 2 weeks ago

On Tue, Jun 24, 2025 at 11:40:05AM +0200, David Hildenbrand wrote:
> On 24.06.25 11:37, Alexander Gordeev wrote:
> > On Mon, Jun 23, 2025 at 09:45:34PM +0200, David Hildenbrand wrote:
> > ...
> > > Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?
> > > 
> > > I see
> > > 
> > > if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
> > > 	arch_sync_kernel_mappings(start, start + size);
> > > 
> > > And then
> > > 
> > > arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
> > > 
> > > 
> > > Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?
> > > 
> > > (same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)
> > 
> > AFAICT it was thought as architecture-specific:
> > 
> > /*
> >   * Architectures can set this mask to a combination of PGTBL_P?D_MODIFIED values
> >   * and let generic vmalloc and ioremap code know when arch_sync_kernel_mappings()
> >   * needs to be called.
> >   */
> > #ifndef ARCH_PAGE_TABLE_SYNC_MASK
> > #define ARCH_PAGE_TABLE_SYNC_MASK 0
> > #endif
> > 
> > Not sure if that needs to be addressed at all.
> 
> Okay, if there are no users of PGTBL_PTE_MODIFIED we could just ... remove
> it. Dead code.

As you noticed, PGTBL_PMD_MODIFIED bit is used only. Thus, all other
bits would have to be removed as well, not just PGTBL_PTE_MODIFIED?

That is more or less revert of at least the below commits and rewriting
it in a PMD-focused manner:

2ba3e6947aed ("mm/vmalloc: track which page-table levels were modified")
d8626138009b ("mm: add functions to track page directory modifications")

That would be a completely different effort, which I am not aming at ;)

> -- 
> Cheers,
> 
> David / dhildenb

Thanks!

Re: [PATCH] mm: move mask update out of the atomic context

Posted by David Hildenbrand 7 months, 2 weeks ago

On 24.06.25 14:00, Alexander Gordeev wrote:
> On Tue, Jun 24, 2025 at 11:40:05AM +0200, David Hildenbrand wrote:
>> On 24.06.25 11:37, Alexander Gordeev wrote:
>>> On Mon, Jun 23, 2025 at 09:45:34PM +0200, David Hildenbrand wrote:
>>> ...
>>>> Let's ask the real questions: who checks PGTBL_PTE_MODIFIED?
>>>>
>>>> I see
>>>>
>>>> if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
>>>> 	arch_sync_kernel_mappings(start, start + size);
>>>>
>>>> And then
>>>>
>>>> arch/arm/include/asm/page.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>>>> arch/x86/include/asm/pgtable-2level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>>>> arch/x86/include/asm/pgtable-3level_types.h:#define ARCH_PAGE_TABLE_SYNC_MASK   PGTBL_PMD_MODIFIED
>>>>
>>>>
>>>> Which makes me wonder why we need PGTBL_PTE_MODIFIED at all? Is there some other check I am missing?
>>>>
>>>> (same question regarding everything excepy PGTBL_PMD_MODIFIED, because that actually seems to be used)
>>>
>>> AFAICT it was thought as architecture-specific:
>>>
>>> /*
>>>    * Architectures can set this mask to a combination of PGTBL_P?D_MODIFIED values
>>>    * and let generic vmalloc and ioremap code know when arch_sync_kernel_mappings()
>>>    * needs to be called.
>>>    */
>>> #ifndef ARCH_PAGE_TABLE_SYNC_MASK
>>> #define ARCH_PAGE_TABLE_SYNC_MASK 0
>>> #endif
>>>
>>> Not sure if that needs to be addressed at all.
>>
>> Okay, if there are no users of PGTBL_PTE_MODIFIED we could just ... remove
>> it. Dead code.
> 
> As you noticed, PGTBL_PMD_MODIFIED bit is used only. Thus, all other
> bits would have to be removed as well, not just PGTBL_PTE_MODIFIED?

Likely, yes.

> 
> That is more or less revert of at least the below commits and rewriting
> it in a PMD-focused manner:
> 
> 2ba3e6947aed ("mm/vmalloc: track which page-table levels were modified")
> d8626138009b ("mm: add functions to track page directory modifications")
> 
> That would be a completely different effort, which I am not aming at ;)

Well, I don't think we should be taking this micro-optimization patch 
here TBH. The real optimization is getting rid of dead code, not 
optimizing dead code.

-- 
Cheers,

David / dhildenb