[PATCH] Initialise mmu_notifier_range correctly

[PATCH] Initialise mmu_notifier_range correctly

[Matthew Wilcox]

One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
incorrectly.

Signed-off-by: Matthew Wilcox <willy@infradead.org>
Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
Tested-by: Dave Chinner <dchinner@redhat.com>

diff --git a/mm/memory.c b/mm/memory.c
index 2dd2f9ab57f4..21a650368be0 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4078,8 +4078,8 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
 		goto out;
 
 	if (range) {
-		range->start = address & PAGE_MASK;
-		range->end = range->start + PAGE_SIZE;
+		mmu_notifier_range_init(range, mm, address & PAGE_MASK,
+				     (address & PAGE_MASK) + PAGE_SIZE);
 		mmu_notifier_invalidate_range_start(range);
 	}
 	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);

Re: [PATCH] Initialise mmu_notifier_range correctly

[Jerome Glisse]

On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
> 
> One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
> incorrectly.
> 
> Signed-off-by: Matthew Wilcox <willy@infradead.org>
> Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
> Tested-by: Dave Chinner <dchinner@redhat.com>

Reviewed-by: Jérôme Glisse <jglisse@redhat.com>

> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 2dd2f9ab57f4..21a650368be0 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -4078,8 +4078,8 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
>  		goto out;
>  
>  	if (range) {
> -		range->start = address & PAGE_MASK;
> -		range->end = range->start + PAGE_SIZE;
> +		mmu_notifier_range_init(range, mm, address & PAGE_MASK,
> +				     (address & PAGE_MASK) + PAGE_SIZE);
>  		mmu_notifier_invalidate_range_start(range);
>  	}
>  	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);

Re: [PATCH] Initialise mmu_notifier_range correctly

[John Hubbard]

On 1/2/19 5:56 PM, Jerome Glisse wrote:
> On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
>>
>> One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
>> incorrectly.
>>
>> Signed-off-by: Matthew Wilcox <willy@infradead.org>
>> Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
>> Tested-by: Dave Chinner <dchinner@redhat.com>
> 
> Reviewed-by: Jérôme Glisse <jglisse@redhat.com>
> 
>>
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 2dd2f9ab57f4..21a650368be0 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4078,8 +4078,8 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
>>  		goto out;
>>  
>>  	if (range) {
>> -		range->start = address & PAGE_MASK;
>> -		range->end = range->start + PAGE_SIZE;
>> +		mmu_notifier_range_init(range, mm, address & PAGE_MASK,
>> +				     (address & PAGE_MASK) + PAGE_SIZE);
>>  		mmu_notifier_invalidate_range_start(range);
>>  	}
>>  	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
> 

Looks correct to me, as well.

Having the range struct declared in separate places from the mmu_notifier_range_init()
calls is not great. But I'm not sure I see a way to make it significantly cleaner, given
that __follow_pte_pmd uses the range pointer as a way to decide to issue the mmn calls.


thanks,
-- 
John Hubbard
NVIDIA

Re: [PATCH] Initialise mmu_notifier_range correctly

[Matthew Wilcox]

On Wed, Jan 02, 2019 at 07:32:08PM -0800, John Hubbard wrote:
> Having the range struct declared in separate places from the mmu_notifier_range_init()
> calls is not great. But I'm not sure I see a way to make it significantly cleaner, given
> that __follow_pte_pmd uses the range pointer as a way to decide to issue the mmn calls.

Yeah, I don't think there's anything we can do.  But I started reviewing
the comments, and they don't make sense together:

                /*
                 * Note because we provide range to follow_pte_pmd it will
                 * call mmu_notifier_invalidate_range_start() on our behalf
                 * before taking any lock.
                 */
                if (follow_pte_pmd(vma->vm_mm, address, &range,
                                   &ptep, &pmdp, &ptl))
                        continue;

                /*
                 * No need to call mmu_notifier_invalidate_range() as we are
                 * downgrading page table protection not changing it to point
                 * to a new page.
                 *
                 * See Documentation/vm/mmu_notifier.rst
                 */

So if we don't call mmu_notifier_invalidate_range, why are we calling
mmu_notifier_invalidate_range_start and mmu_notifier_invalidate_range_end?
ie, why not this ...

diff --git a/fs/dax.c b/fs/dax.c
index 6959837cc465..905340149924 100644
--- a/fs/dax.c
+++ b/fs/dax.c
@@ -777,7 +777,6 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
 
 	i_mmap_lock_read(mapping);
 	vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
-		struct mmu_notifier_range range;
 		unsigned long address;
 
 		cond_resched();
@@ -787,12 +786,7 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
 
 		address = pgoff_address(index, vma);
 
-		/*
-		 * Note because we provide start/end to follow_pte_pmd it will
-		 * call mmu_notifier_invalidate_range_start() on our behalf
-		 * before taking any lock.
-		 */
-		if (follow_pte_pmd(vma->vm_mm, address, &range,
+		if (follow_pte_pmd(vma->vm_mm, address, NULL,
 				   &ptep, &pmdp, &ptl))
 			continue;
 
@@ -834,8 +828,6 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
 unlock_pte:
 			pte_unmap_unlock(ptep, ptl);
 		}
-
-		mmu_notifier_invalidate_range_end(&range);
 	}
 	i_mmap_unlock_read(mapping);
 }

Re: [PATCH] Initialise mmu_notifier_range correctly

[Jerome Glisse]

On Wed, Jan 02, 2019 at 08:18:33PM -0800, Matthew Wilcox wrote:
> On Wed, Jan 02, 2019 at 07:32:08PM -0800, John Hubbard wrote:
> > Having the range struct declared in separate places from the mmu_notifier_range_init()
> > calls is not great. But I'm not sure I see a way to make it significantly cleaner, given
> > that __follow_pte_pmd uses the range pointer as a way to decide to issue the mmn calls.
> 
> Yeah, I don't think there's anything we can do.  But I started reviewing
> the comments, and they don't make sense together:
> 
>                 /*
>                  * Note because we provide range to follow_pte_pmd it will
>                  * call mmu_notifier_invalidate_range_start() on our behalf
>                  * before taking any lock.
>                  */
>                 if (follow_pte_pmd(vma->vm_mm, address, &range,
>                                    &ptep, &pmdp, &ptl))
>                         continue;
> 
>                 /*
>                  * No need to call mmu_notifier_invalidate_range() as we are
>                  * downgrading page table protection not changing it to point
>                  * to a new page.
>                  *
>                  * See Documentation/vm/mmu_notifier.rst
>                  */
> 
> So if we don't call mmu_notifier_invalidate_range, why are we calling
> mmu_notifier_invalidate_range_start and mmu_notifier_invalidate_range_end?
> ie, why not this ...

Thus comments looks wrong to me ... we need to call
mmu_notifier_invalidate_range() those are use by
IOMMU. I might be to blame for those comments thought.


> 
> diff --git a/fs/dax.c b/fs/dax.c
> index 6959837cc465..905340149924 100644
> --- a/fs/dax.c
> +++ b/fs/dax.c
> @@ -777,7 +777,6 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
>  
>  	i_mmap_lock_read(mapping);
>  	vma_interval_tree_foreach(vma, &mapping->i_mmap, index, index) {
> -		struct mmu_notifier_range range;
>  		unsigned long address;
>  
>  		cond_resched();
> @@ -787,12 +786,7 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
>  
>  		address = pgoff_address(index, vma);
>  
> -		/*
> -		 * Note because we provide start/end to follow_pte_pmd it will
> -		 * call mmu_notifier_invalidate_range_start() on our behalf
> -		 * before taking any lock.
> -		 */
> -		if (follow_pte_pmd(vma->vm_mm, address, &range,
> +		if (follow_pte_pmd(vma->vm_mm, address, NULL,
>  				   &ptep, &pmdp, &ptl))
>  			continue;
>  
> @@ -834,8 +828,6 @@ static void dax_entry_mkclean(struct address_space *mapping, pgoff_t index,
>  unlock_pte:
>  			pte_unmap_unlock(ptep, ptl);
>  		}
> -
> -		mmu_notifier_invalidate_range_end(&range);
>  	}
>  	i_mmap_unlock_read(mapping);
>  }

Re: [PATCH] Initialise mmu_notifier_range correctly

[Matthew Wilcox]

On Thu, Jan 03, 2019 at 09:29:59AM -0500, Jerome Glisse wrote:
> On Wed, Jan 02, 2019 at 08:18:33PM -0800, Matthew Wilcox wrote:
> > On Wed, Jan 02, 2019 at 07:32:08PM -0800, John Hubbard wrote:
> > > Having the range struct declared in separate places from the mmu_notifier_range_init()
> > > calls is not great. But I'm not sure I see a way to make it significantly cleaner, given
> > > that __follow_pte_pmd uses the range pointer as a way to decide to issue the mmn calls.
> > 
> > Yeah, I don't think there's anything we can do.  But I started reviewing
> > the comments, and they don't make sense together:
> > 
> >                 /*
> >                  * Note because we provide range to follow_pte_pmd it will
> >                  * call mmu_notifier_invalidate_range_start() on our behalf
> >                  * before taking any lock.
> >                  */
> >                 if (follow_pte_pmd(vma->vm_mm, address, &range,
> >                                    &ptep, &pmdp, &ptl))
> >                         continue;
> > 
> >                 /*
> >                  * No need to call mmu_notifier_invalidate_range() as we are
> >                  * downgrading page table protection not changing it to point
> >                  * to a new page.
> >                  *
> >                  * See Documentation/vm/mmu_notifier.rst
> >                  */
> > 
> > So if we don't call mmu_notifier_invalidate_range, why are we calling
> > mmu_notifier_invalidate_range_start and mmu_notifier_invalidate_range_end?
> > ie, why not this ...
> 
> Thus comments looks wrong to me ... we need to call
> mmu_notifier_invalidate_range() those are use by
> IOMMU. I might be to blame for those comments thought.

Yes, you're to blame for both of them.

a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  791)                 * Note because we provide start/end to follow_pte_pmd it will
a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  792)                 * call mmu_notifier_invalidate_range_start() on our behalf
a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  793)                 * before taking any lock.

0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  794)                 * No need to call mmu_notifier_invalidate_range() as we are
0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  795)                 * downgrading page table protection not changing it to point
0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  796)                 * to a new page.

Re: [PATCH] Initialise mmu_notifier_range correctly

[Jerome Glisse]

On Thu, Jan 03, 2019 at 06:39:08AM -0800, Matthew Wilcox wrote:
> On Thu, Jan 03, 2019 at 09:29:59AM -0500, Jerome Glisse wrote:
> > On Wed, Jan 02, 2019 at 08:18:33PM -0800, Matthew Wilcox wrote:
> > > On Wed, Jan 02, 2019 at 07:32:08PM -0800, John Hubbard wrote:
> > > > Having the range struct declared in separate places from the mmu_notifier_range_init()
> > > > calls is not great. But I'm not sure I see a way to make it significantly cleaner, given
> > > > that __follow_pte_pmd uses the range pointer as a way to decide to issue the mmn calls.
> > > 
> > > Yeah, I don't think there's anything we can do.  But I started reviewing
> > > the comments, and they don't make sense together:
> > > 
> > >                 /*
> > >                  * Note because we provide range to follow_pte_pmd it will
> > >                  * call mmu_notifier_invalidate_range_start() on our behalf
> > >                  * before taking any lock.
> > >                  */
> > >                 if (follow_pte_pmd(vma->vm_mm, address, &range,
> > >                                    &ptep, &pmdp, &ptl))
> > >                         continue;
> > > 
> > >                 /*
> > >                  * No need to call mmu_notifier_invalidate_range() as we are
> > >                  * downgrading page table protection not changing it to point
> > >                  * to a new page.
> > >                  *
> > >                  * See Documentation/vm/mmu_notifier.rst
> > >                  */
> > > 
> > > So if we don't call mmu_notifier_invalidate_range, why are we calling
> > > mmu_notifier_invalidate_range_start and mmu_notifier_invalidate_range_end?
> > > ie, why not this ...
> > 
> > Thus comments looks wrong to me ... we need to call
> > mmu_notifier_invalidate_range() those are use by
> > IOMMU. I might be to blame for those comments thought.
> 
> Yes, you're to blame for both of them.
> 
> a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  791)                 * Note because we provide start/end to follow_pte_pmd it will
> a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  792)                 * call mmu_notifier_invalidate_range_start() on our behalf
> a4d1a88525138 (Jérôme Glisse     2017-08-31 17:17:26 -0400  793)                 * before taking any lock.
> 
> 0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  794)                 * No need to call mmu_notifier_invalidate_range() as we are
> 0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  795)                 * downgrading page table protection not changing it to point
> 0f10851ea475e (Jérôme Glisse     2017-11-15 17:34:07 -0800  796)                 * to a new page.
> 

I remember now we do not need to call invalidate range because
invalidate_range_end() does call invalidate_range so it is fine.
Comments should be better thought. So existing code is fine.

Cheers,
Jérôme

Re: [PATCH] Initialise mmu_notifier_range correctly

[Jerome Glisse]

On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
> 
> One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
> incorrectly.
> 
> Signed-off-by: Matthew Wilcox <willy@infradead.org>
> Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
> Tested-by: Dave Chinner <dchinner@redhat.com>

Actually now that i have read the code again this is not ok to
do so. The caller of follow_pte_pmd() will call range_init and
follow pmd will only update the range address. So existing code
is ok.

I know this is kind of ugly but i do not see a way around that
uglyness.

> 
> diff --git a/mm/memory.c b/mm/memory.c
> index 2dd2f9ab57f4..21a650368be0 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -4078,8 +4078,8 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
>  		goto out;
>  
>  	if (range) {
> -		range->start = address & PAGE_MASK;
> -		range->end = range->start + PAGE_SIZE;
> +		mmu_notifier_range_init(range, mm, address & PAGE_MASK,
> +				     (address & PAGE_MASK) + PAGE_SIZE);
>  		mmu_notifier_invalidate_range_start(range);
>  	}
>  	ptep = pte_offset_map_lock(mm, pmd, address, ptlp);

Re: [PATCH] Initialise mmu_notifier_range correctly

[Matthew Wilcox]

On Thu, Jan 03, 2019 at 09:31:16AM -0500, Jerome Glisse wrote:
> On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
> > 
> > One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
> > incorrectly.
> > 
> > Signed-off-by: Matthew Wilcox <willy@infradead.org>
> > Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
> > Tested-by: Dave Chinner <dchinner@redhat.com>
> 
> Actually now that i have read the code again this is not ok to
> do so. The caller of follow_pte_pmd() will call range_init and
> follow pmd will only update the range address. So existing code
> is ok.

The only caller of follow_pte_pmd() does not call range_init() because it
doesn't know the address.  That's the point of follow_pte_pmd().

> I know this is kind of ugly but i do not see a way around that
> uglyness.

You wrote the code ...

Re: [PATCH] Initialise mmu_notifier_range correctly

[Matthew Wilcox]

On Thu, Jan 03, 2019 at 09:31:16AM -0500, Jerome Glisse wrote:
> On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
> > 
> > One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
> > incorrectly.
> > 
> > Signed-off-by: Matthew Wilcox <willy@infradead.org>
> > Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
> > Tested-by: Dave Chinner <dchinner@redhat.com>
> 
> Actually now that i have read the code again this is not ok to
> do so. The caller of follow_pte_pmd() will call range_init and
> follow pmd will only update the range address. So existing code
> is ok.

I think you need to re-read your own patch.

`git show ac46d4f3c43241ffa23d5bf36153a0830c0e02cc`

@@ -4058,10 +4059,10 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
                if (!pmdpp)
                        goto out;
 
-               if (start && end) {
-                       *start = address & PMD_MASK;
-                       *end = *start + PMD_SIZE;
-                       mmu_notifier_invalidate_range_start(mm, *start, *end);
+               if (range) {
+                       mmu_notifier_range_init(range, mm, address & PMD_MASK,
+                                            (address & PMD_MASK) + PMD_SIZE);
+                       mmu_notifier_invalidate_range_start(range);

... so it's fine to call range_init() *here*.

@@ -4069,17 +4070,17 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsign
ed long address,
[...]
        if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
                goto out;
 
-       if (start && end) {
-               *start = address & PAGE_MASK;
-               *end = *start + PAGE_SIZE;
-               mmu_notifier_invalidate_range_start(mm, *start, *end);
+       if (range) {
+               range->start = address & PAGE_MASK;
+               range->end = range->start + PAGE_SIZE;
+               mmu_notifier_invalidate_range_start(range);

... but then *not* here later in the same function?  You're not making
any sense.

Re: [PATCH] Initialise mmu_notifier_range correctly

[Jerome Glisse]

On Thu, Jan 03, 2019 at 06:43:13AM -0800, Matthew Wilcox wrote:
> On Thu, Jan 03, 2019 at 09:31:16AM -0500, Jerome Glisse wrote:
> > On Wed, Jan 02, 2019 at 04:21:26PM -0800, Matthew Wilcox wrote:
> > > 
> > > One of the paths in follow_pte_pmd() initialised the mmu_notifier_range
> > > incorrectly.
> > > 
> > > Signed-off-by: Matthew Wilcox <willy@infradead.org>
> > > Fixes: ac46d4f3c432 ("mm/mmu_notifier: use structure for invalidate_range_start/end calls v2")
> > > Tested-by: Dave Chinner <dchinner@redhat.com>
> > 
> > Actually now that i have read the code again this is not ok to
> > do so. The caller of follow_pte_pmd() will call range_init and
> > follow pmd will only update the range address. So existing code
> > is ok.
> 
> I think you need to re-read your own patch.
> 
> `git show ac46d4f3c43241ffa23d5bf36153a0830c0e02cc`
> 
> @@ -4058,10 +4059,10 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
>                 if (!pmdpp)
>                         goto out;
>  
> -               if (start && end) {
> -                       *start = address & PMD_MASK;
> -                       *end = *start + PMD_SIZE;
> -                       mmu_notifier_invalidate_range_start(mm, *start, *end);
> +               if (range) {
> +                       mmu_notifier_range_init(range, mm, address & PMD_MASK,
> +                                            (address & PMD_MASK) + PMD_SIZE);
> +                       mmu_notifier_invalidate_range_start(range);
> 
> ... so it's fine to call range_init() *here*.
> 
> @@ -4069,17 +4070,17 @@ static int __follow_pte_pmd(struct mm_struct *mm, unsign
> ed long address,
> [...]
>         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
>                 goto out;
>  
> -       if (start && end) {
> -               *start = address & PAGE_MASK;
> -               *end = *start + PAGE_SIZE;
> -               mmu_notifier_invalidate_range_start(mm, *start, *end);
> +       if (range) {
> +               range->start = address & PAGE_MASK;
> +               range->end = range->start + PAGE_SIZE;
> +               mmu_notifier_invalidate_range_start(range);
> 
> ... but then *not* here later in the same function?  You're not making
> any sense.

Ok i see that the patch that add the reason why mmu notifier is
call have been drop. So yes using range_init in follow_pte_pmd
is fine. With that other patch the reasons is set by the caller
of follow_pte_pmd and using range_init would have overwritten
it.

So this patch is fine for current tree. Sorry i was thinking with
the other patch included in mind.

Cheers,
Jérôme