Re: [PATCH] mm: allow unmapped hole at head side of mbind range

Re: [PATCH] mm: allow unmapped hole at head side of mbind range

[Vlastimil Babka]

+ linux-api

On 10/24/19 9:35 AM, Li Xinhai wrote:
> From: Li Xinhai  <xinhai.li@outlook.com>
> 
> mbind_range silently ignore unmapped hole at middle and tail of the 
> specified range, but report EFAULT if hole at head side.


Hmm that's unfortunate. mbind() manpage says:

EFAULT Part or all of the memory range specified by nodemask and maxnode
points outside your accessible address space.  Or, there was an unmapped
hole in  the  specified  memory range specified by addr and len.

That sounds like any hole inside the specified range should return
EFAULT. But perhaps it can be also interpreted as you suggest, that the
whole range is an unmapped hole. There's some risk of breaking existing
userspace if we change it either way.

> It is more reasonable to support silently ignore holes at any part of 
> the range, only report EFAULT if the whole range is in hole.
> 
> Signed-off-by: Li Xinhai <xinhai.li@outlook.com>
> ---
> 
>  mm/mempolicy.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> 
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> index 4ae967bcf954..ae160d9936d9 100644
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -738,7 +738,7 @@ static int mbind_range(struct mm_struct *mm, unsigned long start,
>         unsigned long vmend;
>  
>         vma = find_vma(mm, start);
> -       if (!vma || vma->vm_start > start)
> +       if (!vma || vma->vm_start >= end)
>                 return -EFAULT;
>  
>         prev = vma->vm_prev;
> 

Re: [PATCH] mm: allow unmapped hole at head side of mbind range

[Hugh Dickins]

On Thu, 24 Oct 2019, Vlastimil Babka wrote:

> + linux-api
> 
> On 10/24/19 9:35 AM, Li Xinhai wrote:
> > From: Li Xinhai  <xinhai.li@outlook.com>
> > 
> > mbind_range silently ignore unmapped hole at middle and tail of the 
> > specified range, but report EFAULT if hole at head side.
> 
> 
> Hmm that's unfortunate. mbind() manpage says:
> 
> EFAULT Part or all of the memory range specified by nodemask and maxnode
> points outside your accessible address space.  Or, there was an unmapped
> hole in  the  specified  memory range specified by addr and len.
> 
> That sounds like any hole inside the specified range should return
> EFAULT.

Yes (though an exception is allowed when restoring to default).

> But perhaps it can be also interpreted as you suggest, that the
> whole range is an unmapped hole.  There's some risk of breaking existing
> userspace if we change it either way.
> 
> > It is more reasonable to support silently ignore holes at any part of 
> > the range, only report EFAULT if the whole range is in hole.
> > 
> > Signed-off-by: Li Xinhai <xinhai.li@outlook.com>

Xinhai, I'm sceptical about this patch: is it something you found
by code inspection, or something you found when using mbind()?

I've not looked long enough to be certain, nor experimented, but:

mbind_range() is only one stage of the mbind() syscall implementation,
and is preceded by queue_pages_range(): look what queue_pages_test_walk()
does when MPOL_MF_DISCONTIG_OK not set.

My impression is that mbind_range() is merely correcting an omission
from the checks already made my queue_pages_test_walk() (an odd way
to proceed, I admit: would be better to check initially than later).

I do think that you should not make this change without considering
MPOL_MF_DISCONTIG_OK and its intention.

Hugh

> > ---
> > 
> >  mm/mempolicy.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > 
> > diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> > index 4ae967bcf954..ae160d9936d9 100644
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -738,7 +738,7 @@ static int mbind_range(struct mm_struct *mm, unsigned long start,
> >         unsigned long vmend;
> >  
> >         vma = find_vma(mm, start);
> > -       if (!vma || vma->vm_start > start)
> > +       if (!vma || vma->vm_start >= end)
> >                 return -EFAULT;
> >  
> >         prev = vma->vm_prev;
> > 

Re: [PATCH] mm: allow unmapped hole at head side of mbind range

[Michal Hocko]

On Thu 24-10-19 07:35:06, Li Xinhai wrote:
> From: Li Xinhai  <xinhai.li@outlook.com>
> 
> mbind_range silently ignore unmapped hole at middle and tail of the 
> specified range, but report EFAULT if hole at head side.
> It is more reasonable to support silently ignore holes at any part of 
> the range, only report EFAULT if the whole range is in hole.

The changelog is a bit cryptic but you are right. find_vma returns the
first vma that ends above the given address. If vm_start > start
then there is still an overlap possible
              [    vma      ]
    [start        end]
and we should mbind [vma->vm_start, end] at least. I haven't checked
whether changing the condition is sufficient for the rest of the code to
work properly. I am pretty sure a test case shouldn't be really hard
to construct and add to the kernel testing machinery.

Btw. when writing a changelog then it is always preferred to describe
user visible effect of the patch. In this case it would be an unexpected
EFAULT on a range that starts before an existing VMA while still
overlapping it. Make sure to note that.

Fixes: 9d8cebd4bcd7 ("mm: fix mbind vma merge problem")
> Signed-off-by: Li Xinhai <xinhai.li@outlook.com>
> ---
> 
>  mm/mempolicy.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> 
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> index 4ae967bcf954..ae160d9936d9 100644
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -738,7 +738,7 @@ static int mbind_range(struct mm_struct *mm, unsigned long start,
>         unsigned long vmend;
>  
>         vma = find_vma(mm, start);
> -       if (!vma || vma->vm_start > start)
> +       if (!vma || vma->vm_start >= end)
>                 return -EFAULT;
>  
>         prev = vma->vm_prev;
> 

-- 
Michal Hocko
SUSE Labs