[PATCH] fix mmap return value when vma is merged after call_mmap()

[PATCH] fix mmap return value when vma is merged after call_mmap()

[Liu Zixian]

On success, mmap should return the begin address of newly mapped area,
but patch "mm: mmap: merge vma after call_mmap() if possible" 
set vm_start of newly merged vma to return value addr.
Users of mmap will get wrong address if vma is merged after call_mmap().
We fix this by moving the assignment to addr before merging vma.

Fixes: d70cec898324 ("mm: mmap: merge vma after call_mmap() if possible")

Signed-off-by: Liu Zixian <liuzixian4@huawei.com>
---
 mm/mmap.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index d91ecb00d38c..9199b5e8cc1e 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -1820,12 +1820,12 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
 				 * and cause general protection fault ultimately.
 				 */
 				fput(vma->vm_file);
-				vm_area_free(vma);
-				vma = merge;
 				/* Update vm_flags and possible addr to pick up the change. We don't
 				 * warn here if addr changed as the vma is not linked by vma_link().
 				 */
 				addr = vma->vm_start;
+				vm_area_free(vma);
+				vma = merge;
 				vm_flags = vma->vm_flags;
 				goto unmap_writable;
 			}
-- 
2.23.0

Re: [PATCH] fix mmap return value when vma is merged after call_mmap()

[David Hildenbrand]

On 01.12.20 14:27, Liu Zixian wrote:
> On success, mmap should return the begin address of newly mapped area,
> but patch "mm: mmap: merge vma after call_mmap() if possible" 
> set vm_start of newly merged vma to return value addr.
> Users of mmap will get wrong address if vma is merged after call_mmap().
> We fix this by moving the assignment to addr before merging vma.
> 
> Fixes: d70cec898324 ("mm: mmap: merge vma after call_mmap() if possible")
> 
> Signed-off-by: Liu Zixian <liuzixian4@huawei.com>
> ---
>  mm/mmap.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index d91ecb00d38c..9199b5e8cc1e 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -1820,12 +1820,12 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>  				 * and cause general protection fault ultimately.
>  				 */
>  				fput(vma->vm_file);
> -				vm_area_free(vma);
> -				vma = merge;
>  				/* Update vm_flags and possible addr to pick up the change. We don't
>  				 * warn here if addr changed as the vma is not linked by vma_link().
>  				 */
>  				addr = vma->vm_start;
> +				vm_area_free(vma);
> +				vma = merge;
>  				vm_flags = vma->vm_flags;
>  				goto unmap_writable;
>  			}
> 

I assume this is quite hard to trigger, right (having vm_flags change)?

"The vm_flags may be changed after call_mmap() because drivers may set
some flags for their own purpose."

Because how you describe it (returning wrong mmap address) this will
completely mess up userspace.

-- 
Thanks,

David / dhildenb

Re: [PATCH] fix mmap return value when vma is merged after call_mmap()

[Jason Gunthorpe]

On Tue, Dec 01, 2020 at 09:27:55PM +0800, Liu Zixian wrote:
> On success, mmap should return the begin address of newly mapped area,
> but patch "mm: mmap: merge vma after call_mmap() if possible" 
> set vm_start of newly merged vma to return value addr.
> Users of mmap will get wrong address if vma is merged after call_mmap().
> We fix this by moving the assignment to addr before merging vma.
> 
> Fixes: d70cec898324 ("mm: mmap: merge vma after call_mmap() if possible")
> 
> Signed-off-by: Liu Zixian <liuzixian4@huawei.com>

No blank line after Fixes:

Consider a Cc: stable since this is quite a significant bug, if you
hit it.

>  mm/mmap.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index d91ecb00d38c..9199b5e8cc1e 100644
> +++ b/mm/mmap.c
> @@ -1820,12 +1820,12 @@ unsigned long mmap_region(struct file *file, unsigned long addr,
>  				 * and cause general protection fault ultimately.
>  				 */
>  				fput(vma->vm_file);
> -				vm_area_free(vma);
> -				vma = merge;
>  				/* Update vm_flags and possible addr to pick up the change. We don't

This comment is now out of date, we don't want to pick up the
'possible addr' change.

>  				 * warn here if addr changed as the vma is not linked by vma_link().
>  				 */
>  				addr = vma->vm_start;
> +				vm_area_free(vma);
> +				vma = merge;
>  				vm_flags = vma->vm_flags;
>  				goto unmap_writable;
>  			}

David's remarks are correct, how was the original patch tested? Did it
always prepend during merge so the new vm_start is OK?

The previous patch really added more spaghetti to this function,
jumping into the label 'unmap_writable' inside an if is aweful.

At least for this fix it would be cleaner if the entire if(unlikely)
block was moved below here:

		WARN_ON_ONCE(addr != vma->vm_start);

		addr = vma->vm_start;

Since we want to do that unconditionally after call_mmap.

Jason

Re: Re: [PATCH] fix mmap return value when vma is merged after call_mmap()

[Liu Zixian]

Thanks for your review, I've sent a v2 patch.