[PATCH] shmem: Update folio if shmem_replace_page() updates the page

[PATCH] shmem: Update folio if shmem_replace_page() updates the page

[Matthew Wilcox (Oracle)]

If we allocate a new page, we need to make sure that our folio matches
that new page.  This will be solved by changing shmem_replace_page()
to shmem_replace_folio(), but this is the minimal fix.

Fixes: da08e9b79323 ("mm/shmem: convert shmem_swapin_page() to shmem_swapin_folio()")
Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
---
 mm/shmem.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mm/shmem.c b/mm/shmem.c
index a6f565308133..bcc0a3c7b5bf 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
 
 	if (shmem_should_replace_folio(folio, gfp)) {
 		error = shmem_replace_page(&page, gfp, info, index);
+		folio = page_folio(page);
 		if (error)
 			goto failed;
 	}
-- 
2.35.1

Re: [PATCH] shmem: Update folio if shmem_replace_page() updates the page

[Andrew Morton]

On Sat, 30 Jul 2022 05:25:18 +0100 "Matthew Wilcox (Oracle)" <willy@infradead.org> wrote:

> If we allocate a new page, we need to make sure that our folio matches
> that new page.  This will be solved by changing shmem_replace_page()
> to shmem_replace_folio(), but this is the minimal fix.
> 
> ...
>
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>  
>  	if (shmem_should_replace_folio(folio, gfp)) {
>  		error = shmem_replace_page(&page, gfp, info, index);
> +		folio = page_folio(page);
>  		if (error)
>  			goto failed;
>  	}

What are the user-visible runtime effects of the bug?

Should we backport this into 5.19.X?

Re: [PATCH] shmem: Update folio if shmem_replace_page() updates the page

[Matthew Wilcox]

On Tue, Aug 02, 2022 at 05:46:37PM -0700, Andrew Morton wrote:
> On Sat, 30 Jul 2022 05:25:18 +0100 "Matthew Wilcox (Oracle)" <willy@infradead.org> wrote:
> 
> > If we allocate a new page, we need to make sure that our folio matches
> > that new page.  This will be solved by changing shmem_replace_page()
> > to shmem_replace_folio(), but this is the minimal fix.
> > 
> > ...
> >
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
> >  
> >  	if (shmem_should_replace_folio(folio, gfp)) {
> >  		error = shmem_replace_page(&page, gfp, info, index);
> > +		folio = page_folio(page);
> >  		if (error)
> >  			goto failed;
> >  	}
> 
> What are the user-visible runtime effects of the bug?
> 
> Should we backport this into 5.19.X?

Definitely should be backported.  The next line not visible in this
patch context says:

        error = shmem_add_to_page_cache(folio, mapping, index,
                                        swp_to_radix_entry(swap), gfp,
                                        charge_mm);

so if we do end up in this path, we store the wrong page in the
shmem inode's page cache, and I would rather imagine that data
corruption ensues.

Re: [PATCH] shmem: Update folio if shmem_replace_page() updates the page

[William Kucharski]

Looks good.

Reviewed-by: William Kucharski <william.kucharski@oracle.com>

> On Aug 2, 2022, at 21:18, Matthew Wilcox <willy@infradead.org> wrote:
> 
> On Tue, Aug 02, 2022 at 05:46:37PM -0700, Andrew Morton wrote:
>>> On Sat, 30 Jul 2022 05:25:18 +0100 "Matthew Wilcox (Oracle)" <willy@infradead.org> wrote:
>>> 
>>> If we allocate a new page, we need to make sure that our folio matches
>>> that new page.  This will be solved by changing shmem_replace_page()
>>> to shmem_replace_folio(), but this is the minimal fix.
>>> 
>>> ...
>>> 
>>> --- a/mm/shmem.c
>>> +++ b/mm/shmem.c
>>> @@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>>> 
>>>    if (shmem_should_replace_folio(folio, gfp)) {
>>>        error = shmem_replace_page(&page, gfp, info, index);
>>> +        folio = page_folio(page);
>>>        if (error)
>>>            goto failed;
>>>    }
>> 
>> What are the user-visible runtime effects of the bug?
>> 
>> Should we backport this into 5.19.X?
> 
> Definitely should be backported.  The next line not visible in this
> patch context says:
> 
>        error = shmem_add_to_page_cache(folio, mapping, index,
>                                        swp_to_radix_entry(swap), gfp,
>                                        charge_mm);
> 
> so if we do end up in this path, we store the wrong page in the
> shmem inode's page cache, and I would rather imagine that data
> corruption ensues.
> 

Re: [PATCH] shmem: Update folio if shmem_replace_page() updates the page

[William Kucharski]

Looks good.

Reviewed-by: William Kucharski <william.kucharski@oracle.com>


> On Jul 29, 2022, at 10:25 PM, Matthew Wilcox (Oracle) <willy@infradead.org> wrote:
> 
> If we allocate a new page, we need to make sure that our folio matches
> that new page.  This will be solved by changing shmem_replace_page()
> to shmem_replace_folio(), but this is the minimal fix.
> 
> Fixes: da08e9b79323 ("mm/shmem: convert shmem_swapin_page() to shmem_swapin_folio()")
> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>
> ---
> mm/shmem.c | 1 +
> 1 file changed, 1 insertion(+)
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index a6f565308133..bcc0a3c7b5bf 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
> 
> if (shmem_should_replace_folio(folio, gfp)) {
> error = shmem_replace_page(&page, gfp, info, index);
> + folio = page_folio(page);
> if (error)
> goto failed;
> }
> -- 
> 2.35.1

Re: [PATCH] shmem: Update folio if shmem_replace_page() updates the page

[Hugh Dickins]

On Sat, 30 Jul 2022, Matthew Wilcox (Oracle) wrote:

> If we allocate a new page, we need to make sure that our folio matches
> that new page.  This will be solved by changing shmem_replace_page()
> to shmem_replace_folio(), but this is the minimal fix.
> 
> Fixes: da08e9b79323 ("mm/shmem: convert shmem_swapin_page() to shmem_swapin_folio()")
> Signed-off-by: Matthew Wilcox (Oracle) <willy@infradead.org>

Acked-by: Hugh Dickins <hughd@google.com>

I hit this myself just once, at about the very time you sent the fix.
But, thinking that shmem_replace_page() was special for gma500, couldn't
understand how I (or most people) would ever get there.  Turns out that
nowadays tmpfs symlinks longer than 128 can come this way on 32-bit (I
had been testing kmap_local stuff for other reasons).

And today I see that Zdenek hit it on 5.19-rc back in June:
https://lore.kernel.org/lkml/584ae788-05e3-5824-8c85-cbb833677850@redhat.com/
so this patch is definitely one for -stable.

Hugh

> ---
>  mm/shmem.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index a6f565308133..bcc0a3c7b5bf 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -1771,6 +1771,7 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>  
>  	if (shmem_should_replace_folio(folio, gfp)) {
>  		error = shmem_replace_page(&page, gfp, info, index);
> +		folio = page_folio(page);
>  		if (error)
>  			goto failed;
>  	}
> -- 
> 2.35.1