filemap: Fix bounds checking in filemap_read()

Commit Message

Trond Myklebust Sept. 13, 2024, 5:57 p.m. UTC

From: Trond Myklebust <[e-mail]>

If the caller supplies an iocb->ki_pos value that is close to the
filesystem upper limit, and an iterator with a count that causes us to
overflow that limit, then filemap_read() enters an infinite loop.

This behaviour was discovered when testing xfstests generic/525 with the
"localio" optimisation for loopback NFS mounts.

Reported-by: Mike Snitzer <[e-mail]>
Fixes: c2a9737f45e2 ("vfs,mm: fix a dead loop in truncate_inode_pages_range()")
Tested-by: Mike Snitzer <[e-mail]>
Signed-off-by: Trond Myklebust <[e-mail]>
---
 mm/filemap.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Matthew Wilcox Sept. 13, 2024, 6:21 p.m. UTC | #1

On Fri, Sep 13, 2024 at 01:57:04PM -0400, [e-mail] wrote:
> If the caller supplies an iocb->ki_pos value that is close to the
> filesystem upper limit, and an iterator with a count that causes us to
> overflow that limit, then filemap_read() enters an infinite loop.

Are we guaranteed that ki_pos lies in the range [0..s_maxbytes)?
I'm not too familiar with the upper paths of the VFS and what guarantees
we can depend on.  If we are guaranteed that, could somebody document
it (and indeed create kernel-doc for struct kiocb)?

>  
> -	iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
> +	iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos);
>  	folio_batch_init(&fbatch);
>  
>  	do {
> -- 
> 2.46.0
>

Dave Chinner Sept. 16, 2024, 9:55 p.m. UTC | #2

On Fri, Sep 13, 2024 at 07:21:54PM +0100, Matthew Wilcox wrote:
> On Fri, Sep 13, 2024 at 01:57:04PM -0400, [e-mail] wrote:
> > If the caller supplies an iocb->ki_pos value that is close to the
> > filesystem upper limit, and an iterator with a count that causes us to
> > overflow that limit, then filemap_read() enters an infinite loop.
> 
> Are we guaranteed that ki_pos lies in the range [0..s_maxbytes)?
> I'm not too familiar with the upper paths of the VFS and what guarantees
> we can depend on.  If we are guaranteed that, could somebody document
> it (and indeed create kernel-doc for struct kiocb)?

filemap_read() checks this itself before doing anything else:

	if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
                return 0;

i.e. there is no guarantee provided by the upper layers, it's first
checked right here in any buffered read path...

-Dave.

Mike Snitzer Sept. 26, 2024, 3:01 p.m. UTC | #3

Hi Willy,

On Tue, Sep 17, 2024 at 07:55:29AM +1000, Dave Chinner wrote:
> On Fri, Sep 13, 2024 at 07:21:54PM +0100, Matthew Wilcox wrote:
> > On Fri, Sep 13, 2024 at 01:57:04PM -0400, [e-mail] wrote:
> > > If the caller supplies an iocb->ki_pos value that is close to the
> > > filesystem upper limit, and an iterator with a count that causes us to
> > > overflow that limit, then filemap_read() enters an infinite loop.
> > 
> > Are we guaranteed that ki_pos lies in the range [0..s_maxbytes)?
> > I'm not too familiar with the upper paths of the VFS and what guarantees
> > we can depend on.  If we are guaranteed that, could somebody document
> > it (and indeed create kernel-doc for struct kiocb)?
> 
> filemap_read() checks this itself before doing anything else:
> 
> 	if (unlikely(iocb->ki_pos >= inode->i_sb->s_maxbytes))
>                 return 0;
> 
> i.e. there is no guarantee provided by the upper layers, it's first
> checked right here in any buffered read path...
> 
> -Dave.
> 
> -- 
> Dave Chinner
> [e-mail]

Linus merged the NFS LOCALIO changes via the NFS client tree a couple
days ago.  LOCALIO teased out this filemap_read infinite loop bug, so
it is important to fix this for 6.12 (probably should get marked for
stable@ too):
https://lore.kernel.org/all/c6f35a8[e-mail]/

(weirdly, Trond's reply to you didn't make it to the linux-nfs or
linux-fsdevel list archives, but Dave's above reply covers the same)

Trond also offered this additional filemap_read negative check:
https://lore.kernel.org/all/482ee0b[e-mail]/

Could be you've been busy with travel or whatever, but for future
reference, should linux-mm and/or Andrew always be cc'd on filemap
fixes?

Thanks,
Mike

Patch

diff --git a/mm/filemap.c b/mm/filemap.c
index d62150418b91..c69227ccdabb 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2605,7 +2605,7 @@  ssize_t filemap_read(struct kiocb *iocb, struct iov_iter *iter,
 	if (unlikely(!iov_iter_count(iter)))
 		return 0;
 
-	iov_iter_truncate(iter, inode->i_sb->s_maxbytes);
+	iov_iter_truncate(iter, inode->i_sb->s_maxbytes - iocb->ki_pos);
 	folio_batch_init(&fbatch);
 
 	do {