set_page_dirty vs truncate

[Matthew Wilcox]

On Fri, Dec 18, 2020 at 10:03:16PM +0000, Matthew Wilcox wrote:
> On Fri, Dec 18, 2020 at 04:05:31PM +0000, Matthew Wilcox wrote:
> > A number of implementations of ->set_page_dirty check whether the page
> > has been truncated (ie page->mapping has become NULL since entering
> > set_page_dirty()).  Several other implementations assume that they can do
> > page->mapping->host to get to the inode.  So either some implementations
> > are doing unnecessary checks or others are vulnerable to a NULL pointer
> > dereference if truncate() races with set_page_dirty().
> > 
> > I'm touching ->set_page_dirty() anyway as part of the page folio
> > conversion.  I'm thinking about passing in the mapping so there's no
> > need to look at page->mapping.
> > 
> > The comments on set_page_dirty() and set_page_dirty_lock() suggests
> > there's no consistency in whether truncation is blocked or not; we're
> > only guaranteed that the inode itself won't go away.  But maybe the
> > comments are stale.
> 
> The comments are, I believe, not stale.  Here's some syzbot
> reports which indicate that ext4 is seeing races between set_page_dirty()
> and truncate():
> 
>  https://groups.google.com/g/syzkaller-lts-bugs/c/s9fHu162zhQ/m/Phnf6ucaAwAJ
> 
> The reproducer includes calls to ftruncate(), so that would suggest
> that's what's going on.

Hmmm ... looks like __set_page_dirty_nobuffers() has a similar problem:

{
        lock_page_memcg(page);
        if (!TestSetPageDirty(page)) {
                struct address_space *mapping = page_mapping(page);
                unsigned long flags;

                if (!mapping) {
                        unlock_page_memcg(page);
                        return 1;
                }

                xa_lock_irqsave(&mapping->i_pages, flags);
                BUG_ON(page_mapping(page) != mapping);

sure, we check that the page wasn't truncated between set_page_dirty()
and the call to TestSetPageDirty(), but we can truncate dirty pages
with no problem.  So between the call to TestSetPageDirty() and
the call to xa_lock_irqsave(), the page can be truncated, and the
BUG_ON should fire.

I haven't been able to find any examples of this, but maybe it's just a very
narrow race.  Does anyone recognise this signature?  Adding the filesystems
which use __set_page_dirty_nobuffers() directly without extra locking.

$ git grep set_page_dirty.*=.*__set_page_dirty_nobuffers
fs/9p/vfs_addr.c:       .set_page_dirty = __set_page_dirty_nobuffers,
fs/cifs/file.c: .set_page_dirty = __set_page_dirty_nobuffers,
fs/cifs/file.c: .set_page_dirty = __set_page_dirty_nobuffers,
fs/fuse/file.c: .set_page_dirty = __set_page_dirty_nobuffers,
fs/hostfs/hostfs_kern.c:        .set_page_dirty = __set_page_dirty_nobuffers,
fs/jfs/jfs_metapage.c:  .set_page_dirty = __set_page_dirty_nobuffers,
fs/nfs/file.c:  .set_page_dirty = __set_page_dirty_nobuffers,
fs/ntfs/aops.c: .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
fs/orangefs/inode.c:    .set_page_dirty = __set_page_dirty_nobuffers,
fs/vboxsf/file.c:       .set_page_dirty = __set_page_dirty_nobuffers,