Re: [PATCH] ovl: use linked upper dentry in copy-up tmpfile

On Wed, Jun 10, 2026 at 1:26 PM Miklos Szeredi <[email protected]> wrote:
>
> On Sun, 3 May 2026 at 22:37, Amir Goldstein <[email protected]> wrote:
> >
> > On Sat, May 2, 2026 at 1:27 AM Souvik Banerjee <[email protected]> wrote:
> > >
> > > ovl_copy_up_tmpfile() stores the disconnected O_TMPFILE dentry as the
> > > overlay's upper dentry reference via ovl_inode_update().  vfs_tmpfile()
> > > allocated this dentry via d_alloc(parentpath->dentry, &slash_name), so
> > > d_name is "/" and d_parent is c->workdir.  Local upper filesystems
> > > (ext4, btrfs, xfs, ...) immediately rename it to "#<inum>" via
> > > d_mark_tmpfile() inside their ->tmpfile() op; FUSE and virtiofs do
> > > not, so both fields stay that way.  Neither identifies the destination
> > > directory and filename where ovl_do_link() actually linked the file.
> > >
> > > When the upper filesystem implements ->d_revalidate() (e.g. FUSE or
> > > virtiofs), ovl_revalidate_real() calls it with the dentry's parent
> > > inode and a snapshot of d_name.  The server tries to look up "/" inside
> > > c->workdir, fails, and overlayfs reports -ESTALE.
> > >
> > > This causes persistent ESTALE errors for any file that was copied up via
> > > the tmpfile path, breaking dpkg, apt, and other tools that do
> > > rename-over-existing on overlayfs with a FUSE/virtiofs upper.
> > >
> > > Before commit 6b52243f633e ("ovl: fold copy-up helpers into callers"),
> > > the tmpfile copy-up path used a dedicated helper ovl_link_tmpfile()
> > > that captured the linked destination dentry returned by ovl_do_link():
> > >
> > >     err = ovl_do_link(temp, udir, upper);
> > >     ...
> > >     if (!err)
> > >         *newdentry = dget(upper);
> > >
> > > and published it via ovl_inode_update(d_inode(c->dentry), newdentry).
> > > The fold inlined ovl_do_link() into ovl_copy_up_tmpfile() but dropped
> > > the dget(upper) capture, and rewrote the publish line as
> > > ovl_inode_update(d_inode(c->dentry), dget(temp)) — where temp is the
> > > disconnected O_TMPFILE dentry.
> > >
> > > Fix by keeping a reference to the linked destination dentry after
> > > ovl_do_link() succeeds, and publishing that dentry at the existing
> > > ovl_inode_update() call site.  The non-tmpfile/workdir path continues to
> > > publish the renamed temporary dentry.
> > >
> > > Reproducer:
> > >   - Mount overlayfs with virtiofs (or a FUSE fs whose server advertises
> > >     FUSE_TMPFILE) as upper
> > >   - Run: dpkg -i <any .deb>
> > >   - Observe: "error installing new file '...': Stale file handle"
> > >
> > > Fixes: 6b52243f633e ("ovl: fold copy-up helpers into callers")
> > > Cc: [email protected] # v4.20+
> > > Signed-off-by: Souvik Banerjee <[email protected]>
> > > ---
> > >  fs/overlayfs/copy_up.c | 12 ++++++++++--
> > >  1 file changed, 10 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/fs/overlayfs/copy_up.c b/fs/overlayfs/copy_up.c
> > > index 13cb60b52bd6..e963701b4c87 100644
> > > --- a/fs/overlayfs/copy_up.c
> > > +++ b/fs/overlayfs/copy_up.c
> > > @@ -853,7 +853,7 @@ static int ovl_copy_up_tmpfile(struct ovl_copy_up_ctx *c)
> > >  {
> > >         struct ovl_fs *ofs = OVL_FS(c->dentry->d_sb);
> > >         struct inode *udir = d_inode(c->destdir);
> > > -       struct dentry *temp, *upper;
> > > +       struct dentry *temp, *upper, *newdentry = NULL;

This init is not needed and confusing because never in this function
using a NULL newdentry is correct.
We rather get an uninit variable warning if that happens in the future.

> > >         struct file *tmpfile;
> > >         int err;
> > >
> > > @@ -889,6 +889,14 @@ static int ovl_copy_up_tmpfile(struct ovl_copy_up_ctx *c)
> > >         err = PTR_ERR(upper);
> > >         if (!IS_ERR(upper)) {
> > >                 err = ovl_do_link(ofs, temp, udir, upper);
> > > +               if (!err) {
> > > +                       /*
> > > +                        * Record the linked dentry -- not the disconnected
> > > +                        * O_TMPFILE dentry -- so that ->d_revalidate() on
> > > +                        * the upper fs sees the real parent/name.
> > > +                        */
> > > +                       newdentry = dget(upper);
> > > +               }
> > >                 end_creating(upper);
> > >         }
> > >
> > > @@ -903,7 +911,7 @@ static int ovl_copy_up_tmpfile(struct ovl_copy_up_ctx *c)
> > >
> > >         if (!c->metacopy)
> > >                 ovl_set_upperdata(d_inode(c->dentry));
> > > -       ovl_inode_update(d_inode(c->dentry), dget(temp));
> > > +       ovl_inode_update(d_inode(c->dentry), newdentry);
> > >
> > >  out:
> > >         ovl_end_write(c->dentry);
> > > --
> > > 2.51.1
> > >
> >
> >
> > Hi Souvik,
> >
> > Thank you for the analysis and the fix.
> > Looks correct to me.
> >
> > Reviewed-by: Amir Goldstein <[email protected]>
> >
> > Christian,
> >
> > Could you pick this up for vfs-fixes?
> > I do not have any other ovl fixes queued up.
>
> Reviewed-by: Miklos Szeredi <[email protected]>
>

I applied the patch without this init to NULL to ovl-fixes, so that it
will be in linux-next.
Kept your RVB. Hope that's ok.

> This seems to have slipped through the cracks.
>
> Christian?

I don't think it is particularly urgent to merge this patch for 7.0
this late in the cycle, so I will hold off on sending an ovl-fixes PR.

Christian,

If you take this ovl fix (and the other one [1]) to vfs-7.2.misc
let me know and I will remove them from ovl-fixes.

Thanks,
Amir.

[1] https://lore.kernel.org/linux-unionfs/[email protected]/