Re: [PATCH 1/2] fcntl: add fcntl(F_CHECK_ORIGINAL_MEMFD)

[Christian Brauner]

On Fri, Sep 01, 2023 at 11:34:32AM -0700, Kees Cook wrote:
> On Fri, Sep 01, 2023 at 04:50:53PM +0200, Michał Cłapiński wrote:
> > On Fri, Sep 1, 2023 at 2:56 PM Christian Brauner <brauner@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Aug 31, 2023 at 10:36:46PM +0200, Michal Clapinski wrote:
> > > > Add a way to check if an fd points to the memfd's original open fd
> > > > (the one created by memfd_create).
> > > > Useful because only the original open fd can be both writable and
> > > > executable.
> > > >
> > > > Signed-off-by: Michal Clapinski <mclapinski@xxxxxxxxxx>
> > > > ---
> > > >  fs/fcntl.c                 | 3 +++
> > > >  include/uapi/linux/fcntl.h | 9 +++++++++
> > > >  2 files changed, 12 insertions(+)
> > > >
> > > > diff --git a/fs/fcntl.c b/fs/fcntl.c
> > > > index e871009f6c88..301527e07a4d 100644
> > > > --- a/fs/fcntl.c
> > > > +++ b/fs/fcntl.c
> > > > @@ -419,6 +419,9 @@ static long do_fcntl(int fd, unsigned int cmd, unsigned long arg,
> > > >       case F_SET_RW_HINT:
> > > >               err = fcntl_rw_hint(filp, cmd, arg);
> > > >               break;
> > > > +     case F_CHECK_ORIGINAL_MEMFD:
> > > > +             err = !(filp->f_mode & FMODE_WRITER);
> > > > +             break;
> > >
> > > Honestly, make this an ioctl on memfds. This is so specific that it
> > > really doesn't belong into fcntl().
> > 
> > I've never touched ioctls but if I'm correct, I can't just add it to
> > memfd. I would have to add it to the underlying fs, so hugetlbfs and
> > shmem (which I think can be defined as ramfs so also there). File
> > sealing fcntl is already memfd specific. Are you sure ioctl will be a
> > better idea?

fcntl() should be generic. Frankly, the sealing stuff should've gone
into an ioctl as well and only upgraded to a fcntl() once multiple fd
types support it.

> 
> Does this check "mean" anything for other files? Because if it's
> generically useful (and got renamed) it maybe would be right for
> fcntl...

For regular files it just means that the file has gotten write access to
the underlying fs and we use this flag to release the necessary
refcounts/protections once the file is closed.

If this check has any meaning beyond that than it only has meaning for
memfd. I'm also not sure why this checks FMODE_WRITER and not
FMODE_WRITE as FMODE_WRITER is almost an entirely internal thing that
only very specific codepaths need to know about.