On 2019-09-18, Aleksa Sarai <cyphar@cyphar.com> wrote:
> On 2019-09-17, Jann Horn <jannh@google.com> wrote:
> > On Wed, Sep 4, 2019 at 10:21 PM Aleksa Sarai <cyphar@cyphar.com> wrote:
> > > The ability for userspace to "re-open" file descriptors through
> > > /proc/self/fd has been a very useful tool for all sorts of usecases
> > > (container runtimes are one common example). However, the current
> > > interface for doing this has resulted in some pretty subtle security
> > > holes. Userspace can re-open a file descriptor with more permissions
> > > than the original, which can result in cases such as /proc/$pid/exe
> > > being re-opened O_RDWR at a later date even though (by definition)
> > > /proc/$pid/exe cannot be opened for writing. When combined with O_PATH
> > > the results can get even more confusing.
> > [...]
> > > Instead we have to restrict it in such a way that it doesn't break
> > > (good) users but does block potential attackers. The solution applied in
> > > this patch is to restrict *re-opening* (not resolution through)
> > > magic-links by requiring that mode of the link be obeyed. Normal
> > > symlinks have modes of a+rwx but magic-links have other modes. These
> > > magic-link modes were historically ignored during path resolution, but
> > > they've now been re-purposed for more useful ends.
> > 
> > Thanks for dealing with this issue!
> > 
> > [...]
> > > diff --git a/fs/namei.c b/fs/namei.c
> > > index 209c51a5226c..54d57dad0f91 100644
> > > --- a/fs/namei.c
> > > +++ b/fs/namei.c
> > > @@ -872,7 +872,7 @@ void nd_jump_link(struct path *path)
> > >
> > >         nd->path = *path;
> > >         nd->inode = nd->path.dentry->d_inode;
> > > -       nd->flags |= LOOKUP_JUMPED;
> > > +       nd->flags |= LOOKUP_JUMPED | LOOKUP_MAGICLINK_JUMPED;
> > >  }
> > [...]
> > > +static int trailing_magiclink(struct nameidata *nd, int acc_mode,
> > > +                             fmode_t *opath_mask)
> > > +{
> > > +       struct inode *inode = nd->link_inode;
> > > +       fmode_t upgrade_mask = 0;
> > > +
> > > +       /* Was the trailing_symlink() a magic-link? */
> > > +       if (!(nd->flags & LOOKUP_MAGICLINK_JUMPED))
> > > +               return 0;
> > > +
> > > +       /*
> > > +        * Figure out the upgrade-mask of the link_inode. Since these aren't
> > > +        * strictly POSIX semantics we don't do an acl_permission_check() here,
> > > +        * so we only care that at least one bit is set for each upgrade-mode.
> > > +        */
> > > +       if (inode->i_mode & S_IRUGO)
> > > +               upgrade_mask |= FMODE_PATH_READ;
> > > +       if (inode->i_mode & S_IWUGO)
> > > +               upgrade_mask |= FMODE_PATH_WRITE;
> > > +       /* Restrict the O_PATH upgrade-mask of the caller. */
> > > +       if (opath_mask)
> > > +               *opath_mask &= upgrade_mask;
> > > +       return may_open_magiclink(upgrade_mask, acc_mode);
> > >  }
> > 
> > This looks racy because entries in the file descriptor table can be
> > switched out as long as task->files->file_lock isn't held. Unless I'm
> > missing something, something like the following (untested) would
> > bypass this restriction:
> 
> You're absolutely right -- good catch!
> 
> > Perhaps you could change nd_jump_link() to "void nd_jump_link(struct
> > path *path, umode_t link_mode)", and let proc_pid_get_link() pass the
> > link_mode through from an out-argument of .proc_get_link()? Then
> > proc_fd_link() could grab the proper mode in a race-free manner. And
> > nd_jump_link() could stash the mode in the nameidata.
> 
> This indeed does appear to be the simplest solution -- I'm currently
> testing a variation of the patch you proposed (with a few extra bits to
> deal with nd_jump_link and proc_get_link being used elsewhere).
> 
> I'll include this change (assuming it fixes the flaw you found) in the
> v13 series I'll send around next week. Thanks, Jann!

In case you're interested -- I've also included a selftest based on this
attack in my series (though it uses CLONE_FILES so that we could also
test O_EMPTYPATH, which wasn't affected because it didn't go through
procfs and thus couldn't hit the "outdated inode->i_mode" problem).

The attack script succeeds around 20% of the time on the original
patchset, and with the updated patchset it doesn't succeed in several
hundred thousand attempts (which I've repeated a few times).

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>