Re: [PATCH v15 3/9] namei: LOOKUP_NO_XDEV: block mountpoint crossing

[Aleksa Sarai <cyphar at cyphar dot com>]

On 2019-11-14, Al Viro <viro@zeniv.linux.org.uk> wrote:
> On Thu, Nov 14, 2019 at 03:49:45PM +1100, Aleksa Sarai wrote:
> > On 2019-11-13, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > > On Tue, Nov 05, 2019 at 08:05:47PM +1100, Aleksa Sarai wrote:
> > > 
> > > > @@ -862,6 +870,8 @@ static int nd_jump_root(struct nameidata *nd)
> > > >  void nd_jump_link(struct path *path)
> > > >  {
> > > >  	struct nameidata *nd = current->nameidata;
> > > > +
> > > > +	nd->last_magiclink.same_mnt = (nd->path.mnt == path->mnt);
> > > >  	path_put(&nd->path);
> > > >  
> > > >  	nd->path = *path;
> > > > @@ -1082,6 +1092,10 @@ const char *get_link(struct nameidata *nd)
> > > >  		if (nd->flags & LOOKUP_MAGICLINK_JUMPED) {
> > > >  			if (unlikely(nd->flags & LOOKUP_NO_MAGICLINKS))
> > > >  				return ERR_PTR(-ELOOP);
> > > > +			if (unlikely(nd->flags & LOOKUP_NO_XDEV)) {
> > > > +				if (!nd->last_magiclink.same_mnt)
> > > > +					return ERR_PTR(-EXDEV);
> > > > +			}
> > > >  		}
> > > 
> > > Ugh...  Wouldn't it be better to take that logics (some equivalent thereof)
> > > into nd_jump_link()?  Or just have nd_jump_link() return an error...
> > 
> > This could be done, but the reason for stashing it away in
> > last_magiclink is because of the future magic-link re-opening patches
> > which can't be implemented like that without putting the open_flags
> > inside nameidata (which was decided to be too ugly a while ago).
> > 
> > My point being that I could implement it this way for this series, but
> > I'd have to implement something like last_magiclink when I end up
> > re-posting the magic-link stuff in a few weeks.
> > 
> > Looking at all the nd_jump_link() users, the other option is to just
> > disallow magic-link crossings entirely for LOOKUP_NO_XDEV. The only
> > thing allowing them permits is to resolve file descriptors that are
> > pointing to the same procfs mount -- and it's unclear to me how useful
> > that really is (apparmorfs and nsfs will always give -EXDEV because
> > aafs_mnt and nsfs_mnt are internal kernel vfsmounts).
> 
> I would rather keep the entire if (nd->flags & LOOKUP_MAGICLINK_JUMPED)
> out of the get_link().  If you want to generate some error if
> nd_jump_link() has been called, just do it right there.  The fewer
> pieces of state need to be carried around, the better...

Sure, I can make nd_jump_link() give -ELOOP and drop the current need
for LOOKUP_MAGICLINK_JUMPED -- if necessary we can re-add it for the
magic-link reopening patches.

> And as for opening them...  Why would you need full open_flags in there?
> Details, please...

I was referring to [1] which has been dropped from this series. I
misspoke -- you don't need the full open_flags, you just need acc_mode
in nameidata -- but from memory you (understandably) weren't in favour
of that either because it further muddled the open semantics with namei.

So the solution I went with was to stash away the i_mode of the
magiclink in nd->last_magiclink.mode (though to avoid a race which Jann
found, you actually need to recalculate it when you call nd_jump_link()
but that's a different topic) and then check it in trailing_magiclink().

However, I've since figured out that we need to restrict things like
bind-mounts and truncate() because they can be used to get around the
restrictions. I dropped that patch from this series so that I could work
on implementing the restrictions for the other relevant VFS syscalls
separately from openat2 (upgrade_mask will be re-added to open_how with
those patches).

My point was that AFAICS we will either have to have nd->acc_mode (or
something similar) or have nd->last_magiclink in order to implement the
magic-link reopening hardening.

[1]: https://lore.kernel.org/lkml/20190930183316.10190-2-cyphar@cyphar.com/

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

Attachment: signature.asc
Description: PGP signature