Re: [PATCH 01/14] fs_parse: add uid & gid option option parsing helpers

[Christian Brauner <brauner@kernel.org>]

On Fri, Jun 28, 2024 at 11:45:17AM GMT, Jan Kara wrote:
> On Thu 27-06-24 19:26:24, Eric Sandeen wrote:
> > Multiple filesystems take uid and gid as options, and the code to
> > create the ID from an integer and validate it is standard boilerplate
> > that can be moved into common helper functions, so do that for
> > consistency and less cut&paste.
> > 
> > This also helps avoid the buggy pattern noted by Seth Jenkins at
> > https://lore.kernel.org/lkml/CALxfFW4BXhEwxR0Q5LSkg-8Vb4r2MONKCcUCVioehXQKr35eHg@mail.gmail.com/
> > because uid/gid parsing will fail before any assignment in most
> > filesystems.
> > 
> > Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
> 
> I like the idea since this seems like a nobrainer but is actually
> surprisingly subtle...
> 
> > diff --git a/fs/fs_parser.c b/fs/fs_parser.c
> > index a4d6ca0b8971..24727ec34e5a 100644
> > --- a/fs/fs_parser.c
> > +++ b/fs/fs_parser.c
> > @@ -308,6 +308,40 @@ int fs_param_is_fd(struct p_log *log, const struct fs_parameter_spec *p,
> >  }
> >  EXPORT_SYMBOL(fs_param_is_fd);
> >  
> > +int fs_param_is_uid(struct p_log *log, const struct fs_parameter_spec *p,
> > +		    struct fs_parameter *param, struct fs_parse_result *result)
> > +{
> > +	kuid_t uid;
> > +
> > +	if (fs_param_is_u32(log, p, param, result) != 0)
> > +		return fs_param_bad_value(log, param);
> > +
> > +	uid = make_kuid(current_user_ns(), result->uint_32);
> 
> But here is the problem: Filesystems mountable in user namespaces need to use
> fc->user_ns for resolving uids / gids (e.g. like fuse_parse_param()).
> Having helpers that work for some filesystems and are subtly broken for
> others is worse than no helpers... Or am I missing something?
> 
> And the problem with fc->user_ns is that currently __fs_parse() does not
> get fs_context as an argument... So that will need some larger work.

Not really. If someone does an fsopen() in a namespace but the process
that actually sets mount options is in another namespace then it's
completely intransparent what uid/gid this will resolve to if it's
resovled according to fsopen().

It's also a bit strange if someone ends up handing off a tmpfs fscontext
that was created in the initial namespace to some random namespace and
they now can set uid/gid options that aren't mapped according to their
namespace but instead are 1:1 resolved according to the intial
namespace. So this would hinder delegation.

The expectation is that uid/gid options are resolved in the caller's
namespace and that shouldn't be any different for fscontexts for
namespace mountable filesystems. The crucial point is to ensure that the
resulting kuid/kgid can be resolved in the namespace the filesystem is
mounted in at the end. That's what was lacking in e.g., tmpfs in commit
0200679fc795 ("tmpfs: verify {g,u}id mount options correctly")

The fuse conversion is the only inconsistency in that regard.