lkml.org 
[lkml]   [2014]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v5 2/4] fuse: Support fuse filesystems outside of init_user_ns
Seth Forshee <seth.forshee@canonical.com> writes:

> On Wed, Nov 19, 2014 at 03:09:11PM +0100, Serge E. Hallyn wrote:
>> Quoting Miklos Szeredi (miklos@szeredi.hu):
>> > On Wed, Nov 19, 2014 at 9:50 AM, Miklos Szeredi <miklos@szeredi.hu> wrote:
>> > > On Tue, Nov 18, 2014 at 4:21 PM, Seth Forshee
>> > > <seth.forshee@canonical.com> wrote:
>> > >>> I asked around a bit, and it turns out there are use cases for nested
>> > >> containers (i.e. a container within a container) where the rootfs for
>> > >> the outer container mounts a filesystem containing the rootfs for the
>> > >> inner container. If that mount is nosuid then suid utilities like ping
>> > >> aren't going to work in the inner container.
>> > >>
>> > >> So since there's a use case for suid in a userns mount and we have what
>> > >> we belive are sufficient protections against using this as a vector to
>> > >> get privileges outside the container, I'm planning to move ahead without
>> > >> the MNT_NOSUID restriction. Any objections?
>> > >
>> > > In the general case how'd we prevent suid executable being tricked to
>> > > do something it shouldn't do by unprivileged mounting into sensitive
>> > > places (i.e. config files) inside the container?
>>
>> The design of the namespaces would prevent that. You cannot manipulate your
>> mounts namespace unless you own it. You cannot manipulate the mounts namespace
>> for a task whose user namespace you do not own. If you can, for instance,
>> bind mount $HOME/shadow onto /etc/shadow, then you already own your user
>> namespace and are root there, so any suid-root program which you mount through
>> fuse will only subjegate your own namespace. Any task which running in the
>> parent user-ns (and therefore parent mount-ns) will not see your bind mount.
>>
>> > > Allowing SUID looks like a slippery slope to me. And there are plenty
>> > > of solutions to the "ping" problem, AFAICS, that don't involve the
>> > > suid bit.
>> >
>> > ping isn't even suid on my system, it has security.capability xattr instead.
>>
>> security.capability xattrs that will have the exact same concerns wrt
>> confusion through bind mounts as suid.
>>
>> > Please just get rid of SUID/SGID. It's a legacy, it's a hack, not
>> > worth the complexity and potential problems arising from that
>> > complexity.
>>
>> Oh boy, I don't know which side to sit on here :) I'm all for replacing
>> suid with some use of file capabilities, but realistically there are reasons
>> why that hasn't happened more widely than it has - tar, package managers,
>> cpio, nfs, etc.
>
> Miklos: I we're all generally in agreement here that suid/sgid is not
> the best solution, but as Serge points out we are unfortunately not yet
> in a place where it can be completely dropped in favor of capabilities.
> In light of this can I convince you to reconsider your position?

Regardless of what fuse does user namespaces must support mounting
filesystems that have the setuid and setgid bits set. Likewise we need
to handle capabilities.

There is a parallel bit of work to the fuse patches that I think at this
point should be completed first.

- Add s_user_ns to struct super. So we can have filesystems whose
labels are not interpreted at a global scope.

- Tweak the file capability code to look at s_user_ns and treat it
properly.

- Tweak the lsms to look at s_user_ns and ignore security labels that
don't come from init_user_ns. (The lsms at their discrection can
be more trusting but the default should be for them to ignore those
labels).

- Tweak the security checks to allow setting file capabilities and
other security xattrs if we have the appropriate capabilities in
s_user_ns.

- Update tmpfs and ramfs to set s_user_ns when being mounted.


When those bits are done we can tweak the fuse patches to also set
s_user_ns.

As for MNT_NO_SUID if fuse wants to enforce that in some way. I don't
particularly care, but I don't think that makes sense as a vfs property.

Eric


\
 
 \ /
  Last update: 2014-11-21 19:41    [W:0.104 / U:0.788 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site