[RFC 17/20] ima: Use integrity_admin_ns_capable() to check corresponding capability

[James Bottomley]

On Wed, 2021-12-01 at 12:35 -0500, Stefan Berger wrote:
> On 12/1/21 11:58, James Bottomley wrote:
> > On Tue, 2021-11-30 at 11:06 -0500, Stefan Berger wrote:
> > > From: Denis Semakin <denis.semakin at huawei.com>
> > > 
> > > Use integrity_admin_ns_capable() to check corresponding
> > > capability to allow read/write IMA policy without CAP_SYS_ADMIN
> > > but with CAP_INTEGRITY_ADMIN.
> > > 
> > > Signed-off-by: Denis Semakin <denis.semakin at huawei.com>
> > > ---
> > >   security/integrity/ima/ima_fs.c | 2 +-
> > >   1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/security/integrity/ima/ima_fs.c
> > > b/security/integrity/ima/ima_fs.c
> > > index fd2798f2d224..6766bb8262f2 100644
> > > --- a/security/integrity/ima/ima_fs.c
> > > +++ b/security/integrity/ima/ima_fs.c
> > > @@ -393,7 +393,7 @@ static int ima_open_policy(struct inode
> > > *inode,
> > > struct file *filp)
> > >   #else
> > >   		if ((filp->f_flags & O_ACCMODE) != O_RDONLY)
> > >   			return -EACCES;
> > > -		if (!ns_capable(ns->user_ns, CAP_SYS_ADMIN))
> > > +		if (!integrity_admin_ns_capable(ns->user_ns))
> > so this one is basically replacing what you did in RFC 16/20, which
> > seems a little redundant.
> > 
> > The question I'd like to ask is: is there still a reason for
> > needing CAP_INTEGRITY_ADMIN?  My thinking is that now IMA is pretty
> > much tied to requiring a user (and a mount, because of
> > securityfs_ns) namespace, there might not be a pressing need for an
> > admin capability separated from CAP_SYS_ADMIN because the owner of
> > the user namespace passes the ns_capable(..., CAP_SYS_ADMIN)
> > check.  The rationale in
> 
> Casey suggested using CAP_MAC_ADMIN, which I think would also work.
> 
>      CAP_MAC_ADMIN (since Linux 2.6.25)
>                Allow MAC configuration or state changes. Implemented
> for
>                the Smack Linux Security Module (LSM).
> 
> 
> Down the road I think we should cover setting file extended
> attributes with the same capability as well for when a user signs
> files or installs packages with file signatures.  A container runtime
> could hold CAP_SYS_ADMIN while setting up a container and mounting
> filesystems and drop it for the first process started there. Since we
> are using the user namespace to spawn an IMA namespace, we would then
> require CAP_SYSTEM_ADMIN to be left available so that the user can do
> IMA related stuff in the container (set or append to the policy,
> write file signatures). I am not sure whether that should be the case
> or rather give the user something finer grained, such as
> CAP_MAC_ADMIN. So, it's about granularity...

It's possible ... any orchestration system that doesn't enter a user
namespace has to strictly regulate capabilities.   I'm probably biased
because I always use a user_ns so I never really had to mess with
capabilities.

> > https://kernsec.org/wiki/index.php/IMA_Namespacing_design_considerations
> > 
> > Is effectively "because CAP_SYS_ADMIN is too powerful" but that's
> > no longer true of the user namespace owner.  It only passes the
> > ns_capable() check not the capable() one, so while it does get
> > CAP_SYS_ADMIN, it can only use it in a few situations which
> > represent quite a power reduction already.
> 
> At least docker containers drop CAP_SYS_ADMIN.

Well docker doesn't use the user_ns.  But even given that,
CAP_SYS_ADMIN is always dropped for most container systems.  What
happens when you enter a user namespace is the ns_capable( ...,
CAP_SYS_ADMIN) check returns true if you're the owner of the user_ns,
in the same way it would for root.  So effectively entering a user
namespace without CAP_SYS_ADMIN but mapping the owner id to 0 (what
unshare -r --user does) gives you back a form of CAP_SYS_ADMIN that
responds only in the places in the kernel that have a ns_capable()
check instead of a capable() one (most of the places you list below). 
This is the principle of how unprivileged containers actually work ...
and the source of some of our security problems if you get back an
ability to do something you shouldn't be allowed to do as an
unprivileged user.

>  I am not sure what the decision was based on but probably they don't
> want to give the user what is not absolutely necessary, but usage of
> user namespaces (with IMA namespaces) would kind of force it to be
> available then to do IMA-related stuff ...
> 
> Following this man page here 
> https://man7.org/linux/man-pages/man7/user_namespaces.7.html
> 
> CAP_SYS_ADMIN in a user namespace is about
> 
> - bind-mounting filesystems
> 
> - mounting /proc filesystems
> 
> - creating nested user namespaces
> 
> - configuring UTS namespace
> 
> - configuring whether setgroups() can be used
> 
> - usage of setns()
> 
> 
> Do we want to add '- only way of *setting up* IMA related stuff' to
> this list?

I don't see why not, but other container people should weigh in
because, as I said, I mostly use the user namespace and unprivileged
containers and don't bother with capabilities.

James