Date:	Thu, 19 May 2016 22:40:48 -0500
From:	"Serge E. Hallyn" <serge@...lyn.com>
To:	Mimi Zohar <zohar@...ux.vnet.ibm.com>
Cc:	"Serge E. Hallyn" <serge@...lyn.com>,
	LKML <linux-kernel@...r.kernel.org>, Jann Horn <jann@...jh.net>,
	"Eric W. Biederman" <ebiederm@...ssion.com>,
	Seth Forshee <seth.forshee@...onical.com>,
	LSM <linux-security-module@...r.kernel.org>,
	"Andrew G. Morgan" <morgan@...nel.org>,
	Kees Cook <keescook@...omium.org>,
	Michael Kerrisk-manpages <mtk.manpages@...il.com>,
	"Serge E. Hallyn" <serge.hallyn@...ntu.com>,
	Linux API <linux-api@...r.kernel.org>,
	Andy Lutomirski <luto@...capital.net>,
	Linux Containers <containers@...ts.linux-foundation.org>
Subject: Re: [PATCH RFC] user-namespaced file capabilities - now with more
 magic

Quoting Mimi Zohar (zohar@...ux.vnet.ibm.com):
> On Wed, 2016-05-18 at 16:57 -0500, Serge E. Hallyn wrote:
> > This patch introduces a new security.nscapability xattr.  It
> > is mostly like security.capability, but also lists a 'rootid'.
> > This is the uid_t (in init_user_ns) of the root id (uid 0 in a
> > namespace) in whose namespaces the file capabilities may take
> > effect.
> > 
> > A privileged (cap_setfcap) process in the initial user ns may
> > set and read this xattr directly.  However, its real intent is
> > to be used as a transparent fallback in user namespaces.
> > 
> > Root in a user ns cannot be trusted to write security.capability
> > xattrs, because any user on the host could map his own uid to root
> > in a namespace, write the xattr, and execute the file with privilege
> > on the host.
> > 
> > With this patch, when root in a user ns asks to write security.capability,
> > the kernel will transparently write a security.nscapability xattr
> > instead, filling in the kuid of the calling user's root uid.  Subsequently,
> > any task executing the file which has the noted k_uid as its root uid,
> > or which is in a descendent user_ns of such a user_ns, will run the
> > file with capabilities.
> > 
> > When reading the security.capability xattr from a non-init user_ns, a valid
> > security.nscapability will be shown if it exists.  Such a task is not
> > allowed to read security.nscapability.  This could be accomodated, however
> 
> Add the word "directly" as "to read security.nscapability directly".

Updated in my git tree.

> > it requires the kernel to convert the kuid_t to a valid uid in the reader's
> > user_ns.  So for now it's simply not supported.
> 
> I really like the idea that the kernel transparently replaces
> nscapability for capability.
> 
> > Only a single security.nscapability xattr may be written.  This patch
> > could be expanded to allow a list of capabilities and rootids, however
> > I do not believe that to be a worthwhile use case.
> 
> Ok
> 
> > This allows a simple setxattr to work, allows tar/untar to
> > work, and allows us to tar in one namespace and untar in
> > another while preserving the capability, without risking
> > leaking privilege into a parent namespace.
> > 
> > Note - listxattr is not being handled here.  So results of that can be
> > inconsistent with get/setxattr.  Fixing that will require yet more
> > deceit in fs/xattr.c.
> > 
> > Note2 - it may be less sneaky to hide all the magic behind the
> > security.nscapability xattr.  So userspace would need to know to
> > use that xattr name when needed, but with the same format as
> > security.capability.  The kuid_t rootid would be filled in by the
> > kernel.  That's a middle ground between my last patch and this one.
> 
> The less userspace needs to differentiate between running in a namespace
> and not, the better.
> 
> Note3 - capability is currently protected by EVM, when enabled.  Should
> ns_capability also be a protected xattr?

Hm - that would protect it from offline attacks, but allow the container
to update it, right?  That sounds good.

> > Signed-off-by: Serge Hallyn <serge.hallyn@...ntu.com>
> > ---
> >  fs/xattr.c                      |  18 ++-
> >  include/linux/capability.h      |   8 +-
> >  include/uapi/linux/capability.h |  19 +++
> >  include/uapi/linux/xattr.h      |   3 +
> >  security/commoncap.c            | 253 ++++++++++++++++++++++++++++++++++++++--
> >  5 files changed, 291 insertions(+), 10 deletions(-)
> > 
> > diff --git a/fs/xattr.c b/fs/xattr.c
> > index 4861322..5c0e7ae 100644
> > --- a/fs/xattr.c
> > +++ b/fs/xattr.c
> > @@ -94,11 +94,26 @@ int __vfs_setxattr_noperm(struct dentry *dentry, const char *name,
> >  {
> >  	struct inode *inode = dentry->d_inode;
> >  	int error = -EOPNOTSUPP;
> > +	void *wvalue = NULL;
> > +	size_t wsize = 0;
> >  	int issec = !strncmp(name, XATTR_SECURITY_PREFIX,
> >  				   XATTR_SECURITY_PREFIX_LEN);
> > 
> > -	if (issec)
> > +	if (issec) {
> >  		inode->i_flags &= ~S_NOSEC;
> > +		/* if root in a non-init user_ns tries to set
> > +		 * security.capability, write a security.nscapability
> > +		 * in its place */
> > +		if (!strcmp(name, "security.capability") &&
> > +				current_user_ns() != &init_user_ns) {
> > +			cap_setxattr_make_nscap(dentry, value, size, &wvalue, &wsize);
> > +			if (!wvalue)
> > +				return -EPERM;
> > +			value = wvalue;
> > +			size = wsize;
> > +			name = "security.nscapability";
> > +		}
> 
> The call to capable_wrt_inode_uidgid() is hidden behind
> cap_setxattr_make_nscap().  Does it make sense to call it here instead,
> before the security.capability test?  This would lay the foundation for
> doing something similar for IMA.

Might make sense to move that.  Though looking at it with fresh eyes I wonder
whether adding less code here at __vfs_setxattr_noperm(), i.e.

		if (!cap_setxattr_makenscap(dentry, &value, &size, &name))
			return -EPERM;

would be cleaner.

> (Will continue reviewing ...)

Awesome, thanks Mimi.

Powered by blists - more mailing lists