lkml.org 
[lkml]   [2014]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: Thoughts on credential switching
On Wed, Mar 26, 2014 at 7:48 PM, Jeff Layton <jlayton@redhat.com> wrote:
> On Wed, 26 Mar 2014 17:23:24 -0700
> Andy Lutomirski <luto@amacapital.net> wrote:
>
>> Hi various people who care about user-space NFS servers and/or
>> security-relevant APIs.
>>
>> I propose the following set of new syscalls:
>>
>> int credfd_create(unsigned int flags): returns a new credfd that
>> corresponds to current's creds.
>>
>> int credfd_activate(int fd, unsigned int flags): Change current's
>> creds to match the creds stored in fd. To be clear, this changes both
>> the "subjective" and "objective" (aka real_cred and cred) because
>> there aren't any real semantics for what happens when userspace code
>> runs with real_cred != cred.
>>
>> Rules:
>>
>> - credfd_activate fails (-EINVAL) if fd is not a credfd.
>> - credfd_activate fails (-EPERM) if the fd's userns doesn't match
>> current's userns. credfd_activate is not intended to be a substitute
>> for setns.
>> - credfd_activate will fail (-EPERM) if LSM does not allow the
>> switch. This probably needs to be a new selinux action --
>> dyntransition is too restrictive.
>>
>>
>> Optional:
>> - credfd_create always sets cloexec, because the alternative is
>> silly.
>> - credfd_activate fails (-EINVAL) if dumpable. This is because we
>> don't want a privileged daemon to be ptraced while impersonating
>> someone else.
>> - optional: both credfd_create and credfd_activate fail if
>> !ns_capable(CAP_SYS_ADMIN) or perhaps !capable(CAP_SETUID).
>>
>> The first question: does this solve Ganesha's problem?
>>
>> The second question: is this safe? I can see two major concerns. The
>> bigger concern is that having these syscalls available will allow
>> users to exploit things that were previously secure. For example,
>> maybe some configuration assumes that a task running as uid==1 can't
>> switch to uid==2, even with uid 2's consent. Similar issues happen
>> with capabilities. If CAP_SYS_ADMIN is not required, then this is no
>> longer really true.
>>
>> Alternatively, something running as uid == 0 with heavy capability
>> restrictions in a mount namespace (but not a uid namespace) could pass
>> a credfd out of the namespace. This could break things like Docker
>> pretty badly. CAP_SYS_ADMIN guards against this to some extent. But
>> I think that Docker is already totally screwed if a Docker root task
>> can receive an O_DIRECTORY or O_PATH fd out of the container, so it's
>> not entirely clear that the situation is any worse, even without
>> requiring CAP_SYS_ADMIN.
>>
>> The second concern is that it may be difficult to use this correctly.
>> There's a reason that real_cred and cred exist, but it's not really
>> well set up for being used.
>>
>> As a simple way to stay safe, Ganesha could only use credfds that have
>> real_uid == 0.
>>
>> --Andy
>
>
> I still don't quite grok why having this special credfd_create call
> buys you anything over simply doing what Al had originally suggested --
> switch creds using all of the different syscalls and then simply caching
> that in a "normal" fd:
>
> fd = open("/dev/null", O_PATH...);
>
> ...it seems to me that the credfd_activate call will still need to do
> the same permission checking that all of the individual set*id() calls
> require (and all of the other stuff like changing selinux contexts,
> etc).
>
> IOW, this fd is just a "handle" for passing around a struct cred, but I
> don't see why having access to that handle would allow you to do
> something you couldn't already do anyway.
>
> Am I missing something obvious here?

Not really. I think I didn't adequately explain a piece of this.

I think that what you're suggesting is for an fd to encode a set of
credentials but not to grant permission to use those credentials. So
switch_creds(fd) is more or less the same thing as switch_creds(ruid,
euid, suid, rgid, egid, sgid, groups, mac label, ...). switch_creds
needs to verify that the caller can dyntransition to the label, set
all the ids, etc., but it avoids allocating anything and running RCU
callbacks.

The trouble with this is that the verification needed is complicated
and expensive. And I think that my proposal is potentially more
useful.

A credfd is like a struct cred, but possession of a credfd carries the
permission to use those credentials. So, for example, credfd_activate
to switch to a given uid might work even if setresuid to that uid
would be disallowed. But, for this to be secure, the act of giving
someone a credfd needs to be explicit. Programs implicitly send other
programs their credentials by means of f_cred all the time, and they
don't expect to allow the receiver to impersonate them.

credfd has other uses. A file server, for example, could actually
delegate creation of the credfds to a separate process, and that
process could validate that the request is for a credfd that the file
server really should be able to obtain. This would enable that
process to make sure that the user in question has actually
authenticated itself, so a file server compromise could only attack
users who connect instead of attacking any user on the system. This
is an argument against requiring CAP_SYS_ADMIN to use credfd_activate.

I'm less confident in whether capabilities should be needed to use
credfd_create.

Is that clearer and/or more convincing?

--Andy


\
 
 \ /
  Last update: 2014-03-27 04:41    [W:0.154 / U:0.416 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site