Re: Login broken with old userspace (was Re: [PATCH v2] selinux: introduce an initial SID for early boot processes)

[Ondrej Mosnacek]

On Fri, Jul 28, 2023 at 5:12 PM Paul Moore <paul@xxxxxxxxxxxxxx> wrote:
>
> On Fri, Jul 28, 2023 at 9:24 AM Christian Göttsche
> <cgzones@xxxxxxxxxxxxxx> wrote:
> >
> > On Fri, 28 Jul 2023 at 15:14, Ondrej Mosnacek <omosnace@xxxxxxxxxx> wrote:
> > >
> > > On Fri, Jul 28, 2023 at 1:52 PM Stephen Smalley
> > > <stephen.smalley.work@xxxxxxxxx> wrote:
> > > >
> > > > On Fri, Jul 28, 2023 at 7:36 AM Ondrej Mosnacek <omosnace@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Jul 28, 2023 at 4:12 AM Michael Ellerman <mpe@xxxxxxxxxxxxxx> wrote:
> > > > > >
> > > > > > Ondrej Mosnacek <omosnace@xxxxxxxxxx> writes:
> > > > > > > Currently, SELinux doesn't allow distinguishing between kernel threads
> > > > > > > and userspace processes that are started before the policy is first
> > > > > > > loaded - both get the label corresponding to the kernel SID. The only
> > > > > > > way a process that persists from early boot can get a meaningful label
> > > > > > > is by doing a voluntary dyntransition or re-executing itself.
> > > > > >
> > > > > > Hi,
> > > > > >
> > > > > > This commit breaks login for me when booting linux-next kernels with old
> > > > > > userspace, specifically Ubuntu 16.04 on ppc64le. 18.04 is OK.
> > > > > >
> > > > > > The symptom is that login never accepts the root password, it just
> > > > > > always says "Login incorrect".
> > > > > >
> > > > > > Bisect points to this commit.
> > > > > >
> > > > > > Reverting this commit on top of next-20230726, fixes the problem
> > > > > > (ie. login works again).
> > > > > >
> > > > > > Booting with selinux=0 also fixes the problem.
> > > > > >
> > > > > > Is this expected? The change log below suggests backward compatibility
> > > > > > was considered, is 16.04 just too old?
> > > > >
> > > > > Hi Michael,
> > > > >
> > > > > I can reproduce it on Fedora 38 when I boot with SELINUX=disabled in
> > > > > /etc/selinux/config (+ a kernel including that commit), so it likely
> > > > > isn't caused by the userspace being old. Can you check what you have
> > > > > in /etc/selinux/config (or if it exists at all)?
> > > > >
> > > > > We have deprecated and removed the "runtime disable" functionality in
> > > > > SELinux recently [1], which was used to implement "disabling" SELinux
> > > > > via the /etc/selinux/config file, so now the situation (selinux=0 +
> > > > > SELINUX=disabled in /etc/selinux/config) leads to a state where
> > > > > SELinux is enabled, but no policy is loaded (and no enforcement is
> > > > > done). Such a state mostly behaves as if SElinux was truly disabled
> > > > > (via kernel command line), but there are some subtle differences and I
> > > > > believe we don't officially support it (Paul might clarify). With
> > > > > latest kernels it is recommended to either disable SELinux via the
> > > > > kernel command line (or Kconfig[2]) or to boot it in Enforcing or
> > > > > Permissive mode with a valid/usable policy installed.
> > > > >
> > > > > So I wonder if Ubuntu ships by default with the bad configuration or
> > > > > if it's just a result of using the custom-built linux-next kernel (or
> > > > > some changes on your part). If Ubuntu's stock kernel is configured to
> > > > > boot with SELinux enabled by default, they should also by default ship
> > > > > a usable policy and SELINUX=permissive/enforcing in
> > > > > /etc/selinux/config (or configure the kernel[2] or bootloader to boot
> > > > > with SELinux disabled by default). (Although if they ship a pre-[1]
> > > > > kernel, they may continue to rely on the runtime disable
> > > > > functionality, but it means people will have to be careful when
> > > > > booting newer or custom kernels.)
> > > > >
> > > > > That said, I'd like to get to the bottom of why the commit causes the
> > > > > login to fail and fix it somehow. I presume something in PAM chokes on
> > > > > the fact that userspace tasks now have "init" instead of "kernel" as
> > > > > the pre-policy-load security context, but so far I haven't been able
> > > > > to pinpoint the problem. I'll keep digging...
> > > > >
> > > > > [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=f22f9aaf6c3d92ebd5ad9e67acc03afebaaeb289
> > > > > [2] via CONFIG_LSM (or CONFIG_SECURITY_SELINUX_BOOTPARAM_VALUE on older kernels)
> > > >
> > > > Prior to selinux userspace commit
> > > > 685f4aeeadc0b60f3770404d4f149610d656e3c8 ("libselinux:
> > > > is_selinux_enabled(): drop no-policy-loaded test.") libselinux was
> > > > checking the result of reading /proc/self/attr/current to see if it
> > > > returned the "kernel" string as a means of detecting a system with
> > > > SELinux enabled but no policy loaded, and treated that as if SELinux
> > > > were disabled. Hence, this does break old userspace. Not sure though
> > > > why you'd see the same behavior with modern libselinux.
> > >
> > > Hm... now I tried booting the stock Fedora kernel (without the early
> > > boot initial SID commit) and I got the same failure to login as with
> > > the new kernel. So if Ubuntu 16.04 ships with pre-685f4aeeadc0
> > > libselinux (quite possible), then it seems that the scenario with
> > > terminal login + SELinux enabled + policy not loaded only works with
> > > pre-685f4aeeadc0 libselinux and pre-5b0eea835d4e kernel, the other
> > > combinations are broken. With pre-685f4aeeadc0 libselinux +
> > > post-5b0eea835d4e kernel it is expected as you say (and probably
> > > inevitable barring some hack on the kernel side), but it's not clear
> > > why also only updating libselinux seems to break it... /sys/fs/selinux
> > > is not mounted in my scenario, so there must be something else coming
> > > into play.
> > >
> > >
> > > --
> > > Ondrej Mosnacek
> > > Senior Software Engineer, Linux Security - SELinux kernel
> > > Red Hat, Inc.
> > >
> >
> > Completely untested:
> >
> > diff --git a/security/selinux/ss/services.c b/security/selinux/ss/services.c
> > index 2c5be06fbada..1ed275bd4551 100644
> > --- a/security/selinux/ss/services.c
> > +++ b/security/selinux/ss/services.c
> > @@ -1322,8 +1322,19 @@ static int security_sid_to_context_core(u32
> > sid, char **scontext,
> >         if (!selinux_initialized()) {
> >                 if (sid <= SECINITSID_NUM) {
> >                         char *scontextp;
> > -                       const char *s = initial_sid_to_string[sid];
> > +                       const char *s;
> >
> > +                       /*
> > +                        * Hide the context split of kernel threads and
> > +                        * userspace threads from userspace before the first
> > +                        * policy is loaded.  Userspace, e.g. libselinux prior
> > +                        * to v2.6 or systemd, depends on the context being
> > +                        * "kernel".
> > +                        */
> > +                       if (sid == SECINITSID_INIT)
> > +                               sid = SECINITSID_KERNEL;
> > +
> > +                       s = initial_sid_to_string[sid];
> >                         if (!s)
> >                                 return -EINVAL;
> >                         *scontext_len = strlen(s) + 1;
>
> I think I'd rather see something that does the following:
>
> 1. Convert all direct access of @initial_sid_to_string to calls to
> security_get_initial_sid_context().  I think we can leave all the
> stuff under scripts/ as-is, but I didn't think about it that hard, so
> some additional thought would be required here.

What should we then do with the reverse translation in
security_context_to_sid_core()? It seems it is currently possible for
a process to e.g. change its SID to another initial SID before the
policy is loaded - would we let it to set itself to INIT and yet still
return back KERNEL afterwards?

> 2. Modify security_get_initial_sid_context() to so something similar
> to what Christian proposed, e.g. translate INIT to KERNEL, but do so
> only when POLICYDB_CAP_USERSPACE_INITIAL_CONTEXT is not enabled.  I
> believe this should cover both the uninitialized and old policy case.

You don't know whether the policycap is enabled or not until the
policy is loaded and at that point it doesn't matter because then you
already have a full context assigned to the SID. So we could only
translate INIT to KERNEL unconditionally, but it feels wrong to lie to
userspace like that (plus there is the reverse translation issue
above)... OTOH, I don't know if we have another choice given the "no
regressions" rule...

> It might be easier if both were done as a single patch, as those that
> want the userspace isid patch will likely want this as a fix, but if
> it becomes to ugly I have no problem splitting #1 and #2 into
> different patches.
>
> What do folks think?  Am I missing something?
>
> --
> paul-moore.com
>

-- 
Ondrej Mosnacek
Senior Software Engineer, Linux Security - SELinux kernel
Red Hat, Inc.