[PATCH v7 00/14] ima: Namespace IMA with audit support in IMA-ns

[Christian Brauner]

On Thu, Dec 16, 2021 at 01:50:27PM +0100, Christian Brauner wrote:
> On Thu, Dec 16, 2021 at 12:43:09AM -0500, Stefan Berger wrote:
> > From: Stefan Berger <stefanb at linux.ibm.com>
> > 
> > The goal of this series of patches is to start with the namespacing of
> > IMA and support auditing within an IMA namespace (IMA-ns) as the first
> > step.
> > 
> > In this series the IMA namespace is piggy backing on the user namespace
> > and therefore an IMA namespace gets created when a user namespace is
> > created. The advantage of this is that the user namespace can provide
> > the keys infrastructure that IMA appraisal support will need later on.
> > 
> > We chose the goal of supporting auditing within an IMA namespace since it
> > requires the least changes to IMA. Following this series, auditing within
> > an IMA namespace can be activated by a user running the following lines
> > that rely on a statically linked busybox to be installed on the host for
> > execution within the minimal container environment:
> > 
> > mkdir -p rootfs/{bin,mnt,proc}
> > cp /sbin/busybox rootfs/bin
> > cp /sbin/busybox rootfs/bin/busybox2
> > echo >> rootfs/bin/busybox2
> > PATH=/bin unshare --user --map-root-user --mount-proc --pid --fork \
> >   --root rootfs busybox sh -c \
> >  "busybox mount -t securityfs /mnt /mnt; \
> >   busybox echo 'audit func=BPRM_CHECK mask=MAY_EXEC' > /mnt/ima/policy; \
> >   busybox2 cat /mnt/ima/policy"
> > 
> > [busybox2 is used to demonstrate 2 measurements; see below]
> > 
> > Following the audit log on the host the last line cat'ing the IMA policy
> > inside the namespace would have been audited. Unfortunately the auditing
> > line is not distinguishable from one stemming from actions on the host.
> > The hope here is that Richard Brigg's container id support for auditing
> > would help resolve the problem.
> > 
> > The following lines added to a suitable IMA policy on the host would
> > cause the execution of the commands inside the container (by uid 1000)
> > to be measured and audited as well on the host, thus leading to two
> > auditing messages for the 'busybox2 cat' above and log entries in IMA's
> > system log.
> > 
> > echo -e "measure func=BPRM_CHECK mask=MAY_EXEC uid=1000\n" \
> >         "audit func=BPRM_CHECK mask=MAY_EXEC uid=1000\n" \
> >     > /sys/kernel/security/ima/policy
> > 
> > The goal of supporting measurement and auditing by the host, of actions
> > occurring within IMA namespaces, is that users, particularly root,
> > should not be able to evade the host's IMA policy just by spawning
> > new IMA namespaces, running programs there, and discarding the namespaces
> > again. This is achieved through 'hierarchical processing' of file
> > accesses that are evaluated against the policy of the namespace where
> > the action occurred and against all namespaces' and their policies leading
> > back to the root IMA namespace (init_ima_ns).
> 
> Note that your worst-case is 32 levels (maximum supported userns
> nesting) where each ima namespace defines a separate policy.
> 
> So make sure you don't run into locking issues when hierarchically
> processing rules. So far I think it's fine since the locks aren't held
> across the hierarchial walk but are dropped and reaqcuired for each
> level.
> 
> But that could still mean a lot of contention on iint->mutex since this
> lock is global, i.e. in this context: for all ima namespaces. You might
> want to consider coming up with some rough ideas for how to solve this
> _if_ this becomes a problem in the future.
> 
> > 
> > The patch series adds support for a virtualized SecurityFS with a few
> > new API calls that are used by IMA namespacing. Only the data relevant
> > to the IMA namespace are shown. The files and directories of other
> > security subsystems (TPM, evm, Tomoyo, safesetid) are not showing
> > up when secruityfs is mounted inside a user namespace.
> > 
> > Much of the code leading up to the virtualization of SecurityFS deals
> > with moving IMA's variables from various files into the IMA namespace
> > structure called 'ima_namespace'. When it comes to determining the
> > current IMA namespace I took the approach to get the current IMA
> > namespace (get_current_ns()) on the top level and pass the pointer all
> > the way down to those functions that now need access to the ima_namespace
> > to get to their variables. This later on comes in handy once hierarchical
> > processing is implemented in this series where we walk the list of
> > namespaces backwards and again need to pass the pointer into functions.
> 
> Just to repeat the point from earlier reviews, all those functions need
> to be guaranteed to call from syscall context. Functions that operate on
> files have different semantics.
> 
> > 
> > This patch also introduces usage of CAP_MAC_ADMIN to allow access to the
> > IMA policy via reduced capabilities. We would again later on use this
> > capability to allow users to set file extended attributes for IMA appraisal
> > support.
> > 
> > My tree with these patches is here:
> > 
> > git fetch https://github.com/stefanberger/linux-ima-namespaces v5.15+imans.v7.posted 
> > 
> > Regards,
> >    Stefan
> > 
> > v7:
> >  - Dropped 2 patches related to key queues; using &init_ima_ns for all calls
> >    from functions related to key queues where calls need ima_namespace
> >  - Moved ima_namespace to security/integrity/ima/ima.h
> >  - Extended API descriptions with ns parameter where needed
> >  - Using init_ima_ns in functions related to appraisal and xattrs
> >  - SecurityFS: Using ima_ns_from_file() to get ns pointer 
> >  - Reformatted to 80 columns per line
> 
> Since we're starting to be fairly along I would ask you to please write
> detailed commit messages for the next revision.
> 
> I would also like to see all links for prior versions of this patchset
> in the commit message since the discussion has been fairly extensive so
> for this series it makes a lot of sense. So something like:
> 
> Link: https://lore.kernel.org/r/$MSGID (v1)
> Link: https://lore.kernel.org/r/$MSGID (v2)
> Link: https://lore.kernel.org/r/$MSGID (v3)
> Link: https://lore.kernel.org/r/$MSGID (v4)
> Link: https://lore.kernel.org/r/$MSGID (v5)
> Link: https://lore.kernel.org/r/$MSGID (v6)
> Link: https://lore.kernel.org/r/$MSGID (v7)
> Signed-off-by: meh
> Signed-off-by: mih
> Signed-off-by: muh
> 
> I find that extremely pleasant in case we need to revisit things later.
> (Technically you can get the same by searching lore via the final link
> but I find it be pretty pleasing to just copy+paste directly from the
> commit message to the discussion for the earlier patch.)

So I looked through the series from a high-level view for once and I
would like to change how it is currently structured.

Currently, it looks a lot like you end up with a half-namespaced ima if
you compile and run a kernel in the middle of this patch series. Not
just is this asking for semantic chaos if we need to debug something it
also makes bisection a giant pain later.

In addition, the fact that you need a hack like

> +struct ima_namespace {
> +	int avoid_zero_size;

in the first patch is another good sign that this should be restructured.

Here's how I would prefer to see this done. I think we should organize
this in three big chunks (bullet points are not meant to signify
individual patches):

1. namespace securityfs
   This patch is thematically standalone and should move to the
   beginning of the series.
   I would strongly recommend to fold patch 9 and 10 into a single patch
   and add a lengthy explanation. You should be able to recycle a lof of
   stuff I wrote in earlier reviews.

2. Introduce struct ima_namespace and pass it through to all callers:
   - introduce struct ima_namespace
   - move all the relevant things into this structure (this also avoids
     the "avoid_zero_size" hack).
   - define, setup, and expose init_ima_ns 
   - introduce get_current_ns() and always have it return &init_ima_ns for now
   - replace all accesses to global variables to go through &init_ima_ns
   - add new infrastructure you'll need later on
   Bonus is that you can extend all the functions that later need access
   to a specific ima namespace to take a struct ima_namespace * argument
   and pass down &init_ima_ns down (retrieved via get_current_ns()). This
   will make the actual namespace patch very easy to follow.

3. namespace ima
   - add a new entry for struct ima_namespace to struct user_namespace
   - add creation helpers, kmem cache etc.
   - create files in securityfs per ns

This way at all points in the series we have clearly defined semantics
where ima namespacing is either fully working or fully not working and
the switch is atomic in the patch(es) part of 3.