On Thu, Dec 16, 2021 at 04:52:47PM -0500, James Bottomley wrote: > On Thu, 2021-12-16 at 15:08 +0100, Christian Brauner wrote: > > On Thu, Dec 16, 2021 at 12:43:10AM -0500, Stefan Berger wrote: > [...] > > > diff --git a/kernel/user_namespace.c b/kernel/user_namespace.c > > > index 6b2e3ca7ee99..6fa01323aac9 100644 > > > --- a/kernel/user_namespace.c > > > +++ b/kernel/user_namespace.c > > > @@ -20,6 +20,7 @@ > > > #include <linux/fs_struct.h> > > > #include <linux/bsearch.h> > > > #include <linux/sort.h> > > > +#include <linux/ima.h> > > > > > > static struct kmem_cache *user_ns_cachep __read_mostly; > > > static DEFINE_MUTEX(userns_state_mutex); > > > @@ -141,8 +142,14 @@ int create_user_ns(struct cred *new) > > > if (!setup_userns_sysctls(ns)) > > > goto fail_keyring; > > > > > > + ret = create_ima_ns(ns); > > > > Instead of greedily allocating a new ima namespace for each new user > > namespace creation and wasting memory that is likely wasted since > > most containers won't use ima (for a long time at least) have you > > considered lazily allocating it like I suggested in one of my first > > reviews? > > > > So under the assumption that the only way for a container to get its > > own ima policy it needs to have mounted a new securityfs instance you > > can move the ima namespace allocation into > > fill_super/ima_fs_ns_init(): > > The current patch set has the ima namespace born with an empty policy, > meaning it can never do anything until a new policy is inserted via a > write to the securityfs, and therefore the IMA namespace could be > lazily allocated. However, that's not quite how the initial IMA > namespace behaves because a policy can be passed in on the kernel > command line (or built into the kernel). If the ima NS were born with > a default policy (say taken from the initial IMA default policy, or > simply inherited from the parent at creation time) then we wouldn't be > able to do lazy allocation. Before we tie ourselves to never being > able to have a default policy for an IMA namespace, perhaps we should > discuss if this is the correct behaviour we want to nail into the > system. If ima in the future decides to do policy inheritance it can simply switch from delayed allocation at mount time to allocation at userns creation time. So we can proceed with lazy allocation without much problem for now. In addition what is happening now is in effect policy inheritance, i.e. each container is bound by the parent ima_ns policy until it decides to setup its own. Aside from that the container manager can and should be responsible for the default ima policy to apply to containers. The default ima policy can be passed through the spec file, configuration file, or - for the hardcore people - compiled into the container manager itself. This is not different from LSMs (e.g. AppArmor, seccomp) where the policy for each container is generated from a fixed template that was built into the container manager binary and then written via /proc/<pid>/attr/current during container setup. During container setup the process that ultimately calls exec to execute the payload does all of the required setup work. The setup process should not be automatically bound by a default policy that gets created when the userns is created. That will just cause problems during container setup. Until the setup process has decided that all preliminary setup steps are done only the ancestor policy should restrict it. This is exactly the same for all LSMs and seccomp. They all are usually setup closely before calling exec. I see no reason for ima to diverge from this model.
