[PATCH v5 15/16] ima: Move dentries into ima_namespace

[Mimi Zohar]

On Fri, 2021-12-10 at 10:32 -0500, Stefan Berger wrote:
> On 12/10/21 10:26, Mimi Zohar wrote:
> > On Fri, 2021-12-10 at 09:26 -0500, James Bottomley wrote:
> >> On Fri, 2021-12-10 at 09:17 -0500, Stefan Berger wrote:
> >>> On 12/10/21 08:02, Mimi Zohar wrote:
> >>>> On Fri, 2021-12-10 at 07:40 -0500, Stefan Berger wrote:
> >>>>> On 12/10/21 07:09, Mimi Zohar wrote:
> >>>>>> On Fri, 2021-12-10 at 12:49 +0100, Christian Brauner wrote:
> >>>>>>>> There's still the problem that if you write the policy,
> >>>>>>>> making the file disappear then unmount and remount
> >>>>>>>> securityfs it will come back.  My guess for fixing this is
> >>>>>>>> that we only stash the policy file reference,
> >>>>>>>> create it if NULL but then set the pointer to PTR_ERR(-
> >>>>>>>> EINVAL) or something and refuse to create it for that
> >>>>>>>> value.
> >>>>>>> Some sort of indicator that gets stashed in struct ima_ns
> >>>>>>> that the file does not get recreated on consecutive mounts.
> >>>>>>> That shouldn't be hard to fix.
> >>>>>> The policy file disappearing is for backwards compatibility,
> >>>>>> prior to being able to extend the custom policy.  For embedded
> >>>>>> usecases, allowing the policy to be written exactly once might
> >>>>>> makes sense.  Do we really want/need to continue to support
> >>>>>> removing the policy in namespaces?
> >>>>> I don't have an answer but should the behavior for the same
> >>>>> #define in this case be different for host and namespaces? Or
> >>>>> should we just 'select IMA_WRITE_POLICY and IMA_READ_POLICY' when
> >>>>> IMA_NS is selected?
> >>>> The latter option sounds good.  Being able to analyze the namespace
> >>>> policy is really important.
> >>> Ok, I will adjust the Kconfig for this then. This then warrants the
> >>> question whether to move the dentry into the ima_namespace. The
> >>> current code looks like this.
> >>>
> >>> #if !defined(CONFIG_IMA_WRITE_POLICY) &&
> >>> !defined(CONFIG_IMA_READ_POLICY)
> >>>           securityfs_remove(ns->policy_dentry);
> >>>           ns->policy_dentry = NULL;
> >>>           ns->policy_dentry_removed = true;
> >>> #elif defined(CONFIG_IMA_WRITE_POLICY)
> >>>
> >>> With IMA_NS selecting IMA_WRITE_POLICY and IMA_READ_POLICY the above
> >>> wouldn't be necessary anymore but I find it 'cleaner' to still have
> >>> the dentry isolated rather than it being a global static as it was
> >>> before...
> >> This is really, really why you don't want the semantics inside the
> >> namespace to differ from those outside, because it creates confusion
> >> for the people reading the code, especially with magically forced
> >> config options like this.  I'd strongly suggest you either keep the
> >> semantic in the namespace or eliminate it entirely.
> >>
> >> If you really, really have to make the namespace behave differently,
> >> then use global variables and put a big comment on that code saying it
> >> can never be reached once CONFIG_IMA_NS is enabled.
> > The problem seems to be with removing the securityfs policy file.
> > Instead of removing it, just make it inacessible for the "if
> > !defined(CONFIG_IMA_WRITE_POLICY) && !defined(CONFIG_IMA_READ_POLICY)"
> > case.
> 
> So we would then leave it up to the one building the kernel to select 
> the proper compile time options (suggested ones being IMA_WRITE_POLICY 
> and IMA_READ_POLICY being enabled?) and behavior of host and IMA 
> namespace is then the same per those options? Removing the file didn't 
> seem the problem to me but more like whether the host should ever behave 
> differently from the namespace.

You proposed "select IMA_WRITE_POLICY and IMA_READ_POLICY'" when IMA_NS
is selected.  At least IMA_READ_POLICY should be enabled for
namespaces.

In addition, if removing the securityfs file after a custom policy is
loaded complicates namespacing, then don't remove it.

thanks,

Mimi