On Tue, Apr 11, 2017 at 6:23 AM, Kees Cook [off-list ref] wrote:

On Sun, Apr 9, 2017 at 3:42 AM, Djalal Harouni [off-list ref] wrote:

[...]

quoted

A userspace request to use a kernel feature that is implemented by modules
that are not loaded may trigger the module auto-load feature to load
these modules in order to satisfy userspace. However as today's Linux use
cases cover embedded systems to containers where applications are running
in their own separate environments, reducing or preventing operations
that may affect external environments is an important constraint.
Therefore, we need a way to control if automatic module loading is
allowed or which applications are allowed to trigger the module
auto-load feature.

It might be worth looking through the last several years of kernel
CVEs to find all the ones that would have been stopped with a feature
like this. There are many examples lately: both DCCP (CVE-2017-6074)
and n_hldc (CVE-2017-2636) come to mind this year alone.

You are right, will do it.

quoted

The ModAutoRestrict LSM allows system administrators or sandbox
mechanisms to control the module auto-load feature and prevent loading
unneeded modules or abuse the interface.

The settings can be applied globally using a sysctl interface which
completes the core kernel interface "modules_disable".

Typo: complements

Perhaps mention why this is a complement (i.e. modules_disable is ALL
modules, and modautorestrict is just auto-loaded, not explicitly
loaded).

Ok will fix it.

quoted

The feature is also available as a prctl() interface. This allows to
apply restrictions when sandboxing processes. On embedded Linux systems,
or containers where only some containers/processes should have the
right privileges to load modules, this allows to restrict those
processes from inserting modules. Only privileged processes can be
allowed to perform so. A more restrictive access can be applied where
the module autoload feature is completely disabled.
In this schema the access rules are per-process and inherited by
children created by fork(2) and clone(2), and preserved across execve(2).

Interface:

*) The per-process prctl() settings are:

 prctl(PR_MOD_AUTO_RESTRICT_OPTS, PR_SET_MOD_AUTO_RESTRICT, value, 0, 0)

 Where value means:

 0 - Classic module auto-load permissions, nothing changes.

 1 - The current process must have CAP_SYS_MODULE to be able to
     auto-load modules. CAP_NET_ADMIN should allow to auto-load
     modules with a 'netdev-%s' alias.

 2 - Current process can not auto-load modules. Once set, this prctl
     value can not be changed.

 The per-process value may only be increased, never decreased, thus ensuring
 that once applied, processes can never relaxe their setting.

Typo: relax

quoted

*) The global sysctl setting can be set by writting an integer value to
   '/proc/sys/kernel/modautorestrict/autoload'

I wonder if this should just be named
/proc/sys/kernel/modules_autoload (to look more like
modules_disabled)?

Indeed that's better, noted.

quoted

 The valid values are:

 0 - Classic module auto-load permissions, nothing changes.

 1 - Processes must have CAP_SYS_MODULE to be able to auto-load modules.
     CAP_NET_ADMIN should allow to auto-load modules with a 'netdev-%s'
     alias.

 2 - Processes can not auto-load modules. Once set, this sysctl value
     can not be changed.

*) Access rules:
   First the prctl() settings are checked, if the access is not denied
   then the global sysctl settings are checked.

The original idea and inspiration is from grsecurity 'GRKERNSEC_MODHARDEN'.


The sample code here can be used to test the feature:
https://gist.githubusercontent.com/tixxdz/39a583358f04d40b4d3e5571f95c075b/raw/7fb416285412891e2637fba149da930fe0898356/modautorestrict_test.c

# TODO list:
*) Confirm the struct task_struct->security stacking mechanism.

If we can't settle on a way to do this, perhaps this LSM should just
live in the core kernel, similar to modules_disabled? We do already
have the hooks to implement it with an LSM, but I'd hate to block this
feature just because we can't solve task_struct sharing.

Maybe add a new bit like no_new_privs? (Though I guess you'd need 2
bits, so maybe a full field like no_new_privs used to be.)

I can go with that too. The feature is really useful! and after
thinking about it, it could also be in task and continue to use LSM
hooks, but implemented more like the capability security module, this
way it is more of a core kernel feature than a proper LSM on its own.

quoted

*) Add a logging mechanism.
*) Remove the use of security_kernel_read_file hook. Use only
   security_kernel_module_request and make sure that we cover all cases.
*) Convert documentation to .rst

Another TODO item could be to add output to /proc/$pid/status as done
for seccomp and no_new_privs? Though perhaps only if this is a core
kernel change...

Ok, noted.

Thank you for the review!

-- 
tixxdz