On Mon, Apr 10, 2017 at 9:26 PM, Casey Schaufler [off-list ref] wrote:

On 4/10/2017 11:30 AM, Djalal Harouni wrote:

quoted

On Mon, Apr 10, 2017 at 5:50 PM, Casey Schaufler [off-list ref] wrote:

quoted

On 4/9/2017 3:42 AM, Djalal Harouni wrote:

quoted

From: Tetsuo Handa <penguin-kernel-JPay3/Yim36HaxMnTkn67Xf5DAMn2ifp@public.gmane.org>

[...]

quoted

quoted

While I appreciate your enthusiasm, I would really like
to see a mechanism for managing all of the blobs as being
potentially shared rather than just the task blob. With
infrastructure blob management being the general case we
could stack any set of existing modules that does not
include both SELinux and Smack. (AppArmor is threatening
to join that set, but we're still waiting on that) I

Yes! most of the other kernel maintainers I discussed the feature with
did not even believe or knew that LSM stacking is possible! Some other
kernel frameworks are being replaced with new ones (I'm not sure if
the old ones were perfect!)... but what I'm trying to say is that we
should make it easy for dynamic LSM plugins model so they can explore
the interface, and maybe after some time the whole framework will even
be replaced with a better one...

I'm not committing to dynamic modules, but I am working
to make sure I don't do anything to prevent someone else
from doing it in the future. There are a good number of
people who find the notion of dynamic security policy
disturbing.

I understand, though that with today's container recycling use cases
some of it may make sense. Anyway you are doing the real work, you
know better.

quoted

Thanks to your effort and others we may achieve it, but I don't think
that delaying features for a perfect implementation is the best way.
During linuxcon 2016 Europe you said that there should be a new kind
of LSMs, that's why I think we should make it easy for that to happen.

I agree that it's better to use a work-around today then to
let the shortcomings of the existing infrastructure stop you
from getting your job done.

Indeed!

quoted

quoted

think it's a bad idea to go ahead with one implementation
for task blobs without taking care of the others.

Ok. Sorry if I missed this, but could you please point me why this
could be a bad idea ?

It's a whole lot easier to maintain one sort of blob
management then a different kind for each blob. It would
be lots cleaner to have a single call that tells the
infrastructure how much space you need for all your blobs
than a separate interface for each.

I perfectly agree here. With Tetsuo patch it was easy to add a
stackable LSM, with a consistent approach that would be even better.
But as you note that should not be a blocker to upstream new stackable
LSMs.

quoted

As I understand it, these are internal details that could be replaced.
My first implementation used resizable concurrent hashtables that are
heavily used in the networking code and it worked, and I easily
converted the module to use Tetsuo's approach since I didn't see any
objection for it, and it continued to work. Maybe the point should be
*which* LSM should use the task->security and how ? The
ModAutoRestrict module is a simple LSM which could easily work with
any provided solution.

So I see.

quoted

That being said, I could convert the module back and stick with
rhashtables for now since it does not conflict with any module and it
works well. I would like to avoid same history where Apparmor, Smack
and TOMOYO all suffered to get mainlined, even Yama due to some
requests...  Then I can convert the module back to use the same LSM
infrastructure if you maintainers think that how it should go, that's
totally fine by me. Yama internally could use the same task blob but
it is avoiding conflicts by using lists to manage its internal data,
that's the same design with ModAutoRestrict and rhashtables.

I think that would be the prudent approach. There is still
the possibility that blob sharing (or full stacking, if you
prefer) won't be accepted any time soon.

Ok Casey! I will wait for more feedback, and if other maintainers do
not object, I will convert it back to rhashtables in next iterations
making sure that it should be simple to convert later to a blob
sharing mechanism.

Thanks again for the comments!


-- 
tixxdz