On Wed, Mar 1, 2017 at 3:28 PM, Mickaël Salaün [off-list ref] wrote:

On 01/03/2017 23:20, Andy Lutomirski wrote:

quoted

On Wed, Mar 1, 2017 at 2:14 PM, Mickaël Salaün [off-list ref] wrote:

quoted

On 28/02/2017 21:01, Andy Lutomirski wrote:

quoted

On Tue, Feb 21, 2017 at 5:26 PM, Mickaël Salaün [off-list ref] wrote:

This design makes it possible for a process to add more constraints to
its children on the fly. I think it is a good feature to have and a
safer default inheritance mechanism, but it could be guarded by an
option flag if we want both mechanism to be available. The same design
could be used by seccomp filter too.

Then let's do it right.

Currently each task has an array of seccomp filter layers.  When a
task forks, the child inherits the layers.  All the layers are
presently immutable.  With Landlock, a layer can logically be a
syscall fitler layer or a Landlock layer.  This fits in to the
existing model just fine.

If we want to have an interface to allow modification of an existing
layer, let's make it so that, when a layer is added, you have to
specify a flag to make the layer modifiable (by current, presumably,
although I can imagine other policies down the road).  Then have a
separate API that modifies a layer.

IOW, I think your patch is bad for three reasons, all fixable:

1. The default is wrong.  A layer should be immutable to avoid an easy
attack in which you try to sandbox *yourself* and then you just modify
the layer to weaken it.

This is not possible, there is only an operation for now:
SECCOMP_ADD_LANDLOCK_RULE. You can only add more rules to the list (as
for seccomp filter). There is no way to weaken a sandbox. The question
is: how do we want to handle the rules *tree* (from the kernel point of
view)?

Fair enough.  But I still think that immutability (like regular
seccomp) should be the default.  For security, simplicity is
important.  I guess there could be two ways to relax immutability:
allowing making the layer stricter and allowing any change at all.

As a default, though, programs should be able to expect that:

seccomp(SECCOMP_ADD_WHATEVER, ...);
fork();

[parent gets compromised]
[in parent]seccomp(anything whatsoever);

will not affect the child,  If the parent wants to relax that, that's
fine, but I think it should be explicit.

quoted

2. The API that adds a layer should be different from the API that
modifies a layer.

Right, but it doesn't apply now because we can only add rules.

That's not what the code appears to do, though.  Sometimes it makes a
new layer without modifying tasks that share the layer and sometimes
it modifies the layer.

Both operations are probably okay, but they're not the same operation
and they shouldn't pretend to be.

quoted

3. The whole modification mechanism should be a separate patch to be
reviewed on its own merits.

For a rule *replacement*, sure!

And for modification of policy for non-current tasks.  That's a big
departure from normal seccomp and should be reviewed as such.