KEYS: make keyctl_invalidate() also require Setattr permission
diff mbox

Message ID 20170329220101.26067-1-ebiggers@google.com
State Not Applicable
Headers show

Commit Message

Eric Biggers March 29, 2017, 10:01 p.m. UTC
Currently, a process only needs Search permission on a key to invalidate
it with keyctl_invalidate(), which has an effect similar to unlinking it
from all keyrings.  Unfortunately, this significantly compromises the
keys permission system because as a result, there is no way to grant
read-only access to keys without also granting de-facto "delete" access.
Even worse, processes may invalidate entire _keyrings_, given only
permission to "search" them.

Key invalidation seems to have intended for cases where keyrings are
used as caches, and keys can be re-requested at any time by an upcall to
/sbin/request-key.  However, keyrings are not always used in this way.
For example, users of ext4, f2fs, and ubifs encryption may wish to grant
many differently-privileged processes access to the same keys, set up in
a shared keyring ahead of time.  Permitting everyone to invalidate keys
in this scenario enables a trivial denial-of-service.  And even users of
keyrings as "just caches" may wish to restrict invalidation because it
may require significant work or user interaction to regenerate keys.

This patch fixes the flaw by requiring both Search and Setattr
permission to invalidate a key rather than just Search permission.
Requiring Setattr permission is appropriate because Setattr permission
also allows revoking (via keyctl_revoke()) or expiring (via
keyctl_set_timeout()) the key, which also make the key inaccessible
regardless of how many keyrings it is in.  Continuing to require Search
permission ensures we do not decrease the level of permissions required.

The problem could also be solved by requiring Write permission, but this
would arguably be less relevant because Write is supposed to deal with
the key payload, not the key itself.  A new "Invalidate" permission
would also work but would require keyctl_setperm() users to start
including a new permission which never existed before.

This is a user-visible API change.  But we don't seem to have much
choice, and any breakage will be limited to users who override the
default key permissions using keyctl_setperm() to remove Setattr
permission, then later call keyctl_invalidate().  There are probably not
many such users, possibly even none at all.  In fact, the only user of
keyctl_invalidate() I could find is nfsidmap, which will be unaffected
because it relies on the "root is permitted to invalidate certain
special keys" exception rather than the actual key permissions.

Cc: linux-api@vger.kernel.org
Cc: linux-cifs@vger.kernel.org
Cc: linux-fscrypt@vger.kernel.org
Cc: linux-nfs@vger.kernel.org
Cc: Gwendal Grignou <gwendal@chromium.org>
Cc: Jaegeuk Kim <jaegeuk@kernel.org>
Cc: Richard Weinberger <richard@nod.at>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org # v3.5+
Signed-off-by: Eric Biggers <ebiggers@google.com>
---
 Documentation/security/keys.txt | 4 ++--
 security/keys/keyctl.c          | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

Comments

David Howells April 3, 2017, 8:20 a.m. UTC | #1
Eric Biggers <ebiggers@google.com> wrote:

> Currently, a process only needs Search permission on a key to invalidate
> it with keyctl_invalidate(), which has an effect similar to unlinking it
> from all keyrings.  Unfortunately, this significantly compromises the
> keys permission system because as a result, there is no way to grant
> read-only access to keys without also granting de-facto "delete" access.
> Even worse, processes may invalidate entire _keyrings_, given only
> permission to "search" them.
> 
> Key invalidation seems to have intended for cases where keyrings are
> used as caches, and keys can be re-requested at any time by an upcall to
> /sbin/request-key.  However, keyrings are not always used in this way.
> For example, users of ext4, f2fs, and ubifs encryption may wish to grant
> many differently-privileged processes access to the same keys, set up in
> a shared keyring ahead of time.  Permitting everyone to invalidate keys
> in this scenario enables a trivial denial-of-service.  And even users of
> keyrings as "just caches" may wish to restrict invalidation because it
> may require significant work or user interaction to regenerate keys.
> 
> This patch fixes the flaw by requiring both Search and Setattr
> permission to invalidate a key rather than just Search permission.
> Requiring Setattr permission is appropriate because Setattr permission
> also allows revoking (via keyctl_revoke()) or expiring (via
> keyctl_set_timeout()) the key, which also make the key inaccessible
> regardless of how many keyrings it is in.  Continuing to require Search
> permission ensures we do not decrease the level of permissions required.

I'm not sure this is the right method either.  It might be better to allocate
one of the remaining two permissions bits for this and/or mark keys that are
invalidateable.

David
--
To unsubscribe from this list: send the line "unsubscribe linux-fscrypt" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Biggers April 4, 2017, 6:44 p.m. UTC | #2
Hi David,

On Mon, Apr 03, 2017 at 09:20:29AM +0100, David Howells wrote:
> Eric Biggers <ebiggers@google.com> wrote:
> 
> > Currently, a process only needs Search permission on a key to invalidate
> > it with keyctl_invalidate(), which has an effect similar to unlinking it
> > from all keyrings.  Unfortunately, this significantly compromises the
> > keys permission system because as a result, there is no way to grant
> > read-only access to keys without also granting de-facto "delete" access.
> > Even worse, processes may invalidate entire _keyrings_, given only
> > permission to "search" them.
> > 
> > Key invalidation seems to have intended for cases where keyrings are
> > used as caches, and keys can be re-requested at any time by an upcall to
> > /sbin/request-key.  However, keyrings are not always used in this way.
> > For example, users of ext4, f2fs, and ubifs encryption may wish to grant
> > many differently-privileged processes access to the same keys, set up in
> > a shared keyring ahead of time.  Permitting everyone to invalidate keys
> > in this scenario enables a trivial denial-of-service.  And even users of
> > keyrings as "just caches" may wish to restrict invalidation because it
> > may require significant work or user interaction to regenerate keys.
> > 
> > This patch fixes the flaw by requiring both Search and Setattr
> > permission to invalidate a key rather than just Search permission.
> > Requiring Setattr permission is appropriate because Setattr permission
> > also allows revoking (via keyctl_revoke()) or expiring (via
> > keyctl_set_timeout()) the key, which also make the key inaccessible
> > regardless of how many keyrings it is in.  Continuing to require Search
> > permission ensures we do not decrease the level of permissions required.
> 
> I'm not sure this is the right method either.  It might be better to allocate
> one of the remaining two permissions bits for this and/or mark keys that are
> invalidateable.
> 

What do you think is the right solution?  We really need to do _something_.  Can
you elaborate on what the use case(s) for "invalidate" were intended to be, and
what users of it are known?

As I noted in the commit message I was hesistant to create a new Invalidate
permission because that would require that users of keyctl_setperm +
keyctl_invalidate to start including a new permission which never existed
before.  Note that the new permission would be rejected with EINVAL by old
kernels, so applications might even need to try it both ways.

However perhaps no one cares and doing that would be a non-issue.

If furthermore the only users of "invalidate" are nfsidmap and the in-kernel
dns_resolver, then it may be fine to rename the KEY_FLAG_ROOT_CAN_INVAL flag to
something like "KEY_FLAG_INVALIDATABLE" and making it so only those keys can be
invalidated.  That would be much better, but I don't know for sure that there
aren't other users of keyctl_invalidate() (you may know more).

- Eric
--
To unsubscribe from this list: send the line "unsubscribe linux-fscrypt" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Biggers May 16, 2017, 9:49 p.m. UTC | #3
Hi David,

On Tue, Apr 04, 2017 at 11:44:33AM -0700, Eric Biggers wrote:
> On Mon, Apr 03, 2017 at 09:20:29AM +0100, David Howells wrote:
> > Eric Biggers <ebiggers@google.com> wrote:
> > 
> > > Currently, a process only needs Search permission on a key to invalidate
> > > it with keyctl_invalidate(), which has an effect similar to unlinking it
> > > from all keyrings.  Unfortunately, this significantly compromises the
> > > keys permission system because as a result, there is no way to grant
> > > read-only access to keys without also granting de-facto "delete" access.
> > > Even worse, processes may invalidate entire _keyrings_, given only
> > > permission to "search" them.
> > > 
> > > Key invalidation seems to have intended for cases where keyrings are
> > > used as caches, and keys can be re-requested at any time by an upcall to
> > > /sbin/request-key.  However, keyrings are not always used in this way.
> > > For example, users of ext4, f2fs, and ubifs encryption may wish to grant
> > > many differently-privileged processes access to the same keys, set up in
> > > a shared keyring ahead of time.  Permitting everyone to invalidate keys
> > > in this scenario enables a trivial denial-of-service.  And even users of
> > > keyrings as "just caches" may wish to restrict invalidation because it
> > > may require significant work or user interaction to regenerate keys.
> > > 
> > > This patch fixes the flaw by requiring both Search and Setattr
> > > permission to invalidate a key rather than just Search permission.
> > > Requiring Setattr permission is appropriate because Setattr permission
> > > also allows revoking (via keyctl_revoke()) or expiring (via
> > > keyctl_set_timeout()) the key, which also make the key inaccessible
> > > regardless of how many keyrings it is in.  Continuing to require Search
> > > permission ensures we do not decrease the level of permissions required.
> > 
> > I'm not sure this is the right method either.  It might be better to allocate
> > one of the remaining two permissions bits for this and/or mark keys that are
> > invalidateable.
> > 
> 
> What do you think is the right solution?  We really need to do _something_.  Can
> you elaborate on what the use case(s) for "invalidate" were intended to be, and
> what users of it are known?
> 
> As I noted in the commit message I was hesistant to create a new Invalidate
> permission because that would require that users of keyctl_setperm +
> keyctl_invalidate to start including a new permission which never existed
> before.  Note that the new permission would be rejected with EINVAL by old
> kernels, so applications might even need to try it both ways.
> 
> However perhaps no one cares and doing that would be a non-issue.
> 
> If furthermore the only users of "invalidate" are nfsidmap and the in-kernel
> dns_resolver, then it may be fine to rename the KEY_FLAG_ROOT_CAN_INVAL flag to
> something like "KEY_FLAG_INVALIDATABLE" and making it so only those keys can be
> invalidated.  That would be much better, but I don't know for sure that there
> aren't other users of keyctl_invalidate() (you may know more).
> 

I'm still thinking my proposed fix to require Setattr permission is the best
solution.  Allocating a new permission bit breaks everyone calling
keyctl_setperm() along with keyctl_invalidate(), while marking only certain keys
as invalidatable breaks everyone using keyctl_invalidate() for some "unintended"
key.

Personally, I've now seen keyctl_invalidate() used in two different places for
removing ext4 encryption keys: in Chrome OS userspace, and in a test script.
The fact is that almost everyone using the keyrings API gets confused about the
difference between revocation, invalidation, expiration, and unlinking, so they
may end up doing "invalidate" when maybe they were "supposed" to do something
else.  And it may not be a good idea to break the people who happened to choose
"invalidate", beyond what we have to do to close the security hole.

What do you say?

- Eric
--
To unsubscribe from this list: send the line "unsubscribe linux-fscrypt" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Eric Biggers May 31, 2017, 7:19 p.m. UTC | #4
On Tue, May 16, 2017 at 02:49:57PM -0700, Eric Biggers wrote:
> Hi David,
> 
> On Tue, Apr 04, 2017 at 11:44:33AM -0700, Eric Biggers wrote:
> > On Mon, Apr 03, 2017 at 09:20:29AM +0100, David Howells wrote:
> > > Eric Biggers <ebiggers@google.com> wrote:
> > > 
> > > > Currently, a process only needs Search permission on a key to invalidate
> > > > it with keyctl_invalidate(), which has an effect similar to unlinking it
> > > > from all keyrings.  Unfortunately, this significantly compromises the
> > > > keys permission system because as a result, there is no way to grant
> > > > read-only access to keys without also granting de-facto "delete" access.
> > > > Even worse, processes may invalidate entire _keyrings_, given only
> > > > permission to "search" them.
> > > > 
> > > > Key invalidation seems to have intended for cases where keyrings are
> > > > used as caches, and keys can be re-requested at any time by an upcall to
> > > > /sbin/request-key.  However, keyrings are not always used in this way.
> > > > For example, users of ext4, f2fs, and ubifs encryption may wish to grant
> > > > many differently-privileged processes access to the same keys, set up in
> > > > a shared keyring ahead of time.  Permitting everyone to invalidate keys
> > > > in this scenario enables a trivial denial-of-service.  And even users of
> > > > keyrings as "just caches" may wish to restrict invalidation because it
> > > > may require significant work or user interaction to regenerate keys.
> > > > 
> > > > This patch fixes the flaw by requiring both Search and Setattr
> > > > permission to invalidate a key rather than just Search permission.
> > > > Requiring Setattr permission is appropriate because Setattr permission
> > > > also allows revoking (via keyctl_revoke()) or expiring (via
> > > > keyctl_set_timeout()) the key, which also make the key inaccessible
> > > > regardless of how many keyrings it is in.  Continuing to require Search
> > > > permission ensures we do not decrease the level of permissions required.
> > > 
> > > I'm not sure this is the right method either.  It might be better to allocate
> > > one of the remaining two permissions bits for this and/or mark keys that are
> > > invalidateable.
> > > 
> > 
> > What do you think is the right solution?  We really need to do _something_.  Can
> > you elaborate on what the use case(s) for "invalidate" were intended to be, and
> > what users of it are known?
> > 
> > As I noted in the commit message I was hesistant to create a new Invalidate
> > permission because that would require that users of keyctl_setperm +
> > keyctl_invalidate to start including a new permission which never existed
> > before.  Note that the new permission would be rejected with EINVAL by old
> > kernels, so applications might even need to try it both ways.
> > 
> > However perhaps no one cares and doing that would be a non-issue.
> > 
> > If furthermore the only users of "invalidate" are nfsidmap and the in-kernel
> > dns_resolver, then it may be fine to rename the KEY_FLAG_ROOT_CAN_INVAL flag to
> > something like "KEY_FLAG_INVALIDATABLE" and making it so only those keys can be
> > invalidated.  That would be much better, but I don't know for sure that there
> > aren't other users of keyctl_invalidate() (you may know more).
> > 
> 
> I'm still thinking my proposed fix to require Setattr permission is the best
> solution.  Allocating a new permission bit breaks everyone calling
> keyctl_setperm() along with keyctl_invalidate(), while marking only certain keys
> as invalidatable breaks everyone using keyctl_invalidate() for some "unintended"
> key.
> 
> Personally, I've now seen keyctl_invalidate() used in two different places for
> removing ext4 encryption keys: in Chrome OS userspace, and in a test script.
> The fact is that almost everyone using the keyrings API gets confused about the
> difference between revocation, invalidation, expiration, and unlinking, so they
> may end up doing "invalidate" when maybe they were "supposed" to do something
> else.  And it may not be a good idea to break the people who happened to choose
> "invalidate", beyond what we have to do to close the security hole.
> 
> What do you say?
> 
> - Eric

David, any update on this?  This really needs to be fixed, as users actually
rely on keyring permissions to work sanely, and not have a gaping security hole.

Eric
--
To unsubscribe from this list: send the line "unsubscribe linux-fscrypt" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/Documentation/security/keys.txt b/Documentation/security/keys.txt
index 0e03baf271bd..b167f352c043 100644
--- a/Documentation/security/keys.txt
+++ b/Documentation/security/keys.txt
@@ -820,8 +820,8 @@  The keyctl syscall functions are:
      immediately, though they are still visible in /proc/keys until deleted
      (they're marked with an 'i' flag).
 
-     A process must have search permission on the key for this function to be
-     successful.
+     A process must have Search and Setattr permission on the key for this
+     function to be successful.
 
  (*) Compute a Diffie-Hellman shared secret or public key
 
diff --git a/security/keys/keyctl.c b/security/keys/keyctl.c
index 52c34532c785..c58674710251 100644
--- a/security/keys/keyctl.c
+++ b/security/keys/keyctl.c
@@ -397,7 +397,7 @@  long keyctl_revoke_key(key_serial_t id)
 /*
  * Invalidate a key.
  *
- * The key must be grant the caller Invalidate permission for this to work.
+ * The key must grant the caller Search and Setattr permission for this to work.
  * The key and any links to the key will be automatically garbage collected
  * immediately.
  *
@@ -413,7 +413,7 @@  long keyctl_invalidate_key(key_serial_t id)
 
 	kenter("%d", id);
 
-	key_ref = lookup_user_key(id, 0, KEY_NEED_SEARCH);
+	key_ref = lookup_user_key(id, 0, KEY_NEED_SEARCH | KEY_NEED_SETATTR);
 	if (IS_ERR(key_ref)) {
 		ret = PTR_ERR(key_ref);