On 05/02/2014 12:48 PM, Petr Spacek wrote:
On 1.5.2014 16:10, Rich Megginson wrote:
On 04/30/2014 10:19 AM, Petr Spacek wrote:
Hello list,

following text summarizes schema & DIT layout for DNSSEC key storage in LDAP.

This is subset of full PKCS#11 schema [0]. It stores bare keys with few
metadata attributes when necessary.

The intention is to make transition to full PKCS#11-in-LDAP schema [0] as easy as possible. This transition should happen in next minor version of

In theory, the transition should be just adding few object classes to
existing objects and populating few new metadata attributes. Related object
classes are marked below with "(in long-term)".

Please comment on it soon. We want to implement it ASAP :-)

- Asymmetric
- Private key is stored in LDAP as encrypted PKCS#8 blob
- Public key is published in LDAP
- Encrypted with symmetric "DNSSEC master key" (see below)
- Private key - represented as LDAP object with object classes:
ipaEPrivateKey  [1] # encrypted data
ipaWrappedKey [2] # pointer to master key, outside scope of pure PKCS#11
ipk11PrivateKey [3] (in long-term) # PKCS#11 metadata
- Public key - represented as LDAP object with object classes:
ipaPublicKey    [1] # public key data
ipk11PublicKey  [3] (in long-term) # PKCS#11 metadata

Master key
- Symmetric
- Stored in LDAP as encrypted blob
- Encrypted with asymmetric "replica key" (see below)
- 1 replica = 1 blob, n replicas = n blobs encrypted with different keys
- A replica uses it's own key for master key en/decryption
- Represented as LDAP object with object classes:
ipaESecretKey  [1]
ipk11SecretKey [3] (in long-term)

Replica key
- Asymmetric
- Private key is stored on replica's disk only
- Public key for all replicas is stored in LDAP
- Represented as LDAP object with object classes:
ipaPublicKey   [1]
ipk11PublicKey [3] (in long-term)

DIT layout
 DNSSEC key material
 - Container: cn=keys, cn=sec, cn=dns, dc=example
- Private and public keys are stored as separate objects to accommodate all
PKCS#11 metadata.
 - We need to decide about object naming:
- One obvious option for RDN is to use uniqueID but I don't like it. It is
hard to read for humans.
- Other option is to use uniqueID+PKCS#11 label or other attributes to make it more readable. Can we use multi-valued RDN? If not, why? What are
technical reasons behind it?

I would encourage you not to use multi-valued RDNs.  There aren't any
technical reasons - multi-valued RDNs are part of the LDAP standards and all conforming LDAP implementations must support them. However, they are hard to deal with - you _must_ have some sort of DN class/api on the client side to handle them, and not all clients do - many clients expect to be able to just
do dnstr.lower() == dnstr2.lower() or possibly do simple escaping.

As far as being human readable - the whole goal is that humans _never_ have to look at a DN. If humans have to look at and understand a DN to accomplish a
task, then we have failed.
I agree, users should not see them. I want to make life easier for administrators and developers *debugging* it.

I'm facing UUIDs-only logs and database in oVirt for more than year now and I can tell you that it is horrible, horrible, horrible. It is PITA when I have to debug something in oVirt because I have to search for UUIDs all the time. I want to scream and jump out of the window when I see single log line with 4 or more different UUIDs... :-)

I sympathize, having to go through this with Designate, parsing up to 4 UUIDs per debug log line . . .

Has the DogTag team reviewed this proposal? Their data storage and workflows
are similar.
That is very good point! Nathan, could somebody from DS team (maybe somebody involved in Password Vault) review this "vault without Vault"?

Thank you!

It is question if we like:
 nsUniqID = 0b0b7e53-957d11e3-a51dc0e5-9a05ecda
 nsUniqID = 8ae4190d-957a11e3-a51dc0e5-9a05ecda
more than:

 DNSSEC key metadata
 - Container (per-zone): cn=keys, idnsname=example.net, cn=dns
 - Key metadata can be linked to key material via DN or ipk11Id.
 - This allows key sharing between zones.
(DNSSEC-metadata will be specified later. That is not important for key

 Replica public keys
- Container: cn=DNS,cn=<replica FQDN>,cn=masters,cn=ipa,cn=etc,dc=example
  - or it's child object like cn=wrappingKey

 Master keys
 - Container: cn=master, cn=keys, cn=sec, cn=dns, dc=example
 - Single key = single object.
 - We can use ipk11Label or ipk11Id for naming:
 ipk11Label=dnssecMaster1, ipk11Label=dnssecMaster2, etc.

Work flows
 Read DNSSEC private key
  1) read DNSSEC private key from LDAP
  2) ipaWrappedKey objectClass is present - key is encrypted
3) read master key denoted by ipaWrappingKey attribute in DNSSEC key object
  4) use local replica key to decrypt master key
  5) use decrypted master key to decrypt DNSSEC private key

 Add DNSSEC private key
  1) use local replica key to decrypt master key
  2) encrypt DNSSEC private key with master key
  3) add ipaWrappingKey attribute pointing to master key
  4) store encrypted blob in a new LDAP object

 Add a replica
  1) generate a new replica-key pair for the new replica
  2) store key pair to replica-file (don't scream yet :-)
  4) add public key for the new replica to LDAP
  3) fetch master key from LDAP
  4) encrypt master key with new replica public key
  5) store resulting master key blob to LDAP
  6) generate a new replica-key pair (!)
  7) store new public key to LDAP
  8) remove old public key (from replica-file) from LDAP
  9) fetch master key
 10) decrypt master key using old private key (from replica-file)
 11) encrypt master key using new private key (generated locally)
 12) replace old master key blob in LDAP with new blob (from step 11)

 Delete a replica
This is the tricky part. New master key has to be generated on some other
replica. What should we do if the ipa-replica-manage command was run on
deleted replica?

I propose to split replica master key roll-over to two phases:
 Any machine in IPA domain (including to-be deleted replica):
  1) Delete public key associated with replica from LDAP
  2) Flip a bit in master key metadata and say "this key needs to be
     (Maybe we can disable ipk11Wrap boolean to indicate that this key
should not be used for key wrapping.)

 Remaining replicas:
  3) Periodically check that master key is obsolete
  4) Wait for (random period of time) to limit probability of collision
5) Check that master key is really obsolete and new one is not present
  6) Generate a new master key
  7) Encrypt new master key with all replica-public-keys stored in LDAP
  8) Store new master key blobs to a new LDAP object
(Conflicts are not a problem up to now because we are not deleting old
key. In worst case, we will have multiple new master keys.)
*What should we do now?*
9) ??? Re-encrypt all DNSSEC keys with a new master key? (What if we have
write conflict now?)
??? Let old keys there and wait until key rotation mechanism replaces all old DNSSEC keys with new DNSSEC keys encrypted with a new master key (~
one year)?
10) Old master key can be deleted when no other object is referencing to it.

Congratulations to people who reached this line and didn't skip anything :-)

[0] http://www.freeipa.org/page/V4/PKCS11_in_LDAP/Schema
[1] http://www.freeipa.org/page/V4/PKCS11_in_LDAP/Schema#Encoded_key_data_2

[3] http://www.freeipa.org/page/V4/PKCS11_in_LDAP/Schema#Storage_objects

Freeipa-devel mailing list

Reply via email to