On 26 Sep 2019, at 00:58, Mark Reynolds mreynolds@redhat.com wrote:

On 9/25/19 10:43 AM, Eugen Lamers wrote:

...

I'm trying to setup a replication with a certificate based authentication between supplier and consumer. The certificates in the certdb at /etc/dirsrv/slapd-XXX contain the very same CA with which the respective server certificates in the certdbs have been signed. The certificates all have the 'u' flag, and the CA has the C and T flag. The replication (on the supplier) has been setup such that TLS and certificate based authentication is used, see extract from the replication agreement object:

objectclass: nsds5ReplicationAgreement nsds5replicahost: <consumer-hostname> nsds5replicaport: 389 nsds5replicatransportinfo: TLS nsds5replicabindmethod: SSLCLIENTAUTH

Trying to initialize the consumer raises this error in the error-log of the supplier (the host sending the starttls connection request): Replication bind with EXTERNAL auth failed: LDAP error 48 (Inappropriate authentication) (missing client certificate)

Replication does not use the server's certificate for SSL Client Authentication. It uses the client certificate in the "Replication Manager" entry. Sounds like you did not add a user certificate to this entry.

Look at this for help with adding a client certificate to a user entry:

https://access.redhat.com/documentation/en-us/red_hat_directory_server/10/ht...

Note - the client certificate you use in the replication manager but be issued by the same CA that the all the replicas are using.

I think there is also a test case somewhere in the server that uses TLS auth for the replication manager

https://pagure.io/389-ds-base/blob/master/f/dirsrvtests/tests/suites/replica...

It might help give you some ideas on how this works,

HTH,

Mark

...

The certificate that the server should have sent can, however, be used with the ldap commandline tools as ldapsearch. In this case a wireshark trace clearly shows that the client sends the certificate during the TLS handshake, while in the above case it doesn't. The TLS handshake defines that the client has to send an "empty certificate" if it does not have a certificate that has been issued by a CA offered by the server during the handshake. I can't see a reason for the client to think the condition isn't met. The certificate with which the server (supplier) is setup is the only one available. Is it possible that the server does not know which certificate it has to use for authentication with the consumer server? And if so, how do I make the server know?

The 389-ds in use is the version 1.4.1.6~git0.5ac5a8aad. The problem was the same with 1.4.0.3.

Thanks, Eugen _______________________________________________ 389-users mailing list -- 389-users@lists.fedoraproject.org To unsubscribe send an email to 389-users-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/389-users@lists.fedoraproject....

--

389 Directory Server Development Team _______________________________________________ 389-users mailing list -- 389-users@lists.fedoraproject.org To unsubscribe send an email to 389-users-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/389-users@lists.fedoraproject....

— Sincerely,

William Brown

Senior Software Engineer, 389 Directory Server SUSE Labs

On 26 Sep 2019, at 19:30, Eugen Lamers eugen.lamers@br-automation.com wrote:

Hi William,

...

I think there is also a test case somewhere in the server that uses TLS auth for the replication manager

https://pagure.io/389-ds-base/blob/master/f/dirsrvtests/tests/suites/repl...

unfortunately we updated to the 1.4.1er 389ds which doesn't have the lib389 python module any longer. What I tried to extract from the source code is the setting of "default:CmapLdapAttr nsCertSubjectDN" in the certmap.conf. It didn't help either.

Can you send in your certmap.conf and any of the relevant parts of dse.ldif so I can have a look at what you configured?

Also 1.4.1 *should* have lib389, so I'm curious what distro and platform you are using? It should be present ....

Regards, Eugen _______________________________________________ 389-users mailing list -- 389-users@lists.fedoraproject.org To unsubscribe send an email to 389-users-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/389-users@lists.fedoraproject....

— Sincerely,

William Brown

Senior Software Engineer, 389 Directory Server SUSE Labs

On 27 Sep 2019, at 17:42, Eugen Lamers eugen.lamers@br-automation.com wrote:

Hi William,

...

Can you send in your certmap.conf and any of the relevant parts of dse.ldif so I can have a look at what you configured?

dn: cn=encryption,cn=config objectClass: top objectClass: nsEncryptionConfig cn: encryption nsSSLSessionTimeout: 0 nsSSLClientAuth: allowed CACertExtractFile: /etc/dirsrv/slapd-hostname1_local/ca.pem numSubordinates: 1

dn: cn=RSA,cn=encryption,cn=config objectClass: top objectClass: nsEncryptionModule cn: RSA nsSSLPersonalitySSL: hostname1.domain.com nsSSLActivation: on nsSSLToken: internal (software)

dn: cn=replica,cn=dc\3Ddomain\2Cdc\3Dcom,cn=mapping tree,cn=config objectClass: top objectClass: nsds5replica objectClass: extensibleObject cn: replica nsDS5ReplicaRoot: dc=domain,dc=com nsDS5ReplicaId: 12 nsDS5ReplicaType: 3 nsDS5Flags: 1 nsds5ReplicaPurgeDelay: 604800 nsDS5ReplicaName: b38325b1-e0f011e9-9cdcc130-abdc2534 numSubordinates: 1

dn: cn=hostname1,cn=replica,cn=dc\3Ddomain\2Cdc\3Dcom,cn=mapping tree,cn=config objectClass: top objectClass: nsds5ReplicationAgreement nsDS5ReplicaHost: hostname2.domain.com nsDS5ReplicaPort: 389 nsDS5ReplicaTransportInfo: TLS nsDS5ReplicaBindMethod: SSLCLIENTAUTH nsDS5ReplicaRoot: dc=domain,dc=com description: Agreement between hostname2 and hostname1 cn: hostname1

Don't know if this extract contains what you need to know...

...

Also 1.4.1 *should* have lib389, so I'm curious what distro and platform you are using? It should be present ....

Problem is, we currently use a 389ds version that is higher than the one "included" in our distro, which is SLE 12 SP3. We need to clarify on our side if the mix of packages (389ds, console, etc.) is consistent...

That shouldn't be a problem I think ....

Would you recommend, as I understood Mark Reynolds, to use a "replication manager" entry on both servers and put the very same certificate to it with a appropriate subjectDN like "cn=replication manager"? Currently I use, as stated in my other mail, a setup as recommended in the admin guide of redhat ds 10, which is a supplier DN with the "cn=<fq-serverhostname>" on the consumer, and to which the certificate of the supplier server is put. But the supplier does not send any certificate to the consumer in the TLS handshake. So I don't this the certmap is involved, since it is used on the receiver of the certificate to decide to which entry the certificate should be bound, but the problem occurs already on the sender side.

Anyway, the certmap:

certmap default default default:DNComps default:FilterComps cn default:verifycert off

Adding "default:CmapLdapAttr nsCertSubjectDN" didn't change anything.

I think you need the replication manager entries. So to explain. When the TLS connection is made by hostname1 to hostname2, on hostname2 the SASL mech of EXTERNAL is provided, so the bind code looks at other possible authentication factors. The client certificate provided is then inspected.

The hostname1 client certificate must be signed by a CA that is accepted and trusted for client auth in the nssdb on hostname2. You can check this with certutil

{root@ldapkdc 9:48} /data/config I0> certutil -L -d .

Certificate Nickname Trust Attributes SSL,S/MIME,JAR/XPI

Self-Signed-CA CT,, Server-Cert u,u,u

Notice the "CT,,". That means it's trusted as a server and client CA. You can see the trust flags from -A:

{root@ldapkdc 9:49} /data/config I0> certutil -A --help ... -t trustargs Set the certificate trust attributes: trustargs is of the form x,y,z where x is for SSL, y is for S/MIME, and z is for code signing. Use ,, for no explicit trust. p prohibited (explicitly distrusted) P trusted peer c valid CA T trusted CA to issue client certs (implies c) C trusted CA to issue server certs (implies c) u user cert w send warning g make step-up cert

And then modify them with -M:

{root@ldapkdc 9:48} /data/config I0> certutil -M --help -M Modify trust attributes of certificate -n cert-name The nickname of the cert to modify -t trustargs Set the certificate trust attributes (see -A above) -d certdir Cert database directory (default is ~/.netscape) -P dbprefix Cert & Key database prefix

So back to TLS processing - now the server has validated the client cert, we head into the certmap code - we *only* have the certificate so we need to be able to look up that certificate to a user somehow. Certmap.conf is processed in a specific order (that I can't remember right now :) ). One of the simplest methods is "DNComp", where the DN of the cert must match and be in order of the DN in the directory.

https://access.redhat.com/documentation/en-us/red_hat_directory_server/10/ht...

But if you set certmap to look like:

certmap default default default:CmapLdapAttr nsCertSubjectDN

So this means: when an incomming SASL external is found, and you have TLS, take the certificate subjectDN and do an ldapsearch for it in each available backend.

This has a specific implication though. You need an entry that looks like:

dn: cn=hostname1,cn=config cn: hostname1 nsCertSubjectDN: <subject of hostname1 cert>

Each server will need these entries for all possible incoming servers.

You can also make these entries in your main database such as:

cn=hostname1,ou=replicationmanagers,dc=example,dc=com cn: hostname1 nsCertSubjectDN: <subject of hostname1 cert>

The benefit of this is you have the replication managers in "one place" and they are literally replicated, so you don't need so much per-server configuration. The issue is a chicken-and-egg bootstrap where you can't start replication until ... you start replication. So youcan bootstrap either with PW auth on the repl manager, or you can do a db2ldif -r on the first master, and import that on the second so the replication metadata is available.

I hope this helps: in general I think TLS auth on replication is really hard to setup and do properly. I did it as part of lib389 to automate the tests and it was a challenging process to setup, and that's coming from someone who is literally a developer on the project. It's also super dense/hard to debug when it goes wrong.

I prefer to either have the simple replication manager + pw setup, or have the per-host replication manager in the maindb but with pw again.

Anyway, hope that helps give you advice on what you may be missing :)

Regards, Eugen _______________________________________________ 389-users mailing list -- 389-users@lists.fedoraproject.org To unsubscribe send an email to 389-users-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/389-users@lists.fedoraproject....

— Sincerely,

William Brown

Senior Software Engineer, 389 Directory Server SUSE Labs

On 30 Sep 2019, at 16:48, Eugen Lamers eugen.lamers@br-automation.com wrote:

Hello William,

...

I think you need the replication manager entries. So to explain. When the TLS connection is made by hostname1 to hostname2, on hostname2 the SASL mech of EXTERNAL is provided, so the bind code looks at other possible authentication factors. The client certificate provided is then inspected.

This obviously is my main problem: The client (supplier in this case) does not provide a certificate. wireshark confirms this. What I wonder is how to tell the supplier which certificate to use or where to find it.

Is your Server-Cert marked with trust flags u? see:

...

The hostname1 client certificate must be signed by a CA that is accepted and trusted for client auth in the nssdb on hostname2. You can check this with certutil

{root@ldapkdc 9:48} /data/config I0> certutil -L -d .

Certificate Nickname Trust Attributes SSL,S/MIME,JAR/XPI

Self-Signed-CA CT,, Server-Cert u,u,u

^ here. Server-Cert needs u,, at least (I think the u,u,u may be incorrect actually).

...

Notice the "CT,,". That means it's trusted as a server and client CA. You can see the trust flags from -A:

{root@ldapkdc 9:49} /data/config I0> certutil -A --help ... -t trustargs Set the certificate trust attributes: trustargs is of the form x,y,z where x is for SSL, y is for S/MIME, and z is for code signing. Use ,, for no explicit trust. p prohibited (explicitly distrusted) P trusted peer c valid CA T trusted CA to issue client certs (implies c) C trusted CA to issue server certs (implies c) u user cert w send warning g make step-up cert

And then modify them with -M:

{root@ldapkdc 9:48} /data/config I0> certutil -M --help -M Modify trust attributes of certificate -n cert-name The nickname of the cert to modify -t trustargs Set the certificate trust attributes (see -A above) -d certdir Cert database directory (default is ~/.netscape) -P dbprefix Cert & Key database prefix

All these conditions are met already.

...

So back to TLS processing - now the server has validated the client cert, we head into the certmap code - we *only* have the certificate so we need to be able to look up that certificate to a user somehow. Certmap.conf is processed in a specific order (that I can't remember right now :) ). One of the simplest methods is "DNComp", where the DN of the cert must match and be in order of the DN in the directory.

https://access.redhat.com/documentation/en-us/red_hat_directory_server/10...

But if you set certmap to look like:

certmap default default default:CmapLdapAttr nsCertSubjectDN

So this means: when an incomming SASL external is found, and you have TLS, take the certificate subjectDN and do an ldapsearch for it in each available backend.

This has a specific implication though. You need an entry that looks like:

dn: cn=hostname1,cn=config cn: hostname1 nsCertSubjectDN: <subject of hostname1 cert>

Each server will need these entries for all possible incoming servers.

You can also make these entries in your main database such as:

cn=hostname1,ou=replicationmanagers,dc=example,dc=com cn: hostname1 nsCertSubjectDN: <subject of hostname1 cert>

The benefit of this is you have the replication managers in "one place" and they are literally replicated, so you don't need so much per-server configuration. The issue is a chicken-and-egg bootstrap where you can't start replication until ... you start replication. So youcan bootstrap either with PW auth on the repl manager, or you can do a db2ldif -r on the first master, and import that on the second so the replication metadata is available.

It shouldn't be a problem to create these entries in the db of the consumer, but since the consumer doesn't see a certificate coming from the supplier, there is nothing to be mapped or bound.

Yep. See above.

...

I hope this helps: in general I think TLS auth on replication is really hard to setup and do properly. I did it as part of lib389 to automate the tests and it was a challenging process to setup, and that's coming from someone who is literally a developer on the project. It's also super dense/hard to debug when it goes wrong.

I prefer to either have the simple replication manager + pw setup, or have the per-host replication manager in the maindb but with pw again.

I will take your advice into consideration, even if I'm somehow going to be successful with the challenge. Is there any way to avoid plaintext passwords for the replication manager entries in LDIF files or the setup.inf for the setup-ds(-admin).pl scripts, I mean other than entering the passwords within the 389-console?

Well, they all end up in dse.ldif anyway. Because of the replication manager's nature it well ... has to be reversible to the process. We need to get the plaintext password somehow to initiate the bind. Additionally, any DN listed in a replication agreement as a replication manager has ACI's not evaluated in many cases due to needing a high level of access to perform it's function.

Saying this you need to say what's your threat model - if someone has access to dse.ldif and the content within in they can easily change the ldapi-autobind flag, then change directory content anyway. They may also have access to pin.txt and the nss certdb so they have your private key you're using for auth for replication anyway .... so is cert really more secure than a pw in dse.ldif? The threat is "someone has access to the machine and /etc as root/dirsrv". Once someone has access to one master in a replication topoolgy, they also can add entries which would give you root access/persistence on other hosts, so attacking replication is a low value target. You'd just put another user into the administrators group in the directory so you can ssh to whatever you want.

If you are worried about MITM, TLS already protects you here as does the proper CA verification which replication will already be doing for you.

I think that will give you some more to thing about in you decision making and risk assessment process. Very happy to discuss what your risks and threats are so we can work out what's going to be practical.

Thanx for you effort. Best, Eugen _______________________________________________ 389-users mailing list -- 389-users@lists.fedoraproject.org To unsubscribe send an email to 389-users-leave@lists.fedoraproject.org Fedora Code of Conduct: https://docs.fedoraproject.org/en-US/project/code-of-conduct/ List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines List Archives: https://lists.fedoraproject.org/archives/list/389-users@lists.fedoraproject....

— Sincerely,

William Brown

Senior Software Engineer, 389 Directory Server SUSE Labs

On 1 Feb 2021, at 23:35, Eugen Lamers eugen.lamers@br-automation.com wrote:

Hi William, it's an old thread, but it's mine, so I will give an update to the situation and a follow-up question. We changed from StartTLS on 389 to SSL on 636 some time ago. Trying to reconsider the topic we found that there is no plaintext password sent via network between the replicants, which was the case in the StartTLS on 389 scenario. This would reduce the problem with plaintext password for the replication manager to the storage, mainly the dse.ldif, I think.

To be sure, I'd need to see your configured replication agreement to understand how it's been configured to authenticate.

Provided you are using SSL (LDAPS) and simple bind, then the password *is* sent to the other server, but it's inside of TLS so it is secure.

We would be glad if you could confirm this thought. It would save us from trying again the use of client auth for replication which hadn't been successful yet. Kind regards,

— Sincerely,

William Brown

Senior Software Engineer, 389 Directory Server SUSE Labs, Australia