Bug #74636 [Com]: [krb5] negotiate auth broken for vhosts because of hostname canonicalization

Edit report at https://bugs.php.net/bug.php?id=74636&edit=1 ID: 74636 Comment by: mbechler at eenterphace dot org Reported by: chanlists at googlemail dot com Summary: [krb5] negotiate auth broken for vhosts because of hostname canonicalization Status: Assigned Type: Bug Package: PECL Operating System: debian 8 PHP Version: 5.6.30 Assigned To: mbechler Block user comment: N Private report: N New Comment: Passing GSS_C_NO_NAME or GSS_C_NO_CREDENTIAL to accept_context respecively has some (possible acceptable) downsides: a) it invokes implementation dependant behavior (spec essentially says implementations can choose whichever keys they like) - but the common and more recent implementations (e.g. mit since 1.7 - which is pretty old) seem have the desired behavior (trying all keys). b) this supposedly will also try the default keytab, which means that anyone with access to one of these other keys (if there are any) can forge tickets for your service. That might be an subtle security issue in some setups. If I were using that personally, I would probably rather go for either redirecting to a canonical host or constructing and passing the proper SPN. But I think I would be okay with adding an option to enable that behavior. Previous Comments: ------------------------------------------------------------------------ [2017-05-24 09:32:21] chanlists at googlemail dot com Thinking a bit more about it - now that dns_canonicalize_hostname is no longer required on the server, we cannot really control what the client has set. Right now, gss_acquire_cred is called like status = gss_acquire_cred(&minor_status, object->servname, 0, GSS_C_NO_OID_SET, GSS_C_ACCEPT, &server_creds, NULL, NULL); It would be great if there were an option to replace obj->servname in the call by GSS_C_NO_NAME. My understanding is that in this case, any principal in the keytab should work. So I could stick HTTP/<vhost_name> and HTTP/<real_hostanme> and HTTP/<short_name> in the keytab, and all should work... What do you think? Happy to test... Christian ------------------------------------------------------------------------ [2017-05-23 22:33:46] chanlists at googlemail dot com Arrgh... I mean principal, not principle... ------------------------------------------------------------------------ [2017-05-23 22:32:24] chanlists at googlemail dot com I can confirm that previously dns canonicalization had to be turned off. With your new code, if I specify the service principle as an argument to the constructor, it works regardless of the dns_canonicalize_hostname setting in /etc/krb5.conf. This seems to resolve the issues I raised. Thanks a lot for fixing it! Cheers, Christian ------------------------------------------------------------------------ [2017-05-23 11:28:33] mbechler at eenterphace dot org Regarding the canonicalization, really looks like GSS_C_NT_HOSTBASED_SERVICE name types may be canonicalized. Can you verify that it is fixed by using GSS_KRB5_NT_PRINCIPAL_NAME and passing "HTTP/%s" instead. Downside might be that domain->realm is probably not applied (therefor always uses the default realm), so I would probably leave it as is for the "magic" case. Just commited a change adding the argument and parsing the name in that case with GSS_KRB5_NT_PRINCIPAL_NAME, that way the default would be the best guess we are able to make while allowing total control when overriding it. ------------------------------------------------------------------------ [2017-05-23 10:03:35] mbechler at eenterphace dot org Absolutely right, that is missing and I had noticed that before and even thought that I had added it - but that does not seem to be the case. Adding a constructor argument to KRB5NegotiateAuth for overriding the SPN sounds good? I think that would be the most obvious choice. Do you really need to disable "dns_canonicalize_hostname" (except for the client if it is run on the same host)? It would really surprise me if the library did perform hostname canonicalization on the server name for acceptor credentials. If you provide a patch, I'll be happy to merge that. ------------------------------------------------------------------------ The remainder of the comments for this report are too long. To view the rest of the comments, please view the bug report online at https://bugs.php.net/bug.php?id=74636 -- Edit this bug report at https://bugs.php.net/bug.php?id=74636&edit=1