[
https://issues.apache.org/jira/browse/HADOOP-13558?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15451663#comment-15451663
]
Steve Loughran commented on HADOOP-13558:
-----------------------------------------
I deny all understanding of how Kerberos works, except for a deep fear of the
UGI class. Maybe [[email protected]] could look at it from the perspective of
somebody who understands this stuff
# I don'tt understand why way the current/new code reuses the {{realuser}}
field. Unless the act of creating the UGI has side effects, the first
assignment is a no-op. If it is done for side effects, comments should declare
this, an separate field used. now is the time to do this.
# test-wise, before the patch to UGI goes in —did the new test case fail? As
that's a good sign that the test can recreate the problem and show it is fixed.
# If possible, it'd be good to extend KDiag with more info here
# This really ought to go through a full build and test run against a
kerberized cluster. For example: can a version of Hadoop built with this auth
with HDFS using keytabs as well as tickets. Volunteers?
> UserGroupInformation created from a Subject incorrectly tries to renew the
> Kerberos ticket
> ------------------------------------------------------------------------------------------
>
> Key: HADOOP-13558
> URL: https://issues.apache.org/jira/browse/HADOOP-13558
> Project: Hadoop Common
> Issue Type: Bug
> Components: security
> Affects Versions: 2.7.2, 2.6.4, 3.0.0-alpha2
> Reporter: Alejandro Abdelnur
> Attachments: HADOOP-13558.01.patch
>
>
> The UGI {{checkTGTAndReloginFromKeytab()}} method checks certain conditions
> and if they are met it invokes the {{reloginFromKeytab()}}. The
> {{reloginFromKeytab()}} method then fails with an {{IOException}}
> "loginUserFromKeyTab must be done first" because there is no keytab
> associated with the UGI.
> The {{checkTGTAndReloginFromKeytab()}} method checks if there is a keytab
> ({{isKeytab}} UGI instance variable) associated with the UGI, if there is one
> it triggers a call to {{reloginFromKeytab()}}. The problem is that the
> {{keytabFile}} UGI instance variable is NULL, and that triggers the mentioned
> {{IOException}}.
> The root of the problem seems to be when creating a UGI via the
> {{UGI.loginUserFromSubject(Subject)}} method, this method uses the
> {{UserGroupInformation(Subject)}} constructor, and this constructor does the
> following to determine if there is a keytab or not.
> {code}
> this.isKeytab = KerberosUtil.hasKerberosKeyTab(subject);
> {code}
> If the {{Subject}} given had a keytab, then the UGI instance will have the
> {{isKeytab}} set to TRUE.
> It sets the UGI instance as it would have a keytab because the Subject has a
> keytab. This has 2 problems:
> First, it does not set the keytab file (and this, having the {{isKeytab}} set
> to TRUE and the {{keytabFile}} set to NULL) is what triggers the
> {{IOException}} in the method {{reloginFromKeytab()}}.
> Second (and even if the first problem is fixed, this still is a problem), it
> assumes that because the subject has a keytab it is up to UGI to do the
> relogin using the keytab. This is incorrect if the UGI was created using the
> {{UGI.loginUserFromSubject(Subject)}} method. In such case, the owner of the
> Subject is not the UGI, but the caller, so the caller is responsible for
> renewing the Kerberos tickets and the UGI should not try to do so.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
