Timo Sirainen wrote:

I wrote a page about this in http://wiki.dovecot.org/moin.cgi/UserIds

Quoting from http://wiki.dovecot.org/moin.cgi/UserIds:

2. Dovecot user dovecot user...is only used while Dovecot is processing user logins [...]
3. Authentication process UID auth_user setting...only specifies the UID under which dovecot-auth process is run. ...you might want to create a separate dovecot-auth user to use here if you're not authenticating with PAM or shadow file which require roots.

This and other documentation seems to suggest that the 'dovecot' user ID shouldn't be reused as the auth_user setting. But based on how you have described the purpose of the two IDs, they appear to be serving a similar purpose, and thus it might make sense to use the same ID.

Could you clarify in the documentation whether this is an acceptable practice, or if not, why?

A scenario in which I could see there being an advantage is one in which the network process (running as 'dovecot') is compromised, and the hacker would then be unable to access the password database due to it providing access only to the UID specified by auth_user. Although it seems that additional security steps are really necessary in order for that to be true. In a typical SQL database setup, the database won't enforce system UID access restrictions, so it'll be up to the administrator to make sure the Dovecot database configuration file (say dovecot-mysql.conf) is not generally readable.

Which raises the question, how restrictive can permissions be on those configuration files? Are they read by Dovecot while root? Or will they be read by the authentication process while running as auth_user? If they can be readable only by root, and the administrator takes advantage of that, then it seems there would be little advantage for auth_user != login_user because any non-root account wouldn't be able to read the config file.

I see my dovecot-mysql.conf is world readable. Oops. Probably not a good thing, given it contains a MySQL password in clear text. You might want to put a security check into the code somewhere for that, in the same way it is common for MTAs to check on the permissions of files they depend on.

-Tom

Timo Sirainen wrote:

...

This and other documentation seems to suggest that the 'dovecot' user ID shouldn't be reused as the auth_user setting.

The point is more about making sure that the untrusted login process can't go directly poking into trusted processes. ... I want to keep login processes in as restricted sandbox as possible.

OK. Something that would be helpful would be a block diagram or table of the various Dovecot processes showing their purpose and privilege. Something like:

process: dovecot user: root purpose: parent process/master daemon

process: imap-login user: [typically 'dovecot'] purpose: handles network interaction for login

process: dovecot-auth user: [typically 'root'] purpose: communicates with user and password databases

process: imap-login user: [typically 'dovecot'] purpose: handles network interaction for login

process: imap user: (obtained from the user database) purpose: implements the IMAP protocol; interacts with the user's mailboxes

etc.

...

Which raises the question, how restrictive can permissions be on those configuration files? Are they read by Dovecot while root? Or will they be read by the authentication process while running as auth_user? I see my dovecot-mysql.conf is world readable. Oops. Probably not a good thing, given it contains a MySQL password in clear text.

Looks like currently the config files are read as auth_user. I changed them now to be read before dropping privileges.

Ah...I better tweak my permissions then. I had switched the MySQL config file to be readable only by root, and hadn't noticed any problems, but that's because my dovecot-auth hasn't been restarted since the change.

...

You might want to put a security check into the code somewhere for that, in the same way it is common for MTAs to check on the permissions of files they depend on.

Hmm. I guess it could be useful.

It could always be relegated to an external script if you'd prefer not to muddy up the code. Though that might require some duplication of effort, such as parsing the config files.

BTW. I'm pretty lazy at updating documentation. ...I wouldn't have minded if you had just changed them directly in Wiki :)

I considered that. After all, that's the point of using a Wiki. Though I figured it'd be a bit more polite to put the ideas out there for discussion before making the changes.

BTW, I have a Dovecot-MySQL howto about half done. I'll finish it up once I'm done with a Postfix project and have time to get back to it.

-Tom