OpenSMTPD advisory dissected
poolp.org
poolp.org
* Five years ago a change causes a logic bug in a not super critical section of code.
* Three and a half years later a change in an entirely different section of code makes the section of code super critical at the same time it makes it exploitable.
Bugs can happen in any of the places you are not looking without actually changing anything in those places. Your code is full of trip mines and you don't know where any of the wires are.
On OpenBSD, mailboxes are mode 600. So, only root can write to them. On Debian with exim4, they are mode 660 -- writable by group "mail." So the MDA only needs to run as group mail, not user root.
Is there some reason why OpenBSD is not doing the same thing?
Using a database doesn't require a privileged daemon, and the worst-case scenario (assuming a minimal threshold of competence) is that any attacker can only screw up the database in some way. With a privileged daemon, screwing up generally means that the attacker can potentially gain full control of your hardware, with all the nastiness that entails. Including, at a minimum, read your email or surreptitiously delete them.
On assuming minimal threshold of competence.. you'd probably want to assume that nobody today plasters strings (while trying to escape the untrusted bits) to make SQL queries; many a database have been compromised this way. Here we are, looking at OpenSMTPD which did essentially the same: pass untrusted string to shell. Whoops. For anyone whose production email system were subject to such an issue, it's bad, really bad, whether it's an SQL query going against unprivileged database with all your data in it, or string going to root shell.
I'm not terribly concerned about a piece of a daemon running as root. It's not the entire daemon, just a small privileged helper performing a tiny set of functions. I'd like to see it go, and I think it could go if we compromise on some features, change the way MDAs are invoked, or use a different permission model. The last one would probably be a relatively small change and allow existing tools to work (perhaps with small changes).
I don't want to introduce a database and a daemon to store messages into it and another to let me read them. That's just adding more code, I want less code. I want a better security model, not shuffling responsibility away into a different daemon that needs to be maintained and audited and protected from a subtle big whoops.
Disconnecting the concept of an email user from a system user definitely improves the situation and is unlikely to introduce new failure points. That's how the biggest email providers work - you don't expect gmail to create a new uid for you when you create an account, right?
As for opensmtpd, it should probably only reject, drop or hand off mail via LMTP. Then another daemon should deal with storage and provide access to mail.
You might look at dbmail.org if you want mail storage in PostgreSQL with IMAP support, for example.
Dovecot can handle incoming mail via LMTP, but needs roit to write to mailboxes, unfortunately.
I suppose you could give each user a suid binary to handle incommming email... Not sure if that'd be a great idea, though.
Ed: ironically, a change along these lines resulted in this bug. Maybe it would be better if opensmtpd only spoke LMTP, and never did any launching of a shell. Conceivably the LMTP daemon might have a similar bug, though..
It was spawning such a process that actually was the problem in this case.
I'm still a little baffled that things could go so wrong in handling input/parameters to system()... But it did.
I'm not saying you're wrong - it's just that this concrete error happened in an odd place, as a result of changing to pretty much the architecture you mention (and only for mbox file delivery).
For one ironic example: Thunderbird is better at picking up local mail from a local IMAP server than it is at talking directly to a ~/Maildir/ or a ~/Mailbox .
Files are easier to backup, version and transparently encrypt. Try syncing a large database continuously with Google Drive, vs being able to sync new files. Files can be read by simply opening them in a text editor, just like email should. Email clients shouldn't have to connect to a database just to view your email.
Linux solved authentication and permissions before MySQL.
For example, locking a file is a joke in POSIX. On Windows, locking a file will keep other processes from being able to read or write it, but POSIX will only keep other processes from being able to lock it. And it doesn't work on NFS anyways, so you have to use lock files in practice if you care about potentially working on NFS.
This of course means that backing up files while running a process that might be using them is extremely dangerous--there's no guarantee that the backup system will capture the filesystem in a consistent state. A proper database will come with its own backup utility that is capable of backing up databases without taking them offline.
> Linux solved authentication and permissions before MySQL.
POSIX file permissions are a joke. You get to define just three levels of permissions (user, group, and everybody), and they apply coarsely to an entire file. Database permissions can define permissions for multiple users or sets of users, and some of them will even let you specify permissions on a per-record basis.
I've worked with email clients, so I know how they're implemented, and how the protocols work, and how (some) servers work. Email needs to be stored in a database; it's what users expect in their UX; it's what protocol implementations require. Building that database on top of a file system might have been best practice in 1990, but it's 2020 now, and we know that filesystem-based databases are a disaster waiting to happen, so you'd be an idiot to do so today.
Then why do you still need a file system for your database files? Windows registry is a good example of how storing configurations in a database is a bad idea.
> For example, locking a file is a joke in POSIX.
I don't even see how that applies here, but you are talking about a bunch of different software vs a single database client used for a single purpose. There are mandatory locks if enabled and other solutions, but generally this isn't a problem unless you have unknown software accessing the same data as your program.
> A proper database will come with its own backup utility that is capable of backing up databases without taking them offline.
A file that is being written to hasn't been saved, therefore I would expect to only get a snapshot of the previous version. I don't see how that again applies to emails, often saved as individual files.
> POSIX file permissions are a joke. You get to define just three levels of permissions (user, group, and everybody), and they apply coarsely to an entire file.
There are ACLs.
> Email needs to be stored in a database
No it doesn't. It is best to cache metadata, but email has traditionally not been saved in a database and there are lot of benefits of saving emails as files that can be read from the file system. It is much easier to copy and sync files that store relevant information, than to take continuous versioned backups of a database that stores gigabytes of data.
And the thing you cache it in is called a database. You may choose to eschew an SQL database, or a no-SQL database, and roll your own implementation of a database, but you are building a database in some fashion.
> there are lot of benefits of saving emails as files that can be read from the file system
And now you're using the filesystem as key-value store, aka, database. Except since you're preserving metadata in a distinct database, now you have to maintain ACID qualities across two databases, one of which is pretty notorious for lying to you about durability. Or you can recognize that you're probably not very good at actually dealing with databases, and delegate the task to a proper database who has people on staff who actually worry about these sorts of scenarios.
> generally this isn't a problem unless you have unknown software accessing the same data as your program.
The most common such program is... backup software. Especially the kind that says backups are "copy all files in these directories," which I'm guessing is the kind of backup software that you advocate for.
But you'll find in general that many applications have people who'll want to read your innards. If you're a browser, your cookie jar is valuable; an email client, your email storage and address book.
> It is much easier to copy and sync files that store relevant information
It is easier to write such an application. It is harder to make such an application correct. To give a concrete example of where things can go wrong: Suppose I'm running my email client when the backup process is running. The backup process finishes copying the contents of my mail folder, and my email client gets a new message, which it dutifully files into that folder. This creates a new file in said folder, and also updates the cached metadata file. Now the backup process indexes the metadata file. Were I to try to restore from this backup, the resulting state is inconsistent, and the message may well be gone forever.