Working with the kernel keyring
The Linux kernel keyring is effectively a mechanism to allow shoving blobs of data into the kernel and then setting access controls on them. It's convenient for a couple of reasons: the first is that these blobs are available to the kernel itself (so it can use them for things like NFSv4 authentication or module signing keys), and the second is that once they're locked down there's no way for even root to modify them.
But there's a corner case that can be somewhat confusing here, and it's one that I managed to crash into multiple times when I was implementing some code that works with this. Keys can be "possessed" by a process, and have permissions that are granted to the possessor orthogonally to any permissions granted to the user or group that owns the key. This is important because it allows for the creation of keyrings that are only visible to specific processes - if my userspace keyring manager is using the kernel keyring as a backing store for decrypted material, I don't want any arbitrary process running as me to be able to obtain those keys. As described in keyrings(7), keyrings exist at the session, process and thread levels of granularity.
This is absolutely fine in the normal case, but gets confusing when you start using sudo. sudo by default doesn't create a new login session - when you're working with sudo, you're still working with key posession that's tied to the original user. This makes sense when you consider that you often want applications you run with sudo to have access to the keys that you own, but it becomes a pain when you're trying to work with keys that need to be accessible to a user no matter whether that user owns the login session or not.
I spent a while talking to David Howells about this and he explained the easiest way to handle this. If you do something like the following:
$ sudo keyctl add user testkey testdata @u
a new key will be created and added to UID 0's user keyring (indicated by @u). This is possible because the keyring defaults to 0x3f3f0000 permissions, giving both the possessor and the user read/write access to the keyring. But if you then try to do something like:
$ sudo keyctl setperm 678913344 0x3f3f0000
where 678913344 is the ID of the key we created in the previous command, you'll get permission denied. This is because the default permissions on a key are 0x3f010000, meaning that the possessor has permission to do anything to the key but the user only has permission to view its attributes. The cause of this confusion is that although we have permission to write to UID 0's keyring (because the permissions are 0x3f3f0000), we don't possess it - the only permissions we have for this key are the user ones, and the default state for user permissions on new keys only gives us permission to view the attributes, not change them.
But! There's a way around this. If we instead do:
$ sudo keyctl add user testkey testdata @s
then the key is added to the current session keyring (@s). Because the session keyring belongs to us, we possess any keys within it and so we have permission to modify the permissions further. We can then do:
$ sudo keyctl setperm 678913344 0x3f3f0000
and it works. Hurrah! Except that if we log in as root, we'll be part of another session and won't be able to see that key. Boo. So, after setting the permissions, we should:
$ sudo keyctl link 678913344 @u
which ties it to UID 0's user keyring. Someone who logs in as root will then be able to see the key, as will any processes running as root via sudo. But we probably also want to remove it from the unprivileged user's session keyring, because that's readable/writable by the unprivileged user - they'd be able to revoke the key from underneath us!
$ sudo keyctl unlink 678913344 @s
will achieve this, and now the key is configured appropriately - UID 0 can read, modify and delete the key, other users can't.
This is part of our ongoing work at CoreOS to make rkt more secure. Moving the signing keys into the kernel is the first step towards rkt no longer having to trust the local writable filesystem. Once keys have been enrolled the keyring can be locked down - rkt will then refuse to run any images unless they're signed with one of these keys, and even root will be unable to alter them.
 (obviously it should also be impossible to ptrace() my userspace keyring manager)
 Part of our Secure Boot work has been the integration of dm-verity into CoreOS. Once deployed this will mean that the /usr partition is cryptographically verified by the kernel at runtime, making it impossible for anybody to modify it underneath the kernel. / remains writable in order to permit local configuration and to act as a data store, and right now rkt stores its trusted keys there.