Re: Modular BIG_KEYS (was: Re: [PATCH v4] security/keys: rewrite all of big_key crypto)

[David Howells <dhowells@redhat.com>]

Eric Biggers <ebiggers3@gmail.com> wrote:

> (I also still need to convince myself that there aren't any race conditions
> in key type unregistering.  It's a little weird how it changes the key type
> to the ".dead" key type, rather than pinning the key type in memory while
> it's still used.)

Keys are converted to the dead type and their destructors called by the gc
whilst it holds a write lock on the key's semaphore, so keyctl() calls
shouldn't be a problem as they all hold a read-lock or a write-lock on the
content.

/proc/keys does things under the key_serial_lock and RCU, so if a key type's
->describe() function looks at the payload, then it should use RCU conditions.
Note that this doesn't affect the keyring_describe() as that only looks at a
value that is stored in the key struct - and the same for user_describe() and
big_key_describe().  asymmetric_key_describe(), OTOH...  Maybe the simplest
thing to do is to take key_serial_lock in key_garbage_collector() around the
code that changes the key type.

Inside the kernel, things are murkier as the kernel can reach inside a key
without taking the semaphore.  If it does so, it must use RCU or some other
mechanism to prevent the key payload from being destroyed under it.

With AF_RXRPC, I think I'm doing this wrong.  I think I need to move the
destruction of the key type to after I've unregistered the network namespace
handling as the latter destroys connections which might be using the key.

David