Currently, cancel request key is a 32-bit token, which isn't very much
entropy. If you want to cancel another session's query, you can
brute-force it. In most environments, an unauthorized cancellation of a
query isn't very serious, but it nevertheless would be nice to have more
protection from it. The attached patch makes it longer. It is an
optional protocol feature, so it's fully backwards-compatible with
clients that don't support longer keys.
If the client requests the "_pq_.extended_query_cancel" protocol
feature, the server will generate a longer 256-bit cancellation key.
However, the new longer key length is no longer hardcoded in the
protocol. The client is expected to deal with variable length keys, up
to some reasonable upper limit (TODO: document the maximum). This
flexibility allows e.g. a connection pooler to add more information to
the cancel key, which could be useful. If the client doesn't request the
protocol feature, the server generates a 32-bit key like before.
One complication with this was that because we no longer know how long
the key should be, 4-bytes or something longer, until the backend has
performed the protocol negotiation, we cannot generate the key in the
postmaster before forking the process anymore. The first patch here
changes things so that the cancellation key is generated later, in the
backend, and the backend advertises the key in the PMSignalState array.
This is similar to how this has always worked in EXEC_BACKEND mode with
the ShmemBackendArray, but instead of having a separate array, I added
fields to the PMSignalState slots. This removes a bunch of
EXEC_BACKEND-specific code, which is nice.
Any thoughts on this? Documentation is still missing, and there's one
TODO on adding a portable time-constant memcmp() function; I'll add
those if there's agreement on this otherwise.
--
Heikki Linnakangas
Neon (https://neon.tech)