On 19.01.23 21:48, Jacob Champion wrote:
I like the existing "caveats" documentation, and I've attached a sample
patch with some more caveats documented, based on some of the upthread
conversation:
- text format makes fixed-length columns leak length information too
- you only get partial protection against the Evil DBA
- RSA-OAEP public key safety
(Feel free to use/remix/discard as desired.)
I have added those in the v15 patch I just posted.
When writing the paragraph on RSA-OAEP I was reminded that we didn't
really dig into the asymmetric/symmetric discussion. Assuming that most
first-time users will pick the builtin CMK encryption method, do we
still want to have an asymmetric scheme implemented first instead of a
symmetric keywrap? I'm still concerned about that public key, since it
can't really be made public.
I had started coding that, but one problem was that the openssl CLI
doesn't really provide any means to work with those kinds of keys. The
"openssl enc" command always wants to mix in a password. Without that,
there is no way to write a test case, and more crucially no way for
users to set up these kinds of keys. Unless we write our own tooling
for this, which, you know, the patch just passed 400k in size.
For the padding caveat:
+ There is no concern if all values are of the same length (e.g., credit
+ card numbers).
I nodded along to this statement last year, and then this year I learned
that CCNs aren't fixed-length. So with a 16-byte block, you're probably
going to be able to figure out who has an American Express card.
Heh. I have removed that parenthetical remark.
The column encryption algorithm is set per-column -- but isn't it
tightly coupled to the CEK, since the key length has to match? From a
layperson perspective, using the same key to encrypt the same plaintext
under two different algorithms (if they happen to have the same key
length) seems like it might be cryptographically risky. Is there a
reason I should be encouraged to do that?
Not really. I was also initially confused by this setup, but that's how
other similar systems are set up, so I thought it would be confusing to
do it differently.
With the loss of \gencr it looks like we also lost a potential way to
force encryption from within psql. Any plans to add that for v1?
\gencr didn't do that either. We could do it. The libpq API supports
it. We just need to come up with some syntax for psql.
While testing, I forgot how the new option worked and connected with
`column_encryption=on` -- and then I accidentally sent unencrypted data
to the server, since `on` means "not enabled". :( The server errors out
after the damage is done, of course, but would it be okay to strictly
validate that option's values?
fixed in v15
Are there plans to document client-side implementation requirements, to
ensure cross-client compatibility? Things like the "PG\x00\x01"
associated data are buried at the moment (or else I've missed them in
the docs). If you're holding off until the feature is more finalized,
that's fine too.
This is documented in the protocol chapter, which I thought was the
right place. Did you want more documentation, or in a different place?
Speaking of cross-client compatibility, I'm still disconcerted by the
ability to write the value "hello world" into an encrypted integer
column. Should clients be required to validate the text format, using
the attrealtypid?
Well, we can ask them to, but we can't really require them, in a
cryptographic sense. I'm not sure what more we can do.
It occurred to me when looking at the "unspecified" CMK scheme that the
CEK doesn't really have to be an encryption key at all. In that case it
can function more like a (possibly signed?) cookie for lookup, or even
be ignored altogether if you don't want to use a wrapping scheme
(similar to JWE's "direct" mode, maybe?). So now you have three ways to
look up or determine a column encryption key (CMK realm, CMK name, CEK
cookie)... is that a concept worth exploring in v1 and/or the documentation?
I don't completely follow this.
