Re: [quicwg/base-drafts] token-based greasing / initial packet protection (#3166)

[David Schinazi <[e-mail]>]

DavidSchinazi commented on this pull request.

Thanks for writing this up! I like the token approach but we might need to tighten how this interacts with version negotiation.

> @@ -2721,6 +2721,66 @@ between endpoints.  Application protocols SHOULD define rules for handling
 streams that are prematurely cancelled by either endpoint.
 
 
+# Alternative Initial Set {#alternative-initial}
+
+In order to avoid ossification of the cleartext and obfuscated fields of QUIC
+packets, a server can announce an alternative set of initial values to be used,
+which is comprised of:
+
+* Alternative version number; a 32-bit unsigned number that is to be presented
+  on wire in place of the version number specified in this document.  This value
+  MUST NOT be a reserved version ({{versions}}).
+
+* Packet type modifier; a two-bit value that is to be applied as a bit-wise
+  exclusive or (XOR) to the most significant bits of the Initial, Handshake,

Shouldn't this be `to the Long Packet Type field of` instead of `to the most significant bits of` ?

> +  0-RTT, and Retry packets. This XOR is applied after the packets are encrypted
+  and before they are decrypted.
+
+* Alternative initial salt; a 16-byte binary blob that is to be used in place of
+  the initial salt defined in section 5.2 of {{QUIC-TLS}}.
+
+A server advertises these values using a NEW_TOKEN frame {{frame-new-token}}.
+The token MUST include or associated with the alternative version number with
+which it can be used.
+
+Typically, a server would pre-allocate a set of unused version numbers as the
+alternative version numbers, associating each of those version numbers with a
+packet type modifier chosen at random.  Then, when issuing a token using a
+NEW_TOKEN frame, the server generates the alternative initial salt by calling a
+pseudo-random function, embeds that initial salt into the token which is then
+encrypted, and sends a NEW_TOKEN frame that comprises of the generated token and

I might be worth adding a bit more detail to how the token is encrypted, perhaps `which is then encrypted using a key only known to the server`

> +NEW_TOKEN frame, the server generates the alternative initial salt by calling a
+pseudo-random function, embeds that initial salt into the token which is then
+encrypted, and sends a NEW_TOKEN frame that comprises of the generated token and
+the alternative initial set.
+
+When the client reconnects to the server by using the provided token and the
+alternative initial set, the server first checks if the version number field of
+the incoming packet contains one of the alternative version numbers it
+advertises, then if that is the case, applies the corresponding packet type
+modifier to recover the correct packet type.  If the recovered packet type is an
+Initial packet and that packet contains a NEW_TOKEN token, the server decrypts
+the embedded token and recovers the alternative initial salt, uses that to
+decrypt the payload of the Initial packet.
+
+When the server is incapable of determining the alternative initial salt, it can
+send a Version Negotiation packet that instructs the client to use the default

In the current spec, the client fails the connection upon receipt of a Version Negotiation packet. So this has a dependency on the version downgrade prevention extension.

But assuming the presence of a version negotiation mechanism with downgrade prevention, I think we have an issue here. If the server rotates the key it uses to encrypt the alternative initial salt, and then receives a token it can no longer decrypt, it sends a version negotiation packet, and the client tries again with QUICv1. This is likely to trip the version downgrade prevention mechanism - how do we distinguish between a server that has forgotten their key and an attacker that is trying to force a downgrade to QUICv1?

> @@ -3940,6 +4007,19 @@ described in {{QUIC-TLS}}.  This protection does not provide confidentiality or
 integrity against on-path attackers, but provides some level of protection
 against off-path attackers.
 
+Additionally, the token is accompanied by a checksum.  This is because when a

Seeing a checksum here makes me uncomfortable. We're adding a checksum to retry packets because they have no other mechanism to detect corruption. Initial packets on the other hand have an auth tag, so a corrupted token will cause the deprotection to fail. In an ideal world we should handle that like any other corruption on the initial packet. But I think this goes back to my other question about version negotiation, I think there be dragons.

-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/pull/3166#pullrequestreview-308689911