aes128gcm: why verify padding?

draft-ietf-httpbis-encryption-encoding<https://github.com/httpwg/http-extensions/blob/master/draft-ietf-httpbis-encryption-encoding.md> allows any amount of padding to be added alongside the content being encrypted. The spec says the receiver (decryptor) MUST confirm that every padding byte is 0x00. Why?

I'm not sure how it helps to check the padding. It doesn't make the encryption deterministic as there is a salt involved, and the length of padding can also vary. It isn't useful for integrity as that is provided by the GCM tag.

Checking the padding could be harmful. Padding (to a block size) has caused lots of security vulnerabilities when the padding is checked before doing a proper integrity check. AEAD algorithms (such as GCM) theoretically solve this, but only when receivers process a whole AEAD decryption in one go, not via a (very common) streaming API.

The aes128gcm Content-Encoding adds "app-layer" padding, but by putting it first with a fixed format it is almost tempting receivers to check it before processing the rest of the ciphertext, and hence introduce vulnerabilities.

3 possible improvements:


1.       Allow padding bytes to have any value. The receiver just skips them.


2.       Calculate the padding length by reducing a 2-byte unsigned integer modulo the maximum possible padding length (plaintext length - 2). That means every 2-byte value will give a valid padding length.



3.       Put the 2-byte padding length field at the end. So the receiver has to have a whole AEAD plaintext block before knowing where the content starts.


A plaintext block would be: plaintext = padding || content || padlen
Len(plaintext) is the record size - 16.
Len(padlen) is 2 bytes.
Len(padding) = padlen % (Len(plaintext) - 2)
Padding bytes have any value.
Content is the remainder of the plaintext block.

The nice feature of this is that every plaintext block is valid. The last 2 bytes can be anything; mod (Len(plaintext) - 2) always gives a valid padding length; padding can be anything; and all that is left is the content to pass to the next layer. With this arrangement there is very little risk that plaintext structure is dangerously checked before performing the AEAD integrity check.

This form is no harder that the current structure to process, but it does encourage proper use of AEAD APIs.

--
James Manger