@node Transport
+@cindex Transport
+@cindex Transport protocol
+@cindex Salsa20
+@cindex PRP
+@cindex Nonce
+@cindex Poly1305
+@cindex XTEA
+@cindex Serial
@section Transport protocol
@verbatim
-[PktLen] + ENCn(SERIAL) + ENC(KEY, ENCn(SERIAL), DATA_SIZE+DATA+NOISE) +
- AUTH(ENCn(SERIAL) + ENC(KEY, ENCn(SERIAL), DATA_SIZE+DATA+NOISE))
+TAG || ENCRYPTED || NONCE --> PACKET
+ ^ ^ ^
+ | | |
+ | | +-------------+
+ | | |
+ | +-------------+ |
+ | | |
+ +--< AUTH(AUTH_KEY, ENCRYPTED || NONCE)
+ ^ ^
+ | |
++------------------------+ |
+| |
+| +---------------+
+| |
++--< ENCRYPT(KEY, NONCE, PAYLOAD)
+ ^ ^
+ | |
+ | +--< DATA || PAD [|| ZEROS]
+ |
+ +--< PRP(PRP_KEY, SERIAL)
@end verbatim
-All transport and handshake messages are indistinguishable from
-pseudo random noise, except when using TCP connections.
-
-@code{PktLen} is used only with TCP connections. It is big-endian
-@emph{uint16} length of the whole packet (except @code{PktLen} itself).
-
@code{SERIAL} is message's serial number. Odds are reserved for
-client(->server) messages, evens for server(->client) messages.
+client (to server) messages, evens for server (to client) messages.
-@code{ENCn} is XTEA block cipher algorithm used here as PRP (pseudo
+@code{PRP} is XTEA block cipher algorithm used here as PRP (pseudo
random permutation function) to obfuscate @code{SERIAL}. Plaintext
@code{SERIAL} state is kept in peers internal state, but encrypted
-before transmission. XTEA is compact and fast enough. Salsa20 is PRF
-function and requires much more code to create PRP from it.
+before transmission.
-XTEA's encryption key is the first 128-bit of Salsa20's output with
-established common key and zero nonce (message nonces start from 1).
+XTEA's encryption key @code{PRP_KEY} is the first 128-bit of Salsa20's
+output with established common key and zero nonce (message nonces start
+from 1).
-@code{ENC} is Salsa20 stream cipher, with established session @code{KEY}
-and obfuscated @code{SERIAL} used as a nonce. First 256 bits of
-Salsa20's output is used as Poly1305 authentication key, next 256 bits
-are ignored. All remaining output is XORed with the data, encrypting it.
+@verbatim
+PRP_KEY = 128bit(ENCRYPT(KEY, 0))
+@end verbatim
-@code{DATA_SIZE} is big-endian @emph{uint16} storing length of the
-@code{DATA}.
+@code{ENCRYPT} is Salsa20 stream cipher, with established session
+@code{KEY} and obfuscated @code{SERIAL} used as a nonce. 512 bit of
+Salsa20's output is ignored and only remaining is XORed with ther data,
+encrypting it.
-@code{NOISE} is optional. It is just some junk data, intended to fill up
-packet to MTU size. This is useful for concealing payload packets length.
+@code{DATA} is padded with @code{PAD} (0x80 byte). Optional @code{ZEROS}
+may follow, to fill up packet to conceal payload packet length.
@code{AUTH} is Poly1305 authentication function. First 256 bits of
-Salsa20 output are used as a one-time key for @code{AUTH}.
+Salsa20's output are used as a one-time key for @code{AUTH}.
+
+@verbatim
+AUTH_KEY = 256bit(ENCRYPT(KEY, NONCE))
+@end verbatim
To prevent replay attacks we must remember received @code{SERIAL}s and
-if meet one, then drop it. Basically we could just store latest number
-and check if received one is greater, but because of UDP packets
-reordering this can lead to valid packets dropping and overall
-performance degradation. We store up to 256 seen nonces in hash
-structure, in two swapping buckets.
+drop when receiving duplicate ones.
+
+In @ref{Encless, encryptionless mode} this scheme is slightly different:
+
+@verbatim
+ PACKET = ENCODED || NONCE
+ENCODED = ENCLESS(DATA || PAD || ZEROS)
+ NONCE = PRP(PRP_KEY, SERIAL)
+@end verbatim
+
+@code{ENCLESS} is AONT and chaffing function. There is no need in
+explicit separate authentication.