The implementation of AES-GCM in #1473 is much slower than the original version, even when the authentication tag is not being computed. This PR modifies the code in gcm() in order to significantly speed up the encryption/decryption time (when authentication tags are not being computed).
This PR adds processing of the Finished messages in TLS 1.3 handshakes. It also addresses some shellcheck issues.
If in debug mode, the HMAC of the transcript hash of the handshake context ($msg_transcript) is computed and compared against the Finished message sent by the server.
If the full server response is parsed and the connection with the server is not to be closed when tls_sockets() completes, then the TLS 1.3 handshake is completed by creating the client Finished message and sending it to the server.
This PR reorganizes the code for deriving TLS 1.3 symmetric keys in order to facilitate implementing the full key schedule. For example, rather than having a single function to derive the handshake traffic keys, this PR creates one function to derive the handshake secret and a separate function to derive the handshake traffic keys. The second function has been generalized so that it can derive either client or server traffic keys. Separating into two functions also makes the handshake_secret available for later use to derive the master secret and then the application traffic secrets and the application traffic keys.
This PR also changes where there message transcript is created, a message transcript will also be needed to derive the application traffic secrets. This PR includes the code to add the messages to the initial message transcript that will be needed for the input to the application traffic secret derivation function.
RFC 8446 specifies cipher suites that use three symmetric encryption algorithms, all of which are Authenticated Encryption with Associated Data (AEAD) algorithms. In each of these algorithms when data is encryption an authentication tag is created, which allows the recipient to verify that the data has not been modified. The authentication may also cover some additional data that was not encrypted.
The current implementations of these algorithms in testssl.sh decrypt the ciphertext, but do not check that the authentication tag is correct (which involves the recipient computing the correct tag for the received data and then comparing it to the provided tag). While testssl.sh can get away with not checking authentication tags when receiving data, the ability to compute authentication tags is needed in order to send encrypted data as TLS servers would reject any encrypted data that did not have a correct authentication tag. Being able to send encrypted data is necessary to be able to complete the TLS 1.3 handshake.
This PR replaces the current implementations of the symmetric encryption algorithms with full implementations of each of the algorithms. These full implementations include the ability to encrypt data for sending, and can also verify the authentication tag when decrypting data. Since the Bash implementations of these algorithms is very slow, the decryption code is designed to only compute and check authentication tags in debug mode.
While the implementation of the code to compute authentication tags for AES-CCM was based on NIST Special Publication 800-38C, I was not able to implement the code for AES-GCM or Poly1305 from their specifications (NIST Special Publication 800-38D and RFC 8439, respectively). So, I would very much like to thank the implementers of https://github.com/mko-x/SharedAES-GCM and https://github.com/floodyberry/poly1305-donna. The implementations of AES-GCM and Poly1305 in the PR were developed by translating the C code in https://github.com/mko-x/SharedAES-GCM and https://github.com/floodyberry/poly1305-donna into Bash. I don't understand what that code is doing, but it seems to work. :-)
I have only tested this code on a computer with a 64-bit operating system. While I have not tested it, I believe that the decryption code will work with 32-bit integers if not in debug mode (i.e., if not trying to compute the authentication tags). I also believe that the AES-CCM code for computing authentication tags will work with 32-bit integers. However, AES-GCM and Poly1305 code for computing authentication tags will definitely only work on systems that have 64-bit integers. So, on systems that do not have 64-bit integers, encryption will not work for AES-GCM or ChaCha20-Poly1305, and decryption will not work for these algorithms if in debug mode.
For a fresh start it seemed a good idea to cleanup
the order of functions and some variables so that
those with the same functionality are somewhat grouped.
Some of the functions have now a header and a foooter
to make it easier to spot and use then. Also for added future
functions the hope is that they will be put where they better
fit
This PR simplifies the code for determining which draft version of TLS 1.3 a server is offering by making use of a simple regular expression and $BASH_REMATCH rather than looping through every possible draft version.
PR #1463 changed run_ssl_poodle() to only run the test if it is known that the server supports SSLv3. However, support for SSLv3 may be unknown at the time run_ssl_poodle() is run (e.g., if the server supports TLS 1 and SSLv3, and run_ssl_poodle() is the first test performed). So, run_ssl_poodle() should perform testing unless it is known that SSLv3 is not supported.
There's a check for >825 days certificate lifetime. That
check emits a debug statement when the lifetime is within
this limit. It does that also when the certificate expired.
This commit adds now the word "total"
DEBUG: all is fine with total certificate life time
to make sure the life time left not is what should be understood.
... by adding the formerly intruoced "DEBUG" statement as a filter.
Note: "DEBUG" can now / should now be taken preferably for extra
output on debug level 1.
Replacing badssl.com by testssl.net. The former needed almost 5 min
for a run, whereas one IP of testssl.net needs ~80 secs. With --fast
even less.
Several vulnerability checks add a time penalty when the server
side only support TLS 1.3 as The TLS 1.3 RFC 8446 and implementations
known so far don't support the flaws being checked for.
This PR adds "shortcut" checks for all TLS 1.3, assuming that the
TLS 1.3 implementation is correct which seems at this time a valid
assumpution. That either saves a TCP connect or at least some logic to
be executed. Also in some cases a TLS 1.3 only server emitted unnecessary
warnings, see #1444.
If $DEBUG -eq 1 then it outputs information that a shortcut was
used. It doesn't do that in other cases because the screen output
seems too obtrusive.
It also adds a shortcut for beast when SSL 3 or TLS 1.0 is is known
not to be supported.
This commit radds 747fb039ed which
was accidenially reverted in 45f28d8166.
It fixes#1462.
See also #1459.
Dirk Wetter
David Cooper
Stu Tomlinson