Pike SSL/TLS Implementation Notes
SSL/TLS Standard Documents and Amount of Coverage
SSL and TLS are specified in quite a few documents;
the following is an attempt to list them all, and
the level of implementation in Pike.
[ ] Not implemented at present.
[-] Not implemented and will not be implemented.
Typically an obsolete or insecure standard.
[/] Partially implemented.
[X] Fully implemented.
NB: Constants from documents marked as not implemented may
still be added to SSL.Constants for debug purposes.
The SSL 2.0 protocol was specified in the following document:
[-] SSL 2.0 draft-hickman-netscape-ssl-00.txt
The SSL 3.0 Protocol was specified in the following draft (later an RFC):
[X] SSL 3.0 draft-freier-ssl-version3-02.txt
[X] SSL 3.0 RFC 6101
The TLS 1.0 Protocol is specified in the following RFCs:
[X] SSL 3.1/TLS 1.0 RFC 2246
[/] Kerberos for TLS 1.0 RFC 2712
[X] AES Ciphers for TLS 1.0 RFC 3268
[/] Extensions for TLS 1.0 RFC 3546
[X] TLS Compression Methods RFC 3749
[ ] LZS Compression for TLS RFC 3943
[X] Camellia Cipher for TLS RFC 4132
[ ] SEED Cipher for TLS 1.0 RFC 4162
[X] Pre-Shared Keys for TLS RFC 4279
The TLS 1.1 Protocol is specified in the following RFCs:
[X] SSL 3.2/TLS 1.1 RFC 4346
[/] Extensions for TLS 1.1 RFC 4366
[X] ECC Ciphers for TLS 1.1 RFC 4492
[ ] Session Resumption RFC 4507
[ ] TLS Handshake Message RFC 4680
[ ] User Mapping Extension RFC 4681
[X] PSK with NULL for TLS 1.1 RFC 4785
[ ] SRP with TLS 1.1 RFC 5054
[ ] Session Resumption RFC 5077
[ ] OpenPGP Authentication RFC 5081
[X] Authenticated Encryption RFC 5116
The DTLS 1.0 Protocol is specified in the following RFCs:
[ ] DTLS 1.0 RFC 4347
[ ] DTLS over DCCP RFC 5238
The TLS 1.2 Protocol is specified in the following RFCs:
[X] SSL 3.3/TLS 1.2 RFC 5246
[X] AES GCM Cipher for TLS RFC 5288
[X] ECC with SHA256/384 & GCM RFC 5289
[/] Suite B Profile for TLS RFC 5430
[X] DES and IDEA for TLS RFC 5469
[X] Pre-Shared Keys with GCM RFC 5487
[X] ECDHE_PSK Cipher for TLS RFC 5489
[ ] Renegotiation Extension RFC 5746
[ ] Authorization Extensions RFC 5878
[X] Camellia Cipher for TLS RFC 5932
[ ] KeyNote Auth for TLS RFC 6042
[ ] TLS Extension Definitions RFC 6066
[ ] OpenPGP Authentication RFC 6091
[ ] ARIA Cipher for TLS RFC 6209
[ ] Additional Master Secrets RFC 6358
[X] Camellia Cipher for TLS RFC 6367
[/] Suite B Profile for TLS RFC 6460
[X] Heartbeat Extension RFC 6520
[X] AES-CCM Cipher for TLS RFC 6655
[ ] Multiple Certificates RFC 6961
[ ] Certificate Transparency RFC 6962
[ ] ECC Brainpool Curves RFC 7027
[ ] Raw Public Keys in (D)TLS RFC 7250
[X] AES-CCM ECC Suites for TLS RFC 7251
[X] TLS ALPN Extension RFC 7301
[X] TLS Encrypt-then-MAC RFC 7366
[ ] Known Attacks on TLS and DTLS RFC 7457
[X] Prohibit RC4 RFC 7465
[X] TLS Fallback SCSV RFC 7507
[ ] Recommendations for TLS and DTLS TFC 7525
[/] Session Hash and Extended MS RFC 7627
[X] TLS Padding RFC 7685
The DTLS 1.2 Protocol is specified in the following RFCs:
[ ] DTLS 1.2 RFC 6347
Drafts (in order of age, oldest first):
[/] 56-bit Export Cipher draft-ietf-tls-56-bit-ciphersuites-01.txt
[-] Next Protocol Negotiation draft-agl-tls-nextprotoneg
[ ] Chacha20Poly1305 draft-agl-tls-chacha20poly1305-04.txt
[X] Negotiated FF-DHE Parameters draft-ietf-tls-negotiated-ff-dhe
[/] SSL 3.4/TLS 1.3 draft-ietf-tls-tls13-06.txt
The TLS parameters registry:
SSL.File and Stdio.File Emulation Overview
The underlying stream object is always in nonblocking
mode, to avoid risk of hanging in Stdio.Buffer.
All I/O-ops are always buffered with Stdio.Buffer.
Internal Callback Handling in Nonblocking Mode:
In nonblocking mode all internal callback handling is
performed directly with the real_backend.
If the main backend has been started (ie master()->asyncp()
is true), we assume that it will handle I/O.
Otherwise if nonthreaded or we are on the backend thread
(master()->backend_thread()), we rotate the real_backend
once with 0.0 timeout per I/O-op.
Note that this may cause problems when using custom
backends without having started the main backend.
Internal Callback Handling in Blocking Mode:
In blocking mode all internal callback handling is
performed with a dedicated local_backend.
The local_backend is created when the SSL.File is
switched to blocking mode.
The local_backend is then rotated until the blocking
call is done.
User Callback Handling:
All user installed callbacks are called via call_out()
on internal_poll() in the real_backend.
Get data from user_read_buffer, install ssl_read_callback
on underflow. In blocking mode rotate the local backend
until all data is available. In nonblocking mode
attempt to rotate the local backend once if no callbacks
Fill write_buffer, install ssl_write_callback.
In blocking mode rotate the local backend until
the write_buffer is empty. In nonblocking mode
attempt to rotate the local backend once if
no callbacks are installed.
Schedule a close packet, and block further calls to
write. If both directions block also further calls
to read. Install both ssl_write_callback and
ssl_read_callback and rotate the local backend until
connection closed from other end, or linger time expires.
Clear user callbacks and switch to nonblocking mode.
Attempt to send a close packet. Terminate the
Decode received data and add it to user_read_buffer.
Schedule read_callback with real_backend call_out.
Uninstall on user_read_buffer full.
Send data from write_buffer, uninstall on write_buffer empty.
Schedule write_callback with real_backend call_out.
On send failure, block futher calls of write.
Schedule close_callback if close() has not been called yet.
Block further calls of read.
As ssl_close_callback, but allow use of stream when done.
Known Problems and Missing Features
Nonblocking mode without callbacks.
Support for set_buffer_mode() et al not yet supported,
neither directly nor in the embedded stream.
Session objects should be possible to serialize with
encode_value() to allow multiple frontend nodes to share the
session cache. (Overloading the session cache functionality in
Context is already easy to do)
The handshake message hash should be streaming and discard raw
data after each packet. Also, it is probably possible to find
only one place in the code where data can be fed to the
Currently Pike will always try to maximize the number of bits
used for certificates, key exchanges, cipher keys and
hashes. Another popular approach that should be supported is
to minimize the bits used, above the set threashold. The
rationale is that everything allowed is good enough and the
capability negotiation should optimize on consumed resources.
Servers with different certificates and parameters can sit on
the same port with different SNI, or with different ALPN. We
should make it possible to select Context based on negotiation
(made tricky, as the negotiation depends on the Context).
It should be possible to lazy load certificates to increase
startup time and reduce memory usage for servers with many
sites. Dynamic loading and unloaded could be part of the same
mechanism as the Context selection mentioned above.
Truncated HMAC is only supported on the server side. It has
however been mentioned on the IETF TLS mailing list that there
are security issues with truncated HMAC, so this is only
lacking for completeness.