stud
is a network proxy that terminates TLS/SSL connections and forwards the
unencrypted traffic to some backend. It's designed to handle 10s of thousands of
connections efficiently on multicore machines.
It follows a process-per-core model; a parent process spawns N children who
each accept()
on a common socket to distribute connected clients among them.
Within each child, asynchronous socket I/O is conducted across the local
connections using libev
and OpenSSL
's nonblocking API. By default,
stud
has an overhead of ~200KB per connection--it preallocates
some buffer space for data in flight between frontend and backend.
stud
has very few features--it's designed to be paired with an intelligent
backend like haproxy or nginx. It maintains a strict 1:1 connection pattern
with this backend handler so that the backend can dictate throttling behavior,
maxmium connection behavior, availability of service, etc.
stud
will optionally write the client IP address as the first few octets
(depending on IPv4 or IPv6) to the backend--or provide that information
using HAProxy's PROXY protocol. In this way, backends who care about the
client IP can still access it even though stud
itself appears to be the
connected client.
Thanks to a contribution from Emeric at Exceliance (the folks behind HAProxy),
a special build of stud
can be made that utilitizes shared memory to
use a common session cache between all child processes. This can speed up
large stud
deployments by avoiding client renegotiation.
Please be aware of the policy regarding releases, code stability, and security:
- In git, the tip of the master branch should always build on Linux and FreeBSD, and is likely to be as stable as any other changeset. A careful review of patches is conducted before being pushed to github.
- Periodically, a version tag will be pushed to github for an old(er)
changeset--0.1, 0.2, etc. These tags mark a particular release of
stud
that has seen heavy testing and several weeks of production stability. Conservative users are advised to use a tag. stud
has an optional build that utilizes shared memory-based SSL contexts and UDP peer communication to keep a session cache between many child processes running on many machines. The use of this build can dramatically speed up SSL handshakes on many-core and/or clustered deployments. However, it's important to acknowledge the inevitable theoretical security tradeoffs associated with the use of this (substantially more complex) binary. Therefore, the deeply paranoid are advised to use only the standardstud
binary at the cost of some performance.
stud
requires:
libev >= 4
openssl (recent, >=1.0.0 recommended)
Stud currently works on Linux, OpenBSD, FreeBSD, and MacOSX. It has been tested the most heavily on Linux/x86_64.
While porting it to other POSIX platforms is likely trivial, it hasn't be done yet. Patches welcome!
If you're handling a large number of connections, you'll
probably want to raise ulimit -n
before running stud
.
It's very strongly recommended to not run stud
as root; ideally, it would
be run as a user ("stud", perhaps) that does nothing but run stud
. Stud
will setuid (using -u) after binding if you need to bind to a low port (< 1024).
To install stud
:
$ make
$ sudo make install
The only required argument is a path to a PEM file that contains the certificate (or a chain of certificates) and private key.
Detail about the entire set of options can be found by invoking stud -h
:
CONFIGURATION:
--config=FILE Load configuration from specified file.
--default-config Prints default configuration to stdout.
ENCRYPTION METHODS:
--tls TLSv1 (default)
--ssl SSLv3 (implies no TLSv1)
-c --ciphers=SUITE Sets allowed ciphers (Default: "")
-e --ssl-engine=NAME Sets OpenSSL engine (Default: "")
-O --prefer-server-ciphers Prefer server list order
SOCKET:
-b --backend=HOST,PORT Backend [connect] (default is "[127.0.0.1]:8000")
-f --frontend=HOST,PORT Frontend [bind] (default is "[*]:8443")
PERFORMANCE:
-n --workers=NUM Number of worker processes (Default: 1)
-B --backlog=NUM Set listen backlog size (Default: 100)
-k --keepalive=SECS TCP keepalive on client socket (Default: 3600)
SECURITY:
-r --chroot=DIR Sets chroot directory (Default: "")
-u --user=USER Set uid/gid after binding the socket (Default: "")
-g --group=GROUP Set gid after binding the socket (Default: "")
LOGGING:
-q --quiet Be quiet; emit only error messages
-s --syslog Send log message to syslog in addition to stderr/stdout
--syslog-facility=FACILITY Syslog facility to use (Default: "daemon")
OTHER OPTIONS:
--daemon Fork into background and become a daemon (Default: off)
--write-ip Write 1 octet with the IP family followed by the IP
address in 4 (IPv4) or 16 (IPv6) octets little-endian
to backend before the actual data
(Default: off)
--write-proxy Write HaProxy's PROXY (IPv4 or IPv6) protocol line
before actual data
(Default: off)
-t --test Test configuration and exit
-V --version Print program version and exit
-h --help This help message
Stud can also use a configuration file that supports all the same options as the
command-line arguments. You can use stud --default-config
to
generate the default configuration on stdout; then, customize your configuration and
pass it to stud --config=FILE
.
If you're using stud
for HTTPS, please make sure to use the --ssl
option!
To use DH with stud, you will need to add some bytes to your pem file:
% openssl dhparam -rand - 1024 >> PEMFILE
Be sure to set your cipher suite appropriately: -c DHE-RSA-AES256-SHA
stud
was originally written by Jamie Turner (@jamwt) and is maintained
by the Bump (http://bu.mp) server team. It currently (12/11) provides
server-side TLS termination for over 85 million Bump users.
Special thanks to Colin Percival (@cperciva) for an early security audit and code review.
Finally, thank you to all the stud contributors, who have taken the program from a good start to a solid project: